arc.c revision 277300
1/* 2 * CDDL HEADER START 3 * 4 * The contents of this file are subject to the terms of the 5 * Common Development and Distribution License (the "License"). 6 * You may not use this file except in compliance with the License. 7 * 8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE 9 * or http://www.opensolaris.org/os/licensing. 10 * See the License for the specific language governing permissions 11 * and limitations under the License. 12 * 13 * When distributing Covered Code, include this CDDL HEADER in each 14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE. 15 * If applicable, add the following below this CDDL HEADER, with the 16 * fields enclosed by brackets "[]" replaced with your own identifying 17 * information: Portions Copyright [yyyy] [name of copyright owner] 18 * 19 * CDDL HEADER END 20 */ 21/* 22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved. 23 * Copyright (c) 2011, 2014 by Delphix. All rights reserved. 24 * Copyright (c) 2014 by Saso Kiselkov. All rights reserved. 25 * Copyright 2014 Nexenta Systems, Inc. All rights reserved. 26 */ 27 28/* 29 * DVA-based Adjustable Replacement Cache 30 * 31 * While much of the theory of operation used here is 32 * based on the self-tuning, low overhead replacement cache 33 * presented by Megiddo and Modha at FAST 2003, there are some 34 * significant differences: 35 * 36 * 1. The Megiddo and Modha model assumes any page is evictable. 37 * Pages in its cache cannot be "locked" into memory. This makes 38 * the eviction algorithm simple: evict the last page in the list. 39 * This also make the performance characteristics easy to reason 40 * about. Our cache is not so simple. At any given moment, some 41 * subset of the blocks in the cache are un-evictable because we 42 * have handed out a reference to them. Blocks are only evictable 43 * when there are no external references active. This makes 44 * eviction far more problematic: we choose to evict the evictable 45 * blocks that are the "lowest" in the list. 46 * 47 * There are times when it is not possible to evict the requested 48 * space. In these circumstances we are unable to adjust the cache 49 * size. To prevent the cache growing unbounded at these times we 50 * implement a "cache throttle" that slows the flow of new data 51 * into the cache until we can make space available. 52 * 53 * 2. The Megiddo and Modha model assumes a fixed cache size. 54 * Pages are evicted when the cache is full and there is a cache 55 * miss. Our model has a variable sized cache. It grows with 56 * high use, but also tries to react to memory pressure from the 57 * operating system: decreasing its size when system memory is 58 * tight. 59 * 60 * 3. The Megiddo and Modha model assumes a fixed page size. All 61 * elements of the cache are therefore exactly the same size. So 62 * when adjusting the cache size following a cache miss, its simply 63 * a matter of choosing a single page to evict. In our model, we 64 * have variable sized cache blocks (rangeing from 512 bytes to 65 * 128K bytes). We therefore choose a set of blocks to evict to make 66 * space for a cache miss that approximates as closely as possible 67 * the space used by the new block. 68 * 69 * See also: "ARC: A Self-Tuning, Low Overhead Replacement Cache" 70 * by N. Megiddo & D. Modha, FAST 2003 71 */ 72 73/* 74 * The locking model: 75 * 76 * A new reference to a cache buffer can be obtained in two 77 * ways: 1) via a hash table lookup using the DVA as a key, 78 * or 2) via one of the ARC lists. The arc_read() interface 79 * uses method 1, while the internal arc algorithms for 80 * adjusting the cache use method 2. We therefore provide two 81 * types of locks: 1) the hash table lock array, and 2) the 82 * arc list locks. 83 * 84 * Buffers do not have their own mutexs, rather they rely on the 85 * hash table mutexs for the bulk of their protection (i.e. most 86 * fields in the arc_buf_hdr_t are protected by these mutexs). 87 * 88 * buf_hash_find() returns the appropriate mutex (held) when it 89 * locates the requested buffer in the hash table. It returns 90 * NULL for the mutex if the buffer was not in the table. 91 * 92 * buf_hash_remove() expects the appropriate hash mutex to be 93 * already held before it is invoked. 94 * 95 * Each arc state also has a mutex which is used to protect the 96 * buffer list associated with the state. When attempting to 97 * obtain a hash table lock while holding an arc list lock you 98 * must use: mutex_tryenter() to avoid deadlock. Also note that 99 * the active state mutex must be held before the ghost state mutex. 100 * 101 * Arc buffers may have an associated eviction callback function. 102 * This function will be invoked prior to removing the buffer (e.g. 103 * in arc_do_user_evicts()). Note however that the data associated 104 * with the buffer may be evicted prior to the callback. The callback 105 * must be made with *no locks held* (to prevent deadlock). Additionally, 106 * the users of callbacks must ensure that their private data is 107 * protected from simultaneous callbacks from arc_clear_callback() 108 * and arc_do_user_evicts(). 109 * 110 * Note that the majority of the performance stats are manipulated 111 * with atomic operations. 112 * 113 * The L2ARC uses the l2arc_buflist_mtx global mutex for the following: 114 * 115 * - L2ARC buflist creation 116 * - L2ARC buflist eviction 117 * - L2ARC write completion, which walks L2ARC buflists 118 * - ARC header destruction, as it removes from L2ARC buflists 119 * - ARC header release, as it removes from L2ARC buflists 120 */ 121 122#include <sys/spa.h> 123#include <sys/zio.h> 124#include <sys/zio_compress.h> 125#include <sys/zfs_context.h> 126#include <sys/arc.h> 127#include <sys/refcount.h> 128#include <sys/vdev.h> 129#include <sys/vdev_impl.h> 130#include <sys/dsl_pool.h> 131#ifdef _KERNEL 132#include <sys/dnlc.h> 133#endif 134#include <sys/callb.h> 135#include <sys/kstat.h> 136#include <sys/trim_map.h> 137#include <zfs_fletcher.h> 138#include <sys/sdt.h> 139 140#include <vm/vm_pageout.h> 141#include <machine/vmparam.h> 142 143#ifdef illumos 144#ifndef _KERNEL 145/* set with ZFS_DEBUG=watch, to enable watchpoints on frozen buffers */ 146boolean_t arc_watch = B_FALSE; 147int arc_procfd; 148#endif 149#endif /* illumos */ 150 151static kmutex_t arc_reclaim_thr_lock; 152static kcondvar_t arc_reclaim_thr_cv; /* used to signal reclaim thr */ 153static uint8_t arc_thread_exit; 154 155#define ARC_REDUCE_DNLC_PERCENT 3 156uint_t arc_reduce_dnlc_percent = ARC_REDUCE_DNLC_PERCENT; 157 158typedef enum arc_reclaim_strategy { 159 ARC_RECLAIM_AGGR, /* Aggressive reclaim strategy */ 160 ARC_RECLAIM_CONS /* Conservative reclaim strategy */ 161} arc_reclaim_strategy_t; 162 163/* 164 * The number of iterations through arc_evict_*() before we 165 * drop & reacquire the lock. 166 */ 167int arc_evict_iterations = 100; 168 169/* number of seconds before growing cache again */ 170static int arc_grow_retry = 60; 171 172/* shift of arc_c for calculating both min and max arc_p */ 173static int arc_p_min_shift = 4; 174 175/* log2(fraction of arc to reclaim) */ 176static int arc_shrink_shift = 5; 177 178/* 179 * minimum lifespan of a prefetch block in clock ticks 180 * (initialized in arc_init()) 181 */ 182static int arc_min_prefetch_lifespan; 183 184/* 185 * If this percent of memory is free, don't throttle. 186 */ 187int arc_lotsfree_percent = 10; 188 189static int arc_dead; 190extern int zfs_prefetch_disable; 191 192/* 193 * The arc has filled available memory and has now warmed up. 194 */ 195static boolean_t arc_warm; 196 197uint64_t zfs_arc_max; 198uint64_t zfs_arc_min; 199uint64_t zfs_arc_meta_limit = 0; 200uint64_t zfs_arc_meta_min = 0; 201int zfs_arc_grow_retry = 0; 202int zfs_arc_shrink_shift = 0; 203int zfs_arc_p_min_shift = 0; 204int zfs_disable_dup_eviction = 0; 205uint64_t zfs_arc_average_blocksize = 8 * 1024; /* 8KB */ 206u_int zfs_arc_free_target = 0; 207 208static int sysctl_vfs_zfs_arc_free_target(SYSCTL_HANDLER_ARGS); 209static int sysctl_vfs_zfs_arc_meta_limit(SYSCTL_HANDLER_ARGS); 210 211#ifdef _KERNEL 212static void 213arc_free_target_init(void *unused __unused) 214{ 215 216 zfs_arc_free_target = vm_pageout_wakeup_thresh; 217} 218SYSINIT(arc_free_target_init, SI_SUB_KTHREAD_PAGE, SI_ORDER_ANY, 219 arc_free_target_init, NULL); 220 221TUNABLE_QUAD("vfs.zfs.arc_meta_limit", &zfs_arc_meta_limit); 222TUNABLE_QUAD("vfs.zfs.arc_meta_min", &zfs_arc_meta_min); 223TUNABLE_INT("vfs.zfs.arc_shrink_shift", &zfs_arc_shrink_shift); 224SYSCTL_DECL(_vfs_zfs); 225SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_max, CTLFLAG_RDTUN, &zfs_arc_max, 0, 226 "Maximum ARC size"); 227SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_min, CTLFLAG_RDTUN, &zfs_arc_min, 0, 228 "Minimum ARC size"); 229SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_average_blocksize, CTLFLAG_RDTUN, 230 &zfs_arc_average_blocksize, 0, 231 "ARC average blocksize"); 232SYSCTL_INT(_vfs_zfs, OID_AUTO, arc_shrink_shift, CTLFLAG_RW, 233 &arc_shrink_shift, 0, 234 "log2(fraction of arc to reclaim)"); 235 236/* 237 * We don't have a tunable for arc_free_target due to the dependency on 238 * pagedaemon initialisation. 239 */ 240SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_free_target, 241 CTLTYPE_UINT | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, sizeof(u_int), 242 sysctl_vfs_zfs_arc_free_target, "IU", 243 "Desired number of free pages below which ARC triggers reclaim"); 244 245static int 246sysctl_vfs_zfs_arc_free_target(SYSCTL_HANDLER_ARGS) 247{ 248 u_int val; 249 int err; 250 251 val = zfs_arc_free_target; 252 err = sysctl_handle_int(oidp, &val, 0, req); 253 if (err != 0 || req->newptr == NULL) 254 return (err); 255 256 if (val < minfree) 257 return (EINVAL); 258 if (val > vm_cnt.v_page_count) 259 return (EINVAL); 260 261 zfs_arc_free_target = val; 262 263 return (0); 264} 265 266/* 267 * Must be declared here, before the definition of corresponding kstat 268 * macro which uses the same names will confuse the compiler. 269 */ 270SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_meta_limit, 271 CTLTYPE_U64 | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, sizeof(uint64_t), 272 sysctl_vfs_zfs_arc_meta_limit, "QU", 273 "ARC metadata limit"); 274#endif 275 276/* 277 * Note that buffers can be in one of 6 states: 278 * ARC_anon - anonymous (discussed below) 279 * ARC_mru - recently used, currently cached 280 * ARC_mru_ghost - recentely used, no longer in cache 281 * ARC_mfu - frequently used, currently cached 282 * ARC_mfu_ghost - frequently used, no longer in cache 283 * ARC_l2c_only - exists in L2ARC but not other states 284 * When there are no active references to the buffer, they are 285 * are linked onto a list in one of these arc states. These are 286 * the only buffers that can be evicted or deleted. Within each 287 * state there are multiple lists, one for meta-data and one for 288 * non-meta-data. Meta-data (indirect blocks, blocks of dnodes, 289 * etc.) is tracked separately so that it can be managed more 290 * explicitly: favored over data, limited explicitly. 291 * 292 * Anonymous buffers are buffers that are not associated with 293 * a DVA. These are buffers that hold dirty block copies 294 * before they are written to stable storage. By definition, 295 * they are "ref'd" and are considered part of arc_mru 296 * that cannot be freed. Generally, they will aquire a DVA 297 * as they are written and migrate onto the arc_mru list. 298 * 299 * The ARC_l2c_only state is for buffers that are in the second 300 * level ARC but no longer in any of the ARC_m* lists. The second 301 * level ARC itself may also contain buffers that are in any of 302 * the ARC_m* states - meaning that a buffer can exist in two 303 * places. The reason for the ARC_l2c_only state is to keep the 304 * buffer header in the hash table, so that reads that hit the 305 * second level ARC benefit from these fast lookups. 306 */ 307 308#define ARCS_LOCK_PAD CACHE_LINE_SIZE 309struct arcs_lock { 310 kmutex_t arcs_lock; 311#ifdef _KERNEL 312 unsigned char pad[(ARCS_LOCK_PAD - sizeof (kmutex_t))]; 313#endif 314}; 315 316/* 317 * must be power of two for mask use to work 318 * 319 */ 320#define ARC_BUFC_NUMDATALISTS 16 321#define ARC_BUFC_NUMMETADATALISTS 16 322#define ARC_BUFC_NUMLISTS (ARC_BUFC_NUMMETADATALISTS + ARC_BUFC_NUMDATALISTS) 323 324typedef struct arc_state { 325 uint64_t arcs_lsize[ARC_BUFC_NUMTYPES]; /* amount of evictable data */ 326 uint64_t arcs_size; /* total amount of data in this state */ 327 list_t arcs_lists[ARC_BUFC_NUMLISTS]; /* list of evictable buffers */ 328 struct arcs_lock arcs_locks[ARC_BUFC_NUMLISTS] __aligned(CACHE_LINE_SIZE); 329} arc_state_t; 330 331#define ARCS_LOCK(s, i) (&((s)->arcs_locks[(i)].arcs_lock)) 332 333/* The 6 states: */ 334static arc_state_t ARC_anon; 335static arc_state_t ARC_mru; 336static arc_state_t ARC_mru_ghost; 337static arc_state_t ARC_mfu; 338static arc_state_t ARC_mfu_ghost; 339static arc_state_t ARC_l2c_only; 340 341typedef struct arc_stats { 342 kstat_named_t arcstat_hits; 343 kstat_named_t arcstat_misses; 344 kstat_named_t arcstat_demand_data_hits; 345 kstat_named_t arcstat_demand_data_misses; 346 kstat_named_t arcstat_demand_metadata_hits; 347 kstat_named_t arcstat_demand_metadata_misses; 348 kstat_named_t arcstat_prefetch_data_hits; 349 kstat_named_t arcstat_prefetch_data_misses; 350 kstat_named_t arcstat_prefetch_metadata_hits; 351 kstat_named_t arcstat_prefetch_metadata_misses; 352 kstat_named_t arcstat_mru_hits; 353 kstat_named_t arcstat_mru_ghost_hits; 354 kstat_named_t arcstat_mfu_hits; 355 kstat_named_t arcstat_mfu_ghost_hits; 356 kstat_named_t arcstat_allocated; 357 kstat_named_t arcstat_deleted; 358 kstat_named_t arcstat_stolen; 359 kstat_named_t arcstat_recycle_miss; 360 /* 361 * Number of buffers that could not be evicted because the hash lock 362 * was held by another thread. The lock may not necessarily be held 363 * by something using the same buffer, since hash locks are shared 364 * by multiple buffers. 365 */ 366 kstat_named_t arcstat_mutex_miss; 367 /* 368 * Number of buffers skipped because they have I/O in progress, are 369 * indrect prefetch buffers that have not lived long enough, or are 370 * not from the spa we're trying to evict from. 371 */ 372 kstat_named_t arcstat_evict_skip; 373 kstat_named_t arcstat_evict_l2_cached; 374 kstat_named_t arcstat_evict_l2_eligible; 375 kstat_named_t arcstat_evict_l2_ineligible; 376 kstat_named_t arcstat_hash_elements; 377 kstat_named_t arcstat_hash_elements_max; 378 kstat_named_t arcstat_hash_collisions; 379 kstat_named_t arcstat_hash_chains; 380 kstat_named_t arcstat_hash_chain_max; 381 kstat_named_t arcstat_p; 382 kstat_named_t arcstat_c; 383 kstat_named_t arcstat_c_min; 384 kstat_named_t arcstat_c_max; 385 kstat_named_t arcstat_size; 386 kstat_named_t arcstat_hdr_size; 387 kstat_named_t arcstat_data_size; 388 kstat_named_t arcstat_other_size; 389 kstat_named_t arcstat_l2_hits; 390 kstat_named_t arcstat_l2_misses; 391 kstat_named_t arcstat_l2_feeds; 392 kstat_named_t arcstat_l2_rw_clash; 393 kstat_named_t arcstat_l2_read_bytes; 394 kstat_named_t arcstat_l2_write_bytes; 395 kstat_named_t arcstat_l2_writes_sent; 396 kstat_named_t arcstat_l2_writes_done; 397 kstat_named_t arcstat_l2_writes_error; 398 kstat_named_t arcstat_l2_writes_hdr_miss; 399 kstat_named_t arcstat_l2_evict_lock_retry; 400 kstat_named_t arcstat_l2_evict_reading; 401 kstat_named_t arcstat_l2_free_on_write; 402 kstat_named_t arcstat_l2_cdata_free_on_write; 403 kstat_named_t arcstat_l2_abort_lowmem; 404 kstat_named_t arcstat_l2_cksum_bad; 405 kstat_named_t arcstat_l2_io_error; 406 kstat_named_t arcstat_l2_size; 407 kstat_named_t arcstat_l2_asize; 408 kstat_named_t arcstat_l2_hdr_size; 409 kstat_named_t arcstat_l2_compress_successes; 410 kstat_named_t arcstat_l2_compress_zeros; 411 kstat_named_t arcstat_l2_compress_failures; 412 kstat_named_t arcstat_l2_write_trylock_fail; 413 kstat_named_t arcstat_l2_write_passed_headroom; 414 kstat_named_t arcstat_l2_write_spa_mismatch; 415 kstat_named_t arcstat_l2_write_in_l2; 416 kstat_named_t arcstat_l2_write_hdr_io_in_progress; 417 kstat_named_t arcstat_l2_write_not_cacheable; 418 kstat_named_t arcstat_l2_write_full; 419 kstat_named_t arcstat_l2_write_buffer_iter; 420 kstat_named_t arcstat_l2_write_pios; 421 kstat_named_t arcstat_l2_write_buffer_bytes_scanned; 422 kstat_named_t arcstat_l2_write_buffer_list_iter; 423 kstat_named_t arcstat_l2_write_buffer_list_null_iter; 424 kstat_named_t arcstat_memory_throttle_count; 425 kstat_named_t arcstat_duplicate_buffers; 426 kstat_named_t arcstat_duplicate_buffers_size; 427 kstat_named_t arcstat_duplicate_reads; 428 kstat_named_t arcstat_meta_used; 429 kstat_named_t arcstat_meta_limit; 430 kstat_named_t arcstat_meta_max; 431 kstat_named_t arcstat_meta_min; 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 { "arc_meta_min", KSTAT_DATA_UINT64 } 514}; 515 516#define ARCSTAT(stat) (arc_stats.stat.value.ui64) 517 518#define ARCSTAT_INCR(stat, val) \ 519 atomic_add_64(&arc_stats.stat.value.ui64, (val)) 520 521#define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1) 522#define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1) 523 524#define ARCSTAT_MAX(stat, val) { \ 525 uint64_t m; \ 526 while ((val) > (m = arc_stats.stat.value.ui64) && \ 527 (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \ 528 continue; \ 529} 530 531#define ARCSTAT_MAXSTAT(stat) \ 532 ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64) 533 534/* 535 * We define a macro to allow ARC hits/misses to be easily broken down by 536 * two separate conditions, giving a total of four different subtypes for 537 * each of hits and misses (so eight statistics total). 538 */ 539#define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \ 540 if (cond1) { \ 541 if (cond2) { \ 542 ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \ 543 } else { \ 544 ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \ 545 } \ 546 } else { \ 547 if (cond2) { \ 548 ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \ 549 } else { \ 550 ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\ 551 } \ 552 } 553 554kstat_t *arc_ksp; 555static arc_state_t *arc_anon; 556static arc_state_t *arc_mru; 557static arc_state_t *arc_mru_ghost; 558static arc_state_t *arc_mfu; 559static arc_state_t *arc_mfu_ghost; 560static arc_state_t *arc_l2c_only; 561 562/* 563 * There are several ARC variables that are critical to export as kstats -- 564 * but we don't want to have to grovel around in the kstat whenever we wish to 565 * manipulate them. For these variables, we therefore define them to be in 566 * terms of the statistic variable. This assures that we are not introducing 567 * the possibility of inconsistency by having shadow copies of the variables, 568 * while still allowing the code to be readable. 569 */ 570#define arc_size ARCSTAT(arcstat_size) /* actual total arc size */ 571#define arc_p ARCSTAT(arcstat_p) /* target size of MRU */ 572#define arc_c ARCSTAT(arcstat_c) /* target size of cache */ 573#define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */ 574#define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */ 575#define arc_meta_limit ARCSTAT(arcstat_meta_limit) /* max size for metadata */ 576#define arc_meta_min ARCSTAT(arcstat_meta_min) /* min size for metadata */ 577#define arc_meta_used ARCSTAT(arcstat_meta_used) /* size of metadata */ 578#define arc_meta_max ARCSTAT(arcstat_meta_max) /* max size of metadata */ 579 580#define L2ARC_IS_VALID_COMPRESS(_c_) \ 581 ((_c_) == ZIO_COMPRESS_LZ4 || (_c_) == ZIO_COMPRESS_EMPTY) 582 583static int arc_no_grow; /* Don't try to grow cache size */ 584static uint64_t arc_tempreserve; 585static uint64_t arc_loaned_bytes; 586 587typedef struct l2arc_buf_hdr l2arc_buf_hdr_t; 588 589typedef struct arc_callback arc_callback_t; 590 591struct arc_callback { 592 void *acb_private; 593 arc_done_func_t *acb_done; 594 arc_buf_t *acb_buf; 595 zio_t *acb_zio_dummy; 596 arc_callback_t *acb_next; 597}; 598 599typedef struct arc_write_callback arc_write_callback_t; 600 601struct arc_write_callback { 602 void *awcb_private; 603 arc_done_func_t *awcb_ready; 604 arc_done_func_t *awcb_physdone; 605 arc_done_func_t *awcb_done; 606 arc_buf_t *awcb_buf; 607}; 608 609struct arc_buf_hdr { 610 /* protected by hash lock */ 611 dva_t b_dva; 612 uint64_t b_birth; 613 uint64_t b_cksum0; 614 615 kmutex_t b_freeze_lock; 616 zio_cksum_t *b_freeze_cksum; 617 void *b_thawed; 618 619 arc_buf_hdr_t *b_hash_next; 620 arc_buf_t *b_buf; 621 arc_flags_t b_flags; 622 uint32_t b_datacnt; 623 624 arc_callback_t *b_acb; 625 kcondvar_t b_cv; 626 627 /* immutable */ 628 arc_buf_contents_t b_type; 629 uint64_t b_size; 630 uint64_t b_spa; 631 632 /* protected by arc state mutex */ 633 arc_state_t *b_state; 634 list_node_t b_arc_node; 635 636 /* updated atomically */ 637 clock_t b_arc_access; 638 639 /* self protecting */ 640 refcount_t b_refcnt; 641 642 l2arc_buf_hdr_t *b_l2hdr; 643 list_node_t b_l2node; 644}; 645 646#ifdef _KERNEL 647static int 648sysctl_vfs_zfs_arc_meta_limit(SYSCTL_HANDLER_ARGS) 649{ 650 uint64_t val; 651 int err; 652 653 val = arc_meta_limit; 654 err = sysctl_handle_64(oidp, &val, 0, req); 655 if (err != 0 || req->newptr == NULL) 656 return (err); 657 658 if (val <= 0 || val > arc_c_max) 659 return (EINVAL); 660 661 arc_meta_limit = val; 662 return (0); 663} 664#endif 665 666static arc_buf_t *arc_eviction_list; 667static kmutex_t arc_eviction_mtx; 668static arc_buf_hdr_t arc_eviction_hdr; 669 670#define GHOST_STATE(state) \ 671 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \ 672 (state) == arc_l2c_only) 673 674#define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_FLAG_IN_HASH_TABLE) 675#define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) 676#define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_FLAG_IO_ERROR) 677#define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_FLAG_PREFETCH) 678#define HDR_FREED_IN_READ(hdr) ((hdr)->b_flags & ARC_FLAG_FREED_IN_READ) 679#define HDR_BUF_AVAILABLE(hdr) ((hdr)->b_flags & ARC_FLAG_BUF_AVAILABLE) 680#define HDR_FREE_IN_PROGRESS(hdr) \ 681 ((hdr)->b_flags & ARC_FLAG_FREE_IN_PROGRESS) 682#define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_FLAG_L2CACHE) 683#define HDR_L2_READING(hdr) \ 684 ((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS && \ 685 (hdr)->b_l2hdr != NULL) 686#define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITING) 687#define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_FLAG_L2_EVICTED) 688#define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITE_HEAD) 689 690/* 691 * Other sizes 692 */ 693 694#define HDR_SIZE ((int64_t)sizeof (arc_buf_hdr_t)) 695#define L2HDR_SIZE ((int64_t)sizeof (l2arc_buf_hdr_t)) 696 697/* 698 * Hash table routines 699 */ 700 701#define HT_LOCK_PAD CACHE_LINE_SIZE 702 703struct ht_lock { 704 kmutex_t ht_lock; 705#ifdef _KERNEL 706 unsigned char pad[(HT_LOCK_PAD - sizeof (kmutex_t))]; 707#endif 708}; 709 710#define BUF_LOCKS 256 711typedef struct buf_hash_table { 712 uint64_t ht_mask; 713 arc_buf_hdr_t **ht_table; 714 struct ht_lock ht_locks[BUF_LOCKS] __aligned(CACHE_LINE_SIZE); 715} buf_hash_table_t; 716 717static buf_hash_table_t buf_hash_table; 718 719#define BUF_HASH_INDEX(spa, dva, birth) \ 720 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask) 721#define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)]) 722#define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock)) 723#define HDR_LOCK(hdr) \ 724 (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth))) 725 726uint64_t zfs_crc64_table[256]; 727 728/* 729 * Level 2 ARC 730 */ 731 732#define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */ 733#define L2ARC_HEADROOM 2 /* num of writes */ 734/* 735 * If we discover during ARC scan any buffers to be compressed, we boost 736 * our headroom for the next scanning cycle by this percentage multiple. 737 */ 738#define L2ARC_HEADROOM_BOOST 200 739#define L2ARC_FEED_SECS 1 /* caching interval secs */ 740#define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */ 741 742#define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent) 743#define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done) 744 745/* L2ARC Performance Tunables */ 746uint64_t l2arc_write_max = L2ARC_WRITE_SIZE; /* default max write size */ 747uint64_t l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra write during warmup */ 748uint64_t l2arc_headroom = L2ARC_HEADROOM; /* number of dev writes */ 749uint64_t l2arc_headroom_boost = L2ARC_HEADROOM_BOOST; 750uint64_t l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */ 751uint64_t l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval milliseconds */ 752boolean_t l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */ 753boolean_t l2arc_feed_again = B_TRUE; /* turbo warmup */ 754boolean_t l2arc_norw = B_TRUE; /* no reads during writes */ 755 756SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_max, CTLFLAG_RW, 757 &l2arc_write_max, 0, "max write size"); 758SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_boost, CTLFLAG_RW, 759 &l2arc_write_boost, 0, "extra write during warmup"); 760SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_headroom, CTLFLAG_RW, 761 &l2arc_headroom, 0, "number of dev writes"); 762SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_secs, CTLFLAG_RW, 763 &l2arc_feed_secs, 0, "interval seconds"); 764SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_min_ms, CTLFLAG_RW, 765 &l2arc_feed_min_ms, 0, "min interval milliseconds"); 766 767SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_noprefetch, CTLFLAG_RW, 768 &l2arc_noprefetch, 0, "don't cache prefetch bufs"); 769SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_feed_again, CTLFLAG_RW, 770 &l2arc_feed_again, 0, "turbo warmup"); 771SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_norw, CTLFLAG_RW, 772 &l2arc_norw, 0, "no reads during writes"); 773 774SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_size, CTLFLAG_RD, 775 &ARC_anon.arcs_size, 0, "size of anonymous state"); 776SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_metadata_lsize, CTLFLAG_RD, 777 &ARC_anon.arcs_lsize[ARC_BUFC_METADATA], 0, "size of anonymous state"); 778SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_data_lsize, CTLFLAG_RD, 779 &ARC_anon.arcs_lsize[ARC_BUFC_DATA], 0, "size of anonymous state"); 780 781SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_size, CTLFLAG_RD, 782 &ARC_mru.arcs_size, 0, "size of mru state"); 783SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_metadata_lsize, CTLFLAG_RD, 784 &ARC_mru.arcs_lsize[ARC_BUFC_METADATA], 0, "size of metadata in mru state"); 785SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_data_lsize, CTLFLAG_RD, 786 &ARC_mru.arcs_lsize[ARC_BUFC_DATA], 0, "size of data in mru state"); 787 788SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_size, CTLFLAG_RD, 789 &ARC_mru_ghost.arcs_size, 0, "size of mru ghost state"); 790SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_metadata_lsize, CTLFLAG_RD, 791 &ARC_mru_ghost.arcs_lsize[ARC_BUFC_METADATA], 0, 792 "size of metadata in mru ghost state"); 793SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_data_lsize, CTLFLAG_RD, 794 &ARC_mru_ghost.arcs_lsize[ARC_BUFC_DATA], 0, 795 "size of data in mru ghost state"); 796 797SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_size, CTLFLAG_RD, 798 &ARC_mfu.arcs_size, 0, "size of mfu state"); 799SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_metadata_lsize, CTLFLAG_RD, 800 &ARC_mfu.arcs_lsize[ARC_BUFC_METADATA], 0, "size of metadata in mfu state"); 801SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_data_lsize, CTLFLAG_RD, 802 &ARC_mfu.arcs_lsize[ARC_BUFC_DATA], 0, "size of data in mfu state"); 803 804SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_size, CTLFLAG_RD, 805 &ARC_mfu_ghost.arcs_size, 0, "size of mfu ghost state"); 806SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_metadata_lsize, CTLFLAG_RD, 807 &ARC_mfu_ghost.arcs_lsize[ARC_BUFC_METADATA], 0, 808 "size of metadata in mfu ghost state"); 809SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_data_lsize, CTLFLAG_RD, 810 &ARC_mfu_ghost.arcs_lsize[ARC_BUFC_DATA], 0, 811 "size of data in mfu ghost state"); 812 813SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2c_only_size, CTLFLAG_RD, 814 &ARC_l2c_only.arcs_size, 0, "size of mru state"); 815 816/* 817 * L2ARC Internals 818 */ 819typedef struct l2arc_dev { 820 vdev_t *l2ad_vdev; /* vdev */ 821 spa_t *l2ad_spa; /* spa */ 822 uint64_t l2ad_hand; /* next write location */ 823 uint64_t l2ad_start; /* first addr on device */ 824 uint64_t l2ad_end; /* last addr on device */ 825 uint64_t l2ad_evict; /* last addr eviction reached */ 826 boolean_t l2ad_first; /* first sweep through */ 827 boolean_t l2ad_writing; /* currently writing */ 828 list_t *l2ad_buflist; /* buffer list */ 829 list_node_t l2ad_node; /* device list node */ 830} l2arc_dev_t; 831 832static list_t L2ARC_dev_list; /* device list */ 833static list_t *l2arc_dev_list; /* device list pointer */ 834static kmutex_t l2arc_dev_mtx; /* device list mutex */ 835static l2arc_dev_t *l2arc_dev_last; /* last device used */ 836static kmutex_t l2arc_buflist_mtx; /* mutex for all buflists */ 837static list_t L2ARC_free_on_write; /* free after write buf list */ 838static list_t *l2arc_free_on_write; /* free after write list ptr */ 839static kmutex_t l2arc_free_on_write_mtx; /* mutex for list */ 840static uint64_t l2arc_ndev; /* number of devices */ 841 842typedef struct l2arc_read_callback { 843 arc_buf_t *l2rcb_buf; /* read buffer */ 844 spa_t *l2rcb_spa; /* spa */ 845 blkptr_t l2rcb_bp; /* original blkptr */ 846 zbookmark_phys_t l2rcb_zb; /* original bookmark */ 847 int l2rcb_flags; /* original flags */ 848 enum zio_compress l2rcb_compress; /* applied compress */ 849} l2arc_read_callback_t; 850 851typedef struct l2arc_write_callback { 852 l2arc_dev_t *l2wcb_dev; /* device info */ 853 arc_buf_hdr_t *l2wcb_head; /* head of write buflist */ 854} l2arc_write_callback_t; 855 856struct l2arc_buf_hdr { 857 /* protected by arc_buf_hdr mutex */ 858 l2arc_dev_t *b_dev; /* L2ARC device */ 859 uint64_t b_daddr; /* disk address, offset byte */ 860 /* compression applied to buffer data */ 861 enum zio_compress b_compress; 862 /* real alloc'd buffer size depending on b_compress applied */ 863 int b_asize; 864 /* temporary buffer holder for in-flight compressed data */ 865 void *b_tmp_cdata; 866}; 867 868typedef struct l2arc_data_free { 869 /* protected by l2arc_free_on_write_mtx */ 870 void *l2df_data; 871 size_t l2df_size; 872 void (*l2df_func)(void *, size_t); 873 list_node_t l2df_list_node; 874} l2arc_data_free_t; 875 876static kmutex_t l2arc_feed_thr_lock; 877static kcondvar_t l2arc_feed_thr_cv; 878static uint8_t l2arc_thread_exit; 879 880static void arc_get_data_buf(arc_buf_t *); 881static void arc_access(arc_buf_hdr_t *, kmutex_t *); 882static int arc_evict_needed(arc_buf_contents_t); 883static void arc_evict_ghost(arc_state_t *, uint64_t, int64_t); 884static void arc_buf_watch(arc_buf_t *); 885 886static boolean_t l2arc_write_eligible(uint64_t, arc_buf_hdr_t *); 887static void l2arc_read_done(zio_t *); 888static void l2arc_hdr_stat_add(void); 889static void l2arc_hdr_stat_remove(void); 890 891static boolean_t l2arc_compress_buf(l2arc_buf_hdr_t *); 892static void l2arc_decompress_zio(zio_t *, arc_buf_hdr_t *, enum zio_compress); 893static void l2arc_release_cdata_buf(arc_buf_hdr_t *); 894 895static uint64_t 896buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth) 897{ 898 uint8_t *vdva = (uint8_t *)dva; 899 uint64_t crc = -1ULL; 900 int i; 901 902 ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY); 903 904 for (i = 0; i < sizeof (dva_t); i++) 905 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ vdva[i]) & 0xFF]; 906 907 crc ^= (spa>>8) ^ birth; 908 909 return (crc); 910} 911 912#define BUF_EMPTY(buf) \ 913 ((buf)->b_dva.dva_word[0] == 0 && \ 914 (buf)->b_dva.dva_word[1] == 0 && \ 915 (buf)->b_cksum0 == 0) 916 917#define BUF_EQUAL(spa, dva, birth, buf) \ 918 ((buf)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \ 919 ((buf)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \ 920 ((buf)->b_birth == birth) && ((buf)->b_spa == spa) 921 922static void 923buf_discard_identity(arc_buf_hdr_t *hdr) 924{ 925 hdr->b_dva.dva_word[0] = 0; 926 hdr->b_dva.dva_word[1] = 0; 927 hdr->b_birth = 0; 928 hdr->b_cksum0 = 0; 929} 930 931static arc_buf_hdr_t * 932buf_hash_find(uint64_t spa, const blkptr_t *bp, kmutex_t **lockp) 933{ 934 const dva_t *dva = BP_IDENTITY(bp); 935 uint64_t birth = BP_PHYSICAL_BIRTH(bp); 936 uint64_t idx = BUF_HASH_INDEX(spa, dva, birth); 937 kmutex_t *hash_lock = BUF_HASH_LOCK(idx); 938 arc_buf_hdr_t *hdr; 939 940 mutex_enter(hash_lock); 941 for (hdr = buf_hash_table.ht_table[idx]; hdr != NULL; 942 hdr = hdr->b_hash_next) { 943 if (BUF_EQUAL(spa, dva, birth, hdr)) { 944 *lockp = hash_lock; 945 return (hdr); 946 } 947 } 948 mutex_exit(hash_lock); 949 *lockp = NULL; 950 return (NULL); 951} 952 953/* 954 * Insert an entry into the hash table. If there is already an element 955 * equal to elem in the hash table, then the already existing element 956 * will be returned and the new element will not be inserted. 957 * Otherwise returns NULL. 958 */ 959static arc_buf_hdr_t * 960buf_hash_insert(arc_buf_hdr_t *hdr, kmutex_t **lockp) 961{ 962 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth); 963 kmutex_t *hash_lock = BUF_HASH_LOCK(idx); 964 arc_buf_hdr_t *fhdr; 965 uint32_t i; 966 967 ASSERT(!DVA_IS_EMPTY(&hdr->b_dva)); 968 ASSERT(hdr->b_birth != 0); 969 ASSERT(!HDR_IN_HASH_TABLE(hdr)); 970 *lockp = hash_lock; 971 mutex_enter(hash_lock); 972 for (fhdr = buf_hash_table.ht_table[idx], i = 0; fhdr != NULL; 973 fhdr = fhdr->b_hash_next, i++) { 974 if (BUF_EQUAL(hdr->b_spa, &hdr->b_dva, hdr->b_birth, fhdr)) 975 return (fhdr); 976 } 977 978 hdr->b_hash_next = buf_hash_table.ht_table[idx]; 979 buf_hash_table.ht_table[idx] = hdr; 980 hdr->b_flags |= ARC_FLAG_IN_HASH_TABLE; 981 982 /* collect some hash table performance data */ 983 if (i > 0) { 984 ARCSTAT_BUMP(arcstat_hash_collisions); 985 if (i == 1) 986 ARCSTAT_BUMP(arcstat_hash_chains); 987 988 ARCSTAT_MAX(arcstat_hash_chain_max, i); 989 } 990 991 ARCSTAT_BUMP(arcstat_hash_elements); 992 ARCSTAT_MAXSTAT(arcstat_hash_elements); 993 994 return (NULL); 995} 996 997static void 998buf_hash_remove(arc_buf_hdr_t *hdr) 999{ 1000 arc_buf_hdr_t *fhdr, **hdrp; 1001 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth); 1002 1003 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx))); 1004 ASSERT(HDR_IN_HASH_TABLE(hdr)); 1005 1006 hdrp = &buf_hash_table.ht_table[idx]; 1007 while ((fhdr = *hdrp) != hdr) { 1008 ASSERT(fhdr != NULL); 1009 hdrp = &fhdr->b_hash_next; 1010 } 1011 *hdrp = hdr->b_hash_next; 1012 hdr->b_hash_next = NULL; 1013 hdr->b_flags &= ~ARC_FLAG_IN_HASH_TABLE; 1014 1015 /* collect some hash table performance data */ 1016 ARCSTAT_BUMPDOWN(arcstat_hash_elements); 1017 1018 if (buf_hash_table.ht_table[idx] && 1019 buf_hash_table.ht_table[idx]->b_hash_next == NULL) 1020 ARCSTAT_BUMPDOWN(arcstat_hash_chains); 1021} 1022 1023/* 1024 * Global data structures and functions for the buf kmem cache. 1025 */ 1026static kmem_cache_t *hdr_cache; 1027static kmem_cache_t *buf_cache; 1028 1029static void 1030buf_fini(void) 1031{ 1032 int i; 1033 1034 kmem_free(buf_hash_table.ht_table, 1035 (buf_hash_table.ht_mask + 1) * sizeof (void *)); 1036 for (i = 0; i < BUF_LOCKS; i++) 1037 mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock); 1038 kmem_cache_destroy(hdr_cache); 1039 kmem_cache_destroy(buf_cache); 1040} 1041 1042/* 1043 * Constructor callback - called when the cache is empty 1044 * and a new buf is requested. 1045 */ 1046/* ARGSUSED */ 1047static int 1048hdr_cons(void *vbuf, void *unused, int kmflag) 1049{ 1050 arc_buf_hdr_t *hdr = vbuf; 1051 1052 bzero(hdr, sizeof (arc_buf_hdr_t)); 1053 refcount_create(&hdr->b_refcnt); 1054 cv_init(&hdr->b_cv, NULL, CV_DEFAULT, NULL); 1055 mutex_init(&hdr->b_freeze_lock, NULL, MUTEX_DEFAULT, NULL); 1056 arc_space_consume(sizeof (arc_buf_hdr_t), ARC_SPACE_HDRS); 1057 1058 return (0); 1059} 1060 1061/* ARGSUSED */ 1062static int 1063buf_cons(void *vbuf, void *unused, int kmflag) 1064{ 1065 arc_buf_t *buf = vbuf; 1066 1067 bzero(buf, sizeof (arc_buf_t)); 1068 mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL); 1069 arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS); 1070 1071 return (0); 1072} 1073 1074/* 1075 * Destructor callback - called when a cached buf is 1076 * no longer required. 1077 */ 1078/* ARGSUSED */ 1079static void 1080hdr_dest(void *vbuf, void *unused) 1081{ 1082 arc_buf_hdr_t *hdr = vbuf; 1083 1084 ASSERT(BUF_EMPTY(hdr)); 1085 refcount_destroy(&hdr->b_refcnt); 1086 cv_destroy(&hdr->b_cv); 1087 mutex_destroy(&hdr->b_freeze_lock); 1088 arc_space_return(sizeof (arc_buf_hdr_t), ARC_SPACE_HDRS); 1089} 1090 1091/* ARGSUSED */ 1092static void 1093buf_dest(void *vbuf, void *unused) 1094{ 1095 arc_buf_t *buf = vbuf; 1096 1097 mutex_destroy(&buf->b_evict_lock); 1098 arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS); 1099} 1100 1101/* 1102 * Reclaim callback -- invoked when memory is low. 1103 */ 1104/* ARGSUSED */ 1105static void 1106hdr_recl(void *unused) 1107{ 1108 dprintf("hdr_recl called\n"); 1109 /* 1110 * umem calls the reclaim func when we destroy the buf cache, 1111 * which is after we do arc_fini(). 1112 */ 1113 if (!arc_dead) 1114 cv_signal(&arc_reclaim_thr_cv); 1115} 1116 1117static void 1118buf_init(void) 1119{ 1120 uint64_t *ct; 1121 uint64_t hsize = 1ULL << 12; 1122 int i, j; 1123 1124 /* 1125 * The hash table is big enough to fill all of physical memory 1126 * with an average block size of zfs_arc_average_blocksize (default 8K). 1127 * By default, the table will take up 1128 * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers). 1129 */ 1130 while (hsize * zfs_arc_average_blocksize < (uint64_t)physmem * PAGESIZE) 1131 hsize <<= 1; 1132retry: 1133 buf_hash_table.ht_mask = hsize - 1; 1134 buf_hash_table.ht_table = 1135 kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP); 1136 if (buf_hash_table.ht_table == NULL) { 1137 ASSERT(hsize > (1ULL << 8)); 1138 hsize >>= 1; 1139 goto retry; 1140 } 1141 1142 hdr_cache = kmem_cache_create("arc_buf_hdr_t", sizeof (arc_buf_hdr_t), 1143 0, hdr_cons, hdr_dest, hdr_recl, NULL, NULL, 0); 1144 buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t), 1145 0, buf_cons, buf_dest, NULL, NULL, NULL, 0); 1146 1147 for (i = 0; i < 256; i++) 1148 for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--) 1149 *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY); 1150 1151 for (i = 0; i < BUF_LOCKS; i++) { 1152 mutex_init(&buf_hash_table.ht_locks[i].ht_lock, 1153 NULL, MUTEX_DEFAULT, NULL); 1154 } 1155} 1156 1157#define ARC_MINTIME (hz>>4) /* 62 ms */ 1158 1159static void 1160arc_cksum_verify(arc_buf_t *buf) 1161{ 1162 zio_cksum_t zc; 1163 1164 if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 1165 return; 1166 1167 mutex_enter(&buf->b_hdr->b_freeze_lock); 1168 if (buf->b_hdr->b_freeze_cksum == NULL || 1169 (buf->b_hdr->b_flags & ARC_FLAG_IO_ERROR)) { 1170 mutex_exit(&buf->b_hdr->b_freeze_lock); 1171 return; 1172 } 1173 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc); 1174 if (!ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc)) 1175 panic("buffer modified while frozen!"); 1176 mutex_exit(&buf->b_hdr->b_freeze_lock); 1177} 1178 1179static int 1180arc_cksum_equal(arc_buf_t *buf) 1181{ 1182 zio_cksum_t zc; 1183 int equal; 1184 1185 mutex_enter(&buf->b_hdr->b_freeze_lock); 1186 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc); 1187 equal = ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc); 1188 mutex_exit(&buf->b_hdr->b_freeze_lock); 1189 1190 return (equal); 1191} 1192 1193static void 1194arc_cksum_compute(arc_buf_t *buf, boolean_t force) 1195{ 1196 if (!force && !(zfs_flags & ZFS_DEBUG_MODIFY)) 1197 return; 1198 1199 mutex_enter(&buf->b_hdr->b_freeze_lock); 1200 if (buf->b_hdr->b_freeze_cksum != NULL) { 1201 mutex_exit(&buf->b_hdr->b_freeze_lock); 1202 return; 1203 } 1204 buf->b_hdr->b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t), KM_SLEEP); 1205 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, 1206 buf->b_hdr->b_freeze_cksum); 1207 mutex_exit(&buf->b_hdr->b_freeze_lock); 1208#ifdef illumos 1209 arc_buf_watch(buf); 1210#endif 1211} 1212 1213#ifdef illumos 1214#ifndef _KERNEL 1215typedef struct procctl { 1216 long cmd; 1217 prwatch_t prwatch; 1218} procctl_t; 1219#endif 1220 1221/* ARGSUSED */ 1222static void 1223arc_buf_unwatch(arc_buf_t *buf) 1224{ 1225#ifndef _KERNEL 1226 if (arc_watch) { 1227 int result; 1228 procctl_t ctl; 1229 ctl.cmd = PCWATCH; 1230 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data; 1231 ctl.prwatch.pr_size = 0; 1232 ctl.prwatch.pr_wflags = 0; 1233 result = write(arc_procfd, &ctl, sizeof (ctl)); 1234 ASSERT3U(result, ==, sizeof (ctl)); 1235 } 1236#endif 1237} 1238 1239/* ARGSUSED */ 1240static void 1241arc_buf_watch(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 = buf->b_hdr->b_size; 1250 ctl.prwatch.pr_wflags = WA_WRITE; 1251 result = write(arc_procfd, &ctl, sizeof (ctl)); 1252 ASSERT3U(result, ==, sizeof (ctl)); 1253 } 1254#endif 1255} 1256#endif /* illumos */ 1257 1258void 1259arc_buf_thaw(arc_buf_t *buf) 1260{ 1261 if (zfs_flags & ZFS_DEBUG_MODIFY) { 1262 if (buf->b_hdr->b_state != arc_anon) 1263 panic("modifying non-anon buffer!"); 1264 if (buf->b_hdr->b_flags & ARC_FLAG_IO_IN_PROGRESS) 1265 panic("modifying buffer while i/o in progress!"); 1266 arc_cksum_verify(buf); 1267 } 1268 1269 mutex_enter(&buf->b_hdr->b_freeze_lock); 1270 if (buf->b_hdr->b_freeze_cksum != NULL) { 1271 kmem_free(buf->b_hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 1272 buf->b_hdr->b_freeze_cksum = NULL; 1273 } 1274 1275 if (zfs_flags & ZFS_DEBUG_MODIFY) { 1276 if (buf->b_hdr->b_thawed) 1277 kmem_free(buf->b_hdr->b_thawed, 1); 1278 buf->b_hdr->b_thawed = kmem_alloc(1, KM_SLEEP); 1279 } 1280 1281 mutex_exit(&buf->b_hdr->b_freeze_lock); 1282 1283#ifdef illumos 1284 arc_buf_unwatch(buf); 1285#endif 1286} 1287 1288void 1289arc_buf_freeze(arc_buf_t *buf) 1290{ 1291 kmutex_t *hash_lock; 1292 1293 if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 1294 return; 1295 1296 hash_lock = HDR_LOCK(buf->b_hdr); 1297 mutex_enter(hash_lock); 1298 1299 ASSERT(buf->b_hdr->b_freeze_cksum != NULL || 1300 buf->b_hdr->b_state == arc_anon); 1301 arc_cksum_compute(buf, B_FALSE); 1302 mutex_exit(hash_lock); 1303 1304} 1305 1306static void 1307get_buf_info(arc_buf_hdr_t *hdr, arc_state_t *state, list_t **list, kmutex_t **lock) 1308{ 1309 uint64_t buf_hashid = buf_hash(hdr->b_spa, &hdr->b_dva, hdr->b_birth); 1310 1311 if (hdr->b_type == ARC_BUFC_METADATA) 1312 buf_hashid &= (ARC_BUFC_NUMMETADATALISTS - 1); 1313 else { 1314 buf_hashid &= (ARC_BUFC_NUMDATALISTS - 1); 1315 buf_hashid += ARC_BUFC_NUMMETADATALISTS; 1316 } 1317 1318 *list = &state->arcs_lists[buf_hashid]; 1319 *lock = ARCS_LOCK(state, buf_hashid); 1320} 1321 1322 1323static void 1324add_reference(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, void *tag) 1325{ 1326 ASSERT(MUTEX_HELD(hash_lock)); 1327 1328 if ((refcount_add(&hdr->b_refcnt, tag) == 1) && 1329 (hdr->b_state != arc_anon)) { 1330 uint64_t delta = hdr->b_size * hdr->b_datacnt; 1331 uint64_t *size = &hdr->b_state->arcs_lsize[hdr->b_type]; 1332 list_t *list; 1333 kmutex_t *lock; 1334 1335 get_buf_info(hdr, hdr->b_state, &list, &lock); 1336 ASSERT(!MUTEX_HELD(lock)); 1337 mutex_enter(lock); 1338 ASSERT(list_link_active(&hdr->b_arc_node)); 1339 list_remove(list, hdr); 1340 if (GHOST_STATE(hdr->b_state)) { 1341 ASSERT0(hdr->b_datacnt); 1342 ASSERT3P(hdr->b_buf, ==, NULL); 1343 delta = hdr->b_size; 1344 } 1345 ASSERT(delta > 0); 1346 ASSERT3U(*size, >=, delta); 1347 atomic_add_64(size, -delta); 1348 mutex_exit(lock); 1349 /* remove the prefetch flag if we get a reference */ 1350 if (hdr->b_flags & ARC_FLAG_PREFETCH) 1351 hdr->b_flags &= ~ARC_FLAG_PREFETCH; 1352 } 1353} 1354 1355static int 1356remove_reference(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, void *tag) 1357{ 1358 int cnt; 1359 arc_state_t *state = hdr->b_state; 1360 1361 ASSERT(state == arc_anon || MUTEX_HELD(hash_lock)); 1362 ASSERT(!GHOST_STATE(state)); 1363 1364 if (((cnt = refcount_remove(&hdr->b_refcnt, tag)) == 0) && 1365 (state != arc_anon)) { 1366 uint64_t *size = &state->arcs_lsize[hdr->b_type]; 1367 list_t *list; 1368 kmutex_t *lock; 1369 1370 get_buf_info(hdr, state, &list, &lock); 1371 ASSERT(!MUTEX_HELD(lock)); 1372 mutex_enter(lock); 1373 ASSERT(!list_link_active(&hdr->b_arc_node)); 1374 list_insert_head(list, hdr); 1375 ASSERT(hdr->b_datacnt > 0); 1376 atomic_add_64(size, hdr->b_size * hdr->b_datacnt); 1377 mutex_exit(lock); 1378 } 1379 return (cnt); 1380} 1381 1382/* 1383 * Move the supplied buffer to the indicated state. The mutex 1384 * for the buffer must be held by the caller. 1385 */ 1386static void 1387arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *hdr, 1388 kmutex_t *hash_lock) 1389{ 1390 arc_state_t *old_state = hdr->b_state; 1391 int64_t refcnt = refcount_count(&hdr->b_refcnt); 1392 uint64_t from_delta, to_delta; 1393 list_t *list; 1394 kmutex_t *lock; 1395 1396 ASSERT(MUTEX_HELD(hash_lock)); 1397 ASSERT3P(new_state, !=, old_state); 1398 ASSERT(refcnt == 0 || hdr->b_datacnt > 0); 1399 ASSERT(hdr->b_datacnt == 0 || !GHOST_STATE(new_state)); 1400 ASSERT(hdr->b_datacnt <= 1 || old_state != arc_anon); 1401 1402 from_delta = to_delta = hdr->b_datacnt * hdr->b_size; 1403 1404 /* 1405 * If this buffer is evictable, transfer it from the 1406 * old state list to the new state list. 1407 */ 1408 if (refcnt == 0) { 1409 if (old_state != arc_anon) { 1410 int use_mutex; 1411 uint64_t *size = &old_state->arcs_lsize[hdr->b_type]; 1412 1413 get_buf_info(hdr, old_state, &list, &lock); 1414 use_mutex = !MUTEX_HELD(lock); 1415 if (use_mutex) 1416 mutex_enter(lock); 1417 1418 ASSERT(list_link_active(&hdr->b_arc_node)); 1419 list_remove(list, hdr); 1420 1421 /* 1422 * If prefetching out of the ghost cache, 1423 * we will have a non-zero datacnt. 1424 */ 1425 if (GHOST_STATE(old_state) && hdr->b_datacnt == 0) { 1426 /* ghost elements have a ghost size */ 1427 ASSERT(hdr->b_buf == NULL); 1428 from_delta = hdr->b_size; 1429 } 1430 ASSERT3U(*size, >=, from_delta); 1431 atomic_add_64(size, -from_delta); 1432 1433 if (use_mutex) 1434 mutex_exit(lock); 1435 } 1436 if (new_state != arc_anon) { 1437 int use_mutex; 1438 uint64_t *size = &new_state->arcs_lsize[hdr->b_type]; 1439 1440 get_buf_info(hdr, new_state, &list, &lock); 1441 use_mutex = !MUTEX_HELD(lock); 1442 if (use_mutex) 1443 mutex_enter(lock); 1444 1445 list_insert_head(list, hdr); 1446 1447 /* ghost elements have a ghost size */ 1448 if (GHOST_STATE(new_state)) { 1449 ASSERT(hdr->b_datacnt == 0); 1450 ASSERT(hdr->b_buf == NULL); 1451 to_delta = hdr->b_size; 1452 } 1453 atomic_add_64(size, to_delta); 1454 1455 if (use_mutex) 1456 mutex_exit(lock); 1457 } 1458 } 1459 1460 ASSERT(!BUF_EMPTY(hdr)); 1461 if (new_state == arc_anon && HDR_IN_HASH_TABLE(hdr)) 1462 buf_hash_remove(hdr); 1463 1464 /* adjust state sizes */ 1465 if (to_delta) 1466 atomic_add_64(&new_state->arcs_size, to_delta); 1467 if (from_delta) { 1468 ASSERT3U(old_state->arcs_size, >=, from_delta); 1469 atomic_add_64(&old_state->arcs_size, -from_delta); 1470 } 1471 hdr->b_state = new_state; 1472 1473 /* adjust l2arc hdr stats */ 1474 if (new_state == arc_l2c_only) 1475 l2arc_hdr_stat_add(); 1476 else if (old_state == arc_l2c_only) 1477 l2arc_hdr_stat_remove(); 1478} 1479 1480void 1481arc_space_consume(uint64_t space, arc_space_type_t type) 1482{ 1483 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES); 1484 1485 switch (type) { 1486 case ARC_SPACE_DATA: 1487 ARCSTAT_INCR(arcstat_data_size, space); 1488 break; 1489 case ARC_SPACE_OTHER: 1490 ARCSTAT_INCR(arcstat_other_size, space); 1491 break; 1492 case ARC_SPACE_HDRS: 1493 ARCSTAT_INCR(arcstat_hdr_size, space); 1494 break; 1495 case ARC_SPACE_L2HDRS: 1496 ARCSTAT_INCR(arcstat_l2_hdr_size, space); 1497 break; 1498 } 1499 1500 ARCSTAT_INCR(arcstat_meta_used, space); 1501 atomic_add_64(&arc_size, space); 1502} 1503 1504void 1505arc_space_return(uint64_t space, arc_space_type_t type) 1506{ 1507 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES); 1508 1509 switch (type) { 1510 case ARC_SPACE_DATA: 1511 ARCSTAT_INCR(arcstat_data_size, -space); 1512 break; 1513 case ARC_SPACE_OTHER: 1514 ARCSTAT_INCR(arcstat_other_size, -space); 1515 break; 1516 case ARC_SPACE_HDRS: 1517 ARCSTAT_INCR(arcstat_hdr_size, -space); 1518 break; 1519 case ARC_SPACE_L2HDRS: 1520 ARCSTAT_INCR(arcstat_l2_hdr_size, -space); 1521 break; 1522 } 1523 1524 ASSERT(arc_meta_used >= space); 1525 if (arc_meta_max < arc_meta_used) 1526 arc_meta_max = arc_meta_used; 1527 ARCSTAT_INCR(arcstat_meta_used, -space); 1528 ASSERT(arc_size >= space); 1529 atomic_add_64(&arc_size, -space); 1530} 1531 1532arc_buf_t * 1533arc_buf_alloc(spa_t *spa, int size, void *tag, arc_buf_contents_t type) 1534{ 1535 arc_buf_hdr_t *hdr; 1536 arc_buf_t *buf; 1537 1538 ASSERT3U(size, >, 0); 1539 hdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE); 1540 ASSERT(BUF_EMPTY(hdr)); 1541 hdr->b_size = size; 1542 hdr->b_type = type; 1543 hdr->b_spa = spa_load_guid(spa); 1544 hdr->b_state = arc_anon; 1545 hdr->b_arc_access = 0; 1546 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 1547 buf->b_hdr = hdr; 1548 buf->b_data = NULL; 1549 buf->b_efunc = NULL; 1550 buf->b_private = NULL; 1551 buf->b_next = NULL; 1552 hdr->b_buf = buf; 1553 arc_get_data_buf(buf); 1554 hdr->b_datacnt = 1; 1555 hdr->b_flags = 0; 1556 ASSERT(refcount_is_zero(&hdr->b_refcnt)); 1557 (void) refcount_add(&hdr->b_refcnt, tag); 1558 1559 return (buf); 1560} 1561 1562static char *arc_onloan_tag = "onloan"; 1563 1564/* 1565 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in 1566 * flight data by arc_tempreserve_space() until they are "returned". Loaned 1567 * buffers must be returned to the arc before they can be used by the DMU or 1568 * freed. 1569 */ 1570arc_buf_t * 1571arc_loan_buf(spa_t *spa, int size) 1572{ 1573 arc_buf_t *buf; 1574 1575 buf = arc_buf_alloc(spa, size, arc_onloan_tag, ARC_BUFC_DATA); 1576 1577 atomic_add_64(&arc_loaned_bytes, size); 1578 return (buf); 1579} 1580 1581/* 1582 * Return a loaned arc buffer to the arc. 1583 */ 1584void 1585arc_return_buf(arc_buf_t *buf, void *tag) 1586{ 1587 arc_buf_hdr_t *hdr = buf->b_hdr; 1588 1589 ASSERT(buf->b_data != NULL); 1590 (void) refcount_add(&hdr->b_refcnt, tag); 1591 (void) refcount_remove(&hdr->b_refcnt, arc_onloan_tag); 1592 1593 atomic_add_64(&arc_loaned_bytes, -hdr->b_size); 1594} 1595 1596/* Detach an arc_buf from a dbuf (tag) */ 1597void 1598arc_loan_inuse_buf(arc_buf_t *buf, void *tag) 1599{ 1600 arc_buf_hdr_t *hdr; 1601 1602 ASSERT(buf->b_data != NULL); 1603 hdr = buf->b_hdr; 1604 (void) refcount_add(&hdr->b_refcnt, arc_onloan_tag); 1605 (void) refcount_remove(&hdr->b_refcnt, tag); 1606 buf->b_efunc = NULL; 1607 buf->b_private = NULL; 1608 1609 atomic_add_64(&arc_loaned_bytes, hdr->b_size); 1610} 1611 1612static arc_buf_t * 1613arc_buf_clone(arc_buf_t *from) 1614{ 1615 arc_buf_t *buf; 1616 arc_buf_hdr_t *hdr = from->b_hdr; 1617 uint64_t size = hdr->b_size; 1618 1619 ASSERT(hdr->b_state != arc_anon); 1620 1621 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 1622 buf->b_hdr = hdr; 1623 buf->b_data = NULL; 1624 buf->b_efunc = NULL; 1625 buf->b_private = NULL; 1626 buf->b_next = hdr->b_buf; 1627 hdr->b_buf = buf; 1628 arc_get_data_buf(buf); 1629 bcopy(from->b_data, buf->b_data, size); 1630 1631 /* 1632 * This buffer already exists in the arc so create a duplicate 1633 * copy for the caller. If the buffer is associated with user data 1634 * then track the size and number of duplicates. These stats will be 1635 * updated as duplicate buffers are created and destroyed. 1636 */ 1637 if (hdr->b_type == ARC_BUFC_DATA) { 1638 ARCSTAT_BUMP(arcstat_duplicate_buffers); 1639 ARCSTAT_INCR(arcstat_duplicate_buffers_size, size); 1640 } 1641 hdr->b_datacnt += 1; 1642 return (buf); 1643} 1644 1645void 1646arc_buf_add_ref(arc_buf_t *buf, void* tag) 1647{ 1648 arc_buf_hdr_t *hdr; 1649 kmutex_t *hash_lock; 1650 1651 /* 1652 * Check to see if this buffer is evicted. Callers 1653 * must verify b_data != NULL to know if the add_ref 1654 * was successful. 1655 */ 1656 mutex_enter(&buf->b_evict_lock); 1657 if (buf->b_data == NULL) { 1658 mutex_exit(&buf->b_evict_lock); 1659 return; 1660 } 1661 hash_lock = HDR_LOCK(buf->b_hdr); 1662 mutex_enter(hash_lock); 1663 hdr = buf->b_hdr; 1664 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 1665 mutex_exit(&buf->b_evict_lock); 1666 1667 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu); 1668 add_reference(hdr, hash_lock, tag); 1669 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr); 1670 arc_access(hdr, hash_lock); 1671 mutex_exit(hash_lock); 1672 ARCSTAT_BUMP(arcstat_hits); 1673 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_FLAG_PREFETCH), 1674 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA, 1675 data, metadata, hits); 1676} 1677 1678static void 1679arc_buf_free_on_write(void *data, size_t size, 1680 void (*free_func)(void *, size_t)) 1681{ 1682 l2arc_data_free_t *df; 1683 1684 df = kmem_alloc(sizeof (l2arc_data_free_t), KM_SLEEP); 1685 df->l2df_data = data; 1686 df->l2df_size = size; 1687 df->l2df_func = free_func; 1688 mutex_enter(&l2arc_free_on_write_mtx); 1689 list_insert_head(l2arc_free_on_write, df); 1690 mutex_exit(&l2arc_free_on_write_mtx); 1691} 1692 1693/* 1694 * Free the arc data buffer. If it is an l2arc write in progress, 1695 * the buffer is placed on l2arc_free_on_write to be freed later. 1696 */ 1697static void 1698arc_buf_data_free(arc_buf_t *buf, void (*free_func)(void *, size_t)) 1699{ 1700 arc_buf_hdr_t *hdr = buf->b_hdr; 1701 1702 if (HDR_L2_WRITING(hdr)) { 1703 arc_buf_free_on_write(buf->b_data, hdr->b_size, free_func); 1704 ARCSTAT_BUMP(arcstat_l2_free_on_write); 1705 } else { 1706 free_func(buf->b_data, hdr->b_size); 1707 } 1708} 1709 1710/* 1711 * Free up buf->b_data and if 'remove' is set, then pull the 1712 * arc_buf_t off of the the arc_buf_hdr_t's list and free it. 1713 */ 1714static void 1715arc_buf_l2_cdata_free(arc_buf_hdr_t *hdr) 1716{ 1717 l2arc_buf_hdr_t *l2hdr = hdr->b_l2hdr; 1718 1719 ASSERT(MUTEX_HELD(&l2arc_buflist_mtx)); 1720 1721 if (l2hdr->b_tmp_cdata == NULL) 1722 return; 1723 1724 ASSERT(HDR_L2_WRITING(hdr)); 1725 arc_buf_free_on_write(l2hdr->b_tmp_cdata, hdr->b_size, 1726 zio_data_buf_free); 1727 ARCSTAT_BUMP(arcstat_l2_cdata_free_on_write); 1728 l2hdr->b_tmp_cdata = NULL; 1729} 1730 1731static void 1732arc_buf_destroy(arc_buf_t *buf, boolean_t recycle, boolean_t remove) 1733{ 1734 arc_buf_t **bufp; 1735 1736 /* free up data associated with the buf */ 1737 if (buf->b_data) { 1738 arc_state_t *state = buf->b_hdr->b_state; 1739 uint64_t size = buf->b_hdr->b_size; 1740 arc_buf_contents_t type = buf->b_hdr->b_type; 1741 1742 arc_cksum_verify(buf); 1743#ifdef illumos 1744 arc_buf_unwatch(buf); 1745#endif 1746 1747 if (!recycle) { 1748 if (type == ARC_BUFC_METADATA) { 1749 arc_buf_data_free(buf, zio_buf_free); 1750 arc_space_return(size, ARC_SPACE_DATA); 1751 } else { 1752 ASSERT(type == ARC_BUFC_DATA); 1753 arc_buf_data_free(buf, zio_data_buf_free); 1754 ARCSTAT_INCR(arcstat_data_size, -size); 1755 atomic_add_64(&arc_size, -size); 1756 } 1757 } 1758 if (list_link_active(&buf->b_hdr->b_arc_node)) { 1759 uint64_t *cnt = &state->arcs_lsize[type]; 1760 1761 ASSERT(refcount_is_zero(&buf->b_hdr->b_refcnt)); 1762 ASSERT(state != arc_anon); 1763 1764 ASSERT3U(*cnt, >=, size); 1765 atomic_add_64(cnt, -size); 1766 } 1767 ASSERT3U(state->arcs_size, >=, size); 1768 atomic_add_64(&state->arcs_size, -size); 1769 buf->b_data = NULL; 1770 1771 /* 1772 * If we're destroying a duplicate buffer make sure 1773 * that the appropriate statistics are updated. 1774 */ 1775 if (buf->b_hdr->b_datacnt > 1 && 1776 buf->b_hdr->b_type == ARC_BUFC_DATA) { 1777 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers); 1778 ARCSTAT_INCR(arcstat_duplicate_buffers_size, -size); 1779 } 1780 ASSERT(buf->b_hdr->b_datacnt > 0); 1781 buf->b_hdr->b_datacnt -= 1; 1782 } 1783 1784 /* only remove the buf if requested */ 1785 if (!remove) 1786 return; 1787 1788 /* remove the buf from the hdr list */ 1789 for (bufp = &buf->b_hdr->b_buf; *bufp != buf; bufp = &(*bufp)->b_next) 1790 continue; 1791 *bufp = buf->b_next; 1792 buf->b_next = NULL; 1793 1794 ASSERT(buf->b_efunc == NULL); 1795 1796 /* clean up the buf */ 1797 buf->b_hdr = NULL; 1798 kmem_cache_free(buf_cache, buf); 1799} 1800 1801static void 1802arc_hdr_destroy(arc_buf_hdr_t *hdr) 1803{ 1804 ASSERT(refcount_is_zero(&hdr->b_refcnt)); 1805 ASSERT3P(hdr->b_state, ==, arc_anon); 1806 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 1807 l2arc_buf_hdr_t *l2hdr = hdr->b_l2hdr; 1808 1809 if (l2hdr != NULL) { 1810 boolean_t buflist_held = MUTEX_HELD(&l2arc_buflist_mtx); 1811 /* 1812 * To prevent arc_free() and l2arc_evict() from 1813 * attempting to free the same buffer at the same time, 1814 * a FREE_IN_PROGRESS flag is given to arc_free() to 1815 * give it priority. l2arc_evict() can't destroy this 1816 * header while we are waiting on l2arc_buflist_mtx. 1817 * 1818 * The hdr may be removed from l2ad_buflist before we 1819 * grab l2arc_buflist_mtx, so b_l2hdr is rechecked. 1820 */ 1821 if (!buflist_held) { 1822 mutex_enter(&l2arc_buflist_mtx); 1823 l2hdr = hdr->b_l2hdr; 1824 } 1825 1826 if (l2hdr != NULL) { 1827 trim_map_free(l2hdr->b_dev->l2ad_vdev, l2hdr->b_daddr, 1828 hdr->b_size, 0); 1829 list_remove(l2hdr->b_dev->l2ad_buflist, hdr); 1830 arc_buf_l2_cdata_free(hdr); 1831 ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size); 1832 ARCSTAT_INCR(arcstat_l2_asize, -l2hdr->b_asize); 1833 vdev_space_update(l2hdr->b_dev->l2ad_vdev, 1834 -l2hdr->b_asize, 0, 0); 1835 kmem_free(l2hdr, sizeof (l2arc_buf_hdr_t)); 1836 if (hdr->b_state == arc_l2c_only) 1837 l2arc_hdr_stat_remove(); 1838 hdr->b_l2hdr = NULL; 1839 } 1840 1841 if (!buflist_held) 1842 mutex_exit(&l2arc_buflist_mtx); 1843 } 1844 1845 if (!BUF_EMPTY(hdr)) { 1846 ASSERT(!HDR_IN_HASH_TABLE(hdr)); 1847 buf_discard_identity(hdr); 1848 } 1849 while (hdr->b_buf) { 1850 arc_buf_t *buf = hdr->b_buf; 1851 1852 if (buf->b_efunc) { 1853 mutex_enter(&arc_eviction_mtx); 1854 mutex_enter(&buf->b_evict_lock); 1855 ASSERT(buf->b_hdr != NULL); 1856 arc_buf_destroy(hdr->b_buf, FALSE, FALSE); 1857 hdr->b_buf = buf->b_next; 1858 buf->b_hdr = &arc_eviction_hdr; 1859 buf->b_next = arc_eviction_list; 1860 arc_eviction_list = buf; 1861 mutex_exit(&buf->b_evict_lock); 1862 mutex_exit(&arc_eviction_mtx); 1863 } else { 1864 arc_buf_destroy(hdr->b_buf, FALSE, TRUE); 1865 } 1866 } 1867 if (hdr->b_freeze_cksum != NULL) { 1868 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 1869 hdr->b_freeze_cksum = NULL; 1870 } 1871 if (hdr->b_thawed) { 1872 kmem_free(hdr->b_thawed, 1); 1873 hdr->b_thawed = NULL; 1874 } 1875 1876 ASSERT(!list_link_active(&hdr->b_arc_node)); 1877 ASSERT3P(hdr->b_hash_next, ==, NULL); 1878 ASSERT3P(hdr->b_acb, ==, NULL); 1879 kmem_cache_free(hdr_cache, hdr); 1880} 1881 1882void 1883arc_buf_free(arc_buf_t *buf, void *tag) 1884{ 1885 arc_buf_hdr_t *hdr = buf->b_hdr; 1886 int hashed = hdr->b_state != arc_anon; 1887 1888 ASSERT(buf->b_efunc == NULL); 1889 ASSERT(buf->b_data != NULL); 1890 1891 if (hashed) { 1892 kmutex_t *hash_lock = HDR_LOCK(hdr); 1893 1894 mutex_enter(hash_lock); 1895 hdr = buf->b_hdr; 1896 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 1897 1898 (void) remove_reference(hdr, hash_lock, tag); 1899 if (hdr->b_datacnt > 1) { 1900 arc_buf_destroy(buf, FALSE, TRUE); 1901 } else { 1902 ASSERT(buf == hdr->b_buf); 1903 ASSERT(buf->b_efunc == NULL); 1904 hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE; 1905 } 1906 mutex_exit(hash_lock); 1907 } else if (HDR_IO_IN_PROGRESS(hdr)) { 1908 int destroy_hdr; 1909 /* 1910 * We are in the middle of an async write. Don't destroy 1911 * this buffer unless the write completes before we finish 1912 * decrementing the reference count. 1913 */ 1914 mutex_enter(&arc_eviction_mtx); 1915 (void) remove_reference(hdr, NULL, tag); 1916 ASSERT(refcount_is_zero(&hdr->b_refcnt)); 1917 destroy_hdr = !HDR_IO_IN_PROGRESS(hdr); 1918 mutex_exit(&arc_eviction_mtx); 1919 if (destroy_hdr) 1920 arc_hdr_destroy(hdr); 1921 } else { 1922 if (remove_reference(hdr, NULL, tag) > 0) 1923 arc_buf_destroy(buf, FALSE, TRUE); 1924 else 1925 arc_hdr_destroy(hdr); 1926 } 1927} 1928 1929boolean_t 1930arc_buf_remove_ref(arc_buf_t *buf, void* tag) 1931{ 1932 arc_buf_hdr_t *hdr = buf->b_hdr; 1933 kmutex_t *hash_lock = HDR_LOCK(hdr); 1934 boolean_t no_callback = (buf->b_efunc == NULL); 1935 1936 if (hdr->b_state == arc_anon) { 1937 ASSERT(hdr->b_datacnt == 1); 1938 arc_buf_free(buf, tag); 1939 return (no_callback); 1940 } 1941 1942 mutex_enter(hash_lock); 1943 hdr = buf->b_hdr; 1944 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 1945 ASSERT(hdr->b_state != arc_anon); 1946 ASSERT(buf->b_data != NULL); 1947 1948 (void) remove_reference(hdr, hash_lock, tag); 1949 if (hdr->b_datacnt > 1) { 1950 if (no_callback) 1951 arc_buf_destroy(buf, FALSE, TRUE); 1952 } else if (no_callback) { 1953 ASSERT(hdr->b_buf == buf && buf->b_next == NULL); 1954 ASSERT(buf->b_efunc == NULL); 1955 hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE; 1956 } 1957 ASSERT(no_callback || hdr->b_datacnt > 1 || 1958 refcount_is_zero(&hdr->b_refcnt)); 1959 mutex_exit(hash_lock); 1960 return (no_callback); 1961} 1962 1963int 1964arc_buf_size(arc_buf_t *buf) 1965{ 1966 return (buf->b_hdr->b_size); 1967} 1968 1969/* 1970 * Called from the DMU to determine if the current buffer should be 1971 * evicted. In order to ensure proper locking, the eviction must be initiated 1972 * from the DMU. Return true if the buffer is associated with user data and 1973 * duplicate buffers still exist. 1974 */ 1975boolean_t 1976arc_buf_eviction_needed(arc_buf_t *buf) 1977{ 1978 arc_buf_hdr_t *hdr; 1979 boolean_t evict_needed = B_FALSE; 1980 1981 if (zfs_disable_dup_eviction) 1982 return (B_FALSE); 1983 1984 mutex_enter(&buf->b_evict_lock); 1985 hdr = buf->b_hdr; 1986 if (hdr == NULL) { 1987 /* 1988 * We are in arc_do_user_evicts(); let that function 1989 * perform the eviction. 1990 */ 1991 ASSERT(buf->b_data == NULL); 1992 mutex_exit(&buf->b_evict_lock); 1993 return (B_FALSE); 1994 } else if (buf->b_data == NULL) { 1995 /* 1996 * We have already been added to the arc eviction list; 1997 * recommend eviction. 1998 */ 1999 ASSERT3P(hdr, ==, &arc_eviction_hdr); 2000 mutex_exit(&buf->b_evict_lock); 2001 return (B_TRUE); 2002 } 2003 2004 if (hdr->b_datacnt > 1 && hdr->b_type == ARC_BUFC_DATA) 2005 evict_needed = B_TRUE; 2006 2007 mutex_exit(&buf->b_evict_lock); 2008 return (evict_needed); 2009} 2010 2011/* 2012 * Evict buffers from list until we've removed the specified number of 2013 * bytes. Move the removed buffers to the appropriate evict state. 2014 * If the recycle flag is set, then attempt to "recycle" a buffer: 2015 * - look for a buffer to evict that is `bytes' long. 2016 * - return the data block from this buffer rather than freeing it. 2017 * This flag is used by callers that are trying to make space for a 2018 * new buffer in a full arc cache. 2019 * 2020 * This function makes a "best effort". It skips over any buffers 2021 * it can't get a hash_lock on, and so may not catch all candidates. 2022 * It may also return without evicting as much space as requested. 2023 */ 2024static void * 2025arc_evict(arc_state_t *state, uint64_t spa, int64_t bytes, boolean_t recycle, 2026 arc_buf_contents_t type) 2027{ 2028 arc_state_t *evicted_state; 2029 uint64_t bytes_evicted = 0, skipped = 0, missed = 0; 2030 int64_t bytes_remaining; 2031 arc_buf_hdr_t *hdr, *hdr_prev = NULL; 2032 list_t *evicted_list, *list, *evicted_list_start, *list_start; 2033 kmutex_t *lock, *evicted_lock; 2034 kmutex_t *hash_lock; 2035 boolean_t have_lock; 2036 void *stolen = NULL; 2037 arc_buf_hdr_t marker = { 0 }; 2038 int count = 0; 2039 static int evict_metadata_offset, evict_data_offset; 2040 int i, idx, offset, list_count, lists; 2041 2042 ASSERT(state == arc_mru || state == arc_mfu); 2043 2044 evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost; 2045 2046 /* 2047 * Decide which "type" (data vs metadata) to recycle from. 2048 * 2049 * If we are over the metadata limit, recycle from metadata. 2050 * If we are under the metadata minimum, recycle from data. 2051 * Otherwise, recycle from whichever type has the oldest (least 2052 * recently accessed) header. This is not yet implemented. 2053 */ 2054 if (recycle) { 2055 arc_buf_contents_t realtype; 2056 if (state->arcs_lsize[ARC_BUFC_DATA] == 0) { 2057 realtype = ARC_BUFC_METADATA; 2058 } else if (state->arcs_lsize[ARC_BUFC_METADATA] == 0) { 2059 realtype = ARC_BUFC_DATA; 2060 } else if (arc_meta_used >= arc_meta_limit) { 2061 realtype = ARC_BUFC_METADATA; 2062 } else if (arc_meta_used <= arc_meta_min) { 2063 realtype = ARC_BUFC_DATA; 2064 } else { 2065#ifdef illumos 2066 if (data_hdr->b_arc_access < 2067 metadata_hdr->b_arc_access) { 2068 realtype = ARC_BUFC_DATA; 2069 } else { 2070 realtype = ARC_BUFC_METADATA; 2071 } 2072#else 2073 /* TODO */ 2074 realtype = type; 2075#endif 2076 } 2077 if (realtype != type) { 2078 /* 2079 * If we want to evict from a different list, 2080 * we can not recycle, because DATA vs METADATA 2081 * buffers are segregated into different kmem 2082 * caches (and vmem arenas). 2083 */ 2084 type = realtype; 2085 recycle = B_FALSE; 2086 } 2087 } 2088 2089 if (type == ARC_BUFC_METADATA) { 2090 offset = 0; 2091 list_count = ARC_BUFC_NUMMETADATALISTS; 2092 list_start = &state->arcs_lists[0]; 2093 evicted_list_start = &evicted_state->arcs_lists[0]; 2094 idx = evict_metadata_offset; 2095 } else { 2096 offset = ARC_BUFC_NUMMETADATALISTS; 2097 list_start = &state->arcs_lists[offset]; 2098 evicted_list_start = &evicted_state->arcs_lists[offset]; 2099 list_count = ARC_BUFC_NUMDATALISTS; 2100 idx = evict_data_offset; 2101 } 2102 bytes_remaining = evicted_state->arcs_lsize[type]; 2103 lists = 0; 2104 2105evict_start: 2106 list = &list_start[idx]; 2107 evicted_list = &evicted_list_start[idx]; 2108 lock = ARCS_LOCK(state, (offset + idx)); 2109 evicted_lock = ARCS_LOCK(evicted_state, (offset + idx)); 2110 2111 mutex_enter(lock); 2112 mutex_enter(evicted_lock); 2113 2114 for (hdr = list_tail(list); hdr; hdr = hdr_prev) { 2115 hdr_prev = list_prev(list, hdr); 2116 bytes_remaining -= (hdr->b_size * hdr->b_datacnt); 2117 /* prefetch buffers have a minimum lifespan */ 2118 if (HDR_IO_IN_PROGRESS(hdr) || 2119 (spa && hdr->b_spa != spa) || 2120 (hdr->b_flags & (ARC_FLAG_PREFETCH | ARC_FLAG_INDIRECT) && 2121 ddi_get_lbolt() - hdr->b_arc_access < 2122 arc_min_prefetch_lifespan)) { 2123 skipped++; 2124 continue; 2125 } 2126 /* "lookahead" for better eviction candidate */ 2127 if (recycle && hdr->b_size != bytes && 2128 hdr_prev && hdr_prev->b_size == bytes) 2129 continue; 2130 2131 /* ignore markers */ 2132 if (hdr->b_spa == 0) 2133 continue; 2134 2135 /* 2136 * It may take a long time to evict all the bufs requested. 2137 * To avoid blocking all arc activity, periodically drop 2138 * the arcs_mtx and give other threads a chance to run 2139 * before reacquiring the lock. 2140 * 2141 * If we are looking for a buffer to recycle, we are in 2142 * the hot code path, so don't sleep. 2143 */ 2144 if (!recycle && count++ > arc_evict_iterations) { 2145 list_insert_after(list, hdr, &marker); 2146 mutex_exit(evicted_lock); 2147 mutex_exit(lock); 2148 kpreempt(KPREEMPT_SYNC); 2149 mutex_enter(lock); 2150 mutex_enter(evicted_lock); 2151 hdr_prev = list_prev(list, &marker); 2152 list_remove(list, &marker); 2153 count = 0; 2154 continue; 2155 } 2156 2157 hash_lock = HDR_LOCK(hdr); 2158 have_lock = MUTEX_HELD(hash_lock); 2159 if (have_lock || mutex_tryenter(hash_lock)) { 2160 ASSERT0(refcount_count(&hdr->b_refcnt)); 2161 ASSERT(hdr->b_datacnt > 0); 2162 while (hdr->b_buf) { 2163 arc_buf_t *buf = hdr->b_buf; 2164 if (!mutex_tryenter(&buf->b_evict_lock)) { 2165 missed += 1; 2166 break; 2167 } 2168 if (buf->b_data) { 2169 bytes_evicted += hdr->b_size; 2170 if (recycle && hdr->b_type == type && 2171 hdr->b_size == bytes && 2172 !HDR_L2_WRITING(hdr)) { 2173 stolen = buf->b_data; 2174 recycle = FALSE; 2175 } 2176 } 2177 if (buf->b_efunc) { 2178 mutex_enter(&arc_eviction_mtx); 2179 arc_buf_destroy(buf, 2180 buf->b_data == stolen, FALSE); 2181 hdr->b_buf = buf->b_next; 2182 buf->b_hdr = &arc_eviction_hdr; 2183 buf->b_next = arc_eviction_list; 2184 arc_eviction_list = buf; 2185 mutex_exit(&arc_eviction_mtx); 2186 mutex_exit(&buf->b_evict_lock); 2187 } else { 2188 mutex_exit(&buf->b_evict_lock); 2189 arc_buf_destroy(buf, 2190 buf->b_data == stolen, TRUE); 2191 } 2192 } 2193 2194 if (hdr->b_l2hdr) { 2195 ARCSTAT_INCR(arcstat_evict_l2_cached, 2196 hdr->b_size); 2197 } else { 2198 if (l2arc_write_eligible(hdr->b_spa, hdr)) { 2199 ARCSTAT_INCR(arcstat_evict_l2_eligible, 2200 hdr->b_size); 2201 } else { 2202 ARCSTAT_INCR( 2203 arcstat_evict_l2_ineligible, 2204 hdr->b_size); 2205 } 2206 } 2207 2208 if (hdr->b_datacnt == 0) { 2209 arc_change_state(evicted_state, hdr, hash_lock); 2210 ASSERT(HDR_IN_HASH_TABLE(hdr)); 2211 hdr->b_flags |= ARC_FLAG_IN_HASH_TABLE; 2212 hdr->b_flags &= ~ARC_FLAG_BUF_AVAILABLE; 2213 DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, hdr); 2214 } 2215 if (!have_lock) 2216 mutex_exit(hash_lock); 2217 if (bytes >= 0 && bytes_evicted >= bytes) 2218 break; 2219 if (bytes_remaining > 0) { 2220 mutex_exit(evicted_lock); 2221 mutex_exit(lock); 2222 idx = ((idx + 1) & (list_count - 1)); 2223 lists++; 2224 goto evict_start; 2225 } 2226 } else { 2227 missed += 1; 2228 } 2229 } 2230 2231 mutex_exit(evicted_lock); 2232 mutex_exit(lock); 2233 2234 idx = ((idx + 1) & (list_count - 1)); 2235 lists++; 2236 2237 if (bytes_evicted < bytes) { 2238 if (lists < list_count) 2239 goto evict_start; 2240 else 2241 dprintf("only evicted %lld bytes from %x", 2242 (longlong_t)bytes_evicted, state); 2243 } 2244 if (type == ARC_BUFC_METADATA) 2245 evict_metadata_offset = idx; 2246 else 2247 evict_data_offset = idx; 2248 2249 if (skipped) 2250 ARCSTAT_INCR(arcstat_evict_skip, skipped); 2251 2252 if (missed) 2253 ARCSTAT_INCR(arcstat_mutex_miss, missed); 2254 2255 /* 2256 * Note: we have just evicted some data into the ghost state, 2257 * potentially putting the ghost size over the desired size. Rather 2258 * that evicting from the ghost list in this hot code path, leave 2259 * this chore to the arc_reclaim_thread(). 2260 */ 2261 2262 if (stolen) 2263 ARCSTAT_BUMP(arcstat_stolen); 2264 return (stolen); 2265} 2266 2267/* 2268 * Remove buffers from list until we've removed the specified number of 2269 * bytes. Destroy the buffers that are removed. 2270 */ 2271static void 2272arc_evict_ghost(arc_state_t *state, uint64_t spa, int64_t bytes) 2273{ 2274 arc_buf_hdr_t *hdr, *hdr_prev; 2275 arc_buf_hdr_t marker = { 0 }; 2276 list_t *list, *list_start; 2277 kmutex_t *hash_lock, *lock; 2278 uint64_t bytes_deleted = 0; 2279 uint64_t bufs_skipped = 0; 2280 int count = 0; 2281 static int evict_offset; 2282 int list_count, idx = evict_offset; 2283 int offset, lists = 0; 2284 2285 ASSERT(GHOST_STATE(state)); 2286 2287 /* 2288 * data lists come after metadata lists 2289 */ 2290 list_start = &state->arcs_lists[ARC_BUFC_NUMMETADATALISTS]; 2291 list_count = ARC_BUFC_NUMDATALISTS; 2292 offset = ARC_BUFC_NUMMETADATALISTS; 2293 2294evict_start: 2295 list = &list_start[idx]; 2296 lock = ARCS_LOCK(state, idx + offset); 2297 2298 mutex_enter(lock); 2299 for (hdr = list_tail(list); hdr; hdr = hdr_prev) { 2300 hdr_prev = list_prev(list, hdr); 2301 if (hdr->b_type > ARC_BUFC_NUMTYPES) 2302 panic("invalid hdr=%p", (void *)hdr); 2303 if (spa && hdr->b_spa != spa) 2304 continue; 2305 2306 /* ignore markers */ 2307 if (hdr->b_spa == 0) 2308 continue; 2309 2310 hash_lock = HDR_LOCK(hdr); 2311 /* caller may be trying to modify this buffer, skip it */ 2312 if (MUTEX_HELD(hash_lock)) 2313 continue; 2314 2315 /* 2316 * It may take a long time to evict all the bufs requested. 2317 * To avoid blocking all arc activity, periodically drop 2318 * the arcs_mtx and give other threads a chance to run 2319 * before reacquiring the lock. 2320 */ 2321 if (count++ > arc_evict_iterations) { 2322 list_insert_after(list, hdr, &marker); 2323 mutex_exit(lock); 2324 kpreempt(KPREEMPT_SYNC); 2325 mutex_enter(lock); 2326 hdr_prev = list_prev(list, &marker); 2327 list_remove(list, &marker); 2328 count = 0; 2329 continue; 2330 } 2331 if (mutex_tryenter(hash_lock)) { 2332 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 2333 ASSERT(hdr->b_buf == NULL); 2334 ARCSTAT_BUMP(arcstat_deleted); 2335 bytes_deleted += hdr->b_size; 2336 2337 if (hdr->b_l2hdr != NULL) { 2338 /* 2339 * This buffer is cached on the 2nd Level ARC; 2340 * don't destroy the header. 2341 */ 2342 arc_change_state(arc_l2c_only, hdr, hash_lock); 2343 mutex_exit(hash_lock); 2344 } else { 2345 arc_change_state(arc_anon, hdr, hash_lock); 2346 mutex_exit(hash_lock); 2347 arc_hdr_destroy(hdr); 2348 } 2349 2350 DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, hdr); 2351 if (bytes >= 0 && bytes_deleted >= bytes) 2352 break; 2353 } else if (bytes < 0) { 2354 /* 2355 * Insert a list marker and then wait for the 2356 * hash lock to become available. Once its 2357 * available, restart from where we left off. 2358 */ 2359 list_insert_after(list, hdr, &marker); 2360 mutex_exit(lock); 2361 mutex_enter(hash_lock); 2362 mutex_exit(hash_lock); 2363 mutex_enter(lock); 2364 hdr_prev = list_prev(list, &marker); 2365 list_remove(list, &marker); 2366 } else { 2367 bufs_skipped += 1; 2368 } 2369 2370 } 2371 mutex_exit(lock); 2372 idx = ((idx + 1) & (ARC_BUFC_NUMDATALISTS - 1)); 2373 lists++; 2374 2375 if (lists < list_count) 2376 goto evict_start; 2377 2378 evict_offset = idx; 2379 if ((uintptr_t)list > (uintptr_t)&state->arcs_lists[ARC_BUFC_NUMMETADATALISTS] && 2380 (bytes < 0 || bytes_deleted < bytes)) { 2381 list_start = &state->arcs_lists[0]; 2382 list_count = ARC_BUFC_NUMMETADATALISTS; 2383 offset = lists = 0; 2384 goto evict_start; 2385 } 2386 2387 if (bufs_skipped) { 2388 ARCSTAT_INCR(arcstat_mutex_miss, bufs_skipped); 2389 ASSERT(bytes >= 0); 2390 } 2391 2392 if (bytes_deleted < bytes) 2393 dprintf("only deleted %lld bytes from %p", 2394 (longlong_t)bytes_deleted, state); 2395} 2396 2397static void 2398arc_adjust(void) 2399{ 2400 int64_t adjustment, delta; 2401 2402 /* 2403 * Adjust MRU size 2404 */ 2405 2406 adjustment = MIN((int64_t)(arc_size - arc_c), 2407 (int64_t)(arc_anon->arcs_size + arc_mru->arcs_size + arc_meta_used - 2408 arc_p)); 2409 2410 if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_DATA] > 0) { 2411 delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_DATA], adjustment); 2412 (void) arc_evict(arc_mru, 0, delta, FALSE, ARC_BUFC_DATA); 2413 adjustment -= delta; 2414 } 2415 2416 if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_METADATA] > 0) { 2417 delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_METADATA], adjustment); 2418 (void) arc_evict(arc_mru, 0, delta, FALSE, 2419 ARC_BUFC_METADATA); 2420 } 2421 2422 /* 2423 * Adjust MFU size 2424 */ 2425 2426 adjustment = arc_size - arc_c; 2427 2428 if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_DATA] > 0) { 2429 delta = MIN(adjustment, arc_mfu->arcs_lsize[ARC_BUFC_DATA]); 2430 (void) arc_evict(arc_mfu, 0, delta, FALSE, ARC_BUFC_DATA); 2431 adjustment -= delta; 2432 } 2433 2434 if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_METADATA] > 0) { 2435 int64_t delta = MIN(adjustment, 2436 arc_mfu->arcs_lsize[ARC_BUFC_METADATA]); 2437 (void) arc_evict(arc_mfu, 0, delta, FALSE, 2438 ARC_BUFC_METADATA); 2439 } 2440 2441 /* 2442 * Adjust ghost lists 2443 */ 2444 2445 adjustment = arc_mru->arcs_size + arc_mru_ghost->arcs_size - arc_c; 2446 2447 if (adjustment > 0 && arc_mru_ghost->arcs_size > 0) { 2448 delta = MIN(arc_mru_ghost->arcs_size, adjustment); 2449 arc_evict_ghost(arc_mru_ghost, 0, delta); 2450 } 2451 2452 adjustment = 2453 arc_mru_ghost->arcs_size + arc_mfu_ghost->arcs_size - arc_c; 2454 2455 if (adjustment > 0 && arc_mfu_ghost->arcs_size > 0) { 2456 delta = MIN(arc_mfu_ghost->arcs_size, adjustment); 2457 arc_evict_ghost(arc_mfu_ghost, 0, delta); 2458 } 2459} 2460 2461static void 2462arc_do_user_evicts(void) 2463{ 2464 static arc_buf_t *tmp_arc_eviction_list; 2465 2466 /* 2467 * Move list over to avoid LOR 2468 */ 2469restart: 2470 mutex_enter(&arc_eviction_mtx); 2471 tmp_arc_eviction_list = arc_eviction_list; 2472 arc_eviction_list = NULL; 2473 mutex_exit(&arc_eviction_mtx); 2474 2475 while (tmp_arc_eviction_list != NULL) { 2476 arc_buf_t *buf = tmp_arc_eviction_list; 2477 tmp_arc_eviction_list = buf->b_next; 2478 mutex_enter(&buf->b_evict_lock); 2479 buf->b_hdr = NULL; 2480 mutex_exit(&buf->b_evict_lock); 2481 2482 if (buf->b_efunc != NULL) 2483 VERIFY0(buf->b_efunc(buf->b_private)); 2484 2485 buf->b_efunc = NULL; 2486 buf->b_private = NULL; 2487 kmem_cache_free(buf_cache, buf); 2488 } 2489 2490 if (arc_eviction_list != NULL) 2491 goto restart; 2492} 2493 2494/* 2495 * Flush all *evictable* data from the cache for the given spa. 2496 * NOTE: this will not touch "active" (i.e. referenced) data. 2497 */ 2498void 2499arc_flush(spa_t *spa) 2500{ 2501 uint64_t guid = 0; 2502 2503 if (spa) 2504 guid = spa_load_guid(spa); 2505 2506 while (arc_mru->arcs_lsize[ARC_BUFC_DATA]) { 2507 (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_DATA); 2508 if (spa) 2509 break; 2510 } 2511 while (arc_mru->arcs_lsize[ARC_BUFC_METADATA]) { 2512 (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_METADATA); 2513 if (spa) 2514 break; 2515 } 2516 while (arc_mfu->arcs_lsize[ARC_BUFC_DATA]) { 2517 (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_DATA); 2518 if (spa) 2519 break; 2520 } 2521 while (arc_mfu->arcs_lsize[ARC_BUFC_METADATA]) { 2522 (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_METADATA); 2523 if (spa) 2524 break; 2525 } 2526 2527 arc_evict_ghost(arc_mru_ghost, guid, -1); 2528 arc_evict_ghost(arc_mfu_ghost, guid, -1); 2529 2530 mutex_enter(&arc_reclaim_thr_lock); 2531 arc_do_user_evicts(); 2532 mutex_exit(&arc_reclaim_thr_lock); 2533 ASSERT(spa || arc_eviction_list == NULL); 2534} 2535 2536void 2537arc_shrink(void) 2538{ 2539 2540 if (arc_c > arc_c_min) { 2541 uint64_t to_free; 2542 2543 DTRACE_PROBE4(arc__shrink, uint64_t, arc_c, uint64_t, 2544 arc_c_min, uint64_t, arc_p, uint64_t, to_free); 2545#ifdef _KERNEL 2546 to_free = arc_c >> arc_shrink_shift; 2547#else 2548 to_free = arc_c >> arc_shrink_shift; 2549#endif 2550 if (arc_c > arc_c_min + to_free) 2551 atomic_add_64(&arc_c, -to_free); 2552 else 2553 arc_c = arc_c_min; 2554 2555 atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift)); 2556 if (arc_c > arc_size) 2557 arc_c = MAX(arc_size, arc_c_min); 2558 if (arc_p > arc_c) 2559 arc_p = (arc_c >> 1); 2560 2561 DTRACE_PROBE2(arc__shrunk, uint64_t, arc_c, uint64_t, 2562 arc_p); 2563 2564 ASSERT(arc_c >= arc_c_min); 2565 ASSERT((int64_t)arc_p >= 0); 2566 } 2567 2568 if (arc_size > arc_c) { 2569 DTRACE_PROBE2(arc__shrink_adjust, uint64_t, arc_size, 2570 uint64_t, arc_c); 2571 arc_adjust(); 2572 } 2573} 2574 2575static int needfree = 0; 2576 2577static int 2578arc_reclaim_needed(void) 2579{ 2580 2581#ifdef _KERNEL 2582 2583 if (needfree) { 2584 DTRACE_PROBE(arc__reclaim_needfree); 2585 return (1); 2586 } 2587 2588 /* 2589 * Cooperate with pagedaemon when it's time for it to scan 2590 * and reclaim some pages. 2591 */ 2592 if (freemem < zfs_arc_free_target) { 2593 DTRACE_PROBE2(arc__reclaim_freemem, uint64_t, 2594 freemem, uint64_t, zfs_arc_free_target); 2595 return (1); 2596 } 2597 2598#ifdef illumos 2599 /* 2600 * take 'desfree' extra pages, so we reclaim sooner, rather than later 2601 */ 2602 extra = desfree; 2603 2604 /* 2605 * check that we're out of range of the pageout scanner. It starts to 2606 * schedule paging if freemem is less than lotsfree and needfree. 2607 * lotsfree is the high-water mark for pageout, and needfree is the 2608 * number of needed free pages. We add extra pages here to make sure 2609 * the scanner doesn't start up while we're freeing memory. 2610 */ 2611 if (freemem < lotsfree + needfree + extra) 2612 return (1); 2613 2614 /* 2615 * check to make sure that swapfs has enough space so that anon 2616 * reservations can still succeed. anon_resvmem() checks that the 2617 * availrmem is greater than swapfs_minfree, and the number of reserved 2618 * swap pages. We also add a bit of extra here just to prevent 2619 * circumstances from getting really dire. 2620 */ 2621 if (availrmem < swapfs_minfree + swapfs_reserve + extra) 2622 return (1); 2623 2624 /* 2625 * Check that we have enough availrmem that memory locking (e.g., via 2626 * mlock(3C) or memcntl(2)) can still succeed. (pages_pp_maximum 2627 * stores the number of pages that cannot be locked; when availrmem 2628 * drops below pages_pp_maximum, page locking mechanisms such as 2629 * page_pp_lock() will fail.) 2630 */ 2631 if (availrmem <= pages_pp_maximum) 2632 return (1); 2633 2634#endif /* illumos */ 2635#if defined(__i386) || !defined(UMA_MD_SMALL_ALLOC) 2636 /* 2637 * If we're on an i386 platform, it's possible that we'll exhaust the 2638 * kernel heap space before we ever run out of available physical 2639 * memory. Most checks of the size of the heap_area compare against 2640 * tune.t_minarmem, which is the minimum available real memory that we 2641 * can have in the system. However, this is generally fixed at 25 pages 2642 * which is so low that it's useless. In this comparison, we seek to 2643 * calculate the total heap-size, and reclaim if more than 3/4ths of the 2644 * heap is allocated. (Or, in the calculation, if less than 1/4th is 2645 * free) 2646 */ 2647 if (vmem_size(heap_arena, VMEM_FREE) < 2648 (vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC) >> 2)) { 2649 DTRACE_PROBE2(arc__reclaim_used, uint64_t, 2650 vmem_size(heap_arena, VMEM_FREE), uint64_t, 2651 (vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC)) >> 2); 2652 return (1); 2653 } 2654#endif 2655#ifdef illumos 2656 /* 2657 * If zio data pages are being allocated out of a separate heap segment, 2658 * then enforce that the size of available vmem for this arena remains 2659 * above about 1/16th free. 2660 * 2661 * Note: The 1/16th arena free requirement was put in place 2662 * to aggressively evict memory from the arc in order to avoid 2663 * memory fragmentation issues. 2664 */ 2665 if (zio_arena != NULL && 2666 vmem_size(zio_arena, VMEM_FREE) < 2667 (vmem_size(zio_arena, VMEM_ALLOC) >> 4)) 2668 return (1); 2669#endif /* illumos */ 2670#else /* _KERNEL */ 2671 if (spa_get_random(100) == 0) 2672 return (1); 2673#endif /* _KERNEL */ 2674 DTRACE_PROBE(arc__reclaim_no); 2675 2676 return (0); 2677} 2678 2679extern kmem_cache_t *zio_buf_cache[]; 2680extern kmem_cache_t *zio_data_buf_cache[]; 2681extern kmem_cache_t *range_seg_cache; 2682 2683static void __noinline 2684arc_kmem_reap_now(arc_reclaim_strategy_t strat) 2685{ 2686 size_t i; 2687 kmem_cache_t *prev_cache = NULL; 2688 kmem_cache_t *prev_data_cache = NULL; 2689 2690 DTRACE_PROBE(arc__kmem_reap_start); 2691#ifdef _KERNEL 2692 if (arc_meta_used >= arc_meta_limit) { 2693 /* 2694 * We are exceeding our meta-data cache limit. 2695 * Purge some DNLC entries to release holds on meta-data. 2696 */ 2697 dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent); 2698 } 2699#if defined(__i386) 2700 /* 2701 * Reclaim unused memory from all kmem caches. 2702 */ 2703 kmem_reap(); 2704#endif 2705#endif 2706 2707 /* 2708 * An aggressive reclamation will shrink the cache size as well as 2709 * reap free buffers from the arc kmem caches. 2710 */ 2711 if (strat == ARC_RECLAIM_AGGR) 2712 arc_shrink(); 2713 2714 for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) { 2715 if (zio_buf_cache[i] != prev_cache) { 2716 prev_cache = zio_buf_cache[i]; 2717 kmem_cache_reap_now(zio_buf_cache[i]); 2718 } 2719 if (zio_data_buf_cache[i] != prev_data_cache) { 2720 prev_data_cache = zio_data_buf_cache[i]; 2721 kmem_cache_reap_now(zio_data_buf_cache[i]); 2722 } 2723 } 2724 kmem_cache_reap_now(buf_cache); 2725 kmem_cache_reap_now(hdr_cache); 2726 kmem_cache_reap_now(range_seg_cache); 2727 2728#ifdef illumos 2729 /* 2730 * Ask the vmem arena to reclaim unused memory from its 2731 * quantum caches. 2732 */ 2733 if (zio_arena != NULL && strat == ARC_RECLAIM_AGGR) 2734 vmem_qcache_reap(zio_arena); 2735#endif 2736 DTRACE_PROBE(arc__kmem_reap_end); 2737} 2738 2739static void 2740arc_reclaim_thread(void *dummy __unused) 2741{ 2742 clock_t growtime = 0; 2743 arc_reclaim_strategy_t last_reclaim = ARC_RECLAIM_CONS; 2744 callb_cpr_t cpr; 2745 2746 CALLB_CPR_INIT(&cpr, &arc_reclaim_thr_lock, callb_generic_cpr, FTAG); 2747 2748 mutex_enter(&arc_reclaim_thr_lock); 2749 while (arc_thread_exit == 0) { 2750 if (arc_reclaim_needed()) { 2751 2752 if (arc_no_grow) { 2753 if (last_reclaim == ARC_RECLAIM_CONS) { 2754 DTRACE_PROBE(arc__reclaim_aggr_no_grow); 2755 last_reclaim = ARC_RECLAIM_AGGR; 2756 } else { 2757 last_reclaim = ARC_RECLAIM_CONS; 2758 } 2759 } else { 2760 arc_no_grow = TRUE; 2761 last_reclaim = ARC_RECLAIM_AGGR; 2762 DTRACE_PROBE(arc__reclaim_aggr); 2763 membar_producer(); 2764 } 2765 2766 /* reset the growth delay for every reclaim */ 2767 growtime = ddi_get_lbolt() + (arc_grow_retry * hz); 2768 2769 if (needfree && last_reclaim == ARC_RECLAIM_CONS) { 2770 /* 2771 * If needfree is TRUE our vm_lowmem hook 2772 * was called and in that case we must free some 2773 * memory, so switch to aggressive mode. 2774 */ 2775 arc_no_grow = TRUE; 2776 last_reclaim = ARC_RECLAIM_AGGR; 2777 } 2778 arc_kmem_reap_now(last_reclaim); 2779 arc_warm = B_TRUE; 2780 2781 } else if (arc_no_grow && ddi_get_lbolt() >= growtime) { 2782 arc_no_grow = FALSE; 2783 } 2784 2785 arc_adjust(); 2786 2787 if (arc_eviction_list != NULL) 2788 arc_do_user_evicts(); 2789 2790#ifdef _KERNEL 2791 if (needfree) { 2792 needfree = 0; 2793 wakeup(&needfree); 2794 } 2795#endif 2796 2797 /* block until needed, or one second, whichever is shorter */ 2798 CALLB_CPR_SAFE_BEGIN(&cpr); 2799 (void) cv_timedwait(&arc_reclaim_thr_cv, 2800 &arc_reclaim_thr_lock, hz); 2801 CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_thr_lock); 2802 } 2803 2804 arc_thread_exit = 0; 2805 cv_broadcast(&arc_reclaim_thr_cv); 2806 CALLB_CPR_EXIT(&cpr); /* drops arc_reclaim_thr_lock */ 2807 thread_exit(); 2808} 2809 2810/* 2811 * Adapt arc info given the number of bytes we are trying to add and 2812 * the state that we are comming from. This function is only called 2813 * when we are adding new content to the cache. 2814 */ 2815static void 2816arc_adapt(int bytes, arc_state_t *state) 2817{ 2818 int mult; 2819 uint64_t arc_p_min = (arc_c >> arc_p_min_shift); 2820 2821 if (state == arc_l2c_only) 2822 return; 2823 2824 ASSERT(bytes > 0); 2825 /* 2826 * Adapt the target size of the MRU list: 2827 * - if we just hit in the MRU ghost list, then increase 2828 * the target size of the MRU list. 2829 * - if we just hit in the MFU ghost list, then increase 2830 * the target size of the MFU list by decreasing the 2831 * target size of the MRU list. 2832 */ 2833 if (state == arc_mru_ghost) { 2834 mult = ((arc_mru_ghost->arcs_size >= arc_mfu_ghost->arcs_size) ? 2835 1 : (arc_mfu_ghost->arcs_size/arc_mru_ghost->arcs_size)); 2836 mult = MIN(mult, 10); /* avoid wild arc_p adjustment */ 2837 2838 arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult); 2839 } else if (state == arc_mfu_ghost) { 2840 uint64_t delta; 2841 2842 mult = ((arc_mfu_ghost->arcs_size >= arc_mru_ghost->arcs_size) ? 2843 1 : (arc_mru_ghost->arcs_size/arc_mfu_ghost->arcs_size)); 2844 mult = MIN(mult, 10); 2845 2846 delta = MIN(bytes * mult, arc_p); 2847 arc_p = MAX(arc_p_min, arc_p - delta); 2848 } 2849 ASSERT((int64_t)arc_p >= 0); 2850 2851 if (arc_reclaim_needed()) { 2852 cv_signal(&arc_reclaim_thr_cv); 2853 return; 2854 } 2855 2856 if (arc_no_grow) 2857 return; 2858 2859 if (arc_c >= arc_c_max) 2860 return; 2861 2862 /* 2863 * If we're within (2 * maxblocksize) bytes of the target 2864 * cache size, increment the target cache size 2865 */ 2866 if (arc_size > arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) { 2867 DTRACE_PROBE1(arc__inc_adapt, int, bytes); 2868 atomic_add_64(&arc_c, (int64_t)bytes); 2869 if (arc_c > arc_c_max) 2870 arc_c = arc_c_max; 2871 else if (state == arc_anon) 2872 atomic_add_64(&arc_p, (int64_t)bytes); 2873 if (arc_p > arc_c) 2874 arc_p = arc_c; 2875 } 2876 ASSERT((int64_t)arc_p >= 0); 2877} 2878 2879/* 2880 * Check if the cache has reached its limits and eviction is required 2881 * prior to insert. 2882 */ 2883static int 2884arc_evict_needed(arc_buf_contents_t type) 2885{ 2886 if (type == ARC_BUFC_METADATA && arc_meta_used >= arc_meta_limit) 2887 return (1); 2888 2889 if (arc_reclaim_needed()) 2890 return (1); 2891 2892 return (arc_size > arc_c); 2893} 2894 2895/* 2896 * The buffer, supplied as the first argument, needs a data block. 2897 * So, if we are at cache max, determine which cache should be victimized. 2898 * We have the following cases: 2899 * 2900 * 1. Insert for MRU, p > sizeof(arc_anon + arc_mru) -> 2901 * In this situation if we're out of space, but the resident size of the MFU is 2902 * under the limit, victimize the MFU cache to satisfy this insertion request. 2903 * 2904 * 2. Insert for MRU, p <= sizeof(arc_anon + arc_mru) -> 2905 * Here, we've used up all of the available space for the MRU, so we need to 2906 * evict from our own cache instead. Evict from the set of resident MRU 2907 * entries. 2908 * 2909 * 3. Insert for MFU (c - p) > sizeof(arc_mfu) -> 2910 * c minus p represents the MFU space in the cache, since p is the size of the 2911 * cache that is dedicated to the MRU. In this situation there's still space on 2912 * the MFU side, so the MRU side needs to be victimized. 2913 * 2914 * 4. Insert for MFU (c - p) < sizeof(arc_mfu) -> 2915 * MFU's resident set is consuming more space than it has been allotted. In 2916 * this situation, we must victimize our own cache, the MFU, for this insertion. 2917 */ 2918static void 2919arc_get_data_buf(arc_buf_t *buf) 2920{ 2921 arc_state_t *state = buf->b_hdr->b_state; 2922 uint64_t size = buf->b_hdr->b_size; 2923 arc_buf_contents_t type = buf->b_hdr->b_type; 2924 2925 arc_adapt(size, state); 2926 2927 /* 2928 * We have not yet reached cache maximum size, 2929 * just allocate a new buffer. 2930 */ 2931 if (!arc_evict_needed(type)) { 2932 if (type == ARC_BUFC_METADATA) { 2933 buf->b_data = zio_buf_alloc(size); 2934 arc_space_consume(size, ARC_SPACE_DATA); 2935 } else { 2936 ASSERT(type == ARC_BUFC_DATA); 2937 buf->b_data = zio_data_buf_alloc(size); 2938 ARCSTAT_INCR(arcstat_data_size, size); 2939 atomic_add_64(&arc_size, size); 2940 } 2941 goto out; 2942 } 2943 2944 /* 2945 * If we are prefetching from the mfu ghost list, this buffer 2946 * will end up on the mru list; so steal space from there. 2947 */ 2948 if (state == arc_mfu_ghost) 2949 state = buf->b_hdr->b_flags & ARC_FLAG_PREFETCH ? 2950 arc_mru : arc_mfu; 2951 else if (state == arc_mru_ghost) 2952 state = arc_mru; 2953 2954 if (state == arc_mru || state == arc_anon) { 2955 uint64_t mru_used = arc_anon->arcs_size + arc_mru->arcs_size; 2956 state = (arc_mfu->arcs_lsize[type] >= size && 2957 arc_p > mru_used) ? arc_mfu : arc_mru; 2958 } else { 2959 /* MFU cases */ 2960 uint64_t mfu_space = arc_c - arc_p; 2961 state = (arc_mru->arcs_lsize[type] >= size && 2962 mfu_space > arc_mfu->arcs_size) ? arc_mru : arc_mfu; 2963 } 2964 if ((buf->b_data = arc_evict(state, 0, size, TRUE, type)) == NULL) { 2965 if (type == ARC_BUFC_METADATA) { 2966 buf->b_data = zio_buf_alloc(size); 2967 arc_space_consume(size, ARC_SPACE_DATA); 2968 } else { 2969 ASSERT(type == ARC_BUFC_DATA); 2970 buf->b_data = zio_data_buf_alloc(size); 2971 ARCSTAT_INCR(arcstat_data_size, size); 2972 atomic_add_64(&arc_size, size); 2973 } 2974 ARCSTAT_BUMP(arcstat_recycle_miss); 2975 } 2976 ASSERT(buf->b_data != NULL); 2977out: 2978 /* 2979 * Update the state size. Note that ghost states have a 2980 * "ghost size" and so don't need to be updated. 2981 */ 2982 if (!GHOST_STATE(buf->b_hdr->b_state)) { 2983 arc_buf_hdr_t *hdr = buf->b_hdr; 2984 2985 atomic_add_64(&hdr->b_state->arcs_size, size); 2986 if (list_link_active(&hdr->b_arc_node)) { 2987 ASSERT(refcount_is_zero(&hdr->b_refcnt)); 2988 atomic_add_64(&hdr->b_state->arcs_lsize[type], size); 2989 } 2990 /* 2991 * If we are growing the cache, and we are adding anonymous 2992 * data, and we have outgrown arc_p, update arc_p 2993 */ 2994 if (arc_size < arc_c && hdr->b_state == arc_anon && 2995 arc_anon->arcs_size + arc_mru->arcs_size > arc_p) 2996 arc_p = MIN(arc_c, arc_p + size); 2997 } 2998 ARCSTAT_BUMP(arcstat_allocated); 2999} 3000 3001/* 3002 * This routine is called whenever a buffer is accessed. 3003 * NOTE: the hash lock is dropped in this function. 3004 */ 3005static void 3006arc_access(arc_buf_hdr_t *hdr, kmutex_t *hash_lock) 3007{ 3008 clock_t now; 3009 3010 ASSERT(MUTEX_HELD(hash_lock)); 3011 3012 if (hdr->b_state == arc_anon) { 3013 /* 3014 * This buffer is not in the cache, and does not 3015 * appear in our "ghost" list. Add the new buffer 3016 * to the MRU state. 3017 */ 3018 3019 ASSERT(hdr->b_arc_access == 0); 3020 hdr->b_arc_access = ddi_get_lbolt(); 3021 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr); 3022 arc_change_state(arc_mru, hdr, hash_lock); 3023 3024 } else if (hdr->b_state == arc_mru) { 3025 now = ddi_get_lbolt(); 3026 3027 /* 3028 * If this buffer is here because of a prefetch, then either: 3029 * - clear the flag if this is a "referencing" read 3030 * (any subsequent access will bump this into the MFU state). 3031 * or 3032 * - move the buffer to the head of the list if this is 3033 * another prefetch (to make it less likely to be evicted). 3034 */ 3035 if ((hdr->b_flags & ARC_FLAG_PREFETCH) != 0) { 3036 if (refcount_count(&hdr->b_refcnt) == 0) { 3037 ASSERT(list_link_active(&hdr->b_arc_node)); 3038 } else { 3039 hdr->b_flags &= ~ARC_FLAG_PREFETCH; 3040 ARCSTAT_BUMP(arcstat_mru_hits); 3041 } 3042 hdr->b_arc_access = now; 3043 return; 3044 } 3045 3046 /* 3047 * This buffer has been "accessed" only once so far, 3048 * but it is still in the cache. Move it to the MFU 3049 * state. 3050 */ 3051 if (now > hdr->b_arc_access + ARC_MINTIME) { 3052 /* 3053 * More than 125ms have passed since we 3054 * instantiated this buffer. Move it to the 3055 * most frequently used state. 3056 */ 3057 hdr->b_arc_access = now; 3058 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 3059 arc_change_state(arc_mfu, hdr, hash_lock); 3060 } 3061 ARCSTAT_BUMP(arcstat_mru_hits); 3062 } else if (hdr->b_state == arc_mru_ghost) { 3063 arc_state_t *new_state; 3064 /* 3065 * This buffer has been "accessed" recently, but 3066 * was evicted from the cache. Move it to the 3067 * MFU state. 3068 */ 3069 3070 if (hdr->b_flags & ARC_FLAG_PREFETCH) { 3071 new_state = arc_mru; 3072 if (refcount_count(&hdr->b_refcnt) > 0) 3073 hdr->b_flags &= ~ARC_FLAG_PREFETCH; 3074 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr); 3075 } else { 3076 new_state = arc_mfu; 3077 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 3078 } 3079 3080 hdr->b_arc_access = ddi_get_lbolt(); 3081 arc_change_state(new_state, hdr, hash_lock); 3082 3083 ARCSTAT_BUMP(arcstat_mru_ghost_hits); 3084 } else if (hdr->b_state == arc_mfu) { 3085 /* 3086 * This buffer has been accessed more than once and is 3087 * still in the cache. Keep it in the MFU state. 3088 * 3089 * NOTE: an add_reference() that occurred when we did 3090 * the arc_read() will have kicked this off the list. 3091 * If it was a prefetch, we will explicitly move it to 3092 * the head of the list now. 3093 */ 3094 if ((hdr->b_flags & ARC_FLAG_PREFETCH) != 0) { 3095 ASSERT(refcount_count(&hdr->b_refcnt) == 0); 3096 ASSERT(list_link_active(&hdr->b_arc_node)); 3097 } 3098 ARCSTAT_BUMP(arcstat_mfu_hits); 3099 hdr->b_arc_access = ddi_get_lbolt(); 3100 } else if (hdr->b_state == arc_mfu_ghost) { 3101 arc_state_t *new_state = arc_mfu; 3102 /* 3103 * This buffer has been accessed more than once but has 3104 * been evicted from the cache. Move it back to the 3105 * MFU state. 3106 */ 3107 3108 if (hdr->b_flags & ARC_FLAG_PREFETCH) { 3109 /* 3110 * This is a prefetch access... 3111 * move this block back to the MRU state. 3112 */ 3113 ASSERT0(refcount_count(&hdr->b_refcnt)); 3114 new_state = arc_mru; 3115 } 3116 3117 hdr->b_arc_access = ddi_get_lbolt(); 3118 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 3119 arc_change_state(new_state, hdr, hash_lock); 3120 3121 ARCSTAT_BUMP(arcstat_mfu_ghost_hits); 3122 } else if (hdr->b_state == arc_l2c_only) { 3123 /* 3124 * This buffer is on the 2nd Level ARC. 3125 */ 3126 3127 hdr->b_arc_access = ddi_get_lbolt(); 3128 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 3129 arc_change_state(arc_mfu, hdr, hash_lock); 3130 } else { 3131 ASSERT(!"invalid arc state"); 3132 } 3133} 3134 3135/* a generic arc_done_func_t which you can use */ 3136/* ARGSUSED */ 3137void 3138arc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg) 3139{ 3140 if (zio == NULL || zio->io_error == 0) 3141 bcopy(buf->b_data, arg, buf->b_hdr->b_size); 3142 VERIFY(arc_buf_remove_ref(buf, arg)); 3143} 3144 3145/* a generic arc_done_func_t */ 3146void 3147arc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg) 3148{ 3149 arc_buf_t **bufp = arg; 3150 if (zio && zio->io_error) { 3151 VERIFY(arc_buf_remove_ref(buf, arg)); 3152 *bufp = NULL; 3153 } else { 3154 *bufp = buf; 3155 ASSERT(buf->b_data); 3156 } 3157} 3158 3159static void 3160arc_read_done(zio_t *zio) 3161{ 3162 arc_buf_hdr_t *hdr; 3163 arc_buf_t *buf; 3164 arc_buf_t *abuf; /* buffer we're assigning to callback */ 3165 kmutex_t *hash_lock = NULL; 3166 arc_callback_t *callback_list, *acb; 3167 int freeable = FALSE; 3168 3169 buf = zio->io_private; 3170 hdr = buf->b_hdr; 3171 3172 /* 3173 * The hdr was inserted into hash-table and removed from lists 3174 * prior to starting I/O. We should find this header, since 3175 * it's in the hash table, and it should be legit since it's 3176 * not possible to evict it during the I/O. The only possible 3177 * reason for it not to be found is if we were freed during the 3178 * read. 3179 */ 3180 if (HDR_IN_HASH_TABLE(hdr)) { 3181 ASSERT3U(hdr->b_birth, ==, BP_PHYSICAL_BIRTH(zio->io_bp)); 3182 ASSERT3U(hdr->b_dva.dva_word[0], ==, 3183 BP_IDENTITY(zio->io_bp)->dva_word[0]); 3184 ASSERT3U(hdr->b_dva.dva_word[1], ==, 3185 BP_IDENTITY(zio->io_bp)->dva_word[1]); 3186 3187 arc_buf_hdr_t *found = buf_hash_find(hdr->b_spa, zio->io_bp, 3188 &hash_lock); 3189 3190 ASSERT((found == NULL && HDR_FREED_IN_READ(hdr) && 3191 hash_lock == NULL) || 3192 (found == hdr && 3193 DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) || 3194 (found == hdr && HDR_L2_READING(hdr))); 3195 } 3196 3197 hdr->b_flags &= ~ARC_FLAG_L2_EVICTED; 3198 if (l2arc_noprefetch && (hdr->b_flags & ARC_FLAG_PREFETCH)) 3199 hdr->b_flags &= ~ARC_FLAG_L2CACHE; 3200 3201 /* byteswap if necessary */ 3202 callback_list = hdr->b_acb; 3203 ASSERT(callback_list != NULL); 3204 if (BP_SHOULD_BYTESWAP(zio->io_bp) && zio->io_error == 0) { 3205 dmu_object_byteswap_t bswap = 3206 DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp)); 3207 arc_byteswap_func_t *func = BP_GET_LEVEL(zio->io_bp) > 0 ? 3208 byteswap_uint64_array : 3209 dmu_ot_byteswap[bswap].ob_func; 3210 func(buf->b_data, hdr->b_size); 3211 } 3212 3213 arc_cksum_compute(buf, B_FALSE); 3214#ifdef illumos 3215 arc_buf_watch(buf); 3216#endif 3217 3218 if (hash_lock && zio->io_error == 0 && hdr->b_state == arc_anon) { 3219 /* 3220 * Only call arc_access on anonymous buffers. This is because 3221 * if we've issued an I/O for an evicted buffer, we've already 3222 * called arc_access (to prevent any simultaneous readers from 3223 * getting confused). 3224 */ 3225 arc_access(hdr, hash_lock); 3226 } 3227 3228 /* create copies of the data buffer for the callers */ 3229 abuf = buf; 3230 for (acb = callback_list; acb; acb = acb->acb_next) { 3231 if (acb->acb_done) { 3232 if (abuf == NULL) { 3233 ARCSTAT_BUMP(arcstat_duplicate_reads); 3234 abuf = arc_buf_clone(buf); 3235 } 3236 acb->acb_buf = abuf; 3237 abuf = NULL; 3238 } 3239 } 3240 hdr->b_acb = NULL; 3241 hdr->b_flags &= ~ARC_FLAG_IO_IN_PROGRESS; 3242 ASSERT(!HDR_BUF_AVAILABLE(hdr)); 3243 if (abuf == buf) { 3244 ASSERT(buf->b_efunc == NULL); 3245 ASSERT(hdr->b_datacnt == 1); 3246 hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE; 3247 } 3248 3249 ASSERT(refcount_is_zero(&hdr->b_refcnt) || callback_list != NULL); 3250 3251 if (zio->io_error != 0) { 3252 hdr->b_flags |= ARC_FLAG_IO_ERROR; 3253 if (hdr->b_state != arc_anon) 3254 arc_change_state(arc_anon, hdr, hash_lock); 3255 if (HDR_IN_HASH_TABLE(hdr)) 3256 buf_hash_remove(hdr); 3257 freeable = refcount_is_zero(&hdr->b_refcnt); 3258 } 3259 3260 /* 3261 * Broadcast before we drop the hash_lock to avoid the possibility 3262 * that the hdr (and hence the cv) might be freed before we get to 3263 * the cv_broadcast(). 3264 */ 3265 cv_broadcast(&hdr->b_cv); 3266 3267 if (hash_lock) { 3268 mutex_exit(hash_lock); 3269 } else { 3270 /* 3271 * This block was freed while we waited for the read to 3272 * complete. It has been removed from the hash table and 3273 * moved to the anonymous state (so that it won't show up 3274 * in the cache). 3275 */ 3276 ASSERT3P(hdr->b_state, ==, arc_anon); 3277 freeable = refcount_is_zero(&hdr->b_refcnt); 3278 } 3279 3280 /* execute each callback and free its structure */ 3281 while ((acb = callback_list) != NULL) { 3282 if (acb->acb_done) 3283 acb->acb_done(zio, acb->acb_buf, acb->acb_private); 3284 3285 if (acb->acb_zio_dummy != NULL) { 3286 acb->acb_zio_dummy->io_error = zio->io_error; 3287 zio_nowait(acb->acb_zio_dummy); 3288 } 3289 3290 callback_list = acb->acb_next; 3291 kmem_free(acb, sizeof (arc_callback_t)); 3292 } 3293 3294 if (freeable) 3295 arc_hdr_destroy(hdr); 3296} 3297 3298/* 3299 * "Read" the block block at the specified DVA (in bp) via the 3300 * cache. If the block is found in the cache, invoke the provided 3301 * callback immediately and return. Note that the `zio' parameter 3302 * in the callback will be NULL in this case, since no IO was 3303 * required. If the block is not in the cache pass the read request 3304 * on to the spa with a substitute callback function, so that the 3305 * requested block will be added to the cache. 3306 * 3307 * If a read request arrives for a block that has a read in-progress, 3308 * either wait for the in-progress read to complete (and return the 3309 * results); or, if this is a read with a "done" func, add a record 3310 * to the read to invoke the "done" func when the read completes, 3311 * and return; or just return. 3312 * 3313 * arc_read_done() will invoke all the requested "done" functions 3314 * for readers of this block. 3315 */ 3316int 3317arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_done_func_t *done, 3318 void *private, zio_priority_t priority, int zio_flags, 3319 arc_flags_t *arc_flags, const zbookmark_phys_t *zb) 3320{ 3321 arc_buf_hdr_t *hdr = NULL; 3322 arc_buf_t *buf = NULL; 3323 kmutex_t *hash_lock = NULL; 3324 zio_t *rzio; 3325 uint64_t guid = spa_load_guid(spa); 3326 3327 ASSERT(!BP_IS_EMBEDDED(bp) || 3328 BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA); 3329 3330top: 3331 if (!BP_IS_EMBEDDED(bp)) { 3332 /* 3333 * Embedded BP's have no DVA and require no I/O to "read". 3334 * Create an anonymous arc buf to back it. 3335 */ 3336 hdr = buf_hash_find(guid, bp, &hash_lock); 3337 } 3338 3339 if (hdr != NULL && hdr->b_datacnt > 0) { 3340 3341 *arc_flags |= ARC_FLAG_CACHED; 3342 3343 if (HDR_IO_IN_PROGRESS(hdr)) { 3344 3345 if (*arc_flags & ARC_FLAG_WAIT) { 3346 cv_wait(&hdr->b_cv, hash_lock); 3347 mutex_exit(hash_lock); 3348 goto top; 3349 } 3350 ASSERT(*arc_flags & ARC_FLAG_NOWAIT); 3351 3352 if (done) { 3353 arc_callback_t *acb = NULL; 3354 3355 acb = kmem_zalloc(sizeof (arc_callback_t), 3356 KM_SLEEP); 3357 acb->acb_done = done; 3358 acb->acb_private = private; 3359 if (pio != NULL) 3360 acb->acb_zio_dummy = zio_null(pio, 3361 spa, NULL, NULL, NULL, zio_flags); 3362 3363 ASSERT(acb->acb_done != NULL); 3364 acb->acb_next = hdr->b_acb; 3365 hdr->b_acb = acb; 3366 add_reference(hdr, hash_lock, private); 3367 mutex_exit(hash_lock); 3368 return (0); 3369 } 3370 mutex_exit(hash_lock); 3371 return (0); 3372 } 3373 3374 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu); 3375 3376 if (done) { 3377 add_reference(hdr, hash_lock, private); 3378 /* 3379 * If this block is already in use, create a new 3380 * copy of the data so that we will be guaranteed 3381 * that arc_release() will always succeed. 3382 */ 3383 buf = hdr->b_buf; 3384 ASSERT(buf); 3385 ASSERT(buf->b_data); 3386 if (HDR_BUF_AVAILABLE(hdr)) { 3387 ASSERT(buf->b_efunc == NULL); 3388 hdr->b_flags &= ~ARC_FLAG_BUF_AVAILABLE; 3389 } else { 3390 buf = arc_buf_clone(buf); 3391 } 3392 3393 } else if (*arc_flags & ARC_FLAG_PREFETCH && 3394 refcount_count(&hdr->b_refcnt) == 0) { 3395 hdr->b_flags |= ARC_FLAG_PREFETCH; 3396 } 3397 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr); 3398 arc_access(hdr, hash_lock); 3399 if (*arc_flags & ARC_FLAG_L2CACHE) 3400 hdr->b_flags |= ARC_FLAG_L2CACHE; 3401 if (*arc_flags & ARC_FLAG_L2COMPRESS) 3402 hdr->b_flags |= ARC_FLAG_L2COMPRESS; 3403 mutex_exit(hash_lock); 3404 ARCSTAT_BUMP(arcstat_hits); 3405 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_FLAG_PREFETCH), 3406 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA, 3407 data, metadata, hits); 3408 3409 if (done) 3410 done(NULL, buf, private); 3411 } else { 3412 uint64_t size = BP_GET_LSIZE(bp); 3413 arc_callback_t *acb; 3414 vdev_t *vd = NULL; 3415 uint64_t addr = 0; 3416 boolean_t devw = B_FALSE; 3417 enum zio_compress b_compress = ZIO_COMPRESS_OFF; 3418 uint64_t b_asize = 0; 3419 3420 if (hdr == NULL) { 3421 /* this block is not in the cache */ 3422 arc_buf_hdr_t *exists = NULL; 3423 arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp); 3424 buf = arc_buf_alloc(spa, size, private, type); 3425 hdr = buf->b_hdr; 3426 if (!BP_IS_EMBEDDED(bp)) { 3427 hdr->b_dva = *BP_IDENTITY(bp); 3428 hdr->b_birth = BP_PHYSICAL_BIRTH(bp); 3429 hdr->b_cksum0 = bp->blk_cksum.zc_word[0]; 3430 exists = buf_hash_insert(hdr, &hash_lock); 3431 } 3432 if (exists != NULL) { 3433 /* somebody beat us to the hash insert */ 3434 mutex_exit(hash_lock); 3435 buf_discard_identity(hdr); 3436 (void) arc_buf_remove_ref(buf, private); 3437 goto top; /* restart the IO request */ 3438 } 3439 3440 /* if this is a prefetch, we don't have a reference */ 3441 if (*arc_flags & ARC_FLAG_PREFETCH) { 3442 (void) remove_reference(hdr, hash_lock, 3443 private); 3444 hdr->b_flags |= ARC_FLAG_PREFETCH; 3445 } 3446 if (*arc_flags & ARC_FLAG_L2CACHE) 3447 hdr->b_flags |= ARC_FLAG_L2CACHE; 3448 if (*arc_flags & ARC_FLAG_L2COMPRESS) 3449 hdr->b_flags |= ARC_FLAG_L2COMPRESS; 3450 if (BP_GET_LEVEL(bp) > 0) 3451 hdr->b_flags |= ARC_FLAG_INDIRECT; 3452 } else { 3453 /* this block is in the ghost cache */ 3454 ASSERT(GHOST_STATE(hdr->b_state)); 3455 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 3456 ASSERT0(refcount_count(&hdr->b_refcnt)); 3457 ASSERT(hdr->b_buf == NULL); 3458 3459 /* if this is a prefetch, we don't have a reference */ 3460 if (*arc_flags & ARC_FLAG_PREFETCH) 3461 hdr->b_flags |= ARC_FLAG_PREFETCH; 3462 else 3463 add_reference(hdr, hash_lock, private); 3464 if (*arc_flags & ARC_FLAG_L2CACHE) 3465 hdr->b_flags |= ARC_FLAG_L2CACHE; 3466 if (*arc_flags & ARC_FLAG_L2COMPRESS) 3467 hdr->b_flags |= ARC_FLAG_L2COMPRESS; 3468 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 3469 buf->b_hdr = hdr; 3470 buf->b_data = NULL; 3471 buf->b_efunc = NULL; 3472 buf->b_private = NULL; 3473 buf->b_next = NULL; 3474 hdr->b_buf = buf; 3475 ASSERT(hdr->b_datacnt == 0); 3476 hdr->b_datacnt = 1; 3477 arc_get_data_buf(buf); 3478 arc_access(hdr, hash_lock); 3479 } 3480 3481 ASSERT(!GHOST_STATE(hdr->b_state)); 3482 3483 acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP); 3484 acb->acb_done = done; 3485 acb->acb_private = private; 3486 3487 ASSERT(hdr->b_acb == NULL); 3488 hdr->b_acb = acb; 3489 hdr->b_flags |= ARC_FLAG_IO_IN_PROGRESS; 3490 3491 if (hdr->b_l2hdr != NULL && 3492 (vd = hdr->b_l2hdr->b_dev->l2ad_vdev) != NULL) { 3493 devw = hdr->b_l2hdr->b_dev->l2ad_writing; 3494 addr = hdr->b_l2hdr->b_daddr; 3495 b_compress = hdr->b_l2hdr->b_compress; 3496 b_asize = hdr->b_l2hdr->b_asize; 3497 /* 3498 * Lock out device removal. 3499 */ 3500 if (vdev_is_dead(vd) || 3501 !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER)) 3502 vd = NULL; 3503 } 3504 3505 if (hash_lock != NULL) 3506 mutex_exit(hash_lock); 3507 3508 /* 3509 * At this point, we have a level 1 cache miss. Try again in 3510 * L2ARC if possible. 3511 */ 3512 ASSERT3U(hdr->b_size, ==, size); 3513 DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp, 3514 uint64_t, size, zbookmark_phys_t *, zb); 3515 ARCSTAT_BUMP(arcstat_misses); 3516 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_FLAG_PREFETCH), 3517 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA, 3518 data, metadata, misses); 3519#ifdef _KERNEL 3520 curthread->td_ru.ru_inblock++; 3521#endif 3522 3523 if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) { 3524 /* 3525 * Read from the L2ARC if the following are true: 3526 * 1. The L2ARC vdev was previously cached. 3527 * 2. This buffer still has L2ARC metadata. 3528 * 3. This buffer isn't currently writing to the L2ARC. 3529 * 4. The L2ARC entry wasn't evicted, which may 3530 * also have invalidated the vdev. 3531 * 5. This isn't prefetch and l2arc_noprefetch is set. 3532 */ 3533 if (hdr->b_l2hdr != NULL && 3534 !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) && 3535 !(l2arc_noprefetch && HDR_PREFETCH(hdr))) { 3536 l2arc_read_callback_t *cb; 3537 3538 DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr); 3539 ARCSTAT_BUMP(arcstat_l2_hits); 3540 3541 cb = kmem_zalloc(sizeof (l2arc_read_callback_t), 3542 KM_SLEEP); 3543 cb->l2rcb_buf = buf; 3544 cb->l2rcb_spa = spa; 3545 cb->l2rcb_bp = *bp; 3546 cb->l2rcb_zb = *zb; 3547 cb->l2rcb_flags = zio_flags; 3548 cb->l2rcb_compress = b_compress; 3549 3550 ASSERT(addr >= VDEV_LABEL_START_SIZE && 3551 addr + size < vd->vdev_psize - 3552 VDEV_LABEL_END_SIZE); 3553 3554 /* 3555 * l2arc read. The SCL_L2ARC lock will be 3556 * released by l2arc_read_done(). 3557 * Issue a null zio if the underlying buffer 3558 * was squashed to zero size by compression. 3559 */ 3560 if (b_compress == ZIO_COMPRESS_EMPTY) { 3561 rzio = zio_null(pio, spa, vd, 3562 l2arc_read_done, cb, 3563 zio_flags | ZIO_FLAG_DONT_CACHE | 3564 ZIO_FLAG_CANFAIL | 3565 ZIO_FLAG_DONT_PROPAGATE | 3566 ZIO_FLAG_DONT_RETRY); 3567 } else { 3568 rzio = zio_read_phys(pio, vd, addr, 3569 b_asize, buf->b_data, 3570 ZIO_CHECKSUM_OFF, 3571 l2arc_read_done, cb, priority, 3572 zio_flags | ZIO_FLAG_DONT_CACHE | 3573 ZIO_FLAG_CANFAIL | 3574 ZIO_FLAG_DONT_PROPAGATE | 3575 ZIO_FLAG_DONT_RETRY, B_FALSE); 3576 } 3577 DTRACE_PROBE2(l2arc__read, vdev_t *, vd, 3578 zio_t *, rzio); 3579 ARCSTAT_INCR(arcstat_l2_read_bytes, b_asize); 3580 3581 if (*arc_flags & ARC_FLAG_NOWAIT) { 3582 zio_nowait(rzio); 3583 return (0); 3584 } 3585 3586 ASSERT(*arc_flags & ARC_FLAG_WAIT); 3587 if (zio_wait(rzio) == 0) 3588 return (0); 3589 3590 /* l2arc read error; goto zio_read() */ 3591 } else { 3592 DTRACE_PROBE1(l2arc__miss, 3593 arc_buf_hdr_t *, hdr); 3594 ARCSTAT_BUMP(arcstat_l2_misses); 3595 if (HDR_L2_WRITING(hdr)) 3596 ARCSTAT_BUMP(arcstat_l2_rw_clash); 3597 spa_config_exit(spa, SCL_L2ARC, vd); 3598 } 3599 } else { 3600 if (vd != NULL) 3601 spa_config_exit(spa, SCL_L2ARC, vd); 3602 if (l2arc_ndev != 0) { 3603 DTRACE_PROBE1(l2arc__miss, 3604 arc_buf_hdr_t *, hdr); 3605 ARCSTAT_BUMP(arcstat_l2_misses); 3606 } 3607 } 3608 3609 rzio = zio_read(pio, spa, bp, buf->b_data, size, 3610 arc_read_done, buf, priority, zio_flags, zb); 3611 3612 if (*arc_flags & ARC_FLAG_WAIT) 3613 return (zio_wait(rzio)); 3614 3615 ASSERT(*arc_flags & ARC_FLAG_NOWAIT); 3616 zio_nowait(rzio); 3617 } 3618 return (0); 3619} 3620 3621void 3622arc_set_callback(arc_buf_t *buf, arc_evict_func_t *func, void *private) 3623{ 3624 ASSERT(buf->b_hdr != NULL); 3625 ASSERT(buf->b_hdr->b_state != arc_anon); 3626 ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt) || func == NULL); 3627 ASSERT(buf->b_efunc == NULL); 3628 ASSERT(!HDR_BUF_AVAILABLE(buf->b_hdr)); 3629 3630 buf->b_efunc = func; 3631 buf->b_private = private; 3632} 3633 3634/* 3635 * Notify the arc that a block was freed, and thus will never be used again. 3636 */ 3637void 3638arc_freed(spa_t *spa, const blkptr_t *bp) 3639{ 3640 arc_buf_hdr_t *hdr; 3641 kmutex_t *hash_lock; 3642 uint64_t guid = spa_load_guid(spa); 3643 3644 ASSERT(!BP_IS_EMBEDDED(bp)); 3645 3646 hdr = buf_hash_find(guid, bp, &hash_lock); 3647 if (hdr == NULL) 3648 return; 3649 if (HDR_BUF_AVAILABLE(hdr)) { 3650 arc_buf_t *buf = hdr->b_buf; 3651 add_reference(hdr, hash_lock, FTAG); 3652 hdr->b_flags &= ~ARC_FLAG_BUF_AVAILABLE; 3653 mutex_exit(hash_lock); 3654 3655 arc_release(buf, FTAG); 3656 (void) arc_buf_remove_ref(buf, FTAG); 3657 } else { 3658 mutex_exit(hash_lock); 3659 } 3660 3661} 3662 3663/* 3664 * Clear the user eviction callback set by arc_set_callback(), first calling 3665 * it if it exists. Because the presence of a callback keeps an arc_buf cached 3666 * clearing the callback may result in the arc_buf being destroyed. However, 3667 * it will not result in the *last* arc_buf being destroyed, hence the data 3668 * will remain cached in the ARC. We make a copy of the arc buffer here so 3669 * that we can process the callback without holding any locks. 3670 * 3671 * It's possible that the callback is already in the process of being cleared 3672 * by another thread. In this case we can not clear the callback. 3673 * 3674 * Returns B_TRUE if the callback was successfully called and cleared. 3675 */ 3676boolean_t 3677arc_clear_callback(arc_buf_t *buf) 3678{ 3679 arc_buf_hdr_t *hdr; 3680 kmutex_t *hash_lock; 3681 arc_evict_func_t *efunc = buf->b_efunc; 3682 void *private = buf->b_private; 3683 list_t *list, *evicted_list; 3684 kmutex_t *lock, *evicted_lock; 3685 3686 mutex_enter(&buf->b_evict_lock); 3687 hdr = buf->b_hdr; 3688 if (hdr == NULL) { 3689 /* 3690 * We are in arc_do_user_evicts(). 3691 */ 3692 ASSERT(buf->b_data == NULL); 3693 mutex_exit(&buf->b_evict_lock); 3694 return (B_FALSE); 3695 } else if (buf->b_data == NULL) { 3696 /* 3697 * We are on the eviction list; process this buffer now 3698 * but let arc_do_user_evicts() do the reaping. 3699 */ 3700 buf->b_efunc = NULL; 3701 mutex_exit(&buf->b_evict_lock); 3702 VERIFY0(efunc(private)); 3703 return (B_TRUE); 3704 } 3705 hash_lock = HDR_LOCK(hdr); 3706 mutex_enter(hash_lock); 3707 hdr = buf->b_hdr; 3708 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 3709 3710 ASSERT3U(refcount_count(&hdr->b_refcnt), <, hdr->b_datacnt); 3711 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu); 3712 3713 buf->b_efunc = NULL; 3714 buf->b_private = NULL; 3715 3716 if (hdr->b_datacnt > 1) { 3717 mutex_exit(&buf->b_evict_lock); 3718 arc_buf_destroy(buf, FALSE, TRUE); 3719 } else { 3720 ASSERT(buf == hdr->b_buf); 3721 hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE; 3722 mutex_exit(&buf->b_evict_lock); 3723 } 3724 3725 mutex_exit(hash_lock); 3726 VERIFY0(efunc(private)); 3727 return (B_TRUE); 3728} 3729 3730/* 3731 * Release this buffer from the cache, making it an anonymous buffer. This 3732 * must be done after a read and prior to modifying the buffer contents. 3733 * If the buffer has more than one reference, we must make 3734 * a new hdr for the buffer. 3735 */ 3736void 3737arc_release(arc_buf_t *buf, void *tag) 3738{ 3739 arc_buf_hdr_t *hdr; 3740 kmutex_t *hash_lock = NULL; 3741 l2arc_buf_hdr_t *l2hdr; 3742 uint64_t buf_size; 3743 3744 /* 3745 * It would be nice to assert that if it's DMU metadata (level > 3746 * 0 || it's the dnode file), then it must be syncing context. 3747 * But we don't know that information at this level. 3748 */ 3749 3750 mutex_enter(&buf->b_evict_lock); 3751 hdr = buf->b_hdr; 3752 3753 /* this buffer is not on any list */ 3754 ASSERT(refcount_count(&hdr->b_refcnt) > 0); 3755 3756 if (hdr->b_state == arc_anon) { 3757 /* this buffer is already released */ 3758 ASSERT(buf->b_efunc == NULL); 3759 } else { 3760 hash_lock = HDR_LOCK(hdr); 3761 mutex_enter(hash_lock); 3762 hdr = buf->b_hdr; 3763 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 3764 } 3765 3766 l2hdr = hdr->b_l2hdr; 3767 if (l2hdr) { 3768 mutex_enter(&l2arc_buflist_mtx); 3769 arc_buf_l2_cdata_free(hdr); 3770 hdr->b_l2hdr = NULL; 3771 list_remove(l2hdr->b_dev->l2ad_buflist, hdr); 3772 } 3773 buf_size = hdr->b_size; 3774 3775 /* 3776 * Do we have more than one buf? 3777 */ 3778 if (hdr->b_datacnt > 1) { 3779 arc_buf_hdr_t *nhdr; 3780 arc_buf_t **bufp; 3781 uint64_t blksz = hdr->b_size; 3782 uint64_t spa = hdr->b_spa; 3783 arc_buf_contents_t type = hdr->b_type; 3784 uint32_t flags = hdr->b_flags; 3785 3786 ASSERT(hdr->b_buf != buf || buf->b_next != NULL); 3787 /* 3788 * Pull the data off of this hdr and attach it to 3789 * a new anonymous hdr. 3790 */ 3791 (void) remove_reference(hdr, hash_lock, tag); 3792 bufp = &hdr->b_buf; 3793 while (*bufp != buf) 3794 bufp = &(*bufp)->b_next; 3795 *bufp = buf->b_next; 3796 buf->b_next = NULL; 3797 3798 ASSERT3U(hdr->b_state->arcs_size, >=, hdr->b_size); 3799 atomic_add_64(&hdr->b_state->arcs_size, -hdr->b_size); 3800 if (refcount_is_zero(&hdr->b_refcnt)) { 3801 uint64_t *size = &hdr->b_state->arcs_lsize[hdr->b_type]; 3802 ASSERT3U(*size, >=, hdr->b_size); 3803 atomic_add_64(size, -hdr->b_size); 3804 } 3805 3806 /* 3807 * We're releasing a duplicate user data buffer, update 3808 * our statistics accordingly. 3809 */ 3810 if (hdr->b_type == ARC_BUFC_DATA) { 3811 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers); 3812 ARCSTAT_INCR(arcstat_duplicate_buffers_size, 3813 -hdr->b_size); 3814 } 3815 hdr->b_datacnt -= 1; 3816 arc_cksum_verify(buf); 3817#ifdef illumos 3818 arc_buf_unwatch(buf); 3819#endif 3820 3821 mutex_exit(hash_lock); 3822 3823 nhdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE); 3824 nhdr->b_size = blksz; 3825 nhdr->b_spa = spa; 3826 nhdr->b_type = type; 3827 nhdr->b_buf = buf; 3828 nhdr->b_state = arc_anon; 3829 nhdr->b_arc_access = 0; 3830 nhdr->b_flags = flags & ARC_FLAG_L2_WRITING; 3831 nhdr->b_l2hdr = NULL; 3832 nhdr->b_datacnt = 1; 3833 nhdr->b_freeze_cksum = NULL; 3834 (void) refcount_add(&nhdr->b_refcnt, tag); 3835 buf->b_hdr = nhdr; 3836 mutex_exit(&buf->b_evict_lock); 3837 atomic_add_64(&arc_anon->arcs_size, blksz); 3838 } else { 3839 mutex_exit(&buf->b_evict_lock); 3840 ASSERT(refcount_count(&hdr->b_refcnt) == 1); 3841 ASSERT(!list_link_active(&hdr->b_arc_node)); 3842 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 3843 if (hdr->b_state != arc_anon) 3844 arc_change_state(arc_anon, hdr, hash_lock); 3845 hdr->b_arc_access = 0; 3846 if (hash_lock) 3847 mutex_exit(hash_lock); 3848 3849 buf_discard_identity(hdr); 3850 arc_buf_thaw(buf); 3851 } 3852 buf->b_efunc = NULL; 3853 buf->b_private = NULL; 3854 3855 if (l2hdr) { 3856 ARCSTAT_INCR(arcstat_l2_asize, -l2hdr->b_asize); 3857 vdev_space_update(l2hdr->b_dev->l2ad_vdev, 3858 -l2hdr->b_asize, 0, 0); 3859 trim_map_free(l2hdr->b_dev->l2ad_vdev, l2hdr->b_daddr, 3860 hdr->b_size, 0); 3861 kmem_free(l2hdr, sizeof (l2arc_buf_hdr_t)); 3862 ARCSTAT_INCR(arcstat_l2_size, -buf_size); 3863 mutex_exit(&l2arc_buflist_mtx); 3864 } 3865} 3866 3867int 3868arc_released(arc_buf_t *buf) 3869{ 3870 int released; 3871 3872 mutex_enter(&buf->b_evict_lock); 3873 released = (buf->b_data != NULL && buf->b_hdr->b_state == arc_anon); 3874 mutex_exit(&buf->b_evict_lock); 3875 return (released); 3876} 3877 3878#ifdef ZFS_DEBUG 3879int 3880arc_referenced(arc_buf_t *buf) 3881{ 3882 int referenced; 3883 3884 mutex_enter(&buf->b_evict_lock); 3885 referenced = (refcount_count(&buf->b_hdr->b_refcnt)); 3886 mutex_exit(&buf->b_evict_lock); 3887 return (referenced); 3888} 3889#endif 3890 3891static void 3892arc_write_ready(zio_t *zio) 3893{ 3894 arc_write_callback_t *callback = zio->io_private; 3895 arc_buf_t *buf = callback->awcb_buf; 3896 arc_buf_hdr_t *hdr = buf->b_hdr; 3897 3898 ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt)); 3899 callback->awcb_ready(zio, buf, callback->awcb_private); 3900 3901 /* 3902 * If the IO is already in progress, then this is a re-write 3903 * attempt, so we need to thaw and re-compute the cksum. 3904 * It is the responsibility of the callback to handle the 3905 * accounting for any re-write attempt. 3906 */ 3907 if (HDR_IO_IN_PROGRESS(hdr)) { 3908 mutex_enter(&hdr->b_freeze_lock); 3909 if (hdr->b_freeze_cksum != NULL) { 3910 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 3911 hdr->b_freeze_cksum = NULL; 3912 } 3913 mutex_exit(&hdr->b_freeze_lock); 3914 } 3915 arc_cksum_compute(buf, B_FALSE); 3916 hdr->b_flags |= ARC_FLAG_IO_IN_PROGRESS; 3917} 3918 3919/* 3920 * The SPA calls this callback for each physical write that happens on behalf 3921 * of a logical write. See the comment in dbuf_write_physdone() for details. 3922 */ 3923static void 3924arc_write_physdone(zio_t *zio) 3925{ 3926 arc_write_callback_t *cb = zio->io_private; 3927 if (cb->awcb_physdone != NULL) 3928 cb->awcb_physdone(zio, cb->awcb_buf, cb->awcb_private); 3929} 3930 3931static void 3932arc_write_done(zio_t *zio) 3933{ 3934 arc_write_callback_t *callback = zio->io_private; 3935 arc_buf_t *buf = callback->awcb_buf; 3936 arc_buf_hdr_t *hdr = buf->b_hdr; 3937 3938 ASSERT(hdr->b_acb == NULL); 3939 3940 if (zio->io_error == 0) { 3941 if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) { 3942 buf_discard_identity(hdr); 3943 } else { 3944 hdr->b_dva = *BP_IDENTITY(zio->io_bp); 3945 hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp); 3946 hdr->b_cksum0 = zio->io_bp->blk_cksum.zc_word[0]; 3947 } 3948 } else { 3949 ASSERT(BUF_EMPTY(hdr)); 3950 } 3951 3952 /* 3953 * If the block to be written was all-zero or compressed enough to be 3954 * embedded in the BP, no write was performed so there will be no 3955 * dva/birth/checksum. The buffer must therefore remain anonymous 3956 * (and uncached). 3957 */ 3958 if (!BUF_EMPTY(hdr)) { 3959 arc_buf_hdr_t *exists; 3960 kmutex_t *hash_lock; 3961 3962 ASSERT(zio->io_error == 0); 3963 3964 arc_cksum_verify(buf); 3965 3966 exists = buf_hash_insert(hdr, &hash_lock); 3967 if (exists) { 3968 /* 3969 * This can only happen if we overwrite for 3970 * sync-to-convergence, because we remove 3971 * buffers from the hash table when we arc_free(). 3972 */ 3973 if (zio->io_flags & ZIO_FLAG_IO_REWRITE) { 3974 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) 3975 panic("bad overwrite, hdr=%p exists=%p", 3976 (void *)hdr, (void *)exists); 3977 ASSERT(refcount_is_zero(&exists->b_refcnt)); 3978 arc_change_state(arc_anon, exists, hash_lock); 3979 mutex_exit(hash_lock); 3980 arc_hdr_destroy(exists); 3981 exists = buf_hash_insert(hdr, &hash_lock); 3982 ASSERT3P(exists, ==, NULL); 3983 } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) { 3984 /* nopwrite */ 3985 ASSERT(zio->io_prop.zp_nopwrite); 3986 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) 3987 panic("bad nopwrite, hdr=%p exists=%p", 3988 (void *)hdr, (void *)exists); 3989 } else { 3990 /* Dedup */ 3991 ASSERT(hdr->b_datacnt == 1); 3992 ASSERT(hdr->b_state == arc_anon); 3993 ASSERT(BP_GET_DEDUP(zio->io_bp)); 3994 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0); 3995 } 3996 } 3997 hdr->b_flags &= ~ARC_FLAG_IO_IN_PROGRESS; 3998 /* if it's not anon, we are doing a scrub */ 3999 if (!exists && hdr->b_state == arc_anon) 4000 arc_access(hdr, hash_lock); 4001 mutex_exit(hash_lock); 4002 } else { 4003 hdr->b_flags &= ~ARC_FLAG_IO_IN_PROGRESS; 4004 } 4005 4006 ASSERT(!refcount_is_zero(&hdr->b_refcnt)); 4007 callback->awcb_done(zio, buf, callback->awcb_private); 4008 4009 kmem_free(callback, sizeof (arc_write_callback_t)); 4010} 4011 4012zio_t * 4013arc_write(zio_t *pio, spa_t *spa, uint64_t txg, 4014 blkptr_t *bp, arc_buf_t *buf, boolean_t l2arc, boolean_t l2arc_compress, 4015 const zio_prop_t *zp, arc_done_func_t *ready, arc_done_func_t *physdone, 4016 arc_done_func_t *done, void *private, zio_priority_t priority, 4017 int zio_flags, const zbookmark_phys_t *zb) 4018{ 4019 arc_buf_hdr_t *hdr = buf->b_hdr; 4020 arc_write_callback_t *callback; 4021 zio_t *zio; 4022 4023 ASSERT(ready != NULL); 4024 ASSERT(done != NULL); 4025 ASSERT(!HDR_IO_ERROR(hdr)); 4026 ASSERT((hdr->b_flags & ARC_FLAG_IO_IN_PROGRESS) == 0); 4027 ASSERT(hdr->b_acb == NULL); 4028 if (l2arc) 4029 hdr->b_flags |= ARC_FLAG_L2CACHE; 4030 if (l2arc_compress) 4031 hdr->b_flags |= ARC_FLAG_L2COMPRESS; 4032 callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP); 4033 callback->awcb_ready = ready; 4034 callback->awcb_physdone = physdone; 4035 callback->awcb_done = done; 4036 callback->awcb_private = private; 4037 callback->awcb_buf = buf; 4038 4039 zio = zio_write(pio, spa, txg, bp, buf->b_data, hdr->b_size, zp, 4040 arc_write_ready, arc_write_physdone, arc_write_done, callback, 4041 priority, zio_flags, zb); 4042 4043 return (zio); 4044} 4045 4046static int 4047arc_memory_throttle(uint64_t reserve, uint64_t txg) 4048{ 4049#ifdef _KERNEL 4050 uint64_t available_memory = ptob(freemem); 4051 static uint64_t page_load = 0; 4052 static uint64_t last_txg = 0; 4053 4054#if defined(__i386) || !defined(UMA_MD_SMALL_ALLOC) 4055 available_memory = 4056 MIN(available_memory, ptob(vmem_size(heap_arena, VMEM_FREE))); 4057#endif 4058 4059 if (freemem > (uint64_t)physmem * arc_lotsfree_percent / 100) 4060 return (0); 4061 4062 if (txg > last_txg) { 4063 last_txg = txg; 4064 page_load = 0; 4065 } 4066 /* 4067 * If we are in pageout, we know that memory is already tight, 4068 * the arc is already going to be evicting, so we just want to 4069 * continue to let page writes occur as quickly as possible. 4070 */ 4071 if (curproc == pageproc) { 4072 if (page_load > MAX(ptob(minfree), available_memory) / 4) 4073 return (SET_ERROR(ERESTART)); 4074 /* Note: reserve is inflated, so we deflate */ 4075 page_load += reserve / 8; 4076 return (0); 4077 } else if (page_load > 0 && arc_reclaim_needed()) { 4078 /* memory is low, delay before restarting */ 4079 ARCSTAT_INCR(arcstat_memory_throttle_count, 1); 4080 return (SET_ERROR(EAGAIN)); 4081 } 4082 page_load = 0; 4083#endif 4084 return (0); 4085} 4086 4087void 4088arc_tempreserve_clear(uint64_t reserve) 4089{ 4090 atomic_add_64(&arc_tempreserve, -reserve); 4091 ASSERT((int64_t)arc_tempreserve >= 0); 4092} 4093 4094int 4095arc_tempreserve_space(uint64_t reserve, uint64_t txg) 4096{ 4097 int error; 4098 uint64_t anon_size; 4099 4100 if (reserve > arc_c/4 && !arc_no_grow) { 4101 arc_c = MIN(arc_c_max, reserve * 4); 4102 DTRACE_PROBE1(arc__set_reserve, uint64_t, arc_c); 4103 } 4104 if (reserve > arc_c) 4105 return (SET_ERROR(ENOMEM)); 4106 4107 /* 4108 * Don't count loaned bufs as in flight dirty data to prevent long 4109 * network delays from blocking transactions that are ready to be 4110 * assigned to a txg. 4111 */ 4112 anon_size = MAX((int64_t)(arc_anon->arcs_size - arc_loaned_bytes), 0); 4113 4114 /* 4115 * Writes will, almost always, require additional memory allocations 4116 * in order to compress/encrypt/etc the data. We therefore need to 4117 * make sure that there is sufficient available memory for this. 4118 */ 4119 error = arc_memory_throttle(reserve, txg); 4120 if (error != 0) 4121 return (error); 4122 4123 /* 4124 * Throttle writes when the amount of dirty data in the cache 4125 * gets too large. We try to keep the cache less than half full 4126 * of dirty blocks so that our sync times don't grow too large. 4127 * Note: if two requests come in concurrently, we might let them 4128 * both succeed, when one of them should fail. Not a huge deal. 4129 */ 4130 4131 if (reserve + arc_tempreserve + anon_size > arc_c / 2 && 4132 anon_size > arc_c / 4) { 4133 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK " 4134 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n", 4135 arc_tempreserve>>10, 4136 arc_anon->arcs_lsize[ARC_BUFC_METADATA]>>10, 4137 arc_anon->arcs_lsize[ARC_BUFC_DATA]>>10, 4138 reserve>>10, arc_c>>10); 4139 return (SET_ERROR(ERESTART)); 4140 } 4141 atomic_add_64(&arc_tempreserve, reserve); 4142 return (0); 4143} 4144 4145static kmutex_t arc_lowmem_lock; 4146#ifdef _KERNEL 4147static eventhandler_tag arc_event_lowmem = NULL; 4148 4149static void 4150arc_lowmem(void *arg __unused, int howto __unused) 4151{ 4152 4153 /* Serialize access via arc_lowmem_lock. */ 4154 mutex_enter(&arc_lowmem_lock); 4155 mutex_enter(&arc_reclaim_thr_lock); 4156 needfree = 1; 4157 DTRACE_PROBE(arc__needfree); 4158 cv_signal(&arc_reclaim_thr_cv); 4159 4160 /* 4161 * It is unsafe to block here in arbitrary threads, because we can come 4162 * here from ARC itself and may hold ARC locks and thus risk a deadlock 4163 * with ARC reclaim thread. 4164 */ 4165 if (curproc == pageproc) { 4166 while (needfree) 4167 msleep(&needfree, &arc_reclaim_thr_lock, 0, "zfs:lowmem", 0); 4168 } 4169 mutex_exit(&arc_reclaim_thr_lock); 4170 mutex_exit(&arc_lowmem_lock); 4171} 4172#endif 4173 4174void 4175arc_init(void) 4176{ 4177 int i, prefetch_tunable_set = 0; 4178 4179 mutex_init(&arc_reclaim_thr_lock, NULL, MUTEX_DEFAULT, NULL); 4180 cv_init(&arc_reclaim_thr_cv, NULL, CV_DEFAULT, NULL); 4181 mutex_init(&arc_lowmem_lock, NULL, MUTEX_DEFAULT, NULL); 4182 4183 /* Convert seconds to clock ticks */ 4184 arc_min_prefetch_lifespan = 1 * hz; 4185 4186 /* Start out with 1/8 of all memory */ 4187 arc_c = kmem_size() / 8; 4188 4189#ifdef illumos 4190#ifdef _KERNEL 4191 /* 4192 * On architectures where the physical memory can be larger 4193 * than the addressable space (intel in 32-bit mode), we may 4194 * need to limit the cache to 1/8 of VM size. 4195 */ 4196 arc_c = MIN(arc_c, vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 8); 4197#endif 4198#endif /* illumos */ 4199 /* set min cache to 1/32 of all memory, or 16MB, whichever is more */ 4200 arc_c_min = MAX(arc_c / 4, 64<<18); 4201 /* set max to 1/2 of all memory, or all but 1GB, whichever is more */ 4202 if (arc_c * 8 >= 1<<30) 4203 arc_c_max = (arc_c * 8) - (1<<30); 4204 else 4205 arc_c_max = arc_c_min; 4206 arc_c_max = MAX(arc_c * 5, arc_c_max); 4207 4208#ifdef _KERNEL 4209 /* 4210 * Allow the tunables to override our calculations if they are 4211 * reasonable (ie. over 16MB) 4212 */ 4213 if (zfs_arc_max > 64<<18 && zfs_arc_max < kmem_size()) 4214 arc_c_max = zfs_arc_max; 4215 if (zfs_arc_min > 64<<18 && zfs_arc_min <= arc_c_max) 4216 arc_c_min = zfs_arc_min; 4217#endif 4218 4219 arc_c = arc_c_max; 4220 arc_p = (arc_c >> 1); 4221 4222 /* limit meta-data to 1/4 of the arc capacity */ 4223 arc_meta_limit = arc_c_max / 4; 4224 4225 /* Allow the tunable to override if it is reasonable */ 4226 if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max) 4227 arc_meta_limit = zfs_arc_meta_limit; 4228 4229 if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0) 4230 arc_c_min = arc_meta_limit / 2; 4231 4232 if (zfs_arc_meta_min > 0) { 4233 arc_meta_min = zfs_arc_meta_min; 4234 } else { 4235 arc_meta_min = arc_c_min / 2; 4236 } 4237 4238 if (zfs_arc_grow_retry > 0) 4239 arc_grow_retry = zfs_arc_grow_retry; 4240 4241 if (zfs_arc_shrink_shift > 0) 4242 arc_shrink_shift = zfs_arc_shrink_shift; 4243 4244 if (zfs_arc_p_min_shift > 0) 4245 arc_p_min_shift = zfs_arc_p_min_shift; 4246 4247 /* if kmem_flags are set, lets try to use less memory */ 4248 if (kmem_debugging()) 4249 arc_c = arc_c / 2; 4250 if (arc_c < arc_c_min) 4251 arc_c = arc_c_min; 4252 4253 zfs_arc_min = arc_c_min; 4254 zfs_arc_max = arc_c_max; 4255 4256 arc_anon = &ARC_anon; 4257 arc_mru = &ARC_mru; 4258 arc_mru_ghost = &ARC_mru_ghost; 4259 arc_mfu = &ARC_mfu; 4260 arc_mfu_ghost = &ARC_mfu_ghost; 4261 arc_l2c_only = &ARC_l2c_only; 4262 arc_size = 0; 4263 4264 for (i = 0; i < ARC_BUFC_NUMLISTS; i++) { 4265 mutex_init(&arc_anon->arcs_locks[i].arcs_lock, 4266 NULL, MUTEX_DEFAULT, NULL); 4267 mutex_init(&arc_mru->arcs_locks[i].arcs_lock, 4268 NULL, MUTEX_DEFAULT, NULL); 4269 mutex_init(&arc_mru_ghost->arcs_locks[i].arcs_lock, 4270 NULL, MUTEX_DEFAULT, NULL); 4271 mutex_init(&arc_mfu->arcs_locks[i].arcs_lock, 4272 NULL, MUTEX_DEFAULT, NULL); 4273 mutex_init(&arc_mfu_ghost->arcs_locks[i].arcs_lock, 4274 NULL, MUTEX_DEFAULT, NULL); 4275 mutex_init(&arc_l2c_only->arcs_locks[i].arcs_lock, 4276 NULL, MUTEX_DEFAULT, NULL); 4277 4278 list_create(&arc_mru->arcs_lists[i], 4279 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4280 list_create(&arc_mru_ghost->arcs_lists[i], 4281 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4282 list_create(&arc_mfu->arcs_lists[i], 4283 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4284 list_create(&arc_mfu_ghost->arcs_lists[i], 4285 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4286 list_create(&arc_mfu_ghost->arcs_lists[i], 4287 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4288 list_create(&arc_l2c_only->arcs_lists[i], 4289 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4290 } 4291 4292 buf_init(); 4293 4294 arc_thread_exit = 0; 4295 arc_eviction_list = NULL; 4296 mutex_init(&arc_eviction_mtx, NULL, MUTEX_DEFAULT, NULL); 4297 bzero(&arc_eviction_hdr, sizeof (arc_buf_hdr_t)); 4298 4299 arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED, 4300 sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL); 4301 4302 if (arc_ksp != NULL) { 4303 arc_ksp->ks_data = &arc_stats; 4304 kstat_install(arc_ksp); 4305 } 4306 4307 (void) thread_create(NULL, 0, arc_reclaim_thread, NULL, 0, &p0, 4308 TS_RUN, minclsyspri); 4309 4310#ifdef _KERNEL 4311 arc_event_lowmem = EVENTHANDLER_REGISTER(vm_lowmem, arc_lowmem, NULL, 4312 EVENTHANDLER_PRI_FIRST); 4313#endif 4314 4315 arc_dead = FALSE; 4316 arc_warm = B_FALSE; 4317 4318 /* 4319 * Calculate maximum amount of dirty data per pool. 4320 * 4321 * If it has been set by /etc/system, take that. 4322 * Otherwise, use a percentage of physical memory defined by 4323 * zfs_dirty_data_max_percent (default 10%) with a cap at 4324 * zfs_dirty_data_max_max (default 4GB). 4325 */ 4326 if (zfs_dirty_data_max == 0) { 4327 zfs_dirty_data_max = ptob(physmem) * 4328 zfs_dirty_data_max_percent / 100; 4329 zfs_dirty_data_max = MIN(zfs_dirty_data_max, 4330 zfs_dirty_data_max_max); 4331 } 4332 4333#ifdef _KERNEL 4334 if (TUNABLE_INT_FETCH("vfs.zfs.prefetch_disable", &zfs_prefetch_disable)) 4335 prefetch_tunable_set = 1; 4336 4337#ifdef __i386__ 4338 if (prefetch_tunable_set == 0) { 4339 printf("ZFS NOTICE: Prefetch is disabled by default on i386 " 4340 "-- to enable,\n"); 4341 printf(" add \"vfs.zfs.prefetch_disable=0\" " 4342 "to /boot/loader.conf.\n"); 4343 zfs_prefetch_disable = 1; 4344 } 4345#else 4346 if ((((uint64_t)physmem * PAGESIZE) < (1ULL << 32)) && 4347 prefetch_tunable_set == 0) { 4348 printf("ZFS NOTICE: Prefetch is disabled by default if less " 4349 "than 4GB of RAM is present;\n" 4350 " to enable, add \"vfs.zfs.prefetch_disable=0\" " 4351 "to /boot/loader.conf.\n"); 4352 zfs_prefetch_disable = 1; 4353 } 4354#endif 4355 /* Warn about ZFS memory and address space requirements. */ 4356 if (((uint64_t)physmem * PAGESIZE) < (256 + 128 + 64) * (1 << 20)) { 4357 printf("ZFS WARNING: Recommended minimum RAM size is 512MB; " 4358 "expect unstable behavior.\n"); 4359 } 4360 if (kmem_size() < 512 * (1 << 20)) { 4361 printf("ZFS WARNING: Recommended minimum kmem_size is 512MB; " 4362 "expect unstable behavior.\n"); 4363 printf(" Consider tuning vm.kmem_size and " 4364 "vm.kmem_size_max\n"); 4365 printf(" in /boot/loader.conf.\n"); 4366 } 4367#endif 4368} 4369 4370void 4371arc_fini(void) 4372{ 4373 int i; 4374 4375 mutex_enter(&arc_reclaim_thr_lock); 4376 arc_thread_exit = 1; 4377 cv_signal(&arc_reclaim_thr_cv); 4378 while (arc_thread_exit != 0) 4379 cv_wait(&arc_reclaim_thr_cv, &arc_reclaim_thr_lock); 4380 mutex_exit(&arc_reclaim_thr_lock); 4381 4382 arc_flush(NULL); 4383 4384 arc_dead = TRUE; 4385 4386 if (arc_ksp != NULL) { 4387 kstat_delete(arc_ksp); 4388 arc_ksp = NULL; 4389 } 4390 4391 mutex_destroy(&arc_eviction_mtx); 4392 mutex_destroy(&arc_reclaim_thr_lock); 4393 cv_destroy(&arc_reclaim_thr_cv); 4394 4395 for (i = 0; i < ARC_BUFC_NUMLISTS; i++) { 4396 list_destroy(&arc_mru->arcs_lists[i]); 4397 list_destroy(&arc_mru_ghost->arcs_lists[i]); 4398 list_destroy(&arc_mfu->arcs_lists[i]); 4399 list_destroy(&arc_mfu_ghost->arcs_lists[i]); 4400 list_destroy(&arc_l2c_only->arcs_lists[i]); 4401 4402 mutex_destroy(&arc_anon->arcs_locks[i].arcs_lock); 4403 mutex_destroy(&arc_mru->arcs_locks[i].arcs_lock); 4404 mutex_destroy(&arc_mru_ghost->arcs_locks[i].arcs_lock); 4405 mutex_destroy(&arc_mfu->arcs_locks[i].arcs_lock); 4406 mutex_destroy(&arc_mfu_ghost->arcs_locks[i].arcs_lock); 4407 mutex_destroy(&arc_l2c_only->arcs_locks[i].arcs_lock); 4408 } 4409 4410 buf_fini(); 4411 4412 ASSERT(arc_loaned_bytes == 0); 4413 4414 mutex_destroy(&arc_lowmem_lock); 4415#ifdef _KERNEL 4416 if (arc_event_lowmem != NULL) 4417 EVENTHANDLER_DEREGISTER(vm_lowmem, arc_event_lowmem); 4418#endif 4419} 4420 4421/* 4422 * Level 2 ARC 4423 * 4424 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk. 4425 * It uses dedicated storage devices to hold cached data, which are populated 4426 * using large infrequent writes. The main role of this cache is to boost 4427 * the performance of random read workloads. The intended L2ARC devices 4428 * include short-stroked disks, solid state disks, and other media with 4429 * substantially faster read latency than disk. 4430 * 4431 * +-----------------------+ 4432 * | ARC | 4433 * +-----------------------+ 4434 * | ^ ^ 4435 * | | | 4436 * l2arc_feed_thread() arc_read() 4437 * | | | 4438 * | l2arc read | 4439 * V | | 4440 * +---------------+ | 4441 * | L2ARC | | 4442 * +---------------+ | 4443 * | ^ | 4444 * l2arc_write() | | 4445 * | | | 4446 * V | | 4447 * +-------+ +-------+ 4448 * | vdev | | vdev | 4449 * | cache | | cache | 4450 * +-------+ +-------+ 4451 * +=========+ .-----. 4452 * : L2ARC : |-_____-| 4453 * : devices : | Disks | 4454 * +=========+ `-_____-' 4455 * 4456 * Read requests are satisfied from the following sources, in order: 4457 * 4458 * 1) ARC 4459 * 2) vdev cache of L2ARC devices 4460 * 3) L2ARC devices 4461 * 4) vdev cache of disks 4462 * 5) disks 4463 * 4464 * Some L2ARC device types exhibit extremely slow write performance. 4465 * To accommodate for this there are some significant differences between 4466 * the L2ARC and traditional cache design: 4467 * 4468 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from 4469 * the ARC behave as usual, freeing buffers and placing headers on ghost 4470 * lists. The ARC does not send buffers to the L2ARC during eviction as 4471 * this would add inflated write latencies for all ARC memory pressure. 4472 * 4473 * 2. The L2ARC attempts to cache data from the ARC before it is evicted. 4474 * It does this by periodically scanning buffers from the eviction-end of 4475 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are 4476 * not already there. It scans until a headroom of buffers is satisfied, 4477 * which itself is a buffer for ARC eviction. If a compressible buffer is 4478 * found during scanning and selected for writing to an L2ARC device, we 4479 * temporarily boost scanning headroom during the next scan cycle to make 4480 * sure we adapt to compression effects (which might significantly reduce 4481 * the data volume we write to L2ARC). The thread that does this is 4482 * l2arc_feed_thread(), illustrated below; example sizes are included to 4483 * provide a better sense of ratio than this diagram: 4484 * 4485 * head --> tail 4486 * +---------------------+----------+ 4487 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC 4488 * +---------------------+----------+ | o L2ARC eligible 4489 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer 4490 * +---------------------+----------+ | 4491 * 15.9 Gbytes ^ 32 Mbytes | 4492 * headroom | 4493 * l2arc_feed_thread() 4494 * | 4495 * l2arc write hand <--[oooo]--' 4496 * | 8 Mbyte 4497 * | write max 4498 * V 4499 * +==============================+ 4500 * L2ARC dev |####|#|###|###| |####| ... | 4501 * +==============================+ 4502 * 32 Gbytes 4503 * 4504 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of 4505 * evicted, then the L2ARC has cached a buffer much sooner than it probably 4506 * needed to, potentially wasting L2ARC device bandwidth and storage. It is 4507 * safe to say that this is an uncommon case, since buffers at the end of 4508 * the ARC lists have moved there due to inactivity. 4509 * 4510 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom, 4511 * then the L2ARC simply misses copying some buffers. This serves as a 4512 * pressure valve to prevent heavy read workloads from both stalling the ARC 4513 * with waits and clogging the L2ARC with writes. This also helps prevent 4514 * the potential for the L2ARC to churn if it attempts to cache content too 4515 * quickly, such as during backups of the entire pool. 4516 * 4517 * 5. After system boot and before the ARC has filled main memory, there are 4518 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru 4519 * lists can remain mostly static. Instead of searching from tail of these 4520 * lists as pictured, the l2arc_feed_thread() will search from the list heads 4521 * for eligible buffers, greatly increasing its chance of finding them. 4522 * 4523 * The L2ARC device write speed is also boosted during this time so that 4524 * the L2ARC warms up faster. Since there have been no ARC evictions yet, 4525 * there are no L2ARC reads, and no fear of degrading read performance 4526 * through increased writes. 4527 * 4528 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that 4529 * the vdev queue can aggregate them into larger and fewer writes. Each 4530 * device is written to in a rotor fashion, sweeping writes through 4531 * available space then repeating. 4532 * 4533 * 7. The L2ARC does not store dirty content. It never needs to flush 4534 * write buffers back to disk based storage. 4535 * 4536 * 8. If an ARC buffer is written (and dirtied) which also exists in the 4537 * L2ARC, the now stale L2ARC buffer is immediately dropped. 4538 * 4539 * The performance of the L2ARC can be tweaked by a number of tunables, which 4540 * may be necessary for different workloads: 4541 * 4542 * l2arc_write_max max write bytes per interval 4543 * l2arc_write_boost extra write bytes during device warmup 4544 * l2arc_noprefetch skip caching prefetched buffers 4545 * l2arc_headroom number of max device writes to precache 4546 * l2arc_headroom_boost when we find compressed buffers during ARC 4547 * scanning, we multiply headroom by this 4548 * percentage factor for the next scan cycle, 4549 * since more compressed buffers are likely to 4550 * be present 4551 * l2arc_feed_secs seconds between L2ARC writing 4552 * 4553 * Tunables may be removed or added as future performance improvements are 4554 * integrated, and also may become zpool properties. 4555 * 4556 * There are three key functions that control how the L2ARC warms up: 4557 * 4558 * l2arc_write_eligible() check if a buffer is eligible to cache 4559 * l2arc_write_size() calculate how much to write 4560 * l2arc_write_interval() calculate sleep delay between writes 4561 * 4562 * These three functions determine what to write, how much, and how quickly 4563 * to send writes. 4564 */ 4565 4566static boolean_t 4567l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *hdr) 4568{ 4569 /* 4570 * A buffer is *not* eligible for the L2ARC if it: 4571 * 1. belongs to a different spa. 4572 * 2. is already cached on the L2ARC. 4573 * 3. has an I/O in progress (it may be an incomplete read). 4574 * 4. is flagged not eligible (zfs property). 4575 */ 4576 if (hdr->b_spa != spa_guid) { 4577 ARCSTAT_BUMP(arcstat_l2_write_spa_mismatch); 4578 return (B_FALSE); 4579 } 4580 if (hdr->b_l2hdr != NULL) { 4581 ARCSTAT_BUMP(arcstat_l2_write_in_l2); 4582 return (B_FALSE); 4583 } 4584 if (HDR_IO_IN_PROGRESS(hdr)) { 4585 ARCSTAT_BUMP(arcstat_l2_write_hdr_io_in_progress); 4586 return (B_FALSE); 4587 } 4588 if (!HDR_L2CACHE(hdr)) { 4589 ARCSTAT_BUMP(arcstat_l2_write_not_cacheable); 4590 return (B_FALSE); 4591 } 4592 4593 return (B_TRUE); 4594} 4595 4596static uint64_t 4597l2arc_write_size(void) 4598{ 4599 uint64_t size; 4600 4601 /* 4602 * Make sure our globals have meaningful values in case the user 4603 * altered them. 4604 */ 4605 size = l2arc_write_max; 4606 if (size == 0) { 4607 cmn_err(CE_NOTE, "Bad value for l2arc_write_max, value must " 4608 "be greater than zero, resetting it to the default (%d)", 4609 L2ARC_WRITE_SIZE); 4610 size = l2arc_write_max = L2ARC_WRITE_SIZE; 4611 } 4612 4613 if (arc_warm == B_FALSE) 4614 size += l2arc_write_boost; 4615 4616 return (size); 4617 4618} 4619 4620static clock_t 4621l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote) 4622{ 4623 clock_t interval, next, now; 4624 4625 /* 4626 * If the ARC lists are busy, increase our write rate; if the 4627 * lists are stale, idle back. This is achieved by checking 4628 * how much we previously wrote - if it was more than half of 4629 * what we wanted, schedule the next write much sooner. 4630 */ 4631 if (l2arc_feed_again && wrote > (wanted / 2)) 4632 interval = (hz * l2arc_feed_min_ms) / 1000; 4633 else 4634 interval = hz * l2arc_feed_secs; 4635 4636 now = ddi_get_lbolt(); 4637 next = MAX(now, MIN(now + interval, began + interval)); 4638 4639 return (next); 4640} 4641 4642static void 4643l2arc_hdr_stat_add(void) 4644{ 4645 ARCSTAT_INCR(arcstat_l2_hdr_size, HDR_SIZE + L2HDR_SIZE); 4646 ARCSTAT_INCR(arcstat_hdr_size, -HDR_SIZE); 4647} 4648 4649static void 4650l2arc_hdr_stat_remove(void) 4651{ 4652 ARCSTAT_INCR(arcstat_l2_hdr_size, -(HDR_SIZE + L2HDR_SIZE)); 4653 ARCSTAT_INCR(arcstat_hdr_size, HDR_SIZE); 4654} 4655 4656/* 4657 * Cycle through L2ARC devices. This is how L2ARC load balances. 4658 * If a device is returned, this also returns holding the spa config lock. 4659 */ 4660static l2arc_dev_t * 4661l2arc_dev_get_next(void) 4662{ 4663 l2arc_dev_t *first, *next = NULL; 4664 4665 /* 4666 * Lock out the removal of spas (spa_namespace_lock), then removal 4667 * of cache devices (l2arc_dev_mtx). Once a device has been selected, 4668 * both locks will be dropped and a spa config lock held instead. 4669 */ 4670 mutex_enter(&spa_namespace_lock); 4671 mutex_enter(&l2arc_dev_mtx); 4672 4673 /* if there are no vdevs, there is nothing to do */ 4674 if (l2arc_ndev == 0) 4675 goto out; 4676 4677 first = NULL; 4678 next = l2arc_dev_last; 4679 do { 4680 /* loop around the list looking for a non-faulted vdev */ 4681 if (next == NULL) { 4682 next = list_head(l2arc_dev_list); 4683 } else { 4684 next = list_next(l2arc_dev_list, next); 4685 if (next == NULL) 4686 next = list_head(l2arc_dev_list); 4687 } 4688 4689 /* if we have come back to the start, bail out */ 4690 if (first == NULL) 4691 first = next; 4692 else if (next == first) 4693 break; 4694 4695 } while (vdev_is_dead(next->l2ad_vdev)); 4696 4697 /* if we were unable to find any usable vdevs, return NULL */ 4698 if (vdev_is_dead(next->l2ad_vdev)) 4699 next = NULL; 4700 4701 l2arc_dev_last = next; 4702 4703out: 4704 mutex_exit(&l2arc_dev_mtx); 4705 4706 /* 4707 * Grab the config lock to prevent the 'next' device from being 4708 * removed while we are writing to it. 4709 */ 4710 if (next != NULL) 4711 spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER); 4712 mutex_exit(&spa_namespace_lock); 4713 4714 return (next); 4715} 4716 4717/* 4718 * Free buffers that were tagged for destruction. 4719 */ 4720static void 4721l2arc_do_free_on_write() 4722{ 4723 list_t *buflist; 4724 l2arc_data_free_t *df, *df_prev; 4725 4726 mutex_enter(&l2arc_free_on_write_mtx); 4727 buflist = l2arc_free_on_write; 4728 4729 for (df = list_tail(buflist); df; df = df_prev) { 4730 df_prev = list_prev(buflist, df); 4731 ASSERT(df->l2df_data != NULL); 4732 ASSERT(df->l2df_func != NULL); 4733 df->l2df_func(df->l2df_data, df->l2df_size); 4734 list_remove(buflist, df); 4735 kmem_free(df, sizeof (l2arc_data_free_t)); 4736 } 4737 4738 mutex_exit(&l2arc_free_on_write_mtx); 4739} 4740 4741/* 4742 * A write to a cache device has completed. Update all headers to allow 4743 * reads from these buffers to begin. 4744 */ 4745static void 4746l2arc_write_done(zio_t *zio) 4747{ 4748 l2arc_write_callback_t *cb; 4749 l2arc_dev_t *dev; 4750 list_t *buflist; 4751 arc_buf_hdr_t *head, *hdr, *hdr_prev; 4752 l2arc_buf_hdr_t *abl2; 4753 kmutex_t *hash_lock; 4754 int64_t bytes_dropped = 0; 4755 4756 cb = zio->io_private; 4757 ASSERT(cb != NULL); 4758 dev = cb->l2wcb_dev; 4759 ASSERT(dev != NULL); 4760 head = cb->l2wcb_head; 4761 ASSERT(head != NULL); 4762 buflist = dev->l2ad_buflist; 4763 ASSERT(buflist != NULL); 4764 DTRACE_PROBE2(l2arc__iodone, zio_t *, zio, 4765 l2arc_write_callback_t *, cb); 4766 4767 if (zio->io_error != 0) 4768 ARCSTAT_BUMP(arcstat_l2_writes_error); 4769 4770 mutex_enter(&l2arc_buflist_mtx); 4771 4772 /* 4773 * All writes completed, or an error was hit. 4774 */ 4775 for (hdr = list_prev(buflist, head); hdr; hdr = hdr_prev) { 4776 hdr_prev = list_prev(buflist, hdr); 4777 abl2 = hdr->b_l2hdr; 4778 4779 /* 4780 * Release the temporary compressed buffer as soon as possible. 4781 */ 4782 if (abl2->b_compress != ZIO_COMPRESS_OFF) 4783 l2arc_release_cdata_buf(hdr); 4784 4785 hash_lock = HDR_LOCK(hdr); 4786 if (!mutex_tryenter(hash_lock)) { 4787 /* 4788 * This buffer misses out. It may be in a stage 4789 * of eviction. Its ARC_L2_WRITING flag will be 4790 * left set, denying reads to this buffer. 4791 */ 4792 ARCSTAT_BUMP(arcstat_l2_writes_hdr_miss); 4793 continue; 4794 } 4795 4796 if (zio->io_error != 0) { 4797 /* 4798 * Error - drop L2ARC entry. 4799 */ 4800 list_remove(buflist, hdr); 4801 ARCSTAT_INCR(arcstat_l2_asize, -abl2->b_asize); 4802 bytes_dropped += abl2->b_asize; 4803 hdr->b_l2hdr = NULL; 4804 trim_map_free(abl2->b_dev->l2ad_vdev, abl2->b_daddr, 4805 hdr->b_size, 0); 4806 kmem_free(abl2, sizeof (l2arc_buf_hdr_t)); 4807 ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size); 4808 } 4809 4810 /* 4811 * Allow ARC to begin reads to this L2ARC entry. 4812 */ 4813 hdr->b_flags &= ~ARC_FLAG_L2_WRITING; 4814 4815 mutex_exit(hash_lock); 4816 } 4817 4818 atomic_inc_64(&l2arc_writes_done); 4819 list_remove(buflist, head); 4820 kmem_cache_free(hdr_cache, head); 4821 mutex_exit(&l2arc_buflist_mtx); 4822 4823 vdev_space_update(dev->l2ad_vdev, -bytes_dropped, 0, 0); 4824 4825 l2arc_do_free_on_write(); 4826 4827 kmem_free(cb, sizeof (l2arc_write_callback_t)); 4828} 4829 4830/* 4831 * A read to a cache device completed. Validate buffer contents before 4832 * handing over to the regular ARC routines. 4833 */ 4834static void 4835l2arc_read_done(zio_t *zio) 4836{ 4837 l2arc_read_callback_t *cb; 4838 arc_buf_hdr_t *hdr; 4839 arc_buf_t *buf; 4840 kmutex_t *hash_lock; 4841 int equal; 4842 4843 ASSERT(zio->io_vd != NULL); 4844 ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE); 4845 4846 spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd); 4847 4848 cb = zio->io_private; 4849 ASSERT(cb != NULL); 4850 buf = cb->l2rcb_buf; 4851 ASSERT(buf != NULL); 4852 4853 hash_lock = HDR_LOCK(buf->b_hdr); 4854 mutex_enter(hash_lock); 4855 hdr = buf->b_hdr; 4856 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 4857 4858 /* 4859 * If the buffer was compressed, decompress it first. 4860 */ 4861 if (cb->l2rcb_compress != ZIO_COMPRESS_OFF) 4862 l2arc_decompress_zio(zio, hdr, cb->l2rcb_compress); 4863 ASSERT(zio->io_data != NULL); 4864 4865 /* 4866 * Check this survived the L2ARC journey. 4867 */ 4868 equal = arc_cksum_equal(buf); 4869 if (equal && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) { 4870 mutex_exit(hash_lock); 4871 zio->io_private = buf; 4872 zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */ 4873 zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */ 4874 arc_read_done(zio); 4875 } else { 4876 mutex_exit(hash_lock); 4877 /* 4878 * Buffer didn't survive caching. Increment stats and 4879 * reissue to the original storage device. 4880 */ 4881 if (zio->io_error != 0) { 4882 ARCSTAT_BUMP(arcstat_l2_io_error); 4883 } else { 4884 zio->io_error = SET_ERROR(EIO); 4885 } 4886 if (!equal) 4887 ARCSTAT_BUMP(arcstat_l2_cksum_bad); 4888 4889 /* 4890 * If there's no waiter, issue an async i/o to the primary 4891 * storage now. If there *is* a waiter, the caller must 4892 * issue the i/o in a context where it's OK to block. 4893 */ 4894 if (zio->io_waiter == NULL) { 4895 zio_t *pio = zio_unique_parent(zio); 4896 4897 ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL); 4898 4899 zio_nowait(zio_read(pio, cb->l2rcb_spa, &cb->l2rcb_bp, 4900 buf->b_data, zio->io_size, arc_read_done, buf, 4901 zio->io_priority, cb->l2rcb_flags, &cb->l2rcb_zb)); 4902 } 4903 } 4904 4905 kmem_free(cb, sizeof (l2arc_read_callback_t)); 4906} 4907 4908/* 4909 * This is the list priority from which the L2ARC will search for pages to 4910 * cache. This is used within loops (0..3) to cycle through lists in the 4911 * desired order. This order can have a significant effect on cache 4912 * performance. 4913 * 4914 * Currently the metadata lists are hit first, MFU then MRU, followed by 4915 * the data lists. This function returns a locked list, and also returns 4916 * the lock pointer. 4917 */ 4918static list_t * 4919l2arc_list_locked(int list_num, kmutex_t **lock) 4920{ 4921 list_t *list = NULL; 4922 int idx; 4923 4924 ASSERT(list_num >= 0 && list_num < 2 * ARC_BUFC_NUMLISTS); 4925 4926 if (list_num < ARC_BUFC_NUMMETADATALISTS) { 4927 idx = list_num; 4928 list = &arc_mfu->arcs_lists[idx]; 4929 *lock = ARCS_LOCK(arc_mfu, idx); 4930 } else if (list_num < ARC_BUFC_NUMMETADATALISTS * 2) { 4931 idx = list_num - ARC_BUFC_NUMMETADATALISTS; 4932 list = &arc_mru->arcs_lists[idx]; 4933 *lock = ARCS_LOCK(arc_mru, idx); 4934 } else if (list_num < (ARC_BUFC_NUMMETADATALISTS * 2 + 4935 ARC_BUFC_NUMDATALISTS)) { 4936 idx = list_num - ARC_BUFC_NUMMETADATALISTS; 4937 list = &arc_mfu->arcs_lists[idx]; 4938 *lock = ARCS_LOCK(arc_mfu, idx); 4939 } else { 4940 idx = list_num - ARC_BUFC_NUMLISTS; 4941 list = &arc_mru->arcs_lists[idx]; 4942 *lock = ARCS_LOCK(arc_mru, idx); 4943 } 4944 4945 ASSERT(!(MUTEX_HELD(*lock))); 4946 mutex_enter(*lock); 4947 return (list); 4948} 4949 4950/* 4951 * Evict buffers from the device write hand to the distance specified in 4952 * bytes. This distance may span populated buffers, it may span nothing. 4953 * This is clearing a region on the L2ARC device ready for writing. 4954 * If the 'all' boolean is set, every buffer is evicted. 4955 */ 4956static void 4957l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all) 4958{ 4959 list_t *buflist; 4960 l2arc_buf_hdr_t *abl2; 4961 arc_buf_hdr_t *hdr, *hdr_prev; 4962 kmutex_t *hash_lock; 4963 uint64_t taddr; 4964 int64_t bytes_evicted = 0; 4965 4966 buflist = dev->l2ad_buflist; 4967 4968 if (buflist == NULL) 4969 return; 4970 4971 if (!all && dev->l2ad_first) { 4972 /* 4973 * This is the first sweep through the device. There is 4974 * nothing to evict. 4975 */ 4976 return; 4977 } 4978 4979 if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) { 4980 /* 4981 * When nearing the end of the device, evict to the end 4982 * before the device write hand jumps to the start. 4983 */ 4984 taddr = dev->l2ad_end; 4985 } else { 4986 taddr = dev->l2ad_hand + distance; 4987 } 4988 DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist, 4989 uint64_t, taddr, boolean_t, all); 4990 4991top: 4992 mutex_enter(&l2arc_buflist_mtx); 4993 for (hdr = list_tail(buflist); hdr; hdr = hdr_prev) { 4994 hdr_prev = list_prev(buflist, hdr); 4995 4996 hash_lock = HDR_LOCK(hdr); 4997 if (!mutex_tryenter(hash_lock)) { 4998 /* 4999 * Missed the hash lock. Retry. 5000 */ 5001 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry); 5002 mutex_exit(&l2arc_buflist_mtx); 5003 mutex_enter(hash_lock); 5004 mutex_exit(hash_lock); 5005 goto top; 5006 } 5007 5008 if (HDR_L2_WRITE_HEAD(hdr)) { 5009 /* 5010 * We hit a write head node. Leave it for 5011 * l2arc_write_done(). 5012 */ 5013 list_remove(buflist, hdr); 5014 mutex_exit(hash_lock); 5015 continue; 5016 } 5017 5018 if (!all && hdr->b_l2hdr != NULL && 5019 (hdr->b_l2hdr->b_daddr > taddr || 5020 hdr->b_l2hdr->b_daddr < dev->l2ad_hand)) { 5021 /* 5022 * We've evicted to the target address, 5023 * or the end of the device. 5024 */ 5025 mutex_exit(hash_lock); 5026 break; 5027 } 5028 5029 if (HDR_FREE_IN_PROGRESS(hdr)) { 5030 /* 5031 * Already on the path to destruction. 5032 */ 5033 mutex_exit(hash_lock); 5034 continue; 5035 } 5036 5037 if (hdr->b_state == arc_l2c_only) { 5038 ASSERT(!HDR_L2_READING(hdr)); 5039 /* 5040 * This doesn't exist in the ARC. Destroy. 5041 * arc_hdr_destroy() will call list_remove() 5042 * and decrement arcstat_l2_size. 5043 */ 5044 arc_change_state(arc_anon, hdr, hash_lock); 5045 arc_hdr_destroy(hdr); 5046 } else { 5047 /* 5048 * Invalidate issued or about to be issued 5049 * reads, since we may be about to write 5050 * over this location. 5051 */ 5052 if (HDR_L2_READING(hdr)) { 5053 ARCSTAT_BUMP(arcstat_l2_evict_reading); 5054 hdr->b_flags |= ARC_FLAG_L2_EVICTED; 5055 } 5056 5057 /* 5058 * Tell ARC this no longer exists in L2ARC. 5059 */ 5060 if (hdr->b_l2hdr != NULL) { 5061 abl2 = hdr->b_l2hdr; 5062 ARCSTAT_INCR(arcstat_l2_asize, -abl2->b_asize); 5063 bytes_evicted += abl2->b_asize; 5064 hdr->b_l2hdr = NULL; 5065 /* 5066 * We are destroying l2hdr, so ensure that 5067 * its compressed buffer, if any, is not leaked. 5068 */ 5069 ASSERT(abl2->b_tmp_cdata == NULL); 5070 kmem_free(abl2, sizeof (l2arc_buf_hdr_t)); 5071 ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size); 5072 } 5073 list_remove(buflist, hdr); 5074 5075 /* 5076 * This may have been leftover after a 5077 * failed write. 5078 */ 5079 hdr->b_flags &= ~ARC_FLAG_L2_WRITING; 5080 } 5081 mutex_exit(hash_lock); 5082 } 5083 mutex_exit(&l2arc_buflist_mtx); 5084 5085 vdev_space_update(dev->l2ad_vdev, -bytes_evicted, 0, 0); 5086 dev->l2ad_evict = taddr; 5087} 5088 5089/* 5090 * Find and write ARC buffers to the L2ARC device. 5091 * 5092 * An ARC_FLAG_L2_WRITING flag is set so that the L2ARC buffers are not valid 5093 * for reading until they have completed writing. 5094 * The headroom_boost is an in-out parameter used to maintain headroom boost 5095 * state between calls to this function. 5096 * 5097 * Returns the number of bytes actually written (which may be smaller than 5098 * the delta by which the device hand has changed due to alignment). 5099 */ 5100static uint64_t 5101l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz, 5102 boolean_t *headroom_boost) 5103{ 5104 arc_buf_hdr_t *hdr, *hdr_prev, *head; 5105 list_t *list; 5106 uint64_t write_asize, write_psize, write_sz, headroom, 5107 buf_compress_minsz; 5108 void *buf_data; 5109 kmutex_t *list_lock; 5110 boolean_t full; 5111 l2arc_write_callback_t *cb; 5112 zio_t *pio, *wzio; 5113 uint64_t guid = spa_load_guid(spa); 5114 const boolean_t do_headroom_boost = *headroom_boost; 5115 int try; 5116 5117 ASSERT(dev->l2ad_vdev != NULL); 5118 5119 /* Lower the flag now, we might want to raise it again later. */ 5120 *headroom_boost = B_FALSE; 5121 5122 pio = NULL; 5123 write_sz = write_asize = write_psize = 0; 5124 full = B_FALSE; 5125 head = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE); 5126 head->b_flags |= ARC_FLAG_L2_WRITE_HEAD; 5127 5128 ARCSTAT_BUMP(arcstat_l2_write_buffer_iter); 5129 /* 5130 * We will want to try to compress buffers that are at least 2x the 5131 * device sector size. 5132 */ 5133 buf_compress_minsz = 2 << dev->l2ad_vdev->vdev_ashift; 5134 5135 /* 5136 * Copy buffers for L2ARC writing. 5137 */ 5138 mutex_enter(&l2arc_buflist_mtx); 5139 for (try = 0; try < 2 * ARC_BUFC_NUMLISTS; try++) { 5140 uint64_t passed_sz = 0; 5141 5142 list = l2arc_list_locked(try, &list_lock); 5143 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_iter); 5144 5145 /* 5146 * L2ARC fast warmup. 5147 * 5148 * Until the ARC is warm and starts to evict, read from the 5149 * head of the ARC lists rather than the tail. 5150 */ 5151 if (arc_warm == B_FALSE) 5152 hdr = list_head(list); 5153 else 5154 hdr = list_tail(list); 5155 if (hdr == NULL) 5156 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_null_iter); 5157 5158 headroom = target_sz * l2arc_headroom * 2 / ARC_BUFC_NUMLISTS; 5159 if (do_headroom_boost) 5160 headroom = (headroom * l2arc_headroom_boost) / 100; 5161 5162 for (; hdr; hdr = hdr_prev) { 5163 l2arc_buf_hdr_t *l2hdr; 5164 kmutex_t *hash_lock; 5165 uint64_t buf_sz; 5166 5167 if (arc_warm == B_FALSE) 5168 hdr_prev = list_next(list, hdr); 5169 else 5170 hdr_prev = list_prev(list, hdr); 5171 ARCSTAT_INCR(arcstat_l2_write_buffer_bytes_scanned, hdr->b_size); 5172 5173 hash_lock = HDR_LOCK(hdr); 5174 if (!mutex_tryenter(hash_lock)) { 5175 ARCSTAT_BUMP(arcstat_l2_write_trylock_fail); 5176 /* 5177 * Skip this buffer rather than waiting. 5178 */ 5179 continue; 5180 } 5181 5182 passed_sz += hdr->b_size; 5183 if (passed_sz > headroom) { 5184 /* 5185 * Searched too far. 5186 */ 5187 mutex_exit(hash_lock); 5188 ARCSTAT_BUMP(arcstat_l2_write_passed_headroom); 5189 break; 5190 } 5191 5192 if (!l2arc_write_eligible(guid, hdr)) { 5193 mutex_exit(hash_lock); 5194 continue; 5195 } 5196 5197 if ((write_sz + hdr->b_size) > target_sz) { 5198 full = B_TRUE; 5199 mutex_exit(hash_lock); 5200 ARCSTAT_BUMP(arcstat_l2_write_full); 5201 break; 5202 } 5203 5204 if (pio == NULL) { 5205 /* 5206 * Insert a dummy header on the buflist so 5207 * l2arc_write_done() can find where the 5208 * write buffers begin without searching. 5209 */ 5210 list_insert_head(dev->l2ad_buflist, head); 5211 5212 cb = kmem_alloc( 5213 sizeof (l2arc_write_callback_t), KM_SLEEP); 5214 cb->l2wcb_dev = dev; 5215 cb->l2wcb_head = head; 5216 pio = zio_root(spa, l2arc_write_done, cb, 5217 ZIO_FLAG_CANFAIL); 5218 ARCSTAT_BUMP(arcstat_l2_write_pios); 5219 } 5220 5221 /* 5222 * Create and add a new L2ARC header. 5223 */ 5224 l2hdr = kmem_zalloc(sizeof (l2arc_buf_hdr_t), KM_SLEEP); 5225 l2hdr->b_dev = dev; 5226 hdr->b_flags |= ARC_FLAG_L2_WRITING; 5227 5228 /* 5229 * Temporarily stash the data buffer in b_tmp_cdata. 5230 * The subsequent write step will pick it up from 5231 * there. This is because can't access hdr->b_buf 5232 * without holding the hash_lock, which we in turn 5233 * can't access without holding the ARC list locks 5234 * (which we want to avoid during compression/writing). 5235 */ 5236 l2hdr->b_compress = ZIO_COMPRESS_OFF; 5237 l2hdr->b_asize = hdr->b_size; 5238 l2hdr->b_tmp_cdata = hdr->b_buf->b_data; 5239 5240 buf_sz = hdr->b_size; 5241 hdr->b_l2hdr = l2hdr; 5242 5243 list_insert_head(dev->l2ad_buflist, hdr); 5244 5245 /* 5246 * Compute and store the buffer cksum before 5247 * writing. On debug the cksum is verified first. 5248 */ 5249 arc_cksum_verify(hdr->b_buf); 5250 arc_cksum_compute(hdr->b_buf, B_TRUE); 5251 5252 mutex_exit(hash_lock); 5253 5254 write_sz += buf_sz; 5255 } 5256 5257 mutex_exit(list_lock); 5258 5259 if (full == B_TRUE) 5260 break; 5261 } 5262 5263 /* No buffers selected for writing? */ 5264 if (pio == NULL) { 5265 ASSERT0(write_sz); 5266 mutex_exit(&l2arc_buflist_mtx); 5267 kmem_cache_free(hdr_cache, head); 5268 return (0); 5269 } 5270 5271 /* 5272 * Now start writing the buffers. We're starting at the write head 5273 * and work backwards, retracing the course of the buffer selector 5274 * loop above. 5275 */ 5276 for (hdr = list_prev(dev->l2ad_buflist, head); hdr; 5277 hdr = list_prev(dev->l2ad_buflist, hdr)) { 5278 l2arc_buf_hdr_t *l2hdr; 5279 uint64_t buf_sz; 5280 5281 /* 5282 * We shouldn't need to lock the buffer here, since we flagged 5283 * it as ARC_FLAG_L2_WRITING in the previous step, but we must 5284 * take care to only access its L2 cache parameters. In 5285 * particular, hdr->b_buf may be invalid by now due to 5286 * ARC eviction. 5287 */ 5288 l2hdr = hdr->b_l2hdr; 5289 l2hdr->b_daddr = dev->l2ad_hand; 5290 5291 if ((hdr->b_flags & ARC_FLAG_L2COMPRESS) && 5292 l2hdr->b_asize >= buf_compress_minsz) { 5293 if (l2arc_compress_buf(l2hdr)) { 5294 /* 5295 * If compression succeeded, enable headroom 5296 * boost on the next scan cycle. 5297 */ 5298 *headroom_boost = B_TRUE; 5299 } 5300 } 5301 5302 /* 5303 * Pick up the buffer data we had previously stashed away 5304 * (and now potentially also compressed). 5305 */ 5306 buf_data = l2hdr->b_tmp_cdata; 5307 buf_sz = l2hdr->b_asize; 5308 5309 /* 5310 * If the data has not been compressed, then clear b_tmp_cdata 5311 * to make sure that it points only to a temporary compression 5312 * buffer. 5313 */ 5314 if (!L2ARC_IS_VALID_COMPRESS(l2hdr->b_compress)) 5315 l2hdr->b_tmp_cdata = NULL; 5316 5317 /* Compression may have squashed the buffer to zero length. */ 5318 if (buf_sz != 0) { 5319 uint64_t buf_p_sz; 5320 5321 wzio = zio_write_phys(pio, dev->l2ad_vdev, 5322 dev->l2ad_hand, buf_sz, buf_data, ZIO_CHECKSUM_OFF, 5323 NULL, NULL, ZIO_PRIORITY_ASYNC_WRITE, 5324 ZIO_FLAG_CANFAIL, B_FALSE); 5325 5326 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev, 5327 zio_t *, wzio); 5328 (void) zio_nowait(wzio); 5329 5330 write_asize += buf_sz; 5331 /* 5332 * Keep the clock hand suitably device-aligned. 5333 */ 5334 buf_p_sz = vdev_psize_to_asize(dev->l2ad_vdev, buf_sz); 5335 write_psize += buf_p_sz; 5336 dev->l2ad_hand += buf_p_sz; 5337 } 5338 } 5339 5340 mutex_exit(&l2arc_buflist_mtx); 5341 5342 ASSERT3U(write_asize, <=, target_sz); 5343 ARCSTAT_BUMP(arcstat_l2_writes_sent); 5344 ARCSTAT_INCR(arcstat_l2_write_bytes, write_asize); 5345 ARCSTAT_INCR(arcstat_l2_size, write_sz); 5346 ARCSTAT_INCR(arcstat_l2_asize, write_asize); 5347 vdev_space_update(dev->l2ad_vdev, write_asize, 0, 0); 5348 5349 /* 5350 * Bump device hand to the device start if it is approaching the end. 5351 * l2arc_evict() will already have evicted ahead for this case. 5352 */ 5353 if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) { 5354 dev->l2ad_hand = dev->l2ad_start; 5355 dev->l2ad_evict = dev->l2ad_start; 5356 dev->l2ad_first = B_FALSE; 5357 } 5358 5359 dev->l2ad_writing = B_TRUE; 5360 (void) zio_wait(pio); 5361 dev->l2ad_writing = B_FALSE; 5362 5363 return (write_asize); 5364} 5365 5366/* 5367 * Compresses an L2ARC buffer. 5368 * The data to be compressed must be prefilled in l2hdr->b_tmp_cdata and its 5369 * size in l2hdr->b_asize. This routine tries to compress the data and 5370 * depending on the compression result there are three possible outcomes: 5371 * *) The buffer was incompressible. The original l2hdr contents were left 5372 * untouched and are ready for writing to an L2 device. 5373 * *) The buffer was all-zeros, so there is no need to write it to an L2 5374 * device. To indicate this situation b_tmp_cdata is NULL'ed, b_asize is 5375 * set to zero and b_compress is set to ZIO_COMPRESS_EMPTY. 5376 * *) Compression succeeded and b_tmp_cdata was replaced with a temporary 5377 * data buffer which holds the compressed data to be written, and b_asize 5378 * tells us how much data there is. b_compress is set to the appropriate 5379 * compression algorithm. Once writing is done, invoke 5380 * l2arc_release_cdata_buf on this l2hdr to free this temporary buffer. 5381 * 5382 * Returns B_TRUE if compression succeeded, or B_FALSE if it didn't (the 5383 * buffer was incompressible). 5384 */ 5385static boolean_t 5386l2arc_compress_buf(l2arc_buf_hdr_t *l2hdr) 5387{ 5388 void *cdata; 5389 size_t csize, len, rounded; 5390 5391 ASSERT(l2hdr->b_compress == ZIO_COMPRESS_OFF); 5392 ASSERT(l2hdr->b_tmp_cdata != NULL); 5393 5394 len = l2hdr->b_asize; 5395 cdata = zio_data_buf_alloc(len); 5396 csize = zio_compress_data(ZIO_COMPRESS_LZ4, l2hdr->b_tmp_cdata, 5397 cdata, l2hdr->b_asize); 5398 5399 if (csize == 0) { 5400 /* zero block, indicate that there's nothing to write */ 5401 zio_data_buf_free(cdata, len); 5402 l2hdr->b_compress = ZIO_COMPRESS_EMPTY; 5403 l2hdr->b_asize = 0; 5404 l2hdr->b_tmp_cdata = NULL; 5405 ARCSTAT_BUMP(arcstat_l2_compress_zeros); 5406 return (B_TRUE); 5407 } 5408 5409 rounded = P2ROUNDUP(csize, 5410 (size_t)1 << l2hdr->b_dev->l2ad_vdev->vdev_ashift); 5411 if (rounded < len) { 5412 /* 5413 * Compression succeeded, we'll keep the cdata around for 5414 * writing and release it afterwards. 5415 */ 5416 if (rounded > csize) { 5417 bzero((char *)cdata + csize, rounded - csize); 5418 csize = rounded; 5419 } 5420 l2hdr->b_compress = ZIO_COMPRESS_LZ4; 5421 l2hdr->b_asize = csize; 5422 l2hdr->b_tmp_cdata = cdata; 5423 ARCSTAT_BUMP(arcstat_l2_compress_successes); 5424 return (B_TRUE); 5425 } else { 5426 /* 5427 * Compression failed, release the compressed buffer. 5428 * l2hdr will be left unmodified. 5429 */ 5430 zio_data_buf_free(cdata, len); 5431 ARCSTAT_BUMP(arcstat_l2_compress_failures); 5432 return (B_FALSE); 5433 } 5434} 5435 5436/* 5437 * Decompresses a zio read back from an l2arc device. On success, the 5438 * underlying zio's io_data buffer is overwritten by the uncompressed 5439 * version. On decompression error (corrupt compressed stream), the 5440 * zio->io_error value is set to signal an I/O error. 5441 * 5442 * Please note that the compressed data stream is not checksummed, so 5443 * if the underlying device is experiencing data corruption, we may feed 5444 * corrupt data to the decompressor, so the decompressor needs to be 5445 * able to handle this situation (LZ4 does). 5446 */ 5447static void 5448l2arc_decompress_zio(zio_t *zio, arc_buf_hdr_t *hdr, enum zio_compress c) 5449{ 5450 ASSERT(L2ARC_IS_VALID_COMPRESS(c)); 5451 5452 if (zio->io_error != 0) { 5453 /* 5454 * An io error has occured, just restore the original io 5455 * size in preparation for a main pool read. 5456 */ 5457 zio->io_orig_size = zio->io_size = hdr->b_size; 5458 return; 5459 } 5460 5461 if (c == ZIO_COMPRESS_EMPTY) { 5462 /* 5463 * An empty buffer results in a null zio, which means we 5464 * need to fill its io_data after we're done restoring the 5465 * buffer's contents. 5466 */ 5467 ASSERT(hdr->b_buf != NULL); 5468 bzero(hdr->b_buf->b_data, hdr->b_size); 5469 zio->io_data = zio->io_orig_data = hdr->b_buf->b_data; 5470 } else { 5471 ASSERT(zio->io_data != NULL); 5472 /* 5473 * We copy the compressed data from the start of the arc buffer 5474 * (the zio_read will have pulled in only what we need, the 5475 * rest is garbage which we will overwrite at decompression) 5476 * and then decompress back to the ARC data buffer. This way we 5477 * can minimize copying by simply decompressing back over the 5478 * original compressed data (rather than decompressing to an 5479 * aux buffer and then copying back the uncompressed buffer, 5480 * which is likely to be much larger). 5481 */ 5482 uint64_t csize; 5483 void *cdata; 5484 5485 csize = zio->io_size; 5486 cdata = zio_data_buf_alloc(csize); 5487 bcopy(zio->io_data, cdata, csize); 5488 if (zio_decompress_data(c, cdata, zio->io_data, csize, 5489 hdr->b_size) != 0) 5490 zio->io_error = EIO; 5491 zio_data_buf_free(cdata, csize); 5492 } 5493 5494 /* Restore the expected uncompressed IO size. */ 5495 zio->io_orig_size = zio->io_size = hdr->b_size; 5496} 5497 5498/* 5499 * Releases the temporary b_tmp_cdata buffer in an l2arc header structure. 5500 * This buffer serves as a temporary holder of compressed data while 5501 * the buffer entry is being written to an l2arc device. Once that is 5502 * done, we can dispose of it. 5503 */ 5504static void 5505l2arc_release_cdata_buf(arc_buf_hdr_t *hdr) 5506{ 5507 l2arc_buf_hdr_t *l2hdr = hdr->b_l2hdr; 5508 5509 ASSERT(L2ARC_IS_VALID_COMPRESS(l2hdr->b_compress)); 5510 if (l2hdr->b_compress != ZIO_COMPRESS_EMPTY) { 5511 /* 5512 * If the data was compressed, then we've allocated a 5513 * temporary buffer for it, so now we need to release it. 5514 */ 5515 ASSERT(l2hdr->b_tmp_cdata != NULL); 5516 zio_data_buf_free(l2hdr->b_tmp_cdata, hdr->b_size); 5517 l2hdr->b_tmp_cdata = NULL; 5518 } else { 5519 ASSERT(l2hdr->b_tmp_cdata == NULL); 5520 } 5521} 5522 5523/* 5524 * This thread feeds the L2ARC at regular intervals. This is the beating 5525 * heart of the L2ARC. 5526 */ 5527static void 5528l2arc_feed_thread(void *dummy __unused) 5529{ 5530 callb_cpr_t cpr; 5531 l2arc_dev_t *dev; 5532 spa_t *spa; 5533 uint64_t size, wrote; 5534 clock_t begin, next = ddi_get_lbolt(); 5535 boolean_t headroom_boost = B_FALSE; 5536 5537 CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG); 5538 5539 mutex_enter(&l2arc_feed_thr_lock); 5540 5541 while (l2arc_thread_exit == 0) { 5542 CALLB_CPR_SAFE_BEGIN(&cpr); 5543 (void) cv_timedwait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock, 5544 next - ddi_get_lbolt()); 5545 CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock); 5546 next = ddi_get_lbolt() + hz; 5547 5548 /* 5549 * Quick check for L2ARC devices. 5550 */ 5551 mutex_enter(&l2arc_dev_mtx); 5552 if (l2arc_ndev == 0) { 5553 mutex_exit(&l2arc_dev_mtx); 5554 continue; 5555 } 5556 mutex_exit(&l2arc_dev_mtx); 5557 begin = ddi_get_lbolt(); 5558 5559 /* 5560 * This selects the next l2arc device to write to, and in 5561 * doing so the next spa to feed from: dev->l2ad_spa. This 5562 * will return NULL if there are now no l2arc devices or if 5563 * they are all faulted. 5564 * 5565 * If a device is returned, its spa's config lock is also 5566 * held to prevent device removal. l2arc_dev_get_next() 5567 * will grab and release l2arc_dev_mtx. 5568 */ 5569 if ((dev = l2arc_dev_get_next()) == NULL) 5570 continue; 5571 5572 spa = dev->l2ad_spa; 5573 ASSERT(spa != NULL); 5574 5575 /* 5576 * If the pool is read-only then force the feed thread to 5577 * sleep a little longer. 5578 */ 5579 if (!spa_writeable(spa)) { 5580 next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz; 5581 spa_config_exit(spa, SCL_L2ARC, dev); 5582 continue; 5583 } 5584 5585 /* 5586 * Avoid contributing to memory pressure. 5587 */ 5588 if (arc_reclaim_needed()) { 5589 ARCSTAT_BUMP(arcstat_l2_abort_lowmem); 5590 spa_config_exit(spa, SCL_L2ARC, dev); 5591 continue; 5592 } 5593 5594 ARCSTAT_BUMP(arcstat_l2_feeds); 5595 5596 size = l2arc_write_size(); 5597 5598 /* 5599 * Evict L2ARC buffers that will be overwritten. 5600 */ 5601 l2arc_evict(dev, size, B_FALSE); 5602 5603 /* 5604 * Write ARC buffers. 5605 */ 5606 wrote = l2arc_write_buffers(spa, dev, size, &headroom_boost); 5607 5608 /* 5609 * Calculate interval between writes. 5610 */ 5611 next = l2arc_write_interval(begin, size, wrote); 5612 spa_config_exit(spa, SCL_L2ARC, dev); 5613 } 5614 5615 l2arc_thread_exit = 0; 5616 cv_broadcast(&l2arc_feed_thr_cv); 5617 CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */ 5618 thread_exit(); 5619} 5620 5621boolean_t 5622l2arc_vdev_present(vdev_t *vd) 5623{ 5624 l2arc_dev_t *dev; 5625 5626 mutex_enter(&l2arc_dev_mtx); 5627 for (dev = list_head(l2arc_dev_list); dev != NULL; 5628 dev = list_next(l2arc_dev_list, dev)) { 5629 if (dev->l2ad_vdev == vd) 5630 break; 5631 } 5632 mutex_exit(&l2arc_dev_mtx); 5633 5634 return (dev != NULL); 5635} 5636 5637/* 5638 * Add a vdev for use by the L2ARC. By this point the spa has already 5639 * validated the vdev and opened it. 5640 */ 5641void 5642l2arc_add_vdev(spa_t *spa, vdev_t *vd) 5643{ 5644 l2arc_dev_t *adddev; 5645 5646 ASSERT(!l2arc_vdev_present(vd)); 5647 5648 vdev_ashift_optimize(vd); 5649 5650 /* 5651 * Create a new l2arc device entry. 5652 */ 5653 adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP); 5654 adddev->l2ad_spa = spa; 5655 adddev->l2ad_vdev = vd; 5656 adddev->l2ad_start = VDEV_LABEL_START_SIZE; 5657 adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd); 5658 adddev->l2ad_hand = adddev->l2ad_start; 5659 adddev->l2ad_evict = adddev->l2ad_start; 5660 adddev->l2ad_first = B_TRUE; 5661 adddev->l2ad_writing = B_FALSE; 5662 5663 /* 5664 * This is a list of all ARC buffers that are still valid on the 5665 * device. 5666 */ 5667 adddev->l2ad_buflist = kmem_zalloc(sizeof (list_t), KM_SLEEP); 5668 list_create(adddev->l2ad_buflist, sizeof (arc_buf_hdr_t), 5669 offsetof(arc_buf_hdr_t, b_l2node)); 5670 5671 vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand); 5672 5673 /* 5674 * Add device to global list 5675 */ 5676 mutex_enter(&l2arc_dev_mtx); 5677 list_insert_head(l2arc_dev_list, adddev); 5678 atomic_inc_64(&l2arc_ndev); 5679 mutex_exit(&l2arc_dev_mtx); 5680} 5681 5682/* 5683 * Remove a vdev from the L2ARC. 5684 */ 5685void 5686l2arc_remove_vdev(vdev_t *vd) 5687{ 5688 l2arc_dev_t *dev, *nextdev, *remdev = NULL; 5689 5690 /* 5691 * Find the device by vdev 5692 */ 5693 mutex_enter(&l2arc_dev_mtx); 5694 for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) { 5695 nextdev = list_next(l2arc_dev_list, dev); 5696 if (vd == dev->l2ad_vdev) { 5697 remdev = dev; 5698 break; 5699 } 5700 } 5701 ASSERT(remdev != NULL); 5702 5703 /* 5704 * Remove device from global list 5705 */ 5706 list_remove(l2arc_dev_list, remdev); 5707 l2arc_dev_last = NULL; /* may have been invalidated */ 5708 atomic_dec_64(&l2arc_ndev); 5709 mutex_exit(&l2arc_dev_mtx); 5710 5711 /* 5712 * Clear all buflists and ARC references. L2ARC device flush. 5713 */ 5714 l2arc_evict(remdev, 0, B_TRUE); 5715 list_destroy(remdev->l2ad_buflist); 5716 kmem_free(remdev->l2ad_buflist, sizeof (list_t)); 5717 kmem_free(remdev, sizeof (l2arc_dev_t)); 5718} 5719 5720void 5721l2arc_init(void) 5722{ 5723 l2arc_thread_exit = 0; 5724 l2arc_ndev = 0; 5725 l2arc_writes_sent = 0; 5726 l2arc_writes_done = 0; 5727 5728 mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL); 5729 cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL); 5730 mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL); 5731 mutex_init(&l2arc_buflist_mtx, NULL, MUTEX_DEFAULT, NULL); 5732 mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL); 5733 5734 l2arc_dev_list = &L2ARC_dev_list; 5735 l2arc_free_on_write = &L2ARC_free_on_write; 5736 list_create(l2arc_dev_list, sizeof (l2arc_dev_t), 5737 offsetof(l2arc_dev_t, l2ad_node)); 5738 list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t), 5739 offsetof(l2arc_data_free_t, l2df_list_node)); 5740} 5741 5742void 5743l2arc_fini(void) 5744{ 5745 /* 5746 * This is called from dmu_fini(), which is called from spa_fini(); 5747 * Because of this, we can assume that all l2arc devices have 5748 * already been removed when the pools themselves were removed. 5749 */ 5750 5751 l2arc_do_free_on_write(); 5752 5753 mutex_destroy(&l2arc_feed_thr_lock); 5754 cv_destroy(&l2arc_feed_thr_cv); 5755 mutex_destroy(&l2arc_dev_mtx); 5756 mutex_destroy(&l2arc_buflist_mtx); 5757 mutex_destroy(&l2arc_free_on_write_mtx); 5758 5759 list_destroy(l2arc_dev_list); 5760 list_destroy(l2arc_free_on_write); 5761} 5762 5763void 5764l2arc_start(void) 5765{ 5766 if (!(spa_mode_global & FWRITE)) 5767 return; 5768 5769 (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0, 5770 TS_RUN, minclsyspri); 5771} 5772 5773void 5774l2arc_stop(void) 5775{ 5776 if (!(spa_mode_global & FWRITE)) 5777 return; 5778 5779 mutex_enter(&l2arc_feed_thr_lock); 5780 cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */ 5781 l2arc_thread_exit = 1; 5782 while (l2arc_thread_exit != 0) 5783 cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock); 5784 mutex_exit(&l2arc_feed_thr_lock); 5785} 5786