arc.c revision 287665
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; 200int zfs_arc_grow_retry = 0; 201int zfs_arc_shrink_shift = 0; 202int zfs_arc_p_min_shift = 0; 203int zfs_disable_dup_eviction = 0; 204uint64_t zfs_arc_average_blocksize = 8 * 1024; /* 8KB */ 205u_int zfs_arc_free_target = 0; 206 207static int sysctl_vfs_zfs_arc_free_target(SYSCTL_HANDLER_ARGS); 208static int sysctl_vfs_zfs_arc_meta_limit(SYSCTL_HANDLER_ARGS); 209 210#ifdef _KERNEL 211static void 212arc_free_target_init(void *unused __unused) 213{ 214 215 zfs_arc_free_target = vm_pageout_wakeup_thresh; 216} 217SYSINIT(arc_free_target_init, SI_SUB_KTHREAD_PAGE, SI_ORDER_ANY, 218 arc_free_target_init, NULL); 219 220TUNABLE_QUAD("vfs.zfs.arc_max", &zfs_arc_max); 221TUNABLE_QUAD("vfs.zfs.arc_min", &zfs_arc_min); 222TUNABLE_QUAD("vfs.zfs.arc_meta_limit", &zfs_arc_meta_limit); 223TUNABLE_QUAD("vfs.zfs.arc_average_blocksize", &zfs_arc_average_blocksize); 224TUNABLE_INT("vfs.zfs.arc_shrink_shift", &zfs_arc_shrink_shift); 225SYSCTL_DECL(_vfs_zfs); 226SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_max, CTLFLAG_RDTUN, &zfs_arc_max, 0, 227 "Maximum ARC size"); 228SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_min, CTLFLAG_RDTUN, &zfs_arc_min, 0, 229 "Minimum ARC size"); 230SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_average_blocksize, CTLFLAG_RDTUN, 231 &zfs_arc_average_blocksize, 0, 232 "ARC average blocksize"); 233SYSCTL_INT(_vfs_zfs, OID_AUTO, arc_shrink_shift, CTLFLAG_RW, 234 &arc_shrink_shift, 0, 235 "log2(fraction of arc to reclaim)"); 236 237/* 238 * We don't have a tunable for arc_free_target due to the dependency on 239 * pagedaemon initialisation. 240 */ 241SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_free_target, 242 CTLTYPE_UINT | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, sizeof(u_int), 243 sysctl_vfs_zfs_arc_free_target, "IU", 244 "Desired number of free pages below which ARC triggers reclaim"); 245 246static int 247sysctl_vfs_zfs_arc_free_target(SYSCTL_HANDLER_ARGS) 248{ 249 u_int val; 250 int err; 251 252 val = zfs_arc_free_target; 253 err = sysctl_handle_int(oidp, &val, 0, req); 254 if (err != 0 || req->newptr == NULL) 255 return (err); 256 257 if (val < minfree) 258 return (EINVAL); 259 if (val > cnt.v_page_count) 260 return (EINVAL); 261 262 zfs_arc_free_target = val; 263 264 return (0); 265} 266 267/* 268 * Must be declared here, before the definition of corresponding kstat 269 * macro which uses the same names will confuse the compiler. 270 */ 271SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_meta_limit, 272 CTLTYPE_U64 | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, sizeof(uint64_t), 273 sysctl_vfs_zfs_arc_meta_limit, "QU", 274 "ARC metadata limit"); 275#endif 276 277/* 278 * Note that buffers can be in one of 6 states: 279 * ARC_anon - anonymous (discussed below) 280 * ARC_mru - recently used, currently cached 281 * ARC_mru_ghost - recentely used, no longer in cache 282 * ARC_mfu - frequently used, currently cached 283 * ARC_mfu_ghost - frequently used, no longer in cache 284 * ARC_l2c_only - exists in L2ARC but not other states 285 * When there are no active references to the buffer, they are 286 * are linked onto a list in one of these arc states. These are 287 * the only buffers that can be evicted or deleted. Within each 288 * state there are multiple lists, one for meta-data and one for 289 * non-meta-data. Meta-data (indirect blocks, blocks of dnodes, 290 * etc.) is tracked separately so that it can be managed more 291 * explicitly: favored over data, limited explicitly. 292 * 293 * Anonymous buffers are buffers that are not associated with 294 * a DVA. These are buffers that hold dirty block copies 295 * before they are written to stable storage. By definition, 296 * they are "ref'd" and are considered part of arc_mru 297 * that cannot be freed. Generally, they will aquire a DVA 298 * as they are written and migrate onto the arc_mru list. 299 * 300 * The ARC_l2c_only state is for buffers that are in the second 301 * level ARC but no longer in any of the ARC_m* lists. The second 302 * level ARC itself may also contain buffers that are in any of 303 * the ARC_m* states - meaning that a buffer can exist in two 304 * places. The reason for the ARC_l2c_only state is to keep the 305 * buffer header in the hash table, so that reads that hit the 306 * second level ARC benefit from these fast lookups. 307 */ 308 309#define ARCS_LOCK_PAD CACHE_LINE_SIZE 310struct arcs_lock { 311 kmutex_t arcs_lock; 312#ifdef _KERNEL 313 unsigned char pad[(ARCS_LOCK_PAD - sizeof (kmutex_t))]; 314#endif 315}; 316 317/* 318 * must be power of two for mask use to work 319 * 320 */ 321#define ARC_BUFC_NUMDATALISTS 16 322#define ARC_BUFC_NUMMETADATALISTS 16 323#define ARC_BUFC_NUMLISTS (ARC_BUFC_NUMMETADATALISTS + ARC_BUFC_NUMDATALISTS) 324 325typedef struct arc_state { 326 uint64_t arcs_lsize[ARC_BUFC_NUMTYPES]; /* amount of evictable data */ 327 uint64_t arcs_size; /* total amount of data in this state */ 328 list_t arcs_lists[ARC_BUFC_NUMLISTS]; /* list of evictable buffers */ 329 struct arcs_lock arcs_locks[ARC_BUFC_NUMLISTS] __aligned(CACHE_LINE_SIZE); 330} arc_state_t; 331 332#define ARCS_LOCK(s, i) (&((s)->arcs_locks[(i)].arcs_lock)) 333 334/* The 6 states: */ 335static arc_state_t ARC_anon; 336static arc_state_t ARC_mru; 337static arc_state_t ARC_mru_ghost; 338static arc_state_t ARC_mfu; 339static arc_state_t ARC_mfu_ghost; 340static arc_state_t ARC_l2c_only; 341 342typedef struct arc_stats { 343 kstat_named_t arcstat_hits; 344 kstat_named_t arcstat_misses; 345 kstat_named_t arcstat_demand_data_hits; 346 kstat_named_t arcstat_demand_data_misses; 347 kstat_named_t arcstat_demand_metadata_hits; 348 kstat_named_t arcstat_demand_metadata_misses; 349 kstat_named_t arcstat_prefetch_data_hits; 350 kstat_named_t arcstat_prefetch_data_misses; 351 kstat_named_t arcstat_prefetch_metadata_hits; 352 kstat_named_t arcstat_prefetch_metadata_misses; 353 kstat_named_t arcstat_mru_hits; 354 kstat_named_t arcstat_mru_ghost_hits; 355 kstat_named_t arcstat_mfu_hits; 356 kstat_named_t arcstat_mfu_ghost_hits; 357 kstat_named_t arcstat_allocated; 358 kstat_named_t arcstat_deleted; 359 kstat_named_t arcstat_stolen; 360 kstat_named_t arcstat_recycle_miss; 361 /* 362 * Number of buffers that could not be evicted because the hash lock 363 * was held by another thread. The lock may not necessarily be held 364 * by something using the same buffer, since hash locks are shared 365 * by multiple buffers. 366 */ 367 kstat_named_t arcstat_mutex_miss; 368 /* 369 * Number of buffers skipped because they have I/O in progress, are 370 * indrect prefetch buffers that have not lived long enough, or are 371 * not from the spa we're trying to evict from. 372 */ 373 kstat_named_t arcstat_evict_skip; 374 kstat_named_t arcstat_evict_l2_cached; 375 kstat_named_t arcstat_evict_l2_eligible; 376 kstat_named_t arcstat_evict_l2_ineligible; 377 kstat_named_t arcstat_hash_elements; 378 kstat_named_t arcstat_hash_elements_max; 379 kstat_named_t arcstat_hash_collisions; 380 kstat_named_t arcstat_hash_chains; 381 kstat_named_t arcstat_hash_chain_max; 382 kstat_named_t arcstat_p; 383 kstat_named_t arcstat_c; 384 kstat_named_t arcstat_c_min; 385 kstat_named_t arcstat_c_max; 386 kstat_named_t arcstat_size; 387 kstat_named_t arcstat_hdr_size; 388 kstat_named_t arcstat_data_size; 389 kstat_named_t arcstat_other_size; 390 kstat_named_t arcstat_l2_hits; 391 kstat_named_t arcstat_l2_misses; 392 kstat_named_t arcstat_l2_feeds; 393 kstat_named_t arcstat_l2_rw_clash; 394 kstat_named_t arcstat_l2_read_bytes; 395 kstat_named_t arcstat_l2_write_bytes; 396 kstat_named_t arcstat_l2_writes_sent; 397 kstat_named_t arcstat_l2_writes_done; 398 kstat_named_t arcstat_l2_writes_error; 399 kstat_named_t arcstat_l2_writes_hdr_miss; 400 kstat_named_t arcstat_l2_evict_lock_retry; 401 kstat_named_t arcstat_l2_evict_reading; 402 kstat_named_t arcstat_l2_free_on_write; 403 kstat_named_t arcstat_l2_cdata_free_on_write; 404 kstat_named_t arcstat_l2_abort_lowmem; 405 kstat_named_t arcstat_l2_cksum_bad; 406 kstat_named_t arcstat_l2_io_error; 407 kstat_named_t arcstat_l2_size; 408 kstat_named_t arcstat_l2_asize; 409 kstat_named_t arcstat_l2_hdr_size; 410 kstat_named_t arcstat_l2_compress_successes; 411 kstat_named_t arcstat_l2_compress_zeros; 412 kstat_named_t arcstat_l2_compress_failures; 413 kstat_named_t arcstat_l2_write_trylock_fail; 414 kstat_named_t arcstat_l2_write_passed_headroom; 415 kstat_named_t arcstat_l2_write_spa_mismatch; 416 kstat_named_t arcstat_l2_write_in_l2; 417 kstat_named_t arcstat_l2_write_hdr_io_in_progress; 418 kstat_named_t arcstat_l2_write_not_cacheable; 419 kstat_named_t arcstat_l2_write_full; 420 kstat_named_t arcstat_l2_write_buffer_iter; 421 kstat_named_t arcstat_l2_write_pios; 422 kstat_named_t arcstat_l2_write_buffer_bytes_scanned; 423 kstat_named_t arcstat_l2_write_buffer_list_iter; 424 kstat_named_t arcstat_l2_write_buffer_list_null_iter; 425 kstat_named_t arcstat_memory_throttle_count; 426 kstat_named_t arcstat_duplicate_buffers; 427 kstat_named_t arcstat_duplicate_buffers_size; 428 kstat_named_t arcstat_duplicate_reads; 429 kstat_named_t arcstat_meta_used; 430 kstat_named_t arcstat_meta_limit; 431 kstat_named_t arcstat_meta_max; 432} arc_stats_t; 433 434static arc_stats_t arc_stats = { 435 { "hits", KSTAT_DATA_UINT64 }, 436 { "misses", KSTAT_DATA_UINT64 }, 437 { "demand_data_hits", KSTAT_DATA_UINT64 }, 438 { "demand_data_misses", KSTAT_DATA_UINT64 }, 439 { "demand_metadata_hits", KSTAT_DATA_UINT64 }, 440 { "demand_metadata_misses", KSTAT_DATA_UINT64 }, 441 { "prefetch_data_hits", KSTAT_DATA_UINT64 }, 442 { "prefetch_data_misses", KSTAT_DATA_UINT64 }, 443 { "prefetch_metadata_hits", KSTAT_DATA_UINT64 }, 444 { "prefetch_metadata_misses", KSTAT_DATA_UINT64 }, 445 { "mru_hits", KSTAT_DATA_UINT64 }, 446 { "mru_ghost_hits", KSTAT_DATA_UINT64 }, 447 { "mfu_hits", KSTAT_DATA_UINT64 }, 448 { "mfu_ghost_hits", KSTAT_DATA_UINT64 }, 449 { "allocated", KSTAT_DATA_UINT64 }, 450 { "deleted", KSTAT_DATA_UINT64 }, 451 { "stolen", KSTAT_DATA_UINT64 }, 452 { "recycle_miss", KSTAT_DATA_UINT64 }, 453 { "mutex_miss", KSTAT_DATA_UINT64 }, 454 { "evict_skip", KSTAT_DATA_UINT64 }, 455 { "evict_l2_cached", KSTAT_DATA_UINT64 }, 456 { "evict_l2_eligible", KSTAT_DATA_UINT64 }, 457 { "evict_l2_ineligible", KSTAT_DATA_UINT64 }, 458 { "hash_elements", KSTAT_DATA_UINT64 }, 459 { "hash_elements_max", KSTAT_DATA_UINT64 }, 460 { "hash_collisions", KSTAT_DATA_UINT64 }, 461 { "hash_chains", KSTAT_DATA_UINT64 }, 462 { "hash_chain_max", KSTAT_DATA_UINT64 }, 463 { "p", KSTAT_DATA_UINT64 }, 464 { "c", KSTAT_DATA_UINT64 }, 465 { "c_min", KSTAT_DATA_UINT64 }, 466 { "c_max", KSTAT_DATA_UINT64 }, 467 { "size", KSTAT_DATA_UINT64 }, 468 { "hdr_size", KSTAT_DATA_UINT64 }, 469 { "data_size", KSTAT_DATA_UINT64 }, 470 { "other_size", KSTAT_DATA_UINT64 }, 471 { "l2_hits", KSTAT_DATA_UINT64 }, 472 { "l2_misses", KSTAT_DATA_UINT64 }, 473 { "l2_feeds", KSTAT_DATA_UINT64 }, 474 { "l2_rw_clash", KSTAT_DATA_UINT64 }, 475 { "l2_read_bytes", KSTAT_DATA_UINT64 }, 476 { "l2_write_bytes", KSTAT_DATA_UINT64 }, 477 { "l2_writes_sent", KSTAT_DATA_UINT64 }, 478 { "l2_writes_done", KSTAT_DATA_UINT64 }, 479 { "l2_writes_error", KSTAT_DATA_UINT64 }, 480 { "l2_writes_hdr_miss", KSTAT_DATA_UINT64 }, 481 { "l2_evict_lock_retry", KSTAT_DATA_UINT64 }, 482 { "l2_evict_reading", KSTAT_DATA_UINT64 }, 483 { "l2_free_on_write", KSTAT_DATA_UINT64 }, 484 { "l2_cdata_free_on_write", KSTAT_DATA_UINT64 }, 485 { "l2_abort_lowmem", KSTAT_DATA_UINT64 }, 486 { "l2_cksum_bad", KSTAT_DATA_UINT64 }, 487 { "l2_io_error", KSTAT_DATA_UINT64 }, 488 { "l2_size", KSTAT_DATA_UINT64 }, 489 { "l2_asize", KSTAT_DATA_UINT64 }, 490 { "l2_hdr_size", KSTAT_DATA_UINT64 }, 491 { "l2_compress_successes", KSTAT_DATA_UINT64 }, 492 { "l2_compress_zeros", KSTAT_DATA_UINT64 }, 493 { "l2_compress_failures", KSTAT_DATA_UINT64 }, 494 { "l2_write_trylock_fail", KSTAT_DATA_UINT64 }, 495 { "l2_write_passed_headroom", KSTAT_DATA_UINT64 }, 496 { "l2_write_spa_mismatch", KSTAT_DATA_UINT64 }, 497 { "l2_write_in_l2", KSTAT_DATA_UINT64 }, 498 { "l2_write_io_in_progress", KSTAT_DATA_UINT64 }, 499 { "l2_write_not_cacheable", KSTAT_DATA_UINT64 }, 500 { "l2_write_full", KSTAT_DATA_UINT64 }, 501 { "l2_write_buffer_iter", KSTAT_DATA_UINT64 }, 502 { "l2_write_pios", KSTAT_DATA_UINT64 }, 503 { "l2_write_buffer_bytes_scanned", KSTAT_DATA_UINT64 }, 504 { "l2_write_buffer_list_iter", KSTAT_DATA_UINT64 }, 505 { "l2_write_buffer_list_null_iter", KSTAT_DATA_UINT64 }, 506 { "memory_throttle_count", KSTAT_DATA_UINT64 }, 507 { "duplicate_buffers", KSTAT_DATA_UINT64 }, 508 { "duplicate_buffers_size", KSTAT_DATA_UINT64 }, 509 { "duplicate_reads", KSTAT_DATA_UINT64 }, 510 { "arc_meta_used", KSTAT_DATA_UINT64 }, 511 { "arc_meta_limit", KSTAT_DATA_UINT64 }, 512 { "arc_meta_max", KSTAT_DATA_UINT64 } 513}; 514 515#define ARCSTAT(stat) (arc_stats.stat.value.ui64) 516 517#define ARCSTAT_INCR(stat, val) \ 518 atomic_add_64(&arc_stats.stat.value.ui64, (val)) 519 520#define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1) 521#define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1) 522 523#define ARCSTAT_MAX(stat, val) { \ 524 uint64_t m; \ 525 while ((val) > (m = arc_stats.stat.value.ui64) && \ 526 (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \ 527 continue; \ 528} 529 530#define ARCSTAT_MAXSTAT(stat) \ 531 ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64) 532 533/* 534 * We define a macro to allow ARC hits/misses to be easily broken down by 535 * two separate conditions, giving a total of four different subtypes for 536 * each of hits and misses (so eight statistics total). 537 */ 538#define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \ 539 if (cond1) { \ 540 if (cond2) { \ 541 ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \ 542 } else { \ 543 ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \ 544 } \ 545 } else { \ 546 if (cond2) { \ 547 ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \ 548 } else { \ 549 ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\ 550 } \ 551 } 552 553kstat_t *arc_ksp; 554static arc_state_t *arc_anon; 555static arc_state_t *arc_mru; 556static arc_state_t *arc_mru_ghost; 557static arc_state_t *arc_mfu; 558static arc_state_t *arc_mfu_ghost; 559static arc_state_t *arc_l2c_only; 560 561/* 562 * There are several ARC variables that are critical to export as kstats -- 563 * but we don't want to have to grovel around in the kstat whenever we wish to 564 * manipulate them. For these variables, we therefore define them to be in 565 * terms of the statistic variable. This assures that we are not introducing 566 * the possibility of inconsistency by having shadow copies of the variables, 567 * while still allowing the code to be readable. 568 */ 569#define arc_size ARCSTAT(arcstat_size) /* actual total arc size */ 570#define arc_p ARCSTAT(arcstat_p) /* target size of MRU */ 571#define arc_c ARCSTAT(arcstat_c) /* target size of cache */ 572#define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */ 573#define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */ 574#define arc_meta_limit ARCSTAT(arcstat_meta_limit) /* max size for metadata */ 575#define arc_meta_used ARCSTAT(arcstat_meta_used) /* size of metadata */ 576#define arc_meta_max ARCSTAT(arcstat_meta_max) /* max size of metadata */ 577 578#define L2ARC_IS_VALID_COMPRESS(_c_) \ 579 ((_c_) == ZIO_COMPRESS_LZ4 || (_c_) == ZIO_COMPRESS_EMPTY) 580 581static int arc_no_grow; /* Don't try to grow cache size */ 582static uint64_t arc_tempreserve; 583static uint64_t arc_loaned_bytes; 584 585typedef struct l2arc_buf_hdr l2arc_buf_hdr_t; 586 587typedef struct arc_callback arc_callback_t; 588 589struct arc_callback { 590 void *acb_private; 591 arc_done_func_t *acb_done; 592 arc_buf_t *acb_buf; 593 zio_t *acb_zio_dummy; 594 arc_callback_t *acb_next; 595}; 596 597typedef struct arc_write_callback arc_write_callback_t; 598 599struct arc_write_callback { 600 void *awcb_private; 601 arc_done_func_t *awcb_ready; 602 arc_done_func_t *awcb_physdone; 603 arc_done_func_t *awcb_done; 604 arc_buf_t *awcb_buf; 605}; 606 607struct arc_buf_hdr { 608 /* protected by hash lock */ 609 dva_t b_dva; 610 uint64_t b_birth; 611 uint64_t b_cksum0; 612 613 kmutex_t b_freeze_lock; 614 zio_cksum_t *b_freeze_cksum; 615 void *b_thawed; 616 617 arc_buf_hdr_t *b_hash_next; 618 arc_buf_t *b_buf; 619 arc_flags_t b_flags; 620 uint32_t b_datacnt; 621 622 arc_callback_t *b_acb; 623 kcondvar_t b_cv; 624 625 /* immutable */ 626 arc_buf_contents_t b_type; 627 uint64_t b_size; 628 uint64_t b_spa; 629 630 /* protected by arc state mutex */ 631 arc_state_t *b_state; 632 list_node_t b_arc_node; 633 634 /* updated atomically */ 635 clock_t b_arc_access; 636 637 /* self protecting */ 638 refcount_t b_refcnt; 639 640 l2arc_buf_hdr_t *b_l2hdr; 641 list_node_t b_l2node; 642}; 643 644#ifdef _KERNEL 645static int 646sysctl_vfs_zfs_arc_meta_limit(SYSCTL_HANDLER_ARGS) 647{ 648 uint64_t val; 649 int err; 650 651 val = arc_meta_limit; 652 err = sysctl_handle_64(oidp, &val, 0, req); 653 if (err != 0 || req->newptr == NULL) 654 return (err); 655 656 if (val <= 0 || val > arc_c_max) 657 return (EINVAL); 658 659 arc_meta_limit = val; 660 return (0); 661} 662#endif 663 664static arc_buf_t *arc_eviction_list; 665static kmutex_t arc_eviction_mtx; 666static arc_buf_hdr_t arc_eviction_hdr; 667 668#define GHOST_STATE(state) \ 669 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \ 670 (state) == arc_l2c_only) 671 672#define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_FLAG_IN_HASH_TABLE) 673#define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) 674#define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_FLAG_IO_ERROR) 675#define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_FLAG_PREFETCH) 676#define HDR_FREED_IN_READ(hdr) ((hdr)->b_flags & ARC_FLAG_FREED_IN_READ) 677#define HDR_BUF_AVAILABLE(hdr) ((hdr)->b_flags & ARC_FLAG_BUF_AVAILABLE) 678#define HDR_FREE_IN_PROGRESS(hdr) \ 679 ((hdr)->b_flags & ARC_FLAG_FREE_IN_PROGRESS) 680#define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_FLAG_L2CACHE) 681#define HDR_L2_READING(hdr) \ 682 ((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS && \ 683 (hdr)->b_l2hdr != NULL) 684#define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITING) 685#define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_FLAG_L2_EVICTED) 686#define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITE_HEAD) 687 688/* 689 * Other sizes 690 */ 691 692#define HDR_SIZE ((int64_t)sizeof (arc_buf_hdr_t)) 693#define L2HDR_SIZE ((int64_t)sizeof (l2arc_buf_hdr_t)) 694 695/* 696 * Hash table routines 697 */ 698 699#define HT_LOCK_PAD CACHE_LINE_SIZE 700 701struct ht_lock { 702 kmutex_t ht_lock; 703#ifdef _KERNEL 704 unsigned char pad[(HT_LOCK_PAD - sizeof (kmutex_t))]; 705#endif 706}; 707 708#define BUF_LOCKS 256 709typedef struct buf_hash_table { 710 uint64_t ht_mask; 711 arc_buf_hdr_t **ht_table; 712 struct ht_lock ht_locks[BUF_LOCKS] __aligned(CACHE_LINE_SIZE); 713} buf_hash_table_t; 714 715static buf_hash_table_t buf_hash_table; 716 717#define BUF_HASH_INDEX(spa, dva, birth) \ 718 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask) 719#define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)]) 720#define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock)) 721#define HDR_LOCK(hdr) \ 722 (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth))) 723 724uint64_t zfs_crc64_table[256]; 725 726/* 727 * Level 2 ARC 728 */ 729 730#define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */ 731#define L2ARC_HEADROOM 2 /* num of writes */ 732/* 733 * If we discover during ARC scan any buffers to be compressed, we boost 734 * our headroom for the next scanning cycle by this percentage multiple. 735 */ 736#define L2ARC_HEADROOM_BOOST 200 737#define L2ARC_FEED_SECS 1 /* caching interval secs */ 738#define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */ 739 740#define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent) 741#define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done) 742 743/* L2ARC Performance Tunables */ 744uint64_t l2arc_write_max = L2ARC_WRITE_SIZE; /* default max write size */ 745uint64_t l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra write during warmup */ 746uint64_t l2arc_headroom = L2ARC_HEADROOM; /* number of dev writes */ 747uint64_t l2arc_headroom_boost = L2ARC_HEADROOM_BOOST; 748uint64_t l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */ 749uint64_t l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval milliseconds */ 750boolean_t l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */ 751boolean_t l2arc_feed_again = B_TRUE; /* turbo warmup */ 752boolean_t l2arc_norw = B_TRUE; /* no reads during writes */ 753 754SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_max, CTLFLAG_RW, 755 &l2arc_write_max, 0, "max write size"); 756SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_boost, CTLFLAG_RW, 757 &l2arc_write_boost, 0, "extra write during warmup"); 758SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_headroom, CTLFLAG_RW, 759 &l2arc_headroom, 0, "number of dev writes"); 760SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_secs, CTLFLAG_RW, 761 &l2arc_feed_secs, 0, "interval seconds"); 762SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_min_ms, CTLFLAG_RW, 763 &l2arc_feed_min_ms, 0, "min interval milliseconds"); 764 765SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_noprefetch, CTLFLAG_RW, 766 &l2arc_noprefetch, 0, "don't cache prefetch bufs"); 767SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_feed_again, CTLFLAG_RW, 768 &l2arc_feed_again, 0, "turbo warmup"); 769SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_norw, CTLFLAG_RW, 770 &l2arc_norw, 0, "no reads during writes"); 771 772SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_size, CTLFLAG_RD, 773 &ARC_anon.arcs_size, 0, "size of anonymous state"); 774SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_metadata_lsize, CTLFLAG_RD, 775 &ARC_anon.arcs_lsize[ARC_BUFC_METADATA], 0, "size of anonymous state"); 776SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_data_lsize, CTLFLAG_RD, 777 &ARC_anon.arcs_lsize[ARC_BUFC_DATA], 0, "size of anonymous state"); 778 779SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_size, CTLFLAG_RD, 780 &ARC_mru.arcs_size, 0, "size of mru state"); 781SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_metadata_lsize, CTLFLAG_RD, 782 &ARC_mru.arcs_lsize[ARC_BUFC_METADATA], 0, "size of metadata in mru state"); 783SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_data_lsize, CTLFLAG_RD, 784 &ARC_mru.arcs_lsize[ARC_BUFC_DATA], 0, "size of data in mru state"); 785 786SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_size, CTLFLAG_RD, 787 &ARC_mru_ghost.arcs_size, 0, "size of mru ghost state"); 788SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_metadata_lsize, CTLFLAG_RD, 789 &ARC_mru_ghost.arcs_lsize[ARC_BUFC_METADATA], 0, 790 "size of metadata in mru ghost state"); 791SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_data_lsize, CTLFLAG_RD, 792 &ARC_mru_ghost.arcs_lsize[ARC_BUFC_DATA], 0, 793 "size of data in mru ghost state"); 794 795SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_size, CTLFLAG_RD, 796 &ARC_mfu.arcs_size, 0, "size of mfu state"); 797SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_metadata_lsize, CTLFLAG_RD, 798 &ARC_mfu.arcs_lsize[ARC_BUFC_METADATA], 0, "size of metadata in mfu state"); 799SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_data_lsize, CTLFLAG_RD, 800 &ARC_mfu.arcs_lsize[ARC_BUFC_DATA], 0, "size of data in mfu state"); 801 802SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_size, CTLFLAG_RD, 803 &ARC_mfu_ghost.arcs_size, 0, "size of mfu ghost state"); 804SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_metadata_lsize, CTLFLAG_RD, 805 &ARC_mfu_ghost.arcs_lsize[ARC_BUFC_METADATA], 0, 806 "size of metadata in mfu ghost state"); 807SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_data_lsize, CTLFLAG_RD, 808 &ARC_mfu_ghost.arcs_lsize[ARC_BUFC_DATA], 0, 809 "size of data in mfu ghost state"); 810 811SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2c_only_size, CTLFLAG_RD, 812 &ARC_l2c_only.arcs_size, 0, "size of mru state"); 813 814/* 815 * L2ARC Internals 816 */ 817typedef struct l2arc_dev { 818 vdev_t *l2ad_vdev; /* vdev */ 819 spa_t *l2ad_spa; /* spa */ 820 uint64_t l2ad_hand; /* next write location */ 821 uint64_t l2ad_start; /* first addr on device */ 822 uint64_t l2ad_end; /* last addr on device */ 823 uint64_t l2ad_evict; /* last addr eviction reached */ 824 boolean_t l2ad_first; /* first sweep through */ 825 boolean_t l2ad_writing; /* currently writing */ 826 list_t *l2ad_buflist; /* buffer list */ 827 list_node_t l2ad_node; /* device list node */ 828} l2arc_dev_t; 829 830static list_t L2ARC_dev_list; /* device list */ 831static list_t *l2arc_dev_list; /* device list pointer */ 832static kmutex_t l2arc_dev_mtx; /* device list mutex */ 833static l2arc_dev_t *l2arc_dev_last; /* last device used */ 834static kmutex_t l2arc_buflist_mtx; /* mutex for all buflists */ 835static list_t L2ARC_free_on_write; /* free after write buf list */ 836static list_t *l2arc_free_on_write; /* free after write list ptr */ 837static kmutex_t l2arc_free_on_write_mtx; /* mutex for list */ 838static uint64_t l2arc_ndev; /* number of devices */ 839 840typedef struct l2arc_read_callback { 841 arc_buf_t *l2rcb_buf; /* read buffer */ 842 spa_t *l2rcb_spa; /* spa */ 843 blkptr_t l2rcb_bp; /* original blkptr */ 844 zbookmark_phys_t l2rcb_zb; /* original bookmark */ 845 int l2rcb_flags; /* original flags */ 846 enum zio_compress l2rcb_compress; /* applied compress */ 847} l2arc_read_callback_t; 848 849typedef struct l2arc_write_callback { 850 l2arc_dev_t *l2wcb_dev; /* device info */ 851 arc_buf_hdr_t *l2wcb_head; /* head of write buflist */ 852} l2arc_write_callback_t; 853 854struct l2arc_buf_hdr { 855 /* protected by arc_buf_hdr mutex */ 856 l2arc_dev_t *b_dev; /* L2ARC device */ 857 uint64_t b_daddr; /* disk address, offset byte */ 858 /* compression applied to buffer data */ 859 enum zio_compress b_compress; 860 /* real alloc'd buffer size depending on b_compress applied */ 861 int b_asize; 862 /* temporary buffer holder for in-flight compressed data */ 863 void *b_tmp_cdata; 864}; 865 866typedef struct l2arc_data_free { 867 /* protected by l2arc_free_on_write_mtx */ 868 void *l2df_data; 869 size_t l2df_size; 870 void (*l2df_func)(void *, size_t); 871 list_node_t l2df_list_node; 872} l2arc_data_free_t; 873 874static kmutex_t l2arc_feed_thr_lock; 875static kcondvar_t l2arc_feed_thr_cv; 876static uint8_t l2arc_thread_exit; 877 878static void arc_get_data_buf(arc_buf_t *); 879static void arc_access(arc_buf_hdr_t *, kmutex_t *); 880static int arc_evict_needed(arc_buf_contents_t); 881static void arc_evict_ghost(arc_state_t *, uint64_t, int64_t); 882static void arc_buf_watch(arc_buf_t *); 883 884static boolean_t l2arc_write_eligible(uint64_t, arc_buf_hdr_t *); 885static void l2arc_read_done(zio_t *); 886static void l2arc_hdr_stat_add(void); 887static void l2arc_hdr_stat_remove(void); 888 889static boolean_t l2arc_compress_buf(l2arc_buf_hdr_t *); 890static void l2arc_decompress_zio(zio_t *, arc_buf_hdr_t *, enum zio_compress); 891static void l2arc_release_cdata_buf(arc_buf_hdr_t *); 892 893static uint64_t 894buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth) 895{ 896 uint8_t *vdva = (uint8_t *)dva; 897 uint64_t crc = -1ULL; 898 int i; 899 900 ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY); 901 902 for (i = 0; i < sizeof (dva_t); i++) 903 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ vdva[i]) & 0xFF]; 904 905 crc ^= (spa>>8) ^ birth; 906 907 return (crc); 908} 909 910#define BUF_EMPTY(buf) \ 911 ((buf)->b_dva.dva_word[0] == 0 && \ 912 (buf)->b_dva.dva_word[1] == 0 && \ 913 (buf)->b_cksum0 == 0) 914 915#define BUF_EQUAL(spa, dva, birth, buf) \ 916 ((buf)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \ 917 ((buf)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \ 918 ((buf)->b_birth == birth) && ((buf)->b_spa == spa) 919 920static void 921buf_discard_identity(arc_buf_hdr_t *hdr) 922{ 923 hdr->b_dva.dva_word[0] = 0; 924 hdr->b_dva.dva_word[1] = 0; 925 hdr->b_birth = 0; 926 hdr->b_cksum0 = 0; 927} 928 929static arc_buf_hdr_t * 930buf_hash_find(uint64_t spa, const blkptr_t *bp, kmutex_t **lockp) 931{ 932 const dva_t *dva = BP_IDENTITY(bp); 933 uint64_t birth = BP_PHYSICAL_BIRTH(bp); 934 uint64_t idx = BUF_HASH_INDEX(spa, dva, birth); 935 kmutex_t *hash_lock = BUF_HASH_LOCK(idx); 936 arc_buf_hdr_t *hdr; 937 938 mutex_enter(hash_lock); 939 for (hdr = buf_hash_table.ht_table[idx]; hdr != NULL; 940 hdr = hdr->b_hash_next) { 941 if (BUF_EQUAL(spa, dva, birth, hdr)) { 942 *lockp = hash_lock; 943 return (hdr); 944 } 945 } 946 mutex_exit(hash_lock); 947 *lockp = NULL; 948 return (NULL); 949} 950 951/* 952 * Insert an entry into the hash table. If there is already an element 953 * equal to elem in the hash table, then the already existing element 954 * will be returned and the new element will not be inserted. 955 * Otherwise returns NULL. 956 */ 957static arc_buf_hdr_t * 958buf_hash_insert(arc_buf_hdr_t *hdr, kmutex_t **lockp) 959{ 960 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth); 961 kmutex_t *hash_lock = BUF_HASH_LOCK(idx); 962 arc_buf_hdr_t *fhdr; 963 uint32_t i; 964 965 ASSERT(!DVA_IS_EMPTY(&hdr->b_dva)); 966 ASSERT(hdr->b_birth != 0); 967 ASSERT(!HDR_IN_HASH_TABLE(hdr)); 968 *lockp = hash_lock; 969 mutex_enter(hash_lock); 970 for (fhdr = buf_hash_table.ht_table[idx], i = 0; fhdr != NULL; 971 fhdr = fhdr->b_hash_next, i++) { 972 if (BUF_EQUAL(hdr->b_spa, &hdr->b_dva, hdr->b_birth, fhdr)) 973 return (fhdr); 974 } 975 976 hdr->b_hash_next = buf_hash_table.ht_table[idx]; 977 buf_hash_table.ht_table[idx] = hdr; 978 hdr->b_flags |= ARC_FLAG_IN_HASH_TABLE; 979 980 /* collect some hash table performance data */ 981 if (i > 0) { 982 ARCSTAT_BUMP(arcstat_hash_collisions); 983 if (i == 1) 984 ARCSTAT_BUMP(arcstat_hash_chains); 985 986 ARCSTAT_MAX(arcstat_hash_chain_max, i); 987 } 988 989 ARCSTAT_BUMP(arcstat_hash_elements); 990 ARCSTAT_MAXSTAT(arcstat_hash_elements); 991 992 return (NULL); 993} 994 995static void 996buf_hash_remove(arc_buf_hdr_t *hdr) 997{ 998 arc_buf_hdr_t *fhdr, **hdrp; 999 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth); 1000 1001 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx))); 1002 ASSERT(HDR_IN_HASH_TABLE(hdr)); 1003 1004 hdrp = &buf_hash_table.ht_table[idx]; 1005 while ((fhdr = *hdrp) != hdr) { 1006 ASSERT(fhdr != NULL); 1007 hdrp = &fhdr->b_hash_next; 1008 } 1009 *hdrp = hdr->b_hash_next; 1010 hdr->b_hash_next = NULL; 1011 hdr->b_flags &= ~ARC_FLAG_IN_HASH_TABLE; 1012 1013 /* collect some hash table performance data */ 1014 ARCSTAT_BUMPDOWN(arcstat_hash_elements); 1015 1016 if (buf_hash_table.ht_table[idx] && 1017 buf_hash_table.ht_table[idx]->b_hash_next == NULL) 1018 ARCSTAT_BUMPDOWN(arcstat_hash_chains); 1019} 1020 1021/* 1022 * Global data structures and functions for the buf kmem cache. 1023 */ 1024static kmem_cache_t *hdr_cache; 1025static kmem_cache_t *buf_cache; 1026 1027static void 1028buf_fini(void) 1029{ 1030 int i; 1031 1032 kmem_free(buf_hash_table.ht_table, 1033 (buf_hash_table.ht_mask + 1) * sizeof (void *)); 1034 for (i = 0; i < BUF_LOCKS; i++) 1035 mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock); 1036 kmem_cache_destroy(hdr_cache); 1037 kmem_cache_destroy(buf_cache); 1038} 1039 1040/* 1041 * Constructor callback - called when the cache is empty 1042 * and a new buf is requested. 1043 */ 1044/* ARGSUSED */ 1045static int 1046hdr_cons(void *vbuf, void *unused, int kmflag) 1047{ 1048 arc_buf_hdr_t *hdr = vbuf; 1049 1050 bzero(hdr, sizeof (arc_buf_hdr_t)); 1051 refcount_create(&hdr->b_refcnt); 1052 cv_init(&hdr->b_cv, NULL, CV_DEFAULT, NULL); 1053 mutex_init(&hdr->b_freeze_lock, NULL, MUTEX_DEFAULT, NULL); 1054 arc_space_consume(sizeof (arc_buf_hdr_t), ARC_SPACE_HDRS); 1055 1056 return (0); 1057} 1058 1059/* ARGSUSED */ 1060static int 1061buf_cons(void *vbuf, void *unused, int kmflag) 1062{ 1063 arc_buf_t *buf = vbuf; 1064 1065 bzero(buf, sizeof (arc_buf_t)); 1066 mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL); 1067 arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS); 1068 1069 return (0); 1070} 1071 1072/* 1073 * Destructor callback - called when a cached buf is 1074 * no longer required. 1075 */ 1076/* ARGSUSED */ 1077static void 1078hdr_dest(void *vbuf, void *unused) 1079{ 1080 arc_buf_hdr_t *hdr = vbuf; 1081 1082 ASSERT(BUF_EMPTY(hdr)); 1083 refcount_destroy(&hdr->b_refcnt); 1084 cv_destroy(&hdr->b_cv); 1085 mutex_destroy(&hdr->b_freeze_lock); 1086 arc_space_return(sizeof (arc_buf_hdr_t), ARC_SPACE_HDRS); 1087} 1088 1089/* ARGSUSED */ 1090static void 1091buf_dest(void *vbuf, void *unused) 1092{ 1093 arc_buf_t *buf = vbuf; 1094 1095 mutex_destroy(&buf->b_evict_lock); 1096 arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS); 1097} 1098 1099/* 1100 * Reclaim callback -- invoked when memory is low. 1101 */ 1102/* ARGSUSED */ 1103static void 1104hdr_recl(void *unused) 1105{ 1106 dprintf("hdr_recl called\n"); 1107 /* 1108 * umem calls the reclaim func when we destroy the buf cache, 1109 * which is after we do arc_fini(). 1110 */ 1111 if (!arc_dead) 1112 cv_signal(&arc_reclaim_thr_cv); 1113} 1114 1115static void 1116buf_init(void) 1117{ 1118 uint64_t *ct; 1119 uint64_t hsize = 1ULL << 12; 1120 int i, j; 1121 1122 /* 1123 * The hash table is big enough to fill all of physical memory 1124 * with an average block size of zfs_arc_average_blocksize (default 8K). 1125 * By default, the table will take up 1126 * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers). 1127 */ 1128 while (hsize * zfs_arc_average_blocksize < (uint64_t)physmem * PAGESIZE) 1129 hsize <<= 1; 1130retry: 1131 buf_hash_table.ht_mask = hsize - 1; 1132 buf_hash_table.ht_table = 1133 kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP); 1134 if (buf_hash_table.ht_table == NULL) { 1135 ASSERT(hsize > (1ULL << 8)); 1136 hsize >>= 1; 1137 goto retry; 1138 } 1139 1140 hdr_cache = kmem_cache_create("arc_buf_hdr_t", sizeof (arc_buf_hdr_t), 1141 0, hdr_cons, hdr_dest, hdr_recl, NULL, NULL, 0); 1142 buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t), 1143 0, buf_cons, buf_dest, NULL, NULL, NULL, 0); 1144 1145 for (i = 0; i < 256; i++) 1146 for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--) 1147 *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY); 1148 1149 for (i = 0; i < BUF_LOCKS; i++) { 1150 mutex_init(&buf_hash_table.ht_locks[i].ht_lock, 1151 NULL, MUTEX_DEFAULT, NULL); 1152 } 1153} 1154 1155#define ARC_MINTIME (hz>>4) /* 62 ms */ 1156 1157static void 1158arc_cksum_verify(arc_buf_t *buf) 1159{ 1160 zio_cksum_t zc; 1161 1162 if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 1163 return; 1164 1165 mutex_enter(&buf->b_hdr->b_freeze_lock); 1166 if (buf->b_hdr->b_freeze_cksum == NULL || 1167 (buf->b_hdr->b_flags & ARC_FLAG_IO_ERROR)) { 1168 mutex_exit(&buf->b_hdr->b_freeze_lock); 1169 return; 1170 } 1171 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc); 1172 if (!ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc)) 1173 panic("buffer modified while frozen!"); 1174 mutex_exit(&buf->b_hdr->b_freeze_lock); 1175} 1176 1177static int 1178arc_cksum_equal(arc_buf_t *buf) 1179{ 1180 zio_cksum_t zc; 1181 int equal; 1182 1183 mutex_enter(&buf->b_hdr->b_freeze_lock); 1184 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc); 1185 equal = ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc); 1186 mutex_exit(&buf->b_hdr->b_freeze_lock); 1187 1188 return (equal); 1189} 1190 1191static void 1192arc_cksum_compute(arc_buf_t *buf, boolean_t force) 1193{ 1194 if (!force && !(zfs_flags & ZFS_DEBUG_MODIFY)) 1195 return; 1196 1197 mutex_enter(&buf->b_hdr->b_freeze_lock); 1198 if (buf->b_hdr->b_freeze_cksum != NULL) { 1199 mutex_exit(&buf->b_hdr->b_freeze_lock); 1200 return; 1201 } 1202 buf->b_hdr->b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t), KM_SLEEP); 1203 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, 1204 buf->b_hdr->b_freeze_cksum); 1205 mutex_exit(&buf->b_hdr->b_freeze_lock); 1206#ifdef illumos 1207 arc_buf_watch(buf); 1208#endif /* illumos */ 1209} 1210 1211#ifdef illumos 1212#ifndef _KERNEL 1213typedef struct procctl { 1214 long cmd; 1215 prwatch_t prwatch; 1216} procctl_t; 1217#endif 1218 1219/* ARGSUSED */ 1220static void 1221arc_buf_unwatch(arc_buf_t *buf) 1222{ 1223#ifndef _KERNEL 1224 if (arc_watch) { 1225 int result; 1226 procctl_t ctl; 1227 ctl.cmd = PCWATCH; 1228 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data; 1229 ctl.prwatch.pr_size = 0; 1230 ctl.prwatch.pr_wflags = 0; 1231 result = write(arc_procfd, &ctl, sizeof (ctl)); 1232 ASSERT3U(result, ==, sizeof (ctl)); 1233 } 1234#endif 1235} 1236 1237/* ARGSUSED */ 1238static void 1239arc_buf_watch(arc_buf_t *buf) 1240{ 1241#ifndef _KERNEL 1242 if (arc_watch) { 1243 int result; 1244 procctl_t ctl; 1245 ctl.cmd = PCWATCH; 1246 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data; 1247 ctl.prwatch.pr_size = buf->b_hdr->b_size; 1248 ctl.prwatch.pr_wflags = WA_WRITE; 1249 result = write(arc_procfd, &ctl, sizeof (ctl)); 1250 ASSERT3U(result, ==, sizeof (ctl)); 1251 } 1252#endif 1253} 1254#endif /* illumos */ 1255 1256void 1257arc_buf_thaw(arc_buf_t *buf) 1258{ 1259 if (zfs_flags & ZFS_DEBUG_MODIFY) { 1260 if (buf->b_hdr->b_state != arc_anon) 1261 panic("modifying non-anon buffer!"); 1262 if (buf->b_hdr->b_flags & ARC_FLAG_IO_IN_PROGRESS) 1263 panic("modifying buffer while i/o in progress!"); 1264 arc_cksum_verify(buf); 1265 } 1266 1267 mutex_enter(&buf->b_hdr->b_freeze_lock); 1268 if (buf->b_hdr->b_freeze_cksum != NULL) { 1269 kmem_free(buf->b_hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 1270 buf->b_hdr->b_freeze_cksum = NULL; 1271 } 1272 1273 if (zfs_flags & ZFS_DEBUG_MODIFY) { 1274 if (buf->b_hdr->b_thawed) 1275 kmem_free(buf->b_hdr->b_thawed, 1); 1276 buf->b_hdr->b_thawed = kmem_alloc(1, KM_SLEEP); 1277 } 1278 1279 mutex_exit(&buf->b_hdr->b_freeze_lock); 1280 1281#ifdef illumos 1282 arc_buf_unwatch(buf); 1283#endif /* illumos */ 1284} 1285 1286void 1287arc_buf_freeze(arc_buf_t *buf) 1288{ 1289 kmutex_t *hash_lock; 1290 1291 if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 1292 return; 1293 1294 hash_lock = HDR_LOCK(buf->b_hdr); 1295 mutex_enter(hash_lock); 1296 1297 ASSERT(buf->b_hdr->b_freeze_cksum != NULL || 1298 buf->b_hdr->b_state == arc_anon); 1299 arc_cksum_compute(buf, B_FALSE); 1300 mutex_exit(hash_lock); 1301 1302} 1303 1304static void 1305get_buf_info(arc_buf_hdr_t *hdr, arc_state_t *state, list_t **list, kmutex_t **lock) 1306{ 1307 uint64_t buf_hashid = buf_hash(hdr->b_spa, &hdr->b_dva, hdr->b_birth); 1308 1309 if (hdr->b_type == ARC_BUFC_METADATA) 1310 buf_hashid &= (ARC_BUFC_NUMMETADATALISTS - 1); 1311 else { 1312 buf_hashid &= (ARC_BUFC_NUMDATALISTS - 1); 1313 buf_hashid += ARC_BUFC_NUMMETADATALISTS; 1314 } 1315 1316 *list = &state->arcs_lists[buf_hashid]; 1317 *lock = ARCS_LOCK(state, buf_hashid); 1318} 1319 1320 1321static void 1322add_reference(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, void *tag) 1323{ 1324 ASSERT(MUTEX_HELD(hash_lock)); 1325 1326 if ((refcount_add(&hdr->b_refcnt, tag) == 1) && 1327 (hdr->b_state != arc_anon)) { 1328 uint64_t delta = hdr->b_size * hdr->b_datacnt; 1329 uint64_t *size = &hdr->b_state->arcs_lsize[hdr->b_type]; 1330 list_t *list; 1331 kmutex_t *lock; 1332 1333 get_buf_info(hdr, hdr->b_state, &list, &lock); 1334 ASSERT(!MUTEX_HELD(lock)); 1335 mutex_enter(lock); 1336 ASSERT(list_link_active(&hdr->b_arc_node)); 1337 list_remove(list, hdr); 1338 if (GHOST_STATE(hdr->b_state)) { 1339 ASSERT0(hdr->b_datacnt); 1340 ASSERT3P(hdr->b_buf, ==, NULL); 1341 delta = hdr->b_size; 1342 } 1343 ASSERT(delta > 0); 1344 ASSERT3U(*size, >=, delta); 1345 atomic_add_64(size, -delta); 1346 mutex_exit(lock); 1347 /* remove the prefetch flag if we get a reference */ 1348 if (hdr->b_flags & ARC_FLAG_PREFETCH) 1349 hdr->b_flags &= ~ARC_FLAG_PREFETCH; 1350 } 1351} 1352 1353static int 1354remove_reference(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, void *tag) 1355{ 1356 int cnt; 1357 arc_state_t *state = hdr->b_state; 1358 1359 ASSERT(state == arc_anon || MUTEX_HELD(hash_lock)); 1360 ASSERT(!GHOST_STATE(state)); 1361 1362 if (((cnt = refcount_remove(&hdr->b_refcnt, tag)) == 0) && 1363 (state != arc_anon)) { 1364 uint64_t *size = &state->arcs_lsize[hdr->b_type]; 1365 list_t *list; 1366 kmutex_t *lock; 1367 1368 get_buf_info(hdr, state, &list, &lock); 1369 ASSERT(!MUTEX_HELD(lock)); 1370 mutex_enter(lock); 1371 ASSERT(!list_link_active(&hdr->b_arc_node)); 1372 list_insert_head(list, hdr); 1373 ASSERT(hdr->b_datacnt > 0); 1374 atomic_add_64(size, hdr->b_size * hdr->b_datacnt); 1375 mutex_exit(lock); 1376 } 1377 return (cnt); 1378} 1379 1380/* 1381 * Move the supplied buffer to the indicated state. The mutex 1382 * for the buffer must be held by the caller. 1383 */ 1384static void 1385arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *hdr, 1386 kmutex_t *hash_lock) 1387{ 1388 arc_state_t *old_state = hdr->b_state; 1389 int64_t refcnt = refcount_count(&hdr->b_refcnt); 1390 uint64_t from_delta, to_delta; 1391 list_t *list; 1392 kmutex_t *lock; 1393 1394 ASSERT(MUTEX_HELD(hash_lock)); 1395 ASSERT3P(new_state, !=, old_state); 1396 ASSERT(refcnt == 0 || hdr->b_datacnt > 0); 1397 ASSERT(hdr->b_datacnt == 0 || !GHOST_STATE(new_state)); 1398 ASSERT(hdr->b_datacnt <= 1 || old_state != arc_anon); 1399 1400 from_delta = to_delta = hdr->b_datacnt * hdr->b_size; 1401 1402 /* 1403 * If this buffer is evictable, transfer it from the 1404 * old state list to the new state list. 1405 */ 1406 if (refcnt == 0) { 1407 if (old_state != arc_anon) { 1408 int use_mutex; 1409 uint64_t *size = &old_state->arcs_lsize[hdr->b_type]; 1410 1411 get_buf_info(hdr, old_state, &list, &lock); 1412 use_mutex = !MUTEX_HELD(lock); 1413 if (use_mutex) 1414 mutex_enter(lock); 1415 1416 ASSERT(list_link_active(&hdr->b_arc_node)); 1417 list_remove(list, hdr); 1418 1419 /* 1420 * If prefetching out of the ghost cache, 1421 * we will have a non-zero datacnt. 1422 */ 1423 if (GHOST_STATE(old_state) && hdr->b_datacnt == 0) { 1424 /* ghost elements have a ghost size */ 1425 ASSERT(hdr->b_buf == NULL); 1426 from_delta = hdr->b_size; 1427 } 1428 ASSERT3U(*size, >=, from_delta); 1429 atomic_add_64(size, -from_delta); 1430 1431 if (use_mutex) 1432 mutex_exit(lock); 1433 } 1434 if (new_state != arc_anon) { 1435 int use_mutex; 1436 uint64_t *size = &new_state->arcs_lsize[hdr->b_type]; 1437 1438 get_buf_info(hdr, new_state, &list, &lock); 1439 use_mutex = !MUTEX_HELD(lock); 1440 if (use_mutex) 1441 mutex_enter(lock); 1442 1443 list_insert_head(list, hdr); 1444 1445 /* ghost elements have a ghost size */ 1446 if (GHOST_STATE(new_state)) { 1447 ASSERT(hdr->b_datacnt == 0); 1448 ASSERT(hdr->b_buf == NULL); 1449 to_delta = hdr->b_size; 1450 } 1451 atomic_add_64(size, to_delta); 1452 1453 if (use_mutex) 1454 mutex_exit(lock); 1455 } 1456 } 1457 1458 ASSERT(!BUF_EMPTY(hdr)); 1459 if (new_state == arc_anon && HDR_IN_HASH_TABLE(hdr)) 1460 buf_hash_remove(hdr); 1461 1462 /* adjust state sizes */ 1463 if (to_delta) 1464 atomic_add_64(&new_state->arcs_size, to_delta); 1465 if (from_delta) { 1466 ASSERT3U(old_state->arcs_size, >=, from_delta); 1467 atomic_add_64(&old_state->arcs_size, -from_delta); 1468 } 1469 hdr->b_state = new_state; 1470 1471 /* adjust l2arc hdr stats */ 1472 if (new_state == arc_l2c_only) 1473 l2arc_hdr_stat_add(); 1474 else if (old_state == arc_l2c_only) 1475 l2arc_hdr_stat_remove(); 1476} 1477 1478void 1479arc_space_consume(uint64_t space, arc_space_type_t type) 1480{ 1481 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES); 1482 1483 switch (type) { 1484 case ARC_SPACE_DATA: 1485 ARCSTAT_INCR(arcstat_data_size, space); 1486 break; 1487 case ARC_SPACE_OTHER: 1488 ARCSTAT_INCR(arcstat_other_size, space); 1489 break; 1490 case ARC_SPACE_HDRS: 1491 ARCSTAT_INCR(arcstat_hdr_size, space); 1492 break; 1493 case ARC_SPACE_L2HDRS: 1494 ARCSTAT_INCR(arcstat_l2_hdr_size, space); 1495 break; 1496 } 1497 1498 ARCSTAT_INCR(arcstat_meta_used, space); 1499 atomic_add_64(&arc_size, space); 1500} 1501 1502void 1503arc_space_return(uint64_t space, arc_space_type_t type) 1504{ 1505 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES); 1506 1507 switch (type) { 1508 case ARC_SPACE_DATA: 1509 ARCSTAT_INCR(arcstat_data_size, -space); 1510 break; 1511 case ARC_SPACE_OTHER: 1512 ARCSTAT_INCR(arcstat_other_size, -space); 1513 break; 1514 case ARC_SPACE_HDRS: 1515 ARCSTAT_INCR(arcstat_hdr_size, -space); 1516 break; 1517 case ARC_SPACE_L2HDRS: 1518 ARCSTAT_INCR(arcstat_l2_hdr_size, -space); 1519 break; 1520 } 1521 1522 ASSERT(arc_meta_used >= space); 1523 if (arc_meta_max < arc_meta_used) 1524 arc_meta_max = arc_meta_used; 1525 ARCSTAT_INCR(arcstat_meta_used, -space); 1526 ASSERT(arc_size >= space); 1527 atomic_add_64(&arc_size, -space); 1528} 1529 1530arc_buf_t * 1531arc_buf_alloc(spa_t *spa, int size, void *tag, arc_buf_contents_t type) 1532{ 1533 arc_buf_hdr_t *hdr; 1534 arc_buf_t *buf; 1535 1536 ASSERT3U(size, >, 0); 1537 hdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE); 1538 ASSERT(BUF_EMPTY(hdr)); 1539 hdr->b_size = size; 1540 hdr->b_type = type; 1541 hdr->b_spa = spa_load_guid(spa); 1542 hdr->b_state = arc_anon; 1543 hdr->b_arc_access = 0; 1544 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 1545 buf->b_hdr = hdr; 1546 buf->b_data = NULL; 1547 buf->b_efunc = NULL; 1548 buf->b_private = NULL; 1549 buf->b_next = NULL; 1550 hdr->b_buf = buf; 1551 arc_get_data_buf(buf); 1552 hdr->b_datacnt = 1; 1553 hdr->b_flags = 0; 1554 ASSERT(refcount_is_zero(&hdr->b_refcnt)); 1555 (void) refcount_add(&hdr->b_refcnt, tag); 1556 1557 return (buf); 1558} 1559 1560static char *arc_onloan_tag = "onloan"; 1561 1562/* 1563 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in 1564 * flight data by arc_tempreserve_space() until they are "returned". Loaned 1565 * buffers must be returned to the arc before they can be used by the DMU or 1566 * freed. 1567 */ 1568arc_buf_t * 1569arc_loan_buf(spa_t *spa, int size) 1570{ 1571 arc_buf_t *buf; 1572 1573 buf = arc_buf_alloc(spa, size, arc_onloan_tag, ARC_BUFC_DATA); 1574 1575 atomic_add_64(&arc_loaned_bytes, size); 1576 return (buf); 1577} 1578 1579/* 1580 * Return a loaned arc buffer to the arc. 1581 */ 1582void 1583arc_return_buf(arc_buf_t *buf, void *tag) 1584{ 1585 arc_buf_hdr_t *hdr = buf->b_hdr; 1586 1587 ASSERT(buf->b_data != NULL); 1588 (void) refcount_add(&hdr->b_refcnt, tag); 1589 (void) refcount_remove(&hdr->b_refcnt, arc_onloan_tag); 1590 1591 atomic_add_64(&arc_loaned_bytes, -hdr->b_size); 1592} 1593 1594/* Detach an arc_buf from a dbuf (tag) */ 1595void 1596arc_loan_inuse_buf(arc_buf_t *buf, void *tag) 1597{ 1598 arc_buf_hdr_t *hdr; 1599 1600 ASSERT(buf->b_data != NULL); 1601 hdr = buf->b_hdr; 1602 (void) refcount_add(&hdr->b_refcnt, arc_onloan_tag); 1603 (void) refcount_remove(&hdr->b_refcnt, tag); 1604 buf->b_efunc = NULL; 1605 buf->b_private = NULL; 1606 1607 atomic_add_64(&arc_loaned_bytes, hdr->b_size); 1608} 1609 1610static arc_buf_t * 1611arc_buf_clone(arc_buf_t *from) 1612{ 1613 arc_buf_t *buf; 1614 arc_buf_hdr_t *hdr = from->b_hdr; 1615 uint64_t size = hdr->b_size; 1616 1617 ASSERT(hdr->b_state != arc_anon); 1618 1619 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 1620 buf->b_hdr = hdr; 1621 buf->b_data = NULL; 1622 buf->b_efunc = NULL; 1623 buf->b_private = NULL; 1624 buf->b_next = hdr->b_buf; 1625 hdr->b_buf = buf; 1626 arc_get_data_buf(buf); 1627 bcopy(from->b_data, buf->b_data, size); 1628 1629 /* 1630 * This buffer already exists in the arc so create a duplicate 1631 * copy for the caller. If the buffer is associated with user data 1632 * then track the size and number of duplicates. These stats will be 1633 * updated as duplicate buffers are created and destroyed. 1634 */ 1635 if (hdr->b_type == ARC_BUFC_DATA) { 1636 ARCSTAT_BUMP(arcstat_duplicate_buffers); 1637 ARCSTAT_INCR(arcstat_duplicate_buffers_size, size); 1638 } 1639 hdr->b_datacnt += 1; 1640 return (buf); 1641} 1642 1643void 1644arc_buf_add_ref(arc_buf_t *buf, void* tag) 1645{ 1646 arc_buf_hdr_t *hdr; 1647 kmutex_t *hash_lock; 1648 1649 /* 1650 * Check to see if this buffer is evicted. Callers 1651 * must verify b_data != NULL to know if the add_ref 1652 * was successful. 1653 */ 1654 mutex_enter(&buf->b_evict_lock); 1655 if (buf->b_data == NULL) { 1656 mutex_exit(&buf->b_evict_lock); 1657 return; 1658 } 1659 hash_lock = HDR_LOCK(buf->b_hdr); 1660 mutex_enter(hash_lock); 1661 hdr = buf->b_hdr; 1662 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 1663 mutex_exit(&buf->b_evict_lock); 1664 1665 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu); 1666 add_reference(hdr, hash_lock, tag); 1667 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr); 1668 arc_access(hdr, hash_lock); 1669 mutex_exit(hash_lock); 1670 ARCSTAT_BUMP(arcstat_hits); 1671 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_FLAG_PREFETCH), 1672 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA, 1673 data, metadata, hits); 1674} 1675 1676static void 1677arc_buf_free_on_write(void *data, size_t size, 1678 void (*free_func)(void *, size_t)) 1679{ 1680 l2arc_data_free_t *df; 1681 1682 df = kmem_alloc(sizeof (l2arc_data_free_t), KM_SLEEP); 1683 df->l2df_data = data; 1684 df->l2df_size = size; 1685 df->l2df_func = free_func; 1686 mutex_enter(&l2arc_free_on_write_mtx); 1687 list_insert_head(l2arc_free_on_write, df); 1688 mutex_exit(&l2arc_free_on_write_mtx); 1689} 1690 1691/* 1692 * Free the arc data buffer. If it is an l2arc write in progress, 1693 * the buffer is placed on l2arc_free_on_write to be freed later. 1694 */ 1695static void 1696arc_buf_data_free(arc_buf_t *buf, void (*free_func)(void *, size_t)) 1697{ 1698 arc_buf_hdr_t *hdr = buf->b_hdr; 1699 1700 if (HDR_L2_WRITING(hdr)) { 1701 arc_buf_free_on_write(buf->b_data, hdr->b_size, free_func); 1702 ARCSTAT_BUMP(arcstat_l2_free_on_write); 1703 } else { 1704 free_func(buf->b_data, hdr->b_size); 1705 } 1706} 1707 1708/* 1709 * Free up buf->b_data and if 'remove' is set, then pull the 1710 * arc_buf_t off of the the arc_buf_hdr_t's list and free it. 1711 */ 1712static void 1713arc_buf_l2_cdata_free(arc_buf_hdr_t *hdr) 1714{ 1715 l2arc_buf_hdr_t *l2hdr = hdr->b_l2hdr; 1716 1717 ASSERT(MUTEX_HELD(&l2arc_buflist_mtx)); 1718 1719 if (l2hdr->b_tmp_cdata == NULL) 1720 return; 1721 1722 ASSERT(HDR_L2_WRITING(hdr)); 1723 arc_buf_free_on_write(l2hdr->b_tmp_cdata, hdr->b_size, 1724 zio_data_buf_free); 1725 ARCSTAT_BUMP(arcstat_l2_cdata_free_on_write); 1726 l2hdr->b_tmp_cdata = NULL; 1727} 1728 1729static void 1730arc_buf_destroy(arc_buf_t *buf, boolean_t recycle, boolean_t remove) 1731{ 1732 arc_buf_t **bufp; 1733 1734 /* free up data associated with the buf */ 1735 if (buf->b_data) { 1736 arc_state_t *state = buf->b_hdr->b_state; 1737 uint64_t size = buf->b_hdr->b_size; 1738 arc_buf_contents_t type = buf->b_hdr->b_type; 1739 1740 arc_cksum_verify(buf); 1741#ifdef illumos 1742 arc_buf_unwatch(buf); 1743#endif /* illumos */ 1744 1745 if (!recycle) { 1746 if (type == ARC_BUFC_METADATA) { 1747 arc_buf_data_free(buf, zio_buf_free); 1748 arc_space_return(size, ARC_SPACE_DATA); 1749 } else { 1750 ASSERT(type == ARC_BUFC_DATA); 1751 arc_buf_data_free(buf, zio_data_buf_free); 1752 ARCSTAT_INCR(arcstat_data_size, -size); 1753 atomic_add_64(&arc_size, -size); 1754 } 1755 } 1756 if (list_link_active(&buf->b_hdr->b_arc_node)) { 1757 uint64_t *cnt = &state->arcs_lsize[type]; 1758 1759 ASSERT(refcount_is_zero(&buf->b_hdr->b_refcnt)); 1760 ASSERT(state != arc_anon); 1761 1762 ASSERT3U(*cnt, >=, size); 1763 atomic_add_64(cnt, -size); 1764 } 1765 ASSERT3U(state->arcs_size, >=, size); 1766 atomic_add_64(&state->arcs_size, -size); 1767 buf->b_data = NULL; 1768 1769 /* 1770 * If we're destroying a duplicate buffer make sure 1771 * that the appropriate statistics are updated. 1772 */ 1773 if (buf->b_hdr->b_datacnt > 1 && 1774 buf->b_hdr->b_type == ARC_BUFC_DATA) { 1775 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers); 1776 ARCSTAT_INCR(arcstat_duplicate_buffers_size, -size); 1777 } 1778 ASSERT(buf->b_hdr->b_datacnt > 0); 1779 buf->b_hdr->b_datacnt -= 1; 1780 } 1781 1782 /* only remove the buf if requested */ 1783 if (!remove) 1784 return; 1785 1786 /* remove the buf from the hdr list */ 1787 for (bufp = &buf->b_hdr->b_buf; *bufp != buf; bufp = &(*bufp)->b_next) 1788 continue; 1789 *bufp = buf->b_next; 1790 buf->b_next = NULL; 1791 1792 ASSERT(buf->b_efunc == NULL); 1793 1794 /* clean up the buf */ 1795 buf->b_hdr = NULL; 1796 kmem_cache_free(buf_cache, buf); 1797} 1798 1799static void 1800arc_hdr_destroy(arc_buf_hdr_t *hdr) 1801{ 1802 ASSERT(refcount_is_zero(&hdr->b_refcnt)); 1803 ASSERT3P(hdr->b_state, ==, arc_anon); 1804 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 1805 l2arc_buf_hdr_t *l2hdr = hdr->b_l2hdr; 1806 1807 if (l2hdr != NULL) { 1808 boolean_t buflist_held = MUTEX_HELD(&l2arc_buflist_mtx); 1809 /* 1810 * To prevent arc_free() and l2arc_evict() from 1811 * attempting to free the same buffer at the same time, 1812 * a FREE_IN_PROGRESS flag is given to arc_free() to 1813 * give it priority. l2arc_evict() can't destroy this 1814 * header while we are waiting on l2arc_buflist_mtx. 1815 * 1816 * The hdr may be removed from l2ad_buflist before we 1817 * grab l2arc_buflist_mtx, so b_l2hdr is rechecked. 1818 */ 1819 if (!buflist_held) { 1820 mutex_enter(&l2arc_buflist_mtx); 1821 l2hdr = hdr->b_l2hdr; 1822 } 1823 1824 if (l2hdr != NULL) { 1825 trim_map_free(l2hdr->b_dev->l2ad_vdev, l2hdr->b_daddr, 1826 l2hdr->b_asize, 0); 1827 list_remove(l2hdr->b_dev->l2ad_buflist, hdr); 1828 arc_buf_l2_cdata_free(hdr); 1829 ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size); 1830 ARCSTAT_INCR(arcstat_l2_asize, -l2hdr->b_asize); 1831 vdev_space_update(l2hdr->b_dev->l2ad_vdev, 1832 -l2hdr->b_asize, 0, 0); 1833 kmem_free(l2hdr, sizeof (l2arc_buf_hdr_t)); 1834 if (hdr->b_state == arc_l2c_only) 1835 l2arc_hdr_stat_remove(); 1836 hdr->b_l2hdr = NULL; 1837 } 1838 1839 if (!buflist_held) 1840 mutex_exit(&l2arc_buflist_mtx); 1841 } 1842 1843 if (!BUF_EMPTY(hdr)) { 1844 ASSERT(!HDR_IN_HASH_TABLE(hdr)); 1845 buf_discard_identity(hdr); 1846 } 1847 while (hdr->b_buf) { 1848 arc_buf_t *buf = hdr->b_buf; 1849 1850 if (buf->b_efunc) { 1851 mutex_enter(&arc_eviction_mtx); 1852 mutex_enter(&buf->b_evict_lock); 1853 ASSERT(buf->b_hdr != NULL); 1854 arc_buf_destroy(hdr->b_buf, FALSE, FALSE); 1855 hdr->b_buf = buf->b_next; 1856 buf->b_hdr = &arc_eviction_hdr; 1857 buf->b_next = arc_eviction_list; 1858 arc_eviction_list = buf; 1859 mutex_exit(&buf->b_evict_lock); 1860 mutex_exit(&arc_eviction_mtx); 1861 } else { 1862 arc_buf_destroy(hdr->b_buf, FALSE, TRUE); 1863 } 1864 } 1865 if (hdr->b_freeze_cksum != NULL) { 1866 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 1867 hdr->b_freeze_cksum = NULL; 1868 } 1869 if (hdr->b_thawed) { 1870 kmem_free(hdr->b_thawed, 1); 1871 hdr->b_thawed = NULL; 1872 } 1873 1874 ASSERT(!list_link_active(&hdr->b_arc_node)); 1875 ASSERT3P(hdr->b_hash_next, ==, NULL); 1876 ASSERT3P(hdr->b_acb, ==, NULL); 1877 kmem_cache_free(hdr_cache, hdr); 1878} 1879 1880void 1881arc_buf_free(arc_buf_t *buf, void *tag) 1882{ 1883 arc_buf_hdr_t *hdr = buf->b_hdr; 1884 int hashed = hdr->b_state != arc_anon; 1885 1886 ASSERT(buf->b_efunc == NULL); 1887 ASSERT(buf->b_data != NULL); 1888 1889 if (hashed) { 1890 kmutex_t *hash_lock = HDR_LOCK(hdr); 1891 1892 mutex_enter(hash_lock); 1893 hdr = buf->b_hdr; 1894 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 1895 1896 (void) remove_reference(hdr, hash_lock, tag); 1897 if (hdr->b_datacnt > 1) { 1898 arc_buf_destroy(buf, FALSE, TRUE); 1899 } else { 1900 ASSERT(buf == hdr->b_buf); 1901 ASSERT(buf->b_efunc == NULL); 1902 hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE; 1903 } 1904 mutex_exit(hash_lock); 1905 } else if (HDR_IO_IN_PROGRESS(hdr)) { 1906 int destroy_hdr; 1907 /* 1908 * We are in the middle of an async write. Don't destroy 1909 * this buffer unless the write completes before we finish 1910 * decrementing the reference count. 1911 */ 1912 mutex_enter(&arc_eviction_mtx); 1913 (void) remove_reference(hdr, NULL, tag); 1914 ASSERT(refcount_is_zero(&hdr->b_refcnt)); 1915 destroy_hdr = !HDR_IO_IN_PROGRESS(hdr); 1916 mutex_exit(&arc_eviction_mtx); 1917 if (destroy_hdr) 1918 arc_hdr_destroy(hdr); 1919 } else { 1920 if (remove_reference(hdr, NULL, tag) > 0) 1921 arc_buf_destroy(buf, FALSE, TRUE); 1922 else 1923 arc_hdr_destroy(hdr); 1924 } 1925} 1926 1927boolean_t 1928arc_buf_remove_ref(arc_buf_t *buf, void* tag) 1929{ 1930 arc_buf_hdr_t *hdr = buf->b_hdr; 1931 kmutex_t *hash_lock = HDR_LOCK(hdr); 1932 boolean_t no_callback = (buf->b_efunc == NULL); 1933 1934 if (hdr->b_state == arc_anon) { 1935 ASSERT(hdr->b_datacnt == 1); 1936 arc_buf_free(buf, tag); 1937 return (no_callback); 1938 } 1939 1940 mutex_enter(hash_lock); 1941 hdr = buf->b_hdr; 1942 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 1943 ASSERT(hdr->b_state != arc_anon); 1944 ASSERT(buf->b_data != NULL); 1945 1946 (void) remove_reference(hdr, hash_lock, tag); 1947 if (hdr->b_datacnt > 1) { 1948 if (no_callback) 1949 arc_buf_destroy(buf, FALSE, TRUE); 1950 } else if (no_callback) { 1951 ASSERT(hdr->b_buf == buf && buf->b_next == NULL); 1952 ASSERT(buf->b_efunc == NULL); 1953 hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE; 1954 } 1955 ASSERT(no_callback || hdr->b_datacnt > 1 || 1956 refcount_is_zero(&hdr->b_refcnt)); 1957 mutex_exit(hash_lock); 1958 return (no_callback); 1959} 1960 1961int 1962arc_buf_size(arc_buf_t *buf) 1963{ 1964 return (buf->b_hdr->b_size); 1965} 1966 1967/* 1968 * Called from the DMU to determine if the current buffer should be 1969 * evicted. In order to ensure proper locking, the eviction must be initiated 1970 * from the DMU. Return true if the buffer is associated with user data and 1971 * duplicate buffers still exist. 1972 */ 1973boolean_t 1974arc_buf_eviction_needed(arc_buf_t *buf) 1975{ 1976 arc_buf_hdr_t *hdr; 1977 boolean_t evict_needed = B_FALSE; 1978 1979 if (zfs_disable_dup_eviction) 1980 return (B_FALSE); 1981 1982 mutex_enter(&buf->b_evict_lock); 1983 hdr = buf->b_hdr; 1984 if (hdr == NULL) { 1985 /* 1986 * We are in arc_do_user_evicts(); let that function 1987 * perform the eviction. 1988 */ 1989 ASSERT(buf->b_data == NULL); 1990 mutex_exit(&buf->b_evict_lock); 1991 return (B_FALSE); 1992 } else if (buf->b_data == NULL) { 1993 /* 1994 * We have already been added to the arc eviction list; 1995 * recommend eviction. 1996 */ 1997 ASSERT3P(hdr, ==, &arc_eviction_hdr); 1998 mutex_exit(&buf->b_evict_lock); 1999 return (B_TRUE); 2000 } 2001 2002 if (hdr->b_datacnt > 1 && hdr->b_type == ARC_BUFC_DATA) 2003 evict_needed = B_TRUE; 2004 2005 mutex_exit(&buf->b_evict_lock); 2006 return (evict_needed); 2007} 2008 2009/* 2010 * Evict buffers from list until we've removed the specified number of 2011 * bytes. Move the removed buffers to the appropriate evict state. 2012 * If the recycle flag is set, then attempt to "recycle" a buffer: 2013 * - look for a buffer to evict that is `bytes' long. 2014 * - return the data block from this buffer rather than freeing it. 2015 * This flag is used by callers that are trying to make space for a 2016 * new buffer in a full arc cache. 2017 * 2018 * This function makes a "best effort". It skips over any buffers 2019 * it can't get a hash_lock on, and so may not catch all candidates. 2020 * It may also return without evicting as much space as requested. 2021 */ 2022static void * 2023arc_evict(arc_state_t *state, uint64_t spa, int64_t bytes, boolean_t recycle, 2024 arc_buf_contents_t type) 2025{ 2026 arc_state_t *evicted_state; 2027 uint64_t bytes_evicted = 0, skipped = 0, missed = 0; 2028 int64_t bytes_remaining; 2029 arc_buf_hdr_t *hdr, *hdr_prev = NULL; 2030 list_t *evicted_list, *list, *evicted_list_start, *list_start; 2031 kmutex_t *lock, *evicted_lock; 2032 kmutex_t *hash_lock; 2033 boolean_t have_lock; 2034 void *stolen = NULL; 2035 arc_buf_hdr_t marker = { 0 }; 2036 int count = 0; 2037 static int evict_metadata_offset, evict_data_offset; 2038 int i, idx, offset, list_count, lists; 2039 2040 ASSERT(state == arc_mru || state == arc_mfu); 2041 2042 evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost; 2043 2044 if (type == ARC_BUFC_METADATA) { 2045 offset = 0; 2046 list_count = ARC_BUFC_NUMMETADATALISTS; 2047 list_start = &state->arcs_lists[0]; 2048 evicted_list_start = &evicted_state->arcs_lists[0]; 2049 idx = evict_metadata_offset; 2050 } else { 2051 offset = ARC_BUFC_NUMMETADATALISTS; 2052 list_start = &state->arcs_lists[offset]; 2053 evicted_list_start = &evicted_state->arcs_lists[offset]; 2054 list_count = ARC_BUFC_NUMDATALISTS; 2055 idx = evict_data_offset; 2056 } 2057 bytes_remaining = evicted_state->arcs_lsize[type]; 2058 lists = 0; 2059 2060evict_start: 2061 list = &list_start[idx]; 2062 evicted_list = &evicted_list_start[idx]; 2063 lock = ARCS_LOCK(state, (offset + idx)); 2064 evicted_lock = ARCS_LOCK(evicted_state, (offset + idx)); 2065 2066 mutex_enter(lock); 2067 mutex_enter(evicted_lock); 2068 2069 for (hdr = list_tail(list); hdr; hdr = hdr_prev) { 2070 hdr_prev = list_prev(list, hdr); 2071 bytes_remaining -= (hdr->b_size * hdr->b_datacnt); 2072 /* prefetch buffers have a minimum lifespan */ 2073 if (HDR_IO_IN_PROGRESS(hdr) || 2074 (spa && hdr->b_spa != spa) || 2075 (hdr->b_flags & (ARC_FLAG_PREFETCH | ARC_FLAG_INDIRECT) && 2076 ddi_get_lbolt() - hdr->b_arc_access < 2077 arc_min_prefetch_lifespan)) { 2078 skipped++; 2079 continue; 2080 } 2081 /* "lookahead" for better eviction candidate */ 2082 if (recycle && hdr->b_size != bytes && 2083 hdr_prev && hdr_prev->b_size == bytes) 2084 continue; 2085 2086 /* ignore markers */ 2087 if (hdr->b_spa == 0) 2088 continue; 2089 2090 /* 2091 * It may take a long time to evict all the bufs requested. 2092 * To avoid blocking all arc activity, periodically drop 2093 * the arcs_mtx and give other threads a chance to run 2094 * before reacquiring the lock. 2095 * 2096 * If we are looking for a buffer to recycle, we are in 2097 * the hot code path, so don't sleep. 2098 */ 2099 if (!recycle && count++ > arc_evict_iterations) { 2100 list_insert_after(list, hdr, &marker); 2101 mutex_exit(evicted_lock); 2102 mutex_exit(lock); 2103 kpreempt(KPREEMPT_SYNC); 2104 mutex_enter(lock); 2105 mutex_enter(evicted_lock); 2106 hdr_prev = list_prev(list, &marker); 2107 list_remove(list, &marker); 2108 count = 0; 2109 continue; 2110 } 2111 2112 hash_lock = HDR_LOCK(hdr); 2113 have_lock = MUTEX_HELD(hash_lock); 2114 if (have_lock || mutex_tryenter(hash_lock)) { 2115 ASSERT0(refcount_count(&hdr->b_refcnt)); 2116 ASSERT(hdr->b_datacnt > 0); 2117 while (hdr->b_buf) { 2118 arc_buf_t *buf = hdr->b_buf; 2119 if (!mutex_tryenter(&buf->b_evict_lock)) { 2120 missed += 1; 2121 break; 2122 } 2123 if (buf->b_data) { 2124 bytes_evicted += hdr->b_size; 2125 if (recycle && hdr->b_type == type && 2126 hdr->b_size == bytes && 2127 !HDR_L2_WRITING(hdr)) { 2128 stolen = buf->b_data; 2129 recycle = FALSE; 2130 } 2131 } 2132 if (buf->b_efunc) { 2133 mutex_enter(&arc_eviction_mtx); 2134 arc_buf_destroy(buf, 2135 buf->b_data == stolen, FALSE); 2136 hdr->b_buf = buf->b_next; 2137 buf->b_hdr = &arc_eviction_hdr; 2138 buf->b_next = arc_eviction_list; 2139 arc_eviction_list = buf; 2140 mutex_exit(&arc_eviction_mtx); 2141 mutex_exit(&buf->b_evict_lock); 2142 } else { 2143 mutex_exit(&buf->b_evict_lock); 2144 arc_buf_destroy(buf, 2145 buf->b_data == stolen, TRUE); 2146 } 2147 } 2148 2149 if (hdr->b_l2hdr) { 2150 ARCSTAT_INCR(arcstat_evict_l2_cached, 2151 hdr->b_size); 2152 } else { 2153 if (l2arc_write_eligible(hdr->b_spa, hdr)) { 2154 ARCSTAT_INCR(arcstat_evict_l2_eligible, 2155 hdr->b_size); 2156 } else { 2157 ARCSTAT_INCR( 2158 arcstat_evict_l2_ineligible, 2159 hdr->b_size); 2160 } 2161 } 2162 2163 if (hdr->b_datacnt == 0) { 2164 arc_change_state(evicted_state, hdr, hash_lock); 2165 ASSERT(HDR_IN_HASH_TABLE(hdr)); 2166 hdr->b_flags |= ARC_FLAG_IN_HASH_TABLE; 2167 hdr->b_flags &= ~ARC_FLAG_BUF_AVAILABLE; 2168 DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, hdr); 2169 } 2170 if (!have_lock) 2171 mutex_exit(hash_lock); 2172 if (bytes >= 0 && bytes_evicted >= bytes) 2173 break; 2174 if (bytes_remaining > 0) { 2175 mutex_exit(evicted_lock); 2176 mutex_exit(lock); 2177 idx = ((idx + 1) & (list_count - 1)); 2178 lists++; 2179 goto evict_start; 2180 } 2181 } else { 2182 missed += 1; 2183 } 2184 } 2185 2186 mutex_exit(evicted_lock); 2187 mutex_exit(lock); 2188 2189 idx = ((idx + 1) & (list_count - 1)); 2190 lists++; 2191 2192 if (bytes_evicted < bytes) { 2193 if (lists < list_count) 2194 goto evict_start; 2195 else 2196 dprintf("only evicted %lld bytes from %x", 2197 (longlong_t)bytes_evicted, state); 2198 } 2199 if (type == ARC_BUFC_METADATA) 2200 evict_metadata_offset = idx; 2201 else 2202 evict_data_offset = idx; 2203 2204 if (skipped) 2205 ARCSTAT_INCR(arcstat_evict_skip, skipped); 2206 2207 if (missed) 2208 ARCSTAT_INCR(arcstat_mutex_miss, missed); 2209 2210 /* 2211 * Note: we have just evicted some data into the ghost state, 2212 * potentially putting the ghost size over the desired size. Rather 2213 * that evicting from the ghost list in this hot code path, leave 2214 * this chore to the arc_reclaim_thread(). 2215 */ 2216 2217 if (stolen) 2218 ARCSTAT_BUMP(arcstat_stolen); 2219 return (stolen); 2220} 2221 2222/* 2223 * Remove buffers from list until we've removed the specified number of 2224 * bytes. Destroy the buffers that are removed. 2225 */ 2226static void 2227arc_evict_ghost(arc_state_t *state, uint64_t spa, int64_t bytes) 2228{ 2229 arc_buf_hdr_t *hdr, *hdr_prev; 2230 arc_buf_hdr_t marker = { 0 }; 2231 list_t *list, *list_start; 2232 kmutex_t *hash_lock, *lock; 2233 uint64_t bytes_deleted = 0; 2234 uint64_t bufs_skipped = 0; 2235 int count = 0; 2236 static int evict_offset; 2237 int list_count, idx = evict_offset; 2238 int offset, lists = 0; 2239 2240 ASSERT(GHOST_STATE(state)); 2241 2242 /* 2243 * data lists come after metadata lists 2244 */ 2245 list_start = &state->arcs_lists[ARC_BUFC_NUMMETADATALISTS]; 2246 list_count = ARC_BUFC_NUMDATALISTS; 2247 offset = ARC_BUFC_NUMMETADATALISTS; 2248 2249evict_start: 2250 list = &list_start[idx]; 2251 lock = ARCS_LOCK(state, idx + offset); 2252 2253 mutex_enter(lock); 2254 for (hdr = list_tail(list); hdr; hdr = hdr_prev) { 2255 hdr_prev = list_prev(list, hdr); 2256 if (hdr->b_type > ARC_BUFC_NUMTYPES) 2257 panic("invalid hdr=%p", (void *)hdr); 2258 if (spa && hdr->b_spa != spa) 2259 continue; 2260 2261 /* ignore markers */ 2262 if (hdr->b_spa == 0) 2263 continue; 2264 2265 hash_lock = HDR_LOCK(hdr); 2266 /* caller may be trying to modify this buffer, skip it */ 2267 if (MUTEX_HELD(hash_lock)) 2268 continue; 2269 2270 /* 2271 * It may take a long time to evict all the bufs requested. 2272 * To avoid blocking all arc activity, periodically drop 2273 * the arcs_mtx and give other threads a chance to run 2274 * before reacquiring the lock. 2275 */ 2276 if (count++ > arc_evict_iterations) { 2277 list_insert_after(list, hdr, &marker); 2278 mutex_exit(lock); 2279 kpreempt(KPREEMPT_SYNC); 2280 mutex_enter(lock); 2281 hdr_prev = list_prev(list, &marker); 2282 list_remove(list, &marker); 2283 count = 0; 2284 continue; 2285 } 2286 if (mutex_tryenter(hash_lock)) { 2287 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 2288 ASSERT(hdr->b_buf == NULL); 2289 ARCSTAT_BUMP(arcstat_deleted); 2290 bytes_deleted += hdr->b_size; 2291 2292 if (hdr->b_l2hdr != NULL) { 2293 /* 2294 * This buffer is cached on the 2nd Level ARC; 2295 * don't destroy the header. 2296 */ 2297 arc_change_state(arc_l2c_only, hdr, hash_lock); 2298 mutex_exit(hash_lock); 2299 } else { 2300 arc_change_state(arc_anon, hdr, hash_lock); 2301 mutex_exit(hash_lock); 2302 arc_hdr_destroy(hdr); 2303 } 2304 2305 DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, hdr); 2306 if (bytes >= 0 && bytes_deleted >= bytes) 2307 break; 2308 } else if (bytes < 0) { 2309 /* 2310 * Insert a list marker and then wait for the 2311 * hash lock to become available. Once its 2312 * available, restart from where we left off. 2313 */ 2314 list_insert_after(list, hdr, &marker); 2315 mutex_exit(lock); 2316 mutex_enter(hash_lock); 2317 mutex_exit(hash_lock); 2318 mutex_enter(lock); 2319 hdr_prev = list_prev(list, &marker); 2320 list_remove(list, &marker); 2321 } else { 2322 bufs_skipped += 1; 2323 } 2324 2325 } 2326 mutex_exit(lock); 2327 idx = ((idx + 1) & (ARC_BUFC_NUMDATALISTS - 1)); 2328 lists++; 2329 2330 if (lists < list_count) 2331 goto evict_start; 2332 2333 evict_offset = idx; 2334 if ((uintptr_t)list > (uintptr_t)&state->arcs_lists[ARC_BUFC_NUMMETADATALISTS] && 2335 (bytes < 0 || bytes_deleted < bytes)) { 2336 list_start = &state->arcs_lists[0]; 2337 list_count = ARC_BUFC_NUMMETADATALISTS; 2338 offset = lists = 0; 2339 goto evict_start; 2340 } 2341 2342 if (bufs_skipped) { 2343 ARCSTAT_INCR(arcstat_mutex_miss, bufs_skipped); 2344 ASSERT(bytes >= 0); 2345 } 2346 2347 if (bytes_deleted < bytes) 2348 dprintf("only deleted %lld bytes from %p", 2349 (longlong_t)bytes_deleted, state); 2350} 2351 2352static void 2353arc_adjust(void) 2354{ 2355 int64_t adjustment, delta; 2356 2357 /* 2358 * Adjust MRU size 2359 */ 2360 2361 adjustment = MIN((int64_t)(arc_size - arc_c), 2362 (int64_t)(arc_anon->arcs_size + arc_mru->arcs_size + arc_meta_used - 2363 arc_p)); 2364 2365 if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_DATA] > 0) { 2366 delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_DATA], adjustment); 2367 (void) arc_evict(arc_mru, 0, delta, FALSE, ARC_BUFC_DATA); 2368 adjustment -= delta; 2369 } 2370 2371 if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_METADATA] > 0) { 2372 delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_METADATA], adjustment); 2373 (void) arc_evict(arc_mru, 0, delta, FALSE, 2374 ARC_BUFC_METADATA); 2375 } 2376 2377 /* 2378 * Adjust MFU size 2379 */ 2380 2381 adjustment = arc_size - arc_c; 2382 2383 if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_DATA] > 0) { 2384 delta = MIN(adjustment, arc_mfu->arcs_lsize[ARC_BUFC_DATA]); 2385 (void) arc_evict(arc_mfu, 0, delta, FALSE, ARC_BUFC_DATA); 2386 adjustment -= delta; 2387 } 2388 2389 if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_METADATA] > 0) { 2390 int64_t delta = MIN(adjustment, 2391 arc_mfu->arcs_lsize[ARC_BUFC_METADATA]); 2392 (void) arc_evict(arc_mfu, 0, delta, FALSE, 2393 ARC_BUFC_METADATA); 2394 } 2395 2396 /* 2397 * Adjust ghost lists 2398 */ 2399 2400 adjustment = arc_mru->arcs_size + arc_mru_ghost->arcs_size - arc_c; 2401 2402 if (adjustment > 0 && arc_mru_ghost->arcs_size > 0) { 2403 delta = MIN(arc_mru_ghost->arcs_size, adjustment); 2404 arc_evict_ghost(arc_mru_ghost, 0, delta); 2405 } 2406 2407 adjustment = 2408 arc_mru_ghost->arcs_size + arc_mfu_ghost->arcs_size - arc_c; 2409 2410 if (adjustment > 0 && arc_mfu_ghost->arcs_size > 0) { 2411 delta = MIN(arc_mfu_ghost->arcs_size, adjustment); 2412 arc_evict_ghost(arc_mfu_ghost, 0, delta); 2413 } 2414} 2415 2416static void 2417arc_do_user_evicts(void) 2418{ 2419 static arc_buf_t *tmp_arc_eviction_list; 2420 2421 /* 2422 * Move list over to avoid LOR 2423 */ 2424restart: 2425 mutex_enter(&arc_eviction_mtx); 2426 tmp_arc_eviction_list = arc_eviction_list; 2427 arc_eviction_list = NULL; 2428 mutex_exit(&arc_eviction_mtx); 2429 2430 while (tmp_arc_eviction_list != NULL) { 2431 arc_buf_t *buf = tmp_arc_eviction_list; 2432 tmp_arc_eviction_list = buf->b_next; 2433 mutex_enter(&buf->b_evict_lock); 2434 buf->b_hdr = NULL; 2435 mutex_exit(&buf->b_evict_lock); 2436 2437 if (buf->b_efunc != NULL) 2438 VERIFY0(buf->b_efunc(buf->b_private)); 2439 2440 buf->b_efunc = NULL; 2441 buf->b_private = NULL; 2442 kmem_cache_free(buf_cache, buf); 2443 } 2444 2445 if (arc_eviction_list != NULL) 2446 goto restart; 2447} 2448 2449/* 2450 * Flush all *evictable* data from the cache for the given spa. 2451 * NOTE: this will not touch "active" (i.e. referenced) data. 2452 */ 2453void 2454arc_flush(spa_t *spa) 2455{ 2456 uint64_t guid = 0; 2457 2458 if (spa) 2459 guid = spa_load_guid(spa); 2460 2461 while (arc_mru->arcs_lsize[ARC_BUFC_DATA]) { 2462 (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_DATA); 2463 if (spa) 2464 break; 2465 } 2466 while (arc_mru->arcs_lsize[ARC_BUFC_METADATA]) { 2467 (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_METADATA); 2468 if (spa) 2469 break; 2470 } 2471 while (arc_mfu->arcs_lsize[ARC_BUFC_DATA]) { 2472 (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_DATA); 2473 if (spa) 2474 break; 2475 } 2476 while (arc_mfu->arcs_lsize[ARC_BUFC_METADATA]) { 2477 (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_METADATA); 2478 if (spa) 2479 break; 2480 } 2481 2482 arc_evict_ghost(arc_mru_ghost, guid, -1); 2483 arc_evict_ghost(arc_mfu_ghost, guid, -1); 2484 2485 mutex_enter(&arc_reclaim_thr_lock); 2486 arc_do_user_evicts(); 2487 mutex_exit(&arc_reclaim_thr_lock); 2488 ASSERT(spa || arc_eviction_list == NULL); 2489} 2490 2491void 2492arc_shrink(void) 2493{ 2494 2495 if (arc_c > arc_c_min) { 2496 uint64_t to_free; 2497 2498 DTRACE_PROBE4(arc__shrink, uint64_t, arc_c, uint64_t, 2499 arc_c_min, uint64_t, arc_p, uint64_t, to_free); 2500#ifdef _KERNEL 2501 to_free = arc_c >> arc_shrink_shift; 2502#else 2503 to_free = arc_c >> arc_shrink_shift; 2504#endif 2505 if (arc_c > arc_c_min + to_free) 2506 atomic_add_64(&arc_c, -to_free); 2507 else 2508 arc_c = arc_c_min; 2509 2510 atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift)); 2511 if (arc_c > arc_size) 2512 arc_c = MAX(arc_size, arc_c_min); 2513 if (arc_p > arc_c) 2514 arc_p = (arc_c >> 1); 2515 2516 DTRACE_PROBE2(arc__shrunk, uint64_t, arc_c, uint64_t, 2517 arc_p); 2518 2519 ASSERT(arc_c >= arc_c_min); 2520 ASSERT((int64_t)arc_p >= 0); 2521 } 2522 2523 if (arc_size > arc_c) { 2524 DTRACE_PROBE2(arc__shrink_adjust, uint64_t, arc_size, 2525 uint64_t, arc_c); 2526 arc_adjust(); 2527 } 2528} 2529 2530static int needfree = 0; 2531 2532static int 2533arc_reclaim_needed(void) 2534{ 2535 2536#ifdef _KERNEL 2537 2538 if (needfree) { 2539 DTRACE_PROBE(arc__reclaim_needfree); 2540 return (1); 2541 } 2542 2543 /* 2544 * Cooperate with pagedaemon when it's time for it to scan 2545 * and reclaim some pages. 2546 */ 2547 if (freemem < zfs_arc_free_target) { 2548 DTRACE_PROBE2(arc__reclaim_freemem, uint64_t, 2549 freemem, uint64_t, zfs_arc_free_target); 2550 return (1); 2551 } 2552 2553#ifdef sun 2554 /* 2555 * take 'desfree' extra pages, so we reclaim sooner, rather than later 2556 */ 2557 extra = desfree; 2558 2559 /* 2560 * check that we're out of range of the pageout scanner. It starts to 2561 * schedule paging if freemem is less than lotsfree and needfree. 2562 * lotsfree is the high-water mark for pageout, and needfree is the 2563 * number of needed free pages. We add extra pages here to make sure 2564 * the scanner doesn't start up while we're freeing memory. 2565 */ 2566 if (freemem < lotsfree + needfree + extra) 2567 return (1); 2568 2569 /* 2570 * check to make sure that swapfs has enough space so that anon 2571 * reservations can still succeed. anon_resvmem() checks that the 2572 * availrmem is greater than swapfs_minfree, and the number of reserved 2573 * swap pages. We also add a bit of extra here just to prevent 2574 * circumstances from getting really dire. 2575 */ 2576 if (availrmem < swapfs_minfree + swapfs_reserve + extra) 2577 return (1); 2578 2579 /* 2580 * Check that we have enough availrmem that memory locking (e.g., via 2581 * mlock(3C) or memcntl(2)) can still succeed. (pages_pp_maximum 2582 * stores the number of pages that cannot be locked; when availrmem 2583 * drops below pages_pp_maximum, page locking mechanisms such as 2584 * page_pp_lock() will fail.) 2585 */ 2586 if (availrmem <= pages_pp_maximum) 2587 return (1); 2588 2589#endif /* sun */ 2590#if defined(__i386) || !defined(UMA_MD_SMALL_ALLOC) 2591 /* 2592 * If we're on an i386 platform, it's possible that we'll exhaust the 2593 * kernel heap space before we ever run out of available physical 2594 * memory. Most checks of the size of the heap_area compare against 2595 * tune.t_minarmem, which is the minimum available real memory that we 2596 * can have in the system. However, this is generally fixed at 25 pages 2597 * which is so low that it's useless. In this comparison, we seek to 2598 * calculate the total heap-size, and reclaim if more than 3/4ths of the 2599 * heap is allocated. (Or, in the calculation, if less than 1/4th is 2600 * free) 2601 */ 2602 if (vmem_size(heap_arena, VMEM_FREE) < 2603 (vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC) >> 2)) { 2604 DTRACE_PROBE2(arc__reclaim_used, uint64_t, 2605 vmem_size(heap_arena, VMEM_FREE), uint64_t, 2606 (vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC)) >> 2); 2607 return (1); 2608 } 2609#define zio_arena NULL 2610#else 2611#define zio_arena heap_arena 2612#endif 2613 2614 /* 2615 * If zio data pages are being allocated out of a separate heap segment, 2616 * then enforce that the size of available vmem for this arena remains 2617 * above about 1/16th free. 2618 * 2619 * Note: The 1/16th arena free requirement was put in place 2620 * to aggressively evict memory from the arc in order to avoid 2621 * memory fragmentation issues. 2622 */ 2623 if (zio_arena != NULL && 2624 vmem_size(zio_arena, VMEM_FREE) < 2625 (vmem_size(zio_arena, VMEM_ALLOC) >> 4)) 2626 return (1); 2627 2628 /* 2629 * Above limits know nothing about real level of KVA fragmentation. 2630 * Start aggressive reclamation if too little sequential KVA left. 2631 */ 2632 if (vmem_size(heap_arena, VMEM_MAXFREE) < zfs_max_recordsize) 2633 return (1); 2634 2635#else /* _KERNEL */ 2636 if (spa_get_random(100) == 0) 2637 return (1); 2638#endif /* _KERNEL */ 2639 DTRACE_PROBE(arc__reclaim_no); 2640 2641 return (0); 2642} 2643 2644extern kmem_cache_t *zio_buf_cache[]; 2645extern kmem_cache_t *zio_data_buf_cache[]; 2646extern kmem_cache_t *range_seg_cache; 2647 2648static __noinline void 2649arc_kmem_reap_now(arc_reclaim_strategy_t strat) 2650{ 2651 size_t i; 2652 kmem_cache_t *prev_cache = NULL; 2653 kmem_cache_t *prev_data_cache = NULL; 2654 2655 DTRACE_PROBE(arc__kmem_reap_start); 2656#ifdef _KERNEL 2657 if (arc_meta_used >= arc_meta_limit) { 2658 /* 2659 * We are exceeding our meta-data cache limit. 2660 * Purge some DNLC entries to release holds on meta-data. 2661 */ 2662 dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent); 2663 } 2664#if defined(__i386) 2665 /* 2666 * Reclaim unused memory from all kmem caches. 2667 */ 2668 kmem_reap(); 2669#endif 2670#endif 2671 2672 /* 2673 * An aggressive reclamation will shrink the cache size as well as 2674 * reap free buffers from the arc kmem caches. 2675 */ 2676 if (strat == ARC_RECLAIM_AGGR) 2677 arc_shrink(); 2678 2679 for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) { 2680 if (zio_buf_cache[i] != prev_cache) { 2681 prev_cache = zio_buf_cache[i]; 2682 kmem_cache_reap_now(zio_buf_cache[i]); 2683 } 2684 if (zio_data_buf_cache[i] != prev_data_cache) { 2685 prev_data_cache = zio_data_buf_cache[i]; 2686 kmem_cache_reap_now(zio_data_buf_cache[i]); 2687 } 2688 } 2689 kmem_cache_reap_now(buf_cache); 2690 kmem_cache_reap_now(hdr_cache); 2691 kmem_cache_reap_now(range_seg_cache); 2692 2693#ifdef sun 2694 /* 2695 * Ask the vmem arena to reclaim unused memory from its 2696 * quantum caches. 2697 */ 2698 if (zio_arena != NULL && strat == ARC_RECLAIM_AGGR) 2699 vmem_qcache_reap(zio_arena); 2700#endif 2701 DTRACE_PROBE(arc__kmem_reap_end); 2702} 2703 2704static void 2705arc_reclaim_thread(void *dummy __unused) 2706{ 2707 clock_t growtime = 0; 2708 arc_reclaim_strategy_t last_reclaim = ARC_RECLAIM_CONS; 2709 callb_cpr_t cpr; 2710 2711 CALLB_CPR_INIT(&cpr, &arc_reclaim_thr_lock, callb_generic_cpr, FTAG); 2712 2713 mutex_enter(&arc_reclaim_thr_lock); 2714 while (arc_thread_exit == 0) { 2715 if (arc_reclaim_needed()) { 2716 2717 if (arc_no_grow) { 2718 if (last_reclaim == ARC_RECLAIM_CONS) { 2719 DTRACE_PROBE(arc__reclaim_aggr_no_grow); 2720 last_reclaim = ARC_RECLAIM_AGGR; 2721 } else { 2722 last_reclaim = ARC_RECLAIM_CONS; 2723 } 2724 } else { 2725 arc_no_grow = TRUE; 2726 last_reclaim = ARC_RECLAIM_AGGR; 2727 DTRACE_PROBE(arc__reclaim_aggr); 2728 membar_producer(); 2729 } 2730 2731 /* reset the growth delay for every reclaim */ 2732 growtime = ddi_get_lbolt() + (arc_grow_retry * hz); 2733 2734 if (needfree && last_reclaim == ARC_RECLAIM_CONS) { 2735 /* 2736 * If needfree is TRUE our vm_lowmem hook 2737 * was called and in that case we must free some 2738 * memory, so switch to aggressive mode. 2739 */ 2740 arc_no_grow = TRUE; 2741 last_reclaim = ARC_RECLAIM_AGGR; 2742 } 2743 arc_kmem_reap_now(last_reclaim); 2744 arc_warm = B_TRUE; 2745 2746 } else if (arc_no_grow && ddi_get_lbolt() >= growtime) { 2747 arc_no_grow = FALSE; 2748 } 2749 2750 arc_adjust(); 2751 2752 if (arc_eviction_list != NULL) 2753 arc_do_user_evicts(); 2754 2755#ifdef _KERNEL 2756 if (needfree) { 2757 needfree = 0; 2758 wakeup(&needfree); 2759 } 2760#endif 2761 2762 /* block until needed, or one second, whichever is shorter */ 2763 CALLB_CPR_SAFE_BEGIN(&cpr); 2764 (void) cv_timedwait(&arc_reclaim_thr_cv, 2765 &arc_reclaim_thr_lock, hz); 2766 CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_thr_lock); 2767 } 2768 2769 arc_thread_exit = 0; 2770 cv_broadcast(&arc_reclaim_thr_cv); 2771 CALLB_CPR_EXIT(&cpr); /* drops arc_reclaim_thr_lock */ 2772 thread_exit(); 2773} 2774 2775/* 2776 * Adapt arc info given the number of bytes we are trying to add and 2777 * the state that we are comming from. This function is only called 2778 * when we are adding new content to the cache. 2779 */ 2780static void 2781arc_adapt(int bytes, arc_state_t *state) 2782{ 2783 int mult; 2784 uint64_t arc_p_min = (arc_c >> arc_p_min_shift); 2785 2786 if (state == arc_l2c_only) 2787 return; 2788 2789 ASSERT(bytes > 0); 2790 /* 2791 * Adapt the target size of the MRU list: 2792 * - if we just hit in the MRU ghost list, then increase 2793 * the target size of the MRU list. 2794 * - if we just hit in the MFU ghost list, then increase 2795 * the target size of the MFU list by decreasing the 2796 * target size of the MRU list. 2797 */ 2798 if (state == arc_mru_ghost) { 2799 mult = ((arc_mru_ghost->arcs_size >= arc_mfu_ghost->arcs_size) ? 2800 1 : (arc_mfu_ghost->arcs_size/arc_mru_ghost->arcs_size)); 2801 mult = MIN(mult, 10); /* avoid wild arc_p adjustment */ 2802 2803 arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult); 2804 } else if (state == arc_mfu_ghost) { 2805 uint64_t delta; 2806 2807 mult = ((arc_mfu_ghost->arcs_size >= arc_mru_ghost->arcs_size) ? 2808 1 : (arc_mru_ghost->arcs_size/arc_mfu_ghost->arcs_size)); 2809 mult = MIN(mult, 10); 2810 2811 delta = MIN(bytes * mult, arc_p); 2812 arc_p = MAX(arc_p_min, arc_p - delta); 2813 } 2814 ASSERT((int64_t)arc_p >= 0); 2815 2816 if (arc_reclaim_needed()) { 2817 cv_signal(&arc_reclaim_thr_cv); 2818 return; 2819 } 2820 2821 if (arc_no_grow) 2822 return; 2823 2824 if (arc_c >= arc_c_max) 2825 return; 2826 2827 /* 2828 * If we're within (2 * maxblocksize) bytes of the target 2829 * cache size, increment the target cache size 2830 */ 2831 if (arc_size > arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) { 2832 DTRACE_PROBE1(arc__inc_adapt, int, bytes); 2833 atomic_add_64(&arc_c, (int64_t)bytes); 2834 if (arc_c > arc_c_max) 2835 arc_c = arc_c_max; 2836 else if (state == arc_anon) 2837 atomic_add_64(&arc_p, (int64_t)bytes); 2838 if (arc_p > arc_c) 2839 arc_p = arc_c; 2840 } 2841 ASSERT((int64_t)arc_p >= 0); 2842} 2843 2844/* 2845 * Check if the cache has reached its limits and eviction is required 2846 * prior to insert. 2847 */ 2848static int 2849arc_evict_needed(arc_buf_contents_t type) 2850{ 2851 if (type == ARC_BUFC_METADATA && arc_meta_used >= arc_meta_limit) 2852 return (1); 2853 2854 if (arc_reclaim_needed()) 2855 return (1); 2856 2857 return (arc_size > arc_c); 2858} 2859 2860/* 2861 * The buffer, supplied as the first argument, needs a data block. 2862 * So, if we are at cache max, determine which cache should be victimized. 2863 * We have the following cases: 2864 * 2865 * 1. Insert for MRU, p > sizeof(arc_anon + arc_mru) -> 2866 * In this situation if we're out of space, but the resident size of the MFU is 2867 * under the limit, victimize the MFU cache to satisfy this insertion request. 2868 * 2869 * 2. Insert for MRU, p <= sizeof(arc_anon + arc_mru) -> 2870 * Here, we've used up all of the available space for the MRU, so we need to 2871 * evict from our own cache instead. Evict from the set of resident MRU 2872 * entries. 2873 * 2874 * 3. Insert for MFU (c - p) > sizeof(arc_mfu) -> 2875 * c minus p represents the MFU space in the cache, since p is the size of the 2876 * cache that is dedicated to the MRU. In this situation there's still space on 2877 * the MFU side, so the MRU side needs to be victimized. 2878 * 2879 * 4. Insert for MFU (c - p) < sizeof(arc_mfu) -> 2880 * MFU's resident set is consuming more space than it has been allotted. In 2881 * this situation, we must victimize our own cache, the MFU, for this insertion. 2882 */ 2883static void 2884arc_get_data_buf(arc_buf_t *buf) 2885{ 2886 arc_state_t *state = buf->b_hdr->b_state; 2887 uint64_t size = buf->b_hdr->b_size; 2888 arc_buf_contents_t type = buf->b_hdr->b_type; 2889 2890 arc_adapt(size, state); 2891 2892 /* 2893 * We have not yet reached cache maximum size, 2894 * just allocate a new buffer. 2895 */ 2896 if (!arc_evict_needed(type)) { 2897 if (type == ARC_BUFC_METADATA) { 2898 buf->b_data = zio_buf_alloc(size); 2899 arc_space_consume(size, ARC_SPACE_DATA); 2900 } else { 2901 ASSERT(type == ARC_BUFC_DATA); 2902 buf->b_data = zio_data_buf_alloc(size); 2903 ARCSTAT_INCR(arcstat_data_size, size); 2904 atomic_add_64(&arc_size, size); 2905 } 2906 goto out; 2907 } 2908 2909 /* 2910 * If we are prefetching from the mfu ghost list, this buffer 2911 * will end up on the mru list; so steal space from there. 2912 */ 2913 if (state == arc_mfu_ghost) 2914 state = buf->b_hdr->b_flags & ARC_FLAG_PREFETCH ? 2915 arc_mru : arc_mfu; 2916 else if (state == arc_mru_ghost) 2917 state = arc_mru; 2918 2919 if (state == arc_mru || state == arc_anon) { 2920 uint64_t mru_used = arc_anon->arcs_size + arc_mru->arcs_size; 2921 state = (arc_mfu->arcs_lsize[type] >= size && 2922 arc_p > mru_used) ? arc_mfu : arc_mru; 2923 } else { 2924 /* MFU cases */ 2925 uint64_t mfu_space = arc_c - arc_p; 2926 state = (arc_mru->arcs_lsize[type] >= size && 2927 mfu_space > arc_mfu->arcs_size) ? arc_mru : arc_mfu; 2928 } 2929 if ((buf->b_data = arc_evict(state, 0, size, TRUE, type)) == NULL) { 2930 if (type == ARC_BUFC_METADATA) { 2931 buf->b_data = zio_buf_alloc(size); 2932 arc_space_consume(size, ARC_SPACE_DATA); 2933 } else { 2934 ASSERT(type == ARC_BUFC_DATA); 2935 buf->b_data = zio_data_buf_alloc(size); 2936 ARCSTAT_INCR(arcstat_data_size, size); 2937 atomic_add_64(&arc_size, size); 2938 } 2939 ARCSTAT_BUMP(arcstat_recycle_miss); 2940 } 2941 ASSERT(buf->b_data != NULL); 2942out: 2943 /* 2944 * Update the state size. Note that ghost states have a 2945 * "ghost size" and so don't need to be updated. 2946 */ 2947 if (!GHOST_STATE(buf->b_hdr->b_state)) { 2948 arc_buf_hdr_t *hdr = buf->b_hdr; 2949 2950 atomic_add_64(&hdr->b_state->arcs_size, size); 2951 if (list_link_active(&hdr->b_arc_node)) { 2952 ASSERT(refcount_is_zero(&hdr->b_refcnt)); 2953 atomic_add_64(&hdr->b_state->arcs_lsize[type], size); 2954 } 2955 /* 2956 * If we are growing the cache, and we are adding anonymous 2957 * data, and we have outgrown arc_p, update arc_p 2958 */ 2959 if (arc_size < arc_c && hdr->b_state == arc_anon && 2960 arc_anon->arcs_size + arc_mru->arcs_size > arc_p) 2961 arc_p = MIN(arc_c, arc_p + size); 2962 } 2963 ARCSTAT_BUMP(arcstat_allocated); 2964} 2965 2966/* 2967 * This routine is called whenever a buffer is accessed. 2968 * NOTE: the hash lock is dropped in this function. 2969 */ 2970static void 2971arc_access(arc_buf_hdr_t *hdr, kmutex_t *hash_lock) 2972{ 2973 clock_t now; 2974 2975 ASSERT(MUTEX_HELD(hash_lock)); 2976 2977 if (hdr->b_state == arc_anon) { 2978 /* 2979 * This buffer is not in the cache, and does not 2980 * appear in our "ghost" list. Add the new buffer 2981 * to the MRU state. 2982 */ 2983 2984 ASSERT(hdr->b_arc_access == 0); 2985 hdr->b_arc_access = ddi_get_lbolt(); 2986 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr); 2987 arc_change_state(arc_mru, hdr, hash_lock); 2988 2989 } else if (hdr->b_state == arc_mru) { 2990 now = ddi_get_lbolt(); 2991 2992 /* 2993 * If this buffer is here because of a prefetch, then either: 2994 * - clear the flag if this is a "referencing" read 2995 * (any subsequent access will bump this into the MFU state). 2996 * or 2997 * - move the buffer to the head of the list if this is 2998 * another prefetch (to make it less likely to be evicted). 2999 */ 3000 if ((hdr->b_flags & ARC_FLAG_PREFETCH) != 0) { 3001 if (refcount_count(&hdr->b_refcnt) == 0) { 3002 ASSERT(list_link_active(&hdr->b_arc_node)); 3003 } else { 3004 hdr->b_flags &= ~ARC_FLAG_PREFETCH; 3005 ARCSTAT_BUMP(arcstat_mru_hits); 3006 } 3007 hdr->b_arc_access = now; 3008 return; 3009 } 3010 3011 /* 3012 * This buffer has been "accessed" only once so far, 3013 * but it is still in the cache. Move it to the MFU 3014 * state. 3015 */ 3016 if (now > hdr->b_arc_access + ARC_MINTIME) { 3017 /* 3018 * More than 125ms have passed since we 3019 * instantiated this buffer. Move it to the 3020 * most frequently used state. 3021 */ 3022 hdr->b_arc_access = now; 3023 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 3024 arc_change_state(arc_mfu, hdr, hash_lock); 3025 } 3026 ARCSTAT_BUMP(arcstat_mru_hits); 3027 } else if (hdr->b_state == arc_mru_ghost) { 3028 arc_state_t *new_state; 3029 /* 3030 * This buffer has been "accessed" recently, but 3031 * was evicted from the cache. Move it to the 3032 * MFU state. 3033 */ 3034 3035 if (hdr->b_flags & ARC_FLAG_PREFETCH) { 3036 new_state = arc_mru; 3037 if (refcount_count(&hdr->b_refcnt) > 0) 3038 hdr->b_flags &= ~ARC_FLAG_PREFETCH; 3039 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr); 3040 } else { 3041 new_state = arc_mfu; 3042 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 3043 } 3044 3045 hdr->b_arc_access = ddi_get_lbolt(); 3046 arc_change_state(new_state, hdr, hash_lock); 3047 3048 ARCSTAT_BUMP(arcstat_mru_ghost_hits); 3049 } else if (hdr->b_state == arc_mfu) { 3050 /* 3051 * This buffer has been accessed more than once and is 3052 * still in the cache. Keep it in the MFU state. 3053 * 3054 * NOTE: an add_reference() that occurred when we did 3055 * the arc_read() will have kicked this off the list. 3056 * If it was a prefetch, we will explicitly move it to 3057 * the head of the list now. 3058 */ 3059 if ((hdr->b_flags & ARC_FLAG_PREFETCH) != 0) { 3060 ASSERT(refcount_count(&hdr->b_refcnt) == 0); 3061 ASSERT(list_link_active(&hdr->b_arc_node)); 3062 } 3063 ARCSTAT_BUMP(arcstat_mfu_hits); 3064 hdr->b_arc_access = ddi_get_lbolt(); 3065 } else if (hdr->b_state == arc_mfu_ghost) { 3066 arc_state_t *new_state = arc_mfu; 3067 /* 3068 * This buffer has been accessed more than once but has 3069 * been evicted from the cache. Move it back to the 3070 * MFU state. 3071 */ 3072 3073 if (hdr->b_flags & ARC_FLAG_PREFETCH) { 3074 /* 3075 * This is a prefetch access... 3076 * move this block back to the MRU state. 3077 */ 3078 ASSERT0(refcount_count(&hdr->b_refcnt)); 3079 new_state = arc_mru; 3080 } 3081 3082 hdr->b_arc_access = ddi_get_lbolt(); 3083 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 3084 arc_change_state(new_state, hdr, hash_lock); 3085 3086 ARCSTAT_BUMP(arcstat_mfu_ghost_hits); 3087 } else if (hdr->b_state == arc_l2c_only) { 3088 /* 3089 * This buffer is on the 2nd Level ARC. 3090 */ 3091 3092 hdr->b_arc_access = ddi_get_lbolt(); 3093 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 3094 arc_change_state(arc_mfu, hdr, hash_lock); 3095 } else { 3096 ASSERT(!"invalid arc state"); 3097 } 3098} 3099 3100/* a generic arc_done_func_t which you can use */ 3101/* ARGSUSED */ 3102void 3103arc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg) 3104{ 3105 if (zio == NULL || zio->io_error == 0) 3106 bcopy(buf->b_data, arg, buf->b_hdr->b_size); 3107 VERIFY(arc_buf_remove_ref(buf, arg)); 3108} 3109 3110/* a generic arc_done_func_t */ 3111void 3112arc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg) 3113{ 3114 arc_buf_t **bufp = arg; 3115 if (zio && zio->io_error) { 3116 VERIFY(arc_buf_remove_ref(buf, arg)); 3117 *bufp = NULL; 3118 } else { 3119 *bufp = buf; 3120 ASSERT(buf->b_data); 3121 } 3122} 3123 3124static void 3125arc_read_done(zio_t *zio) 3126{ 3127 arc_buf_hdr_t *hdr; 3128 arc_buf_t *buf; 3129 arc_buf_t *abuf; /* buffer we're assigning to callback */ 3130 kmutex_t *hash_lock = NULL; 3131 arc_callback_t *callback_list, *acb; 3132 int freeable = FALSE; 3133 3134 buf = zio->io_private; 3135 hdr = buf->b_hdr; 3136 3137 /* 3138 * The hdr was inserted into hash-table and removed from lists 3139 * prior to starting I/O. We should find this header, since 3140 * it's in the hash table, and it should be legit since it's 3141 * not possible to evict it during the I/O. The only possible 3142 * reason for it not to be found is if we were freed during the 3143 * read. 3144 */ 3145 if (HDR_IN_HASH_TABLE(hdr)) { 3146 ASSERT3U(hdr->b_birth, ==, BP_PHYSICAL_BIRTH(zio->io_bp)); 3147 ASSERT3U(hdr->b_dva.dva_word[0], ==, 3148 BP_IDENTITY(zio->io_bp)->dva_word[0]); 3149 ASSERT3U(hdr->b_dva.dva_word[1], ==, 3150 BP_IDENTITY(zio->io_bp)->dva_word[1]); 3151 3152 arc_buf_hdr_t *found = buf_hash_find(hdr->b_spa, zio->io_bp, 3153 &hash_lock); 3154 3155 ASSERT((found == NULL && HDR_FREED_IN_READ(hdr) && 3156 hash_lock == NULL) || 3157 (found == hdr && 3158 DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) || 3159 (found == hdr && HDR_L2_READING(hdr))); 3160 } 3161 3162 hdr->b_flags &= ~ARC_FLAG_L2_EVICTED; 3163 if (l2arc_noprefetch && (hdr->b_flags & ARC_FLAG_PREFETCH)) 3164 hdr->b_flags &= ~ARC_FLAG_L2CACHE; 3165 3166 /* byteswap if necessary */ 3167 callback_list = hdr->b_acb; 3168 ASSERT(callback_list != NULL); 3169 if (BP_SHOULD_BYTESWAP(zio->io_bp) && zio->io_error == 0) { 3170 dmu_object_byteswap_t bswap = 3171 DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp)); 3172 arc_byteswap_func_t *func = BP_GET_LEVEL(zio->io_bp) > 0 ? 3173 byteswap_uint64_array : 3174 dmu_ot_byteswap[bswap].ob_func; 3175 func(buf->b_data, hdr->b_size); 3176 } 3177 3178 arc_cksum_compute(buf, B_FALSE); 3179#ifdef illumos 3180 arc_buf_watch(buf); 3181#endif /* illumos */ 3182 3183 if (hash_lock && zio->io_error == 0 && hdr->b_state == arc_anon) { 3184 /* 3185 * Only call arc_access on anonymous buffers. This is because 3186 * if we've issued an I/O for an evicted buffer, we've already 3187 * called arc_access (to prevent any simultaneous readers from 3188 * getting confused). 3189 */ 3190 arc_access(hdr, hash_lock); 3191 } 3192 3193 /* create copies of the data buffer for the callers */ 3194 abuf = buf; 3195 for (acb = callback_list; acb; acb = acb->acb_next) { 3196 if (acb->acb_done) { 3197 if (abuf == NULL) { 3198 ARCSTAT_BUMP(arcstat_duplicate_reads); 3199 abuf = arc_buf_clone(buf); 3200 } 3201 acb->acb_buf = abuf; 3202 abuf = NULL; 3203 } 3204 } 3205 hdr->b_acb = NULL; 3206 hdr->b_flags &= ~ARC_FLAG_IO_IN_PROGRESS; 3207 ASSERT(!HDR_BUF_AVAILABLE(hdr)); 3208 if (abuf == buf) { 3209 ASSERT(buf->b_efunc == NULL); 3210 ASSERT(hdr->b_datacnt == 1); 3211 hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE; 3212 } 3213 3214 ASSERT(refcount_is_zero(&hdr->b_refcnt) || callback_list != NULL); 3215 3216 if (zio->io_error != 0) { 3217 hdr->b_flags |= ARC_FLAG_IO_ERROR; 3218 if (hdr->b_state != arc_anon) 3219 arc_change_state(arc_anon, hdr, hash_lock); 3220 if (HDR_IN_HASH_TABLE(hdr)) 3221 buf_hash_remove(hdr); 3222 freeable = refcount_is_zero(&hdr->b_refcnt); 3223 } 3224 3225 /* 3226 * Broadcast before we drop the hash_lock to avoid the possibility 3227 * that the hdr (and hence the cv) might be freed before we get to 3228 * the cv_broadcast(). 3229 */ 3230 cv_broadcast(&hdr->b_cv); 3231 3232 if (hash_lock) { 3233 mutex_exit(hash_lock); 3234 } else { 3235 /* 3236 * This block was freed while we waited for the read to 3237 * complete. It has been removed from the hash table and 3238 * moved to the anonymous state (so that it won't show up 3239 * in the cache). 3240 */ 3241 ASSERT3P(hdr->b_state, ==, arc_anon); 3242 freeable = refcount_is_zero(&hdr->b_refcnt); 3243 } 3244 3245 /* execute each callback and free its structure */ 3246 while ((acb = callback_list) != NULL) { 3247 if (acb->acb_done) 3248 acb->acb_done(zio, acb->acb_buf, acb->acb_private); 3249 3250 if (acb->acb_zio_dummy != NULL) { 3251 acb->acb_zio_dummy->io_error = zio->io_error; 3252 zio_nowait(acb->acb_zio_dummy); 3253 } 3254 3255 callback_list = acb->acb_next; 3256 kmem_free(acb, sizeof (arc_callback_t)); 3257 } 3258 3259 if (freeable) 3260 arc_hdr_destroy(hdr); 3261} 3262 3263/* 3264 * "Read" the block block at the specified DVA (in bp) via the 3265 * cache. If the block is found in the cache, invoke the provided 3266 * callback immediately and return. Note that the `zio' parameter 3267 * in the callback will be NULL in this case, since no IO was 3268 * required. If the block is not in the cache pass the read request 3269 * on to the spa with a substitute callback function, so that the 3270 * requested block will be added to the cache. 3271 * 3272 * If a read request arrives for a block that has a read in-progress, 3273 * either wait for the in-progress read to complete (and return the 3274 * results); or, if this is a read with a "done" func, add a record 3275 * to the read to invoke the "done" func when the read completes, 3276 * and return; or just return. 3277 * 3278 * arc_read_done() will invoke all the requested "done" functions 3279 * for readers of this block. 3280 */ 3281int 3282arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_done_func_t *done, 3283 void *private, zio_priority_t priority, int zio_flags, 3284 arc_flags_t *arc_flags, const zbookmark_phys_t *zb) 3285{ 3286 arc_buf_hdr_t *hdr = NULL; 3287 arc_buf_t *buf = NULL; 3288 kmutex_t *hash_lock = NULL; 3289 zio_t *rzio; 3290 uint64_t guid = spa_load_guid(spa); 3291 3292 ASSERT(!BP_IS_EMBEDDED(bp) || 3293 BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA); 3294 3295top: 3296 if (!BP_IS_EMBEDDED(bp)) { 3297 /* 3298 * Embedded BP's have no DVA and require no I/O to "read". 3299 * Create an anonymous arc buf to back it. 3300 */ 3301 hdr = buf_hash_find(guid, bp, &hash_lock); 3302 } 3303 3304 if (hdr != NULL && hdr->b_datacnt > 0) { 3305 3306 *arc_flags |= ARC_FLAG_CACHED; 3307 3308 if (HDR_IO_IN_PROGRESS(hdr)) { 3309 3310 if (*arc_flags & ARC_FLAG_WAIT) { 3311 cv_wait(&hdr->b_cv, hash_lock); 3312 mutex_exit(hash_lock); 3313 goto top; 3314 } 3315 ASSERT(*arc_flags & ARC_FLAG_NOWAIT); 3316 3317 if (done) { 3318 arc_callback_t *acb = NULL; 3319 3320 acb = kmem_zalloc(sizeof (arc_callback_t), 3321 KM_SLEEP); 3322 acb->acb_done = done; 3323 acb->acb_private = private; 3324 if (pio != NULL) 3325 acb->acb_zio_dummy = zio_null(pio, 3326 spa, NULL, NULL, NULL, zio_flags); 3327 3328 ASSERT(acb->acb_done != NULL); 3329 acb->acb_next = hdr->b_acb; 3330 hdr->b_acb = acb; 3331 add_reference(hdr, hash_lock, private); 3332 mutex_exit(hash_lock); 3333 return (0); 3334 } 3335 mutex_exit(hash_lock); 3336 return (0); 3337 } 3338 3339 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu); 3340 3341 if (done) { 3342 add_reference(hdr, hash_lock, private); 3343 /* 3344 * If this block is already in use, create a new 3345 * copy of the data so that we will be guaranteed 3346 * that arc_release() will always succeed. 3347 */ 3348 buf = hdr->b_buf; 3349 ASSERT(buf); 3350 ASSERT(buf->b_data); 3351 if (HDR_BUF_AVAILABLE(hdr)) { 3352 ASSERT(buf->b_efunc == NULL); 3353 hdr->b_flags &= ~ARC_FLAG_BUF_AVAILABLE; 3354 } else { 3355 buf = arc_buf_clone(buf); 3356 } 3357 3358 } else if (*arc_flags & ARC_FLAG_PREFETCH && 3359 refcount_count(&hdr->b_refcnt) == 0) { 3360 hdr->b_flags |= ARC_FLAG_PREFETCH; 3361 } 3362 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr); 3363 arc_access(hdr, hash_lock); 3364 if (*arc_flags & ARC_FLAG_L2CACHE) 3365 hdr->b_flags |= ARC_FLAG_L2CACHE; 3366 if (*arc_flags & ARC_FLAG_L2COMPRESS) 3367 hdr->b_flags |= ARC_FLAG_L2COMPRESS; 3368 mutex_exit(hash_lock); 3369 ARCSTAT_BUMP(arcstat_hits); 3370 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_FLAG_PREFETCH), 3371 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA, 3372 data, metadata, hits); 3373 3374 if (done) 3375 done(NULL, buf, private); 3376 } else { 3377 uint64_t size = BP_GET_LSIZE(bp); 3378 arc_callback_t *acb; 3379 vdev_t *vd = NULL; 3380 uint64_t addr = 0; 3381 boolean_t devw = B_FALSE; 3382 enum zio_compress b_compress = ZIO_COMPRESS_OFF; 3383 uint64_t b_asize = 0; 3384 3385 if (hdr == NULL) { 3386 /* this block is not in the cache */ 3387 arc_buf_hdr_t *exists = NULL; 3388 arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp); 3389 buf = arc_buf_alloc(spa, size, private, type); 3390 hdr = buf->b_hdr; 3391 if (!BP_IS_EMBEDDED(bp)) { 3392 hdr->b_dva = *BP_IDENTITY(bp); 3393 hdr->b_birth = BP_PHYSICAL_BIRTH(bp); 3394 hdr->b_cksum0 = bp->blk_cksum.zc_word[0]; 3395 exists = buf_hash_insert(hdr, &hash_lock); 3396 } 3397 if (exists != NULL) { 3398 /* somebody beat us to the hash insert */ 3399 mutex_exit(hash_lock); 3400 buf_discard_identity(hdr); 3401 (void) arc_buf_remove_ref(buf, private); 3402 goto top; /* restart the IO request */ 3403 } 3404 3405 /* if this is a prefetch, we don't have a reference */ 3406 if (*arc_flags & ARC_FLAG_PREFETCH) { 3407 (void) remove_reference(hdr, hash_lock, 3408 private); 3409 hdr->b_flags |= ARC_FLAG_PREFETCH; 3410 } 3411 if (*arc_flags & ARC_FLAG_L2CACHE) 3412 hdr->b_flags |= ARC_FLAG_L2CACHE; 3413 if (*arc_flags & ARC_FLAG_L2COMPRESS) 3414 hdr->b_flags |= ARC_FLAG_L2COMPRESS; 3415 if (BP_GET_LEVEL(bp) > 0) 3416 hdr->b_flags |= ARC_FLAG_INDIRECT; 3417 } else { 3418 /* this block is in the ghost cache */ 3419 ASSERT(GHOST_STATE(hdr->b_state)); 3420 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 3421 ASSERT0(refcount_count(&hdr->b_refcnt)); 3422 ASSERT(hdr->b_buf == NULL); 3423 3424 /* if this is a prefetch, we don't have a reference */ 3425 if (*arc_flags & ARC_FLAG_PREFETCH) 3426 hdr->b_flags |= ARC_FLAG_PREFETCH; 3427 else 3428 add_reference(hdr, hash_lock, private); 3429 if (*arc_flags & ARC_FLAG_L2CACHE) 3430 hdr->b_flags |= ARC_FLAG_L2CACHE; 3431 if (*arc_flags & ARC_FLAG_L2COMPRESS) 3432 hdr->b_flags |= ARC_FLAG_L2COMPRESS; 3433 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 3434 buf->b_hdr = hdr; 3435 buf->b_data = NULL; 3436 buf->b_efunc = NULL; 3437 buf->b_private = NULL; 3438 buf->b_next = NULL; 3439 hdr->b_buf = buf; 3440 ASSERT(hdr->b_datacnt == 0); 3441 hdr->b_datacnt = 1; 3442 arc_get_data_buf(buf); 3443 arc_access(hdr, hash_lock); 3444 } 3445 3446 ASSERT(!GHOST_STATE(hdr->b_state)); 3447 3448 acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP); 3449 acb->acb_done = done; 3450 acb->acb_private = private; 3451 3452 ASSERT(hdr->b_acb == NULL); 3453 hdr->b_acb = acb; 3454 hdr->b_flags |= ARC_FLAG_IO_IN_PROGRESS; 3455 3456 if (hdr->b_l2hdr != NULL && 3457 (vd = hdr->b_l2hdr->b_dev->l2ad_vdev) != NULL) { 3458 devw = hdr->b_l2hdr->b_dev->l2ad_writing; 3459 addr = hdr->b_l2hdr->b_daddr; 3460 b_compress = hdr->b_l2hdr->b_compress; 3461 b_asize = hdr->b_l2hdr->b_asize; 3462 /* 3463 * Lock out device removal. 3464 */ 3465 if (vdev_is_dead(vd) || 3466 !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER)) 3467 vd = NULL; 3468 } 3469 3470 if (hash_lock != NULL) 3471 mutex_exit(hash_lock); 3472 3473 /* 3474 * At this point, we have a level 1 cache miss. Try again in 3475 * L2ARC if possible. 3476 */ 3477 ASSERT3U(hdr->b_size, ==, size); 3478 DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp, 3479 uint64_t, size, zbookmark_phys_t *, zb); 3480 ARCSTAT_BUMP(arcstat_misses); 3481 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_FLAG_PREFETCH), 3482 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA, 3483 data, metadata, misses); 3484#ifdef _KERNEL 3485 curthread->td_ru.ru_inblock++; 3486#endif 3487 3488 if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) { 3489 /* 3490 * Read from the L2ARC if the following are true: 3491 * 1. The L2ARC vdev was previously cached. 3492 * 2. This buffer still has L2ARC metadata. 3493 * 3. This buffer isn't currently writing to the L2ARC. 3494 * 4. The L2ARC entry wasn't evicted, which may 3495 * also have invalidated the vdev. 3496 * 5. This isn't prefetch and l2arc_noprefetch is set. 3497 */ 3498 if (hdr->b_l2hdr != NULL && 3499 !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) && 3500 !(l2arc_noprefetch && HDR_PREFETCH(hdr))) { 3501 l2arc_read_callback_t *cb; 3502 3503 DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr); 3504 ARCSTAT_BUMP(arcstat_l2_hits); 3505 3506 cb = kmem_zalloc(sizeof (l2arc_read_callback_t), 3507 KM_SLEEP); 3508 cb->l2rcb_buf = buf; 3509 cb->l2rcb_spa = spa; 3510 cb->l2rcb_bp = *bp; 3511 cb->l2rcb_zb = *zb; 3512 cb->l2rcb_flags = zio_flags; 3513 cb->l2rcb_compress = b_compress; 3514 3515 ASSERT(addr >= VDEV_LABEL_START_SIZE && 3516 addr + size < vd->vdev_psize - 3517 VDEV_LABEL_END_SIZE); 3518 3519 /* 3520 * l2arc read. The SCL_L2ARC lock will be 3521 * released by l2arc_read_done(). 3522 * Issue a null zio if the underlying buffer 3523 * was squashed to zero size by compression. 3524 */ 3525 if (b_compress == ZIO_COMPRESS_EMPTY) { 3526 rzio = zio_null(pio, spa, vd, 3527 l2arc_read_done, cb, 3528 zio_flags | ZIO_FLAG_DONT_CACHE | 3529 ZIO_FLAG_CANFAIL | 3530 ZIO_FLAG_DONT_PROPAGATE | 3531 ZIO_FLAG_DONT_RETRY); 3532 } else { 3533 rzio = zio_read_phys(pio, vd, addr, 3534 b_asize, buf->b_data, 3535 ZIO_CHECKSUM_OFF, 3536 l2arc_read_done, cb, priority, 3537 zio_flags | ZIO_FLAG_DONT_CACHE | 3538 ZIO_FLAG_CANFAIL | 3539 ZIO_FLAG_DONT_PROPAGATE | 3540 ZIO_FLAG_DONT_RETRY, B_FALSE); 3541 } 3542 DTRACE_PROBE2(l2arc__read, vdev_t *, vd, 3543 zio_t *, rzio); 3544 ARCSTAT_INCR(arcstat_l2_read_bytes, b_asize); 3545 3546 if (*arc_flags & ARC_FLAG_NOWAIT) { 3547 zio_nowait(rzio); 3548 return (0); 3549 } 3550 3551 ASSERT(*arc_flags & ARC_FLAG_WAIT); 3552 if (zio_wait(rzio) == 0) 3553 return (0); 3554 3555 /* l2arc read error; goto zio_read() */ 3556 } else { 3557 DTRACE_PROBE1(l2arc__miss, 3558 arc_buf_hdr_t *, hdr); 3559 ARCSTAT_BUMP(arcstat_l2_misses); 3560 if (HDR_L2_WRITING(hdr)) 3561 ARCSTAT_BUMP(arcstat_l2_rw_clash); 3562 spa_config_exit(spa, SCL_L2ARC, vd); 3563 } 3564 } else { 3565 if (vd != NULL) 3566 spa_config_exit(spa, SCL_L2ARC, vd); 3567 if (l2arc_ndev != 0) { 3568 DTRACE_PROBE1(l2arc__miss, 3569 arc_buf_hdr_t *, hdr); 3570 ARCSTAT_BUMP(arcstat_l2_misses); 3571 } 3572 } 3573 3574 rzio = zio_read(pio, spa, bp, buf->b_data, size, 3575 arc_read_done, buf, priority, zio_flags, zb); 3576 3577 if (*arc_flags & ARC_FLAG_WAIT) 3578 return (zio_wait(rzio)); 3579 3580 ASSERT(*arc_flags & ARC_FLAG_NOWAIT); 3581 zio_nowait(rzio); 3582 } 3583 return (0); 3584} 3585 3586void 3587arc_set_callback(arc_buf_t *buf, arc_evict_func_t *func, void *private) 3588{ 3589 ASSERT(buf->b_hdr != NULL); 3590 ASSERT(buf->b_hdr->b_state != arc_anon); 3591 ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt) || func == NULL); 3592 ASSERT(buf->b_efunc == NULL); 3593 ASSERT(!HDR_BUF_AVAILABLE(buf->b_hdr)); 3594 3595 buf->b_efunc = func; 3596 buf->b_private = private; 3597} 3598 3599/* 3600 * Notify the arc that a block was freed, and thus will never be used again. 3601 */ 3602void 3603arc_freed(spa_t *spa, const blkptr_t *bp) 3604{ 3605 arc_buf_hdr_t *hdr; 3606 kmutex_t *hash_lock; 3607 uint64_t guid = spa_load_guid(spa); 3608 3609 ASSERT(!BP_IS_EMBEDDED(bp)); 3610 3611 hdr = buf_hash_find(guid, bp, &hash_lock); 3612 if (hdr == NULL) 3613 return; 3614 if (HDR_BUF_AVAILABLE(hdr)) { 3615 arc_buf_t *buf = hdr->b_buf; 3616 add_reference(hdr, hash_lock, FTAG); 3617 hdr->b_flags &= ~ARC_FLAG_BUF_AVAILABLE; 3618 mutex_exit(hash_lock); 3619 3620 arc_release(buf, FTAG); 3621 (void) arc_buf_remove_ref(buf, FTAG); 3622 } else { 3623 mutex_exit(hash_lock); 3624 } 3625 3626} 3627 3628/* 3629 * Clear the user eviction callback set by arc_set_callback(), first calling 3630 * it if it exists. Because the presence of a callback keeps an arc_buf cached 3631 * clearing the callback may result in the arc_buf being destroyed. However, 3632 * it will not result in the *last* arc_buf being destroyed, hence the data 3633 * will remain cached in the ARC. We make a copy of the arc buffer here so 3634 * that we can process the callback without holding any locks. 3635 * 3636 * It's possible that the callback is already in the process of being cleared 3637 * by another thread. In this case we can not clear the callback. 3638 * 3639 * Returns B_TRUE if the callback was successfully called and cleared. 3640 */ 3641boolean_t 3642arc_clear_callback(arc_buf_t *buf) 3643{ 3644 arc_buf_hdr_t *hdr; 3645 kmutex_t *hash_lock; 3646 arc_evict_func_t *efunc = buf->b_efunc; 3647 void *private = buf->b_private; 3648 list_t *list, *evicted_list; 3649 kmutex_t *lock, *evicted_lock; 3650 3651 mutex_enter(&buf->b_evict_lock); 3652 hdr = buf->b_hdr; 3653 if (hdr == NULL) { 3654 /* 3655 * We are in arc_do_user_evicts(). 3656 */ 3657 ASSERT(buf->b_data == NULL); 3658 mutex_exit(&buf->b_evict_lock); 3659 return (B_FALSE); 3660 } else if (buf->b_data == NULL) { 3661 /* 3662 * We are on the eviction list; process this buffer now 3663 * but let arc_do_user_evicts() do the reaping. 3664 */ 3665 buf->b_efunc = NULL; 3666 mutex_exit(&buf->b_evict_lock); 3667 VERIFY0(efunc(private)); 3668 return (B_TRUE); 3669 } 3670 hash_lock = HDR_LOCK(hdr); 3671 mutex_enter(hash_lock); 3672 hdr = buf->b_hdr; 3673 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 3674 3675 ASSERT3U(refcount_count(&hdr->b_refcnt), <, hdr->b_datacnt); 3676 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu); 3677 3678 buf->b_efunc = NULL; 3679 buf->b_private = NULL; 3680 3681 if (hdr->b_datacnt > 1) { 3682 mutex_exit(&buf->b_evict_lock); 3683 arc_buf_destroy(buf, FALSE, TRUE); 3684 } else { 3685 ASSERT(buf == hdr->b_buf); 3686 hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE; 3687 mutex_exit(&buf->b_evict_lock); 3688 } 3689 3690 mutex_exit(hash_lock); 3691 VERIFY0(efunc(private)); 3692 return (B_TRUE); 3693} 3694 3695/* 3696 * Release this buffer from the cache, making it an anonymous buffer. This 3697 * must be done after a read and prior to modifying the buffer contents. 3698 * If the buffer has more than one reference, we must make 3699 * a new hdr for the buffer. 3700 */ 3701void 3702arc_release(arc_buf_t *buf, void *tag) 3703{ 3704 arc_buf_hdr_t *hdr; 3705 kmutex_t *hash_lock = NULL; 3706 l2arc_buf_hdr_t *l2hdr; 3707 uint64_t buf_size; 3708 3709 /* 3710 * It would be nice to assert that if it's DMU metadata (level > 3711 * 0 || it's the dnode file), then it must be syncing context. 3712 * But we don't know that information at this level. 3713 */ 3714 3715 mutex_enter(&buf->b_evict_lock); 3716 hdr = buf->b_hdr; 3717 3718 /* this buffer is not on any list */ 3719 ASSERT(refcount_count(&hdr->b_refcnt) > 0); 3720 3721 if (hdr->b_state == arc_anon) { 3722 /* this buffer is already released */ 3723 ASSERT(buf->b_efunc == NULL); 3724 } else { 3725 hash_lock = HDR_LOCK(hdr); 3726 mutex_enter(hash_lock); 3727 hdr = buf->b_hdr; 3728 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 3729 } 3730 3731 l2hdr = hdr->b_l2hdr; 3732 if (l2hdr) { 3733 mutex_enter(&l2arc_buflist_mtx); 3734 arc_buf_l2_cdata_free(hdr); 3735 hdr->b_l2hdr = NULL; 3736 list_remove(l2hdr->b_dev->l2ad_buflist, hdr); 3737 } 3738 buf_size = hdr->b_size; 3739 3740 /* 3741 * Do we have more than one buf? 3742 */ 3743 if (hdr->b_datacnt > 1) { 3744 arc_buf_hdr_t *nhdr; 3745 arc_buf_t **bufp; 3746 uint64_t blksz = hdr->b_size; 3747 uint64_t spa = hdr->b_spa; 3748 arc_buf_contents_t type = hdr->b_type; 3749 uint32_t flags = hdr->b_flags; 3750 3751 ASSERT(hdr->b_buf != buf || buf->b_next != NULL); 3752 /* 3753 * Pull the data off of this hdr and attach it to 3754 * a new anonymous hdr. 3755 */ 3756 (void) remove_reference(hdr, hash_lock, tag); 3757 bufp = &hdr->b_buf; 3758 while (*bufp != buf) 3759 bufp = &(*bufp)->b_next; 3760 *bufp = buf->b_next; 3761 buf->b_next = NULL; 3762 3763 ASSERT3U(hdr->b_state->arcs_size, >=, hdr->b_size); 3764 atomic_add_64(&hdr->b_state->arcs_size, -hdr->b_size); 3765 if (refcount_is_zero(&hdr->b_refcnt)) { 3766 uint64_t *size = &hdr->b_state->arcs_lsize[hdr->b_type]; 3767 ASSERT3U(*size, >=, hdr->b_size); 3768 atomic_add_64(size, -hdr->b_size); 3769 } 3770 3771 /* 3772 * We're releasing a duplicate user data buffer, update 3773 * our statistics accordingly. 3774 */ 3775 if (hdr->b_type == ARC_BUFC_DATA) { 3776 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers); 3777 ARCSTAT_INCR(arcstat_duplicate_buffers_size, 3778 -hdr->b_size); 3779 } 3780 hdr->b_datacnt -= 1; 3781 arc_cksum_verify(buf); 3782#ifdef illumos 3783 arc_buf_unwatch(buf); 3784#endif /* illumos */ 3785 3786 mutex_exit(hash_lock); 3787 3788 nhdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE); 3789 nhdr->b_size = blksz; 3790 nhdr->b_spa = spa; 3791 nhdr->b_type = type; 3792 nhdr->b_buf = buf; 3793 nhdr->b_state = arc_anon; 3794 nhdr->b_arc_access = 0; 3795 nhdr->b_flags = flags & ARC_FLAG_L2_WRITING; 3796 nhdr->b_l2hdr = NULL; 3797 nhdr->b_datacnt = 1; 3798 nhdr->b_freeze_cksum = NULL; 3799 (void) refcount_add(&nhdr->b_refcnt, tag); 3800 buf->b_hdr = nhdr; 3801 mutex_exit(&buf->b_evict_lock); 3802 atomic_add_64(&arc_anon->arcs_size, blksz); 3803 } else { 3804 mutex_exit(&buf->b_evict_lock); 3805 ASSERT(refcount_count(&hdr->b_refcnt) == 1); 3806 ASSERT(!list_link_active(&hdr->b_arc_node)); 3807 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 3808 if (hdr->b_state != arc_anon) 3809 arc_change_state(arc_anon, hdr, hash_lock); 3810 hdr->b_arc_access = 0; 3811 if (hash_lock) 3812 mutex_exit(hash_lock); 3813 3814 buf_discard_identity(hdr); 3815 arc_buf_thaw(buf); 3816 } 3817 buf->b_efunc = NULL; 3818 buf->b_private = NULL; 3819 3820 if (l2hdr) { 3821 ARCSTAT_INCR(arcstat_l2_asize, -l2hdr->b_asize); 3822 vdev_space_update(l2hdr->b_dev->l2ad_vdev, 3823 -l2hdr->b_asize, 0, 0); 3824 trim_map_free(l2hdr->b_dev->l2ad_vdev, l2hdr->b_daddr, 3825 l2hdr->b_asize, 0); 3826 kmem_free(l2hdr, sizeof (l2arc_buf_hdr_t)); 3827 ARCSTAT_INCR(arcstat_l2_size, -buf_size); 3828 mutex_exit(&l2arc_buflist_mtx); 3829 } 3830} 3831 3832int 3833arc_released(arc_buf_t *buf) 3834{ 3835 int released; 3836 3837 mutex_enter(&buf->b_evict_lock); 3838 released = (buf->b_data != NULL && buf->b_hdr->b_state == arc_anon); 3839 mutex_exit(&buf->b_evict_lock); 3840 return (released); 3841} 3842 3843#ifdef ZFS_DEBUG 3844int 3845arc_referenced(arc_buf_t *buf) 3846{ 3847 int referenced; 3848 3849 mutex_enter(&buf->b_evict_lock); 3850 referenced = (refcount_count(&buf->b_hdr->b_refcnt)); 3851 mutex_exit(&buf->b_evict_lock); 3852 return (referenced); 3853} 3854#endif 3855 3856static void 3857arc_write_ready(zio_t *zio) 3858{ 3859 arc_write_callback_t *callback = zio->io_private; 3860 arc_buf_t *buf = callback->awcb_buf; 3861 arc_buf_hdr_t *hdr = buf->b_hdr; 3862 3863 ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt)); 3864 callback->awcb_ready(zio, buf, callback->awcb_private); 3865 3866 /* 3867 * If the IO is already in progress, then this is a re-write 3868 * attempt, so we need to thaw and re-compute the cksum. 3869 * It is the responsibility of the callback to handle the 3870 * accounting for any re-write attempt. 3871 */ 3872 if (HDR_IO_IN_PROGRESS(hdr)) { 3873 mutex_enter(&hdr->b_freeze_lock); 3874 if (hdr->b_freeze_cksum != NULL) { 3875 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 3876 hdr->b_freeze_cksum = NULL; 3877 } 3878 mutex_exit(&hdr->b_freeze_lock); 3879 } 3880 arc_cksum_compute(buf, B_FALSE); 3881 hdr->b_flags |= ARC_FLAG_IO_IN_PROGRESS; 3882} 3883 3884/* 3885 * The SPA calls this callback for each physical write that happens on behalf 3886 * of a logical write. See the comment in dbuf_write_physdone() for details. 3887 */ 3888static void 3889arc_write_physdone(zio_t *zio) 3890{ 3891 arc_write_callback_t *cb = zio->io_private; 3892 if (cb->awcb_physdone != NULL) 3893 cb->awcb_physdone(zio, cb->awcb_buf, cb->awcb_private); 3894} 3895 3896static void 3897arc_write_done(zio_t *zio) 3898{ 3899 arc_write_callback_t *callback = zio->io_private; 3900 arc_buf_t *buf = callback->awcb_buf; 3901 arc_buf_hdr_t *hdr = buf->b_hdr; 3902 3903 ASSERT(hdr->b_acb == NULL); 3904 3905 if (zio->io_error == 0) { 3906 if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) { 3907 buf_discard_identity(hdr); 3908 } else { 3909 hdr->b_dva = *BP_IDENTITY(zio->io_bp); 3910 hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp); 3911 hdr->b_cksum0 = zio->io_bp->blk_cksum.zc_word[0]; 3912 } 3913 } else { 3914 ASSERT(BUF_EMPTY(hdr)); 3915 } 3916 3917 /* 3918 * If the block to be written was all-zero or compressed enough to be 3919 * embedded in the BP, no write was performed so there will be no 3920 * dva/birth/checksum. The buffer must therefore remain anonymous 3921 * (and uncached). 3922 */ 3923 if (!BUF_EMPTY(hdr)) { 3924 arc_buf_hdr_t *exists; 3925 kmutex_t *hash_lock; 3926 3927 ASSERT(zio->io_error == 0); 3928 3929 arc_cksum_verify(buf); 3930 3931 exists = buf_hash_insert(hdr, &hash_lock); 3932 if (exists) { 3933 /* 3934 * This can only happen if we overwrite for 3935 * sync-to-convergence, because we remove 3936 * buffers from the hash table when we arc_free(). 3937 */ 3938 if (zio->io_flags & ZIO_FLAG_IO_REWRITE) { 3939 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) 3940 panic("bad overwrite, hdr=%p exists=%p", 3941 (void *)hdr, (void *)exists); 3942 ASSERT(refcount_is_zero(&exists->b_refcnt)); 3943 arc_change_state(arc_anon, exists, hash_lock); 3944 mutex_exit(hash_lock); 3945 arc_hdr_destroy(exists); 3946 exists = buf_hash_insert(hdr, &hash_lock); 3947 ASSERT3P(exists, ==, NULL); 3948 } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) { 3949 /* nopwrite */ 3950 ASSERT(zio->io_prop.zp_nopwrite); 3951 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) 3952 panic("bad nopwrite, hdr=%p exists=%p", 3953 (void *)hdr, (void *)exists); 3954 } else { 3955 /* Dedup */ 3956 ASSERT(hdr->b_datacnt == 1); 3957 ASSERT(hdr->b_state == arc_anon); 3958 ASSERT(BP_GET_DEDUP(zio->io_bp)); 3959 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0); 3960 } 3961 } 3962 hdr->b_flags &= ~ARC_FLAG_IO_IN_PROGRESS; 3963 /* if it's not anon, we are doing a scrub */ 3964 if (!exists && hdr->b_state == arc_anon) 3965 arc_access(hdr, hash_lock); 3966 mutex_exit(hash_lock); 3967 } else { 3968 hdr->b_flags &= ~ARC_FLAG_IO_IN_PROGRESS; 3969 } 3970 3971 ASSERT(!refcount_is_zero(&hdr->b_refcnt)); 3972 callback->awcb_done(zio, buf, callback->awcb_private); 3973 3974 kmem_free(callback, sizeof (arc_write_callback_t)); 3975} 3976 3977zio_t * 3978arc_write(zio_t *pio, spa_t *spa, uint64_t txg, 3979 blkptr_t *bp, arc_buf_t *buf, boolean_t l2arc, boolean_t l2arc_compress, 3980 const zio_prop_t *zp, arc_done_func_t *ready, arc_done_func_t *physdone, 3981 arc_done_func_t *done, void *private, zio_priority_t priority, 3982 int zio_flags, const zbookmark_phys_t *zb) 3983{ 3984 arc_buf_hdr_t *hdr = buf->b_hdr; 3985 arc_write_callback_t *callback; 3986 zio_t *zio; 3987 3988 ASSERT(ready != NULL); 3989 ASSERT(done != NULL); 3990 ASSERT(!HDR_IO_ERROR(hdr)); 3991 ASSERT((hdr->b_flags & ARC_FLAG_IO_IN_PROGRESS) == 0); 3992 ASSERT(hdr->b_acb == NULL); 3993 if (l2arc) 3994 hdr->b_flags |= ARC_FLAG_L2CACHE; 3995 if (l2arc_compress) 3996 hdr->b_flags |= ARC_FLAG_L2COMPRESS; 3997 callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP); 3998 callback->awcb_ready = ready; 3999 callback->awcb_physdone = physdone; 4000 callback->awcb_done = done; 4001 callback->awcb_private = private; 4002 callback->awcb_buf = buf; 4003 4004 zio = zio_write(pio, spa, txg, bp, buf->b_data, hdr->b_size, zp, 4005 arc_write_ready, arc_write_physdone, arc_write_done, callback, 4006 priority, zio_flags, zb); 4007 4008 return (zio); 4009} 4010 4011static int 4012arc_memory_throttle(uint64_t reserve, uint64_t txg) 4013{ 4014#ifdef _KERNEL 4015 uint64_t available_memory = ptob(freemem); 4016 static uint64_t page_load = 0; 4017 static uint64_t last_txg = 0; 4018 4019#if defined(__i386) || !defined(UMA_MD_SMALL_ALLOC) 4020 available_memory = 4021 MIN(available_memory, ptob(vmem_size(heap_arena, VMEM_FREE))); 4022#endif 4023 4024 if (freemem > (uint64_t)physmem * arc_lotsfree_percent / 100) 4025 return (0); 4026 4027 if (txg > last_txg) { 4028 last_txg = txg; 4029 page_load = 0; 4030 } 4031 /* 4032 * If we are in pageout, we know that memory is already tight, 4033 * the arc is already going to be evicting, so we just want to 4034 * continue to let page writes occur as quickly as possible. 4035 */ 4036 if (curproc == pageproc) { 4037 if (page_load > MAX(ptob(minfree), available_memory) / 4) 4038 return (SET_ERROR(ERESTART)); 4039 /* Note: reserve is inflated, so we deflate */ 4040 page_load += reserve / 8; 4041 return (0); 4042 } else if (page_load > 0 && arc_reclaim_needed()) { 4043 /* memory is low, delay before restarting */ 4044 ARCSTAT_INCR(arcstat_memory_throttle_count, 1); 4045 return (SET_ERROR(EAGAIN)); 4046 } 4047 page_load = 0; 4048#endif 4049 return (0); 4050} 4051 4052void 4053arc_tempreserve_clear(uint64_t reserve) 4054{ 4055 atomic_add_64(&arc_tempreserve, -reserve); 4056 ASSERT((int64_t)arc_tempreserve >= 0); 4057} 4058 4059int 4060arc_tempreserve_space(uint64_t reserve, uint64_t txg) 4061{ 4062 int error; 4063 uint64_t anon_size; 4064 4065 if (reserve > arc_c/4 && !arc_no_grow) { 4066 arc_c = MIN(arc_c_max, reserve * 4); 4067 DTRACE_PROBE1(arc__set_reserve, uint64_t, arc_c); 4068 } 4069 if (reserve > arc_c) 4070 return (SET_ERROR(ENOMEM)); 4071 4072 /* 4073 * Don't count loaned bufs as in flight dirty data to prevent long 4074 * network delays from blocking transactions that are ready to be 4075 * assigned to a txg. 4076 */ 4077 anon_size = MAX((int64_t)(arc_anon->arcs_size - arc_loaned_bytes), 0); 4078 4079 /* 4080 * Writes will, almost always, require additional memory allocations 4081 * in order to compress/encrypt/etc the data. We therefore need to 4082 * make sure that there is sufficient available memory for this. 4083 */ 4084 error = arc_memory_throttle(reserve, txg); 4085 if (error != 0) 4086 return (error); 4087 4088 /* 4089 * Throttle writes when the amount of dirty data in the cache 4090 * gets too large. We try to keep the cache less than half full 4091 * of dirty blocks so that our sync times don't grow too large. 4092 * Note: if two requests come in concurrently, we might let them 4093 * both succeed, when one of them should fail. Not a huge deal. 4094 */ 4095 4096 if (reserve + arc_tempreserve + anon_size > arc_c / 2 && 4097 anon_size > arc_c / 4) { 4098 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK " 4099 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n", 4100 arc_tempreserve>>10, 4101 arc_anon->arcs_lsize[ARC_BUFC_METADATA]>>10, 4102 arc_anon->arcs_lsize[ARC_BUFC_DATA]>>10, 4103 reserve>>10, arc_c>>10); 4104 return (SET_ERROR(ERESTART)); 4105 } 4106 atomic_add_64(&arc_tempreserve, reserve); 4107 return (0); 4108} 4109 4110static kmutex_t arc_lowmem_lock; 4111#ifdef _KERNEL 4112static eventhandler_tag arc_event_lowmem = NULL; 4113 4114static void 4115arc_lowmem(void *arg __unused, int howto __unused) 4116{ 4117 4118 /* Serialize access via arc_lowmem_lock. */ 4119 mutex_enter(&arc_lowmem_lock); 4120 mutex_enter(&arc_reclaim_thr_lock); 4121 needfree = 1; 4122 DTRACE_PROBE(arc__needfree); 4123 cv_signal(&arc_reclaim_thr_cv); 4124 4125 /* 4126 * It is unsafe to block here in arbitrary threads, because we can come 4127 * here from ARC itself and may hold ARC locks and thus risk a deadlock 4128 * with ARC reclaim thread. 4129 */ 4130 if (curproc == pageproc) { 4131 while (needfree) 4132 msleep(&needfree, &arc_reclaim_thr_lock, 0, "zfs:lowmem", 0); 4133 } 4134 mutex_exit(&arc_reclaim_thr_lock); 4135 mutex_exit(&arc_lowmem_lock); 4136} 4137#endif 4138 4139void 4140arc_init(void) 4141{ 4142 int i, prefetch_tunable_set = 0; 4143 4144 mutex_init(&arc_reclaim_thr_lock, NULL, MUTEX_DEFAULT, NULL); 4145 cv_init(&arc_reclaim_thr_cv, NULL, CV_DEFAULT, NULL); 4146 mutex_init(&arc_lowmem_lock, NULL, MUTEX_DEFAULT, NULL); 4147 4148 /* Convert seconds to clock ticks */ 4149 arc_min_prefetch_lifespan = 1 * hz; 4150 4151 /* Start out with 1/8 of all memory */ 4152 arc_c = kmem_size() / 8; 4153 4154#ifdef sun 4155#ifdef _KERNEL 4156 /* 4157 * On architectures where the physical memory can be larger 4158 * than the addressable space (intel in 32-bit mode), we may 4159 * need to limit the cache to 1/8 of VM size. 4160 */ 4161 arc_c = MIN(arc_c, vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 8); 4162#endif 4163#endif /* sun */ 4164 /* set min cache to 1/32 of all memory, or 16MB, whichever is more */ 4165 arc_c_min = MAX(arc_c / 4, 16 << 20); 4166 /* set max to 1/2 of all memory, or all but 1GB, whichever is more */ 4167 if (arc_c * 8 >= 1 << 30) 4168 arc_c_max = (arc_c * 8) - (1 << 30); 4169 else 4170 arc_c_max = arc_c_min; 4171 arc_c_max = MAX(arc_c * 5, arc_c_max); 4172 4173#ifdef _KERNEL 4174 /* 4175 * Allow the tunables to override our calculations if they are 4176 * reasonable (ie. over 16MB) 4177 */ 4178 if (zfs_arc_max > 16 << 20 && zfs_arc_max < kmem_size()) 4179 arc_c_max = zfs_arc_max; 4180 if (zfs_arc_min > 16 << 20 && zfs_arc_min <= arc_c_max) 4181 arc_c_min = zfs_arc_min; 4182#endif 4183 4184 arc_c = arc_c_max; 4185 arc_p = (arc_c >> 1); 4186 4187 /* limit meta-data to 1/4 of the arc capacity */ 4188 arc_meta_limit = arc_c_max / 4; 4189 4190 /* Allow the tunable to override if it is reasonable */ 4191 if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max) 4192 arc_meta_limit = zfs_arc_meta_limit; 4193 4194 if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0) 4195 arc_c_min = arc_meta_limit / 2; 4196 4197 if (zfs_arc_grow_retry > 0) 4198 arc_grow_retry = zfs_arc_grow_retry; 4199 4200 if (zfs_arc_shrink_shift > 0) 4201 arc_shrink_shift = zfs_arc_shrink_shift; 4202 4203 if (zfs_arc_p_min_shift > 0) 4204 arc_p_min_shift = zfs_arc_p_min_shift; 4205 4206 /* if kmem_flags are set, lets try to use less memory */ 4207 if (kmem_debugging()) 4208 arc_c = arc_c / 2; 4209 if (arc_c < arc_c_min) 4210 arc_c = arc_c_min; 4211 4212 zfs_arc_min = arc_c_min; 4213 zfs_arc_max = arc_c_max; 4214 4215 arc_anon = &ARC_anon; 4216 arc_mru = &ARC_mru; 4217 arc_mru_ghost = &ARC_mru_ghost; 4218 arc_mfu = &ARC_mfu; 4219 arc_mfu_ghost = &ARC_mfu_ghost; 4220 arc_l2c_only = &ARC_l2c_only; 4221 arc_size = 0; 4222 4223 for (i = 0; i < ARC_BUFC_NUMLISTS; i++) { 4224 mutex_init(&arc_anon->arcs_locks[i].arcs_lock, 4225 NULL, MUTEX_DEFAULT, NULL); 4226 mutex_init(&arc_mru->arcs_locks[i].arcs_lock, 4227 NULL, MUTEX_DEFAULT, NULL); 4228 mutex_init(&arc_mru_ghost->arcs_locks[i].arcs_lock, 4229 NULL, MUTEX_DEFAULT, NULL); 4230 mutex_init(&arc_mfu->arcs_locks[i].arcs_lock, 4231 NULL, MUTEX_DEFAULT, NULL); 4232 mutex_init(&arc_mfu_ghost->arcs_locks[i].arcs_lock, 4233 NULL, MUTEX_DEFAULT, NULL); 4234 mutex_init(&arc_l2c_only->arcs_locks[i].arcs_lock, 4235 NULL, MUTEX_DEFAULT, NULL); 4236 4237 list_create(&arc_mru->arcs_lists[i], 4238 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4239 list_create(&arc_mru_ghost->arcs_lists[i], 4240 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4241 list_create(&arc_mfu->arcs_lists[i], 4242 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4243 list_create(&arc_mfu_ghost->arcs_lists[i], 4244 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4245 list_create(&arc_mfu_ghost->arcs_lists[i], 4246 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4247 list_create(&arc_l2c_only->arcs_lists[i], 4248 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4249 } 4250 4251 buf_init(); 4252 4253 arc_thread_exit = 0; 4254 arc_eviction_list = NULL; 4255 mutex_init(&arc_eviction_mtx, NULL, MUTEX_DEFAULT, NULL); 4256 bzero(&arc_eviction_hdr, sizeof (arc_buf_hdr_t)); 4257 4258 arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED, 4259 sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL); 4260 4261 if (arc_ksp != NULL) { 4262 arc_ksp->ks_data = &arc_stats; 4263 kstat_install(arc_ksp); 4264 } 4265 4266 (void) thread_create(NULL, 0, arc_reclaim_thread, NULL, 0, &p0, 4267 TS_RUN, minclsyspri); 4268 4269#ifdef _KERNEL 4270 arc_event_lowmem = EVENTHANDLER_REGISTER(vm_lowmem, arc_lowmem, NULL, 4271 EVENTHANDLER_PRI_FIRST); 4272#endif 4273 4274 arc_dead = FALSE; 4275 arc_warm = B_FALSE; 4276 4277 /* 4278 * Calculate maximum amount of dirty data per pool. 4279 * 4280 * If it has been set by /etc/system, take that. 4281 * Otherwise, use a percentage of physical memory defined by 4282 * zfs_dirty_data_max_percent (default 10%) with a cap at 4283 * zfs_dirty_data_max_max (default 4GB). 4284 */ 4285 if (zfs_dirty_data_max == 0) { 4286 zfs_dirty_data_max = ptob(physmem) * 4287 zfs_dirty_data_max_percent / 100; 4288 zfs_dirty_data_max = MIN(zfs_dirty_data_max, 4289 zfs_dirty_data_max_max); 4290 } 4291 4292#ifdef _KERNEL 4293 if (TUNABLE_INT_FETCH("vfs.zfs.prefetch_disable", &zfs_prefetch_disable)) 4294 prefetch_tunable_set = 1; 4295 4296#ifdef __i386__ 4297 if (prefetch_tunable_set == 0) { 4298 printf("ZFS NOTICE: Prefetch is disabled by default on i386 " 4299 "-- to enable,\n"); 4300 printf(" add \"vfs.zfs.prefetch_disable=0\" " 4301 "to /boot/loader.conf.\n"); 4302 zfs_prefetch_disable = 1; 4303 } 4304#else 4305 if ((((uint64_t)physmem * PAGESIZE) < (1ULL << 32)) && 4306 prefetch_tunable_set == 0) { 4307 printf("ZFS NOTICE: Prefetch is disabled by default if less " 4308 "than 4GB of RAM is present;\n" 4309 " to enable, add \"vfs.zfs.prefetch_disable=0\" " 4310 "to /boot/loader.conf.\n"); 4311 zfs_prefetch_disable = 1; 4312 } 4313#endif 4314 /* Warn about ZFS memory and address space requirements. */ 4315 if (((uint64_t)physmem * PAGESIZE) < (256 + 128 + 64) * (1 << 20)) { 4316 printf("ZFS WARNING: Recommended minimum RAM size is 512MB; " 4317 "expect unstable behavior.\n"); 4318 } 4319 if (kmem_size() < 512 * (1 << 20)) { 4320 printf("ZFS WARNING: Recommended minimum kmem_size is 512MB; " 4321 "expect unstable behavior.\n"); 4322 printf(" Consider tuning vm.kmem_size and " 4323 "vm.kmem_size_max\n"); 4324 printf(" in /boot/loader.conf.\n"); 4325 } 4326#endif 4327} 4328 4329void 4330arc_fini(void) 4331{ 4332 int i; 4333 4334 mutex_enter(&arc_reclaim_thr_lock); 4335 arc_thread_exit = 1; 4336 cv_signal(&arc_reclaim_thr_cv); 4337 while (arc_thread_exit != 0) 4338 cv_wait(&arc_reclaim_thr_cv, &arc_reclaim_thr_lock); 4339 mutex_exit(&arc_reclaim_thr_lock); 4340 4341 arc_flush(NULL); 4342 4343 arc_dead = TRUE; 4344 4345 if (arc_ksp != NULL) { 4346 kstat_delete(arc_ksp); 4347 arc_ksp = NULL; 4348 } 4349 4350 mutex_destroy(&arc_eviction_mtx); 4351 mutex_destroy(&arc_reclaim_thr_lock); 4352 cv_destroy(&arc_reclaim_thr_cv); 4353 4354 for (i = 0; i < ARC_BUFC_NUMLISTS; i++) { 4355 list_destroy(&arc_mru->arcs_lists[i]); 4356 list_destroy(&arc_mru_ghost->arcs_lists[i]); 4357 list_destroy(&arc_mfu->arcs_lists[i]); 4358 list_destroy(&arc_mfu_ghost->arcs_lists[i]); 4359 list_destroy(&arc_l2c_only->arcs_lists[i]); 4360 4361 mutex_destroy(&arc_anon->arcs_locks[i].arcs_lock); 4362 mutex_destroy(&arc_mru->arcs_locks[i].arcs_lock); 4363 mutex_destroy(&arc_mru_ghost->arcs_locks[i].arcs_lock); 4364 mutex_destroy(&arc_mfu->arcs_locks[i].arcs_lock); 4365 mutex_destroy(&arc_mfu_ghost->arcs_locks[i].arcs_lock); 4366 mutex_destroy(&arc_l2c_only->arcs_locks[i].arcs_lock); 4367 } 4368 4369 buf_fini(); 4370 4371 ASSERT(arc_loaned_bytes == 0); 4372 4373 mutex_destroy(&arc_lowmem_lock); 4374#ifdef _KERNEL 4375 if (arc_event_lowmem != NULL) 4376 EVENTHANDLER_DEREGISTER(vm_lowmem, arc_event_lowmem); 4377#endif 4378} 4379 4380/* 4381 * Level 2 ARC 4382 * 4383 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk. 4384 * It uses dedicated storage devices to hold cached data, which are populated 4385 * using large infrequent writes. The main role of this cache is to boost 4386 * the performance of random read workloads. The intended L2ARC devices 4387 * include short-stroked disks, solid state disks, and other media with 4388 * substantially faster read latency than disk. 4389 * 4390 * +-----------------------+ 4391 * | ARC | 4392 * +-----------------------+ 4393 * | ^ ^ 4394 * | | | 4395 * l2arc_feed_thread() arc_read() 4396 * | | | 4397 * | l2arc read | 4398 * V | | 4399 * +---------------+ | 4400 * | L2ARC | | 4401 * +---------------+ | 4402 * | ^ | 4403 * l2arc_write() | | 4404 * | | | 4405 * V | | 4406 * +-------+ +-------+ 4407 * | vdev | | vdev | 4408 * | cache | | cache | 4409 * +-------+ +-------+ 4410 * +=========+ .-----. 4411 * : L2ARC : |-_____-| 4412 * : devices : | Disks | 4413 * +=========+ `-_____-' 4414 * 4415 * Read requests are satisfied from the following sources, in order: 4416 * 4417 * 1) ARC 4418 * 2) vdev cache of L2ARC devices 4419 * 3) L2ARC devices 4420 * 4) vdev cache of disks 4421 * 5) disks 4422 * 4423 * Some L2ARC device types exhibit extremely slow write performance. 4424 * To accommodate for this there are some significant differences between 4425 * the L2ARC and traditional cache design: 4426 * 4427 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from 4428 * the ARC behave as usual, freeing buffers and placing headers on ghost 4429 * lists. The ARC does not send buffers to the L2ARC during eviction as 4430 * this would add inflated write latencies for all ARC memory pressure. 4431 * 4432 * 2. The L2ARC attempts to cache data from the ARC before it is evicted. 4433 * It does this by periodically scanning buffers from the eviction-end of 4434 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are 4435 * not already there. It scans until a headroom of buffers is satisfied, 4436 * which itself is a buffer for ARC eviction. If a compressible buffer is 4437 * found during scanning and selected for writing to an L2ARC device, we 4438 * temporarily boost scanning headroom during the next scan cycle to make 4439 * sure we adapt to compression effects (which might significantly reduce 4440 * the data volume we write to L2ARC). The thread that does this is 4441 * l2arc_feed_thread(), illustrated below; example sizes are included to 4442 * provide a better sense of ratio than this diagram: 4443 * 4444 * head --> tail 4445 * +---------------------+----------+ 4446 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC 4447 * +---------------------+----------+ | o L2ARC eligible 4448 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer 4449 * +---------------------+----------+ | 4450 * 15.9 Gbytes ^ 32 Mbytes | 4451 * headroom | 4452 * l2arc_feed_thread() 4453 * | 4454 * l2arc write hand <--[oooo]--' 4455 * | 8 Mbyte 4456 * | write max 4457 * V 4458 * +==============================+ 4459 * L2ARC dev |####|#|###|###| |####| ... | 4460 * +==============================+ 4461 * 32 Gbytes 4462 * 4463 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of 4464 * evicted, then the L2ARC has cached a buffer much sooner than it probably 4465 * needed to, potentially wasting L2ARC device bandwidth and storage. It is 4466 * safe to say that this is an uncommon case, since buffers at the end of 4467 * the ARC lists have moved there due to inactivity. 4468 * 4469 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom, 4470 * then the L2ARC simply misses copying some buffers. This serves as a 4471 * pressure valve to prevent heavy read workloads from both stalling the ARC 4472 * with waits and clogging the L2ARC with writes. This also helps prevent 4473 * the potential for the L2ARC to churn if it attempts to cache content too 4474 * quickly, such as during backups of the entire pool. 4475 * 4476 * 5. After system boot and before the ARC has filled main memory, there are 4477 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru 4478 * lists can remain mostly static. Instead of searching from tail of these 4479 * lists as pictured, the l2arc_feed_thread() will search from the list heads 4480 * for eligible buffers, greatly increasing its chance of finding them. 4481 * 4482 * The L2ARC device write speed is also boosted during this time so that 4483 * the L2ARC warms up faster. Since there have been no ARC evictions yet, 4484 * there are no L2ARC reads, and no fear of degrading read performance 4485 * through increased writes. 4486 * 4487 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that 4488 * the vdev queue can aggregate them into larger and fewer writes. Each 4489 * device is written to in a rotor fashion, sweeping writes through 4490 * available space then repeating. 4491 * 4492 * 7. The L2ARC does not store dirty content. It never needs to flush 4493 * write buffers back to disk based storage. 4494 * 4495 * 8. If an ARC buffer is written (and dirtied) which also exists in the 4496 * L2ARC, the now stale L2ARC buffer is immediately dropped. 4497 * 4498 * The performance of the L2ARC can be tweaked by a number of tunables, which 4499 * may be necessary for different workloads: 4500 * 4501 * l2arc_write_max max write bytes per interval 4502 * l2arc_write_boost extra write bytes during device warmup 4503 * l2arc_noprefetch skip caching prefetched buffers 4504 * l2arc_headroom number of max device writes to precache 4505 * l2arc_headroom_boost when we find compressed buffers during ARC 4506 * scanning, we multiply headroom by this 4507 * percentage factor for the next scan cycle, 4508 * since more compressed buffers are likely to 4509 * be present 4510 * l2arc_feed_secs seconds between L2ARC writing 4511 * 4512 * Tunables may be removed or added as future performance improvements are 4513 * integrated, and also may become zpool properties. 4514 * 4515 * There are three key functions that control how the L2ARC warms up: 4516 * 4517 * l2arc_write_eligible() check if a buffer is eligible to cache 4518 * l2arc_write_size() calculate how much to write 4519 * l2arc_write_interval() calculate sleep delay between writes 4520 * 4521 * These three functions determine what to write, how much, and how quickly 4522 * to send writes. 4523 */ 4524 4525static boolean_t 4526l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *hdr) 4527{ 4528 /* 4529 * A buffer is *not* eligible for the L2ARC if it: 4530 * 1. belongs to a different spa. 4531 * 2. is already cached on the L2ARC. 4532 * 3. has an I/O in progress (it may be an incomplete read). 4533 * 4. is flagged not eligible (zfs property). 4534 */ 4535 if (hdr->b_spa != spa_guid) { 4536 ARCSTAT_BUMP(arcstat_l2_write_spa_mismatch); 4537 return (B_FALSE); 4538 } 4539 if (hdr->b_l2hdr != NULL) { 4540 ARCSTAT_BUMP(arcstat_l2_write_in_l2); 4541 return (B_FALSE); 4542 } 4543 if (HDR_IO_IN_PROGRESS(hdr)) { 4544 ARCSTAT_BUMP(arcstat_l2_write_hdr_io_in_progress); 4545 return (B_FALSE); 4546 } 4547 if (!HDR_L2CACHE(hdr)) { 4548 ARCSTAT_BUMP(arcstat_l2_write_not_cacheable); 4549 return (B_FALSE); 4550 } 4551 4552 return (B_TRUE); 4553} 4554 4555static uint64_t 4556l2arc_write_size(void) 4557{ 4558 uint64_t size; 4559 4560 /* 4561 * Make sure our globals have meaningful values in case the user 4562 * altered them. 4563 */ 4564 size = l2arc_write_max; 4565 if (size == 0) { 4566 cmn_err(CE_NOTE, "Bad value for l2arc_write_max, value must " 4567 "be greater than zero, resetting it to the default (%d)", 4568 L2ARC_WRITE_SIZE); 4569 size = l2arc_write_max = L2ARC_WRITE_SIZE; 4570 } 4571 4572 if (arc_warm == B_FALSE) 4573 size += l2arc_write_boost; 4574 4575 return (size); 4576 4577} 4578 4579static clock_t 4580l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote) 4581{ 4582 clock_t interval, next, now; 4583 4584 /* 4585 * If the ARC lists are busy, increase our write rate; if the 4586 * lists are stale, idle back. This is achieved by checking 4587 * how much we previously wrote - if it was more than half of 4588 * what we wanted, schedule the next write much sooner. 4589 */ 4590 if (l2arc_feed_again && wrote > (wanted / 2)) 4591 interval = (hz * l2arc_feed_min_ms) / 1000; 4592 else 4593 interval = hz * l2arc_feed_secs; 4594 4595 now = ddi_get_lbolt(); 4596 next = MAX(now, MIN(now + interval, began + interval)); 4597 4598 return (next); 4599} 4600 4601static void 4602l2arc_hdr_stat_add(void) 4603{ 4604 ARCSTAT_INCR(arcstat_l2_hdr_size, HDR_SIZE + L2HDR_SIZE); 4605 ARCSTAT_INCR(arcstat_hdr_size, -HDR_SIZE); 4606} 4607 4608static void 4609l2arc_hdr_stat_remove(void) 4610{ 4611 ARCSTAT_INCR(arcstat_l2_hdr_size, -(HDR_SIZE + L2HDR_SIZE)); 4612 ARCSTAT_INCR(arcstat_hdr_size, HDR_SIZE); 4613} 4614 4615/* 4616 * Cycle through L2ARC devices. This is how L2ARC load balances. 4617 * If a device is returned, this also returns holding the spa config lock. 4618 */ 4619static l2arc_dev_t * 4620l2arc_dev_get_next(void) 4621{ 4622 l2arc_dev_t *first, *next = NULL; 4623 4624 /* 4625 * Lock out the removal of spas (spa_namespace_lock), then removal 4626 * of cache devices (l2arc_dev_mtx). Once a device has been selected, 4627 * both locks will be dropped and a spa config lock held instead. 4628 */ 4629 mutex_enter(&spa_namespace_lock); 4630 mutex_enter(&l2arc_dev_mtx); 4631 4632 /* if there are no vdevs, there is nothing to do */ 4633 if (l2arc_ndev == 0) 4634 goto out; 4635 4636 first = NULL; 4637 next = l2arc_dev_last; 4638 do { 4639 /* loop around the list looking for a non-faulted vdev */ 4640 if (next == NULL) { 4641 next = list_head(l2arc_dev_list); 4642 } else { 4643 next = list_next(l2arc_dev_list, next); 4644 if (next == NULL) 4645 next = list_head(l2arc_dev_list); 4646 } 4647 4648 /* if we have come back to the start, bail out */ 4649 if (first == NULL) 4650 first = next; 4651 else if (next == first) 4652 break; 4653 4654 } while (vdev_is_dead(next->l2ad_vdev)); 4655 4656 /* if we were unable to find any usable vdevs, return NULL */ 4657 if (vdev_is_dead(next->l2ad_vdev)) 4658 next = NULL; 4659 4660 l2arc_dev_last = next; 4661 4662out: 4663 mutex_exit(&l2arc_dev_mtx); 4664 4665 /* 4666 * Grab the config lock to prevent the 'next' device from being 4667 * removed while we are writing to it. 4668 */ 4669 if (next != NULL) 4670 spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER); 4671 mutex_exit(&spa_namespace_lock); 4672 4673 return (next); 4674} 4675 4676/* 4677 * Free buffers that were tagged for destruction. 4678 */ 4679static void 4680l2arc_do_free_on_write() 4681{ 4682 list_t *buflist; 4683 l2arc_data_free_t *df, *df_prev; 4684 4685 mutex_enter(&l2arc_free_on_write_mtx); 4686 buflist = l2arc_free_on_write; 4687 4688 for (df = list_tail(buflist); df; df = df_prev) { 4689 df_prev = list_prev(buflist, df); 4690 ASSERT(df->l2df_data != NULL); 4691 ASSERT(df->l2df_func != NULL); 4692 df->l2df_func(df->l2df_data, df->l2df_size); 4693 list_remove(buflist, df); 4694 kmem_free(df, sizeof (l2arc_data_free_t)); 4695 } 4696 4697 mutex_exit(&l2arc_free_on_write_mtx); 4698} 4699 4700/* 4701 * A write to a cache device has completed. Update all headers to allow 4702 * reads from these buffers to begin. 4703 */ 4704static void 4705l2arc_write_done(zio_t *zio) 4706{ 4707 l2arc_write_callback_t *cb; 4708 l2arc_dev_t *dev; 4709 list_t *buflist; 4710 arc_buf_hdr_t *head, *hdr, *hdr_prev; 4711 l2arc_buf_hdr_t *abl2; 4712 kmutex_t *hash_lock; 4713 int64_t bytes_dropped = 0; 4714 4715 cb = zio->io_private; 4716 ASSERT(cb != NULL); 4717 dev = cb->l2wcb_dev; 4718 ASSERT(dev != NULL); 4719 head = cb->l2wcb_head; 4720 ASSERT(head != NULL); 4721 buflist = dev->l2ad_buflist; 4722 ASSERT(buflist != NULL); 4723 DTRACE_PROBE2(l2arc__iodone, zio_t *, zio, 4724 l2arc_write_callback_t *, cb); 4725 4726 if (zio->io_error != 0) 4727 ARCSTAT_BUMP(arcstat_l2_writes_error); 4728 4729 mutex_enter(&l2arc_buflist_mtx); 4730 4731 /* 4732 * All writes completed, or an error was hit. 4733 */ 4734 for (hdr = list_prev(buflist, head); hdr; hdr = hdr_prev) { 4735 hdr_prev = list_prev(buflist, hdr); 4736 abl2 = hdr->b_l2hdr; 4737 4738 /* 4739 * Release the temporary compressed buffer as soon as possible. 4740 */ 4741 if (abl2->b_compress != ZIO_COMPRESS_OFF) 4742 l2arc_release_cdata_buf(hdr); 4743 4744 hash_lock = HDR_LOCK(hdr); 4745 if (!mutex_tryenter(hash_lock)) { 4746 /* 4747 * This buffer misses out. It may be in a stage 4748 * of eviction. Its ARC_L2_WRITING flag will be 4749 * left set, denying reads to this buffer. 4750 */ 4751 ARCSTAT_BUMP(arcstat_l2_writes_hdr_miss); 4752 continue; 4753 } 4754 4755 if (zio->io_error != 0) { 4756 /* 4757 * Error - drop L2ARC entry. 4758 */ 4759 list_remove(buflist, hdr); 4760 ARCSTAT_INCR(arcstat_l2_asize, -abl2->b_asize); 4761 bytes_dropped += abl2->b_asize; 4762 hdr->b_l2hdr = NULL; 4763 trim_map_free(abl2->b_dev->l2ad_vdev, abl2->b_daddr, 4764 abl2->b_asize, 0); 4765 kmem_free(abl2, sizeof (l2arc_buf_hdr_t)); 4766 ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size); 4767 } 4768 4769 /* 4770 * Allow ARC to begin reads to this L2ARC entry. 4771 */ 4772 hdr->b_flags &= ~ARC_FLAG_L2_WRITING; 4773 4774 mutex_exit(hash_lock); 4775 } 4776 4777 atomic_inc_64(&l2arc_writes_done); 4778 list_remove(buflist, head); 4779 kmem_cache_free(hdr_cache, head); 4780 mutex_exit(&l2arc_buflist_mtx); 4781 4782 vdev_space_update(dev->l2ad_vdev, -bytes_dropped, 0, 0); 4783 4784 l2arc_do_free_on_write(); 4785 4786 kmem_free(cb, sizeof (l2arc_write_callback_t)); 4787} 4788 4789/* 4790 * A read to a cache device completed. Validate buffer contents before 4791 * handing over to the regular ARC routines. 4792 */ 4793static void 4794l2arc_read_done(zio_t *zio) 4795{ 4796 l2arc_read_callback_t *cb; 4797 arc_buf_hdr_t *hdr; 4798 arc_buf_t *buf; 4799 kmutex_t *hash_lock; 4800 int equal; 4801 4802 ASSERT(zio->io_vd != NULL); 4803 ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE); 4804 4805 spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd); 4806 4807 cb = zio->io_private; 4808 ASSERT(cb != NULL); 4809 buf = cb->l2rcb_buf; 4810 ASSERT(buf != NULL); 4811 4812 hash_lock = HDR_LOCK(buf->b_hdr); 4813 mutex_enter(hash_lock); 4814 hdr = buf->b_hdr; 4815 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 4816 4817 /* 4818 * If the buffer was compressed, decompress it first. 4819 */ 4820 if (cb->l2rcb_compress != ZIO_COMPRESS_OFF) 4821 l2arc_decompress_zio(zio, hdr, cb->l2rcb_compress); 4822 ASSERT(zio->io_data != NULL); 4823 4824 /* 4825 * Check this survived the L2ARC journey. 4826 */ 4827 equal = arc_cksum_equal(buf); 4828 if (equal && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) { 4829 mutex_exit(hash_lock); 4830 zio->io_private = buf; 4831 zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */ 4832 zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */ 4833 arc_read_done(zio); 4834 } else { 4835 mutex_exit(hash_lock); 4836 /* 4837 * Buffer didn't survive caching. Increment stats and 4838 * reissue to the original storage device. 4839 */ 4840 if (zio->io_error != 0) { 4841 ARCSTAT_BUMP(arcstat_l2_io_error); 4842 } else { 4843 zio->io_error = SET_ERROR(EIO); 4844 } 4845 if (!equal) 4846 ARCSTAT_BUMP(arcstat_l2_cksum_bad); 4847 4848 /* 4849 * If there's no waiter, issue an async i/o to the primary 4850 * storage now. If there *is* a waiter, the caller must 4851 * issue the i/o in a context where it's OK to block. 4852 */ 4853 if (zio->io_waiter == NULL) { 4854 zio_t *pio = zio_unique_parent(zio); 4855 4856 ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL); 4857 4858 zio_nowait(zio_read(pio, cb->l2rcb_spa, &cb->l2rcb_bp, 4859 buf->b_data, zio->io_size, arc_read_done, buf, 4860 zio->io_priority, cb->l2rcb_flags, &cb->l2rcb_zb)); 4861 } 4862 } 4863 4864 kmem_free(cb, sizeof (l2arc_read_callback_t)); 4865} 4866 4867/* 4868 * This is the list priority from which the L2ARC will search for pages to 4869 * cache. This is used within loops (0..3) to cycle through lists in the 4870 * desired order. This order can have a significant effect on cache 4871 * performance. 4872 * 4873 * Currently the metadata lists are hit first, MFU then MRU, followed by 4874 * the data lists. This function returns a locked list, and also returns 4875 * the lock pointer. 4876 */ 4877static list_t * 4878l2arc_list_locked(int list_num, kmutex_t **lock) 4879{ 4880 list_t *list = NULL; 4881 int idx; 4882 4883 ASSERT(list_num >= 0 && list_num < 2 * ARC_BUFC_NUMLISTS); 4884 4885 if (list_num < ARC_BUFC_NUMMETADATALISTS) { 4886 idx = list_num; 4887 list = &arc_mfu->arcs_lists[idx]; 4888 *lock = ARCS_LOCK(arc_mfu, idx); 4889 } else if (list_num < ARC_BUFC_NUMMETADATALISTS * 2) { 4890 idx = list_num - ARC_BUFC_NUMMETADATALISTS; 4891 list = &arc_mru->arcs_lists[idx]; 4892 *lock = ARCS_LOCK(arc_mru, idx); 4893 } else if (list_num < (ARC_BUFC_NUMMETADATALISTS * 2 + 4894 ARC_BUFC_NUMDATALISTS)) { 4895 idx = list_num - ARC_BUFC_NUMMETADATALISTS; 4896 list = &arc_mfu->arcs_lists[idx]; 4897 *lock = ARCS_LOCK(arc_mfu, idx); 4898 } else { 4899 idx = list_num - ARC_BUFC_NUMLISTS; 4900 list = &arc_mru->arcs_lists[idx]; 4901 *lock = ARCS_LOCK(arc_mru, idx); 4902 } 4903 4904 ASSERT(!(MUTEX_HELD(*lock))); 4905 mutex_enter(*lock); 4906 return (list); 4907} 4908 4909/* 4910 * Evict buffers from the device write hand to the distance specified in 4911 * bytes. This distance may span populated buffers, it may span nothing. 4912 * This is clearing a region on the L2ARC device ready for writing. 4913 * If the 'all' boolean is set, every buffer is evicted. 4914 */ 4915static void 4916l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all) 4917{ 4918 list_t *buflist; 4919 l2arc_buf_hdr_t *abl2; 4920 arc_buf_hdr_t *hdr, *hdr_prev; 4921 kmutex_t *hash_lock; 4922 uint64_t taddr; 4923 int64_t bytes_evicted = 0; 4924 4925 buflist = dev->l2ad_buflist; 4926 4927 if (buflist == NULL) 4928 return; 4929 4930 if (!all && dev->l2ad_first) { 4931 /* 4932 * This is the first sweep through the device. There is 4933 * nothing to evict. 4934 */ 4935 return; 4936 } 4937 4938 if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) { 4939 /* 4940 * When nearing the end of the device, evict to the end 4941 * before the device write hand jumps to the start. 4942 */ 4943 taddr = dev->l2ad_end; 4944 } else { 4945 taddr = dev->l2ad_hand + distance; 4946 } 4947 DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist, 4948 uint64_t, taddr, boolean_t, all); 4949 4950top: 4951 mutex_enter(&l2arc_buflist_mtx); 4952 for (hdr = list_tail(buflist); hdr; hdr = hdr_prev) { 4953 hdr_prev = list_prev(buflist, hdr); 4954 4955 hash_lock = HDR_LOCK(hdr); 4956 if (!mutex_tryenter(hash_lock)) { 4957 /* 4958 * Missed the hash lock. Retry. 4959 */ 4960 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry); 4961 mutex_exit(&l2arc_buflist_mtx); 4962 mutex_enter(hash_lock); 4963 mutex_exit(hash_lock); 4964 goto top; 4965 } 4966 4967 if (HDR_L2_WRITE_HEAD(hdr)) { 4968 /* 4969 * We hit a write head node. Leave it for 4970 * l2arc_write_done(). 4971 */ 4972 list_remove(buflist, hdr); 4973 mutex_exit(hash_lock); 4974 continue; 4975 } 4976 4977 if (!all && hdr->b_l2hdr != NULL && 4978 (hdr->b_l2hdr->b_daddr > taddr || 4979 hdr->b_l2hdr->b_daddr < dev->l2ad_hand)) { 4980 /* 4981 * We've evicted to the target address, 4982 * or the end of the device. 4983 */ 4984 mutex_exit(hash_lock); 4985 break; 4986 } 4987 4988 if (HDR_FREE_IN_PROGRESS(hdr)) { 4989 /* 4990 * Already on the path to destruction. 4991 */ 4992 mutex_exit(hash_lock); 4993 continue; 4994 } 4995 4996 if (hdr->b_state == arc_l2c_only) { 4997 ASSERT(!HDR_L2_READING(hdr)); 4998 /* 4999 * This doesn't exist in the ARC. Destroy. 5000 * arc_hdr_destroy() will call list_remove() 5001 * and decrement arcstat_l2_size. 5002 */ 5003 arc_change_state(arc_anon, hdr, hash_lock); 5004 arc_hdr_destroy(hdr); 5005 } else { 5006 /* 5007 * Invalidate issued or about to be issued 5008 * reads, since we may be about to write 5009 * over this location. 5010 */ 5011 if (HDR_L2_READING(hdr)) { 5012 ARCSTAT_BUMP(arcstat_l2_evict_reading); 5013 hdr->b_flags |= ARC_FLAG_L2_EVICTED; 5014 } 5015 5016 /* 5017 * Tell ARC this no longer exists in L2ARC. 5018 */ 5019 if (hdr->b_l2hdr != NULL) { 5020 abl2 = hdr->b_l2hdr; 5021 ARCSTAT_INCR(arcstat_l2_asize, -abl2->b_asize); 5022 bytes_evicted += abl2->b_asize; 5023 hdr->b_l2hdr = NULL; 5024 /* 5025 * We are destroying l2hdr, so ensure that 5026 * its compressed buffer, if any, is not leaked. 5027 */ 5028 ASSERT(abl2->b_tmp_cdata == NULL); 5029 kmem_free(abl2, sizeof (l2arc_buf_hdr_t)); 5030 ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size); 5031 } 5032 list_remove(buflist, hdr); 5033 5034 /* 5035 * This may have been leftover after a 5036 * failed write. 5037 */ 5038 hdr->b_flags &= ~ARC_FLAG_L2_WRITING; 5039 } 5040 mutex_exit(hash_lock); 5041 } 5042 mutex_exit(&l2arc_buflist_mtx); 5043 5044 vdev_space_update(dev->l2ad_vdev, -bytes_evicted, 0, 0); 5045 dev->l2ad_evict = taddr; 5046} 5047 5048/* 5049 * Find and write ARC buffers to the L2ARC device. 5050 * 5051 * An ARC_FLAG_L2_WRITING flag is set so that the L2ARC buffers are not valid 5052 * for reading until they have completed writing. 5053 * The headroom_boost is an in-out parameter used to maintain headroom boost 5054 * state between calls to this function. 5055 * 5056 * Returns the number of bytes actually written (which may be smaller than 5057 * the delta by which the device hand has changed due to alignment). 5058 */ 5059static uint64_t 5060l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz, 5061 boolean_t *headroom_boost) 5062{ 5063 arc_buf_hdr_t *hdr, *hdr_prev, *head; 5064 list_t *list; 5065 uint64_t write_asize, write_sz, headroom, buf_compress_minsz; 5066 void *buf_data; 5067 kmutex_t *list_lock; 5068 boolean_t full; 5069 l2arc_write_callback_t *cb; 5070 zio_t *pio, *wzio; 5071 uint64_t guid = spa_load_guid(spa); 5072 const boolean_t do_headroom_boost = *headroom_boost; 5073 int try; 5074 5075 ASSERT(dev->l2ad_vdev != NULL); 5076 5077 /* Lower the flag now, we might want to raise it again later. */ 5078 *headroom_boost = B_FALSE; 5079 5080 pio = NULL; 5081 write_sz = write_asize = 0; 5082 full = B_FALSE; 5083 head = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE); 5084 head->b_flags |= ARC_FLAG_L2_WRITE_HEAD; 5085 5086 ARCSTAT_BUMP(arcstat_l2_write_buffer_iter); 5087 /* 5088 * We will want to try to compress buffers that are at least 2x the 5089 * device sector size. 5090 */ 5091 buf_compress_minsz = 2 << dev->l2ad_vdev->vdev_ashift; 5092 5093 /* 5094 * Copy buffers for L2ARC writing. 5095 */ 5096 mutex_enter(&l2arc_buflist_mtx); 5097 for (try = 0; try < 2 * ARC_BUFC_NUMLISTS; try++) { 5098 uint64_t passed_sz = 0; 5099 5100 list = l2arc_list_locked(try, &list_lock); 5101 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_iter); 5102 5103 /* 5104 * L2ARC fast warmup. 5105 * 5106 * Until the ARC is warm and starts to evict, read from the 5107 * head of the ARC lists rather than the tail. 5108 */ 5109 if (arc_warm == B_FALSE) 5110 hdr = list_head(list); 5111 else 5112 hdr = list_tail(list); 5113 if (hdr == NULL) 5114 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_null_iter); 5115 5116 headroom = target_sz * l2arc_headroom * 2 / ARC_BUFC_NUMLISTS; 5117 if (do_headroom_boost) 5118 headroom = (headroom * l2arc_headroom_boost) / 100; 5119 5120 for (; hdr; hdr = hdr_prev) { 5121 l2arc_buf_hdr_t *l2hdr; 5122 kmutex_t *hash_lock; 5123 uint64_t buf_sz; 5124 uint64_t buf_a_sz; 5125 5126 if (arc_warm == B_FALSE) 5127 hdr_prev = list_next(list, hdr); 5128 else 5129 hdr_prev = list_prev(list, hdr); 5130 ARCSTAT_INCR(arcstat_l2_write_buffer_bytes_scanned, hdr->b_size); 5131 5132 hash_lock = HDR_LOCK(hdr); 5133 if (!mutex_tryenter(hash_lock)) { 5134 ARCSTAT_BUMP(arcstat_l2_write_trylock_fail); 5135 /* 5136 * Skip this buffer rather than waiting. 5137 */ 5138 continue; 5139 } 5140 5141 passed_sz += hdr->b_size; 5142 if (passed_sz > headroom) { 5143 /* 5144 * Searched too far. 5145 */ 5146 mutex_exit(hash_lock); 5147 ARCSTAT_BUMP(arcstat_l2_write_passed_headroom); 5148 break; 5149 } 5150 5151 if (!l2arc_write_eligible(guid, hdr)) { 5152 mutex_exit(hash_lock); 5153 continue; 5154 } 5155 5156 /* 5157 * Assume that the buffer is not going to be compressed 5158 * and could take more space on disk because of a larger 5159 * disk block size. 5160 */ 5161 buf_sz = hdr->b_size; 5162 buf_a_sz = vdev_psize_to_asize(dev->l2ad_vdev, buf_sz); 5163 5164 if ((write_asize + buf_a_sz) > target_sz) { 5165 full = B_TRUE; 5166 mutex_exit(hash_lock); 5167 ARCSTAT_BUMP(arcstat_l2_write_full); 5168 break; 5169 } 5170 5171 if (pio == NULL) { 5172 /* 5173 * Insert a dummy header on the buflist so 5174 * l2arc_write_done() can find where the 5175 * write buffers begin without searching. 5176 */ 5177 list_insert_head(dev->l2ad_buflist, head); 5178 5179 cb = kmem_alloc( 5180 sizeof (l2arc_write_callback_t), KM_SLEEP); 5181 cb->l2wcb_dev = dev; 5182 cb->l2wcb_head = head; 5183 pio = zio_root(spa, l2arc_write_done, cb, 5184 ZIO_FLAG_CANFAIL); 5185 ARCSTAT_BUMP(arcstat_l2_write_pios); 5186 } 5187 5188 /* 5189 * Create and add a new L2ARC header. 5190 */ 5191 l2hdr = kmem_zalloc(sizeof (l2arc_buf_hdr_t), KM_SLEEP); 5192 l2hdr->b_dev = dev; 5193 hdr->b_flags |= ARC_FLAG_L2_WRITING; 5194 5195 /* 5196 * Temporarily stash the data buffer in b_tmp_cdata. 5197 * The subsequent write step will pick it up from 5198 * there. This is because can't access hdr->b_buf 5199 * without holding the hash_lock, which we in turn 5200 * can't access without holding the ARC list locks 5201 * (which we want to avoid during compression/writing). 5202 */ 5203 l2hdr->b_compress = ZIO_COMPRESS_OFF; 5204 l2hdr->b_asize = hdr->b_size; 5205 l2hdr->b_tmp_cdata = hdr->b_buf->b_data; 5206 5207 hdr->b_l2hdr = l2hdr; 5208 5209 list_insert_head(dev->l2ad_buflist, hdr); 5210 5211 /* 5212 * Compute and store the buffer cksum before 5213 * writing. On debug the cksum is verified first. 5214 */ 5215 arc_cksum_verify(hdr->b_buf); 5216 arc_cksum_compute(hdr->b_buf, B_TRUE); 5217 5218 mutex_exit(hash_lock); 5219 5220 write_sz += buf_sz; 5221 write_asize += buf_a_sz; 5222 } 5223 5224 mutex_exit(list_lock); 5225 5226 if (full == B_TRUE) 5227 break; 5228 } 5229 5230 /* No buffers selected for writing? */ 5231 if (pio == NULL) { 5232 ASSERT0(write_sz); 5233 mutex_exit(&l2arc_buflist_mtx); 5234 kmem_cache_free(hdr_cache, head); 5235 return (0); 5236 } 5237 5238 /* 5239 * Note that elsewhere in this file arcstat_l2_asize 5240 * and the used space on l2ad_vdev are updated using b_asize, 5241 * which is not necessarily rounded up to the device block size. 5242 * Too keep accounting consistent we do the same here as well: 5243 * stats_size accumulates the sum of b_asize of the written buffers, 5244 * while write_asize accumulates the sum of b_asize rounded up 5245 * to the device block size. 5246 * The latter sum is used only to validate the corectness of the code. 5247 */ 5248 uint64_t stats_size = 0; 5249 write_asize = 0; 5250 5251 /* 5252 * Now start writing the buffers. We're starting at the write head 5253 * and work backwards, retracing the course of the buffer selector 5254 * loop above. 5255 */ 5256 for (hdr = list_prev(dev->l2ad_buflist, head); hdr; 5257 hdr = list_prev(dev->l2ad_buflist, hdr)) { 5258 l2arc_buf_hdr_t *l2hdr; 5259 uint64_t buf_sz; 5260 5261 /* 5262 * We shouldn't need to lock the buffer here, since we flagged 5263 * it as ARC_FLAG_L2_WRITING in the previous step, but we must 5264 * take care to only access its L2 cache parameters. In 5265 * particular, hdr->b_buf may be invalid by now due to 5266 * ARC eviction. 5267 */ 5268 l2hdr = hdr->b_l2hdr; 5269 l2hdr->b_daddr = dev->l2ad_hand; 5270 5271 if ((hdr->b_flags & ARC_FLAG_L2COMPRESS) && 5272 l2hdr->b_asize >= buf_compress_minsz) { 5273 if (l2arc_compress_buf(l2hdr)) { 5274 /* 5275 * If compression succeeded, enable headroom 5276 * boost on the next scan cycle. 5277 */ 5278 *headroom_boost = B_TRUE; 5279 } 5280 } 5281 5282 /* 5283 * Pick up the buffer data we had previously stashed away 5284 * (and now potentially also compressed). 5285 */ 5286 buf_data = l2hdr->b_tmp_cdata; 5287 buf_sz = l2hdr->b_asize; 5288 5289 /* 5290 * If the data has not been compressed, then clear b_tmp_cdata 5291 * to make sure that it points only to a temporary compression 5292 * buffer. 5293 */ 5294 if (!L2ARC_IS_VALID_COMPRESS(l2hdr->b_compress)) 5295 l2hdr->b_tmp_cdata = NULL; 5296 5297 /* Compression may have squashed the buffer to zero length. */ 5298 if (buf_sz != 0) { 5299 uint64_t buf_a_sz; 5300 5301 wzio = zio_write_phys(pio, dev->l2ad_vdev, 5302 dev->l2ad_hand, buf_sz, buf_data, ZIO_CHECKSUM_OFF, 5303 NULL, NULL, ZIO_PRIORITY_ASYNC_WRITE, 5304 ZIO_FLAG_CANFAIL, B_FALSE); 5305 5306 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev, 5307 zio_t *, wzio); 5308 (void) zio_nowait(wzio); 5309 5310 stats_size += buf_sz; 5311 /* 5312 * Keep the clock hand suitably device-aligned. 5313 */ 5314 buf_a_sz = vdev_psize_to_asize(dev->l2ad_vdev, buf_sz); 5315 write_asize += buf_a_sz; 5316 dev->l2ad_hand += buf_a_sz; 5317 } 5318 } 5319 5320 mutex_exit(&l2arc_buflist_mtx); 5321 5322 ASSERT3U(write_asize, <=, target_sz); 5323 ARCSTAT_BUMP(arcstat_l2_writes_sent); 5324 ARCSTAT_INCR(arcstat_l2_write_bytes, write_asize); 5325 ARCSTAT_INCR(arcstat_l2_size, write_sz); 5326 ARCSTAT_INCR(arcstat_l2_asize, stats_size); 5327 vdev_space_update(dev->l2ad_vdev, stats_size, 0, 0); 5328 5329 /* 5330 * Bump device hand to the device start if it is approaching the end. 5331 * l2arc_evict() will already have evicted ahead for this case. 5332 */ 5333 if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) { 5334 dev->l2ad_hand = dev->l2ad_start; 5335 dev->l2ad_evict = dev->l2ad_start; 5336 dev->l2ad_first = B_FALSE; 5337 } 5338 5339 dev->l2ad_writing = B_TRUE; 5340 (void) zio_wait(pio); 5341 dev->l2ad_writing = B_FALSE; 5342 5343 return (write_asize); 5344} 5345 5346/* 5347 * Compresses an L2ARC buffer. 5348 * The data to be compressed must be prefilled in l2hdr->b_tmp_cdata and its 5349 * size in l2hdr->b_asize. This routine tries to compress the data and 5350 * depending on the compression result there are three possible outcomes: 5351 * *) The buffer was incompressible. The original l2hdr contents were left 5352 * untouched and are ready for writing to an L2 device. 5353 * *) The buffer was all-zeros, so there is no need to write it to an L2 5354 * device. To indicate this situation b_tmp_cdata is NULL'ed, b_asize is 5355 * set to zero and b_compress is set to ZIO_COMPRESS_EMPTY. 5356 * *) Compression succeeded and b_tmp_cdata was replaced with a temporary 5357 * data buffer which holds the compressed data to be written, and b_asize 5358 * tells us how much data there is. b_compress is set to the appropriate 5359 * compression algorithm. Once writing is done, invoke 5360 * l2arc_release_cdata_buf on this l2hdr to free this temporary buffer. 5361 * 5362 * Returns B_TRUE if compression succeeded, or B_FALSE if it didn't (the 5363 * buffer was incompressible). 5364 */ 5365static boolean_t 5366l2arc_compress_buf(l2arc_buf_hdr_t *l2hdr) 5367{ 5368 void *cdata; 5369 size_t csize, len, rounded; 5370 5371 ASSERT(l2hdr->b_compress == ZIO_COMPRESS_OFF); 5372 ASSERT(l2hdr->b_tmp_cdata != NULL); 5373 5374 len = l2hdr->b_asize; 5375 cdata = zio_data_buf_alloc(len); 5376 csize = zio_compress_data(ZIO_COMPRESS_LZ4, l2hdr->b_tmp_cdata, 5377 cdata, l2hdr->b_asize); 5378 5379 if (csize == 0) { 5380 /* zero block, indicate that there's nothing to write */ 5381 zio_data_buf_free(cdata, len); 5382 l2hdr->b_compress = ZIO_COMPRESS_EMPTY; 5383 l2hdr->b_asize = 0; 5384 l2hdr->b_tmp_cdata = NULL; 5385 ARCSTAT_BUMP(arcstat_l2_compress_zeros); 5386 return (B_TRUE); 5387 } 5388 5389 rounded = P2ROUNDUP(csize, 5390 (size_t)1 << l2hdr->b_dev->l2ad_vdev->vdev_ashift); 5391 if (rounded < len) { 5392 /* 5393 * Compression succeeded, we'll keep the cdata around for 5394 * writing and release it afterwards. 5395 */ 5396 if (rounded > csize) { 5397 bzero((char *)cdata + csize, rounded - csize); 5398 csize = rounded; 5399 } 5400 l2hdr->b_compress = ZIO_COMPRESS_LZ4; 5401 l2hdr->b_asize = csize; 5402 l2hdr->b_tmp_cdata = cdata; 5403 ARCSTAT_BUMP(arcstat_l2_compress_successes); 5404 return (B_TRUE); 5405 } else { 5406 /* 5407 * Compression failed, release the compressed buffer. 5408 * l2hdr will be left unmodified. 5409 */ 5410 zio_data_buf_free(cdata, len); 5411 ARCSTAT_BUMP(arcstat_l2_compress_failures); 5412 return (B_FALSE); 5413 } 5414} 5415 5416/* 5417 * Decompresses a zio read back from an l2arc device. On success, the 5418 * underlying zio's io_data buffer is overwritten by the uncompressed 5419 * version. On decompression error (corrupt compressed stream), the 5420 * zio->io_error value is set to signal an I/O error. 5421 * 5422 * Please note that the compressed data stream is not checksummed, so 5423 * if the underlying device is experiencing data corruption, we may feed 5424 * corrupt data to the decompressor, so the decompressor needs to be 5425 * able to handle this situation (LZ4 does). 5426 */ 5427static void 5428l2arc_decompress_zio(zio_t *zio, arc_buf_hdr_t *hdr, enum zio_compress c) 5429{ 5430 ASSERT(L2ARC_IS_VALID_COMPRESS(c)); 5431 5432 if (zio->io_error != 0) { 5433 /* 5434 * An io error has occured, just restore the original io 5435 * size in preparation for a main pool read. 5436 */ 5437 zio->io_orig_size = zio->io_size = hdr->b_size; 5438 return; 5439 } 5440 5441 if (c == ZIO_COMPRESS_EMPTY) { 5442 /* 5443 * An empty buffer results in a null zio, which means we 5444 * need to fill its io_data after we're done restoring the 5445 * buffer's contents. 5446 */ 5447 ASSERT(hdr->b_buf != NULL); 5448 bzero(hdr->b_buf->b_data, hdr->b_size); 5449 zio->io_data = zio->io_orig_data = hdr->b_buf->b_data; 5450 } else { 5451 ASSERT(zio->io_data != NULL); 5452 /* 5453 * We copy the compressed data from the start of the arc buffer 5454 * (the zio_read will have pulled in only what we need, the 5455 * rest is garbage which we will overwrite at decompression) 5456 * and then decompress back to the ARC data buffer. This way we 5457 * can minimize copying by simply decompressing back over the 5458 * original compressed data (rather than decompressing to an 5459 * aux buffer and then copying back the uncompressed buffer, 5460 * which is likely to be much larger). 5461 */ 5462 uint64_t csize; 5463 void *cdata; 5464 5465 csize = zio->io_size; 5466 cdata = zio_data_buf_alloc(csize); 5467 bcopy(zio->io_data, cdata, csize); 5468 if (zio_decompress_data(c, cdata, zio->io_data, csize, 5469 hdr->b_size) != 0) 5470 zio->io_error = EIO; 5471 zio_data_buf_free(cdata, csize); 5472 } 5473 5474 /* Restore the expected uncompressed IO size. */ 5475 zio->io_orig_size = zio->io_size = hdr->b_size; 5476} 5477 5478/* 5479 * Releases the temporary b_tmp_cdata buffer in an l2arc header structure. 5480 * This buffer serves as a temporary holder of compressed data while 5481 * the buffer entry is being written to an l2arc device. Once that is 5482 * done, we can dispose of it. 5483 */ 5484static void 5485l2arc_release_cdata_buf(arc_buf_hdr_t *hdr) 5486{ 5487 l2arc_buf_hdr_t *l2hdr = hdr->b_l2hdr; 5488 5489 ASSERT(L2ARC_IS_VALID_COMPRESS(l2hdr->b_compress)); 5490 if (l2hdr->b_compress != ZIO_COMPRESS_EMPTY) { 5491 /* 5492 * If the data was compressed, then we've allocated a 5493 * temporary buffer for it, so now we need to release it. 5494 */ 5495 ASSERT(l2hdr->b_tmp_cdata != NULL); 5496 zio_data_buf_free(l2hdr->b_tmp_cdata, hdr->b_size); 5497 l2hdr->b_tmp_cdata = NULL; 5498 } else { 5499 ASSERT(l2hdr->b_tmp_cdata == NULL); 5500 } 5501} 5502 5503/* 5504 * This thread feeds the L2ARC at regular intervals. This is the beating 5505 * heart of the L2ARC. 5506 */ 5507static void 5508l2arc_feed_thread(void *dummy __unused) 5509{ 5510 callb_cpr_t cpr; 5511 l2arc_dev_t *dev; 5512 spa_t *spa; 5513 uint64_t size, wrote; 5514 clock_t begin, next = ddi_get_lbolt(); 5515 boolean_t headroom_boost = B_FALSE; 5516 5517 CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG); 5518 5519 mutex_enter(&l2arc_feed_thr_lock); 5520 5521 while (l2arc_thread_exit == 0) { 5522 CALLB_CPR_SAFE_BEGIN(&cpr); 5523 (void) cv_timedwait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock, 5524 next - ddi_get_lbolt()); 5525 CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock); 5526 next = ddi_get_lbolt() + hz; 5527 5528 /* 5529 * Quick check for L2ARC devices. 5530 */ 5531 mutex_enter(&l2arc_dev_mtx); 5532 if (l2arc_ndev == 0) { 5533 mutex_exit(&l2arc_dev_mtx); 5534 continue; 5535 } 5536 mutex_exit(&l2arc_dev_mtx); 5537 begin = ddi_get_lbolt(); 5538 5539 /* 5540 * This selects the next l2arc device to write to, and in 5541 * doing so the next spa to feed from: dev->l2ad_spa. This 5542 * will return NULL if there are now no l2arc devices or if 5543 * they are all faulted. 5544 * 5545 * If a device is returned, its spa's config lock is also 5546 * held to prevent device removal. l2arc_dev_get_next() 5547 * will grab and release l2arc_dev_mtx. 5548 */ 5549 if ((dev = l2arc_dev_get_next()) == NULL) 5550 continue; 5551 5552 spa = dev->l2ad_spa; 5553 ASSERT(spa != NULL); 5554 5555 /* 5556 * If the pool is read-only then force the feed thread to 5557 * sleep a little longer. 5558 */ 5559 if (!spa_writeable(spa)) { 5560 next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz; 5561 spa_config_exit(spa, SCL_L2ARC, dev); 5562 continue; 5563 } 5564 5565 /* 5566 * Avoid contributing to memory pressure. 5567 */ 5568 if (arc_reclaim_needed()) { 5569 ARCSTAT_BUMP(arcstat_l2_abort_lowmem); 5570 spa_config_exit(spa, SCL_L2ARC, dev); 5571 continue; 5572 } 5573 5574 ARCSTAT_BUMP(arcstat_l2_feeds); 5575 5576 size = l2arc_write_size(); 5577 5578 /* 5579 * Evict L2ARC buffers that will be overwritten. 5580 */ 5581 l2arc_evict(dev, size, B_FALSE); 5582 5583 /* 5584 * Write ARC buffers. 5585 */ 5586 wrote = l2arc_write_buffers(spa, dev, size, &headroom_boost); 5587 5588 /* 5589 * Calculate interval between writes. 5590 */ 5591 next = l2arc_write_interval(begin, size, wrote); 5592 spa_config_exit(spa, SCL_L2ARC, dev); 5593 } 5594 5595 l2arc_thread_exit = 0; 5596 cv_broadcast(&l2arc_feed_thr_cv); 5597 CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */ 5598 thread_exit(); 5599} 5600 5601boolean_t 5602l2arc_vdev_present(vdev_t *vd) 5603{ 5604 l2arc_dev_t *dev; 5605 5606 mutex_enter(&l2arc_dev_mtx); 5607 for (dev = list_head(l2arc_dev_list); dev != NULL; 5608 dev = list_next(l2arc_dev_list, dev)) { 5609 if (dev->l2ad_vdev == vd) 5610 break; 5611 } 5612 mutex_exit(&l2arc_dev_mtx); 5613 5614 return (dev != NULL); 5615} 5616 5617/* 5618 * Add a vdev for use by the L2ARC. By this point the spa has already 5619 * validated the vdev and opened it. 5620 */ 5621void 5622l2arc_add_vdev(spa_t *spa, vdev_t *vd) 5623{ 5624 l2arc_dev_t *adddev; 5625 5626 ASSERT(!l2arc_vdev_present(vd)); 5627 5628 vdev_ashift_optimize(vd); 5629 5630 /* 5631 * Create a new l2arc device entry. 5632 */ 5633 adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP); 5634 adddev->l2ad_spa = spa; 5635 adddev->l2ad_vdev = vd; 5636 adddev->l2ad_start = VDEV_LABEL_START_SIZE; 5637 adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd); 5638 adddev->l2ad_hand = adddev->l2ad_start; 5639 adddev->l2ad_evict = adddev->l2ad_start; 5640 adddev->l2ad_first = B_TRUE; 5641 adddev->l2ad_writing = B_FALSE; 5642 5643 /* 5644 * This is a list of all ARC buffers that are still valid on the 5645 * device. 5646 */ 5647 adddev->l2ad_buflist = kmem_zalloc(sizeof (list_t), KM_SLEEP); 5648 list_create(adddev->l2ad_buflist, sizeof (arc_buf_hdr_t), 5649 offsetof(arc_buf_hdr_t, b_l2node)); 5650 5651 vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand); 5652 5653 /* 5654 * Add device to global list 5655 */ 5656 mutex_enter(&l2arc_dev_mtx); 5657 list_insert_head(l2arc_dev_list, adddev); 5658 atomic_inc_64(&l2arc_ndev); 5659 mutex_exit(&l2arc_dev_mtx); 5660} 5661 5662/* 5663 * Remove a vdev from the L2ARC. 5664 */ 5665void 5666l2arc_remove_vdev(vdev_t *vd) 5667{ 5668 l2arc_dev_t *dev, *nextdev, *remdev = NULL; 5669 5670 /* 5671 * Find the device by vdev 5672 */ 5673 mutex_enter(&l2arc_dev_mtx); 5674 for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) { 5675 nextdev = list_next(l2arc_dev_list, dev); 5676 if (vd == dev->l2ad_vdev) { 5677 remdev = dev; 5678 break; 5679 } 5680 } 5681 ASSERT(remdev != NULL); 5682 5683 /* 5684 * Remove device from global list 5685 */ 5686 list_remove(l2arc_dev_list, remdev); 5687 l2arc_dev_last = NULL; /* may have been invalidated */ 5688 atomic_dec_64(&l2arc_ndev); 5689 mutex_exit(&l2arc_dev_mtx); 5690 5691 /* 5692 * Clear all buflists and ARC references. L2ARC device flush. 5693 */ 5694 l2arc_evict(remdev, 0, B_TRUE); 5695 list_destroy(remdev->l2ad_buflist); 5696 kmem_free(remdev->l2ad_buflist, sizeof (list_t)); 5697 kmem_free(remdev, sizeof (l2arc_dev_t)); 5698} 5699 5700void 5701l2arc_init(void) 5702{ 5703 l2arc_thread_exit = 0; 5704 l2arc_ndev = 0; 5705 l2arc_writes_sent = 0; 5706 l2arc_writes_done = 0; 5707 5708 mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL); 5709 cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL); 5710 mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL); 5711 mutex_init(&l2arc_buflist_mtx, NULL, MUTEX_DEFAULT, NULL); 5712 mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL); 5713 5714 l2arc_dev_list = &L2ARC_dev_list; 5715 l2arc_free_on_write = &L2ARC_free_on_write; 5716 list_create(l2arc_dev_list, sizeof (l2arc_dev_t), 5717 offsetof(l2arc_dev_t, l2ad_node)); 5718 list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t), 5719 offsetof(l2arc_data_free_t, l2df_list_node)); 5720} 5721 5722void 5723l2arc_fini(void) 5724{ 5725 /* 5726 * This is called from dmu_fini(), which is called from spa_fini(); 5727 * Because of this, we can assume that all l2arc devices have 5728 * already been removed when the pools themselves were removed. 5729 */ 5730 5731 l2arc_do_free_on_write(); 5732 5733 mutex_destroy(&l2arc_feed_thr_lock); 5734 cv_destroy(&l2arc_feed_thr_cv); 5735 mutex_destroy(&l2arc_dev_mtx); 5736 mutex_destroy(&l2arc_buflist_mtx); 5737 mutex_destroy(&l2arc_free_on_write_mtx); 5738 5739 list_destroy(l2arc_dev_list); 5740 list_destroy(l2arc_free_on_write); 5741} 5742 5743void 5744l2arc_start(void) 5745{ 5746 if (!(spa_mode_global & FWRITE)) 5747 return; 5748 5749 (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0, 5750 TS_RUN, minclsyspri); 5751} 5752 5753void 5754l2arc_stop(void) 5755{ 5756 if (!(spa_mode_global & FWRITE)) 5757 return; 5758 5759 mutex_enter(&l2arc_feed_thr_lock); 5760 cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */ 5761 l2arc_thread_exit = 1; 5762 while (l2arc_thread_exit != 0) 5763 cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock); 5764 mutex_exit(&l2arc_feed_thr_lock); 5765} 5766