arc.c revision 277583
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 uint32_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; 667static void arc_get_data_buf(arc_buf_t *buf); 668static void arc_access(arc_buf_hdr_t *buf, kmutex_t *hash_lock); 669static int arc_evict_needed(arc_buf_contents_t type); 670static void arc_evict_ghost(arc_state_t *state, uint64_t spa, int64_t bytes); 671#ifdef illumos 672static void arc_buf_watch(arc_buf_t *buf); 673#endif /* illumos */ 674 675static boolean_t l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *ab); 676 677#define GHOST_STATE(state) \ 678 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \ 679 (state) == arc_l2c_only) 680 681/* 682 * Private ARC flags. These flags are private ARC only flags that will show up 683 * in b_flags in the arc_hdr_buf_t. Some flags are publicly declared, and can 684 * be passed in as arc_flags in things like arc_read. However, these flags 685 * should never be passed and should only be set by ARC code. When adding new 686 * public flags, make sure not to smash the private ones. 687 */ 688 689#define ARC_IN_HASH_TABLE (1 << 9) /* this buffer is hashed */ 690#define ARC_IO_IN_PROGRESS (1 << 10) /* I/O in progress for buf */ 691#define ARC_IO_ERROR (1 << 11) /* I/O failed for buf */ 692#define ARC_FREED_IN_READ (1 << 12) /* buf freed while in read */ 693#define ARC_BUF_AVAILABLE (1 << 13) /* block not in active use */ 694#define ARC_INDIRECT (1 << 14) /* this is an indirect block */ 695#define ARC_FREE_IN_PROGRESS (1 << 15) /* hdr about to be freed */ 696#define ARC_L2_WRITING (1 << 16) /* L2ARC write in progress */ 697#define ARC_L2_EVICTED (1 << 17) /* evicted during I/O */ 698#define ARC_L2_WRITE_HEAD (1 << 18) /* head of write list */ 699 700#define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_IN_HASH_TABLE) 701#define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_IO_IN_PROGRESS) 702#define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_IO_ERROR) 703#define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_PREFETCH) 704#define HDR_FREED_IN_READ(hdr) ((hdr)->b_flags & ARC_FREED_IN_READ) 705#define HDR_BUF_AVAILABLE(hdr) ((hdr)->b_flags & ARC_BUF_AVAILABLE) 706#define HDR_FREE_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FREE_IN_PROGRESS) 707#define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_L2CACHE) 708#define HDR_L2_READING(hdr) ((hdr)->b_flags & ARC_IO_IN_PROGRESS && \ 709 (hdr)->b_l2hdr != NULL) 710#define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_L2_WRITING) 711#define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_L2_EVICTED) 712#define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_L2_WRITE_HEAD) 713 714/* 715 * Other sizes 716 */ 717 718#define HDR_SIZE ((int64_t)sizeof (arc_buf_hdr_t)) 719#define L2HDR_SIZE ((int64_t)sizeof (l2arc_buf_hdr_t)) 720 721/* 722 * Hash table routines 723 */ 724 725#define HT_LOCK_PAD CACHE_LINE_SIZE 726 727struct ht_lock { 728 kmutex_t ht_lock; 729#ifdef _KERNEL 730 unsigned char pad[(HT_LOCK_PAD - sizeof (kmutex_t))]; 731#endif 732}; 733 734#define BUF_LOCKS 256 735typedef struct buf_hash_table { 736 uint64_t ht_mask; 737 arc_buf_hdr_t **ht_table; 738 struct ht_lock ht_locks[BUF_LOCKS] __aligned(CACHE_LINE_SIZE); 739} buf_hash_table_t; 740 741static buf_hash_table_t buf_hash_table; 742 743#define BUF_HASH_INDEX(spa, dva, birth) \ 744 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask) 745#define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)]) 746#define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock)) 747#define HDR_LOCK(hdr) \ 748 (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth))) 749 750uint64_t zfs_crc64_table[256]; 751 752/* 753 * Level 2 ARC 754 */ 755 756#define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */ 757#define L2ARC_HEADROOM 2 /* num of writes */ 758/* 759 * If we discover during ARC scan any buffers to be compressed, we boost 760 * our headroom for the next scanning cycle by this percentage multiple. 761 */ 762#define L2ARC_HEADROOM_BOOST 200 763#define L2ARC_FEED_SECS 1 /* caching interval secs */ 764#define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */ 765 766#define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent) 767#define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done) 768 769/* L2ARC Performance Tunables */ 770uint64_t l2arc_write_max = L2ARC_WRITE_SIZE; /* default max write size */ 771uint64_t l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra write during warmup */ 772uint64_t l2arc_headroom = L2ARC_HEADROOM; /* number of dev writes */ 773uint64_t l2arc_headroom_boost = L2ARC_HEADROOM_BOOST; 774uint64_t l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */ 775uint64_t l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval milliseconds */ 776boolean_t l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */ 777boolean_t l2arc_feed_again = B_TRUE; /* turbo warmup */ 778boolean_t l2arc_norw = B_TRUE; /* no reads during writes */ 779 780SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_max, CTLFLAG_RW, 781 &l2arc_write_max, 0, "max write size"); 782SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_boost, CTLFLAG_RW, 783 &l2arc_write_boost, 0, "extra write during warmup"); 784SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_headroom, CTLFLAG_RW, 785 &l2arc_headroom, 0, "number of dev writes"); 786SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_secs, CTLFLAG_RW, 787 &l2arc_feed_secs, 0, "interval seconds"); 788SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_min_ms, CTLFLAG_RW, 789 &l2arc_feed_min_ms, 0, "min interval milliseconds"); 790 791SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_noprefetch, CTLFLAG_RW, 792 &l2arc_noprefetch, 0, "don't cache prefetch bufs"); 793SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_feed_again, CTLFLAG_RW, 794 &l2arc_feed_again, 0, "turbo warmup"); 795SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_norw, CTLFLAG_RW, 796 &l2arc_norw, 0, "no reads during writes"); 797 798SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_size, CTLFLAG_RD, 799 &ARC_anon.arcs_size, 0, "size of anonymous state"); 800SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_metadata_lsize, CTLFLAG_RD, 801 &ARC_anon.arcs_lsize[ARC_BUFC_METADATA], 0, "size of anonymous state"); 802SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_data_lsize, CTLFLAG_RD, 803 &ARC_anon.arcs_lsize[ARC_BUFC_DATA], 0, "size of anonymous state"); 804 805SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_size, CTLFLAG_RD, 806 &ARC_mru.arcs_size, 0, "size of mru state"); 807SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_metadata_lsize, CTLFLAG_RD, 808 &ARC_mru.arcs_lsize[ARC_BUFC_METADATA], 0, "size of metadata in mru state"); 809SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_data_lsize, CTLFLAG_RD, 810 &ARC_mru.arcs_lsize[ARC_BUFC_DATA], 0, "size of data in mru state"); 811 812SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_size, CTLFLAG_RD, 813 &ARC_mru_ghost.arcs_size, 0, "size of mru ghost state"); 814SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_metadata_lsize, CTLFLAG_RD, 815 &ARC_mru_ghost.arcs_lsize[ARC_BUFC_METADATA], 0, 816 "size of metadata in mru ghost state"); 817SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_data_lsize, CTLFLAG_RD, 818 &ARC_mru_ghost.arcs_lsize[ARC_BUFC_DATA], 0, 819 "size of data in mru ghost state"); 820 821SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_size, CTLFLAG_RD, 822 &ARC_mfu.arcs_size, 0, "size of mfu state"); 823SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_metadata_lsize, CTLFLAG_RD, 824 &ARC_mfu.arcs_lsize[ARC_BUFC_METADATA], 0, "size of metadata in mfu state"); 825SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_data_lsize, CTLFLAG_RD, 826 &ARC_mfu.arcs_lsize[ARC_BUFC_DATA], 0, "size of data in mfu state"); 827 828SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_size, CTLFLAG_RD, 829 &ARC_mfu_ghost.arcs_size, 0, "size of mfu ghost state"); 830SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_metadata_lsize, CTLFLAG_RD, 831 &ARC_mfu_ghost.arcs_lsize[ARC_BUFC_METADATA], 0, 832 "size of metadata in mfu ghost state"); 833SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_data_lsize, CTLFLAG_RD, 834 &ARC_mfu_ghost.arcs_lsize[ARC_BUFC_DATA], 0, 835 "size of data in mfu ghost state"); 836 837SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2c_only_size, CTLFLAG_RD, 838 &ARC_l2c_only.arcs_size, 0, "size of mru state"); 839 840/* 841 * L2ARC Internals 842 */ 843typedef struct l2arc_dev { 844 vdev_t *l2ad_vdev; /* vdev */ 845 spa_t *l2ad_spa; /* spa */ 846 uint64_t l2ad_hand; /* next write location */ 847 uint64_t l2ad_start; /* first addr on device */ 848 uint64_t l2ad_end; /* last addr on device */ 849 uint64_t l2ad_evict; /* last addr eviction reached */ 850 boolean_t l2ad_first; /* first sweep through */ 851 boolean_t l2ad_writing; /* currently writing */ 852 list_t *l2ad_buflist; /* buffer list */ 853 list_node_t l2ad_node; /* device list node */ 854} l2arc_dev_t; 855 856static list_t L2ARC_dev_list; /* device list */ 857static list_t *l2arc_dev_list; /* device list pointer */ 858static kmutex_t l2arc_dev_mtx; /* device list mutex */ 859static l2arc_dev_t *l2arc_dev_last; /* last device used */ 860static kmutex_t l2arc_buflist_mtx; /* mutex for all buflists */ 861static list_t L2ARC_free_on_write; /* free after write buf list */ 862static list_t *l2arc_free_on_write; /* free after write list ptr */ 863static kmutex_t l2arc_free_on_write_mtx; /* mutex for list */ 864static uint64_t l2arc_ndev; /* number of devices */ 865 866typedef struct l2arc_read_callback { 867 arc_buf_t *l2rcb_buf; /* read buffer */ 868 spa_t *l2rcb_spa; /* spa */ 869 blkptr_t l2rcb_bp; /* original blkptr */ 870 zbookmark_phys_t l2rcb_zb; /* original bookmark */ 871 int l2rcb_flags; /* original flags */ 872 enum zio_compress l2rcb_compress; /* applied compress */ 873} l2arc_read_callback_t; 874 875typedef struct l2arc_write_callback { 876 l2arc_dev_t *l2wcb_dev; /* device info */ 877 arc_buf_hdr_t *l2wcb_head; /* head of write buflist */ 878} l2arc_write_callback_t; 879 880struct l2arc_buf_hdr { 881 /* protected by arc_buf_hdr mutex */ 882 l2arc_dev_t *b_dev; /* L2ARC device */ 883 uint64_t b_daddr; /* disk address, offset byte */ 884 /* compression applied to buffer data */ 885 enum zio_compress b_compress; 886 /* real alloc'd buffer size depending on b_compress applied */ 887 int b_asize; 888 /* temporary buffer holder for in-flight compressed data */ 889 void *b_tmp_cdata; 890}; 891 892typedef struct l2arc_data_free { 893 /* protected by l2arc_free_on_write_mtx */ 894 void *l2df_data; 895 size_t l2df_size; 896 void (*l2df_func)(void *, size_t); 897 list_node_t l2df_list_node; 898} l2arc_data_free_t; 899 900static kmutex_t l2arc_feed_thr_lock; 901static kcondvar_t l2arc_feed_thr_cv; 902static uint8_t l2arc_thread_exit; 903 904static void l2arc_read_done(zio_t *zio); 905static void l2arc_hdr_stat_add(void); 906static void l2arc_hdr_stat_remove(void); 907 908static boolean_t l2arc_compress_buf(l2arc_buf_hdr_t *l2hdr); 909static void l2arc_decompress_zio(zio_t *zio, arc_buf_hdr_t *hdr, 910 enum zio_compress c); 911static void l2arc_release_cdata_buf(arc_buf_hdr_t *ab); 912 913static uint64_t 914buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth) 915{ 916 uint8_t *vdva = (uint8_t *)dva; 917 uint64_t crc = -1ULL; 918 int i; 919 920 ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY); 921 922 for (i = 0; i < sizeof (dva_t); i++) 923 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ vdva[i]) & 0xFF]; 924 925 crc ^= (spa>>8) ^ birth; 926 927 return (crc); 928} 929 930#define BUF_EMPTY(buf) \ 931 ((buf)->b_dva.dva_word[0] == 0 && \ 932 (buf)->b_dva.dva_word[1] == 0 && \ 933 (buf)->b_cksum0 == 0) 934 935#define BUF_EQUAL(spa, dva, birth, buf) \ 936 ((buf)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \ 937 ((buf)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \ 938 ((buf)->b_birth == birth) && ((buf)->b_spa == spa) 939 940static void 941buf_discard_identity(arc_buf_hdr_t *hdr) 942{ 943 hdr->b_dva.dva_word[0] = 0; 944 hdr->b_dva.dva_word[1] = 0; 945 hdr->b_birth = 0; 946 hdr->b_cksum0 = 0; 947} 948 949static arc_buf_hdr_t * 950buf_hash_find(uint64_t spa, const blkptr_t *bp, kmutex_t **lockp) 951{ 952 const dva_t *dva = BP_IDENTITY(bp); 953 uint64_t birth = BP_PHYSICAL_BIRTH(bp); 954 uint64_t idx = BUF_HASH_INDEX(spa, dva, birth); 955 kmutex_t *hash_lock = BUF_HASH_LOCK(idx); 956 arc_buf_hdr_t *buf; 957 958 mutex_enter(hash_lock); 959 for (buf = buf_hash_table.ht_table[idx]; buf != NULL; 960 buf = buf->b_hash_next) { 961 if (BUF_EQUAL(spa, dva, birth, buf)) { 962 *lockp = hash_lock; 963 return (buf); 964 } 965 } 966 mutex_exit(hash_lock); 967 *lockp = NULL; 968 return (NULL); 969} 970 971/* 972 * Insert an entry into the hash table. If there is already an element 973 * equal to elem in the hash table, then the already existing element 974 * will be returned and the new element will not be inserted. 975 * Otherwise returns NULL. 976 */ 977static arc_buf_hdr_t * 978buf_hash_insert(arc_buf_hdr_t *buf, kmutex_t **lockp) 979{ 980 uint64_t idx = BUF_HASH_INDEX(buf->b_spa, &buf->b_dva, buf->b_birth); 981 kmutex_t *hash_lock = BUF_HASH_LOCK(idx); 982 arc_buf_hdr_t *fbuf; 983 uint32_t i; 984 985 ASSERT(!DVA_IS_EMPTY(&buf->b_dva)); 986 ASSERT(buf->b_birth != 0); 987 ASSERT(!HDR_IN_HASH_TABLE(buf)); 988 *lockp = hash_lock; 989 mutex_enter(hash_lock); 990 for (fbuf = buf_hash_table.ht_table[idx], i = 0; fbuf != NULL; 991 fbuf = fbuf->b_hash_next, i++) { 992 if (BUF_EQUAL(buf->b_spa, &buf->b_dva, buf->b_birth, fbuf)) 993 return (fbuf); 994 } 995 996 buf->b_hash_next = buf_hash_table.ht_table[idx]; 997 buf_hash_table.ht_table[idx] = buf; 998 buf->b_flags |= ARC_IN_HASH_TABLE; 999 1000 /* collect some hash table performance data */ 1001 if (i > 0) { 1002 ARCSTAT_BUMP(arcstat_hash_collisions); 1003 if (i == 1) 1004 ARCSTAT_BUMP(arcstat_hash_chains); 1005 1006 ARCSTAT_MAX(arcstat_hash_chain_max, i); 1007 } 1008 1009 ARCSTAT_BUMP(arcstat_hash_elements); 1010 ARCSTAT_MAXSTAT(arcstat_hash_elements); 1011 1012 return (NULL); 1013} 1014 1015static void 1016buf_hash_remove(arc_buf_hdr_t *buf) 1017{ 1018 arc_buf_hdr_t *fbuf, **bufp; 1019 uint64_t idx = BUF_HASH_INDEX(buf->b_spa, &buf->b_dva, buf->b_birth); 1020 1021 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx))); 1022 ASSERT(HDR_IN_HASH_TABLE(buf)); 1023 1024 bufp = &buf_hash_table.ht_table[idx]; 1025 while ((fbuf = *bufp) != buf) { 1026 ASSERT(fbuf != NULL); 1027 bufp = &fbuf->b_hash_next; 1028 } 1029 *bufp = buf->b_hash_next; 1030 buf->b_hash_next = NULL; 1031 buf->b_flags &= ~ARC_IN_HASH_TABLE; 1032 1033 /* collect some hash table performance data */ 1034 ARCSTAT_BUMPDOWN(arcstat_hash_elements); 1035 1036 if (buf_hash_table.ht_table[idx] && 1037 buf_hash_table.ht_table[idx]->b_hash_next == NULL) 1038 ARCSTAT_BUMPDOWN(arcstat_hash_chains); 1039} 1040 1041/* 1042 * Global data structures and functions for the buf kmem cache. 1043 */ 1044static kmem_cache_t *hdr_cache; 1045static kmem_cache_t *buf_cache; 1046 1047static void 1048buf_fini(void) 1049{ 1050 int i; 1051 1052 kmem_free(buf_hash_table.ht_table, 1053 (buf_hash_table.ht_mask + 1) * sizeof (void *)); 1054 for (i = 0; i < BUF_LOCKS; i++) 1055 mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock); 1056 kmem_cache_destroy(hdr_cache); 1057 kmem_cache_destroy(buf_cache); 1058} 1059 1060/* 1061 * Constructor callback - called when the cache is empty 1062 * and a new buf is requested. 1063 */ 1064/* ARGSUSED */ 1065static int 1066hdr_cons(void *vbuf, void *unused, int kmflag) 1067{ 1068 arc_buf_hdr_t *buf = vbuf; 1069 1070 bzero(buf, sizeof (arc_buf_hdr_t)); 1071 refcount_create(&buf->b_refcnt); 1072 cv_init(&buf->b_cv, NULL, CV_DEFAULT, NULL); 1073 mutex_init(&buf->b_freeze_lock, NULL, MUTEX_DEFAULT, NULL); 1074 arc_space_consume(sizeof (arc_buf_hdr_t), ARC_SPACE_HDRS); 1075 1076 return (0); 1077} 1078 1079/* ARGSUSED */ 1080static int 1081buf_cons(void *vbuf, void *unused, int kmflag) 1082{ 1083 arc_buf_t *buf = vbuf; 1084 1085 bzero(buf, sizeof (arc_buf_t)); 1086 mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL); 1087 arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS); 1088 1089 return (0); 1090} 1091 1092/* 1093 * Destructor callback - called when a cached buf is 1094 * no longer required. 1095 */ 1096/* ARGSUSED */ 1097static void 1098hdr_dest(void *vbuf, void *unused) 1099{ 1100 arc_buf_hdr_t *buf = vbuf; 1101 1102 ASSERT(BUF_EMPTY(buf)); 1103 refcount_destroy(&buf->b_refcnt); 1104 cv_destroy(&buf->b_cv); 1105 mutex_destroy(&buf->b_freeze_lock); 1106 arc_space_return(sizeof (arc_buf_hdr_t), ARC_SPACE_HDRS); 1107} 1108 1109/* ARGSUSED */ 1110static void 1111buf_dest(void *vbuf, void *unused) 1112{ 1113 arc_buf_t *buf = vbuf; 1114 1115 mutex_destroy(&buf->b_evict_lock); 1116 arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS); 1117} 1118 1119/* 1120 * Reclaim callback -- invoked when memory is low. 1121 */ 1122/* ARGSUSED */ 1123static void 1124hdr_recl(void *unused) 1125{ 1126 dprintf("hdr_recl called\n"); 1127 /* 1128 * umem calls the reclaim func when we destroy the buf cache, 1129 * which is after we do arc_fini(). 1130 */ 1131 if (!arc_dead) 1132 cv_signal(&arc_reclaim_thr_cv); 1133} 1134 1135static void 1136buf_init(void) 1137{ 1138 uint64_t *ct; 1139 uint64_t hsize = 1ULL << 12; 1140 int i, j; 1141 1142 /* 1143 * The hash table is big enough to fill all of physical memory 1144 * with an average block size of zfs_arc_average_blocksize (default 8K). 1145 * By default, the table will take up 1146 * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers). 1147 */ 1148 while (hsize * zfs_arc_average_blocksize < (uint64_t)physmem * PAGESIZE) 1149 hsize <<= 1; 1150retry: 1151 buf_hash_table.ht_mask = hsize - 1; 1152 buf_hash_table.ht_table = 1153 kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP); 1154 if (buf_hash_table.ht_table == NULL) { 1155 ASSERT(hsize > (1ULL << 8)); 1156 hsize >>= 1; 1157 goto retry; 1158 } 1159 1160 hdr_cache = kmem_cache_create("arc_buf_hdr_t", sizeof (arc_buf_hdr_t), 1161 0, hdr_cons, hdr_dest, hdr_recl, NULL, NULL, 0); 1162 buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t), 1163 0, buf_cons, buf_dest, NULL, NULL, NULL, 0); 1164 1165 for (i = 0; i < 256; i++) 1166 for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--) 1167 *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY); 1168 1169 for (i = 0; i < BUF_LOCKS; i++) { 1170 mutex_init(&buf_hash_table.ht_locks[i].ht_lock, 1171 NULL, MUTEX_DEFAULT, NULL); 1172 } 1173} 1174 1175#define ARC_MINTIME (hz>>4) /* 62 ms */ 1176 1177static void 1178arc_cksum_verify(arc_buf_t *buf) 1179{ 1180 zio_cksum_t zc; 1181 1182 if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 1183 return; 1184 1185 mutex_enter(&buf->b_hdr->b_freeze_lock); 1186 if (buf->b_hdr->b_freeze_cksum == NULL || 1187 (buf->b_hdr->b_flags & ARC_IO_ERROR)) { 1188 mutex_exit(&buf->b_hdr->b_freeze_lock); 1189 return; 1190 } 1191 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc); 1192 if (!ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc)) 1193 panic("buffer modified while frozen!"); 1194 mutex_exit(&buf->b_hdr->b_freeze_lock); 1195} 1196 1197static int 1198arc_cksum_equal(arc_buf_t *buf) 1199{ 1200 zio_cksum_t zc; 1201 int equal; 1202 1203 mutex_enter(&buf->b_hdr->b_freeze_lock); 1204 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc); 1205 equal = ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc); 1206 mutex_exit(&buf->b_hdr->b_freeze_lock); 1207 1208 return (equal); 1209} 1210 1211static void 1212arc_cksum_compute(arc_buf_t *buf, boolean_t force) 1213{ 1214 if (!force && !(zfs_flags & ZFS_DEBUG_MODIFY)) 1215 return; 1216 1217 mutex_enter(&buf->b_hdr->b_freeze_lock); 1218 if (buf->b_hdr->b_freeze_cksum != NULL) { 1219 mutex_exit(&buf->b_hdr->b_freeze_lock); 1220 return; 1221 } 1222 buf->b_hdr->b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t), KM_SLEEP); 1223 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, 1224 buf->b_hdr->b_freeze_cksum); 1225 mutex_exit(&buf->b_hdr->b_freeze_lock); 1226#ifdef illumos 1227 arc_buf_watch(buf); 1228#endif /* illumos */ 1229} 1230 1231#ifdef illumos 1232#ifndef _KERNEL 1233typedef struct procctl { 1234 long cmd; 1235 prwatch_t prwatch; 1236} procctl_t; 1237#endif 1238 1239/* ARGSUSED */ 1240static void 1241arc_buf_unwatch(arc_buf_t *buf) 1242{ 1243#ifndef _KERNEL 1244 if (arc_watch) { 1245 int result; 1246 procctl_t ctl; 1247 ctl.cmd = PCWATCH; 1248 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data; 1249 ctl.prwatch.pr_size = 0; 1250 ctl.prwatch.pr_wflags = 0; 1251 result = write(arc_procfd, &ctl, sizeof (ctl)); 1252 ASSERT3U(result, ==, sizeof (ctl)); 1253 } 1254#endif 1255} 1256 1257/* ARGSUSED */ 1258static void 1259arc_buf_watch(arc_buf_t *buf) 1260{ 1261#ifndef _KERNEL 1262 if (arc_watch) { 1263 int result; 1264 procctl_t ctl; 1265 ctl.cmd = PCWATCH; 1266 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data; 1267 ctl.prwatch.pr_size = buf->b_hdr->b_size; 1268 ctl.prwatch.pr_wflags = WA_WRITE; 1269 result = write(arc_procfd, &ctl, sizeof (ctl)); 1270 ASSERT3U(result, ==, sizeof (ctl)); 1271 } 1272#endif 1273} 1274#endif /* illumos */ 1275 1276void 1277arc_buf_thaw(arc_buf_t *buf) 1278{ 1279 if (zfs_flags & ZFS_DEBUG_MODIFY) { 1280 if (buf->b_hdr->b_state != arc_anon) 1281 panic("modifying non-anon buffer!"); 1282 if (buf->b_hdr->b_flags & ARC_IO_IN_PROGRESS) 1283 panic("modifying buffer while i/o in progress!"); 1284 arc_cksum_verify(buf); 1285 } 1286 1287 mutex_enter(&buf->b_hdr->b_freeze_lock); 1288 if (buf->b_hdr->b_freeze_cksum != NULL) { 1289 kmem_free(buf->b_hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 1290 buf->b_hdr->b_freeze_cksum = NULL; 1291 } 1292 1293 if (zfs_flags & ZFS_DEBUG_MODIFY) { 1294 if (buf->b_hdr->b_thawed) 1295 kmem_free(buf->b_hdr->b_thawed, 1); 1296 buf->b_hdr->b_thawed = kmem_alloc(1, KM_SLEEP); 1297 } 1298 1299 mutex_exit(&buf->b_hdr->b_freeze_lock); 1300 1301#ifdef illumos 1302 arc_buf_unwatch(buf); 1303#endif /* illumos */ 1304} 1305 1306void 1307arc_buf_freeze(arc_buf_t *buf) 1308{ 1309 kmutex_t *hash_lock; 1310 1311 if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 1312 return; 1313 1314 hash_lock = HDR_LOCK(buf->b_hdr); 1315 mutex_enter(hash_lock); 1316 1317 ASSERT(buf->b_hdr->b_freeze_cksum != NULL || 1318 buf->b_hdr->b_state == arc_anon); 1319 arc_cksum_compute(buf, B_FALSE); 1320 mutex_exit(hash_lock); 1321 1322} 1323 1324static void 1325get_buf_info(arc_buf_hdr_t *ab, arc_state_t *state, list_t **list, kmutex_t **lock) 1326{ 1327 uint64_t buf_hashid = buf_hash(ab->b_spa, &ab->b_dva, ab->b_birth); 1328 1329 if (ab->b_type == ARC_BUFC_METADATA) 1330 buf_hashid &= (ARC_BUFC_NUMMETADATALISTS - 1); 1331 else { 1332 buf_hashid &= (ARC_BUFC_NUMDATALISTS - 1); 1333 buf_hashid += ARC_BUFC_NUMMETADATALISTS; 1334 } 1335 1336 *list = &state->arcs_lists[buf_hashid]; 1337 *lock = ARCS_LOCK(state, buf_hashid); 1338} 1339 1340 1341static void 1342add_reference(arc_buf_hdr_t *ab, kmutex_t *hash_lock, void *tag) 1343{ 1344 ASSERT(MUTEX_HELD(hash_lock)); 1345 1346 if ((refcount_add(&ab->b_refcnt, tag) == 1) && 1347 (ab->b_state != arc_anon)) { 1348 uint64_t delta = ab->b_size * ab->b_datacnt; 1349 uint64_t *size = &ab->b_state->arcs_lsize[ab->b_type]; 1350 list_t *list; 1351 kmutex_t *lock; 1352 1353 get_buf_info(ab, ab->b_state, &list, &lock); 1354 ASSERT(!MUTEX_HELD(lock)); 1355 mutex_enter(lock); 1356 ASSERT(list_link_active(&ab->b_arc_node)); 1357 list_remove(list, ab); 1358 if (GHOST_STATE(ab->b_state)) { 1359 ASSERT0(ab->b_datacnt); 1360 ASSERT3P(ab->b_buf, ==, NULL); 1361 delta = ab->b_size; 1362 } 1363 ASSERT(delta > 0); 1364 ASSERT3U(*size, >=, delta); 1365 atomic_add_64(size, -delta); 1366 mutex_exit(lock); 1367 /* remove the prefetch flag if we get a reference */ 1368 if (ab->b_flags & ARC_PREFETCH) 1369 ab->b_flags &= ~ARC_PREFETCH; 1370 } 1371} 1372 1373static int 1374remove_reference(arc_buf_hdr_t *ab, kmutex_t *hash_lock, void *tag) 1375{ 1376 int cnt; 1377 arc_state_t *state = ab->b_state; 1378 1379 ASSERT(state == arc_anon || MUTEX_HELD(hash_lock)); 1380 ASSERT(!GHOST_STATE(state)); 1381 1382 if (((cnt = refcount_remove(&ab->b_refcnt, tag)) == 0) && 1383 (state != arc_anon)) { 1384 uint64_t *size = &state->arcs_lsize[ab->b_type]; 1385 list_t *list; 1386 kmutex_t *lock; 1387 1388 get_buf_info(ab, state, &list, &lock); 1389 ASSERT(!MUTEX_HELD(lock)); 1390 mutex_enter(lock); 1391 ASSERT(!list_link_active(&ab->b_arc_node)); 1392 list_insert_head(list, ab); 1393 ASSERT(ab->b_datacnt > 0); 1394 atomic_add_64(size, ab->b_size * ab->b_datacnt); 1395 mutex_exit(lock); 1396 } 1397 return (cnt); 1398} 1399 1400/* 1401 * Move the supplied buffer to the indicated state. The mutex 1402 * for the buffer must be held by the caller. 1403 */ 1404static void 1405arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *ab, kmutex_t *hash_lock) 1406{ 1407 arc_state_t *old_state = ab->b_state; 1408 int64_t refcnt = refcount_count(&ab->b_refcnt); 1409 uint64_t from_delta, to_delta; 1410 list_t *list; 1411 kmutex_t *lock; 1412 1413 ASSERT(MUTEX_HELD(hash_lock)); 1414 ASSERT3P(new_state, !=, old_state); 1415 ASSERT(refcnt == 0 || ab->b_datacnt > 0); 1416 ASSERT(ab->b_datacnt == 0 || !GHOST_STATE(new_state)); 1417 ASSERT(ab->b_datacnt <= 1 || old_state != arc_anon); 1418 1419 from_delta = to_delta = ab->b_datacnt * ab->b_size; 1420 1421 /* 1422 * If this buffer is evictable, transfer it from the 1423 * old state list to the new state list. 1424 */ 1425 if (refcnt == 0) { 1426 if (old_state != arc_anon) { 1427 int use_mutex; 1428 uint64_t *size = &old_state->arcs_lsize[ab->b_type]; 1429 1430 get_buf_info(ab, old_state, &list, &lock); 1431 use_mutex = !MUTEX_HELD(lock); 1432 if (use_mutex) 1433 mutex_enter(lock); 1434 1435 ASSERT(list_link_active(&ab->b_arc_node)); 1436 list_remove(list, ab); 1437 1438 /* 1439 * If prefetching out of the ghost cache, 1440 * we will have a non-zero datacnt. 1441 */ 1442 if (GHOST_STATE(old_state) && ab->b_datacnt == 0) { 1443 /* ghost elements have a ghost size */ 1444 ASSERT(ab->b_buf == NULL); 1445 from_delta = ab->b_size; 1446 } 1447 ASSERT3U(*size, >=, from_delta); 1448 atomic_add_64(size, -from_delta); 1449 1450 if (use_mutex) 1451 mutex_exit(lock); 1452 } 1453 if (new_state != arc_anon) { 1454 int use_mutex; 1455 uint64_t *size = &new_state->arcs_lsize[ab->b_type]; 1456 1457 get_buf_info(ab, new_state, &list, &lock); 1458 use_mutex = !MUTEX_HELD(lock); 1459 if (use_mutex) 1460 mutex_enter(lock); 1461 1462 list_insert_head(list, ab); 1463 1464 /* ghost elements have a ghost size */ 1465 if (GHOST_STATE(new_state)) { 1466 ASSERT(ab->b_datacnt == 0); 1467 ASSERT(ab->b_buf == NULL); 1468 to_delta = ab->b_size; 1469 } 1470 atomic_add_64(size, to_delta); 1471 1472 if (use_mutex) 1473 mutex_exit(lock); 1474 } 1475 } 1476 1477 ASSERT(!BUF_EMPTY(ab)); 1478 if (new_state == arc_anon && HDR_IN_HASH_TABLE(ab)) 1479 buf_hash_remove(ab); 1480 1481 /* adjust state sizes */ 1482 if (to_delta) 1483 atomic_add_64(&new_state->arcs_size, to_delta); 1484 if (from_delta) { 1485 ASSERT3U(old_state->arcs_size, >=, from_delta); 1486 atomic_add_64(&old_state->arcs_size, -from_delta); 1487 } 1488 ab->b_state = new_state; 1489 1490 /* adjust l2arc hdr stats */ 1491 if (new_state == arc_l2c_only) 1492 l2arc_hdr_stat_add(); 1493 else if (old_state == arc_l2c_only) 1494 l2arc_hdr_stat_remove(); 1495} 1496 1497void 1498arc_space_consume(uint64_t space, arc_space_type_t type) 1499{ 1500 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES); 1501 1502 switch (type) { 1503 case ARC_SPACE_DATA: 1504 ARCSTAT_INCR(arcstat_data_size, space); 1505 break; 1506 case ARC_SPACE_OTHER: 1507 ARCSTAT_INCR(arcstat_other_size, space); 1508 break; 1509 case ARC_SPACE_HDRS: 1510 ARCSTAT_INCR(arcstat_hdr_size, space); 1511 break; 1512 case ARC_SPACE_L2HDRS: 1513 ARCSTAT_INCR(arcstat_l2_hdr_size, space); 1514 break; 1515 } 1516 1517 ARCSTAT_INCR(arcstat_meta_used, space); 1518 atomic_add_64(&arc_size, space); 1519} 1520 1521void 1522arc_space_return(uint64_t space, arc_space_type_t type) 1523{ 1524 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES); 1525 1526 switch (type) { 1527 case ARC_SPACE_DATA: 1528 ARCSTAT_INCR(arcstat_data_size, -space); 1529 break; 1530 case ARC_SPACE_OTHER: 1531 ARCSTAT_INCR(arcstat_other_size, -space); 1532 break; 1533 case ARC_SPACE_HDRS: 1534 ARCSTAT_INCR(arcstat_hdr_size, -space); 1535 break; 1536 case ARC_SPACE_L2HDRS: 1537 ARCSTAT_INCR(arcstat_l2_hdr_size, -space); 1538 break; 1539 } 1540 1541 ASSERT(arc_meta_used >= space); 1542 if (arc_meta_max < arc_meta_used) 1543 arc_meta_max = arc_meta_used; 1544 ARCSTAT_INCR(arcstat_meta_used, -space); 1545 ASSERT(arc_size >= space); 1546 atomic_add_64(&arc_size, -space); 1547} 1548 1549arc_buf_t * 1550arc_buf_alloc(spa_t *spa, int size, void *tag, arc_buf_contents_t type) 1551{ 1552 arc_buf_hdr_t *hdr; 1553 arc_buf_t *buf; 1554 1555 ASSERT3U(size, >, 0); 1556 hdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE); 1557 ASSERT(BUF_EMPTY(hdr)); 1558 hdr->b_size = size; 1559 hdr->b_type = type; 1560 hdr->b_spa = spa_load_guid(spa); 1561 hdr->b_state = arc_anon; 1562 hdr->b_arc_access = 0; 1563 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 1564 buf->b_hdr = hdr; 1565 buf->b_data = NULL; 1566 buf->b_efunc = NULL; 1567 buf->b_private = NULL; 1568 buf->b_next = NULL; 1569 hdr->b_buf = buf; 1570 arc_get_data_buf(buf); 1571 hdr->b_datacnt = 1; 1572 hdr->b_flags = 0; 1573 ASSERT(refcount_is_zero(&hdr->b_refcnt)); 1574 (void) refcount_add(&hdr->b_refcnt, tag); 1575 1576 return (buf); 1577} 1578 1579static char *arc_onloan_tag = "onloan"; 1580 1581/* 1582 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in 1583 * flight data by arc_tempreserve_space() until they are "returned". Loaned 1584 * buffers must be returned to the arc before they can be used by the DMU or 1585 * freed. 1586 */ 1587arc_buf_t * 1588arc_loan_buf(spa_t *spa, int size) 1589{ 1590 arc_buf_t *buf; 1591 1592 buf = arc_buf_alloc(spa, size, arc_onloan_tag, ARC_BUFC_DATA); 1593 1594 atomic_add_64(&arc_loaned_bytes, size); 1595 return (buf); 1596} 1597 1598/* 1599 * Return a loaned arc buffer to the arc. 1600 */ 1601void 1602arc_return_buf(arc_buf_t *buf, void *tag) 1603{ 1604 arc_buf_hdr_t *hdr = buf->b_hdr; 1605 1606 ASSERT(buf->b_data != NULL); 1607 (void) refcount_add(&hdr->b_refcnt, tag); 1608 (void) refcount_remove(&hdr->b_refcnt, arc_onloan_tag); 1609 1610 atomic_add_64(&arc_loaned_bytes, -hdr->b_size); 1611} 1612 1613/* Detach an arc_buf from a dbuf (tag) */ 1614void 1615arc_loan_inuse_buf(arc_buf_t *buf, void *tag) 1616{ 1617 arc_buf_hdr_t *hdr; 1618 1619 ASSERT(buf->b_data != NULL); 1620 hdr = buf->b_hdr; 1621 (void) refcount_add(&hdr->b_refcnt, arc_onloan_tag); 1622 (void) refcount_remove(&hdr->b_refcnt, tag); 1623 buf->b_efunc = NULL; 1624 buf->b_private = NULL; 1625 1626 atomic_add_64(&arc_loaned_bytes, hdr->b_size); 1627} 1628 1629static arc_buf_t * 1630arc_buf_clone(arc_buf_t *from) 1631{ 1632 arc_buf_t *buf; 1633 arc_buf_hdr_t *hdr = from->b_hdr; 1634 uint64_t size = hdr->b_size; 1635 1636 ASSERT(hdr->b_state != arc_anon); 1637 1638 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 1639 buf->b_hdr = hdr; 1640 buf->b_data = NULL; 1641 buf->b_efunc = NULL; 1642 buf->b_private = NULL; 1643 buf->b_next = hdr->b_buf; 1644 hdr->b_buf = buf; 1645 arc_get_data_buf(buf); 1646 bcopy(from->b_data, buf->b_data, size); 1647 1648 /* 1649 * This buffer already exists in the arc so create a duplicate 1650 * copy for the caller. If the buffer is associated with user data 1651 * then track the size and number of duplicates. These stats will be 1652 * updated as duplicate buffers are created and destroyed. 1653 */ 1654 if (hdr->b_type == ARC_BUFC_DATA) { 1655 ARCSTAT_BUMP(arcstat_duplicate_buffers); 1656 ARCSTAT_INCR(arcstat_duplicate_buffers_size, size); 1657 } 1658 hdr->b_datacnt += 1; 1659 return (buf); 1660} 1661 1662void 1663arc_buf_add_ref(arc_buf_t *buf, void* tag) 1664{ 1665 arc_buf_hdr_t *hdr; 1666 kmutex_t *hash_lock; 1667 1668 /* 1669 * Check to see if this buffer is evicted. Callers 1670 * must verify b_data != NULL to know if the add_ref 1671 * was successful. 1672 */ 1673 mutex_enter(&buf->b_evict_lock); 1674 if (buf->b_data == NULL) { 1675 mutex_exit(&buf->b_evict_lock); 1676 return; 1677 } 1678 hash_lock = HDR_LOCK(buf->b_hdr); 1679 mutex_enter(hash_lock); 1680 hdr = buf->b_hdr; 1681 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 1682 mutex_exit(&buf->b_evict_lock); 1683 1684 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu); 1685 add_reference(hdr, hash_lock, tag); 1686 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr); 1687 arc_access(hdr, hash_lock); 1688 mutex_exit(hash_lock); 1689 ARCSTAT_BUMP(arcstat_hits); 1690 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH), 1691 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA, 1692 data, metadata, hits); 1693} 1694 1695static void 1696arc_buf_free_on_write(void *data, size_t size, 1697 void (*free_func)(void *, size_t)) 1698{ 1699 l2arc_data_free_t *df; 1700 1701 df = kmem_alloc(sizeof (l2arc_data_free_t), KM_SLEEP); 1702 df->l2df_data = data; 1703 df->l2df_size = size; 1704 df->l2df_func = free_func; 1705 mutex_enter(&l2arc_free_on_write_mtx); 1706 list_insert_head(l2arc_free_on_write, df); 1707 mutex_exit(&l2arc_free_on_write_mtx); 1708} 1709 1710/* 1711 * Free the arc data buffer. If it is an l2arc write in progress, 1712 * the buffer is placed on l2arc_free_on_write to be freed later. 1713 */ 1714static void 1715arc_buf_data_free(arc_buf_t *buf, void (*free_func)(void *, size_t)) 1716{ 1717 arc_buf_hdr_t *hdr = buf->b_hdr; 1718 1719 if (HDR_L2_WRITING(hdr)) { 1720 arc_buf_free_on_write(buf->b_data, hdr->b_size, free_func); 1721 ARCSTAT_BUMP(arcstat_l2_free_on_write); 1722 } else { 1723 free_func(buf->b_data, hdr->b_size); 1724 } 1725} 1726 1727/* 1728 * Free up buf->b_data and if 'remove' is set, then pull the 1729 * arc_buf_t off of the the arc_buf_hdr_t's list and free it. 1730 */ 1731static void 1732arc_buf_l2_cdata_free(arc_buf_hdr_t *hdr) 1733{ 1734 l2arc_buf_hdr_t *l2hdr = hdr->b_l2hdr; 1735 1736 ASSERT(MUTEX_HELD(&l2arc_buflist_mtx)); 1737 1738 if (l2hdr->b_tmp_cdata == NULL) 1739 return; 1740 1741 ASSERT(HDR_L2_WRITING(hdr)); 1742 arc_buf_free_on_write(l2hdr->b_tmp_cdata, hdr->b_size, 1743 zio_data_buf_free); 1744 ARCSTAT_BUMP(arcstat_l2_cdata_free_on_write); 1745 l2hdr->b_tmp_cdata = NULL; 1746} 1747 1748static void 1749arc_buf_destroy(arc_buf_t *buf, boolean_t recycle, boolean_t remove) 1750{ 1751 arc_buf_t **bufp; 1752 1753 /* free up data associated with the buf */ 1754 if (buf->b_data) { 1755 arc_state_t *state = buf->b_hdr->b_state; 1756 uint64_t size = buf->b_hdr->b_size; 1757 arc_buf_contents_t type = buf->b_hdr->b_type; 1758 1759 arc_cksum_verify(buf); 1760#ifdef illumos 1761 arc_buf_unwatch(buf); 1762#endif /* illumos */ 1763 1764 if (!recycle) { 1765 if (type == ARC_BUFC_METADATA) { 1766 arc_buf_data_free(buf, zio_buf_free); 1767 arc_space_return(size, ARC_SPACE_DATA); 1768 } else { 1769 ASSERT(type == ARC_BUFC_DATA); 1770 arc_buf_data_free(buf, zio_data_buf_free); 1771 ARCSTAT_INCR(arcstat_data_size, -size); 1772 atomic_add_64(&arc_size, -size); 1773 } 1774 } 1775 if (list_link_active(&buf->b_hdr->b_arc_node)) { 1776 uint64_t *cnt = &state->arcs_lsize[type]; 1777 1778 ASSERT(refcount_is_zero(&buf->b_hdr->b_refcnt)); 1779 ASSERT(state != arc_anon); 1780 1781 ASSERT3U(*cnt, >=, size); 1782 atomic_add_64(cnt, -size); 1783 } 1784 ASSERT3U(state->arcs_size, >=, size); 1785 atomic_add_64(&state->arcs_size, -size); 1786 buf->b_data = NULL; 1787 1788 /* 1789 * If we're destroying a duplicate buffer make sure 1790 * that the appropriate statistics are updated. 1791 */ 1792 if (buf->b_hdr->b_datacnt > 1 && 1793 buf->b_hdr->b_type == ARC_BUFC_DATA) { 1794 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers); 1795 ARCSTAT_INCR(arcstat_duplicate_buffers_size, -size); 1796 } 1797 ASSERT(buf->b_hdr->b_datacnt > 0); 1798 buf->b_hdr->b_datacnt -= 1; 1799 } 1800 1801 /* only remove the buf if requested */ 1802 if (!remove) 1803 return; 1804 1805 /* remove the buf from the hdr list */ 1806 for (bufp = &buf->b_hdr->b_buf; *bufp != buf; bufp = &(*bufp)->b_next) 1807 continue; 1808 *bufp = buf->b_next; 1809 buf->b_next = NULL; 1810 1811 ASSERT(buf->b_efunc == NULL); 1812 1813 /* clean up the buf */ 1814 buf->b_hdr = NULL; 1815 kmem_cache_free(buf_cache, buf); 1816} 1817 1818static void 1819arc_hdr_destroy(arc_buf_hdr_t *hdr) 1820{ 1821 ASSERT(refcount_is_zero(&hdr->b_refcnt)); 1822 ASSERT3P(hdr->b_state, ==, arc_anon); 1823 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 1824 l2arc_buf_hdr_t *l2hdr = hdr->b_l2hdr; 1825 1826 if (l2hdr != NULL) { 1827 boolean_t buflist_held = MUTEX_HELD(&l2arc_buflist_mtx); 1828 /* 1829 * To prevent arc_free() and l2arc_evict() from 1830 * attempting to free the same buffer at the same time, 1831 * a FREE_IN_PROGRESS flag is given to arc_free() to 1832 * give it priority. l2arc_evict() can't destroy this 1833 * header while we are waiting on l2arc_buflist_mtx. 1834 * 1835 * The hdr may be removed from l2ad_buflist before we 1836 * grab l2arc_buflist_mtx, so b_l2hdr is rechecked. 1837 */ 1838 if (!buflist_held) { 1839 mutex_enter(&l2arc_buflist_mtx); 1840 l2hdr = hdr->b_l2hdr; 1841 } 1842 1843 if (l2hdr != NULL) { 1844 trim_map_free(l2hdr->b_dev->l2ad_vdev, l2hdr->b_daddr, 1845 hdr->b_size, 0); 1846 list_remove(l2hdr->b_dev->l2ad_buflist, hdr); 1847 arc_buf_l2_cdata_free(hdr); 1848 ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size); 1849 ARCSTAT_INCR(arcstat_l2_asize, -l2hdr->b_asize); 1850 vdev_space_update(l2hdr->b_dev->l2ad_vdev, 1851 -l2hdr->b_asize, 0, 0); 1852 kmem_free(l2hdr, sizeof (l2arc_buf_hdr_t)); 1853 if (hdr->b_state == arc_l2c_only) 1854 l2arc_hdr_stat_remove(); 1855 hdr->b_l2hdr = NULL; 1856 } 1857 1858 if (!buflist_held) 1859 mutex_exit(&l2arc_buflist_mtx); 1860 } 1861 1862 if (!BUF_EMPTY(hdr)) { 1863 ASSERT(!HDR_IN_HASH_TABLE(hdr)); 1864 buf_discard_identity(hdr); 1865 } 1866 while (hdr->b_buf) { 1867 arc_buf_t *buf = hdr->b_buf; 1868 1869 if (buf->b_efunc) { 1870 mutex_enter(&arc_eviction_mtx); 1871 mutex_enter(&buf->b_evict_lock); 1872 ASSERT(buf->b_hdr != NULL); 1873 arc_buf_destroy(hdr->b_buf, FALSE, FALSE); 1874 hdr->b_buf = buf->b_next; 1875 buf->b_hdr = &arc_eviction_hdr; 1876 buf->b_next = arc_eviction_list; 1877 arc_eviction_list = buf; 1878 mutex_exit(&buf->b_evict_lock); 1879 mutex_exit(&arc_eviction_mtx); 1880 } else { 1881 arc_buf_destroy(hdr->b_buf, FALSE, TRUE); 1882 } 1883 } 1884 if (hdr->b_freeze_cksum != NULL) { 1885 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 1886 hdr->b_freeze_cksum = NULL; 1887 } 1888 if (hdr->b_thawed) { 1889 kmem_free(hdr->b_thawed, 1); 1890 hdr->b_thawed = NULL; 1891 } 1892 1893 ASSERT(!list_link_active(&hdr->b_arc_node)); 1894 ASSERT3P(hdr->b_hash_next, ==, NULL); 1895 ASSERT3P(hdr->b_acb, ==, NULL); 1896 kmem_cache_free(hdr_cache, hdr); 1897} 1898 1899void 1900arc_buf_free(arc_buf_t *buf, void *tag) 1901{ 1902 arc_buf_hdr_t *hdr = buf->b_hdr; 1903 int hashed = hdr->b_state != arc_anon; 1904 1905 ASSERT(buf->b_efunc == NULL); 1906 ASSERT(buf->b_data != NULL); 1907 1908 if (hashed) { 1909 kmutex_t *hash_lock = HDR_LOCK(hdr); 1910 1911 mutex_enter(hash_lock); 1912 hdr = buf->b_hdr; 1913 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 1914 1915 (void) remove_reference(hdr, hash_lock, tag); 1916 if (hdr->b_datacnt > 1) { 1917 arc_buf_destroy(buf, FALSE, TRUE); 1918 } else { 1919 ASSERT(buf == hdr->b_buf); 1920 ASSERT(buf->b_efunc == NULL); 1921 hdr->b_flags |= ARC_BUF_AVAILABLE; 1922 } 1923 mutex_exit(hash_lock); 1924 } else if (HDR_IO_IN_PROGRESS(hdr)) { 1925 int destroy_hdr; 1926 /* 1927 * We are in the middle of an async write. Don't destroy 1928 * this buffer unless the write completes before we finish 1929 * decrementing the reference count. 1930 */ 1931 mutex_enter(&arc_eviction_mtx); 1932 (void) remove_reference(hdr, NULL, tag); 1933 ASSERT(refcount_is_zero(&hdr->b_refcnt)); 1934 destroy_hdr = !HDR_IO_IN_PROGRESS(hdr); 1935 mutex_exit(&arc_eviction_mtx); 1936 if (destroy_hdr) 1937 arc_hdr_destroy(hdr); 1938 } else { 1939 if (remove_reference(hdr, NULL, tag) > 0) 1940 arc_buf_destroy(buf, FALSE, TRUE); 1941 else 1942 arc_hdr_destroy(hdr); 1943 } 1944} 1945 1946boolean_t 1947arc_buf_remove_ref(arc_buf_t *buf, void* tag) 1948{ 1949 arc_buf_hdr_t *hdr = buf->b_hdr; 1950 kmutex_t *hash_lock = HDR_LOCK(hdr); 1951 boolean_t no_callback = (buf->b_efunc == NULL); 1952 1953 if (hdr->b_state == arc_anon) { 1954 ASSERT(hdr->b_datacnt == 1); 1955 arc_buf_free(buf, tag); 1956 return (no_callback); 1957 } 1958 1959 mutex_enter(hash_lock); 1960 hdr = buf->b_hdr; 1961 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 1962 ASSERT(hdr->b_state != arc_anon); 1963 ASSERT(buf->b_data != NULL); 1964 1965 (void) remove_reference(hdr, hash_lock, tag); 1966 if (hdr->b_datacnt > 1) { 1967 if (no_callback) 1968 arc_buf_destroy(buf, FALSE, TRUE); 1969 } else if (no_callback) { 1970 ASSERT(hdr->b_buf == buf && buf->b_next == NULL); 1971 ASSERT(buf->b_efunc == NULL); 1972 hdr->b_flags |= ARC_BUF_AVAILABLE; 1973 } 1974 ASSERT(no_callback || hdr->b_datacnt > 1 || 1975 refcount_is_zero(&hdr->b_refcnt)); 1976 mutex_exit(hash_lock); 1977 return (no_callback); 1978} 1979 1980int 1981arc_buf_size(arc_buf_t *buf) 1982{ 1983 return (buf->b_hdr->b_size); 1984} 1985 1986/* 1987 * Called from the DMU to determine if the current buffer should be 1988 * evicted. In order to ensure proper locking, the eviction must be initiated 1989 * from the DMU. Return true if the buffer is associated with user data and 1990 * duplicate buffers still exist. 1991 */ 1992boolean_t 1993arc_buf_eviction_needed(arc_buf_t *buf) 1994{ 1995 arc_buf_hdr_t *hdr; 1996 boolean_t evict_needed = B_FALSE; 1997 1998 if (zfs_disable_dup_eviction) 1999 return (B_FALSE); 2000 2001 mutex_enter(&buf->b_evict_lock); 2002 hdr = buf->b_hdr; 2003 if (hdr == NULL) { 2004 /* 2005 * We are in arc_do_user_evicts(); let that function 2006 * perform the eviction. 2007 */ 2008 ASSERT(buf->b_data == NULL); 2009 mutex_exit(&buf->b_evict_lock); 2010 return (B_FALSE); 2011 } else if (buf->b_data == NULL) { 2012 /* 2013 * We have already been added to the arc eviction list; 2014 * recommend eviction. 2015 */ 2016 ASSERT3P(hdr, ==, &arc_eviction_hdr); 2017 mutex_exit(&buf->b_evict_lock); 2018 return (B_TRUE); 2019 } 2020 2021 if (hdr->b_datacnt > 1 && hdr->b_type == ARC_BUFC_DATA) 2022 evict_needed = B_TRUE; 2023 2024 mutex_exit(&buf->b_evict_lock); 2025 return (evict_needed); 2026} 2027 2028/* 2029 * Evict buffers from list until we've removed the specified number of 2030 * bytes. Move the removed buffers to the appropriate evict state. 2031 * If the recycle flag is set, then attempt to "recycle" a buffer: 2032 * - look for a buffer to evict that is `bytes' long. 2033 * - return the data block from this buffer rather than freeing it. 2034 * This flag is used by callers that are trying to make space for a 2035 * new buffer in a full arc cache. 2036 * 2037 * This function makes a "best effort". It skips over any buffers 2038 * it can't get a hash_lock on, and so may not catch all candidates. 2039 * It may also return without evicting as much space as requested. 2040 */ 2041static void * 2042arc_evict(arc_state_t *state, uint64_t spa, int64_t bytes, boolean_t recycle, 2043 arc_buf_contents_t type) 2044{ 2045 arc_state_t *evicted_state; 2046 uint64_t bytes_evicted = 0, skipped = 0, missed = 0; 2047 int64_t bytes_remaining; 2048 arc_buf_hdr_t *ab, *ab_prev = NULL; 2049 list_t *evicted_list, *list, *evicted_list_start, *list_start; 2050 kmutex_t *lock, *evicted_lock; 2051 kmutex_t *hash_lock; 2052 boolean_t have_lock; 2053 void *stolen = NULL; 2054 arc_buf_hdr_t marker = { 0 }; 2055 int count = 0; 2056 static int evict_metadata_offset, evict_data_offset; 2057 int i, idx, offset, list_count, lists; 2058 2059 ASSERT(state == arc_mru || state == arc_mfu); 2060 2061 evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost; 2062 2063 if (type == ARC_BUFC_METADATA) { 2064 offset = 0; 2065 list_count = ARC_BUFC_NUMMETADATALISTS; 2066 list_start = &state->arcs_lists[0]; 2067 evicted_list_start = &evicted_state->arcs_lists[0]; 2068 idx = evict_metadata_offset; 2069 } else { 2070 offset = ARC_BUFC_NUMMETADATALISTS; 2071 list_start = &state->arcs_lists[offset]; 2072 evicted_list_start = &evicted_state->arcs_lists[offset]; 2073 list_count = ARC_BUFC_NUMDATALISTS; 2074 idx = evict_data_offset; 2075 } 2076 bytes_remaining = evicted_state->arcs_lsize[type]; 2077 lists = 0; 2078 2079evict_start: 2080 list = &list_start[idx]; 2081 evicted_list = &evicted_list_start[idx]; 2082 lock = ARCS_LOCK(state, (offset + idx)); 2083 evicted_lock = ARCS_LOCK(evicted_state, (offset + idx)); 2084 2085 mutex_enter(lock); 2086 mutex_enter(evicted_lock); 2087 2088 for (ab = list_tail(list); ab; ab = ab_prev) { 2089 ab_prev = list_prev(list, ab); 2090 bytes_remaining -= (ab->b_size * ab->b_datacnt); 2091 /* prefetch buffers have a minimum lifespan */ 2092 if (HDR_IO_IN_PROGRESS(ab) || 2093 (spa && ab->b_spa != spa) || 2094 (ab->b_flags & (ARC_PREFETCH|ARC_INDIRECT) && 2095 ddi_get_lbolt() - ab->b_arc_access < 2096 arc_min_prefetch_lifespan)) { 2097 skipped++; 2098 continue; 2099 } 2100 /* "lookahead" for better eviction candidate */ 2101 if (recycle && ab->b_size != bytes && 2102 ab_prev && ab_prev->b_size == bytes) 2103 continue; 2104 2105 /* ignore markers */ 2106 if (ab->b_spa == 0) 2107 continue; 2108 2109 /* 2110 * It may take a long time to evict all the bufs requested. 2111 * To avoid blocking all arc activity, periodically drop 2112 * the arcs_mtx and give other threads a chance to run 2113 * before reacquiring the lock. 2114 * 2115 * If we are looking for a buffer to recycle, we are in 2116 * the hot code path, so don't sleep. 2117 */ 2118 if (!recycle && count++ > arc_evict_iterations) { 2119 list_insert_after(list, ab, &marker); 2120 mutex_exit(evicted_lock); 2121 mutex_exit(lock); 2122 kpreempt(KPREEMPT_SYNC); 2123 mutex_enter(lock); 2124 mutex_enter(evicted_lock); 2125 ab_prev = list_prev(list, &marker); 2126 list_remove(list, &marker); 2127 count = 0; 2128 continue; 2129 } 2130 2131 hash_lock = HDR_LOCK(ab); 2132 have_lock = MUTEX_HELD(hash_lock); 2133 if (have_lock || mutex_tryenter(hash_lock)) { 2134 ASSERT0(refcount_count(&ab->b_refcnt)); 2135 ASSERT(ab->b_datacnt > 0); 2136 while (ab->b_buf) { 2137 arc_buf_t *buf = ab->b_buf; 2138 if (!mutex_tryenter(&buf->b_evict_lock)) { 2139 missed += 1; 2140 break; 2141 } 2142 if (buf->b_data) { 2143 bytes_evicted += ab->b_size; 2144 if (recycle && ab->b_type == type && 2145 ab->b_size == bytes && 2146 !HDR_L2_WRITING(ab)) { 2147 stolen = buf->b_data; 2148 recycle = FALSE; 2149 } 2150 } 2151 if (buf->b_efunc) { 2152 mutex_enter(&arc_eviction_mtx); 2153 arc_buf_destroy(buf, 2154 buf->b_data == stolen, FALSE); 2155 ab->b_buf = buf->b_next; 2156 buf->b_hdr = &arc_eviction_hdr; 2157 buf->b_next = arc_eviction_list; 2158 arc_eviction_list = buf; 2159 mutex_exit(&arc_eviction_mtx); 2160 mutex_exit(&buf->b_evict_lock); 2161 } else { 2162 mutex_exit(&buf->b_evict_lock); 2163 arc_buf_destroy(buf, 2164 buf->b_data == stolen, TRUE); 2165 } 2166 } 2167 2168 if (ab->b_l2hdr) { 2169 ARCSTAT_INCR(arcstat_evict_l2_cached, 2170 ab->b_size); 2171 } else { 2172 if (l2arc_write_eligible(ab->b_spa, ab)) { 2173 ARCSTAT_INCR(arcstat_evict_l2_eligible, 2174 ab->b_size); 2175 } else { 2176 ARCSTAT_INCR( 2177 arcstat_evict_l2_ineligible, 2178 ab->b_size); 2179 } 2180 } 2181 2182 if (ab->b_datacnt == 0) { 2183 arc_change_state(evicted_state, ab, hash_lock); 2184 ASSERT(HDR_IN_HASH_TABLE(ab)); 2185 ab->b_flags |= ARC_IN_HASH_TABLE; 2186 ab->b_flags &= ~ARC_BUF_AVAILABLE; 2187 DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, ab); 2188 } 2189 if (!have_lock) 2190 mutex_exit(hash_lock); 2191 if (bytes >= 0 && bytes_evicted >= bytes) 2192 break; 2193 if (bytes_remaining > 0) { 2194 mutex_exit(evicted_lock); 2195 mutex_exit(lock); 2196 idx = ((idx + 1) & (list_count - 1)); 2197 lists++; 2198 goto evict_start; 2199 } 2200 } else { 2201 missed += 1; 2202 } 2203 } 2204 2205 mutex_exit(evicted_lock); 2206 mutex_exit(lock); 2207 2208 idx = ((idx + 1) & (list_count - 1)); 2209 lists++; 2210 2211 if (bytes_evicted < bytes) { 2212 if (lists < list_count) 2213 goto evict_start; 2214 else 2215 dprintf("only evicted %lld bytes from %x", 2216 (longlong_t)bytes_evicted, state); 2217 } 2218 if (type == ARC_BUFC_METADATA) 2219 evict_metadata_offset = idx; 2220 else 2221 evict_data_offset = idx; 2222 2223 if (skipped) 2224 ARCSTAT_INCR(arcstat_evict_skip, skipped); 2225 2226 if (missed) 2227 ARCSTAT_INCR(arcstat_mutex_miss, missed); 2228 2229 /* 2230 * Note: we have just evicted some data into the ghost state, 2231 * potentially putting the ghost size over the desired size. Rather 2232 * that evicting from the ghost list in this hot code path, leave 2233 * this chore to the arc_reclaim_thread(). 2234 */ 2235 2236 if (stolen) 2237 ARCSTAT_BUMP(arcstat_stolen); 2238 return (stolen); 2239} 2240 2241/* 2242 * Remove buffers from list until we've removed the specified number of 2243 * bytes. Destroy the buffers that are removed. 2244 */ 2245static void 2246arc_evict_ghost(arc_state_t *state, uint64_t spa, int64_t bytes) 2247{ 2248 arc_buf_hdr_t *ab, *ab_prev; 2249 arc_buf_hdr_t marker = { 0 }; 2250 list_t *list, *list_start; 2251 kmutex_t *hash_lock, *lock; 2252 uint64_t bytes_deleted = 0; 2253 uint64_t bufs_skipped = 0; 2254 int count = 0; 2255 static int evict_offset; 2256 int list_count, idx = evict_offset; 2257 int offset, lists = 0; 2258 2259 ASSERT(GHOST_STATE(state)); 2260 2261 /* 2262 * data lists come after metadata lists 2263 */ 2264 list_start = &state->arcs_lists[ARC_BUFC_NUMMETADATALISTS]; 2265 list_count = ARC_BUFC_NUMDATALISTS; 2266 offset = ARC_BUFC_NUMMETADATALISTS; 2267 2268evict_start: 2269 list = &list_start[idx]; 2270 lock = ARCS_LOCK(state, idx + offset); 2271 2272 mutex_enter(lock); 2273 for (ab = list_tail(list); ab; ab = ab_prev) { 2274 ab_prev = list_prev(list, ab); 2275 if (ab->b_type > ARC_BUFC_NUMTYPES) 2276 panic("invalid ab=%p", (void *)ab); 2277 if (spa && ab->b_spa != spa) 2278 continue; 2279 2280 /* ignore markers */ 2281 if (ab->b_spa == 0) 2282 continue; 2283 2284 hash_lock = HDR_LOCK(ab); 2285 /* caller may be trying to modify this buffer, skip it */ 2286 if (MUTEX_HELD(hash_lock)) 2287 continue; 2288 2289 /* 2290 * It may take a long time to evict all the bufs requested. 2291 * To avoid blocking all arc activity, periodically drop 2292 * the arcs_mtx and give other threads a chance to run 2293 * before reacquiring the lock. 2294 */ 2295 if (count++ > arc_evict_iterations) { 2296 list_insert_after(list, ab, &marker); 2297 mutex_exit(lock); 2298 kpreempt(KPREEMPT_SYNC); 2299 mutex_enter(lock); 2300 ab_prev = list_prev(list, &marker); 2301 list_remove(list, &marker); 2302 count = 0; 2303 continue; 2304 } 2305 if (mutex_tryenter(hash_lock)) { 2306 ASSERT(!HDR_IO_IN_PROGRESS(ab)); 2307 ASSERT(ab->b_buf == NULL); 2308 ARCSTAT_BUMP(arcstat_deleted); 2309 bytes_deleted += ab->b_size; 2310 2311 if (ab->b_l2hdr != NULL) { 2312 /* 2313 * This buffer is cached on the 2nd Level ARC; 2314 * don't destroy the header. 2315 */ 2316 arc_change_state(arc_l2c_only, ab, hash_lock); 2317 mutex_exit(hash_lock); 2318 } else { 2319 arc_change_state(arc_anon, ab, hash_lock); 2320 mutex_exit(hash_lock); 2321 arc_hdr_destroy(ab); 2322 } 2323 2324 DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, ab); 2325 if (bytes >= 0 && bytes_deleted >= bytes) 2326 break; 2327 } else if (bytes < 0) { 2328 /* 2329 * Insert a list marker and then wait for the 2330 * hash lock to become available. Once its 2331 * available, restart from where we left off. 2332 */ 2333 list_insert_after(list, ab, &marker); 2334 mutex_exit(lock); 2335 mutex_enter(hash_lock); 2336 mutex_exit(hash_lock); 2337 mutex_enter(lock); 2338 ab_prev = list_prev(list, &marker); 2339 list_remove(list, &marker); 2340 } else { 2341 bufs_skipped += 1; 2342 } 2343 2344 } 2345 mutex_exit(lock); 2346 idx = ((idx + 1) & (ARC_BUFC_NUMDATALISTS - 1)); 2347 lists++; 2348 2349 if (lists < list_count) 2350 goto evict_start; 2351 2352 evict_offset = idx; 2353 if ((uintptr_t)list > (uintptr_t)&state->arcs_lists[ARC_BUFC_NUMMETADATALISTS] && 2354 (bytes < 0 || bytes_deleted < bytes)) { 2355 list_start = &state->arcs_lists[0]; 2356 list_count = ARC_BUFC_NUMMETADATALISTS; 2357 offset = lists = 0; 2358 goto evict_start; 2359 } 2360 2361 if (bufs_skipped) { 2362 ARCSTAT_INCR(arcstat_mutex_miss, bufs_skipped); 2363 ASSERT(bytes >= 0); 2364 } 2365 2366 if (bytes_deleted < bytes) 2367 dprintf("only deleted %lld bytes from %p", 2368 (longlong_t)bytes_deleted, state); 2369} 2370 2371static void 2372arc_adjust(void) 2373{ 2374 int64_t adjustment, delta; 2375 2376 /* 2377 * Adjust MRU size 2378 */ 2379 2380 adjustment = MIN((int64_t)(arc_size - arc_c), 2381 (int64_t)(arc_anon->arcs_size + arc_mru->arcs_size + arc_meta_used - 2382 arc_p)); 2383 2384 if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_DATA] > 0) { 2385 delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_DATA], adjustment); 2386 (void) arc_evict(arc_mru, 0, delta, FALSE, ARC_BUFC_DATA); 2387 adjustment -= delta; 2388 } 2389 2390 if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_METADATA] > 0) { 2391 delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_METADATA], adjustment); 2392 (void) arc_evict(arc_mru, 0, delta, FALSE, 2393 ARC_BUFC_METADATA); 2394 } 2395 2396 /* 2397 * Adjust MFU size 2398 */ 2399 2400 adjustment = arc_size - arc_c; 2401 2402 if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_DATA] > 0) { 2403 delta = MIN(adjustment, arc_mfu->arcs_lsize[ARC_BUFC_DATA]); 2404 (void) arc_evict(arc_mfu, 0, delta, FALSE, ARC_BUFC_DATA); 2405 adjustment -= delta; 2406 } 2407 2408 if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_METADATA] > 0) { 2409 int64_t delta = MIN(adjustment, 2410 arc_mfu->arcs_lsize[ARC_BUFC_METADATA]); 2411 (void) arc_evict(arc_mfu, 0, delta, FALSE, 2412 ARC_BUFC_METADATA); 2413 } 2414 2415 /* 2416 * Adjust ghost lists 2417 */ 2418 2419 adjustment = arc_mru->arcs_size + arc_mru_ghost->arcs_size - arc_c; 2420 2421 if (adjustment > 0 && arc_mru_ghost->arcs_size > 0) { 2422 delta = MIN(arc_mru_ghost->arcs_size, adjustment); 2423 arc_evict_ghost(arc_mru_ghost, 0, delta); 2424 } 2425 2426 adjustment = 2427 arc_mru_ghost->arcs_size + arc_mfu_ghost->arcs_size - arc_c; 2428 2429 if (adjustment > 0 && arc_mfu_ghost->arcs_size > 0) { 2430 delta = MIN(arc_mfu_ghost->arcs_size, adjustment); 2431 arc_evict_ghost(arc_mfu_ghost, 0, delta); 2432 } 2433} 2434 2435static void 2436arc_do_user_evicts(void) 2437{ 2438 static arc_buf_t *tmp_arc_eviction_list; 2439 2440 /* 2441 * Move list over to avoid LOR 2442 */ 2443restart: 2444 mutex_enter(&arc_eviction_mtx); 2445 tmp_arc_eviction_list = arc_eviction_list; 2446 arc_eviction_list = NULL; 2447 mutex_exit(&arc_eviction_mtx); 2448 2449 while (tmp_arc_eviction_list != NULL) { 2450 arc_buf_t *buf = tmp_arc_eviction_list; 2451 tmp_arc_eviction_list = buf->b_next; 2452 mutex_enter(&buf->b_evict_lock); 2453 buf->b_hdr = NULL; 2454 mutex_exit(&buf->b_evict_lock); 2455 2456 if (buf->b_efunc != NULL) 2457 VERIFY0(buf->b_efunc(buf->b_private)); 2458 2459 buf->b_efunc = NULL; 2460 buf->b_private = NULL; 2461 kmem_cache_free(buf_cache, buf); 2462 } 2463 2464 if (arc_eviction_list != NULL) 2465 goto restart; 2466} 2467 2468/* 2469 * Flush all *evictable* data from the cache for the given spa. 2470 * NOTE: this will not touch "active" (i.e. referenced) data. 2471 */ 2472void 2473arc_flush(spa_t *spa) 2474{ 2475 uint64_t guid = 0; 2476 2477 if (spa) 2478 guid = spa_load_guid(spa); 2479 2480 while (arc_mru->arcs_lsize[ARC_BUFC_DATA]) { 2481 (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_DATA); 2482 if (spa) 2483 break; 2484 } 2485 while (arc_mru->arcs_lsize[ARC_BUFC_METADATA]) { 2486 (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_METADATA); 2487 if (spa) 2488 break; 2489 } 2490 while (arc_mfu->arcs_lsize[ARC_BUFC_DATA]) { 2491 (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_DATA); 2492 if (spa) 2493 break; 2494 } 2495 while (arc_mfu->arcs_lsize[ARC_BUFC_METADATA]) { 2496 (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_METADATA); 2497 if (spa) 2498 break; 2499 } 2500 2501 arc_evict_ghost(arc_mru_ghost, guid, -1); 2502 arc_evict_ghost(arc_mfu_ghost, guid, -1); 2503 2504 mutex_enter(&arc_reclaim_thr_lock); 2505 arc_do_user_evicts(); 2506 mutex_exit(&arc_reclaim_thr_lock); 2507 ASSERT(spa || arc_eviction_list == NULL); 2508} 2509 2510void 2511arc_shrink(void) 2512{ 2513 2514 if (arc_c > arc_c_min) { 2515 uint64_t to_free; 2516 2517 DTRACE_PROBE4(arc__shrink, uint64_t, arc_c, uint64_t, 2518 arc_c_min, uint64_t, arc_p, uint64_t, to_free); 2519#ifdef _KERNEL 2520 to_free = arc_c >> arc_shrink_shift; 2521#else 2522 to_free = arc_c >> arc_shrink_shift; 2523#endif 2524 if (arc_c > arc_c_min + to_free) 2525 atomic_add_64(&arc_c, -to_free); 2526 else 2527 arc_c = arc_c_min; 2528 2529 atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift)); 2530 if (arc_c > arc_size) 2531 arc_c = MAX(arc_size, arc_c_min); 2532 if (arc_p > arc_c) 2533 arc_p = (arc_c >> 1); 2534 2535 DTRACE_PROBE2(arc__shrunk, uint64_t, arc_c, uint64_t, 2536 arc_p); 2537 2538 ASSERT(arc_c >= arc_c_min); 2539 ASSERT((int64_t)arc_p >= 0); 2540 } 2541 2542 if (arc_size > arc_c) { 2543 DTRACE_PROBE2(arc__shrink_adjust, uint64_t, arc_size, 2544 uint64_t, arc_c); 2545 arc_adjust(); 2546 } 2547} 2548 2549static int needfree = 0; 2550 2551static int 2552arc_reclaim_needed(void) 2553{ 2554 2555#ifdef _KERNEL 2556 2557 if (needfree) { 2558 DTRACE_PROBE(arc__reclaim_needfree); 2559 return (1); 2560 } 2561 2562 /* 2563 * Cooperate with pagedaemon when it's time for it to scan 2564 * and reclaim some pages. 2565 */ 2566 if (freemem < zfs_arc_free_target) { 2567 DTRACE_PROBE2(arc__reclaim_freemem, uint64_t, 2568 freemem, uint64_t, zfs_arc_free_target); 2569 return (1); 2570 } 2571 2572#ifdef sun 2573 /* 2574 * take 'desfree' extra pages, so we reclaim sooner, rather than later 2575 */ 2576 extra = desfree; 2577 2578 /* 2579 * check that we're out of range of the pageout scanner. It starts to 2580 * schedule paging if freemem is less than lotsfree and needfree. 2581 * lotsfree is the high-water mark for pageout, and needfree is the 2582 * number of needed free pages. We add extra pages here to make sure 2583 * the scanner doesn't start up while we're freeing memory. 2584 */ 2585 if (freemem < lotsfree + needfree + extra) 2586 return (1); 2587 2588 /* 2589 * check to make sure that swapfs has enough space so that anon 2590 * reservations can still succeed. anon_resvmem() checks that the 2591 * availrmem is greater than swapfs_minfree, and the number of reserved 2592 * swap pages. We also add a bit of extra here just to prevent 2593 * circumstances from getting really dire. 2594 */ 2595 if (availrmem < swapfs_minfree + swapfs_reserve + extra) 2596 return (1); 2597 2598 /* 2599 * Check that we have enough availrmem that memory locking (e.g., via 2600 * mlock(3C) or memcntl(2)) can still succeed. (pages_pp_maximum 2601 * stores the number of pages that cannot be locked; when availrmem 2602 * drops below pages_pp_maximum, page locking mechanisms such as 2603 * page_pp_lock() will fail.) 2604 */ 2605 if (availrmem <= pages_pp_maximum) 2606 return (1); 2607 2608#endif /* sun */ 2609#if defined(__i386) || !defined(UMA_MD_SMALL_ALLOC) 2610 /* 2611 * If we're on an i386 platform, it's possible that we'll exhaust the 2612 * kernel heap space before we ever run out of available physical 2613 * memory. Most checks of the size of the heap_area compare against 2614 * tune.t_minarmem, which is the minimum available real memory that we 2615 * can have in the system. However, this is generally fixed at 25 pages 2616 * which is so low that it's useless. In this comparison, we seek to 2617 * calculate the total heap-size, and reclaim if more than 3/4ths of the 2618 * heap is allocated. (Or, in the calculation, if less than 1/4th is 2619 * free) 2620 */ 2621 if (vmem_size(heap_arena, VMEM_FREE) < 2622 (vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC) >> 2)) { 2623 DTRACE_PROBE2(arc__reclaim_used, uint64_t, 2624 vmem_size(heap_arena, VMEM_FREE), uint64_t, 2625 (vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC)) >> 2); 2626 return (1); 2627 } 2628#endif 2629#ifdef sun 2630 /* 2631 * If zio data pages are being allocated out of a separate heap segment, 2632 * then enforce that the size of available vmem for this arena remains 2633 * above about 1/16th free. 2634 * 2635 * Note: The 1/16th arena free requirement was put in place 2636 * to aggressively evict memory from the arc in order to avoid 2637 * memory fragmentation issues. 2638 */ 2639 if (zio_arena != NULL && 2640 vmem_size(zio_arena, VMEM_FREE) < 2641 (vmem_size(zio_arena, VMEM_ALLOC) >> 4)) 2642 return (1); 2643#endif /* sun */ 2644#else /* _KERNEL */ 2645 if (spa_get_random(100) == 0) 2646 return (1); 2647#endif /* _KERNEL */ 2648 DTRACE_PROBE(arc__reclaim_no); 2649 2650 return (0); 2651} 2652 2653extern kmem_cache_t *zio_buf_cache[]; 2654extern kmem_cache_t *zio_data_buf_cache[]; 2655extern kmem_cache_t *range_seg_cache; 2656 2657static void __noinline 2658arc_kmem_reap_now(arc_reclaim_strategy_t strat) 2659{ 2660 size_t i; 2661 kmem_cache_t *prev_cache = NULL; 2662 kmem_cache_t *prev_data_cache = NULL; 2663 2664 DTRACE_PROBE(arc__kmem_reap_start); 2665#ifdef _KERNEL 2666 if (arc_meta_used >= arc_meta_limit) { 2667 /* 2668 * We are exceeding our meta-data cache limit. 2669 * Purge some DNLC entries to release holds on meta-data. 2670 */ 2671 dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent); 2672 } 2673#if defined(__i386) 2674 /* 2675 * Reclaim unused memory from all kmem caches. 2676 */ 2677 kmem_reap(); 2678#endif 2679#endif 2680 2681 /* 2682 * An aggressive reclamation will shrink the cache size as well as 2683 * reap free buffers from the arc kmem caches. 2684 */ 2685 if (strat == ARC_RECLAIM_AGGR) 2686 arc_shrink(); 2687 2688 for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) { 2689 if (zio_buf_cache[i] != prev_cache) { 2690 prev_cache = zio_buf_cache[i]; 2691 kmem_cache_reap_now(zio_buf_cache[i]); 2692 } 2693 if (zio_data_buf_cache[i] != prev_data_cache) { 2694 prev_data_cache = zio_data_buf_cache[i]; 2695 kmem_cache_reap_now(zio_data_buf_cache[i]); 2696 } 2697 } 2698 kmem_cache_reap_now(buf_cache); 2699 kmem_cache_reap_now(hdr_cache); 2700 kmem_cache_reap_now(range_seg_cache); 2701 2702#ifdef sun 2703 /* 2704 * Ask the vmem arena to reclaim unused memory from its 2705 * quantum caches. 2706 */ 2707 if (zio_arena != NULL && strat == ARC_RECLAIM_AGGR) 2708 vmem_qcache_reap(zio_arena); 2709#endif 2710 DTRACE_PROBE(arc__kmem_reap_end); 2711} 2712 2713static void 2714arc_reclaim_thread(void *dummy __unused) 2715{ 2716 clock_t growtime = 0; 2717 arc_reclaim_strategy_t last_reclaim = ARC_RECLAIM_CONS; 2718 callb_cpr_t cpr; 2719 2720 CALLB_CPR_INIT(&cpr, &arc_reclaim_thr_lock, callb_generic_cpr, FTAG); 2721 2722 mutex_enter(&arc_reclaim_thr_lock); 2723 while (arc_thread_exit == 0) { 2724 if (arc_reclaim_needed()) { 2725 2726 if (arc_no_grow) { 2727 if (last_reclaim == ARC_RECLAIM_CONS) { 2728 DTRACE_PROBE(arc__reclaim_aggr_no_grow); 2729 last_reclaim = ARC_RECLAIM_AGGR; 2730 } else { 2731 last_reclaim = ARC_RECLAIM_CONS; 2732 } 2733 } else { 2734 arc_no_grow = TRUE; 2735 last_reclaim = ARC_RECLAIM_AGGR; 2736 DTRACE_PROBE(arc__reclaim_aggr); 2737 membar_producer(); 2738 } 2739 2740 /* reset the growth delay for every reclaim */ 2741 growtime = ddi_get_lbolt() + (arc_grow_retry * hz); 2742 2743 if (needfree && last_reclaim == ARC_RECLAIM_CONS) { 2744 /* 2745 * If needfree is TRUE our vm_lowmem hook 2746 * was called and in that case we must free some 2747 * memory, so switch to aggressive mode. 2748 */ 2749 arc_no_grow = TRUE; 2750 last_reclaim = ARC_RECLAIM_AGGR; 2751 } 2752 arc_kmem_reap_now(last_reclaim); 2753 arc_warm = B_TRUE; 2754 2755 } else if (arc_no_grow && ddi_get_lbolt() >= growtime) { 2756 arc_no_grow = FALSE; 2757 } 2758 2759 arc_adjust(); 2760 2761 if (arc_eviction_list != NULL) 2762 arc_do_user_evicts(); 2763 2764#ifdef _KERNEL 2765 if (needfree) { 2766 needfree = 0; 2767 wakeup(&needfree); 2768 } 2769#endif 2770 2771 /* block until needed, or one second, whichever is shorter */ 2772 CALLB_CPR_SAFE_BEGIN(&cpr); 2773 (void) cv_timedwait(&arc_reclaim_thr_cv, 2774 &arc_reclaim_thr_lock, hz); 2775 CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_thr_lock); 2776 } 2777 2778 arc_thread_exit = 0; 2779 cv_broadcast(&arc_reclaim_thr_cv); 2780 CALLB_CPR_EXIT(&cpr); /* drops arc_reclaim_thr_lock */ 2781 thread_exit(); 2782} 2783 2784/* 2785 * Adapt arc info given the number of bytes we are trying to add and 2786 * the state that we are comming from. This function is only called 2787 * when we are adding new content to the cache. 2788 */ 2789static void 2790arc_adapt(int bytes, arc_state_t *state) 2791{ 2792 int mult; 2793 uint64_t arc_p_min = (arc_c >> arc_p_min_shift); 2794 2795 if (state == arc_l2c_only) 2796 return; 2797 2798 ASSERT(bytes > 0); 2799 /* 2800 * Adapt the target size of the MRU list: 2801 * - if we just hit in the MRU ghost list, then increase 2802 * the target size of the MRU list. 2803 * - if we just hit in the MFU ghost list, then increase 2804 * the target size of the MFU list by decreasing the 2805 * target size of the MRU list. 2806 */ 2807 if (state == arc_mru_ghost) { 2808 mult = ((arc_mru_ghost->arcs_size >= arc_mfu_ghost->arcs_size) ? 2809 1 : (arc_mfu_ghost->arcs_size/arc_mru_ghost->arcs_size)); 2810 mult = MIN(mult, 10); /* avoid wild arc_p adjustment */ 2811 2812 arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult); 2813 } else if (state == arc_mfu_ghost) { 2814 uint64_t delta; 2815 2816 mult = ((arc_mfu_ghost->arcs_size >= arc_mru_ghost->arcs_size) ? 2817 1 : (arc_mru_ghost->arcs_size/arc_mfu_ghost->arcs_size)); 2818 mult = MIN(mult, 10); 2819 2820 delta = MIN(bytes * mult, arc_p); 2821 arc_p = MAX(arc_p_min, arc_p - delta); 2822 } 2823 ASSERT((int64_t)arc_p >= 0); 2824 2825 if (arc_reclaim_needed()) { 2826 cv_signal(&arc_reclaim_thr_cv); 2827 return; 2828 } 2829 2830 if (arc_no_grow) 2831 return; 2832 2833 if (arc_c >= arc_c_max) 2834 return; 2835 2836 /* 2837 * If we're within (2 * maxblocksize) bytes of the target 2838 * cache size, increment the target cache size 2839 */ 2840 if (arc_size > arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) { 2841 DTRACE_PROBE1(arc__inc_adapt, int, bytes); 2842 atomic_add_64(&arc_c, (int64_t)bytes); 2843 if (arc_c > arc_c_max) 2844 arc_c = arc_c_max; 2845 else if (state == arc_anon) 2846 atomic_add_64(&arc_p, (int64_t)bytes); 2847 if (arc_p > arc_c) 2848 arc_p = arc_c; 2849 } 2850 ASSERT((int64_t)arc_p >= 0); 2851} 2852 2853/* 2854 * Check if the cache has reached its limits and eviction is required 2855 * prior to insert. 2856 */ 2857static int 2858arc_evict_needed(arc_buf_contents_t type) 2859{ 2860 if (type == ARC_BUFC_METADATA && arc_meta_used >= arc_meta_limit) 2861 return (1); 2862 2863 if (arc_reclaim_needed()) 2864 return (1); 2865 2866 return (arc_size > arc_c); 2867} 2868 2869/* 2870 * The buffer, supplied as the first argument, needs a data block. 2871 * So, if we are at cache max, determine which cache should be victimized. 2872 * We have the following cases: 2873 * 2874 * 1. Insert for MRU, p > sizeof(arc_anon + arc_mru) -> 2875 * In this situation if we're out of space, but the resident size of the MFU is 2876 * under the limit, victimize the MFU cache to satisfy this insertion request. 2877 * 2878 * 2. Insert for MRU, p <= sizeof(arc_anon + arc_mru) -> 2879 * Here, we've used up all of the available space for the MRU, so we need to 2880 * evict from our own cache instead. Evict from the set of resident MRU 2881 * entries. 2882 * 2883 * 3. Insert for MFU (c - p) > sizeof(arc_mfu) -> 2884 * c minus p represents the MFU space in the cache, since p is the size of the 2885 * cache that is dedicated to the MRU. In this situation there's still space on 2886 * the MFU side, so the MRU side needs to be victimized. 2887 * 2888 * 4. Insert for MFU (c - p) < sizeof(arc_mfu) -> 2889 * MFU's resident set is consuming more space than it has been allotted. In 2890 * this situation, we must victimize our own cache, the MFU, for this insertion. 2891 */ 2892static void 2893arc_get_data_buf(arc_buf_t *buf) 2894{ 2895 arc_state_t *state = buf->b_hdr->b_state; 2896 uint64_t size = buf->b_hdr->b_size; 2897 arc_buf_contents_t type = buf->b_hdr->b_type; 2898 2899 arc_adapt(size, state); 2900 2901 /* 2902 * We have not yet reached cache maximum size, 2903 * just allocate a new buffer. 2904 */ 2905 if (!arc_evict_needed(type)) { 2906 if (type == ARC_BUFC_METADATA) { 2907 buf->b_data = zio_buf_alloc(size); 2908 arc_space_consume(size, ARC_SPACE_DATA); 2909 } else { 2910 ASSERT(type == ARC_BUFC_DATA); 2911 buf->b_data = zio_data_buf_alloc(size); 2912 ARCSTAT_INCR(arcstat_data_size, size); 2913 atomic_add_64(&arc_size, size); 2914 } 2915 goto out; 2916 } 2917 2918 /* 2919 * If we are prefetching from the mfu ghost list, this buffer 2920 * will end up on the mru list; so steal space from there. 2921 */ 2922 if (state == arc_mfu_ghost) 2923 state = buf->b_hdr->b_flags & ARC_PREFETCH ? arc_mru : arc_mfu; 2924 else if (state == arc_mru_ghost) 2925 state = arc_mru; 2926 2927 if (state == arc_mru || state == arc_anon) { 2928 uint64_t mru_used = arc_anon->arcs_size + arc_mru->arcs_size; 2929 state = (arc_mfu->arcs_lsize[type] >= size && 2930 arc_p > mru_used) ? arc_mfu : arc_mru; 2931 } else { 2932 /* MFU cases */ 2933 uint64_t mfu_space = arc_c - arc_p; 2934 state = (arc_mru->arcs_lsize[type] >= size && 2935 mfu_space > arc_mfu->arcs_size) ? arc_mru : arc_mfu; 2936 } 2937 if ((buf->b_data = arc_evict(state, 0, size, TRUE, type)) == NULL) { 2938 if (type == ARC_BUFC_METADATA) { 2939 buf->b_data = zio_buf_alloc(size); 2940 arc_space_consume(size, ARC_SPACE_DATA); 2941 } else { 2942 ASSERT(type == ARC_BUFC_DATA); 2943 buf->b_data = zio_data_buf_alloc(size); 2944 ARCSTAT_INCR(arcstat_data_size, size); 2945 atomic_add_64(&arc_size, size); 2946 } 2947 ARCSTAT_BUMP(arcstat_recycle_miss); 2948 } 2949 ASSERT(buf->b_data != NULL); 2950out: 2951 /* 2952 * Update the state size. Note that ghost states have a 2953 * "ghost size" and so don't need to be updated. 2954 */ 2955 if (!GHOST_STATE(buf->b_hdr->b_state)) { 2956 arc_buf_hdr_t *hdr = buf->b_hdr; 2957 2958 atomic_add_64(&hdr->b_state->arcs_size, size); 2959 if (list_link_active(&hdr->b_arc_node)) { 2960 ASSERT(refcount_is_zero(&hdr->b_refcnt)); 2961 atomic_add_64(&hdr->b_state->arcs_lsize[type], size); 2962 } 2963 /* 2964 * If we are growing the cache, and we are adding anonymous 2965 * data, and we have outgrown arc_p, update arc_p 2966 */ 2967 if (arc_size < arc_c && hdr->b_state == arc_anon && 2968 arc_anon->arcs_size + arc_mru->arcs_size > arc_p) 2969 arc_p = MIN(arc_c, arc_p + size); 2970 } 2971 ARCSTAT_BUMP(arcstat_allocated); 2972} 2973 2974/* 2975 * This routine is called whenever a buffer is accessed. 2976 * NOTE: the hash lock is dropped in this function. 2977 */ 2978static void 2979arc_access(arc_buf_hdr_t *buf, kmutex_t *hash_lock) 2980{ 2981 clock_t now; 2982 2983 ASSERT(MUTEX_HELD(hash_lock)); 2984 2985 if (buf->b_state == arc_anon) { 2986 /* 2987 * This buffer is not in the cache, and does not 2988 * appear in our "ghost" list. Add the new buffer 2989 * to the MRU state. 2990 */ 2991 2992 ASSERT(buf->b_arc_access == 0); 2993 buf->b_arc_access = ddi_get_lbolt(); 2994 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, buf); 2995 arc_change_state(arc_mru, buf, hash_lock); 2996 2997 } else if (buf->b_state == arc_mru) { 2998 now = ddi_get_lbolt(); 2999 3000 /* 3001 * If this buffer is here because of a prefetch, then either: 3002 * - clear the flag if this is a "referencing" read 3003 * (any subsequent access will bump this into the MFU state). 3004 * or 3005 * - move the buffer to the head of the list if this is 3006 * another prefetch (to make it less likely to be evicted). 3007 */ 3008 if ((buf->b_flags & ARC_PREFETCH) != 0) { 3009 if (refcount_count(&buf->b_refcnt) == 0) { 3010 ASSERT(list_link_active(&buf->b_arc_node)); 3011 } else { 3012 buf->b_flags &= ~ARC_PREFETCH; 3013 ARCSTAT_BUMP(arcstat_mru_hits); 3014 } 3015 buf->b_arc_access = now; 3016 return; 3017 } 3018 3019 /* 3020 * This buffer has been "accessed" only once so far, 3021 * but it is still in the cache. Move it to the MFU 3022 * state. 3023 */ 3024 if (now > buf->b_arc_access + ARC_MINTIME) { 3025 /* 3026 * More than 125ms have passed since we 3027 * instantiated this buffer. Move it to the 3028 * most frequently used state. 3029 */ 3030 buf->b_arc_access = now; 3031 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf); 3032 arc_change_state(arc_mfu, buf, hash_lock); 3033 } 3034 ARCSTAT_BUMP(arcstat_mru_hits); 3035 } else if (buf->b_state == arc_mru_ghost) { 3036 arc_state_t *new_state; 3037 /* 3038 * This buffer has been "accessed" recently, but 3039 * was evicted from the cache. Move it to the 3040 * MFU state. 3041 */ 3042 3043 if (buf->b_flags & ARC_PREFETCH) { 3044 new_state = arc_mru; 3045 if (refcount_count(&buf->b_refcnt) > 0) 3046 buf->b_flags &= ~ARC_PREFETCH; 3047 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, buf); 3048 } else { 3049 new_state = arc_mfu; 3050 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf); 3051 } 3052 3053 buf->b_arc_access = ddi_get_lbolt(); 3054 arc_change_state(new_state, buf, hash_lock); 3055 3056 ARCSTAT_BUMP(arcstat_mru_ghost_hits); 3057 } else if (buf->b_state == arc_mfu) { 3058 /* 3059 * This buffer has been accessed more than once and is 3060 * still in the cache. Keep it in the MFU state. 3061 * 3062 * NOTE: an add_reference() that occurred when we did 3063 * the arc_read() will have kicked this off the list. 3064 * If it was a prefetch, we will explicitly move it to 3065 * the head of the list now. 3066 */ 3067 if ((buf->b_flags & ARC_PREFETCH) != 0) { 3068 ASSERT(refcount_count(&buf->b_refcnt) == 0); 3069 ASSERT(list_link_active(&buf->b_arc_node)); 3070 } 3071 ARCSTAT_BUMP(arcstat_mfu_hits); 3072 buf->b_arc_access = ddi_get_lbolt(); 3073 } else if (buf->b_state == arc_mfu_ghost) { 3074 arc_state_t *new_state = arc_mfu; 3075 /* 3076 * This buffer has been accessed more than once but has 3077 * been evicted from the cache. Move it back to the 3078 * MFU state. 3079 */ 3080 3081 if (buf->b_flags & ARC_PREFETCH) { 3082 /* 3083 * This is a prefetch access... 3084 * move this block back to the MRU state. 3085 */ 3086 ASSERT0(refcount_count(&buf->b_refcnt)); 3087 new_state = arc_mru; 3088 } 3089 3090 buf->b_arc_access = ddi_get_lbolt(); 3091 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf); 3092 arc_change_state(new_state, buf, hash_lock); 3093 3094 ARCSTAT_BUMP(arcstat_mfu_ghost_hits); 3095 } else if (buf->b_state == arc_l2c_only) { 3096 /* 3097 * This buffer is on the 2nd Level ARC. 3098 */ 3099 3100 buf->b_arc_access = ddi_get_lbolt(); 3101 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf); 3102 arc_change_state(arc_mfu, buf, hash_lock); 3103 } else { 3104 ASSERT(!"invalid arc state"); 3105 } 3106} 3107 3108/* a generic arc_done_func_t which you can use */ 3109/* ARGSUSED */ 3110void 3111arc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg) 3112{ 3113 if (zio == NULL || zio->io_error == 0) 3114 bcopy(buf->b_data, arg, buf->b_hdr->b_size); 3115 VERIFY(arc_buf_remove_ref(buf, arg)); 3116} 3117 3118/* a generic arc_done_func_t */ 3119void 3120arc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg) 3121{ 3122 arc_buf_t **bufp = arg; 3123 if (zio && zio->io_error) { 3124 VERIFY(arc_buf_remove_ref(buf, arg)); 3125 *bufp = NULL; 3126 } else { 3127 *bufp = buf; 3128 ASSERT(buf->b_data); 3129 } 3130} 3131 3132static void 3133arc_read_done(zio_t *zio) 3134{ 3135 arc_buf_hdr_t *hdr; 3136 arc_buf_t *buf; 3137 arc_buf_t *abuf; /* buffer we're assigning to callback */ 3138 kmutex_t *hash_lock = NULL; 3139 arc_callback_t *callback_list, *acb; 3140 int freeable = FALSE; 3141 3142 buf = zio->io_private; 3143 hdr = buf->b_hdr; 3144 3145 /* 3146 * The hdr was inserted into hash-table and removed from lists 3147 * prior to starting I/O. We should find this header, since 3148 * it's in the hash table, and it should be legit since it's 3149 * not possible to evict it during the I/O. The only possible 3150 * reason for it not to be found is if we were freed during the 3151 * read. 3152 */ 3153 if (HDR_IN_HASH_TABLE(hdr)) { 3154 ASSERT3U(hdr->b_birth, ==, BP_PHYSICAL_BIRTH(zio->io_bp)); 3155 ASSERT3U(hdr->b_dva.dva_word[0], ==, 3156 BP_IDENTITY(zio->io_bp)->dva_word[0]); 3157 ASSERT3U(hdr->b_dva.dva_word[1], ==, 3158 BP_IDENTITY(zio->io_bp)->dva_word[1]); 3159 3160 arc_buf_hdr_t *found = buf_hash_find(hdr->b_spa, zio->io_bp, 3161 &hash_lock); 3162 3163 ASSERT((found == NULL && HDR_FREED_IN_READ(hdr) && 3164 hash_lock == NULL) || 3165 (found == hdr && 3166 DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) || 3167 (found == hdr && HDR_L2_READING(hdr))); 3168 } 3169 3170 hdr->b_flags &= ~ARC_L2_EVICTED; 3171 if (l2arc_noprefetch && (hdr->b_flags & ARC_PREFETCH)) 3172 hdr->b_flags &= ~ARC_L2CACHE; 3173 3174 /* byteswap if necessary */ 3175 callback_list = hdr->b_acb; 3176 ASSERT(callback_list != NULL); 3177 if (BP_SHOULD_BYTESWAP(zio->io_bp) && zio->io_error == 0) { 3178 dmu_object_byteswap_t bswap = 3179 DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp)); 3180 arc_byteswap_func_t *func = BP_GET_LEVEL(zio->io_bp) > 0 ? 3181 byteswap_uint64_array : 3182 dmu_ot_byteswap[bswap].ob_func; 3183 func(buf->b_data, hdr->b_size); 3184 } 3185 3186 arc_cksum_compute(buf, B_FALSE); 3187#ifdef illumos 3188 arc_buf_watch(buf); 3189#endif /* illumos */ 3190 3191 if (hash_lock && zio->io_error == 0 && hdr->b_state == arc_anon) { 3192 /* 3193 * Only call arc_access on anonymous buffers. This is because 3194 * if we've issued an I/O for an evicted buffer, we've already 3195 * called arc_access (to prevent any simultaneous readers from 3196 * getting confused). 3197 */ 3198 arc_access(hdr, hash_lock); 3199 } 3200 3201 /* create copies of the data buffer for the callers */ 3202 abuf = buf; 3203 for (acb = callback_list; acb; acb = acb->acb_next) { 3204 if (acb->acb_done) { 3205 if (abuf == NULL) { 3206 ARCSTAT_BUMP(arcstat_duplicate_reads); 3207 abuf = arc_buf_clone(buf); 3208 } 3209 acb->acb_buf = abuf; 3210 abuf = NULL; 3211 } 3212 } 3213 hdr->b_acb = NULL; 3214 hdr->b_flags &= ~ARC_IO_IN_PROGRESS; 3215 ASSERT(!HDR_BUF_AVAILABLE(hdr)); 3216 if (abuf == buf) { 3217 ASSERT(buf->b_efunc == NULL); 3218 ASSERT(hdr->b_datacnt == 1); 3219 hdr->b_flags |= ARC_BUF_AVAILABLE; 3220 } 3221 3222 ASSERT(refcount_is_zero(&hdr->b_refcnt) || callback_list != NULL); 3223 3224 if (zio->io_error != 0) { 3225 hdr->b_flags |= ARC_IO_ERROR; 3226 if (hdr->b_state != arc_anon) 3227 arc_change_state(arc_anon, hdr, hash_lock); 3228 if (HDR_IN_HASH_TABLE(hdr)) 3229 buf_hash_remove(hdr); 3230 freeable = refcount_is_zero(&hdr->b_refcnt); 3231 } 3232 3233 /* 3234 * Broadcast before we drop the hash_lock to avoid the possibility 3235 * that the hdr (and hence the cv) might be freed before we get to 3236 * the cv_broadcast(). 3237 */ 3238 cv_broadcast(&hdr->b_cv); 3239 3240 if (hash_lock) { 3241 mutex_exit(hash_lock); 3242 } else { 3243 /* 3244 * This block was freed while we waited for the read to 3245 * complete. It has been removed from the hash table and 3246 * moved to the anonymous state (so that it won't show up 3247 * in the cache). 3248 */ 3249 ASSERT3P(hdr->b_state, ==, arc_anon); 3250 freeable = refcount_is_zero(&hdr->b_refcnt); 3251 } 3252 3253 /* execute each callback and free its structure */ 3254 while ((acb = callback_list) != NULL) { 3255 if (acb->acb_done) 3256 acb->acb_done(zio, acb->acb_buf, acb->acb_private); 3257 3258 if (acb->acb_zio_dummy != NULL) { 3259 acb->acb_zio_dummy->io_error = zio->io_error; 3260 zio_nowait(acb->acb_zio_dummy); 3261 } 3262 3263 callback_list = acb->acb_next; 3264 kmem_free(acb, sizeof (arc_callback_t)); 3265 } 3266 3267 if (freeable) 3268 arc_hdr_destroy(hdr); 3269} 3270 3271/* 3272 * "Read" the block block at the specified DVA (in bp) via the 3273 * cache. If the block is found in the cache, invoke the provided 3274 * callback immediately and return. Note that the `zio' parameter 3275 * in the callback will be NULL in this case, since no IO was 3276 * required. If the block is not in the cache pass the read request 3277 * on to the spa with a substitute callback function, so that the 3278 * requested block will be added to the cache. 3279 * 3280 * If a read request arrives for a block that has a read in-progress, 3281 * either wait for the in-progress read to complete (and return the 3282 * results); or, if this is a read with a "done" func, add a record 3283 * to the read to invoke the "done" func when the read completes, 3284 * and return; or just return. 3285 * 3286 * arc_read_done() will invoke all the requested "done" functions 3287 * for readers of this block. 3288 */ 3289int 3290arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_done_func_t *done, 3291 void *private, zio_priority_t priority, int zio_flags, uint32_t *arc_flags, 3292 const zbookmark_phys_t *zb) 3293{ 3294 arc_buf_hdr_t *hdr = NULL; 3295 arc_buf_t *buf = NULL; 3296 kmutex_t *hash_lock = NULL; 3297 zio_t *rzio; 3298 uint64_t guid = spa_load_guid(spa); 3299 3300 ASSERT(!BP_IS_EMBEDDED(bp) || 3301 BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA); 3302 3303top: 3304 if (!BP_IS_EMBEDDED(bp)) { 3305 /* 3306 * Embedded BP's have no DVA and require no I/O to "read". 3307 * Create an anonymous arc buf to back it. 3308 */ 3309 hdr = buf_hash_find(guid, bp, &hash_lock); 3310 } 3311 3312 if (hdr != NULL && hdr->b_datacnt > 0) { 3313 3314 *arc_flags |= ARC_CACHED; 3315 3316 if (HDR_IO_IN_PROGRESS(hdr)) { 3317 3318 if (*arc_flags & ARC_WAIT) { 3319 cv_wait(&hdr->b_cv, hash_lock); 3320 mutex_exit(hash_lock); 3321 goto top; 3322 } 3323 ASSERT(*arc_flags & ARC_NOWAIT); 3324 3325 if (done) { 3326 arc_callback_t *acb = NULL; 3327 3328 acb = kmem_zalloc(sizeof (arc_callback_t), 3329 KM_SLEEP); 3330 acb->acb_done = done; 3331 acb->acb_private = private; 3332 if (pio != NULL) 3333 acb->acb_zio_dummy = zio_null(pio, 3334 spa, NULL, NULL, NULL, zio_flags); 3335 3336 ASSERT(acb->acb_done != NULL); 3337 acb->acb_next = hdr->b_acb; 3338 hdr->b_acb = acb; 3339 add_reference(hdr, hash_lock, private); 3340 mutex_exit(hash_lock); 3341 return (0); 3342 } 3343 mutex_exit(hash_lock); 3344 return (0); 3345 } 3346 3347 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu); 3348 3349 if (done) { 3350 add_reference(hdr, hash_lock, private); 3351 /* 3352 * If this block is already in use, create a new 3353 * copy of the data so that we will be guaranteed 3354 * that arc_release() will always succeed. 3355 */ 3356 buf = hdr->b_buf; 3357 ASSERT(buf); 3358 ASSERT(buf->b_data); 3359 if (HDR_BUF_AVAILABLE(hdr)) { 3360 ASSERT(buf->b_efunc == NULL); 3361 hdr->b_flags &= ~ARC_BUF_AVAILABLE; 3362 } else { 3363 buf = arc_buf_clone(buf); 3364 } 3365 3366 } else if (*arc_flags & ARC_PREFETCH && 3367 refcount_count(&hdr->b_refcnt) == 0) { 3368 hdr->b_flags |= ARC_PREFETCH; 3369 } 3370 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr); 3371 arc_access(hdr, hash_lock); 3372 if (*arc_flags & ARC_L2CACHE) 3373 hdr->b_flags |= ARC_L2CACHE; 3374 if (*arc_flags & ARC_L2COMPRESS) 3375 hdr->b_flags |= ARC_L2COMPRESS; 3376 mutex_exit(hash_lock); 3377 ARCSTAT_BUMP(arcstat_hits); 3378 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH), 3379 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA, 3380 data, metadata, hits); 3381 3382 if (done) 3383 done(NULL, buf, private); 3384 } else { 3385 uint64_t size = BP_GET_LSIZE(bp); 3386 arc_callback_t *acb; 3387 vdev_t *vd = NULL; 3388 uint64_t addr = 0; 3389 boolean_t devw = B_FALSE; 3390 enum zio_compress b_compress = ZIO_COMPRESS_OFF; 3391 uint64_t b_asize = 0; 3392 3393 if (hdr == NULL) { 3394 /* this block is not in the cache */ 3395 arc_buf_hdr_t *exists = NULL; 3396 arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp); 3397 buf = arc_buf_alloc(spa, size, private, type); 3398 hdr = buf->b_hdr; 3399 if (!BP_IS_EMBEDDED(bp)) { 3400 hdr->b_dva = *BP_IDENTITY(bp); 3401 hdr->b_birth = BP_PHYSICAL_BIRTH(bp); 3402 hdr->b_cksum0 = bp->blk_cksum.zc_word[0]; 3403 exists = buf_hash_insert(hdr, &hash_lock); 3404 } 3405 if (exists != NULL) { 3406 /* somebody beat us to the hash insert */ 3407 mutex_exit(hash_lock); 3408 buf_discard_identity(hdr); 3409 (void) arc_buf_remove_ref(buf, private); 3410 goto top; /* restart the IO request */ 3411 } 3412 /* if this is a prefetch, we don't have a reference */ 3413 if (*arc_flags & ARC_PREFETCH) { 3414 (void) remove_reference(hdr, hash_lock, 3415 private); 3416 hdr->b_flags |= ARC_PREFETCH; 3417 } 3418 if (*arc_flags & ARC_L2CACHE) 3419 hdr->b_flags |= ARC_L2CACHE; 3420 if (*arc_flags & ARC_L2COMPRESS) 3421 hdr->b_flags |= ARC_L2COMPRESS; 3422 if (BP_GET_LEVEL(bp) > 0) 3423 hdr->b_flags |= ARC_INDIRECT; 3424 } else { 3425 /* this block is in the ghost cache */ 3426 ASSERT(GHOST_STATE(hdr->b_state)); 3427 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 3428 ASSERT0(refcount_count(&hdr->b_refcnt)); 3429 ASSERT(hdr->b_buf == NULL); 3430 3431 /* if this is a prefetch, we don't have a reference */ 3432 if (*arc_flags & ARC_PREFETCH) 3433 hdr->b_flags |= ARC_PREFETCH; 3434 else 3435 add_reference(hdr, hash_lock, private); 3436 if (*arc_flags & ARC_L2CACHE) 3437 hdr->b_flags |= ARC_L2CACHE; 3438 if (*arc_flags & ARC_L2COMPRESS) 3439 hdr->b_flags |= ARC_L2COMPRESS; 3440 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 3441 buf->b_hdr = hdr; 3442 buf->b_data = NULL; 3443 buf->b_efunc = NULL; 3444 buf->b_private = NULL; 3445 buf->b_next = NULL; 3446 hdr->b_buf = buf; 3447 ASSERT(hdr->b_datacnt == 0); 3448 hdr->b_datacnt = 1; 3449 arc_get_data_buf(buf); 3450 arc_access(hdr, hash_lock); 3451 } 3452 3453 ASSERT(!GHOST_STATE(hdr->b_state)); 3454 3455 acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP); 3456 acb->acb_done = done; 3457 acb->acb_private = private; 3458 3459 ASSERT(hdr->b_acb == NULL); 3460 hdr->b_acb = acb; 3461 hdr->b_flags |= ARC_IO_IN_PROGRESS; 3462 3463 if (hdr->b_l2hdr != NULL && 3464 (vd = hdr->b_l2hdr->b_dev->l2ad_vdev) != NULL) { 3465 devw = hdr->b_l2hdr->b_dev->l2ad_writing; 3466 addr = hdr->b_l2hdr->b_daddr; 3467 b_compress = hdr->b_l2hdr->b_compress; 3468 b_asize = hdr->b_l2hdr->b_asize; 3469 /* 3470 * Lock out device removal. 3471 */ 3472 if (vdev_is_dead(vd) || 3473 !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER)) 3474 vd = NULL; 3475 } 3476 3477 if (hash_lock != NULL) 3478 mutex_exit(hash_lock); 3479 3480 /* 3481 * At this point, we have a level 1 cache miss. Try again in 3482 * L2ARC if possible. 3483 */ 3484 ASSERT3U(hdr->b_size, ==, size); 3485 DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp, 3486 uint64_t, size, zbookmark_phys_t *, zb); 3487 ARCSTAT_BUMP(arcstat_misses); 3488 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH), 3489 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA, 3490 data, metadata, misses); 3491#ifdef _KERNEL 3492 curthread->td_ru.ru_inblock++; 3493#endif 3494 3495 if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) { 3496 /* 3497 * Read from the L2ARC if the following are true: 3498 * 1. The L2ARC vdev was previously cached. 3499 * 2. This buffer still has L2ARC metadata. 3500 * 3. This buffer isn't currently writing to the L2ARC. 3501 * 4. The L2ARC entry wasn't evicted, which may 3502 * also have invalidated the vdev. 3503 * 5. This isn't prefetch and l2arc_noprefetch is set. 3504 */ 3505 if (hdr->b_l2hdr != NULL && 3506 !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) && 3507 !(l2arc_noprefetch && HDR_PREFETCH(hdr))) { 3508 l2arc_read_callback_t *cb; 3509 3510 DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr); 3511 ARCSTAT_BUMP(arcstat_l2_hits); 3512 3513 cb = kmem_zalloc(sizeof (l2arc_read_callback_t), 3514 KM_SLEEP); 3515 cb->l2rcb_buf = buf; 3516 cb->l2rcb_spa = spa; 3517 cb->l2rcb_bp = *bp; 3518 cb->l2rcb_zb = *zb; 3519 cb->l2rcb_flags = zio_flags; 3520 cb->l2rcb_compress = b_compress; 3521 3522 ASSERT(addr >= VDEV_LABEL_START_SIZE && 3523 addr + size < vd->vdev_psize - 3524 VDEV_LABEL_END_SIZE); 3525 3526 /* 3527 * l2arc read. The SCL_L2ARC lock will be 3528 * released by l2arc_read_done(). 3529 * Issue a null zio if the underlying buffer 3530 * was squashed to zero size by compression. 3531 */ 3532 if (b_compress == ZIO_COMPRESS_EMPTY) { 3533 rzio = zio_null(pio, spa, vd, 3534 l2arc_read_done, cb, 3535 zio_flags | ZIO_FLAG_DONT_CACHE | 3536 ZIO_FLAG_CANFAIL | 3537 ZIO_FLAG_DONT_PROPAGATE | 3538 ZIO_FLAG_DONT_RETRY); 3539 } else { 3540 rzio = zio_read_phys(pio, vd, addr, 3541 b_asize, buf->b_data, 3542 ZIO_CHECKSUM_OFF, 3543 l2arc_read_done, cb, priority, 3544 zio_flags | ZIO_FLAG_DONT_CACHE | 3545 ZIO_FLAG_CANFAIL | 3546 ZIO_FLAG_DONT_PROPAGATE | 3547 ZIO_FLAG_DONT_RETRY, B_FALSE); 3548 } 3549 DTRACE_PROBE2(l2arc__read, vdev_t *, vd, 3550 zio_t *, rzio); 3551 ARCSTAT_INCR(arcstat_l2_read_bytes, b_asize); 3552 3553 if (*arc_flags & ARC_NOWAIT) { 3554 zio_nowait(rzio); 3555 return (0); 3556 } 3557 3558 ASSERT(*arc_flags & ARC_WAIT); 3559 if (zio_wait(rzio) == 0) 3560 return (0); 3561 3562 /* l2arc read error; goto zio_read() */ 3563 } else { 3564 DTRACE_PROBE1(l2arc__miss, 3565 arc_buf_hdr_t *, hdr); 3566 ARCSTAT_BUMP(arcstat_l2_misses); 3567 if (HDR_L2_WRITING(hdr)) 3568 ARCSTAT_BUMP(arcstat_l2_rw_clash); 3569 spa_config_exit(spa, SCL_L2ARC, vd); 3570 } 3571 } else { 3572 if (vd != NULL) 3573 spa_config_exit(spa, SCL_L2ARC, vd); 3574 if (l2arc_ndev != 0) { 3575 DTRACE_PROBE1(l2arc__miss, 3576 arc_buf_hdr_t *, hdr); 3577 ARCSTAT_BUMP(arcstat_l2_misses); 3578 } 3579 } 3580 3581 rzio = zio_read(pio, spa, bp, buf->b_data, size, 3582 arc_read_done, buf, priority, zio_flags, zb); 3583 3584 if (*arc_flags & ARC_WAIT) 3585 return (zio_wait(rzio)); 3586 3587 ASSERT(*arc_flags & ARC_NOWAIT); 3588 zio_nowait(rzio); 3589 } 3590 return (0); 3591} 3592 3593void 3594arc_set_callback(arc_buf_t *buf, arc_evict_func_t *func, void *private) 3595{ 3596 ASSERT(buf->b_hdr != NULL); 3597 ASSERT(buf->b_hdr->b_state != arc_anon); 3598 ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt) || func == NULL); 3599 ASSERT(buf->b_efunc == NULL); 3600 ASSERT(!HDR_BUF_AVAILABLE(buf->b_hdr)); 3601 3602 buf->b_efunc = func; 3603 buf->b_private = private; 3604} 3605 3606/* 3607 * Notify the arc that a block was freed, and thus will never be used again. 3608 */ 3609void 3610arc_freed(spa_t *spa, const blkptr_t *bp) 3611{ 3612 arc_buf_hdr_t *hdr; 3613 kmutex_t *hash_lock; 3614 uint64_t guid = spa_load_guid(spa); 3615 3616 ASSERT(!BP_IS_EMBEDDED(bp)); 3617 3618 hdr = buf_hash_find(guid, bp, &hash_lock); 3619 if (hdr == NULL) 3620 return; 3621 if (HDR_BUF_AVAILABLE(hdr)) { 3622 arc_buf_t *buf = hdr->b_buf; 3623 add_reference(hdr, hash_lock, FTAG); 3624 hdr->b_flags &= ~ARC_BUF_AVAILABLE; 3625 mutex_exit(hash_lock); 3626 3627 arc_release(buf, FTAG); 3628 (void) arc_buf_remove_ref(buf, FTAG); 3629 } else { 3630 mutex_exit(hash_lock); 3631 } 3632 3633} 3634 3635/* 3636 * Clear the user eviction callback set by arc_set_callback(), first calling 3637 * it if it exists. Because the presence of a callback keeps an arc_buf cached 3638 * clearing the callback may result in the arc_buf being destroyed. However, 3639 * it will not result in the *last* arc_buf being destroyed, hence the data 3640 * will remain cached in the ARC. We make a copy of the arc buffer here so 3641 * that we can process the callback without holding any locks. 3642 * 3643 * It's possible that the callback is already in the process of being cleared 3644 * by another thread. In this case we can not clear the callback. 3645 * 3646 * Returns B_TRUE if the callback was successfully called and cleared. 3647 */ 3648boolean_t 3649arc_clear_callback(arc_buf_t *buf) 3650{ 3651 arc_buf_hdr_t *hdr; 3652 kmutex_t *hash_lock; 3653 arc_evict_func_t *efunc = buf->b_efunc; 3654 void *private = buf->b_private; 3655 list_t *list, *evicted_list; 3656 kmutex_t *lock, *evicted_lock; 3657 3658 mutex_enter(&buf->b_evict_lock); 3659 hdr = buf->b_hdr; 3660 if (hdr == NULL) { 3661 /* 3662 * We are in arc_do_user_evicts(). 3663 */ 3664 ASSERT(buf->b_data == NULL); 3665 mutex_exit(&buf->b_evict_lock); 3666 return (B_FALSE); 3667 } else if (buf->b_data == NULL) { 3668 /* 3669 * We are on the eviction list; process this buffer now 3670 * but let arc_do_user_evicts() do the reaping. 3671 */ 3672 buf->b_efunc = NULL; 3673 mutex_exit(&buf->b_evict_lock); 3674 VERIFY0(efunc(private)); 3675 return (B_TRUE); 3676 } 3677 hash_lock = HDR_LOCK(hdr); 3678 mutex_enter(hash_lock); 3679 hdr = buf->b_hdr; 3680 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 3681 3682 ASSERT3U(refcount_count(&hdr->b_refcnt), <, hdr->b_datacnt); 3683 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu); 3684 3685 buf->b_efunc = NULL; 3686 buf->b_private = NULL; 3687 3688 if (hdr->b_datacnt > 1) { 3689 mutex_exit(&buf->b_evict_lock); 3690 arc_buf_destroy(buf, FALSE, TRUE); 3691 } else { 3692 ASSERT(buf == hdr->b_buf); 3693 hdr->b_flags |= ARC_BUF_AVAILABLE; 3694 mutex_exit(&buf->b_evict_lock); 3695 } 3696 3697 mutex_exit(hash_lock); 3698 VERIFY0(efunc(private)); 3699 return (B_TRUE); 3700} 3701 3702/* 3703 * Release this buffer from the cache, making it an anonymous buffer. This 3704 * must be done after a read and prior to modifying the buffer contents. 3705 * If the buffer has more than one reference, we must make 3706 * a new hdr for the buffer. 3707 */ 3708void 3709arc_release(arc_buf_t *buf, void *tag) 3710{ 3711 arc_buf_hdr_t *hdr; 3712 kmutex_t *hash_lock = NULL; 3713 l2arc_buf_hdr_t *l2hdr; 3714 uint64_t buf_size; 3715 3716 /* 3717 * It would be nice to assert that if it's DMU metadata (level > 3718 * 0 || it's the dnode file), then it must be syncing context. 3719 * But we don't know that information at this level. 3720 */ 3721 3722 mutex_enter(&buf->b_evict_lock); 3723 hdr = buf->b_hdr; 3724 3725 /* this buffer is not on any list */ 3726 ASSERT(refcount_count(&hdr->b_refcnt) > 0); 3727 3728 if (hdr->b_state == arc_anon) { 3729 /* this buffer is already released */ 3730 ASSERT(buf->b_efunc == NULL); 3731 } else { 3732 hash_lock = HDR_LOCK(hdr); 3733 mutex_enter(hash_lock); 3734 hdr = buf->b_hdr; 3735 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 3736 } 3737 3738 l2hdr = hdr->b_l2hdr; 3739 if (l2hdr) { 3740 mutex_enter(&l2arc_buflist_mtx); 3741 arc_buf_l2_cdata_free(hdr); 3742 hdr->b_l2hdr = NULL; 3743 list_remove(l2hdr->b_dev->l2ad_buflist, hdr); 3744 } 3745 buf_size = hdr->b_size; 3746 3747 /* 3748 * Do we have more than one buf? 3749 */ 3750 if (hdr->b_datacnt > 1) { 3751 arc_buf_hdr_t *nhdr; 3752 arc_buf_t **bufp; 3753 uint64_t blksz = hdr->b_size; 3754 uint64_t spa = hdr->b_spa; 3755 arc_buf_contents_t type = hdr->b_type; 3756 uint32_t flags = hdr->b_flags; 3757 3758 ASSERT(hdr->b_buf != buf || buf->b_next != NULL); 3759 /* 3760 * Pull the data off of this hdr and attach it to 3761 * a new anonymous hdr. 3762 */ 3763 (void) remove_reference(hdr, hash_lock, tag); 3764 bufp = &hdr->b_buf; 3765 while (*bufp != buf) 3766 bufp = &(*bufp)->b_next; 3767 *bufp = buf->b_next; 3768 buf->b_next = NULL; 3769 3770 ASSERT3U(hdr->b_state->arcs_size, >=, hdr->b_size); 3771 atomic_add_64(&hdr->b_state->arcs_size, -hdr->b_size); 3772 if (refcount_is_zero(&hdr->b_refcnt)) { 3773 uint64_t *size = &hdr->b_state->arcs_lsize[hdr->b_type]; 3774 ASSERT3U(*size, >=, hdr->b_size); 3775 atomic_add_64(size, -hdr->b_size); 3776 } 3777 3778 /* 3779 * We're releasing a duplicate user data buffer, update 3780 * our statistics accordingly. 3781 */ 3782 if (hdr->b_type == ARC_BUFC_DATA) { 3783 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers); 3784 ARCSTAT_INCR(arcstat_duplicate_buffers_size, 3785 -hdr->b_size); 3786 } 3787 hdr->b_datacnt -= 1; 3788 arc_cksum_verify(buf); 3789#ifdef illumos 3790 arc_buf_unwatch(buf); 3791#endif /* illumos */ 3792 3793 mutex_exit(hash_lock); 3794 3795 nhdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE); 3796 nhdr->b_size = blksz; 3797 nhdr->b_spa = spa; 3798 nhdr->b_type = type; 3799 nhdr->b_buf = buf; 3800 nhdr->b_state = arc_anon; 3801 nhdr->b_arc_access = 0; 3802 nhdr->b_flags = flags & ARC_L2_WRITING; 3803 nhdr->b_l2hdr = NULL; 3804 nhdr->b_datacnt = 1; 3805 nhdr->b_freeze_cksum = NULL; 3806 (void) refcount_add(&nhdr->b_refcnt, tag); 3807 buf->b_hdr = nhdr; 3808 mutex_exit(&buf->b_evict_lock); 3809 atomic_add_64(&arc_anon->arcs_size, blksz); 3810 } else { 3811 mutex_exit(&buf->b_evict_lock); 3812 ASSERT(refcount_count(&hdr->b_refcnt) == 1); 3813 ASSERT(!list_link_active(&hdr->b_arc_node)); 3814 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 3815 if (hdr->b_state != arc_anon) 3816 arc_change_state(arc_anon, hdr, hash_lock); 3817 hdr->b_arc_access = 0; 3818 if (hash_lock) 3819 mutex_exit(hash_lock); 3820 3821 buf_discard_identity(hdr); 3822 arc_buf_thaw(buf); 3823 } 3824 buf->b_efunc = NULL; 3825 buf->b_private = NULL; 3826 3827 if (l2hdr) { 3828 ARCSTAT_INCR(arcstat_l2_asize, -l2hdr->b_asize); 3829 vdev_space_update(l2hdr->b_dev->l2ad_vdev, 3830 -l2hdr->b_asize, 0, 0); 3831 trim_map_free(l2hdr->b_dev->l2ad_vdev, l2hdr->b_daddr, 3832 hdr->b_size, 0); 3833 kmem_free(l2hdr, sizeof (l2arc_buf_hdr_t)); 3834 ARCSTAT_INCR(arcstat_l2_size, -buf_size); 3835 mutex_exit(&l2arc_buflist_mtx); 3836 } 3837} 3838 3839int 3840arc_released(arc_buf_t *buf) 3841{ 3842 int released; 3843 3844 mutex_enter(&buf->b_evict_lock); 3845 released = (buf->b_data != NULL && buf->b_hdr->b_state == arc_anon); 3846 mutex_exit(&buf->b_evict_lock); 3847 return (released); 3848} 3849 3850#ifdef ZFS_DEBUG 3851int 3852arc_referenced(arc_buf_t *buf) 3853{ 3854 int referenced; 3855 3856 mutex_enter(&buf->b_evict_lock); 3857 referenced = (refcount_count(&buf->b_hdr->b_refcnt)); 3858 mutex_exit(&buf->b_evict_lock); 3859 return (referenced); 3860} 3861#endif 3862 3863static void 3864arc_write_ready(zio_t *zio) 3865{ 3866 arc_write_callback_t *callback = zio->io_private; 3867 arc_buf_t *buf = callback->awcb_buf; 3868 arc_buf_hdr_t *hdr = buf->b_hdr; 3869 3870 ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt)); 3871 callback->awcb_ready(zio, buf, callback->awcb_private); 3872 3873 /* 3874 * If the IO is already in progress, then this is a re-write 3875 * attempt, so we need to thaw and re-compute the cksum. 3876 * It is the responsibility of the callback to handle the 3877 * accounting for any re-write attempt. 3878 */ 3879 if (HDR_IO_IN_PROGRESS(hdr)) { 3880 mutex_enter(&hdr->b_freeze_lock); 3881 if (hdr->b_freeze_cksum != NULL) { 3882 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 3883 hdr->b_freeze_cksum = NULL; 3884 } 3885 mutex_exit(&hdr->b_freeze_lock); 3886 } 3887 arc_cksum_compute(buf, B_FALSE); 3888 hdr->b_flags |= ARC_IO_IN_PROGRESS; 3889} 3890 3891/* 3892 * The SPA calls this callback for each physical write that happens on behalf 3893 * of a logical write. See the comment in dbuf_write_physdone() for details. 3894 */ 3895static void 3896arc_write_physdone(zio_t *zio) 3897{ 3898 arc_write_callback_t *cb = zio->io_private; 3899 if (cb->awcb_physdone != NULL) 3900 cb->awcb_physdone(zio, cb->awcb_buf, cb->awcb_private); 3901} 3902 3903static void 3904arc_write_done(zio_t *zio) 3905{ 3906 arc_write_callback_t *callback = zio->io_private; 3907 arc_buf_t *buf = callback->awcb_buf; 3908 arc_buf_hdr_t *hdr = buf->b_hdr; 3909 3910 ASSERT(hdr->b_acb == NULL); 3911 3912 if (zio->io_error == 0) { 3913 if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) { 3914 buf_discard_identity(hdr); 3915 } else { 3916 hdr->b_dva = *BP_IDENTITY(zio->io_bp); 3917 hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp); 3918 hdr->b_cksum0 = zio->io_bp->blk_cksum.zc_word[0]; 3919 } 3920 } else { 3921 ASSERT(BUF_EMPTY(hdr)); 3922 } 3923 3924 /* 3925 * If the block to be written was all-zero or compressed enough to be 3926 * embedded in the BP, no write was performed so there will be no 3927 * dva/birth/checksum. The buffer must therefore remain anonymous 3928 * (and uncached). 3929 */ 3930 if (!BUF_EMPTY(hdr)) { 3931 arc_buf_hdr_t *exists; 3932 kmutex_t *hash_lock; 3933 3934 ASSERT(zio->io_error == 0); 3935 3936 arc_cksum_verify(buf); 3937 3938 exists = buf_hash_insert(hdr, &hash_lock); 3939 if (exists) { 3940 /* 3941 * This can only happen if we overwrite for 3942 * sync-to-convergence, because we remove 3943 * buffers from the hash table when we arc_free(). 3944 */ 3945 if (zio->io_flags & ZIO_FLAG_IO_REWRITE) { 3946 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) 3947 panic("bad overwrite, hdr=%p exists=%p", 3948 (void *)hdr, (void *)exists); 3949 ASSERT(refcount_is_zero(&exists->b_refcnt)); 3950 arc_change_state(arc_anon, exists, hash_lock); 3951 mutex_exit(hash_lock); 3952 arc_hdr_destroy(exists); 3953 exists = buf_hash_insert(hdr, &hash_lock); 3954 ASSERT3P(exists, ==, NULL); 3955 } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) { 3956 /* nopwrite */ 3957 ASSERT(zio->io_prop.zp_nopwrite); 3958 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) 3959 panic("bad nopwrite, hdr=%p exists=%p", 3960 (void *)hdr, (void *)exists); 3961 } else { 3962 /* Dedup */ 3963 ASSERT(hdr->b_datacnt == 1); 3964 ASSERT(hdr->b_state == arc_anon); 3965 ASSERT(BP_GET_DEDUP(zio->io_bp)); 3966 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0); 3967 } 3968 } 3969 hdr->b_flags &= ~ARC_IO_IN_PROGRESS; 3970 /* if it's not anon, we are doing a scrub */ 3971 if (!exists && hdr->b_state == arc_anon) 3972 arc_access(hdr, hash_lock); 3973 mutex_exit(hash_lock); 3974 } else { 3975 hdr->b_flags &= ~ARC_IO_IN_PROGRESS; 3976 } 3977 3978 ASSERT(!refcount_is_zero(&hdr->b_refcnt)); 3979 callback->awcb_done(zio, buf, callback->awcb_private); 3980 3981 kmem_free(callback, sizeof (arc_write_callback_t)); 3982} 3983 3984zio_t * 3985arc_write(zio_t *pio, spa_t *spa, uint64_t txg, 3986 blkptr_t *bp, arc_buf_t *buf, boolean_t l2arc, boolean_t l2arc_compress, 3987 const zio_prop_t *zp, arc_done_func_t *ready, arc_done_func_t *physdone, 3988 arc_done_func_t *done, void *private, zio_priority_t priority, 3989 int zio_flags, const zbookmark_phys_t *zb) 3990{ 3991 arc_buf_hdr_t *hdr = buf->b_hdr; 3992 arc_write_callback_t *callback; 3993 zio_t *zio; 3994 3995 ASSERT(ready != NULL); 3996 ASSERT(done != NULL); 3997 ASSERT(!HDR_IO_ERROR(hdr)); 3998 ASSERT((hdr->b_flags & ARC_IO_IN_PROGRESS) == 0); 3999 ASSERT(hdr->b_acb == NULL); 4000 if (l2arc) 4001 hdr->b_flags |= ARC_L2CACHE; 4002 if (l2arc_compress) 4003 hdr->b_flags |= ARC_L2COMPRESS; 4004 callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP); 4005 callback->awcb_ready = ready; 4006 callback->awcb_physdone = physdone; 4007 callback->awcb_done = done; 4008 callback->awcb_private = private; 4009 callback->awcb_buf = buf; 4010 4011 zio = zio_write(pio, spa, txg, bp, buf->b_data, hdr->b_size, zp, 4012 arc_write_ready, arc_write_physdone, arc_write_done, callback, 4013 priority, zio_flags, zb); 4014 4015 return (zio); 4016} 4017 4018static int 4019arc_memory_throttle(uint64_t reserve, uint64_t txg) 4020{ 4021#ifdef _KERNEL 4022 uint64_t available_memory = ptob(freemem); 4023 static uint64_t page_load = 0; 4024 static uint64_t last_txg = 0; 4025 4026#if defined(__i386) || !defined(UMA_MD_SMALL_ALLOC) 4027 available_memory = 4028 MIN(available_memory, ptob(vmem_size(heap_arena, VMEM_FREE))); 4029#endif 4030 4031 if (freemem > (uint64_t)physmem * arc_lotsfree_percent / 100) 4032 return (0); 4033 4034 if (txg > last_txg) { 4035 last_txg = txg; 4036 page_load = 0; 4037 } 4038 /* 4039 * If we are in pageout, we know that memory is already tight, 4040 * the arc is already going to be evicting, so we just want to 4041 * continue to let page writes occur as quickly as possible. 4042 */ 4043 if (curproc == pageproc) { 4044 if (page_load > MAX(ptob(minfree), available_memory) / 4) 4045 return (SET_ERROR(ERESTART)); 4046 /* Note: reserve is inflated, so we deflate */ 4047 page_load += reserve / 8; 4048 return (0); 4049 } else if (page_load > 0 && arc_reclaim_needed()) { 4050 /* memory is low, delay before restarting */ 4051 ARCSTAT_INCR(arcstat_memory_throttle_count, 1); 4052 return (SET_ERROR(EAGAIN)); 4053 } 4054 page_load = 0; 4055#endif 4056 return (0); 4057} 4058 4059void 4060arc_tempreserve_clear(uint64_t reserve) 4061{ 4062 atomic_add_64(&arc_tempreserve, -reserve); 4063 ASSERT((int64_t)arc_tempreserve >= 0); 4064} 4065 4066int 4067arc_tempreserve_space(uint64_t reserve, uint64_t txg) 4068{ 4069 int error; 4070 uint64_t anon_size; 4071 4072 if (reserve > arc_c/4 && !arc_no_grow) { 4073 arc_c = MIN(arc_c_max, reserve * 4); 4074 DTRACE_PROBE1(arc__set_reserve, uint64_t, arc_c); 4075 } 4076 if (reserve > arc_c) 4077 return (SET_ERROR(ENOMEM)); 4078 4079 /* 4080 * Don't count loaned bufs as in flight dirty data to prevent long 4081 * network delays from blocking transactions that are ready to be 4082 * assigned to a txg. 4083 */ 4084 anon_size = MAX((int64_t)(arc_anon->arcs_size - arc_loaned_bytes), 0); 4085 4086 /* 4087 * Writes will, almost always, require additional memory allocations 4088 * in order to compress/encrypt/etc the data. We therefore need to 4089 * make sure that there is sufficient available memory for this. 4090 */ 4091 error = arc_memory_throttle(reserve, txg); 4092 if (error != 0) 4093 return (error); 4094 4095 /* 4096 * Throttle writes when the amount of dirty data in the cache 4097 * gets too large. We try to keep the cache less than half full 4098 * of dirty blocks so that our sync times don't grow too large. 4099 * Note: if two requests come in concurrently, we might let them 4100 * both succeed, when one of them should fail. Not a huge deal. 4101 */ 4102 4103 if (reserve + arc_tempreserve + anon_size > arc_c / 2 && 4104 anon_size > arc_c / 4) { 4105 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK " 4106 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n", 4107 arc_tempreserve>>10, 4108 arc_anon->arcs_lsize[ARC_BUFC_METADATA]>>10, 4109 arc_anon->arcs_lsize[ARC_BUFC_DATA]>>10, 4110 reserve>>10, arc_c>>10); 4111 return (SET_ERROR(ERESTART)); 4112 } 4113 atomic_add_64(&arc_tempreserve, reserve); 4114 return (0); 4115} 4116 4117static kmutex_t arc_lowmem_lock; 4118#ifdef _KERNEL 4119static eventhandler_tag arc_event_lowmem = NULL; 4120 4121static void 4122arc_lowmem(void *arg __unused, int howto __unused) 4123{ 4124 4125 /* Serialize access via arc_lowmem_lock. */ 4126 mutex_enter(&arc_lowmem_lock); 4127 mutex_enter(&arc_reclaim_thr_lock); 4128 needfree = 1; 4129 DTRACE_PROBE(arc__needfree); 4130 cv_signal(&arc_reclaim_thr_cv); 4131 4132 /* 4133 * It is unsafe to block here in arbitrary threads, because we can come 4134 * here from ARC itself and may hold ARC locks and thus risk a deadlock 4135 * with ARC reclaim thread. 4136 */ 4137 if (curproc == pageproc) { 4138 while (needfree) 4139 msleep(&needfree, &arc_reclaim_thr_lock, 0, "zfs:lowmem", 0); 4140 } 4141 mutex_exit(&arc_reclaim_thr_lock); 4142 mutex_exit(&arc_lowmem_lock); 4143} 4144#endif 4145 4146void 4147arc_init(void) 4148{ 4149 int i, prefetch_tunable_set = 0; 4150 4151 mutex_init(&arc_reclaim_thr_lock, NULL, MUTEX_DEFAULT, NULL); 4152 cv_init(&arc_reclaim_thr_cv, NULL, CV_DEFAULT, NULL); 4153 mutex_init(&arc_lowmem_lock, NULL, MUTEX_DEFAULT, NULL); 4154 4155 /* Convert seconds to clock ticks */ 4156 arc_min_prefetch_lifespan = 1 * hz; 4157 4158 /* Start out with 1/8 of all memory */ 4159 arc_c = kmem_size() / 8; 4160 4161#ifdef sun 4162#ifdef _KERNEL 4163 /* 4164 * On architectures where the physical memory can be larger 4165 * than the addressable space (intel in 32-bit mode), we may 4166 * need to limit the cache to 1/8 of VM size. 4167 */ 4168 arc_c = MIN(arc_c, vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 8); 4169#endif 4170#endif /* sun */ 4171 /* set min cache to 1/32 of all memory, or 16MB, whichever is more */ 4172 arc_c_min = MAX(arc_c / 4, 64<<18); 4173 /* set max to 1/2 of all memory, or all but 1GB, whichever is more */ 4174 if (arc_c * 8 >= 1<<30) 4175 arc_c_max = (arc_c * 8) - (1<<30); 4176 else 4177 arc_c_max = arc_c_min; 4178 arc_c_max = MAX(arc_c * 5, arc_c_max); 4179 4180#ifdef _KERNEL 4181 /* 4182 * Allow the tunables to override our calculations if they are 4183 * reasonable (ie. over 16MB) 4184 */ 4185 if (zfs_arc_max > 64<<18 && zfs_arc_max < kmem_size()) 4186 arc_c_max = zfs_arc_max; 4187 if (zfs_arc_min > 64<<18 && zfs_arc_min <= arc_c_max) 4188 arc_c_min = zfs_arc_min; 4189#endif 4190 4191 arc_c = arc_c_max; 4192 arc_p = (arc_c >> 1); 4193 4194 /* limit meta-data to 1/4 of the arc capacity */ 4195 arc_meta_limit = arc_c_max / 4; 4196 4197 /* Allow the tunable to override if it is reasonable */ 4198 if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max) 4199 arc_meta_limit = zfs_arc_meta_limit; 4200 4201 if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0) 4202 arc_c_min = arc_meta_limit / 2; 4203 4204 if (zfs_arc_grow_retry > 0) 4205 arc_grow_retry = zfs_arc_grow_retry; 4206 4207 if (zfs_arc_shrink_shift > 0) 4208 arc_shrink_shift = zfs_arc_shrink_shift; 4209 4210 if (zfs_arc_p_min_shift > 0) 4211 arc_p_min_shift = zfs_arc_p_min_shift; 4212 4213 /* if kmem_flags are set, lets try to use less memory */ 4214 if (kmem_debugging()) 4215 arc_c = arc_c / 2; 4216 if (arc_c < arc_c_min) 4217 arc_c = arc_c_min; 4218 4219 zfs_arc_min = arc_c_min; 4220 zfs_arc_max = arc_c_max; 4221 4222 arc_anon = &ARC_anon; 4223 arc_mru = &ARC_mru; 4224 arc_mru_ghost = &ARC_mru_ghost; 4225 arc_mfu = &ARC_mfu; 4226 arc_mfu_ghost = &ARC_mfu_ghost; 4227 arc_l2c_only = &ARC_l2c_only; 4228 arc_size = 0; 4229 4230 for (i = 0; i < ARC_BUFC_NUMLISTS; i++) { 4231 mutex_init(&arc_anon->arcs_locks[i].arcs_lock, 4232 NULL, MUTEX_DEFAULT, NULL); 4233 mutex_init(&arc_mru->arcs_locks[i].arcs_lock, 4234 NULL, MUTEX_DEFAULT, NULL); 4235 mutex_init(&arc_mru_ghost->arcs_locks[i].arcs_lock, 4236 NULL, MUTEX_DEFAULT, NULL); 4237 mutex_init(&arc_mfu->arcs_locks[i].arcs_lock, 4238 NULL, MUTEX_DEFAULT, NULL); 4239 mutex_init(&arc_mfu_ghost->arcs_locks[i].arcs_lock, 4240 NULL, MUTEX_DEFAULT, NULL); 4241 mutex_init(&arc_l2c_only->arcs_locks[i].arcs_lock, 4242 NULL, MUTEX_DEFAULT, NULL); 4243 4244 list_create(&arc_mru->arcs_lists[i], 4245 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4246 list_create(&arc_mru_ghost->arcs_lists[i], 4247 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4248 list_create(&arc_mfu->arcs_lists[i], 4249 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4250 list_create(&arc_mfu_ghost->arcs_lists[i], 4251 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4252 list_create(&arc_mfu_ghost->arcs_lists[i], 4253 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4254 list_create(&arc_l2c_only->arcs_lists[i], 4255 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4256 } 4257 4258 buf_init(); 4259 4260 arc_thread_exit = 0; 4261 arc_eviction_list = NULL; 4262 mutex_init(&arc_eviction_mtx, NULL, MUTEX_DEFAULT, NULL); 4263 bzero(&arc_eviction_hdr, sizeof (arc_buf_hdr_t)); 4264 4265 arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED, 4266 sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL); 4267 4268 if (arc_ksp != NULL) { 4269 arc_ksp->ks_data = &arc_stats; 4270 kstat_install(arc_ksp); 4271 } 4272 4273 (void) thread_create(NULL, 0, arc_reclaim_thread, NULL, 0, &p0, 4274 TS_RUN, minclsyspri); 4275 4276#ifdef _KERNEL 4277 arc_event_lowmem = EVENTHANDLER_REGISTER(vm_lowmem, arc_lowmem, NULL, 4278 EVENTHANDLER_PRI_FIRST); 4279#endif 4280 4281 arc_dead = FALSE; 4282 arc_warm = B_FALSE; 4283 4284 /* 4285 * Calculate maximum amount of dirty data per pool. 4286 * 4287 * If it has been set by /etc/system, take that. 4288 * Otherwise, use a percentage of physical memory defined by 4289 * zfs_dirty_data_max_percent (default 10%) with a cap at 4290 * zfs_dirty_data_max_max (default 4GB). 4291 */ 4292 if (zfs_dirty_data_max == 0) { 4293 zfs_dirty_data_max = ptob(physmem) * 4294 zfs_dirty_data_max_percent / 100; 4295 zfs_dirty_data_max = MIN(zfs_dirty_data_max, 4296 zfs_dirty_data_max_max); 4297 } 4298 4299#ifdef _KERNEL 4300 if (TUNABLE_INT_FETCH("vfs.zfs.prefetch_disable", &zfs_prefetch_disable)) 4301 prefetch_tunable_set = 1; 4302 4303#ifdef __i386__ 4304 if (prefetch_tunable_set == 0) { 4305 printf("ZFS NOTICE: Prefetch is disabled by default on i386 " 4306 "-- to enable,\n"); 4307 printf(" add \"vfs.zfs.prefetch_disable=0\" " 4308 "to /boot/loader.conf.\n"); 4309 zfs_prefetch_disable = 1; 4310 } 4311#else 4312 if ((((uint64_t)physmem * PAGESIZE) < (1ULL << 32)) && 4313 prefetch_tunable_set == 0) { 4314 printf("ZFS NOTICE: Prefetch is disabled by default if less " 4315 "than 4GB of RAM is present;\n" 4316 " to enable, add \"vfs.zfs.prefetch_disable=0\" " 4317 "to /boot/loader.conf.\n"); 4318 zfs_prefetch_disable = 1; 4319 } 4320#endif 4321 /* Warn about ZFS memory and address space requirements. */ 4322 if (((uint64_t)physmem * PAGESIZE) < (256 + 128 + 64) * (1 << 20)) { 4323 printf("ZFS WARNING: Recommended minimum RAM size is 512MB; " 4324 "expect unstable behavior.\n"); 4325 } 4326 if (kmem_size() < 512 * (1 << 20)) { 4327 printf("ZFS WARNING: Recommended minimum kmem_size is 512MB; " 4328 "expect unstable behavior.\n"); 4329 printf(" Consider tuning vm.kmem_size and " 4330 "vm.kmem_size_max\n"); 4331 printf(" in /boot/loader.conf.\n"); 4332 } 4333#endif 4334} 4335 4336void 4337arc_fini(void) 4338{ 4339 int i; 4340 4341 mutex_enter(&arc_reclaim_thr_lock); 4342 arc_thread_exit = 1; 4343 cv_signal(&arc_reclaim_thr_cv); 4344 while (arc_thread_exit != 0) 4345 cv_wait(&arc_reclaim_thr_cv, &arc_reclaim_thr_lock); 4346 mutex_exit(&arc_reclaim_thr_lock); 4347 4348 arc_flush(NULL); 4349 4350 arc_dead = TRUE; 4351 4352 if (arc_ksp != NULL) { 4353 kstat_delete(arc_ksp); 4354 arc_ksp = NULL; 4355 } 4356 4357 mutex_destroy(&arc_eviction_mtx); 4358 mutex_destroy(&arc_reclaim_thr_lock); 4359 cv_destroy(&arc_reclaim_thr_cv); 4360 4361 for (i = 0; i < ARC_BUFC_NUMLISTS; i++) { 4362 list_destroy(&arc_mru->arcs_lists[i]); 4363 list_destroy(&arc_mru_ghost->arcs_lists[i]); 4364 list_destroy(&arc_mfu->arcs_lists[i]); 4365 list_destroy(&arc_mfu_ghost->arcs_lists[i]); 4366 list_destroy(&arc_l2c_only->arcs_lists[i]); 4367 4368 mutex_destroy(&arc_anon->arcs_locks[i].arcs_lock); 4369 mutex_destroy(&arc_mru->arcs_locks[i].arcs_lock); 4370 mutex_destroy(&arc_mru_ghost->arcs_locks[i].arcs_lock); 4371 mutex_destroy(&arc_mfu->arcs_locks[i].arcs_lock); 4372 mutex_destroy(&arc_mfu_ghost->arcs_locks[i].arcs_lock); 4373 mutex_destroy(&arc_l2c_only->arcs_locks[i].arcs_lock); 4374 } 4375 4376 buf_fini(); 4377 4378 ASSERT(arc_loaned_bytes == 0); 4379 4380 mutex_destroy(&arc_lowmem_lock); 4381#ifdef _KERNEL 4382 if (arc_event_lowmem != NULL) 4383 EVENTHANDLER_DEREGISTER(vm_lowmem, arc_event_lowmem); 4384#endif 4385} 4386 4387/* 4388 * Level 2 ARC 4389 * 4390 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk. 4391 * It uses dedicated storage devices to hold cached data, which are populated 4392 * using large infrequent writes. The main role of this cache is to boost 4393 * the performance of random read workloads. The intended L2ARC devices 4394 * include short-stroked disks, solid state disks, and other media with 4395 * substantially faster read latency than disk. 4396 * 4397 * +-----------------------+ 4398 * | ARC | 4399 * +-----------------------+ 4400 * | ^ ^ 4401 * | | | 4402 * l2arc_feed_thread() arc_read() 4403 * | | | 4404 * | l2arc read | 4405 * V | | 4406 * +---------------+ | 4407 * | L2ARC | | 4408 * +---------------+ | 4409 * | ^ | 4410 * l2arc_write() | | 4411 * | | | 4412 * V | | 4413 * +-------+ +-------+ 4414 * | vdev | | vdev | 4415 * | cache | | cache | 4416 * +-------+ +-------+ 4417 * +=========+ .-----. 4418 * : L2ARC : |-_____-| 4419 * : devices : | Disks | 4420 * +=========+ `-_____-' 4421 * 4422 * Read requests are satisfied from the following sources, in order: 4423 * 4424 * 1) ARC 4425 * 2) vdev cache of L2ARC devices 4426 * 3) L2ARC devices 4427 * 4) vdev cache of disks 4428 * 5) disks 4429 * 4430 * Some L2ARC device types exhibit extremely slow write performance. 4431 * To accommodate for this there are some significant differences between 4432 * the L2ARC and traditional cache design: 4433 * 4434 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from 4435 * the ARC behave as usual, freeing buffers and placing headers on ghost 4436 * lists. The ARC does not send buffers to the L2ARC during eviction as 4437 * this would add inflated write latencies for all ARC memory pressure. 4438 * 4439 * 2. The L2ARC attempts to cache data from the ARC before it is evicted. 4440 * It does this by periodically scanning buffers from the eviction-end of 4441 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are 4442 * not already there. It scans until a headroom of buffers is satisfied, 4443 * which itself is a buffer for ARC eviction. If a compressible buffer is 4444 * found during scanning and selected for writing to an L2ARC device, we 4445 * temporarily boost scanning headroom during the next scan cycle to make 4446 * sure we adapt to compression effects (which might significantly reduce 4447 * the data volume we write to L2ARC). The thread that does this is 4448 * l2arc_feed_thread(), illustrated below; example sizes are included to 4449 * provide a better sense of ratio than this diagram: 4450 * 4451 * head --> tail 4452 * +---------------------+----------+ 4453 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC 4454 * +---------------------+----------+ | o L2ARC eligible 4455 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer 4456 * +---------------------+----------+ | 4457 * 15.9 Gbytes ^ 32 Mbytes | 4458 * headroom | 4459 * l2arc_feed_thread() 4460 * | 4461 * l2arc write hand <--[oooo]--' 4462 * | 8 Mbyte 4463 * | write max 4464 * V 4465 * +==============================+ 4466 * L2ARC dev |####|#|###|###| |####| ... | 4467 * +==============================+ 4468 * 32 Gbytes 4469 * 4470 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of 4471 * evicted, then the L2ARC has cached a buffer much sooner than it probably 4472 * needed to, potentially wasting L2ARC device bandwidth and storage. It is 4473 * safe to say that this is an uncommon case, since buffers at the end of 4474 * the ARC lists have moved there due to inactivity. 4475 * 4476 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom, 4477 * then the L2ARC simply misses copying some buffers. This serves as a 4478 * pressure valve to prevent heavy read workloads from both stalling the ARC 4479 * with waits and clogging the L2ARC with writes. This also helps prevent 4480 * the potential for the L2ARC to churn if it attempts to cache content too 4481 * quickly, such as during backups of the entire pool. 4482 * 4483 * 5. After system boot and before the ARC has filled main memory, there are 4484 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru 4485 * lists can remain mostly static. Instead of searching from tail of these 4486 * lists as pictured, the l2arc_feed_thread() will search from the list heads 4487 * for eligible buffers, greatly increasing its chance of finding them. 4488 * 4489 * The L2ARC device write speed is also boosted during this time so that 4490 * the L2ARC warms up faster. Since there have been no ARC evictions yet, 4491 * there are no L2ARC reads, and no fear of degrading read performance 4492 * through increased writes. 4493 * 4494 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that 4495 * the vdev queue can aggregate them into larger and fewer writes. Each 4496 * device is written to in a rotor fashion, sweeping writes through 4497 * available space then repeating. 4498 * 4499 * 7. The L2ARC does not store dirty content. It never needs to flush 4500 * write buffers back to disk based storage. 4501 * 4502 * 8. If an ARC buffer is written (and dirtied) which also exists in the 4503 * L2ARC, the now stale L2ARC buffer is immediately dropped. 4504 * 4505 * The performance of the L2ARC can be tweaked by a number of tunables, which 4506 * may be necessary for different workloads: 4507 * 4508 * l2arc_write_max max write bytes per interval 4509 * l2arc_write_boost extra write bytes during device warmup 4510 * l2arc_noprefetch skip caching prefetched buffers 4511 * l2arc_headroom number of max device writes to precache 4512 * l2arc_headroom_boost when we find compressed buffers during ARC 4513 * scanning, we multiply headroom by this 4514 * percentage factor for the next scan cycle, 4515 * since more compressed buffers are likely to 4516 * be present 4517 * l2arc_feed_secs seconds between L2ARC writing 4518 * 4519 * Tunables may be removed or added as future performance improvements are 4520 * integrated, and also may become zpool properties. 4521 * 4522 * There are three key functions that control how the L2ARC warms up: 4523 * 4524 * l2arc_write_eligible() check if a buffer is eligible to cache 4525 * l2arc_write_size() calculate how much to write 4526 * l2arc_write_interval() calculate sleep delay between writes 4527 * 4528 * These three functions determine what to write, how much, and how quickly 4529 * to send writes. 4530 */ 4531 4532static boolean_t 4533l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *ab) 4534{ 4535 /* 4536 * A buffer is *not* eligible for the L2ARC if it: 4537 * 1. belongs to a different spa. 4538 * 2. is already cached on the L2ARC. 4539 * 3. has an I/O in progress (it may be an incomplete read). 4540 * 4. is flagged not eligible (zfs property). 4541 */ 4542 if (ab->b_spa != spa_guid) { 4543 ARCSTAT_BUMP(arcstat_l2_write_spa_mismatch); 4544 return (B_FALSE); 4545 } 4546 if (ab->b_l2hdr != NULL) { 4547 ARCSTAT_BUMP(arcstat_l2_write_in_l2); 4548 return (B_FALSE); 4549 } 4550 if (HDR_IO_IN_PROGRESS(ab)) { 4551 ARCSTAT_BUMP(arcstat_l2_write_hdr_io_in_progress); 4552 return (B_FALSE); 4553 } 4554 if (!HDR_L2CACHE(ab)) { 4555 ARCSTAT_BUMP(arcstat_l2_write_not_cacheable); 4556 return (B_FALSE); 4557 } 4558 4559 return (B_TRUE); 4560} 4561 4562static uint64_t 4563l2arc_write_size(void) 4564{ 4565 uint64_t size; 4566 4567 /* 4568 * Make sure our globals have meaningful values in case the user 4569 * altered them. 4570 */ 4571 size = l2arc_write_max; 4572 if (size == 0) { 4573 cmn_err(CE_NOTE, "Bad value for l2arc_write_max, value must " 4574 "be greater than zero, resetting it to the default (%d)", 4575 L2ARC_WRITE_SIZE); 4576 size = l2arc_write_max = L2ARC_WRITE_SIZE; 4577 } 4578 4579 if (arc_warm == B_FALSE) 4580 size += l2arc_write_boost; 4581 4582 return (size); 4583 4584} 4585 4586static clock_t 4587l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote) 4588{ 4589 clock_t interval, next, now; 4590 4591 /* 4592 * If the ARC lists are busy, increase our write rate; if the 4593 * lists are stale, idle back. This is achieved by checking 4594 * how much we previously wrote - if it was more than half of 4595 * what we wanted, schedule the next write much sooner. 4596 */ 4597 if (l2arc_feed_again && wrote > (wanted / 2)) 4598 interval = (hz * l2arc_feed_min_ms) / 1000; 4599 else 4600 interval = hz * l2arc_feed_secs; 4601 4602 now = ddi_get_lbolt(); 4603 next = MAX(now, MIN(now + interval, began + interval)); 4604 4605 return (next); 4606} 4607 4608static void 4609l2arc_hdr_stat_add(void) 4610{ 4611 ARCSTAT_INCR(arcstat_l2_hdr_size, HDR_SIZE + L2HDR_SIZE); 4612 ARCSTAT_INCR(arcstat_hdr_size, -HDR_SIZE); 4613} 4614 4615static void 4616l2arc_hdr_stat_remove(void) 4617{ 4618 ARCSTAT_INCR(arcstat_l2_hdr_size, -(HDR_SIZE + L2HDR_SIZE)); 4619 ARCSTAT_INCR(arcstat_hdr_size, HDR_SIZE); 4620} 4621 4622/* 4623 * Cycle through L2ARC devices. This is how L2ARC load balances. 4624 * If a device is returned, this also returns holding the spa config lock. 4625 */ 4626static l2arc_dev_t * 4627l2arc_dev_get_next(void) 4628{ 4629 l2arc_dev_t *first, *next = NULL; 4630 4631 /* 4632 * Lock out the removal of spas (spa_namespace_lock), then removal 4633 * of cache devices (l2arc_dev_mtx). Once a device has been selected, 4634 * both locks will be dropped and a spa config lock held instead. 4635 */ 4636 mutex_enter(&spa_namespace_lock); 4637 mutex_enter(&l2arc_dev_mtx); 4638 4639 /* if there are no vdevs, there is nothing to do */ 4640 if (l2arc_ndev == 0) 4641 goto out; 4642 4643 first = NULL; 4644 next = l2arc_dev_last; 4645 do { 4646 /* loop around the list looking for a non-faulted vdev */ 4647 if (next == NULL) { 4648 next = list_head(l2arc_dev_list); 4649 } else { 4650 next = list_next(l2arc_dev_list, next); 4651 if (next == NULL) 4652 next = list_head(l2arc_dev_list); 4653 } 4654 4655 /* if we have come back to the start, bail out */ 4656 if (first == NULL) 4657 first = next; 4658 else if (next == first) 4659 break; 4660 4661 } while (vdev_is_dead(next->l2ad_vdev)); 4662 4663 /* if we were unable to find any usable vdevs, return NULL */ 4664 if (vdev_is_dead(next->l2ad_vdev)) 4665 next = NULL; 4666 4667 l2arc_dev_last = next; 4668 4669out: 4670 mutex_exit(&l2arc_dev_mtx); 4671 4672 /* 4673 * Grab the config lock to prevent the 'next' device from being 4674 * removed while we are writing to it. 4675 */ 4676 if (next != NULL) 4677 spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER); 4678 mutex_exit(&spa_namespace_lock); 4679 4680 return (next); 4681} 4682 4683/* 4684 * Free buffers that were tagged for destruction. 4685 */ 4686static void 4687l2arc_do_free_on_write() 4688{ 4689 list_t *buflist; 4690 l2arc_data_free_t *df, *df_prev; 4691 4692 mutex_enter(&l2arc_free_on_write_mtx); 4693 buflist = l2arc_free_on_write; 4694 4695 for (df = list_tail(buflist); df; df = df_prev) { 4696 df_prev = list_prev(buflist, df); 4697 ASSERT(df->l2df_data != NULL); 4698 ASSERT(df->l2df_func != NULL); 4699 df->l2df_func(df->l2df_data, df->l2df_size); 4700 list_remove(buflist, df); 4701 kmem_free(df, sizeof (l2arc_data_free_t)); 4702 } 4703 4704 mutex_exit(&l2arc_free_on_write_mtx); 4705} 4706 4707/* 4708 * A write to a cache device has completed. Update all headers to allow 4709 * reads from these buffers to begin. 4710 */ 4711static void 4712l2arc_write_done(zio_t *zio) 4713{ 4714 l2arc_write_callback_t *cb; 4715 l2arc_dev_t *dev; 4716 list_t *buflist; 4717 arc_buf_hdr_t *head, *ab, *ab_prev; 4718 l2arc_buf_hdr_t *abl2; 4719 kmutex_t *hash_lock; 4720 int64_t bytes_dropped = 0; 4721 4722 cb = zio->io_private; 4723 ASSERT(cb != NULL); 4724 dev = cb->l2wcb_dev; 4725 ASSERT(dev != NULL); 4726 head = cb->l2wcb_head; 4727 ASSERT(head != NULL); 4728 buflist = dev->l2ad_buflist; 4729 ASSERT(buflist != NULL); 4730 DTRACE_PROBE2(l2arc__iodone, zio_t *, zio, 4731 l2arc_write_callback_t *, cb); 4732 4733 if (zio->io_error != 0) 4734 ARCSTAT_BUMP(arcstat_l2_writes_error); 4735 4736 mutex_enter(&l2arc_buflist_mtx); 4737 4738 /* 4739 * All writes completed, or an error was hit. 4740 */ 4741 for (ab = list_prev(buflist, head); ab; ab = ab_prev) { 4742 ab_prev = list_prev(buflist, ab); 4743 abl2 = ab->b_l2hdr; 4744 4745 /* 4746 * Release the temporary compressed buffer as soon as possible. 4747 */ 4748 if (abl2->b_compress != ZIO_COMPRESS_OFF) 4749 l2arc_release_cdata_buf(ab); 4750 4751 hash_lock = HDR_LOCK(ab); 4752 if (!mutex_tryenter(hash_lock)) { 4753 /* 4754 * This buffer misses out. It may be in a stage 4755 * of eviction. Its ARC_L2_WRITING flag will be 4756 * left set, denying reads to this buffer. 4757 */ 4758 ARCSTAT_BUMP(arcstat_l2_writes_hdr_miss); 4759 continue; 4760 } 4761 4762 if (zio->io_error != 0) { 4763 /* 4764 * Error - drop L2ARC entry. 4765 */ 4766 list_remove(buflist, ab); 4767 ARCSTAT_INCR(arcstat_l2_asize, -abl2->b_asize); 4768 bytes_dropped += abl2->b_asize; 4769 ab->b_l2hdr = NULL; 4770 trim_map_free(abl2->b_dev->l2ad_vdev, abl2->b_daddr, 4771 ab->b_size, 0); 4772 kmem_free(abl2, sizeof (l2arc_buf_hdr_t)); 4773 ARCSTAT_INCR(arcstat_l2_size, -ab->b_size); 4774 } 4775 4776 /* 4777 * Allow ARC to begin reads to this L2ARC entry. 4778 */ 4779 ab->b_flags &= ~ARC_L2_WRITING; 4780 4781 mutex_exit(hash_lock); 4782 } 4783 4784 atomic_inc_64(&l2arc_writes_done); 4785 list_remove(buflist, head); 4786 kmem_cache_free(hdr_cache, head); 4787 mutex_exit(&l2arc_buflist_mtx); 4788 4789 vdev_space_update(dev->l2ad_vdev, -bytes_dropped, 0, 0); 4790 4791 l2arc_do_free_on_write(); 4792 4793 kmem_free(cb, sizeof (l2arc_write_callback_t)); 4794} 4795 4796/* 4797 * A read to a cache device completed. Validate buffer contents before 4798 * handing over to the regular ARC routines. 4799 */ 4800static void 4801l2arc_read_done(zio_t *zio) 4802{ 4803 l2arc_read_callback_t *cb; 4804 arc_buf_hdr_t *hdr; 4805 arc_buf_t *buf; 4806 kmutex_t *hash_lock; 4807 int equal; 4808 4809 ASSERT(zio->io_vd != NULL); 4810 ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE); 4811 4812 spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd); 4813 4814 cb = zio->io_private; 4815 ASSERT(cb != NULL); 4816 buf = cb->l2rcb_buf; 4817 ASSERT(buf != NULL); 4818 4819 hash_lock = HDR_LOCK(buf->b_hdr); 4820 mutex_enter(hash_lock); 4821 hdr = buf->b_hdr; 4822 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 4823 4824 /* 4825 * If the buffer was compressed, decompress it first. 4826 */ 4827 if (cb->l2rcb_compress != ZIO_COMPRESS_OFF) 4828 l2arc_decompress_zio(zio, hdr, cb->l2rcb_compress); 4829 ASSERT(zio->io_data != NULL); 4830 4831 /* 4832 * Check this survived the L2ARC journey. 4833 */ 4834 equal = arc_cksum_equal(buf); 4835 if (equal && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) { 4836 mutex_exit(hash_lock); 4837 zio->io_private = buf; 4838 zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */ 4839 zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */ 4840 arc_read_done(zio); 4841 } else { 4842 mutex_exit(hash_lock); 4843 /* 4844 * Buffer didn't survive caching. Increment stats and 4845 * reissue to the original storage device. 4846 */ 4847 if (zio->io_error != 0) { 4848 ARCSTAT_BUMP(arcstat_l2_io_error); 4849 } else { 4850 zio->io_error = SET_ERROR(EIO); 4851 } 4852 if (!equal) 4853 ARCSTAT_BUMP(arcstat_l2_cksum_bad); 4854 4855 /* 4856 * If there's no waiter, issue an async i/o to the primary 4857 * storage now. If there *is* a waiter, the caller must 4858 * issue the i/o in a context where it's OK to block. 4859 */ 4860 if (zio->io_waiter == NULL) { 4861 zio_t *pio = zio_unique_parent(zio); 4862 4863 ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL); 4864 4865 zio_nowait(zio_read(pio, cb->l2rcb_spa, &cb->l2rcb_bp, 4866 buf->b_data, zio->io_size, arc_read_done, buf, 4867 zio->io_priority, cb->l2rcb_flags, &cb->l2rcb_zb)); 4868 } 4869 } 4870 4871 kmem_free(cb, sizeof (l2arc_read_callback_t)); 4872} 4873 4874/* 4875 * This is the list priority from which the L2ARC will search for pages to 4876 * cache. This is used within loops (0..3) to cycle through lists in the 4877 * desired order. This order can have a significant effect on cache 4878 * performance. 4879 * 4880 * Currently the metadata lists are hit first, MFU then MRU, followed by 4881 * the data lists. This function returns a locked list, and also returns 4882 * the lock pointer. 4883 */ 4884static list_t * 4885l2arc_list_locked(int list_num, kmutex_t **lock) 4886{ 4887 list_t *list = NULL; 4888 int idx; 4889 4890 ASSERT(list_num >= 0 && list_num < 2 * ARC_BUFC_NUMLISTS); 4891 4892 if (list_num < ARC_BUFC_NUMMETADATALISTS) { 4893 idx = list_num; 4894 list = &arc_mfu->arcs_lists[idx]; 4895 *lock = ARCS_LOCK(arc_mfu, idx); 4896 } else if (list_num < ARC_BUFC_NUMMETADATALISTS * 2) { 4897 idx = list_num - ARC_BUFC_NUMMETADATALISTS; 4898 list = &arc_mru->arcs_lists[idx]; 4899 *lock = ARCS_LOCK(arc_mru, idx); 4900 } else if (list_num < (ARC_BUFC_NUMMETADATALISTS * 2 + 4901 ARC_BUFC_NUMDATALISTS)) { 4902 idx = list_num - ARC_BUFC_NUMMETADATALISTS; 4903 list = &arc_mfu->arcs_lists[idx]; 4904 *lock = ARCS_LOCK(arc_mfu, idx); 4905 } else { 4906 idx = list_num - ARC_BUFC_NUMLISTS; 4907 list = &arc_mru->arcs_lists[idx]; 4908 *lock = ARCS_LOCK(arc_mru, idx); 4909 } 4910 4911 ASSERT(!(MUTEX_HELD(*lock))); 4912 mutex_enter(*lock); 4913 return (list); 4914} 4915 4916/* 4917 * Evict buffers from the device write hand to the distance specified in 4918 * bytes. This distance may span populated buffers, it may span nothing. 4919 * This is clearing a region on the L2ARC device ready for writing. 4920 * If the 'all' boolean is set, every buffer is evicted. 4921 */ 4922static void 4923l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all) 4924{ 4925 list_t *buflist; 4926 l2arc_buf_hdr_t *abl2; 4927 arc_buf_hdr_t *ab, *ab_prev; 4928 kmutex_t *hash_lock; 4929 uint64_t taddr; 4930 int64_t bytes_evicted = 0; 4931 4932 buflist = dev->l2ad_buflist; 4933 4934 if (buflist == NULL) 4935 return; 4936 4937 if (!all && dev->l2ad_first) { 4938 /* 4939 * This is the first sweep through the device. There is 4940 * nothing to evict. 4941 */ 4942 return; 4943 } 4944 4945 if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) { 4946 /* 4947 * When nearing the end of the device, evict to the end 4948 * before the device write hand jumps to the start. 4949 */ 4950 taddr = dev->l2ad_end; 4951 } else { 4952 taddr = dev->l2ad_hand + distance; 4953 } 4954 DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist, 4955 uint64_t, taddr, boolean_t, all); 4956 4957top: 4958 mutex_enter(&l2arc_buflist_mtx); 4959 for (ab = list_tail(buflist); ab; ab = ab_prev) { 4960 ab_prev = list_prev(buflist, ab); 4961 4962 hash_lock = HDR_LOCK(ab); 4963 if (!mutex_tryenter(hash_lock)) { 4964 /* 4965 * Missed the hash lock. Retry. 4966 */ 4967 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry); 4968 mutex_exit(&l2arc_buflist_mtx); 4969 mutex_enter(hash_lock); 4970 mutex_exit(hash_lock); 4971 goto top; 4972 } 4973 4974 if (HDR_L2_WRITE_HEAD(ab)) { 4975 /* 4976 * We hit a write head node. Leave it for 4977 * l2arc_write_done(). 4978 */ 4979 list_remove(buflist, ab); 4980 mutex_exit(hash_lock); 4981 continue; 4982 } 4983 4984 if (!all && ab->b_l2hdr != NULL && 4985 (ab->b_l2hdr->b_daddr > taddr || 4986 ab->b_l2hdr->b_daddr < dev->l2ad_hand)) { 4987 /* 4988 * We've evicted to the target address, 4989 * or the end of the device. 4990 */ 4991 mutex_exit(hash_lock); 4992 break; 4993 } 4994 4995 if (HDR_FREE_IN_PROGRESS(ab)) { 4996 /* 4997 * Already on the path to destruction. 4998 */ 4999 mutex_exit(hash_lock); 5000 continue; 5001 } 5002 5003 if (ab->b_state == arc_l2c_only) { 5004 ASSERT(!HDR_L2_READING(ab)); 5005 /* 5006 * This doesn't exist in the ARC. Destroy. 5007 * arc_hdr_destroy() will call list_remove() 5008 * and decrement arcstat_l2_size. 5009 */ 5010 arc_change_state(arc_anon, ab, hash_lock); 5011 arc_hdr_destroy(ab); 5012 } else { 5013 /* 5014 * Invalidate issued or about to be issued 5015 * reads, since we may be about to write 5016 * over this location. 5017 */ 5018 if (HDR_L2_READING(ab)) { 5019 ARCSTAT_BUMP(arcstat_l2_evict_reading); 5020 ab->b_flags |= ARC_L2_EVICTED; 5021 } 5022 5023 /* 5024 * Tell ARC this no longer exists in L2ARC. 5025 */ 5026 if (ab->b_l2hdr != NULL) { 5027 abl2 = ab->b_l2hdr; 5028 ARCSTAT_INCR(arcstat_l2_asize, -abl2->b_asize); 5029 bytes_evicted += abl2->b_asize; 5030 ab->b_l2hdr = NULL; 5031 /* 5032 * We are destroying l2hdr, so ensure that 5033 * its compressed buffer, if any, is not leaked. 5034 */ 5035 ASSERT(abl2->b_tmp_cdata == NULL); 5036 kmem_free(abl2, sizeof (l2arc_buf_hdr_t)); 5037 ARCSTAT_INCR(arcstat_l2_size, -ab->b_size); 5038 } 5039 list_remove(buflist, ab); 5040 5041 /* 5042 * This may have been leftover after a 5043 * failed write. 5044 */ 5045 ab->b_flags &= ~ARC_L2_WRITING; 5046 } 5047 mutex_exit(hash_lock); 5048 } 5049 mutex_exit(&l2arc_buflist_mtx); 5050 5051 vdev_space_update(dev->l2ad_vdev, -bytes_evicted, 0, 0); 5052 dev->l2ad_evict = taddr; 5053} 5054 5055/* 5056 * Find and write ARC buffers to the L2ARC device. 5057 * 5058 * An ARC_L2_WRITING flag is set so that the L2ARC buffers are not valid 5059 * for reading until they have completed writing. 5060 * The headroom_boost is an in-out parameter used to maintain headroom boost 5061 * state between calls to this function. 5062 * 5063 * Returns the number of bytes actually written (which may be smaller than 5064 * the delta by which the device hand has changed due to alignment). 5065 */ 5066static uint64_t 5067l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz, 5068 boolean_t *headroom_boost) 5069{ 5070 arc_buf_hdr_t *ab, *ab_prev, *head; 5071 list_t *list; 5072 uint64_t write_asize, write_psize, write_sz, headroom, 5073 buf_compress_minsz; 5074 void *buf_data; 5075 kmutex_t *list_lock; 5076 boolean_t full; 5077 l2arc_write_callback_t *cb; 5078 zio_t *pio, *wzio; 5079 uint64_t guid = spa_load_guid(spa); 5080 const boolean_t do_headroom_boost = *headroom_boost; 5081 int try; 5082 5083 ASSERT(dev->l2ad_vdev != NULL); 5084 5085 /* Lower the flag now, we might want to raise it again later. */ 5086 *headroom_boost = B_FALSE; 5087 5088 pio = NULL; 5089 write_sz = write_asize = write_psize = 0; 5090 full = B_FALSE; 5091 head = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE); 5092 head->b_flags |= ARC_L2_WRITE_HEAD; 5093 5094 ARCSTAT_BUMP(arcstat_l2_write_buffer_iter); 5095 /* 5096 * We will want to try to compress buffers that are at least 2x the 5097 * device sector size. 5098 */ 5099 buf_compress_minsz = 2 << dev->l2ad_vdev->vdev_ashift; 5100 5101 /* 5102 * Copy buffers for L2ARC writing. 5103 */ 5104 mutex_enter(&l2arc_buflist_mtx); 5105 for (try = 0; try < 2 * ARC_BUFC_NUMLISTS; try++) { 5106 uint64_t passed_sz = 0; 5107 5108 list = l2arc_list_locked(try, &list_lock); 5109 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_iter); 5110 5111 /* 5112 * L2ARC fast warmup. 5113 * 5114 * Until the ARC is warm and starts to evict, read from the 5115 * head of the ARC lists rather than the tail. 5116 */ 5117 if (arc_warm == B_FALSE) 5118 ab = list_head(list); 5119 else 5120 ab = list_tail(list); 5121 if (ab == NULL) 5122 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_null_iter); 5123 5124 headroom = target_sz * l2arc_headroom * 2 / ARC_BUFC_NUMLISTS; 5125 if (do_headroom_boost) 5126 headroom = (headroom * l2arc_headroom_boost) / 100; 5127 5128 for (; ab; ab = ab_prev) { 5129 l2arc_buf_hdr_t *l2hdr; 5130 kmutex_t *hash_lock; 5131 uint64_t buf_sz; 5132 5133 if (arc_warm == B_FALSE) 5134 ab_prev = list_next(list, ab); 5135 else 5136 ab_prev = list_prev(list, ab); 5137 ARCSTAT_INCR(arcstat_l2_write_buffer_bytes_scanned, ab->b_size); 5138 5139 hash_lock = HDR_LOCK(ab); 5140 if (!mutex_tryenter(hash_lock)) { 5141 ARCSTAT_BUMP(arcstat_l2_write_trylock_fail); 5142 /* 5143 * Skip this buffer rather than waiting. 5144 */ 5145 continue; 5146 } 5147 5148 passed_sz += ab->b_size; 5149 if (passed_sz > headroom) { 5150 /* 5151 * Searched too far. 5152 */ 5153 mutex_exit(hash_lock); 5154 ARCSTAT_BUMP(arcstat_l2_write_passed_headroom); 5155 break; 5156 } 5157 5158 if (!l2arc_write_eligible(guid, ab)) { 5159 mutex_exit(hash_lock); 5160 continue; 5161 } 5162 5163 if ((write_sz + ab->b_size) > target_sz) { 5164 full = B_TRUE; 5165 mutex_exit(hash_lock); 5166 ARCSTAT_BUMP(arcstat_l2_write_full); 5167 break; 5168 } 5169 5170 if (pio == NULL) { 5171 /* 5172 * Insert a dummy header on the buflist so 5173 * l2arc_write_done() can find where the 5174 * write buffers begin without searching. 5175 */ 5176 list_insert_head(dev->l2ad_buflist, head); 5177 5178 cb = kmem_alloc( 5179 sizeof (l2arc_write_callback_t), KM_SLEEP); 5180 cb->l2wcb_dev = dev; 5181 cb->l2wcb_head = head; 5182 pio = zio_root(spa, l2arc_write_done, cb, 5183 ZIO_FLAG_CANFAIL); 5184 ARCSTAT_BUMP(arcstat_l2_write_pios); 5185 } 5186 5187 /* 5188 * Create and add a new L2ARC header. 5189 */ 5190 l2hdr = kmem_zalloc(sizeof (l2arc_buf_hdr_t), KM_SLEEP); 5191 l2hdr->b_dev = dev; 5192 ab->b_flags |= ARC_L2_WRITING; 5193 5194 /* 5195 * Temporarily stash the data buffer in b_tmp_cdata. 5196 * The subsequent write step will pick it up from 5197 * there. This is because can't access ab->b_buf 5198 * without holding the hash_lock, which we in turn 5199 * can't access without holding the ARC list locks 5200 * (which we want to avoid during compression/writing). 5201 */ 5202 l2hdr->b_compress = ZIO_COMPRESS_OFF; 5203 l2hdr->b_asize = ab->b_size; 5204 l2hdr->b_tmp_cdata = ab->b_buf->b_data; 5205 5206 buf_sz = ab->b_size; 5207 ab->b_l2hdr = l2hdr; 5208 5209 list_insert_head(dev->l2ad_buflist, ab); 5210 5211 /* 5212 * Compute and store the buffer cksum before 5213 * writing. On debug the cksum is verified first. 5214 */ 5215 arc_cksum_verify(ab->b_buf); 5216 arc_cksum_compute(ab->b_buf, B_TRUE); 5217 5218 mutex_exit(hash_lock); 5219 5220 write_sz += buf_sz; 5221 } 5222 5223 mutex_exit(list_lock); 5224 5225 if (full == B_TRUE) 5226 break; 5227 } 5228 5229 /* No buffers selected for writing? */ 5230 if (pio == NULL) { 5231 ASSERT0(write_sz); 5232 mutex_exit(&l2arc_buflist_mtx); 5233 kmem_cache_free(hdr_cache, head); 5234 return (0); 5235 } 5236 5237 /* 5238 * Now start writing the buffers. We're starting at the write head 5239 * and work backwards, retracing the course of the buffer selector 5240 * loop above. 5241 */ 5242 for (ab = list_prev(dev->l2ad_buflist, head); ab; 5243 ab = list_prev(dev->l2ad_buflist, ab)) { 5244 l2arc_buf_hdr_t *l2hdr; 5245 uint64_t buf_sz; 5246 5247 /* 5248 * We shouldn't need to lock the buffer here, since we flagged 5249 * it as ARC_L2_WRITING in the previous step, but we must take 5250 * care to only access its L2 cache parameters. In particular, 5251 * ab->b_buf may be invalid by now due to ARC eviction. 5252 */ 5253 l2hdr = ab->b_l2hdr; 5254 l2hdr->b_daddr = dev->l2ad_hand; 5255 5256 if ((ab->b_flags & ARC_L2COMPRESS) && 5257 l2hdr->b_asize >= buf_compress_minsz) { 5258 if (l2arc_compress_buf(l2hdr)) { 5259 /* 5260 * If compression succeeded, enable headroom 5261 * boost on the next scan cycle. 5262 */ 5263 *headroom_boost = B_TRUE; 5264 } 5265 } 5266 5267 /* 5268 * Pick up the buffer data we had previously stashed away 5269 * (and now potentially also compressed). 5270 */ 5271 buf_data = l2hdr->b_tmp_cdata; 5272 buf_sz = l2hdr->b_asize; 5273 5274 /* 5275 * If the data has not been compressed, then clear b_tmp_cdata 5276 * to make sure that it points only to a temporary compression 5277 * buffer. 5278 */ 5279 if (!L2ARC_IS_VALID_COMPRESS(l2hdr->b_compress)) 5280 l2hdr->b_tmp_cdata = NULL; 5281 5282 /* Compression may have squashed the buffer to zero length. */ 5283 if (buf_sz != 0) { 5284 uint64_t buf_p_sz; 5285 5286 wzio = zio_write_phys(pio, dev->l2ad_vdev, 5287 dev->l2ad_hand, buf_sz, buf_data, ZIO_CHECKSUM_OFF, 5288 NULL, NULL, ZIO_PRIORITY_ASYNC_WRITE, 5289 ZIO_FLAG_CANFAIL, B_FALSE); 5290 5291 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev, 5292 zio_t *, wzio); 5293 (void) zio_nowait(wzio); 5294 5295 write_asize += buf_sz; 5296 /* 5297 * Keep the clock hand suitably device-aligned. 5298 */ 5299 buf_p_sz = vdev_psize_to_asize(dev->l2ad_vdev, buf_sz); 5300 write_psize += buf_p_sz; 5301 dev->l2ad_hand += buf_p_sz; 5302 } 5303 } 5304 5305 mutex_exit(&l2arc_buflist_mtx); 5306 5307 ASSERT3U(write_asize, <=, target_sz); 5308 ARCSTAT_BUMP(arcstat_l2_writes_sent); 5309 ARCSTAT_INCR(arcstat_l2_write_bytes, write_asize); 5310 ARCSTAT_INCR(arcstat_l2_size, write_sz); 5311 ARCSTAT_INCR(arcstat_l2_asize, write_asize); 5312 vdev_space_update(dev->l2ad_vdev, write_psize, 0, 0); 5313 5314 /* 5315 * Bump device hand to the device start if it is approaching the end. 5316 * l2arc_evict() will already have evicted ahead for this case. 5317 */ 5318 if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) { 5319 dev->l2ad_hand = dev->l2ad_start; 5320 dev->l2ad_evict = dev->l2ad_start; 5321 dev->l2ad_first = B_FALSE; 5322 } 5323 5324 dev->l2ad_writing = B_TRUE; 5325 (void) zio_wait(pio); 5326 dev->l2ad_writing = B_FALSE; 5327 5328 return (write_asize); 5329} 5330 5331/* 5332 * Compresses an L2ARC buffer. 5333 * The data to be compressed must be prefilled in l2hdr->b_tmp_cdata and its 5334 * size in l2hdr->b_asize. This routine tries to compress the data and 5335 * depending on the compression result there are three possible outcomes: 5336 * *) The buffer was incompressible. The original l2hdr contents were left 5337 * untouched and are ready for writing to an L2 device. 5338 * *) The buffer was all-zeros, so there is no need to write it to an L2 5339 * device. To indicate this situation b_tmp_cdata is NULL'ed, b_asize is 5340 * set to zero and b_compress is set to ZIO_COMPRESS_EMPTY. 5341 * *) Compression succeeded and b_tmp_cdata was replaced with a temporary 5342 * data buffer which holds the compressed data to be written, and b_asize 5343 * tells us how much data there is. b_compress is set to the appropriate 5344 * compression algorithm. Once writing is done, invoke 5345 * l2arc_release_cdata_buf on this l2hdr to free this temporary buffer. 5346 * 5347 * Returns B_TRUE if compression succeeded, or B_FALSE if it didn't (the 5348 * buffer was incompressible). 5349 */ 5350static boolean_t 5351l2arc_compress_buf(l2arc_buf_hdr_t *l2hdr) 5352{ 5353 void *cdata; 5354 size_t csize, len, rounded; 5355 5356 ASSERT(l2hdr->b_compress == ZIO_COMPRESS_OFF); 5357 ASSERT(l2hdr->b_tmp_cdata != NULL); 5358 5359 len = l2hdr->b_asize; 5360 cdata = zio_data_buf_alloc(len); 5361 csize = zio_compress_data(ZIO_COMPRESS_LZ4, l2hdr->b_tmp_cdata, 5362 cdata, l2hdr->b_asize); 5363 5364 if (csize == 0) { 5365 /* zero block, indicate that there's nothing to write */ 5366 zio_data_buf_free(cdata, len); 5367 l2hdr->b_compress = ZIO_COMPRESS_EMPTY; 5368 l2hdr->b_asize = 0; 5369 l2hdr->b_tmp_cdata = NULL; 5370 ARCSTAT_BUMP(arcstat_l2_compress_zeros); 5371 return (B_TRUE); 5372 } 5373 5374 rounded = P2ROUNDUP(csize, 5375 (size_t)1 << l2hdr->b_dev->l2ad_vdev->vdev_ashift); 5376 if (rounded < len) { 5377 /* 5378 * Compression succeeded, we'll keep the cdata around for 5379 * writing and release it afterwards. 5380 */ 5381 if (rounded > csize) { 5382 bzero((char *)cdata + csize, rounded - csize); 5383 csize = rounded; 5384 } 5385 l2hdr->b_compress = ZIO_COMPRESS_LZ4; 5386 l2hdr->b_asize = csize; 5387 l2hdr->b_tmp_cdata = cdata; 5388 ARCSTAT_BUMP(arcstat_l2_compress_successes); 5389 return (B_TRUE); 5390 } else { 5391 /* 5392 * Compression failed, release the compressed buffer. 5393 * l2hdr will be left unmodified. 5394 */ 5395 zio_data_buf_free(cdata, len); 5396 ARCSTAT_BUMP(arcstat_l2_compress_failures); 5397 return (B_FALSE); 5398 } 5399} 5400 5401/* 5402 * Decompresses a zio read back from an l2arc device. On success, the 5403 * underlying zio's io_data buffer is overwritten by the uncompressed 5404 * version. On decompression error (corrupt compressed stream), the 5405 * zio->io_error value is set to signal an I/O error. 5406 * 5407 * Please note that the compressed data stream is not checksummed, so 5408 * if the underlying device is experiencing data corruption, we may feed 5409 * corrupt data to the decompressor, so the decompressor needs to be 5410 * able to handle this situation (LZ4 does). 5411 */ 5412static void 5413l2arc_decompress_zio(zio_t *zio, arc_buf_hdr_t *hdr, enum zio_compress c) 5414{ 5415 ASSERT(L2ARC_IS_VALID_COMPRESS(c)); 5416 5417 if (zio->io_error != 0) { 5418 /* 5419 * An io error has occured, just restore the original io 5420 * size in preparation for a main pool read. 5421 */ 5422 zio->io_orig_size = zio->io_size = hdr->b_size; 5423 return; 5424 } 5425 5426 if (c == ZIO_COMPRESS_EMPTY) { 5427 /* 5428 * An empty buffer results in a null zio, which means we 5429 * need to fill its io_data after we're done restoring the 5430 * buffer's contents. 5431 */ 5432 ASSERT(hdr->b_buf != NULL); 5433 bzero(hdr->b_buf->b_data, hdr->b_size); 5434 zio->io_data = zio->io_orig_data = hdr->b_buf->b_data; 5435 } else { 5436 ASSERT(zio->io_data != NULL); 5437 /* 5438 * We copy the compressed data from the start of the arc buffer 5439 * (the zio_read will have pulled in only what we need, the 5440 * rest is garbage which we will overwrite at decompression) 5441 * and then decompress back to the ARC data buffer. This way we 5442 * can minimize copying by simply decompressing back over the 5443 * original compressed data (rather than decompressing to an 5444 * aux buffer and then copying back the uncompressed buffer, 5445 * which is likely to be much larger). 5446 */ 5447 uint64_t csize; 5448 void *cdata; 5449 5450 csize = zio->io_size; 5451 cdata = zio_data_buf_alloc(csize); 5452 bcopy(zio->io_data, cdata, csize); 5453 if (zio_decompress_data(c, cdata, zio->io_data, csize, 5454 hdr->b_size) != 0) 5455 zio->io_error = EIO; 5456 zio_data_buf_free(cdata, csize); 5457 } 5458 5459 /* Restore the expected uncompressed IO size. */ 5460 zio->io_orig_size = zio->io_size = hdr->b_size; 5461} 5462 5463/* 5464 * Releases the temporary b_tmp_cdata buffer in an l2arc header structure. 5465 * This buffer serves as a temporary holder of compressed data while 5466 * the buffer entry is being written to an l2arc device. Once that is 5467 * done, we can dispose of it. 5468 */ 5469static void 5470l2arc_release_cdata_buf(arc_buf_hdr_t *ab) 5471{ 5472 l2arc_buf_hdr_t *l2hdr = ab->b_l2hdr; 5473 5474 ASSERT(L2ARC_IS_VALID_COMPRESS(l2hdr->b_compress)); 5475 if (l2hdr->b_compress != ZIO_COMPRESS_EMPTY) { 5476 /* 5477 * If the data was compressed, then we've allocated a 5478 * temporary buffer for it, so now we need to release it. 5479 */ 5480 ASSERT(l2hdr->b_tmp_cdata != NULL); 5481 zio_data_buf_free(l2hdr->b_tmp_cdata, ab->b_size); 5482 l2hdr->b_tmp_cdata = NULL; 5483 } else { 5484 ASSERT(l2hdr->b_tmp_cdata == NULL); 5485 } 5486} 5487 5488/* 5489 * This thread feeds the L2ARC at regular intervals. This is the beating 5490 * heart of the L2ARC. 5491 */ 5492static void 5493l2arc_feed_thread(void *dummy __unused) 5494{ 5495 callb_cpr_t cpr; 5496 l2arc_dev_t *dev; 5497 spa_t *spa; 5498 uint64_t size, wrote; 5499 clock_t begin, next = ddi_get_lbolt(); 5500 boolean_t headroom_boost = B_FALSE; 5501 5502 CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG); 5503 5504 mutex_enter(&l2arc_feed_thr_lock); 5505 5506 while (l2arc_thread_exit == 0) { 5507 CALLB_CPR_SAFE_BEGIN(&cpr); 5508 (void) cv_timedwait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock, 5509 next - ddi_get_lbolt()); 5510 CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock); 5511 next = ddi_get_lbolt() + hz; 5512 5513 /* 5514 * Quick check for L2ARC devices. 5515 */ 5516 mutex_enter(&l2arc_dev_mtx); 5517 if (l2arc_ndev == 0) { 5518 mutex_exit(&l2arc_dev_mtx); 5519 continue; 5520 } 5521 mutex_exit(&l2arc_dev_mtx); 5522 begin = ddi_get_lbolt(); 5523 5524 /* 5525 * This selects the next l2arc device to write to, and in 5526 * doing so the next spa to feed from: dev->l2ad_spa. This 5527 * will return NULL if there are now no l2arc devices or if 5528 * they are all faulted. 5529 * 5530 * If a device is returned, its spa's config lock is also 5531 * held to prevent device removal. l2arc_dev_get_next() 5532 * will grab and release l2arc_dev_mtx. 5533 */ 5534 if ((dev = l2arc_dev_get_next()) == NULL) 5535 continue; 5536 5537 spa = dev->l2ad_spa; 5538 ASSERT(spa != NULL); 5539 5540 /* 5541 * If the pool is read-only then force the feed thread to 5542 * sleep a little longer. 5543 */ 5544 if (!spa_writeable(spa)) { 5545 next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz; 5546 spa_config_exit(spa, SCL_L2ARC, dev); 5547 continue; 5548 } 5549 5550 /* 5551 * Avoid contributing to memory pressure. 5552 */ 5553 if (arc_reclaim_needed()) { 5554 ARCSTAT_BUMP(arcstat_l2_abort_lowmem); 5555 spa_config_exit(spa, SCL_L2ARC, dev); 5556 continue; 5557 } 5558 5559 ARCSTAT_BUMP(arcstat_l2_feeds); 5560 5561 size = l2arc_write_size(); 5562 5563 /* 5564 * Evict L2ARC buffers that will be overwritten. 5565 */ 5566 l2arc_evict(dev, size, B_FALSE); 5567 5568 /* 5569 * Write ARC buffers. 5570 */ 5571 wrote = l2arc_write_buffers(spa, dev, size, &headroom_boost); 5572 5573 /* 5574 * Calculate interval between writes. 5575 */ 5576 next = l2arc_write_interval(begin, size, wrote); 5577 spa_config_exit(spa, SCL_L2ARC, dev); 5578 } 5579 5580 l2arc_thread_exit = 0; 5581 cv_broadcast(&l2arc_feed_thr_cv); 5582 CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */ 5583 thread_exit(); 5584} 5585 5586boolean_t 5587l2arc_vdev_present(vdev_t *vd) 5588{ 5589 l2arc_dev_t *dev; 5590 5591 mutex_enter(&l2arc_dev_mtx); 5592 for (dev = list_head(l2arc_dev_list); dev != NULL; 5593 dev = list_next(l2arc_dev_list, dev)) { 5594 if (dev->l2ad_vdev == vd) 5595 break; 5596 } 5597 mutex_exit(&l2arc_dev_mtx); 5598 5599 return (dev != NULL); 5600} 5601 5602/* 5603 * Add a vdev for use by the L2ARC. By this point the spa has already 5604 * validated the vdev and opened it. 5605 */ 5606void 5607l2arc_add_vdev(spa_t *spa, vdev_t *vd) 5608{ 5609 l2arc_dev_t *adddev; 5610 5611 ASSERT(!l2arc_vdev_present(vd)); 5612 5613 vdev_ashift_optimize(vd); 5614 5615 /* 5616 * Create a new l2arc device entry. 5617 */ 5618 adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP); 5619 adddev->l2ad_spa = spa; 5620 adddev->l2ad_vdev = vd; 5621 adddev->l2ad_start = VDEV_LABEL_START_SIZE; 5622 adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd); 5623 adddev->l2ad_hand = adddev->l2ad_start; 5624 adddev->l2ad_evict = adddev->l2ad_start; 5625 adddev->l2ad_first = B_TRUE; 5626 adddev->l2ad_writing = B_FALSE; 5627 5628 /* 5629 * This is a list of all ARC buffers that are still valid on the 5630 * device. 5631 */ 5632 adddev->l2ad_buflist = kmem_zalloc(sizeof (list_t), KM_SLEEP); 5633 list_create(adddev->l2ad_buflist, sizeof (arc_buf_hdr_t), 5634 offsetof(arc_buf_hdr_t, b_l2node)); 5635 5636 vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand); 5637 5638 /* 5639 * Add device to global list 5640 */ 5641 mutex_enter(&l2arc_dev_mtx); 5642 list_insert_head(l2arc_dev_list, adddev); 5643 atomic_inc_64(&l2arc_ndev); 5644 mutex_exit(&l2arc_dev_mtx); 5645} 5646 5647/* 5648 * Remove a vdev from the L2ARC. 5649 */ 5650void 5651l2arc_remove_vdev(vdev_t *vd) 5652{ 5653 l2arc_dev_t *dev, *nextdev, *remdev = NULL; 5654 5655 /* 5656 * Find the device by vdev 5657 */ 5658 mutex_enter(&l2arc_dev_mtx); 5659 for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) { 5660 nextdev = list_next(l2arc_dev_list, dev); 5661 if (vd == dev->l2ad_vdev) { 5662 remdev = dev; 5663 break; 5664 } 5665 } 5666 ASSERT(remdev != NULL); 5667 5668 /* 5669 * Remove device from global list 5670 */ 5671 list_remove(l2arc_dev_list, remdev); 5672 l2arc_dev_last = NULL; /* may have been invalidated */ 5673 atomic_dec_64(&l2arc_ndev); 5674 mutex_exit(&l2arc_dev_mtx); 5675 5676 /* 5677 * Clear all buflists and ARC references. L2ARC device flush. 5678 */ 5679 l2arc_evict(remdev, 0, B_TRUE); 5680 list_destroy(remdev->l2ad_buflist); 5681 kmem_free(remdev->l2ad_buflist, sizeof (list_t)); 5682 kmem_free(remdev, sizeof (l2arc_dev_t)); 5683} 5684 5685void 5686l2arc_init(void) 5687{ 5688 l2arc_thread_exit = 0; 5689 l2arc_ndev = 0; 5690 l2arc_writes_sent = 0; 5691 l2arc_writes_done = 0; 5692 5693 mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL); 5694 cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL); 5695 mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL); 5696 mutex_init(&l2arc_buflist_mtx, NULL, MUTEX_DEFAULT, NULL); 5697 mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL); 5698 5699 l2arc_dev_list = &L2ARC_dev_list; 5700 l2arc_free_on_write = &L2ARC_free_on_write; 5701 list_create(l2arc_dev_list, sizeof (l2arc_dev_t), 5702 offsetof(l2arc_dev_t, l2ad_node)); 5703 list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t), 5704 offsetof(l2arc_data_free_t, l2df_list_node)); 5705} 5706 5707void 5708l2arc_fini(void) 5709{ 5710 /* 5711 * This is called from dmu_fini(), which is called from spa_fini(); 5712 * Because of this, we can assume that all l2arc devices have 5713 * already been removed when the pools themselves were removed. 5714 */ 5715 5716 l2arc_do_free_on_write(); 5717 5718 mutex_destroy(&l2arc_feed_thr_lock); 5719 cv_destroy(&l2arc_feed_thr_cv); 5720 mutex_destroy(&l2arc_dev_mtx); 5721 mutex_destroy(&l2arc_buflist_mtx); 5722 mutex_destroy(&l2arc_free_on_write_mtx); 5723 5724 list_destroy(l2arc_dev_list); 5725 list_destroy(l2arc_free_on_write); 5726} 5727 5728void 5729l2arc_start(void) 5730{ 5731 if (!(spa_mode_global & FWRITE)) 5732 return; 5733 5734 (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0, 5735 TS_RUN, minclsyspri); 5736} 5737 5738void 5739l2arc_stop(void) 5740{ 5741 if (!(spa_mode_global & FWRITE)) 5742 return; 5743 5744 mutex_enter(&l2arc_feed_thr_lock); 5745 cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */ 5746 l2arc_thread_exit = 1; 5747 while (l2arc_thread_exit != 0) 5748 cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock); 5749 mutex_exit(&l2arc_feed_thr_lock); 5750} 5751