arc.c revision 275748
1/* 2 * CDDL HEADER START 3 * 4 * The contents of this file are subject to the terms of the 5 * Common Development and Distribution License (the "License"). 6 * You may not use this file except in compliance with the License. 7 * 8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE 9 * or http://www.opensolaris.org/os/licensing. 10 * See the License for the specific language governing permissions 11 * and limitations under the License. 12 * 13 * When distributing Covered Code, include this CDDL HEADER in each 14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE. 15 * If applicable, add the following below this CDDL HEADER, with the 16 * fields enclosed by brackets "[]" replaced with your own identifying 17 * information: Portions Copyright [yyyy] [name of copyright owner] 18 * 19 * CDDL HEADER END 20 */ 21/* 22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved. 23 * Copyright (c) 2011, 2014 by Delphix. All rights reserved. 24 * Copyright (c) 2014 by Saso Kiselkov. All rights reserved. 25 * Copyright 2014 Nexenta Systems, Inc. All rights reserved. 26 */ 27 28/* 29 * DVA-based Adjustable Replacement Cache 30 * 31 * While much of the theory of operation used here is 32 * based on the self-tuning, low overhead replacement cache 33 * presented by Megiddo and Modha at FAST 2003, there are some 34 * significant differences: 35 * 36 * 1. The Megiddo and Modha model assumes any page is evictable. 37 * Pages in its cache cannot be "locked" into memory. This makes 38 * the eviction algorithm simple: evict the last page in the list. 39 * This also make the performance characteristics easy to reason 40 * about. Our cache is not so simple. At any given moment, some 41 * subset of the blocks in the cache are un-evictable because we 42 * have handed out a reference to them. Blocks are only evictable 43 * when there are no external references active. This makes 44 * eviction far more problematic: we choose to evict the evictable 45 * blocks that are the "lowest" in the list. 46 * 47 * There are times when it is not possible to evict the requested 48 * space. In these circumstances we are unable to adjust the cache 49 * size. To prevent the cache growing unbounded at these times we 50 * implement a "cache throttle" that slows the flow of new data 51 * into the cache until we can make space available. 52 * 53 * 2. The Megiddo and Modha model assumes a fixed cache size. 54 * Pages are evicted when the cache is full and there is a cache 55 * miss. Our model has a variable sized cache. It grows with 56 * high use, but also tries to react to memory pressure from the 57 * operating system: decreasing its size when system memory is 58 * tight. 59 * 60 * 3. The Megiddo and Modha model assumes a fixed page size. All 61 * elements of the cache are therefore exactly the same size. So 62 * when adjusting the cache size following a cache miss, its simply 63 * a matter of choosing a single page to evict. In our model, we 64 * have variable sized cache blocks (rangeing from 512 bytes to 65 * 128K bytes). We therefore choose a set of blocks to evict to make 66 * space for a cache miss that approximates as closely as possible 67 * the space used by the new block. 68 * 69 * See also: "ARC: A Self-Tuning, Low Overhead Replacement Cache" 70 * by N. Megiddo & D. Modha, FAST 2003 71 */ 72 73/* 74 * The locking model: 75 * 76 * A new reference to a cache buffer can be obtained in two 77 * ways: 1) via a hash table lookup using the DVA as a key, 78 * or 2) via one of the ARC lists. The arc_read() interface 79 * uses method 1, while the internal arc algorithms for 80 * adjusting the cache use method 2. We therefore provide two 81 * types of locks: 1) the hash table lock array, and 2) the 82 * arc list locks. 83 * 84 * Buffers do not have their own mutexs, rather they rely on the 85 * hash table mutexs for the bulk of their protection (i.e. most 86 * fields in the arc_buf_hdr_t are protected by these mutexs). 87 * 88 * buf_hash_find() returns the appropriate mutex (held) when it 89 * locates the requested buffer in the hash table. It returns 90 * NULL for the mutex if the buffer was not in the table. 91 * 92 * buf_hash_remove() expects the appropriate hash mutex to be 93 * already held before it is invoked. 94 * 95 * Each arc state also has a mutex which is used to protect the 96 * buffer list associated with the state. When attempting to 97 * obtain a hash table lock while holding an arc list lock you 98 * must use: mutex_tryenter() to avoid deadlock. Also note that 99 * the active state mutex must be held before the ghost state mutex. 100 * 101 * Arc buffers may have an associated eviction callback function. 102 * This function will be invoked prior to removing the buffer (e.g. 103 * in arc_do_user_evicts()). Note however that the data associated 104 * with the buffer may be evicted prior to the callback. The callback 105 * must be made with *no locks held* (to prevent deadlock). Additionally, 106 * the users of callbacks must ensure that their private data is 107 * protected from simultaneous callbacks from arc_clear_callback() 108 * and arc_do_user_evicts(). 109 * 110 * Note that the majority of the performance stats are manipulated 111 * with atomic operations. 112 * 113 * The L2ARC uses the l2arc_buflist_mtx global mutex for the following: 114 * 115 * - L2ARC buflist creation 116 * - L2ARC buflist eviction 117 * - L2ARC write completion, which walks L2ARC buflists 118 * - ARC header destruction, as it removes from L2ARC buflists 119 * - ARC header release, as it removes from L2ARC buflists 120 */ 121 122#include <sys/spa.h> 123#include <sys/zio.h> 124#include <sys/zio_compress.h> 125#include <sys/zfs_context.h> 126#include <sys/arc.h> 127#include <sys/refcount.h> 128#include <sys/vdev.h> 129#include <sys/vdev_impl.h> 130#include <sys/dsl_pool.h> 131#ifdef _KERNEL 132#include <sys/dnlc.h> 133#endif 134#include <sys/callb.h> 135#include <sys/kstat.h> 136#include <sys/trim_map.h> 137#include <zfs_fletcher.h> 138#include <sys/sdt.h> 139 140#include <vm/vm_pageout.h> 141#include <machine/vmparam.h> 142 143#ifdef illumos 144#ifndef _KERNEL 145/* set with ZFS_DEBUG=watch, to enable watchpoints on frozen buffers */ 146boolean_t arc_watch = B_FALSE; 147int arc_procfd; 148#endif 149#endif /* illumos */ 150 151static kmutex_t arc_reclaim_thr_lock; 152static kcondvar_t arc_reclaim_thr_cv; /* used to signal reclaim thr */ 153static uint8_t arc_thread_exit; 154 155#define ARC_REDUCE_DNLC_PERCENT 3 156uint_t arc_reduce_dnlc_percent = ARC_REDUCE_DNLC_PERCENT; 157 158typedef enum arc_reclaim_strategy { 159 ARC_RECLAIM_AGGR, /* Aggressive reclaim strategy */ 160 ARC_RECLAIM_CONS /* Conservative reclaim strategy */ 161} arc_reclaim_strategy_t; 162 163/* 164 * The number of iterations through arc_evict_*() before we 165 * drop & reacquire the lock. 166 */ 167int arc_evict_iterations = 100; 168 169/* number of seconds before growing cache again */ 170static int arc_grow_retry = 60; 171 172/* shift of arc_c for calculating both min and max arc_p */ 173static int arc_p_min_shift = 4; 174 175/* log2(fraction of arc to reclaim) */ 176static int arc_shrink_shift = 5; 177 178/* 179 * minimum lifespan of a prefetch block in clock ticks 180 * (initialized in arc_init()) 181 */ 182static int arc_min_prefetch_lifespan; 183 184/* 185 * If this percent of memory is free, don't throttle. 186 */ 187int arc_lotsfree_percent = 10; 188 189static int arc_dead; 190extern int zfs_prefetch_disable; 191 192/* 193 * The arc has filled available memory and has now warmed up. 194 */ 195static boolean_t arc_warm; 196 197uint64_t zfs_arc_max; 198uint64_t zfs_arc_min; 199uint64_t zfs_arc_meta_limit = 0; 200int zfs_arc_grow_retry = 0; 201int zfs_arc_shrink_shift = 0; 202int zfs_arc_p_min_shift = 0; 203int zfs_disable_dup_eviction = 0; 204uint64_t zfs_arc_average_blocksize = 8 * 1024; /* 8KB */ 205u_int zfs_arc_free_target = 0; 206 207static int sysctl_vfs_zfs_arc_free_target(SYSCTL_HANDLER_ARGS); 208static int sysctl_vfs_zfs_arc_meta_limit(SYSCTL_HANDLER_ARGS); 209 210#ifdef _KERNEL 211static void 212arc_free_target_init(void *unused __unused) 213{ 214 215 zfs_arc_free_target = vm_pageout_wakeup_thresh; 216} 217SYSINIT(arc_free_target_init, SI_SUB_KTHREAD_PAGE, SI_ORDER_ANY, 218 arc_free_target_init, NULL); 219 220TUNABLE_QUAD("vfs.zfs.arc_meta_limit", &zfs_arc_meta_limit); 221TUNABLE_INT("vfs.zfs.arc_shrink_shift", &zfs_arc_shrink_shift); 222SYSCTL_DECL(_vfs_zfs); 223SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_max, CTLFLAG_RDTUN, &zfs_arc_max, 0, 224 "Maximum ARC size"); 225SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_min, CTLFLAG_RDTUN, &zfs_arc_min, 0, 226 "Minimum ARC size"); 227SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_average_blocksize, CTLFLAG_RDTUN, 228 &zfs_arc_average_blocksize, 0, 229 "ARC average blocksize"); 230SYSCTL_INT(_vfs_zfs, OID_AUTO, arc_shrink_shift, CTLFLAG_RW, 231 &arc_shrink_shift, 0, 232 "log2(fraction of arc to reclaim)"); 233 234/* 235 * We don't have a tunable for arc_free_target due to the dependency on 236 * pagedaemon initialisation. 237 */ 238SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_free_target, 239 CTLTYPE_UINT | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, sizeof(u_int), 240 sysctl_vfs_zfs_arc_free_target, "IU", 241 "Desired number of free pages below which ARC triggers reclaim"); 242 243static int 244sysctl_vfs_zfs_arc_free_target(SYSCTL_HANDLER_ARGS) 245{ 246 u_int val; 247 int err; 248 249 val = zfs_arc_free_target; 250 err = sysctl_handle_int(oidp, &val, 0, req); 251 if (err != 0 || req->newptr == NULL) 252 return (err); 253 254 if (val < minfree) 255 return (EINVAL); 256 if (val > vm_cnt.v_page_count) 257 return (EINVAL); 258 259 zfs_arc_free_target = val; 260 261 return (0); 262} 263 264/* 265 * Must be declared here, before the definition of corresponding kstat 266 * macro which uses the same names will confuse the compiler. 267 */ 268SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_meta_limit, 269 CTLTYPE_U64 | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, sizeof(uint64_t), 270 sysctl_vfs_zfs_arc_meta_limit, "QU", 271 "ARC metadata limit"); 272#endif 273 274/* 275 * Note that buffers can be in one of 6 states: 276 * ARC_anon - anonymous (discussed below) 277 * ARC_mru - recently used, currently cached 278 * ARC_mru_ghost - recentely used, no longer in cache 279 * ARC_mfu - frequently used, currently cached 280 * ARC_mfu_ghost - frequently used, no longer in cache 281 * ARC_l2c_only - exists in L2ARC but not other states 282 * When there are no active references to the buffer, they are 283 * are linked onto a list in one of these arc states. These are 284 * the only buffers that can be evicted or deleted. Within each 285 * state there are multiple lists, one for meta-data and one for 286 * non-meta-data. Meta-data (indirect blocks, blocks of dnodes, 287 * etc.) is tracked separately so that it can be managed more 288 * explicitly: favored over data, limited explicitly. 289 * 290 * Anonymous buffers are buffers that are not associated with 291 * a DVA. These are buffers that hold dirty block copies 292 * before they are written to stable storage. By definition, 293 * they are "ref'd" and are considered part of arc_mru 294 * that cannot be freed. Generally, they will aquire a DVA 295 * as they are written and migrate onto the arc_mru list. 296 * 297 * The ARC_l2c_only state is for buffers that are in the second 298 * level ARC but no longer in any of the ARC_m* lists. The second 299 * level ARC itself may also contain buffers that are in any of 300 * the ARC_m* states - meaning that a buffer can exist in two 301 * places. The reason for the ARC_l2c_only state is to keep the 302 * buffer header in the hash table, so that reads that hit the 303 * second level ARC benefit from these fast lookups. 304 */ 305 306#define ARCS_LOCK_PAD CACHE_LINE_SIZE 307struct arcs_lock { 308 kmutex_t arcs_lock; 309#ifdef _KERNEL 310 unsigned char pad[(ARCS_LOCK_PAD - sizeof (kmutex_t))]; 311#endif 312}; 313 314/* 315 * must be power of two for mask use to work 316 * 317 */ 318#define ARC_BUFC_NUMDATALISTS 16 319#define ARC_BUFC_NUMMETADATALISTS 16 320#define ARC_BUFC_NUMLISTS (ARC_BUFC_NUMMETADATALISTS + ARC_BUFC_NUMDATALISTS) 321 322typedef struct arc_state { 323 uint64_t arcs_lsize[ARC_BUFC_NUMTYPES]; /* amount of evictable data */ 324 uint64_t arcs_size; /* total amount of data in this state */ 325 list_t arcs_lists[ARC_BUFC_NUMLISTS]; /* list of evictable buffers */ 326 struct arcs_lock arcs_locks[ARC_BUFC_NUMLISTS] __aligned(CACHE_LINE_SIZE); 327} arc_state_t; 328 329#define ARCS_LOCK(s, i) (&((s)->arcs_locks[(i)].arcs_lock)) 330 331/* The 6 states: */ 332static arc_state_t ARC_anon; 333static arc_state_t ARC_mru; 334static arc_state_t ARC_mru_ghost; 335static arc_state_t ARC_mfu; 336static arc_state_t ARC_mfu_ghost; 337static arc_state_t ARC_l2c_only; 338 339typedef struct arc_stats { 340 kstat_named_t arcstat_hits; 341 kstat_named_t arcstat_misses; 342 kstat_named_t arcstat_demand_data_hits; 343 kstat_named_t arcstat_demand_data_misses; 344 kstat_named_t arcstat_demand_metadata_hits; 345 kstat_named_t arcstat_demand_metadata_misses; 346 kstat_named_t arcstat_prefetch_data_hits; 347 kstat_named_t arcstat_prefetch_data_misses; 348 kstat_named_t arcstat_prefetch_metadata_hits; 349 kstat_named_t arcstat_prefetch_metadata_misses; 350 kstat_named_t arcstat_mru_hits; 351 kstat_named_t arcstat_mru_ghost_hits; 352 kstat_named_t arcstat_mfu_hits; 353 kstat_named_t arcstat_mfu_ghost_hits; 354 kstat_named_t arcstat_allocated; 355 kstat_named_t arcstat_deleted; 356 kstat_named_t arcstat_stolen; 357 kstat_named_t arcstat_recycle_miss; 358 /* 359 * Number of buffers that could not be evicted because the hash lock 360 * was held by another thread. The lock may not necessarily be held 361 * by something using the same buffer, since hash locks are shared 362 * by multiple buffers. 363 */ 364 kstat_named_t arcstat_mutex_miss; 365 /* 366 * Number of buffers skipped because they have I/O in progress, are 367 * indrect prefetch buffers that have not lived long enough, or are 368 * not from the spa we're trying to evict from. 369 */ 370 kstat_named_t arcstat_evict_skip; 371 kstat_named_t arcstat_evict_l2_cached; 372 kstat_named_t arcstat_evict_l2_eligible; 373 kstat_named_t arcstat_evict_l2_ineligible; 374 kstat_named_t arcstat_hash_elements; 375 kstat_named_t arcstat_hash_elements_max; 376 kstat_named_t arcstat_hash_collisions; 377 kstat_named_t arcstat_hash_chains; 378 kstat_named_t arcstat_hash_chain_max; 379 kstat_named_t arcstat_p; 380 kstat_named_t arcstat_c; 381 kstat_named_t arcstat_c_min; 382 kstat_named_t arcstat_c_max; 383 kstat_named_t arcstat_size; 384 kstat_named_t arcstat_hdr_size; 385 kstat_named_t arcstat_data_size; 386 kstat_named_t arcstat_other_size; 387 kstat_named_t arcstat_l2_hits; 388 kstat_named_t arcstat_l2_misses; 389 kstat_named_t arcstat_l2_feeds; 390 kstat_named_t arcstat_l2_rw_clash; 391 kstat_named_t arcstat_l2_read_bytes; 392 kstat_named_t arcstat_l2_write_bytes; 393 kstat_named_t arcstat_l2_writes_sent; 394 kstat_named_t arcstat_l2_writes_done; 395 kstat_named_t arcstat_l2_writes_error; 396 kstat_named_t arcstat_l2_writes_hdr_miss; 397 kstat_named_t arcstat_l2_evict_lock_retry; 398 kstat_named_t arcstat_l2_evict_reading; 399 kstat_named_t arcstat_l2_free_on_write; 400 kstat_named_t arcstat_l2_cdata_free_on_write; 401 kstat_named_t arcstat_l2_abort_lowmem; 402 kstat_named_t arcstat_l2_cksum_bad; 403 kstat_named_t arcstat_l2_io_error; 404 kstat_named_t arcstat_l2_size; 405 kstat_named_t arcstat_l2_asize; 406 kstat_named_t arcstat_l2_hdr_size; 407 kstat_named_t arcstat_l2_compress_successes; 408 kstat_named_t arcstat_l2_compress_zeros; 409 kstat_named_t arcstat_l2_compress_failures; 410 kstat_named_t arcstat_l2_write_trylock_fail; 411 kstat_named_t arcstat_l2_write_passed_headroom; 412 kstat_named_t arcstat_l2_write_spa_mismatch; 413 kstat_named_t arcstat_l2_write_in_l2; 414 kstat_named_t arcstat_l2_write_hdr_io_in_progress; 415 kstat_named_t arcstat_l2_write_not_cacheable; 416 kstat_named_t arcstat_l2_write_full; 417 kstat_named_t arcstat_l2_write_buffer_iter; 418 kstat_named_t arcstat_l2_write_pios; 419 kstat_named_t arcstat_l2_write_buffer_bytes_scanned; 420 kstat_named_t arcstat_l2_write_buffer_list_iter; 421 kstat_named_t arcstat_l2_write_buffer_list_null_iter; 422 kstat_named_t arcstat_memory_throttle_count; 423 kstat_named_t arcstat_duplicate_buffers; 424 kstat_named_t arcstat_duplicate_buffers_size; 425 kstat_named_t arcstat_duplicate_reads; 426 kstat_named_t arcstat_meta_used; 427 kstat_named_t arcstat_meta_limit; 428 kstat_named_t arcstat_meta_max; 429} arc_stats_t; 430 431static arc_stats_t arc_stats = { 432 { "hits", KSTAT_DATA_UINT64 }, 433 { "misses", KSTAT_DATA_UINT64 }, 434 { "demand_data_hits", KSTAT_DATA_UINT64 }, 435 { "demand_data_misses", KSTAT_DATA_UINT64 }, 436 { "demand_metadata_hits", KSTAT_DATA_UINT64 }, 437 { "demand_metadata_misses", KSTAT_DATA_UINT64 }, 438 { "prefetch_data_hits", KSTAT_DATA_UINT64 }, 439 { "prefetch_data_misses", KSTAT_DATA_UINT64 }, 440 { "prefetch_metadata_hits", KSTAT_DATA_UINT64 }, 441 { "prefetch_metadata_misses", KSTAT_DATA_UINT64 }, 442 { "mru_hits", KSTAT_DATA_UINT64 }, 443 { "mru_ghost_hits", KSTAT_DATA_UINT64 }, 444 { "mfu_hits", KSTAT_DATA_UINT64 }, 445 { "mfu_ghost_hits", KSTAT_DATA_UINT64 }, 446 { "allocated", KSTAT_DATA_UINT64 }, 447 { "deleted", KSTAT_DATA_UINT64 }, 448 { "stolen", KSTAT_DATA_UINT64 }, 449 { "recycle_miss", KSTAT_DATA_UINT64 }, 450 { "mutex_miss", KSTAT_DATA_UINT64 }, 451 { "evict_skip", KSTAT_DATA_UINT64 }, 452 { "evict_l2_cached", KSTAT_DATA_UINT64 }, 453 { "evict_l2_eligible", KSTAT_DATA_UINT64 }, 454 { "evict_l2_ineligible", KSTAT_DATA_UINT64 }, 455 { "hash_elements", KSTAT_DATA_UINT64 }, 456 { "hash_elements_max", KSTAT_DATA_UINT64 }, 457 { "hash_collisions", KSTAT_DATA_UINT64 }, 458 { "hash_chains", KSTAT_DATA_UINT64 }, 459 { "hash_chain_max", KSTAT_DATA_UINT64 }, 460 { "p", KSTAT_DATA_UINT64 }, 461 { "c", KSTAT_DATA_UINT64 }, 462 { "c_min", KSTAT_DATA_UINT64 }, 463 { "c_max", KSTAT_DATA_UINT64 }, 464 { "size", KSTAT_DATA_UINT64 }, 465 { "hdr_size", KSTAT_DATA_UINT64 }, 466 { "data_size", KSTAT_DATA_UINT64 }, 467 { "other_size", KSTAT_DATA_UINT64 }, 468 { "l2_hits", KSTAT_DATA_UINT64 }, 469 { "l2_misses", KSTAT_DATA_UINT64 }, 470 { "l2_feeds", KSTAT_DATA_UINT64 }, 471 { "l2_rw_clash", KSTAT_DATA_UINT64 }, 472 { "l2_read_bytes", KSTAT_DATA_UINT64 }, 473 { "l2_write_bytes", KSTAT_DATA_UINT64 }, 474 { "l2_writes_sent", KSTAT_DATA_UINT64 }, 475 { "l2_writes_done", KSTAT_DATA_UINT64 }, 476 { "l2_writes_error", KSTAT_DATA_UINT64 }, 477 { "l2_writes_hdr_miss", KSTAT_DATA_UINT64 }, 478 { "l2_evict_lock_retry", KSTAT_DATA_UINT64 }, 479 { "l2_evict_reading", KSTAT_DATA_UINT64 }, 480 { "l2_free_on_write", KSTAT_DATA_UINT64 }, 481 { "l2_cdata_free_on_write", KSTAT_DATA_UINT64 }, 482 { "l2_abort_lowmem", KSTAT_DATA_UINT64 }, 483 { "l2_cksum_bad", KSTAT_DATA_UINT64 }, 484 { "l2_io_error", KSTAT_DATA_UINT64 }, 485 { "l2_size", KSTAT_DATA_UINT64 }, 486 { "l2_asize", KSTAT_DATA_UINT64 }, 487 { "l2_hdr_size", KSTAT_DATA_UINT64 }, 488 { "l2_compress_successes", KSTAT_DATA_UINT64 }, 489 { "l2_compress_zeros", KSTAT_DATA_UINT64 }, 490 { "l2_compress_failures", KSTAT_DATA_UINT64 }, 491 { "l2_write_trylock_fail", KSTAT_DATA_UINT64 }, 492 { "l2_write_passed_headroom", KSTAT_DATA_UINT64 }, 493 { "l2_write_spa_mismatch", KSTAT_DATA_UINT64 }, 494 { "l2_write_in_l2", KSTAT_DATA_UINT64 }, 495 { "l2_write_io_in_progress", KSTAT_DATA_UINT64 }, 496 { "l2_write_not_cacheable", KSTAT_DATA_UINT64 }, 497 { "l2_write_full", KSTAT_DATA_UINT64 }, 498 { "l2_write_buffer_iter", KSTAT_DATA_UINT64 }, 499 { "l2_write_pios", KSTAT_DATA_UINT64 }, 500 { "l2_write_buffer_bytes_scanned", KSTAT_DATA_UINT64 }, 501 { "l2_write_buffer_list_iter", KSTAT_DATA_UINT64 }, 502 { "l2_write_buffer_list_null_iter", KSTAT_DATA_UINT64 }, 503 { "memory_throttle_count", KSTAT_DATA_UINT64 }, 504 { "duplicate_buffers", KSTAT_DATA_UINT64 }, 505 { "duplicate_buffers_size", KSTAT_DATA_UINT64 }, 506 { "duplicate_reads", KSTAT_DATA_UINT64 }, 507 { "arc_meta_used", KSTAT_DATA_UINT64 }, 508 { "arc_meta_limit", KSTAT_DATA_UINT64 }, 509 { "arc_meta_max", KSTAT_DATA_UINT64 } 510}; 511 512#define ARCSTAT(stat) (arc_stats.stat.value.ui64) 513 514#define ARCSTAT_INCR(stat, val) \ 515 atomic_add_64(&arc_stats.stat.value.ui64, (val)) 516 517#define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1) 518#define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1) 519 520#define ARCSTAT_MAX(stat, val) { \ 521 uint64_t m; \ 522 while ((val) > (m = arc_stats.stat.value.ui64) && \ 523 (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \ 524 continue; \ 525} 526 527#define ARCSTAT_MAXSTAT(stat) \ 528 ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64) 529 530/* 531 * We define a macro to allow ARC hits/misses to be easily broken down by 532 * two separate conditions, giving a total of four different subtypes for 533 * each of hits and misses (so eight statistics total). 534 */ 535#define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \ 536 if (cond1) { \ 537 if (cond2) { \ 538 ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \ 539 } else { \ 540 ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \ 541 } \ 542 } else { \ 543 if (cond2) { \ 544 ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \ 545 } else { \ 546 ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\ 547 } \ 548 } 549 550kstat_t *arc_ksp; 551static arc_state_t *arc_anon; 552static arc_state_t *arc_mru; 553static arc_state_t *arc_mru_ghost; 554static arc_state_t *arc_mfu; 555static arc_state_t *arc_mfu_ghost; 556static arc_state_t *arc_l2c_only; 557 558/* 559 * There are several ARC variables that are critical to export as kstats -- 560 * but we don't want to have to grovel around in the kstat whenever we wish to 561 * manipulate them. For these variables, we therefore define them to be in 562 * terms of the statistic variable. This assures that we are not introducing 563 * the possibility of inconsistency by having shadow copies of the variables, 564 * while still allowing the code to be readable. 565 */ 566#define arc_size ARCSTAT(arcstat_size) /* actual total arc size */ 567#define arc_p ARCSTAT(arcstat_p) /* target size of MRU */ 568#define arc_c ARCSTAT(arcstat_c) /* target size of cache */ 569#define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */ 570#define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */ 571#define arc_meta_limit ARCSTAT(arcstat_meta_limit) /* max size for metadata */ 572#define arc_meta_used ARCSTAT(arcstat_meta_used) /* size of metadata */ 573#define arc_meta_max ARCSTAT(arcstat_meta_max) /* max size of metadata */ 574 575#define L2ARC_IS_VALID_COMPRESS(_c_) \ 576 ((_c_) == ZIO_COMPRESS_LZ4 || (_c_) == ZIO_COMPRESS_EMPTY) 577 578static int arc_no_grow; /* Don't try to grow cache size */ 579static uint64_t arc_tempreserve; 580static uint64_t arc_loaned_bytes; 581 582typedef struct l2arc_buf_hdr l2arc_buf_hdr_t; 583 584typedef struct arc_callback arc_callback_t; 585 586struct arc_callback { 587 void *acb_private; 588 arc_done_func_t *acb_done; 589 arc_buf_t *acb_buf; 590 zio_t *acb_zio_dummy; 591 arc_callback_t *acb_next; 592}; 593 594typedef struct arc_write_callback arc_write_callback_t; 595 596struct arc_write_callback { 597 void *awcb_private; 598 arc_done_func_t *awcb_ready; 599 arc_done_func_t *awcb_physdone; 600 arc_done_func_t *awcb_done; 601 arc_buf_t *awcb_buf; 602}; 603 604struct arc_buf_hdr { 605 /* protected by hash lock */ 606 dva_t b_dva; 607 uint64_t b_birth; 608 uint64_t b_cksum0; 609 610 kmutex_t b_freeze_lock; 611 zio_cksum_t *b_freeze_cksum; 612 void *b_thawed; 613 614 arc_buf_hdr_t *b_hash_next; 615 arc_buf_t *b_buf; 616 uint32_t b_flags; 617 uint32_t b_datacnt; 618 619 arc_callback_t *b_acb; 620 kcondvar_t b_cv; 621 622 /* immutable */ 623 arc_buf_contents_t b_type; 624 uint64_t b_size; 625 uint64_t b_spa; 626 627 /* protected by arc state mutex */ 628 arc_state_t *b_state; 629 list_node_t b_arc_node; 630 631 /* updated atomically */ 632 clock_t b_arc_access; 633 634 /* self protecting */ 635 refcount_t b_refcnt; 636 637 l2arc_buf_hdr_t *b_l2hdr; 638 list_node_t b_l2node; 639}; 640 641#ifdef _KERNEL 642static int 643sysctl_vfs_zfs_arc_meta_limit(SYSCTL_HANDLER_ARGS) 644{ 645 uint64_t val; 646 int err; 647 648 val = arc_meta_limit; 649 err = sysctl_handle_64(oidp, &val, 0, req); 650 if (err != 0 || req->newptr == NULL) 651 return (err); 652 653 if (val <= 0 || val > arc_c_max) 654 return (EINVAL); 655 656 arc_meta_limit = val; 657 return (0); 658} 659#endif 660 661static arc_buf_t *arc_eviction_list; 662static kmutex_t arc_eviction_mtx; 663static arc_buf_hdr_t arc_eviction_hdr; 664static void arc_get_data_buf(arc_buf_t *buf); 665static void arc_access(arc_buf_hdr_t *buf, kmutex_t *hash_lock); 666static int arc_evict_needed(arc_buf_contents_t type); 667static void arc_evict_ghost(arc_state_t *state, uint64_t spa, int64_t bytes); 668#ifdef illumos 669static void arc_buf_watch(arc_buf_t *buf); 670#endif /* illumos */ 671 672static boolean_t l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *ab); 673 674#define GHOST_STATE(state) \ 675 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \ 676 (state) == arc_l2c_only) 677 678/* 679 * Private ARC flags. These flags are private ARC only flags that will show up 680 * in b_flags in the arc_hdr_buf_t. Some flags are publicly declared, and can 681 * be passed in as arc_flags in things like arc_read. However, these flags 682 * should never be passed and should only be set by ARC code. When adding new 683 * public flags, make sure not to smash the private ones. 684 */ 685 686#define ARC_IN_HASH_TABLE (1 << 9) /* this buffer is hashed */ 687#define ARC_IO_IN_PROGRESS (1 << 10) /* I/O in progress for buf */ 688#define ARC_IO_ERROR (1 << 11) /* I/O failed for buf */ 689#define ARC_FREED_IN_READ (1 << 12) /* buf freed while in read */ 690#define ARC_BUF_AVAILABLE (1 << 13) /* block not in active use */ 691#define ARC_INDIRECT (1 << 14) /* this is an indirect block */ 692#define ARC_FREE_IN_PROGRESS (1 << 15) /* hdr about to be freed */ 693#define ARC_L2_WRITING (1 << 16) /* L2ARC write in progress */ 694#define ARC_L2_EVICTED (1 << 17) /* evicted during I/O */ 695#define ARC_L2_WRITE_HEAD (1 << 18) /* head of write list */ 696 697#define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_IN_HASH_TABLE) 698#define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_IO_IN_PROGRESS) 699#define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_IO_ERROR) 700#define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_PREFETCH) 701#define HDR_FREED_IN_READ(hdr) ((hdr)->b_flags & ARC_FREED_IN_READ) 702#define HDR_BUF_AVAILABLE(hdr) ((hdr)->b_flags & ARC_BUF_AVAILABLE) 703#define HDR_FREE_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FREE_IN_PROGRESS) 704#define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_L2CACHE) 705#define HDR_L2_READING(hdr) ((hdr)->b_flags & ARC_IO_IN_PROGRESS && \ 706 (hdr)->b_l2hdr != NULL) 707#define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_L2_WRITING) 708#define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_L2_EVICTED) 709#define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_L2_WRITE_HEAD) 710 711/* 712 * Other sizes 713 */ 714 715#define HDR_SIZE ((int64_t)sizeof (arc_buf_hdr_t)) 716#define L2HDR_SIZE ((int64_t)sizeof (l2arc_buf_hdr_t)) 717 718/* 719 * Hash table routines 720 */ 721 722#define HT_LOCK_PAD CACHE_LINE_SIZE 723 724struct ht_lock { 725 kmutex_t ht_lock; 726#ifdef _KERNEL 727 unsigned char pad[(HT_LOCK_PAD - sizeof (kmutex_t))]; 728#endif 729}; 730 731#define BUF_LOCKS 256 732typedef struct buf_hash_table { 733 uint64_t ht_mask; 734 arc_buf_hdr_t **ht_table; 735 struct ht_lock ht_locks[BUF_LOCKS] __aligned(CACHE_LINE_SIZE); 736} buf_hash_table_t; 737 738static buf_hash_table_t buf_hash_table; 739 740#define BUF_HASH_INDEX(spa, dva, birth) \ 741 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask) 742#define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)]) 743#define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock)) 744#define HDR_LOCK(hdr) \ 745 (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth))) 746 747uint64_t zfs_crc64_table[256]; 748 749/* 750 * Level 2 ARC 751 */ 752 753#define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */ 754#define L2ARC_HEADROOM 2 /* num of writes */ 755/* 756 * If we discover during ARC scan any buffers to be compressed, we boost 757 * our headroom for the next scanning cycle by this percentage multiple. 758 */ 759#define L2ARC_HEADROOM_BOOST 200 760#define L2ARC_FEED_SECS 1 /* caching interval secs */ 761#define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */ 762 763#define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent) 764#define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done) 765 766/* L2ARC Performance Tunables */ 767uint64_t l2arc_write_max = L2ARC_WRITE_SIZE; /* default max write size */ 768uint64_t l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra write during warmup */ 769uint64_t l2arc_headroom = L2ARC_HEADROOM; /* number of dev writes */ 770uint64_t l2arc_headroom_boost = L2ARC_HEADROOM_BOOST; 771uint64_t l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */ 772uint64_t l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval milliseconds */ 773boolean_t l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */ 774boolean_t l2arc_feed_again = B_TRUE; /* turbo warmup */ 775boolean_t l2arc_norw = B_TRUE; /* no reads during writes */ 776 777SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_max, CTLFLAG_RW, 778 &l2arc_write_max, 0, "max write size"); 779SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_boost, CTLFLAG_RW, 780 &l2arc_write_boost, 0, "extra write during warmup"); 781SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_headroom, CTLFLAG_RW, 782 &l2arc_headroom, 0, "number of dev writes"); 783SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_secs, CTLFLAG_RW, 784 &l2arc_feed_secs, 0, "interval seconds"); 785SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_min_ms, CTLFLAG_RW, 786 &l2arc_feed_min_ms, 0, "min interval milliseconds"); 787 788SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_noprefetch, CTLFLAG_RW, 789 &l2arc_noprefetch, 0, "don't cache prefetch bufs"); 790SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_feed_again, CTLFLAG_RW, 791 &l2arc_feed_again, 0, "turbo warmup"); 792SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_norw, CTLFLAG_RW, 793 &l2arc_norw, 0, "no reads during writes"); 794 795SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_size, CTLFLAG_RD, 796 &ARC_anon.arcs_size, 0, "size of anonymous state"); 797SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_metadata_lsize, CTLFLAG_RD, 798 &ARC_anon.arcs_lsize[ARC_BUFC_METADATA], 0, "size of anonymous state"); 799SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_data_lsize, CTLFLAG_RD, 800 &ARC_anon.arcs_lsize[ARC_BUFC_DATA], 0, "size of anonymous state"); 801 802SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_size, CTLFLAG_RD, 803 &ARC_mru.arcs_size, 0, "size of mru state"); 804SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_metadata_lsize, CTLFLAG_RD, 805 &ARC_mru.arcs_lsize[ARC_BUFC_METADATA], 0, "size of metadata in mru state"); 806SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_data_lsize, CTLFLAG_RD, 807 &ARC_mru.arcs_lsize[ARC_BUFC_DATA], 0, "size of data in mru state"); 808 809SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_size, CTLFLAG_RD, 810 &ARC_mru_ghost.arcs_size, 0, "size of mru ghost state"); 811SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_metadata_lsize, CTLFLAG_RD, 812 &ARC_mru_ghost.arcs_lsize[ARC_BUFC_METADATA], 0, 813 "size of metadata in mru ghost state"); 814SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_data_lsize, CTLFLAG_RD, 815 &ARC_mru_ghost.arcs_lsize[ARC_BUFC_DATA], 0, 816 "size of data in mru ghost state"); 817 818SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_size, CTLFLAG_RD, 819 &ARC_mfu.arcs_size, 0, "size of mfu state"); 820SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_metadata_lsize, CTLFLAG_RD, 821 &ARC_mfu.arcs_lsize[ARC_BUFC_METADATA], 0, "size of metadata in mfu state"); 822SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_data_lsize, CTLFLAG_RD, 823 &ARC_mfu.arcs_lsize[ARC_BUFC_DATA], 0, "size of data in mfu state"); 824 825SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_size, CTLFLAG_RD, 826 &ARC_mfu_ghost.arcs_size, 0, "size of mfu ghost state"); 827SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_metadata_lsize, CTLFLAG_RD, 828 &ARC_mfu_ghost.arcs_lsize[ARC_BUFC_METADATA], 0, 829 "size of metadata in mfu ghost state"); 830SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_data_lsize, CTLFLAG_RD, 831 &ARC_mfu_ghost.arcs_lsize[ARC_BUFC_DATA], 0, 832 "size of data in mfu ghost state"); 833 834SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2c_only_size, CTLFLAG_RD, 835 &ARC_l2c_only.arcs_size, 0, "size of mru state"); 836 837/* 838 * L2ARC Internals 839 */ 840typedef struct l2arc_dev { 841 vdev_t *l2ad_vdev; /* vdev */ 842 spa_t *l2ad_spa; /* spa */ 843 uint64_t l2ad_hand; /* next write location */ 844 uint64_t l2ad_start; /* first addr on device */ 845 uint64_t l2ad_end; /* last addr on device */ 846 uint64_t l2ad_evict; /* last addr eviction reached */ 847 boolean_t l2ad_first; /* first sweep through */ 848 boolean_t l2ad_writing; /* currently writing */ 849 list_t *l2ad_buflist; /* buffer list */ 850 list_node_t l2ad_node; /* device list node */ 851} l2arc_dev_t; 852 853static list_t L2ARC_dev_list; /* device list */ 854static list_t *l2arc_dev_list; /* device list pointer */ 855static kmutex_t l2arc_dev_mtx; /* device list mutex */ 856static l2arc_dev_t *l2arc_dev_last; /* last device used */ 857static kmutex_t l2arc_buflist_mtx; /* mutex for all buflists */ 858static list_t L2ARC_free_on_write; /* free after write buf list */ 859static list_t *l2arc_free_on_write; /* free after write list ptr */ 860static kmutex_t l2arc_free_on_write_mtx; /* mutex for list */ 861static uint64_t l2arc_ndev; /* number of devices */ 862 863typedef struct l2arc_read_callback { 864 arc_buf_t *l2rcb_buf; /* read buffer */ 865 spa_t *l2rcb_spa; /* spa */ 866 blkptr_t l2rcb_bp; /* original blkptr */ 867 zbookmark_phys_t l2rcb_zb; /* original bookmark */ 868 int l2rcb_flags; /* original flags */ 869 enum zio_compress l2rcb_compress; /* applied compress */ 870} l2arc_read_callback_t; 871 872typedef struct l2arc_write_callback { 873 l2arc_dev_t *l2wcb_dev; /* device info */ 874 arc_buf_hdr_t *l2wcb_head; /* head of write buflist */ 875} l2arc_write_callback_t; 876 877struct l2arc_buf_hdr { 878 /* protected by arc_buf_hdr mutex */ 879 l2arc_dev_t *b_dev; /* L2ARC device */ 880 uint64_t b_daddr; /* disk address, offset byte */ 881 /* compression applied to buffer data */ 882 enum zio_compress b_compress; 883 /* real alloc'd buffer size depending on b_compress applied */ 884 int b_asize; 885 /* temporary buffer holder for in-flight compressed data */ 886 void *b_tmp_cdata; 887}; 888 889typedef struct l2arc_data_free { 890 /* protected by l2arc_free_on_write_mtx */ 891 void *l2df_data; 892 size_t l2df_size; 893 void (*l2df_func)(void *, size_t); 894 list_node_t l2df_list_node; 895} l2arc_data_free_t; 896 897static kmutex_t l2arc_feed_thr_lock; 898static kcondvar_t l2arc_feed_thr_cv; 899static uint8_t l2arc_thread_exit; 900 901static void l2arc_read_done(zio_t *zio); 902static void l2arc_hdr_stat_add(void); 903static void l2arc_hdr_stat_remove(void); 904 905static boolean_t l2arc_compress_buf(l2arc_buf_hdr_t *l2hdr); 906static void l2arc_decompress_zio(zio_t *zio, arc_buf_hdr_t *hdr, 907 enum zio_compress c); 908static void l2arc_release_cdata_buf(arc_buf_hdr_t *ab); 909 910static uint64_t 911buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth) 912{ 913 uint8_t *vdva = (uint8_t *)dva; 914 uint64_t crc = -1ULL; 915 int i; 916 917 ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY); 918 919 for (i = 0; i < sizeof (dva_t); i++) 920 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ vdva[i]) & 0xFF]; 921 922 crc ^= (spa>>8) ^ birth; 923 924 return (crc); 925} 926 927#define BUF_EMPTY(buf) \ 928 ((buf)->b_dva.dva_word[0] == 0 && \ 929 (buf)->b_dva.dva_word[1] == 0 && \ 930 (buf)->b_cksum0 == 0) 931 932#define BUF_EQUAL(spa, dva, birth, buf) \ 933 ((buf)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \ 934 ((buf)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \ 935 ((buf)->b_birth == birth) && ((buf)->b_spa == spa) 936 937static void 938buf_discard_identity(arc_buf_hdr_t *hdr) 939{ 940 hdr->b_dva.dva_word[0] = 0; 941 hdr->b_dva.dva_word[1] = 0; 942 hdr->b_birth = 0; 943 hdr->b_cksum0 = 0; 944} 945 946static arc_buf_hdr_t * 947buf_hash_find(uint64_t spa, const blkptr_t *bp, kmutex_t **lockp) 948{ 949 const dva_t *dva = BP_IDENTITY(bp); 950 uint64_t birth = BP_PHYSICAL_BIRTH(bp); 951 uint64_t idx = BUF_HASH_INDEX(spa, dva, birth); 952 kmutex_t *hash_lock = BUF_HASH_LOCK(idx); 953 arc_buf_hdr_t *buf; 954 955 mutex_enter(hash_lock); 956 for (buf = buf_hash_table.ht_table[idx]; buf != NULL; 957 buf = buf->b_hash_next) { 958 if (BUF_EQUAL(spa, dva, birth, buf)) { 959 *lockp = hash_lock; 960 return (buf); 961 } 962 } 963 mutex_exit(hash_lock); 964 *lockp = NULL; 965 return (NULL); 966} 967 968/* 969 * Insert an entry into the hash table. If there is already an element 970 * equal to elem in the hash table, then the already existing element 971 * will be returned and the new element will not be inserted. 972 * Otherwise returns NULL. 973 */ 974static arc_buf_hdr_t * 975buf_hash_insert(arc_buf_hdr_t *buf, kmutex_t **lockp) 976{ 977 uint64_t idx = BUF_HASH_INDEX(buf->b_spa, &buf->b_dva, buf->b_birth); 978 kmutex_t *hash_lock = BUF_HASH_LOCK(idx); 979 arc_buf_hdr_t *fbuf; 980 uint32_t i; 981 982 ASSERT(!DVA_IS_EMPTY(&buf->b_dva)); 983 ASSERT(buf->b_birth != 0); 984 ASSERT(!HDR_IN_HASH_TABLE(buf)); 985 *lockp = hash_lock; 986 mutex_enter(hash_lock); 987 for (fbuf = buf_hash_table.ht_table[idx], i = 0; fbuf != NULL; 988 fbuf = fbuf->b_hash_next, i++) { 989 if (BUF_EQUAL(buf->b_spa, &buf->b_dva, buf->b_birth, fbuf)) 990 return (fbuf); 991 } 992 993 buf->b_hash_next = buf_hash_table.ht_table[idx]; 994 buf_hash_table.ht_table[idx] = buf; 995 buf->b_flags |= ARC_IN_HASH_TABLE; 996 997 /* collect some hash table performance data */ 998 if (i > 0) { 999 ARCSTAT_BUMP(arcstat_hash_collisions); 1000 if (i == 1) 1001 ARCSTAT_BUMP(arcstat_hash_chains); 1002 1003 ARCSTAT_MAX(arcstat_hash_chain_max, i); 1004 } 1005 1006 ARCSTAT_BUMP(arcstat_hash_elements); 1007 ARCSTAT_MAXSTAT(arcstat_hash_elements); 1008 1009 return (NULL); 1010} 1011 1012static void 1013buf_hash_remove(arc_buf_hdr_t *buf) 1014{ 1015 arc_buf_hdr_t *fbuf, **bufp; 1016 uint64_t idx = BUF_HASH_INDEX(buf->b_spa, &buf->b_dva, buf->b_birth); 1017 1018 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx))); 1019 ASSERT(HDR_IN_HASH_TABLE(buf)); 1020 1021 bufp = &buf_hash_table.ht_table[idx]; 1022 while ((fbuf = *bufp) != buf) { 1023 ASSERT(fbuf != NULL); 1024 bufp = &fbuf->b_hash_next; 1025 } 1026 *bufp = buf->b_hash_next; 1027 buf->b_hash_next = NULL; 1028 buf->b_flags &= ~ARC_IN_HASH_TABLE; 1029 1030 /* collect some hash table performance data */ 1031 ARCSTAT_BUMPDOWN(arcstat_hash_elements); 1032 1033 if (buf_hash_table.ht_table[idx] && 1034 buf_hash_table.ht_table[idx]->b_hash_next == NULL) 1035 ARCSTAT_BUMPDOWN(arcstat_hash_chains); 1036} 1037 1038/* 1039 * Global data structures and functions for the buf kmem cache. 1040 */ 1041static kmem_cache_t *hdr_cache; 1042static kmem_cache_t *buf_cache; 1043 1044static void 1045buf_fini(void) 1046{ 1047 int i; 1048 1049 kmem_free(buf_hash_table.ht_table, 1050 (buf_hash_table.ht_mask + 1) * sizeof (void *)); 1051 for (i = 0; i < BUF_LOCKS; i++) 1052 mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock); 1053 kmem_cache_destroy(hdr_cache); 1054 kmem_cache_destroy(buf_cache); 1055} 1056 1057/* 1058 * Constructor callback - called when the cache is empty 1059 * and a new buf is requested. 1060 */ 1061/* ARGSUSED */ 1062static int 1063hdr_cons(void *vbuf, void *unused, int kmflag) 1064{ 1065 arc_buf_hdr_t *buf = vbuf; 1066 1067 bzero(buf, sizeof (arc_buf_hdr_t)); 1068 refcount_create(&buf->b_refcnt); 1069 cv_init(&buf->b_cv, NULL, CV_DEFAULT, NULL); 1070 mutex_init(&buf->b_freeze_lock, NULL, MUTEX_DEFAULT, NULL); 1071 arc_space_consume(sizeof (arc_buf_hdr_t), ARC_SPACE_HDRS); 1072 1073 return (0); 1074} 1075 1076/* ARGSUSED */ 1077static int 1078buf_cons(void *vbuf, void *unused, int kmflag) 1079{ 1080 arc_buf_t *buf = vbuf; 1081 1082 bzero(buf, sizeof (arc_buf_t)); 1083 mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL); 1084 arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS); 1085 1086 return (0); 1087} 1088 1089/* 1090 * Destructor callback - called when a cached buf is 1091 * no longer required. 1092 */ 1093/* ARGSUSED */ 1094static void 1095hdr_dest(void *vbuf, void *unused) 1096{ 1097 arc_buf_hdr_t *buf = vbuf; 1098 1099 ASSERT(BUF_EMPTY(buf)); 1100 refcount_destroy(&buf->b_refcnt); 1101 cv_destroy(&buf->b_cv); 1102 mutex_destroy(&buf->b_freeze_lock); 1103 arc_space_return(sizeof (arc_buf_hdr_t), ARC_SPACE_HDRS); 1104} 1105 1106/* ARGSUSED */ 1107static void 1108buf_dest(void *vbuf, void *unused) 1109{ 1110 arc_buf_t *buf = vbuf; 1111 1112 mutex_destroy(&buf->b_evict_lock); 1113 arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS); 1114} 1115 1116/* 1117 * Reclaim callback -- invoked when memory is low. 1118 */ 1119/* ARGSUSED */ 1120static void 1121hdr_recl(void *unused) 1122{ 1123 dprintf("hdr_recl called\n"); 1124 /* 1125 * umem calls the reclaim func when we destroy the buf cache, 1126 * which is after we do arc_fini(). 1127 */ 1128 if (!arc_dead) 1129 cv_signal(&arc_reclaim_thr_cv); 1130} 1131 1132static void 1133buf_init(void) 1134{ 1135 uint64_t *ct; 1136 uint64_t hsize = 1ULL << 12; 1137 int i, j; 1138 1139 /* 1140 * The hash table is big enough to fill all of physical memory 1141 * with an average block size of zfs_arc_average_blocksize (default 8K). 1142 * By default, the table will take up 1143 * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers). 1144 */ 1145 while (hsize * zfs_arc_average_blocksize < (uint64_t)physmem * PAGESIZE) 1146 hsize <<= 1; 1147retry: 1148 buf_hash_table.ht_mask = hsize - 1; 1149 buf_hash_table.ht_table = 1150 kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP); 1151 if (buf_hash_table.ht_table == NULL) { 1152 ASSERT(hsize > (1ULL << 8)); 1153 hsize >>= 1; 1154 goto retry; 1155 } 1156 1157 hdr_cache = kmem_cache_create("arc_buf_hdr_t", sizeof (arc_buf_hdr_t), 1158 0, hdr_cons, hdr_dest, hdr_recl, NULL, NULL, 0); 1159 buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t), 1160 0, buf_cons, buf_dest, NULL, NULL, NULL, 0); 1161 1162 for (i = 0; i < 256; i++) 1163 for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--) 1164 *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY); 1165 1166 for (i = 0; i < BUF_LOCKS; i++) { 1167 mutex_init(&buf_hash_table.ht_locks[i].ht_lock, 1168 NULL, MUTEX_DEFAULT, NULL); 1169 } 1170} 1171 1172#define ARC_MINTIME (hz>>4) /* 62 ms */ 1173 1174static void 1175arc_cksum_verify(arc_buf_t *buf) 1176{ 1177 zio_cksum_t zc; 1178 1179 if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 1180 return; 1181 1182 mutex_enter(&buf->b_hdr->b_freeze_lock); 1183 if (buf->b_hdr->b_freeze_cksum == NULL || 1184 (buf->b_hdr->b_flags & ARC_IO_ERROR)) { 1185 mutex_exit(&buf->b_hdr->b_freeze_lock); 1186 return; 1187 } 1188 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc); 1189 if (!ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc)) 1190 panic("buffer modified while frozen!"); 1191 mutex_exit(&buf->b_hdr->b_freeze_lock); 1192} 1193 1194static int 1195arc_cksum_equal(arc_buf_t *buf) 1196{ 1197 zio_cksum_t zc; 1198 int equal; 1199 1200 mutex_enter(&buf->b_hdr->b_freeze_lock); 1201 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc); 1202 equal = ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc); 1203 mutex_exit(&buf->b_hdr->b_freeze_lock); 1204 1205 return (equal); 1206} 1207 1208static void 1209arc_cksum_compute(arc_buf_t *buf, boolean_t force) 1210{ 1211 if (!force && !(zfs_flags & ZFS_DEBUG_MODIFY)) 1212 return; 1213 1214 mutex_enter(&buf->b_hdr->b_freeze_lock); 1215 if (buf->b_hdr->b_freeze_cksum != NULL) { 1216 mutex_exit(&buf->b_hdr->b_freeze_lock); 1217 return; 1218 } 1219 buf->b_hdr->b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t), KM_SLEEP); 1220 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, 1221 buf->b_hdr->b_freeze_cksum); 1222 mutex_exit(&buf->b_hdr->b_freeze_lock); 1223#ifdef illumos 1224 arc_buf_watch(buf); 1225#endif /* illumos */ 1226} 1227 1228#ifdef illumos 1229#ifndef _KERNEL 1230typedef struct procctl { 1231 long cmd; 1232 prwatch_t prwatch; 1233} procctl_t; 1234#endif 1235 1236/* ARGSUSED */ 1237static void 1238arc_buf_unwatch(arc_buf_t *buf) 1239{ 1240#ifndef _KERNEL 1241 if (arc_watch) { 1242 int result; 1243 procctl_t ctl; 1244 ctl.cmd = PCWATCH; 1245 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data; 1246 ctl.prwatch.pr_size = 0; 1247 ctl.prwatch.pr_wflags = 0; 1248 result = write(arc_procfd, &ctl, sizeof (ctl)); 1249 ASSERT3U(result, ==, sizeof (ctl)); 1250 } 1251#endif 1252} 1253 1254/* ARGSUSED */ 1255static void 1256arc_buf_watch(arc_buf_t *buf) 1257{ 1258#ifndef _KERNEL 1259 if (arc_watch) { 1260 int result; 1261 procctl_t ctl; 1262 ctl.cmd = PCWATCH; 1263 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data; 1264 ctl.prwatch.pr_size = buf->b_hdr->b_size; 1265 ctl.prwatch.pr_wflags = WA_WRITE; 1266 result = write(arc_procfd, &ctl, sizeof (ctl)); 1267 ASSERT3U(result, ==, sizeof (ctl)); 1268 } 1269#endif 1270} 1271#endif /* illumos */ 1272 1273void 1274arc_buf_thaw(arc_buf_t *buf) 1275{ 1276 if (zfs_flags & ZFS_DEBUG_MODIFY) { 1277 if (buf->b_hdr->b_state != arc_anon) 1278 panic("modifying non-anon buffer!"); 1279 if (buf->b_hdr->b_flags & ARC_IO_IN_PROGRESS) 1280 panic("modifying buffer while i/o in progress!"); 1281 arc_cksum_verify(buf); 1282 } 1283 1284 mutex_enter(&buf->b_hdr->b_freeze_lock); 1285 if (buf->b_hdr->b_freeze_cksum != NULL) { 1286 kmem_free(buf->b_hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 1287 buf->b_hdr->b_freeze_cksum = NULL; 1288 } 1289 1290 if (zfs_flags & ZFS_DEBUG_MODIFY) { 1291 if (buf->b_hdr->b_thawed) 1292 kmem_free(buf->b_hdr->b_thawed, 1); 1293 buf->b_hdr->b_thawed = kmem_alloc(1, KM_SLEEP); 1294 } 1295 1296 mutex_exit(&buf->b_hdr->b_freeze_lock); 1297 1298#ifdef illumos 1299 arc_buf_unwatch(buf); 1300#endif /* illumos */ 1301} 1302 1303void 1304arc_buf_freeze(arc_buf_t *buf) 1305{ 1306 kmutex_t *hash_lock; 1307 1308 if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 1309 return; 1310 1311 hash_lock = HDR_LOCK(buf->b_hdr); 1312 mutex_enter(hash_lock); 1313 1314 ASSERT(buf->b_hdr->b_freeze_cksum != NULL || 1315 buf->b_hdr->b_state == arc_anon); 1316 arc_cksum_compute(buf, B_FALSE); 1317 mutex_exit(hash_lock); 1318 1319} 1320 1321static void 1322get_buf_info(arc_buf_hdr_t *ab, arc_state_t *state, list_t **list, kmutex_t **lock) 1323{ 1324 uint64_t buf_hashid = buf_hash(ab->b_spa, &ab->b_dva, ab->b_birth); 1325 1326 if (ab->b_type == ARC_BUFC_METADATA) 1327 buf_hashid &= (ARC_BUFC_NUMMETADATALISTS - 1); 1328 else { 1329 buf_hashid &= (ARC_BUFC_NUMDATALISTS - 1); 1330 buf_hashid += ARC_BUFC_NUMMETADATALISTS; 1331 } 1332 1333 *list = &state->arcs_lists[buf_hashid]; 1334 *lock = ARCS_LOCK(state, buf_hashid); 1335} 1336 1337 1338static void 1339add_reference(arc_buf_hdr_t *ab, kmutex_t *hash_lock, void *tag) 1340{ 1341 ASSERT(MUTEX_HELD(hash_lock)); 1342 1343 if ((refcount_add(&ab->b_refcnt, tag) == 1) && 1344 (ab->b_state != arc_anon)) { 1345 uint64_t delta = ab->b_size * ab->b_datacnt; 1346 uint64_t *size = &ab->b_state->arcs_lsize[ab->b_type]; 1347 list_t *list; 1348 kmutex_t *lock; 1349 1350 get_buf_info(ab, ab->b_state, &list, &lock); 1351 ASSERT(!MUTEX_HELD(lock)); 1352 mutex_enter(lock); 1353 ASSERT(list_link_active(&ab->b_arc_node)); 1354 list_remove(list, ab); 1355 if (GHOST_STATE(ab->b_state)) { 1356 ASSERT0(ab->b_datacnt); 1357 ASSERT3P(ab->b_buf, ==, NULL); 1358 delta = ab->b_size; 1359 } 1360 ASSERT(delta > 0); 1361 ASSERT3U(*size, >=, delta); 1362 atomic_add_64(size, -delta); 1363 mutex_exit(lock); 1364 /* remove the prefetch flag if we get a reference */ 1365 if (ab->b_flags & ARC_PREFETCH) 1366 ab->b_flags &= ~ARC_PREFETCH; 1367 } 1368} 1369 1370static int 1371remove_reference(arc_buf_hdr_t *ab, kmutex_t *hash_lock, void *tag) 1372{ 1373 int cnt; 1374 arc_state_t *state = ab->b_state; 1375 1376 ASSERT(state == arc_anon || MUTEX_HELD(hash_lock)); 1377 ASSERT(!GHOST_STATE(state)); 1378 1379 if (((cnt = refcount_remove(&ab->b_refcnt, tag)) == 0) && 1380 (state != arc_anon)) { 1381 uint64_t *size = &state->arcs_lsize[ab->b_type]; 1382 list_t *list; 1383 kmutex_t *lock; 1384 1385 get_buf_info(ab, state, &list, &lock); 1386 ASSERT(!MUTEX_HELD(lock)); 1387 mutex_enter(lock); 1388 ASSERT(!list_link_active(&ab->b_arc_node)); 1389 list_insert_head(list, ab); 1390 ASSERT(ab->b_datacnt > 0); 1391 atomic_add_64(size, ab->b_size * ab->b_datacnt); 1392 mutex_exit(lock); 1393 } 1394 return (cnt); 1395} 1396 1397/* 1398 * Move the supplied buffer to the indicated state. The mutex 1399 * for the buffer must be held by the caller. 1400 */ 1401static void 1402arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *ab, kmutex_t *hash_lock) 1403{ 1404 arc_state_t *old_state = ab->b_state; 1405 int64_t refcnt = refcount_count(&ab->b_refcnt); 1406 uint64_t from_delta, to_delta; 1407 list_t *list; 1408 kmutex_t *lock; 1409 1410 ASSERT(MUTEX_HELD(hash_lock)); 1411 ASSERT3P(new_state, !=, old_state); 1412 ASSERT(refcnt == 0 || ab->b_datacnt > 0); 1413 ASSERT(ab->b_datacnt == 0 || !GHOST_STATE(new_state)); 1414 ASSERT(ab->b_datacnt <= 1 || old_state != arc_anon); 1415 1416 from_delta = to_delta = ab->b_datacnt * ab->b_size; 1417 1418 /* 1419 * If this buffer is evictable, transfer it from the 1420 * old state list to the new state list. 1421 */ 1422 if (refcnt == 0) { 1423 if (old_state != arc_anon) { 1424 int use_mutex; 1425 uint64_t *size = &old_state->arcs_lsize[ab->b_type]; 1426 1427 get_buf_info(ab, old_state, &list, &lock); 1428 use_mutex = !MUTEX_HELD(lock); 1429 if (use_mutex) 1430 mutex_enter(lock); 1431 1432 ASSERT(list_link_active(&ab->b_arc_node)); 1433 list_remove(list, ab); 1434 1435 /* 1436 * If prefetching out of the ghost cache, 1437 * we will have a non-zero datacnt. 1438 */ 1439 if (GHOST_STATE(old_state) && ab->b_datacnt == 0) { 1440 /* ghost elements have a ghost size */ 1441 ASSERT(ab->b_buf == NULL); 1442 from_delta = ab->b_size; 1443 } 1444 ASSERT3U(*size, >=, from_delta); 1445 atomic_add_64(size, -from_delta); 1446 1447 if (use_mutex) 1448 mutex_exit(lock); 1449 } 1450 if (new_state != arc_anon) { 1451 int use_mutex; 1452 uint64_t *size = &new_state->arcs_lsize[ab->b_type]; 1453 1454 get_buf_info(ab, new_state, &list, &lock); 1455 use_mutex = !MUTEX_HELD(lock); 1456 if (use_mutex) 1457 mutex_enter(lock); 1458 1459 list_insert_head(list, ab); 1460 1461 /* ghost elements have a ghost size */ 1462 if (GHOST_STATE(new_state)) { 1463 ASSERT(ab->b_datacnt == 0); 1464 ASSERT(ab->b_buf == NULL); 1465 to_delta = ab->b_size; 1466 } 1467 atomic_add_64(size, to_delta); 1468 1469 if (use_mutex) 1470 mutex_exit(lock); 1471 } 1472 } 1473 1474 ASSERT(!BUF_EMPTY(ab)); 1475 if (new_state == arc_anon && HDR_IN_HASH_TABLE(ab)) 1476 buf_hash_remove(ab); 1477 1478 /* adjust state sizes */ 1479 if (to_delta) 1480 atomic_add_64(&new_state->arcs_size, to_delta); 1481 if (from_delta) { 1482 ASSERT3U(old_state->arcs_size, >=, from_delta); 1483 atomic_add_64(&old_state->arcs_size, -from_delta); 1484 } 1485 ab->b_state = new_state; 1486 1487 /* adjust l2arc hdr stats */ 1488 if (new_state == arc_l2c_only) 1489 l2arc_hdr_stat_add(); 1490 else if (old_state == arc_l2c_only) 1491 l2arc_hdr_stat_remove(); 1492} 1493 1494void 1495arc_space_consume(uint64_t space, arc_space_type_t type) 1496{ 1497 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES); 1498 1499 switch (type) { 1500 case ARC_SPACE_DATA: 1501 ARCSTAT_INCR(arcstat_data_size, space); 1502 break; 1503 case ARC_SPACE_OTHER: 1504 ARCSTAT_INCR(arcstat_other_size, space); 1505 break; 1506 case ARC_SPACE_HDRS: 1507 ARCSTAT_INCR(arcstat_hdr_size, space); 1508 break; 1509 case ARC_SPACE_L2HDRS: 1510 ARCSTAT_INCR(arcstat_l2_hdr_size, space); 1511 break; 1512 } 1513 1514 ARCSTAT_INCR(arcstat_meta_used, space); 1515 atomic_add_64(&arc_size, space); 1516} 1517 1518void 1519arc_space_return(uint64_t space, arc_space_type_t type) 1520{ 1521 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES); 1522 1523 switch (type) { 1524 case ARC_SPACE_DATA: 1525 ARCSTAT_INCR(arcstat_data_size, -space); 1526 break; 1527 case ARC_SPACE_OTHER: 1528 ARCSTAT_INCR(arcstat_other_size, -space); 1529 break; 1530 case ARC_SPACE_HDRS: 1531 ARCSTAT_INCR(arcstat_hdr_size, -space); 1532 break; 1533 case ARC_SPACE_L2HDRS: 1534 ARCSTAT_INCR(arcstat_l2_hdr_size, -space); 1535 break; 1536 } 1537 1538 ASSERT(arc_meta_used >= space); 1539 if (arc_meta_max < arc_meta_used) 1540 arc_meta_max = arc_meta_used; 1541 ARCSTAT_INCR(arcstat_meta_used, -space); 1542 ASSERT(arc_size >= space); 1543 atomic_add_64(&arc_size, -space); 1544} 1545 1546arc_buf_t * 1547arc_buf_alloc(spa_t *spa, int size, void *tag, arc_buf_contents_t type) 1548{ 1549 arc_buf_hdr_t *hdr; 1550 arc_buf_t *buf; 1551 1552 ASSERT3U(size, >, 0); 1553 hdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE); 1554 ASSERT(BUF_EMPTY(hdr)); 1555 hdr->b_size = size; 1556 hdr->b_type = type; 1557 hdr->b_spa = spa_load_guid(spa); 1558 hdr->b_state = arc_anon; 1559 hdr->b_arc_access = 0; 1560 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 1561 buf->b_hdr = hdr; 1562 buf->b_data = NULL; 1563 buf->b_efunc = NULL; 1564 buf->b_private = NULL; 1565 buf->b_next = NULL; 1566 hdr->b_buf = buf; 1567 arc_get_data_buf(buf); 1568 hdr->b_datacnt = 1; 1569 hdr->b_flags = 0; 1570 ASSERT(refcount_is_zero(&hdr->b_refcnt)); 1571 (void) refcount_add(&hdr->b_refcnt, tag); 1572 1573 return (buf); 1574} 1575 1576static char *arc_onloan_tag = "onloan"; 1577 1578/* 1579 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in 1580 * flight data by arc_tempreserve_space() until they are "returned". Loaned 1581 * buffers must be returned to the arc before they can be used by the DMU or 1582 * freed. 1583 */ 1584arc_buf_t * 1585arc_loan_buf(spa_t *spa, int size) 1586{ 1587 arc_buf_t *buf; 1588 1589 buf = arc_buf_alloc(spa, size, arc_onloan_tag, ARC_BUFC_DATA); 1590 1591 atomic_add_64(&arc_loaned_bytes, size); 1592 return (buf); 1593} 1594 1595/* 1596 * Return a loaned arc buffer to the arc. 1597 */ 1598void 1599arc_return_buf(arc_buf_t *buf, void *tag) 1600{ 1601 arc_buf_hdr_t *hdr = buf->b_hdr; 1602 1603 ASSERT(buf->b_data != NULL); 1604 (void) refcount_add(&hdr->b_refcnt, tag); 1605 (void) refcount_remove(&hdr->b_refcnt, arc_onloan_tag); 1606 1607 atomic_add_64(&arc_loaned_bytes, -hdr->b_size); 1608} 1609 1610/* Detach an arc_buf from a dbuf (tag) */ 1611void 1612arc_loan_inuse_buf(arc_buf_t *buf, void *tag) 1613{ 1614 arc_buf_hdr_t *hdr; 1615 1616 ASSERT(buf->b_data != NULL); 1617 hdr = buf->b_hdr; 1618 (void) refcount_add(&hdr->b_refcnt, arc_onloan_tag); 1619 (void) refcount_remove(&hdr->b_refcnt, tag); 1620 buf->b_efunc = NULL; 1621 buf->b_private = NULL; 1622 1623 atomic_add_64(&arc_loaned_bytes, hdr->b_size); 1624} 1625 1626static arc_buf_t * 1627arc_buf_clone(arc_buf_t *from) 1628{ 1629 arc_buf_t *buf; 1630 arc_buf_hdr_t *hdr = from->b_hdr; 1631 uint64_t size = hdr->b_size; 1632 1633 ASSERT(hdr->b_state != arc_anon); 1634 1635 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 1636 buf->b_hdr = hdr; 1637 buf->b_data = NULL; 1638 buf->b_efunc = NULL; 1639 buf->b_private = NULL; 1640 buf->b_next = hdr->b_buf; 1641 hdr->b_buf = buf; 1642 arc_get_data_buf(buf); 1643 bcopy(from->b_data, buf->b_data, size); 1644 1645 /* 1646 * This buffer already exists in the arc so create a duplicate 1647 * copy for the caller. If the buffer is associated with user data 1648 * then track the size and number of duplicates. These stats will be 1649 * updated as duplicate buffers are created and destroyed. 1650 */ 1651 if (hdr->b_type == ARC_BUFC_DATA) { 1652 ARCSTAT_BUMP(arcstat_duplicate_buffers); 1653 ARCSTAT_INCR(arcstat_duplicate_buffers_size, size); 1654 } 1655 hdr->b_datacnt += 1; 1656 return (buf); 1657} 1658 1659void 1660arc_buf_add_ref(arc_buf_t *buf, void* tag) 1661{ 1662 arc_buf_hdr_t *hdr; 1663 kmutex_t *hash_lock; 1664 1665 /* 1666 * Check to see if this buffer is evicted. Callers 1667 * must verify b_data != NULL to know if the add_ref 1668 * was successful. 1669 */ 1670 mutex_enter(&buf->b_evict_lock); 1671 if (buf->b_data == NULL) { 1672 mutex_exit(&buf->b_evict_lock); 1673 return; 1674 } 1675 hash_lock = HDR_LOCK(buf->b_hdr); 1676 mutex_enter(hash_lock); 1677 hdr = buf->b_hdr; 1678 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 1679 mutex_exit(&buf->b_evict_lock); 1680 1681 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu); 1682 add_reference(hdr, hash_lock, tag); 1683 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr); 1684 arc_access(hdr, hash_lock); 1685 mutex_exit(hash_lock); 1686 ARCSTAT_BUMP(arcstat_hits); 1687 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH), 1688 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA, 1689 data, metadata, hits); 1690} 1691 1692static void 1693arc_buf_free_on_write(void *data, size_t size, 1694 void (*free_func)(void *, size_t)) 1695{ 1696 l2arc_data_free_t *df; 1697 1698 df = kmem_alloc(sizeof (l2arc_data_free_t), KM_SLEEP); 1699 df->l2df_data = data; 1700 df->l2df_size = size; 1701 df->l2df_func = free_func; 1702 mutex_enter(&l2arc_free_on_write_mtx); 1703 list_insert_head(l2arc_free_on_write, df); 1704 mutex_exit(&l2arc_free_on_write_mtx); 1705} 1706 1707/* 1708 * Free the arc data buffer. If it is an l2arc write in progress, 1709 * the buffer is placed on l2arc_free_on_write to be freed later. 1710 */ 1711static void 1712arc_buf_data_free(arc_buf_t *buf, void (*free_func)(void *, size_t)) 1713{ 1714 arc_buf_hdr_t *hdr = buf->b_hdr; 1715 1716 if (HDR_L2_WRITING(hdr)) { 1717 arc_buf_free_on_write(buf->b_data, hdr->b_size, free_func); 1718 ARCSTAT_BUMP(arcstat_l2_free_on_write); 1719 } else { 1720 free_func(buf->b_data, hdr->b_size); 1721 } 1722} 1723 1724/* 1725 * Free up buf->b_data and if 'remove' is set, then pull the 1726 * arc_buf_t off of the the arc_buf_hdr_t's list and free it. 1727 */ 1728static void 1729arc_buf_l2_cdata_free(arc_buf_hdr_t *hdr) 1730{ 1731 l2arc_buf_hdr_t *l2hdr = hdr->b_l2hdr; 1732 1733 ASSERT(MUTEX_HELD(&l2arc_buflist_mtx)); 1734 1735 if (l2hdr->b_tmp_cdata == NULL) 1736 return; 1737 1738 ASSERT(HDR_L2_WRITING(hdr)); 1739 arc_buf_free_on_write(l2hdr->b_tmp_cdata, hdr->b_size, 1740 zio_data_buf_free); 1741 ARCSTAT_BUMP(arcstat_l2_cdata_free_on_write); 1742 l2hdr->b_tmp_cdata = NULL; 1743} 1744 1745static void 1746arc_buf_destroy(arc_buf_t *buf, boolean_t recycle, boolean_t remove) 1747{ 1748 arc_buf_t **bufp; 1749 1750 /* free up data associated with the buf */ 1751 if (buf->b_data) { 1752 arc_state_t *state = buf->b_hdr->b_state; 1753 uint64_t size = buf->b_hdr->b_size; 1754 arc_buf_contents_t type = buf->b_hdr->b_type; 1755 1756 arc_cksum_verify(buf); 1757#ifdef illumos 1758 arc_buf_unwatch(buf); 1759#endif /* illumos */ 1760 1761 if (!recycle) { 1762 if (type == ARC_BUFC_METADATA) { 1763 arc_buf_data_free(buf, zio_buf_free); 1764 arc_space_return(size, ARC_SPACE_DATA); 1765 } else { 1766 ASSERT(type == ARC_BUFC_DATA); 1767 arc_buf_data_free(buf, zio_data_buf_free); 1768 ARCSTAT_INCR(arcstat_data_size, -size); 1769 atomic_add_64(&arc_size, -size); 1770 } 1771 } 1772 if (list_link_active(&buf->b_hdr->b_arc_node)) { 1773 uint64_t *cnt = &state->arcs_lsize[type]; 1774 1775 ASSERT(refcount_is_zero(&buf->b_hdr->b_refcnt)); 1776 ASSERT(state != arc_anon); 1777 1778 ASSERT3U(*cnt, >=, size); 1779 atomic_add_64(cnt, -size); 1780 } 1781 ASSERT3U(state->arcs_size, >=, size); 1782 atomic_add_64(&state->arcs_size, -size); 1783 buf->b_data = NULL; 1784 1785 /* 1786 * If we're destroying a duplicate buffer make sure 1787 * that the appropriate statistics are updated. 1788 */ 1789 if (buf->b_hdr->b_datacnt > 1 && 1790 buf->b_hdr->b_type == ARC_BUFC_DATA) { 1791 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers); 1792 ARCSTAT_INCR(arcstat_duplicate_buffers_size, -size); 1793 } 1794 ASSERT(buf->b_hdr->b_datacnt > 0); 1795 buf->b_hdr->b_datacnt -= 1; 1796 } 1797 1798 /* only remove the buf if requested */ 1799 if (!remove) 1800 return; 1801 1802 /* remove the buf from the hdr list */ 1803 for (bufp = &buf->b_hdr->b_buf; *bufp != buf; bufp = &(*bufp)->b_next) 1804 continue; 1805 *bufp = buf->b_next; 1806 buf->b_next = NULL; 1807 1808 ASSERT(buf->b_efunc == NULL); 1809 1810 /* clean up the buf */ 1811 buf->b_hdr = NULL; 1812 kmem_cache_free(buf_cache, buf); 1813} 1814 1815static void 1816arc_hdr_destroy(arc_buf_hdr_t *hdr) 1817{ 1818 ASSERT(refcount_is_zero(&hdr->b_refcnt)); 1819 ASSERT3P(hdr->b_state, ==, arc_anon); 1820 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 1821 l2arc_buf_hdr_t *l2hdr = hdr->b_l2hdr; 1822 1823 if (l2hdr != NULL) { 1824 boolean_t buflist_held = MUTEX_HELD(&l2arc_buflist_mtx); 1825 /* 1826 * To prevent arc_free() and l2arc_evict() from 1827 * attempting to free the same buffer at the same time, 1828 * a FREE_IN_PROGRESS flag is given to arc_free() to 1829 * give it priority. l2arc_evict() can't destroy this 1830 * header while we are waiting on l2arc_buflist_mtx. 1831 * 1832 * The hdr may be removed from l2ad_buflist before we 1833 * grab l2arc_buflist_mtx, so b_l2hdr is rechecked. 1834 */ 1835 if (!buflist_held) { 1836 mutex_enter(&l2arc_buflist_mtx); 1837 l2hdr = hdr->b_l2hdr; 1838 } 1839 1840 if (l2hdr != NULL) { 1841 trim_map_free(l2hdr->b_dev->l2ad_vdev, l2hdr->b_daddr, 1842 hdr->b_size, 0); 1843 list_remove(l2hdr->b_dev->l2ad_buflist, hdr); 1844 arc_buf_l2_cdata_free(hdr); 1845 ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size); 1846 ARCSTAT_INCR(arcstat_l2_asize, -l2hdr->b_asize); 1847 vdev_space_update(l2hdr->b_dev->l2ad_vdev, 1848 -l2hdr->b_asize, 0, 0); 1849 kmem_free(l2hdr, sizeof (l2arc_buf_hdr_t)); 1850 if (hdr->b_state == arc_l2c_only) 1851 l2arc_hdr_stat_remove(); 1852 hdr->b_l2hdr = NULL; 1853 } 1854 1855 if (!buflist_held) 1856 mutex_exit(&l2arc_buflist_mtx); 1857 } 1858 1859 if (!BUF_EMPTY(hdr)) { 1860 ASSERT(!HDR_IN_HASH_TABLE(hdr)); 1861 buf_discard_identity(hdr); 1862 } 1863 while (hdr->b_buf) { 1864 arc_buf_t *buf = hdr->b_buf; 1865 1866 if (buf->b_efunc) { 1867 mutex_enter(&arc_eviction_mtx); 1868 mutex_enter(&buf->b_evict_lock); 1869 ASSERT(buf->b_hdr != NULL); 1870 arc_buf_destroy(hdr->b_buf, FALSE, FALSE); 1871 hdr->b_buf = buf->b_next; 1872 buf->b_hdr = &arc_eviction_hdr; 1873 buf->b_next = arc_eviction_list; 1874 arc_eviction_list = buf; 1875 mutex_exit(&buf->b_evict_lock); 1876 mutex_exit(&arc_eviction_mtx); 1877 } else { 1878 arc_buf_destroy(hdr->b_buf, FALSE, TRUE); 1879 } 1880 } 1881 if (hdr->b_freeze_cksum != NULL) { 1882 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 1883 hdr->b_freeze_cksum = NULL; 1884 } 1885 if (hdr->b_thawed) { 1886 kmem_free(hdr->b_thawed, 1); 1887 hdr->b_thawed = NULL; 1888 } 1889 1890 ASSERT(!list_link_active(&hdr->b_arc_node)); 1891 ASSERT3P(hdr->b_hash_next, ==, NULL); 1892 ASSERT3P(hdr->b_acb, ==, NULL); 1893 kmem_cache_free(hdr_cache, hdr); 1894} 1895 1896void 1897arc_buf_free(arc_buf_t *buf, void *tag) 1898{ 1899 arc_buf_hdr_t *hdr = buf->b_hdr; 1900 int hashed = hdr->b_state != arc_anon; 1901 1902 ASSERT(buf->b_efunc == NULL); 1903 ASSERT(buf->b_data != NULL); 1904 1905 if (hashed) { 1906 kmutex_t *hash_lock = HDR_LOCK(hdr); 1907 1908 mutex_enter(hash_lock); 1909 hdr = buf->b_hdr; 1910 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 1911 1912 (void) remove_reference(hdr, hash_lock, tag); 1913 if (hdr->b_datacnt > 1) { 1914 arc_buf_destroy(buf, FALSE, TRUE); 1915 } else { 1916 ASSERT(buf == hdr->b_buf); 1917 ASSERT(buf->b_efunc == NULL); 1918 hdr->b_flags |= ARC_BUF_AVAILABLE; 1919 } 1920 mutex_exit(hash_lock); 1921 } else if (HDR_IO_IN_PROGRESS(hdr)) { 1922 int destroy_hdr; 1923 /* 1924 * We are in the middle of an async write. Don't destroy 1925 * this buffer unless the write completes before we finish 1926 * decrementing the reference count. 1927 */ 1928 mutex_enter(&arc_eviction_mtx); 1929 (void) remove_reference(hdr, NULL, tag); 1930 ASSERT(refcount_is_zero(&hdr->b_refcnt)); 1931 destroy_hdr = !HDR_IO_IN_PROGRESS(hdr); 1932 mutex_exit(&arc_eviction_mtx); 1933 if (destroy_hdr) 1934 arc_hdr_destroy(hdr); 1935 } else { 1936 if (remove_reference(hdr, NULL, tag) > 0) 1937 arc_buf_destroy(buf, FALSE, TRUE); 1938 else 1939 arc_hdr_destroy(hdr); 1940 } 1941} 1942 1943boolean_t 1944arc_buf_remove_ref(arc_buf_t *buf, void* tag) 1945{ 1946 arc_buf_hdr_t *hdr = buf->b_hdr; 1947 kmutex_t *hash_lock = HDR_LOCK(hdr); 1948 boolean_t no_callback = (buf->b_efunc == NULL); 1949 1950 if (hdr->b_state == arc_anon) { 1951 ASSERT(hdr->b_datacnt == 1); 1952 arc_buf_free(buf, tag); 1953 return (no_callback); 1954 } 1955 1956 mutex_enter(hash_lock); 1957 hdr = buf->b_hdr; 1958 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 1959 ASSERT(hdr->b_state != arc_anon); 1960 ASSERT(buf->b_data != NULL); 1961 1962 (void) remove_reference(hdr, hash_lock, tag); 1963 if (hdr->b_datacnt > 1) { 1964 if (no_callback) 1965 arc_buf_destroy(buf, FALSE, TRUE); 1966 } else if (no_callback) { 1967 ASSERT(hdr->b_buf == buf && buf->b_next == NULL); 1968 ASSERT(buf->b_efunc == NULL); 1969 hdr->b_flags |= ARC_BUF_AVAILABLE; 1970 } 1971 ASSERT(no_callback || hdr->b_datacnt > 1 || 1972 refcount_is_zero(&hdr->b_refcnt)); 1973 mutex_exit(hash_lock); 1974 return (no_callback); 1975} 1976 1977int 1978arc_buf_size(arc_buf_t *buf) 1979{ 1980 return (buf->b_hdr->b_size); 1981} 1982 1983/* 1984 * Called from the DMU to determine if the current buffer should be 1985 * evicted. In order to ensure proper locking, the eviction must be initiated 1986 * from the DMU. Return true if the buffer is associated with user data and 1987 * duplicate buffers still exist. 1988 */ 1989boolean_t 1990arc_buf_eviction_needed(arc_buf_t *buf) 1991{ 1992 arc_buf_hdr_t *hdr; 1993 boolean_t evict_needed = B_FALSE; 1994 1995 if (zfs_disable_dup_eviction) 1996 return (B_FALSE); 1997 1998 mutex_enter(&buf->b_evict_lock); 1999 hdr = buf->b_hdr; 2000 if (hdr == NULL) { 2001 /* 2002 * We are in arc_do_user_evicts(); let that function 2003 * perform the eviction. 2004 */ 2005 ASSERT(buf->b_data == NULL); 2006 mutex_exit(&buf->b_evict_lock); 2007 return (B_FALSE); 2008 } else if (buf->b_data == NULL) { 2009 /* 2010 * We have already been added to the arc eviction list; 2011 * recommend eviction. 2012 */ 2013 ASSERT3P(hdr, ==, &arc_eviction_hdr); 2014 mutex_exit(&buf->b_evict_lock); 2015 return (B_TRUE); 2016 } 2017 2018 if (hdr->b_datacnt > 1 && hdr->b_type == ARC_BUFC_DATA) 2019 evict_needed = B_TRUE; 2020 2021 mutex_exit(&buf->b_evict_lock); 2022 return (evict_needed); 2023} 2024 2025/* 2026 * Evict buffers from list until we've removed the specified number of 2027 * bytes. Move the removed buffers to the appropriate evict state. 2028 * If the recycle flag is set, then attempt to "recycle" a buffer: 2029 * - look for a buffer to evict that is `bytes' long. 2030 * - return the data block from this buffer rather than freeing it. 2031 * This flag is used by callers that are trying to make space for a 2032 * new buffer in a full arc cache. 2033 * 2034 * This function makes a "best effort". It skips over any buffers 2035 * it can't get a hash_lock on, and so may not catch all candidates. 2036 * It may also return without evicting as much space as requested. 2037 */ 2038static void * 2039arc_evict(arc_state_t *state, uint64_t spa, int64_t bytes, boolean_t recycle, 2040 arc_buf_contents_t type) 2041{ 2042 arc_state_t *evicted_state; 2043 uint64_t bytes_evicted = 0, skipped = 0, missed = 0; 2044 int64_t bytes_remaining; 2045 arc_buf_hdr_t *ab, *ab_prev = NULL; 2046 list_t *evicted_list, *list, *evicted_list_start, *list_start; 2047 kmutex_t *lock, *evicted_lock; 2048 kmutex_t *hash_lock; 2049 boolean_t have_lock; 2050 void *stolen = NULL; 2051 arc_buf_hdr_t marker = { 0 }; 2052 int count = 0; 2053 static int evict_metadata_offset, evict_data_offset; 2054 int i, idx, offset, list_count, lists; 2055 2056 ASSERT(state == arc_mru || state == arc_mfu); 2057 2058 evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost; 2059 2060 if (type == ARC_BUFC_METADATA) { 2061 offset = 0; 2062 list_count = ARC_BUFC_NUMMETADATALISTS; 2063 list_start = &state->arcs_lists[0]; 2064 evicted_list_start = &evicted_state->arcs_lists[0]; 2065 idx = evict_metadata_offset; 2066 } else { 2067 offset = ARC_BUFC_NUMMETADATALISTS; 2068 list_start = &state->arcs_lists[offset]; 2069 evicted_list_start = &evicted_state->arcs_lists[offset]; 2070 list_count = ARC_BUFC_NUMDATALISTS; 2071 idx = evict_data_offset; 2072 } 2073 bytes_remaining = evicted_state->arcs_lsize[type]; 2074 lists = 0; 2075 2076evict_start: 2077 list = &list_start[idx]; 2078 evicted_list = &evicted_list_start[idx]; 2079 lock = ARCS_LOCK(state, (offset + idx)); 2080 evicted_lock = ARCS_LOCK(evicted_state, (offset + idx)); 2081 2082 mutex_enter(lock); 2083 mutex_enter(evicted_lock); 2084 2085 for (ab = list_tail(list); ab; ab = ab_prev) { 2086 ab_prev = list_prev(list, ab); 2087 bytes_remaining -= (ab->b_size * ab->b_datacnt); 2088 /* prefetch buffers have a minimum lifespan */ 2089 if (HDR_IO_IN_PROGRESS(ab) || 2090 (spa && ab->b_spa != spa) || 2091 (ab->b_flags & (ARC_PREFETCH|ARC_INDIRECT) && 2092 ddi_get_lbolt() - ab->b_arc_access < 2093 arc_min_prefetch_lifespan)) { 2094 skipped++; 2095 continue; 2096 } 2097 /* "lookahead" for better eviction candidate */ 2098 if (recycle && ab->b_size != bytes && 2099 ab_prev && ab_prev->b_size == bytes) 2100 continue; 2101 2102 /* ignore markers */ 2103 if (ab->b_spa == 0) 2104 continue; 2105 2106 /* 2107 * It may take a long time to evict all the bufs requested. 2108 * To avoid blocking all arc activity, periodically drop 2109 * the arcs_mtx and give other threads a chance to run 2110 * before reacquiring the lock. 2111 * 2112 * If we are looking for a buffer to recycle, we are in 2113 * the hot code path, so don't sleep. 2114 */ 2115 if (!recycle && count++ > arc_evict_iterations) { 2116 list_insert_after(list, ab, &marker); 2117 mutex_exit(evicted_lock); 2118 mutex_exit(lock); 2119 kpreempt(KPREEMPT_SYNC); 2120 mutex_enter(lock); 2121 mutex_enter(evicted_lock); 2122 ab_prev = list_prev(list, &marker); 2123 list_remove(list, &marker); 2124 count = 0; 2125 continue; 2126 } 2127 2128 hash_lock = HDR_LOCK(ab); 2129 have_lock = MUTEX_HELD(hash_lock); 2130 if (have_lock || mutex_tryenter(hash_lock)) { 2131 ASSERT0(refcount_count(&ab->b_refcnt)); 2132 ASSERT(ab->b_datacnt > 0); 2133 while (ab->b_buf) { 2134 arc_buf_t *buf = ab->b_buf; 2135 if (!mutex_tryenter(&buf->b_evict_lock)) { 2136 missed += 1; 2137 break; 2138 } 2139 if (buf->b_data) { 2140 bytes_evicted += ab->b_size; 2141 if (recycle && ab->b_type == type && 2142 ab->b_size == bytes && 2143 !HDR_L2_WRITING(ab)) { 2144 stolen = buf->b_data; 2145 recycle = FALSE; 2146 } 2147 } 2148 if (buf->b_efunc) { 2149 mutex_enter(&arc_eviction_mtx); 2150 arc_buf_destroy(buf, 2151 buf->b_data == stolen, FALSE); 2152 ab->b_buf = buf->b_next; 2153 buf->b_hdr = &arc_eviction_hdr; 2154 buf->b_next = arc_eviction_list; 2155 arc_eviction_list = buf; 2156 mutex_exit(&arc_eviction_mtx); 2157 mutex_exit(&buf->b_evict_lock); 2158 } else { 2159 mutex_exit(&buf->b_evict_lock); 2160 arc_buf_destroy(buf, 2161 buf->b_data == stolen, TRUE); 2162 } 2163 } 2164 2165 if (ab->b_l2hdr) { 2166 ARCSTAT_INCR(arcstat_evict_l2_cached, 2167 ab->b_size); 2168 } else { 2169 if (l2arc_write_eligible(ab->b_spa, ab)) { 2170 ARCSTAT_INCR(arcstat_evict_l2_eligible, 2171 ab->b_size); 2172 } else { 2173 ARCSTAT_INCR( 2174 arcstat_evict_l2_ineligible, 2175 ab->b_size); 2176 } 2177 } 2178 2179 if (ab->b_datacnt == 0) { 2180 arc_change_state(evicted_state, ab, hash_lock); 2181 ASSERT(HDR_IN_HASH_TABLE(ab)); 2182 ab->b_flags |= ARC_IN_HASH_TABLE; 2183 ab->b_flags &= ~ARC_BUF_AVAILABLE; 2184 DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, ab); 2185 } 2186 if (!have_lock) 2187 mutex_exit(hash_lock); 2188 if (bytes >= 0 && bytes_evicted >= bytes) 2189 break; 2190 if (bytes_remaining > 0) { 2191 mutex_exit(evicted_lock); 2192 mutex_exit(lock); 2193 idx = ((idx + 1) & (list_count - 1)); 2194 lists++; 2195 goto evict_start; 2196 } 2197 } else { 2198 missed += 1; 2199 } 2200 } 2201 2202 mutex_exit(evicted_lock); 2203 mutex_exit(lock); 2204 2205 idx = ((idx + 1) & (list_count - 1)); 2206 lists++; 2207 2208 if (bytes_evicted < bytes) { 2209 if (lists < list_count) 2210 goto evict_start; 2211 else 2212 dprintf("only evicted %lld bytes from %x", 2213 (longlong_t)bytes_evicted, state); 2214 } 2215 if (type == ARC_BUFC_METADATA) 2216 evict_metadata_offset = idx; 2217 else 2218 evict_data_offset = idx; 2219 2220 if (skipped) 2221 ARCSTAT_INCR(arcstat_evict_skip, skipped); 2222 2223 if (missed) 2224 ARCSTAT_INCR(arcstat_mutex_miss, missed); 2225 2226 /* 2227 * Note: we have just evicted some data into the ghost state, 2228 * potentially putting the ghost size over the desired size. Rather 2229 * that evicting from the ghost list in this hot code path, leave 2230 * this chore to the arc_reclaim_thread(). 2231 */ 2232 2233 if (stolen) 2234 ARCSTAT_BUMP(arcstat_stolen); 2235 return (stolen); 2236} 2237 2238/* 2239 * Remove buffers from list until we've removed the specified number of 2240 * bytes. Destroy the buffers that are removed. 2241 */ 2242static void 2243arc_evict_ghost(arc_state_t *state, uint64_t spa, int64_t bytes) 2244{ 2245 arc_buf_hdr_t *ab, *ab_prev; 2246 arc_buf_hdr_t marker = { 0 }; 2247 list_t *list, *list_start; 2248 kmutex_t *hash_lock, *lock; 2249 uint64_t bytes_deleted = 0; 2250 uint64_t bufs_skipped = 0; 2251 int count = 0; 2252 static int evict_offset; 2253 int list_count, idx = evict_offset; 2254 int offset, lists = 0; 2255 2256 ASSERT(GHOST_STATE(state)); 2257 2258 /* 2259 * data lists come after metadata lists 2260 */ 2261 list_start = &state->arcs_lists[ARC_BUFC_NUMMETADATALISTS]; 2262 list_count = ARC_BUFC_NUMDATALISTS; 2263 offset = ARC_BUFC_NUMMETADATALISTS; 2264 2265evict_start: 2266 list = &list_start[idx]; 2267 lock = ARCS_LOCK(state, idx + offset); 2268 2269 mutex_enter(lock); 2270 for (ab = list_tail(list); ab; ab = ab_prev) { 2271 ab_prev = list_prev(list, ab); 2272 if (ab->b_type > ARC_BUFC_NUMTYPES) 2273 panic("invalid ab=%p", (void *)ab); 2274 if (spa && ab->b_spa != spa) 2275 continue; 2276 2277 /* ignore markers */ 2278 if (ab->b_spa == 0) 2279 continue; 2280 2281 hash_lock = HDR_LOCK(ab); 2282 /* caller may be trying to modify this buffer, skip it */ 2283 if (MUTEX_HELD(hash_lock)) 2284 continue; 2285 2286 /* 2287 * It may take a long time to evict all the bufs requested. 2288 * To avoid blocking all arc activity, periodically drop 2289 * the arcs_mtx and give other threads a chance to run 2290 * before reacquiring the lock. 2291 */ 2292 if (count++ > arc_evict_iterations) { 2293 list_insert_after(list, ab, &marker); 2294 mutex_exit(lock); 2295 kpreempt(KPREEMPT_SYNC); 2296 mutex_enter(lock); 2297 ab_prev = list_prev(list, &marker); 2298 list_remove(list, &marker); 2299 count = 0; 2300 continue; 2301 } 2302 if (mutex_tryenter(hash_lock)) { 2303 ASSERT(!HDR_IO_IN_PROGRESS(ab)); 2304 ASSERT(ab->b_buf == NULL); 2305 ARCSTAT_BUMP(arcstat_deleted); 2306 bytes_deleted += ab->b_size; 2307 2308 if (ab->b_l2hdr != NULL) { 2309 /* 2310 * This buffer is cached on the 2nd Level ARC; 2311 * don't destroy the header. 2312 */ 2313 arc_change_state(arc_l2c_only, ab, hash_lock); 2314 mutex_exit(hash_lock); 2315 } else { 2316 arc_change_state(arc_anon, ab, hash_lock); 2317 mutex_exit(hash_lock); 2318 arc_hdr_destroy(ab); 2319 } 2320 2321 DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, ab); 2322 if (bytes >= 0 && bytes_deleted >= bytes) 2323 break; 2324 } else if (bytes < 0) { 2325 /* 2326 * Insert a list marker and then wait for the 2327 * hash lock to become available. Once its 2328 * available, restart from where we left off. 2329 */ 2330 list_insert_after(list, ab, &marker); 2331 mutex_exit(lock); 2332 mutex_enter(hash_lock); 2333 mutex_exit(hash_lock); 2334 mutex_enter(lock); 2335 ab_prev = list_prev(list, &marker); 2336 list_remove(list, &marker); 2337 } else { 2338 bufs_skipped += 1; 2339 } 2340 2341 } 2342 mutex_exit(lock); 2343 idx = ((idx + 1) & (ARC_BUFC_NUMDATALISTS - 1)); 2344 lists++; 2345 2346 if (lists < list_count) 2347 goto evict_start; 2348 2349 evict_offset = idx; 2350 if ((uintptr_t)list > (uintptr_t)&state->arcs_lists[ARC_BUFC_NUMMETADATALISTS] && 2351 (bytes < 0 || bytes_deleted < bytes)) { 2352 list_start = &state->arcs_lists[0]; 2353 list_count = ARC_BUFC_NUMMETADATALISTS; 2354 offset = lists = 0; 2355 goto evict_start; 2356 } 2357 2358 if (bufs_skipped) { 2359 ARCSTAT_INCR(arcstat_mutex_miss, bufs_skipped); 2360 ASSERT(bytes >= 0); 2361 } 2362 2363 if (bytes_deleted < bytes) 2364 dprintf("only deleted %lld bytes from %p", 2365 (longlong_t)bytes_deleted, state); 2366} 2367 2368static void 2369arc_adjust(void) 2370{ 2371 int64_t adjustment, delta; 2372 2373 /* 2374 * Adjust MRU size 2375 */ 2376 2377 adjustment = MIN((int64_t)(arc_size - arc_c), 2378 (int64_t)(arc_anon->arcs_size + arc_mru->arcs_size + arc_meta_used - 2379 arc_p)); 2380 2381 if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_DATA] > 0) { 2382 delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_DATA], adjustment); 2383 (void) arc_evict(arc_mru, 0, delta, FALSE, ARC_BUFC_DATA); 2384 adjustment -= delta; 2385 } 2386 2387 if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_METADATA] > 0) { 2388 delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_METADATA], adjustment); 2389 (void) arc_evict(arc_mru, 0, delta, FALSE, 2390 ARC_BUFC_METADATA); 2391 } 2392 2393 /* 2394 * Adjust MFU size 2395 */ 2396 2397 adjustment = arc_size - arc_c; 2398 2399 if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_DATA] > 0) { 2400 delta = MIN(adjustment, arc_mfu->arcs_lsize[ARC_BUFC_DATA]); 2401 (void) arc_evict(arc_mfu, 0, delta, FALSE, ARC_BUFC_DATA); 2402 adjustment -= delta; 2403 } 2404 2405 if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_METADATA] > 0) { 2406 int64_t delta = MIN(adjustment, 2407 arc_mfu->arcs_lsize[ARC_BUFC_METADATA]); 2408 (void) arc_evict(arc_mfu, 0, delta, FALSE, 2409 ARC_BUFC_METADATA); 2410 } 2411 2412 /* 2413 * Adjust ghost lists 2414 */ 2415 2416 adjustment = arc_mru->arcs_size + arc_mru_ghost->arcs_size - arc_c; 2417 2418 if (adjustment > 0 && arc_mru_ghost->arcs_size > 0) { 2419 delta = MIN(arc_mru_ghost->arcs_size, adjustment); 2420 arc_evict_ghost(arc_mru_ghost, 0, delta); 2421 } 2422 2423 adjustment = 2424 arc_mru_ghost->arcs_size + arc_mfu_ghost->arcs_size - arc_c; 2425 2426 if (adjustment > 0 && arc_mfu_ghost->arcs_size > 0) { 2427 delta = MIN(arc_mfu_ghost->arcs_size, adjustment); 2428 arc_evict_ghost(arc_mfu_ghost, 0, delta); 2429 } 2430} 2431 2432static void 2433arc_do_user_evicts(void) 2434{ 2435 static arc_buf_t *tmp_arc_eviction_list; 2436 2437 /* 2438 * Move list over to avoid LOR 2439 */ 2440restart: 2441 mutex_enter(&arc_eviction_mtx); 2442 tmp_arc_eviction_list = arc_eviction_list; 2443 arc_eviction_list = NULL; 2444 mutex_exit(&arc_eviction_mtx); 2445 2446 while (tmp_arc_eviction_list != NULL) { 2447 arc_buf_t *buf = tmp_arc_eviction_list; 2448 tmp_arc_eviction_list = buf->b_next; 2449 mutex_enter(&buf->b_evict_lock); 2450 buf->b_hdr = NULL; 2451 mutex_exit(&buf->b_evict_lock); 2452 2453 if (buf->b_efunc != NULL) 2454 VERIFY0(buf->b_efunc(buf->b_private)); 2455 2456 buf->b_efunc = NULL; 2457 buf->b_private = NULL; 2458 kmem_cache_free(buf_cache, buf); 2459 } 2460 2461 if (arc_eviction_list != NULL) 2462 goto restart; 2463} 2464 2465/* 2466 * Flush all *evictable* data from the cache for the given spa. 2467 * NOTE: this will not touch "active" (i.e. referenced) data. 2468 */ 2469void 2470arc_flush(spa_t *spa) 2471{ 2472 uint64_t guid = 0; 2473 2474 if (spa) 2475 guid = spa_load_guid(spa); 2476 2477 while (arc_mru->arcs_lsize[ARC_BUFC_DATA]) { 2478 (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_DATA); 2479 if (spa) 2480 break; 2481 } 2482 while (arc_mru->arcs_lsize[ARC_BUFC_METADATA]) { 2483 (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_METADATA); 2484 if (spa) 2485 break; 2486 } 2487 while (arc_mfu->arcs_lsize[ARC_BUFC_DATA]) { 2488 (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_DATA); 2489 if (spa) 2490 break; 2491 } 2492 while (arc_mfu->arcs_lsize[ARC_BUFC_METADATA]) { 2493 (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_METADATA); 2494 if (spa) 2495 break; 2496 } 2497 2498 arc_evict_ghost(arc_mru_ghost, guid, -1); 2499 arc_evict_ghost(arc_mfu_ghost, guid, -1); 2500 2501 mutex_enter(&arc_reclaim_thr_lock); 2502 arc_do_user_evicts(); 2503 mutex_exit(&arc_reclaim_thr_lock); 2504 ASSERT(spa || arc_eviction_list == NULL); 2505} 2506 2507void 2508arc_shrink(void) 2509{ 2510 2511 if (arc_c > arc_c_min) { 2512 uint64_t to_free; 2513 2514 DTRACE_PROBE4(arc__shrink, uint64_t, arc_c, uint64_t, 2515 arc_c_min, uint64_t, arc_p, uint64_t, to_free); 2516#ifdef _KERNEL 2517 to_free = arc_c >> arc_shrink_shift; 2518#else 2519 to_free = arc_c >> arc_shrink_shift; 2520#endif 2521 if (arc_c > arc_c_min + to_free) 2522 atomic_add_64(&arc_c, -to_free); 2523 else 2524 arc_c = arc_c_min; 2525 2526 atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift)); 2527 if (arc_c > arc_size) 2528 arc_c = MAX(arc_size, arc_c_min); 2529 if (arc_p > arc_c) 2530 arc_p = (arc_c >> 1); 2531 2532 DTRACE_PROBE2(arc__shrunk, uint64_t, arc_c, uint64_t, 2533 arc_p); 2534 2535 ASSERT(arc_c >= arc_c_min); 2536 ASSERT((int64_t)arc_p >= 0); 2537 } 2538 2539 if (arc_size > arc_c) { 2540 DTRACE_PROBE2(arc__shrink_adjust, uint64_t, arc_size, 2541 uint64_t, arc_c); 2542 arc_adjust(); 2543 } 2544} 2545 2546static int needfree = 0; 2547 2548static int 2549arc_reclaim_needed(void) 2550{ 2551 2552#ifdef _KERNEL 2553 2554 if (needfree) { 2555 DTRACE_PROBE(arc__reclaim_needfree); 2556 return (1); 2557 } 2558 2559 /* 2560 * Cooperate with pagedaemon when it's time for it to scan 2561 * and reclaim some pages. 2562 */ 2563 if (freemem < zfs_arc_free_target) { 2564 DTRACE_PROBE2(arc__reclaim_freemem, uint64_t, 2565 freemem, uint64_t, zfs_arc_free_target); 2566 return (1); 2567 } 2568 2569#ifdef sun 2570 /* 2571 * take 'desfree' extra pages, so we reclaim sooner, rather than later 2572 */ 2573 extra = desfree; 2574 2575 /* 2576 * check that we're out of range of the pageout scanner. It starts to 2577 * schedule paging if freemem is less than lotsfree and needfree. 2578 * lotsfree is the high-water mark for pageout, and needfree is the 2579 * number of needed free pages. We add extra pages here to make sure 2580 * the scanner doesn't start up while we're freeing memory. 2581 */ 2582 if (freemem < lotsfree + needfree + extra) 2583 return (1); 2584 2585 /* 2586 * check to make sure that swapfs has enough space so that anon 2587 * reservations can still succeed. anon_resvmem() checks that the 2588 * availrmem is greater than swapfs_minfree, and the number of reserved 2589 * swap pages. We also add a bit of extra here just to prevent 2590 * circumstances from getting really dire. 2591 */ 2592 if (availrmem < swapfs_minfree + swapfs_reserve + extra) 2593 return (1); 2594 2595 /* 2596 * Check that we have enough availrmem that memory locking (e.g., via 2597 * mlock(3C) or memcntl(2)) can still succeed. (pages_pp_maximum 2598 * stores the number of pages that cannot be locked; when availrmem 2599 * drops below pages_pp_maximum, page locking mechanisms such as 2600 * page_pp_lock() will fail.) 2601 */ 2602 if (availrmem <= pages_pp_maximum) 2603 return (1); 2604 2605#endif /* sun */ 2606#if defined(__i386) || !defined(UMA_MD_SMALL_ALLOC) 2607 /* 2608 * If we're on an i386 platform, it's possible that we'll exhaust the 2609 * kernel heap space before we ever run out of available physical 2610 * memory. Most checks of the size of the heap_area compare against 2611 * tune.t_minarmem, which is the minimum available real memory that we 2612 * can have in the system. However, this is generally fixed at 25 pages 2613 * which is so low that it's useless. In this comparison, we seek to 2614 * calculate the total heap-size, and reclaim if more than 3/4ths of the 2615 * heap is allocated. (Or, in the calculation, if less than 1/4th is 2616 * free) 2617 */ 2618 if (vmem_size(heap_arena, VMEM_FREE) < 2619 (vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC) >> 2)) { 2620 DTRACE_PROBE2(arc__reclaim_used, uint64_t, 2621 vmem_size(heap_arena, VMEM_FREE), uint64_t, 2622 (vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC)) >> 2); 2623 return (1); 2624 } 2625#endif 2626#ifdef sun 2627 /* 2628 * If zio data pages are being allocated out of a separate heap segment, 2629 * then enforce that the size of available vmem for this arena remains 2630 * above about 1/16th free. 2631 * 2632 * Note: The 1/16th arena free requirement was put in place 2633 * to aggressively evict memory from the arc in order to avoid 2634 * memory fragmentation issues. 2635 */ 2636 if (zio_arena != NULL && 2637 vmem_size(zio_arena, VMEM_FREE) < 2638 (vmem_size(zio_arena, VMEM_ALLOC) >> 4)) 2639 return (1); 2640#endif /* sun */ 2641#else /* _KERNEL */ 2642 if (spa_get_random(100) == 0) 2643 return (1); 2644#endif /* _KERNEL */ 2645 DTRACE_PROBE(arc__reclaim_no); 2646 2647 return (0); 2648} 2649 2650extern kmem_cache_t *zio_buf_cache[]; 2651extern kmem_cache_t *zio_data_buf_cache[]; 2652extern kmem_cache_t *range_seg_cache; 2653 2654static void __noinline 2655arc_kmem_reap_now(arc_reclaim_strategy_t strat) 2656{ 2657 size_t i; 2658 kmem_cache_t *prev_cache = NULL; 2659 kmem_cache_t *prev_data_cache = NULL; 2660 2661 DTRACE_PROBE(arc__kmem_reap_start); 2662#ifdef _KERNEL 2663 if (arc_meta_used >= arc_meta_limit) { 2664 /* 2665 * We are exceeding our meta-data cache limit. 2666 * Purge some DNLC entries to release holds on meta-data. 2667 */ 2668 dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent); 2669 } 2670#if defined(__i386) 2671 /* 2672 * Reclaim unused memory from all kmem caches. 2673 */ 2674 kmem_reap(); 2675#endif 2676#endif 2677 2678 /* 2679 * An aggressive reclamation will shrink the cache size as well as 2680 * reap free buffers from the arc kmem caches. 2681 */ 2682 if (strat == ARC_RECLAIM_AGGR) 2683 arc_shrink(); 2684 2685 for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) { 2686 if (zio_buf_cache[i] != prev_cache) { 2687 prev_cache = zio_buf_cache[i]; 2688 kmem_cache_reap_now(zio_buf_cache[i]); 2689 } 2690 if (zio_data_buf_cache[i] != prev_data_cache) { 2691 prev_data_cache = zio_data_buf_cache[i]; 2692 kmem_cache_reap_now(zio_data_buf_cache[i]); 2693 } 2694 } 2695 kmem_cache_reap_now(buf_cache); 2696 kmem_cache_reap_now(hdr_cache); 2697 kmem_cache_reap_now(range_seg_cache); 2698 2699#ifdef sun 2700 /* 2701 * Ask the vmem arena to reclaim unused memory from its 2702 * quantum caches. 2703 */ 2704 if (zio_arena != NULL && strat == ARC_RECLAIM_AGGR) 2705 vmem_qcache_reap(zio_arena); 2706#endif 2707 DTRACE_PROBE(arc__kmem_reap_end); 2708} 2709 2710static void 2711arc_reclaim_thread(void *dummy __unused) 2712{ 2713 clock_t growtime = 0; 2714 arc_reclaim_strategy_t last_reclaim = ARC_RECLAIM_CONS; 2715 callb_cpr_t cpr; 2716 2717 CALLB_CPR_INIT(&cpr, &arc_reclaim_thr_lock, callb_generic_cpr, FTAG); 2718 2719 mutex_enter(&arc_reclaim_thr_lock); 2720 while (arc_thread_exit == 0) { 2721 if (arc_reclaim_needed()) { 2722 2723 if (arc_no_grow) { 2724 if (last_reclaim == ARC_RECLAIM_CONS) { 2725 DTRACE_PROBE(arc__reclaim_aggr_no_grow); 2726 last_reclaim = ARC_RECLAIM_AGGR; 2727 } else { 2728 last_reclaim = ARC_RECLAIM_CONS; 2729 } 2730 } else { 2731 arc_no_grow = TRUE; 2732 last_reclaim = ARC_RECLAIM_AGGR; 2733 DTRACE_PROBE(arc__reclaim_aggr); 2734 membar_producer(); 2735 } 2736 2737 /* reset the growth delay for every reclaim */ 2738 growtime = ddi_get_lbolt() + (arc_grow_retry * hz); 2739 2740 if (needfree && last_reclaim == ARC_RECLAIM_CONS) { 2741 /* 2742 * If needfree is TRUE our vm_lowmem hook 2743 * was called and in that case we must free some 2744 * memory, so switch to aggressive mode. 2745 */ 2746 arc_no_grow = TRUE; 2747 last_reclaim = ARC_RECLAIM_AGGR; 2748 } 2749 arc_kmem_reap_now(last_reclaim); 2750 arc_warm = B_TRUE; 2751 2752 } else if (arc_no_grow && ddi_get_lbolt() >= growtime) { 2753 arc_no_grow = FALSE; 2754 } 2755 2756 arc_adjust(); 2757 2758 if (arc_eviction_list != NULL) 2759 arc_do_user_evicts(); 2760 2761#ifdef _KERNEL 2762 if (needfree) { 2763 needfree = 0; 2764 wakeup(&needfree); 2765 } 2766#endif 2767 2768 /* block until needed, or one second, whichever is shorter */ 2769 CALLB_CPR_SAFE_BEGIN(&cpr); 2770 (void) cv_timedwait(&arc_reclaim_thr_cv, 2771 &arc_reclaim_thr_lock, hz); 2772 CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_thr_lock); 2773 } 2774 2775 arc_thread_exit = 0; 2776 cv_broadcast(&arc_reclaim_thr_cv); 2777 CALLB_CPR_EXIT(&cpr); /* drops arc_reclaim_thr_lock */ 2778 thread_exit(); 2779} 2780 2781/* 2782 * Adapt arc info given the number of bytes we are trying to add and 2783 * the state that we are comming from. This function is only called 2784 * when we are adding new content to the cache. 2785 */ 2786static void 2787arc_adapt(int bytes, arc_state_t *state) 2788{ 2789 int mult; 2790 uint64_t arc_p_min = (arc_c >> arc_p_min_shift); 2791 2792 if (state == arc_l2c_only) 2793 return; 2794 2795 ASSERT(bytes > 0); 2796 /* 2797 * Adapt the target size of the MRU list: 2798 * - if we just hit in the MRU ghost list, then increase 2799 * the target size of the MRU list. 2800 * - if we just hit in the MFU ghost list, then increase 2801 * the target size of the MFU list by decreasing the 2802 * target size of the MRU list. 2803 */ 2804 if (state == arc_mru_ghost) { 2805 mult = ((arc_mru_ghost->arcs_size >= arc_mfu_ghost->arcs_size) ? 2806 1 : (arc_mfu_ghost->arcs_size/arc_mru_ghost->arcs_size)); 2807 mult = MIN(mult, 10); /* avoid wild arc_p adjustment */ 2808 2809 arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult); 2810 } else if (state == arc_mfu_ghost) { 2811 uint64_t delta; 2812 2813 mult = ((arc_mfu_ghost->arcs_size >= arc_mru_ghost->arcs_size) ? 2814 1 : (arc_mru_ghost->arcs_size/arc_mfu_ghost->arcs_size)); 2815 mult = MIN(mult, 10); 2816 2817 delta = MIN(bytes * mult, arc_p); 2818 arc_p = MAX(arc_p_min, arc_p - delta); 2819 } 2820 ASSERT((int64_t)arc_p >= 0); 2821 2822 if (arc_reclaim_needed()) { 2823 cv_signal(&arc_reclaim_thr_cv); 2824 return; 2825 } 2826 2827 if (arc_no_grow) 2828 return; 2829 2830 if (arc_c >= arc_c_max) 2831 return; 2832 2833 /* 2834 * If we're within (2 * maxblocksize) bytes of the target 2835 * cache size, increment the target cache size 2836 */ 2837 if (arc_size > arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) { 2838 DTRACE_PROBE1(arc__inc_adapt, int, bytes); 2839 atomic_add_64(&arc_c, (int64_t)bytes); 2840 if (arc_c > arc_c_max) 2841 arc_c = arc_c_max; 2842 else if (state == arc_anon) 2843 atomic_add_64(&arc_p, (int64_t)bytes); 2844 if (arc_p > arc_c) 2845 arc_p = arc_c; 2846 } 2847 ASSERT((int64_t)arc_p >= 0); 2848} 2849 2850/* 2851 * Check if the cache has reached its limits and eviction is required 2852 * prior to insert. 2853 */ 2854static int 2855arc_evict_needed(arc_buf_contents_t type) 2856{ 2857 if (type == ARC_BUFC_METADATA && arc_meta_used >= arc_meta_limit) 2858 return (1); 2859 2860 if (arc_reclaim_needed()) 2861 return (1); 2862 2863 return (arc_size > arc_c); 2864} 2865 2866/* 2867 * The buffer, supplied as the first argument, needs a data block. 2868 * So, if we are at cache max, determine which cache should be victimized. 2869 * We have the following cases: 2870 * 2871 * 1. Insert for MRU, p > sizeof(arc_anon + arc_mru) -> 2872 * In this situation if we're out of space, but the resident size of the MFU is 2873 * under the limit, victimize the MFU cache to satisfy this insertion request. 2874 * 2875 * 2. Insert for MRU, p <= sizeof(arc_anon + arc_mru) -> 2876 * Here, we've used up all of the available space for the MRU, so we need to 2877 * evict from our own cache instead. Evict from the set of resident MRU 2878 * entries. 2879 * 2880 * 3. Insert for MFU (c - p) > sizeof(arc_mfu) -> 2881 * c minus p represents the MFU space in the cache, since p is the size of the 2882 * cache that is dedicated to the MRU. In this situation there's still space on 2883 * the MFU side, so the MRU side needs to be victimized. 2884 * 2885 * 4. Insert for MFU (c - p) < sizeof(arc_mfu) -> 2886 * MFU's resident set is consuming more space than it has been allotted. In 2887 * this situation, we must victimize our own cache, the MFU, for this insertion. 2888 */ 2889static void 2890arc_get_data_buf(arc_buf_t *buf) 2891{ 2892 arc_state_t *state = buf->b_hdr->b_state; 2893 uint64_t size = buf->b_hdr->b_size; 2894 arc_buf_contents_t type = buf->b_hdr->b_type; 2895 2896 arc_adapt(size, state); 2897 2898 /* 2899 * We have not yet reached cache maximum size, 2900 * just allocate a new buffer. 2901 */ 2902 if (!arc_evict_needed(type)) { 2903 if (type == ARC_BUFC_METADATA) { 2904 buf->b_data = zio_buf_alloc(size); 2905 arc_space_consume(size, ARC_SPACE_DATA); 2906 } else { 2907 ASSERT(type == ARC_BUFC_DATA); 2908 buf->b_data = zio_data_buf_alloc(size); 2909 ARCSTAT_INCR(arcstat_data_size, size); 2910 atomic_add_64(&arc_size, size); 2911 } 2912 goto out; 2913 } 2914 2915 /* 2916 * If we are prefetching from the mfu ghost list, this buffer 2917 * will end up on the mru list; so steal space from there. 2918 */ 2919 if (state == arc_mfu_ghost) 2920 state = buf->b_hdr->b_flags & ARC_PREFETCH ? arc_mru : arc_mfu; 2921 else if (state == arc_mru_ghost) 2922 state = arc_mru; 2923 2924 if (state == arc_mru || state == arc_anon) { 2925 uint64_t mru_used = arc_anon->arcs_size + arc_mru->arcs_size; 2926 state = (arc_mfu->arcs_lsize[type] >= size && 2927 arc_p > mru_used) ? arc_mfu : arc_mru; 2928 } else { 2929 /* MFU cases */ 2930 uint64_t mfu_space = arc_c - arc_p; 2931 state = (arc_mru->arcs_lsize[type] >= size && 2932 mfu_space > arc_mfu->arcs_size) ? arc_mru : arc_mfu; 2933 } 2934 if ((buf->b_data = arc_evict(state, 0, size, TRUE, type)) == NULL) { 2935 if (type == ARC_BUFC_METADATA) { 2936 buf->b_data = zio_buf_alloc(size); 2937 arc_space_consume(size, ARC_SPACE_DATA); 2938 } else { 2939 ASSERT(type == ARC_BUFC_DATA); 2940 buf->b_data = zio_data_buf_alloc(size); 2941 ARCSTAT_INCR(arcstat_data_size, size); 2942 atomic_add_64(&arc_size, size); 2943 } 2944 ARCSTAT_BUMP(arcstat_recycle_miss); 2945 } 2946 ASSERT(buf->b_data != NULL); 2947out: 2948 /* 2949 * Update the state size. Note that ghost states have a 2950 * "ghost size" and so don't need to be updated. 2951 */ 2952 if (!GHOST_STATE(buf->b_hdr->b_state)) { 2953 arc_buf_hdr_t *hdr = buf->b_hdr; 2954 2955 atomic_add_64(&hdr->b_state->arcs_size, size); 2956 if (list_link_active(&hdr->b_arc_node)) { 2957 ASSERT(refcount_is_zero(&hdr->b_refcnt)); 2958 atomic_add_64(&hdr->b_state->arcs_lsize[type], size); 2959 } 2960 /* 2961 * If we are growing the cache, and we are adding anonymous 2962 * data, and we have outgrown arc_p, update arc_p 2963 */ 2964 if (arc_size < arc_c && hdr->b_state == arc_anon && 2965 arc_anon->arcs_size + arc_mru->arcs_size > arc_p) 2966 arc_p = MIN(arc_c, arc_p + size); 2967 } 2968 ARCSTAT_BUMP(arcstat_allocated); 2969} 2970 2971/* 2972 * This routine is called whenever a buffer is accessed. 2973 * NOTE: the hash lock is dropped in this function. 2974 */ 2975static void 2976arc_access(arc_buf_hdr_t *buf, kmutex_t *hash_lock) 2977{ 2978 clock_t now; 2979 2980 ASSERT(MUTEX_HELD(hash_lock)); 2981 2982 if (buf->b_state == arc_anon) { 2983 /* 2984 * This buffer is not in the cache, and does not 2985 * appear in our "ghost" list. Add the new buffer 2986 * to the MRU state. 2987 */ 2988 2989 ASSERT(buf->b_arc_access == 0); 2990 buf->b_arc_access = ddi_get_lbolt(); 2991 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, buf); 2992 arc_change_state(arc_mru, buf, hash_lock); 2993 2994 } else if (buf->b_state == arc_mru) { 2995 now = ddi_get_lbolt(); 2996 2997 /* 2998 * If this buffer is here because of a prefetch, then either: 2999 * - clear the flag if this is a "referencing" read 3000 * (any subsequent access will bump this into the MFU state). 3001 * or 3002 * - move the buffer to the head of the list if this is 3003 * another prefetch (to make it less likely to be evicted). 3004 */ 3005 if ((buf->b_flags & ARC_PREFETCH) != 0) { 3006 if (refcount_count(&buf->b_refcnt) == 0) { 3007 ASSERT(list_link_active(&buf->b_arc_node)); 3008 } else { 3009 buf->b_flags &= ~ARC_PREFETCH; 3010 ARCSTAT_BUMP(arcstat_mru_hits); 3011 } 3012 buf->b_arc_access = now; 3013 return; 3014 } 3015 3016 /* 3017 * This buffer has been "accessed" only once so far, 3018 * but it is still in the cache. Move it to the MFU 3019 * state. 3020 */ 3021 if (now > buf->b_arc_access + ARC_MINTIME) { 3022 /* 3023 * More than 125ms have passed since we 3024 * instantiated this buffer. Move it to the 3025 * most frequently used state. 3026 */ 3027 buf->b_arc_access = now; 3028 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf); 3029 arc_change_state(arc_mfu, buf, hash_lock); 3030 } 3031 ARCSTAT_BUMP(arcstat_mru_hits); 3032 } else if (buf->b_state == arc_mru_ghost) { 3033 arc_state_t *new_state; 3034 /* 3035 * This buffer has been "accessed" recently, but 3036 * was evicted from the cache. Move it to the 3037 * MFU state. 3038 */ 3039 3040 if (buf->b_flags & ARC_PREFETCH) { 3041 new_state = arc_mru; 3042 if (refcount_count(&buf->b_refcnt) > 0) 3043 buf->b_flags &= ~ARC_PREFETCH; 3044 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, buf); 3045 } else { 3046 new_state = arc_mfu; 3047 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf); 3048 } 3049 3050 buf->b_arc_access = ddi_get_lbolt(); 3051 arc_change_state(new_state, buf, hash_lock); 3052 3053 ARCSTAT_BUMP(arcstat_mru_ghost_hits); 3054 } else if (buf->b_state == arc_mfu) { 3055 /* 3056 * This buffer has been accessed more than once and is 3057 * still in the cache. Keep it in the MFU state. 3058 * 3059 * NOTE: an add_reference() that occurred when we did 3060 * the arc_read() will have kicked this off the list. 3061 * If it was a prefetch, we will explicitly move it to 3062 * the head of the list now. 3063 */ 3064 if ((buf->b_flags & ARC_PREFETCH) != 0) { 3065 ASSERT(refcount_count(&buf->b_refcnt) == 0); 3066 ASSERT(list_link_active(&buf->b_arc_node)); 3067 } 3068 ARCSTAT_BUMP(arcstat_mfu_hits); 3069 buf->b_arc_access = ddi_get_lbolt(); 3070 } else if (buf->b_state == arc_mfu_ghost) { 3071 arc_state_t *new_state = arc_mfu; 3072 /* 3073 * This buffer has been accessed more than once but has 3074 * been evicted from the cache. Move it back to the 3075 * MFU state. 3076 */ 3077 3078 if (buf->b_flags & ARC_PREFETCH) { 3079 /* 3080 * This is a prefetch access... 3081 * move this block back to the MRU state. 3082 */ 3083 ASSERT0(refcount_count(&buf->b_refcnt)); 3084 new_state = arc_mru; 3085 } 3086 3087 buf->b_arc_access = ddi_get_lbolt(); 3088 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf); 3089 arc_change_state(new_state, buf, hash_lock); 3090 3091 ARCSTAT_BUMP(arcstat_mfu_ghost_hits); 3092 } else if (buf->b_state == arc_l2c_only) { 3093 /* 3094 * This buffer is on the 2nd Level ARC. 3095 */ 3096 3097 buf->b_arc_access = ddi_get_lbolt(); 3098 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf); 3099 arc_change_state(arc_mfu, buf, hash_lock); 3100 } else { 3101 ASSERT(!"invalid arc state"); 3102 } 3103} 3104 3105/* a generic arc_done_func_t which you can use */ 3106/* ARGSUSED */ 3107void 3108arc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg) 3109{ 3110 if (zio == NULL || zio->io_error == 0) 3111 bcopy(buf->b_data, arg, buf->b_hdr->b_size); 3112 VERIFY(arc_buf_remove_ref(buf, arg)); 3113} 3114 3115/* a generic arc_done_func_t */ 3116void 3117arc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg) 3118{ 3119 arc_buf_t **bufp = arg; 3120 if (zio && zio->io_error) { 3121 VERIFY(arc_buf_remove_ref(buf, arg)); 3122 *bufp = NULL; 3123 } else { 3124 *bufp = buf; 3125 ASSERT(buf->b_data); 3126 } 3127} 3128 3129static void 3130arc_read_done(zio_t *zio) 3131{ 3132 arc_buf_hdr_t *hdr; 3133 arc_buf_t *buf; 3134 arc_buf_t *abuf; /* buffer we're assigning to callback */ 3135 kmutex_t *hash_lock = NULL; 3136 arc_callback_t *callback_list, *acb; 3137 int freeable = FALSE; 3138 3139 buf = zio->io_private; 3140 hdr = buf->b_hdr; 3141 3142 /* 3143 * The hdr was inserted into hash-table and removed from lists 3144 * prior to starting I/O. We should find this header, since 3145 * it's in the hash table, and it should be legit since it's 3146 * not possible to evict it during the I/O. The only possible 3147 * reason for it not to be found is if we were freed during the 3148 * read. 3149 */ 3150 if (HDR_IN_HASH_TABLE(hdr)) { 3151 ASSERT3U(hdr->b_birth, ==, BP_PHYSICAL_BIRTH(zio->io_bp)); 3152 ASSERT3U(hdr->b_dva.dva_word[0], ==, 3153 BP_IDENTITY(zio->io_bp)->dva_word[0]); 3154 ASSERT3U(hdr->b_dva.dva_word[1], ==, 3155 BP_IDENTITY(zio->io_bp)->dva_word[1]); 3156 3157 arc_buf_hdr_t *found = buf_hash_find(hdr->b_spa, zio->io_bp, 3158 &hash_lock); 3159 3160 ASSERT((found == NULL && HDR_FREED_IN_READ(hdr) && 3161 hash_lock == NULL) || 3162 (found == hdr && 3163 DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) || 3164 (found == hdr && HDR_L2_READING(hdr))); 3165 } 3166 3167 hdr->b_flags &= ~ARC_L2_EVICTED; 3168 if (l2arc_noprefetch && (hdr->b_flags & ARC_PREFETCH)) 3169 hdr->b_flags &= ~ARC_L2CACHE; 3170 3171 /* byteswap if necessary */ 3172 callback_list = hdr->b_acb; 3173 ASSERT(callback_list != NULL); 3174 if (BP_SHOULD_BYTESWAP(zio->io_bp) && zio->io_error == 0) { 3175 dmu_object_byteswap_t bswap = 3176 DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp)); 3177 arc_byteswap_func_t *func = BP_GET_LEVEL(zio->io_bp) > 0 ? 3178 byteswap_uint64_array : 3179 dmu_ot_byteswap[bswap].ob_func; 3180 func(buf->b_data, hdr->b_size); 3181 } 3182 3183 arc_cksum_compute(buf, B_FALSE); 3184#ifdef illumos 3185 arc_buf_watch(buf); 3186#endif /* illumos */ 3187 3188 if (hash_lock && zio->io_error == 0 && hdr->b_state == arc_anon) { 3189 /* 3190 * Only call arc_access on anonymous buffers. This is because 3191 * if we've issued an I/O for an evicted buffer, we've already 3192 * called arc_access (to prevent any simultaneous readers from 3193 * getting confused). 3194 */ 3195 arc_access(hdr, hash_lock); 3196 } 3197 3198 /* create copies of the data buffer for the callers */ 3199 abuf = buf; 3200 for (acb = callback_list; acb; acb = acb->acb_next) { 3201 if (acb->acb_done) { 3202 if (abuf == NULL) { 3203 ARCSTAT_BUMP(arcstat_duplicate_reads); 3204 abuf = arc_buf_clone(buf); 3205 } 3206 acb->acb_buf = abuf; 3207 abuf = NULL; 3208 } 3209 } 3210 hdr->b_acb = NULL; 3211 hdr->b_flags &= ~ARC_IO_IN_PROGRESS; 3212 ASSERT(!HDR_BUF_AVAILABLE(hdr)); 3213 if (abuf == buf) { 3214 ASSERT(buf->b_efunc == NULL); 3215 ASSERT(hdr->b_datacnt == 1); 3216 hdr->b_flags |= ARC_BUF_AVAILABLE; 3217 } 3218 3219 ASSERT(refcount_is_zero(&hdr->b_refcnt) || callback_list != NULL); 3220 3221 if (zio->io_error != 0) { 3222 hdr->b_flags |= ARC_IO_ERROR; 3223 if (hdr->b_state != arc_anon) 3224 arc_change_state(arc_anon, hdr, hash_lock); 3225 if (HDR_IN_HASH_TABLE(hdr)) 3226 buf_hash_remove(hdr); 3227 freeable = refcount_is_zero(&hdr->b_refcnt); 3228 } 3229 3230 /* 3231 * Broadcast before we drop the hash_lock to avoid the possibility 3232 * that the hdr (and hence the cv) might be freed before we get to 3233 * the cv_broadcast(). 3234 */ 3235 cv_broadcast(&hdr->b_cv); 3236 3237 if (hash_lock) { 3238 mutex_exit(hash_lock); 3239 } else { 3240 /* 3241 * This block was freed while we waited for the read to 3242 * complete. It has been removed from the hash table and 3243 * moved to the anonymous state (so that it won't show up 3244 * in the cache). 3245 */ 3246 ASSERT3P(hdr->b_state, ==, arc_anon); 3247 freeable = refcount_is_zero(&hdr->b_refcnt); 3248 } 3249 3250 /* execute each callback and free its structure */ 3251 while ((acb = callback_list) != NULL) { 3252 if (acb->acb_done) 3253 acb->acb_done(zio, acb->acb_buf, acb->acb_private); 3254 3255 if (acb->acb_zio_dummy != NULL) { 3256 acb->acb_zio_dummy->io_error = zio->io_error; 3257 zio_nowait(acb->acb_zio_dummy); 3258 } 3259 3260 callback_list = acb->acb_next; 3261 kmem_free(acb, sizeof (arc_callback_t)); 3262 } 3263 3264 if (freeable) 3265 arc_hdr_destroy(hdr); 3266} 3267 3268/* 3269 * "Read" the block block at the specified DVA (in bp) via the 3270 * cache. If the block is found in the cache, invoke the provided 3271 * callback immediately and return. Note that the `zio' parameter 3272 * in the callback will be NULL in this case, since no IO was 3273 * required. If the block is not in the cache pass the read request 3274 * on to the spa with a substitute callback function, so that the 3275 * requested block will be added to the cache. 3276 * 3277 * If a read request arrives for a block that has a read in-progress, 3278 * either wait for the in-progress read to complete (and return the 3279 * results); or, if this is a read with a "done" func, add a record 3280 * to the read to invoke the "done" func when the read completes, 3281 * and return; or just return. 3282 * 3283 * arc_read_done() will invoke all the requested "done" functions 3284 * for readers of this block. 3285 */ 3286int 3287arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_done_func_t *done, 3288 void *private, zio_priority_t priority, int zio_flags, uint32_t *arc_flags, 3289 const zbookmark_phys_t *zb) 3290{ 3291 arc_buf_hdr_t *hdr = NULL; 3292 arc_buf_t *buf = NULL; 3293 kmutex_t *hash_lock = NULL; 3294 zio_t *rzio; 3295 uint64_t guid = spa_load_guid(spa); 3296 3297 ASSERT(!BP_IS_EMBEDDED(bp) || 3298 BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA); 3299 3300top: 3301 if (!BP_IS_EMBEDDED(bp)) { 3302 /* 3303 * Embedded BP's have no DVA and require no I/O to "read". 3304 * Create an anonymous arc buf to back it. 3305 */ 3306 hdr = buf_hash_find(guid, bp, &hash_lock); 3307 } 3308 3309 if (hdr != NULL && hdr->b_datacnt > 0) { 3310 3311 *arc_flags |= ARC_CACHED; 3312 3313 if (HDR_IO_IN_PROGRESS(hdr)) { 3314 3315 if (*arc_flags & ARC_WAIT) { 3316 cv_wait(&hdr->b_cv, hash_lock); 3317 mutex_exit(hash_lock); 3318 goto top; 3319 } 3320 ASSERT(*arc_flags & ARC_NOWAIT); 3321 3322 if (done) { 3323 arc_callback_t *acb = NULL; 3324 3325 acb = kmem_zalloc(sizeof (arc_callback_t), 3326 KM_SLEEP); 3327 acb->acb_done = done; 3328 acb->acb_private = private; 3329 if (pio != NULL) 3330 acb->acb_zio_dummy = zio_null(pio, 3331 spa, NULL, NULL, NULL, zio_flags); 3332 3333 ASSERT(acb->acb_done != NULL); 3334 acb->acb_next = hdr->b_acb; 3335 hdr->b_acb = acb; 3336 add_reference(hdr, hash_lock, private); 3337 mutex_exit(hash_lock); 3338 return (0); 3339 } 3340 mutex_exit(hash_lock); 3341 return (0); 3342 } 3343 3344 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu); 3345 3346 if (done) { 3347 add_reference(hdr, hash_lock, private); 3348 /* 3349 * If this block is already in use, create a new 3350 * copy of the data so that we will be guaranteed 3351 * that arc_release() will always succeed. 3352 */ 3353 buf = hdr->b_buf; 3354 ASSERT(buf); 3355 ASSERT(buf->b_data); 3356 if (HDR_BUF_AVAILABLE(hdr)) { 3357 ASSERT(buf->b_efunc == NULL); 3358 hdr->b_flags &= ~ARC_BUF_AVAILABLE; 3359 } else { 3360 buf = arc_buf_clone(buf); 3361 } 3362 3363 } else if (*arc_flags & ARC_PREFETCH && 3364 refcount_count(&hdr->b_refcnt) == 0) { 3365 hdr->b_flags |= ARC_PREFETCH; 3366 } 3367 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr); 3368 arc_access(hdr, hash_lock); 3369 if (*arc_flags & ARC_L2CACHE) 3370 hdr->b_flags |= ARC_L2CACHE; 3371 if (*arc_flags & ARC_L2COMPRESS) 3372 hdr->b_flags |= ARC_L2COMPRESS; 3373 mutex_exit(hash_lock); 3374 ARCSTAT_BUMP(arcstat_hits); 3375 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH), 3376 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA, 3377 data, metadata, hits); 3378 3379 if (done) 3380 done(NULL, buf, private); 3381 } else { 3382 uint64_t size = BP_GET_LSIZE(bp); 3383 arc_callback_t *acb; 3384 vdev_t *vd = NULL; 3385 uint64_t addr = 0; 3386 boolean_t devw = B_FALSE; 3387 enum zio_compress b_compress = ZIO_COMPRESS_OFF; 3388 uint64_t b_asize = 0; 3389 3390 if (hdr == NULL) { 3391 /* this block is not in the cache */ 3392 arc_buf_hdr_t *exists = NULL; 3393 arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp); 3394 buf = arc_buf_alloc(spa, size, private, type); 3395 hdr = buf->b_hdr; 3396 if (!BP_IS_EMBEDDED(bp)) { 3397 hdr->b_dva = *BP_IDENTITY(bp); 3398 hdr->b_birth = BP_PHYSICAL_BIRTH(bp); 3399 hdr->b_cksum0 = bp->blk_cksum.zc_word[0]; 3400 exists = buf_hash_insert(hdr, &hash_lock); 3401 } 3402 if (exists != NULL) { 3403 /* somebody beat us to the hash insert */ 3404 mutex_exit(hash_lock); 3405 buf_discard_identity(hdr); 3406 (void) arc_buf_remove_ref(buf, private); 3407 goto top; /* restart the IO request */ 3408 } 3409 /* if this is a prefetch, we don't have a reference */ 3410 if (*arc_flags & ARC_PREFETCH) { 3411 (void) remove_reference(hdr, hash_lock, 3412 private); 3413 hdr->b_flags |= ARC_PREFETCH; 3414 } 3415 if (*arc_flags & ARC_L2CACHE) 3416 hdr->b_flags |= ARC_L2CACHE; 3417 if (*arc_flags & ARC_L2COMPRESS) 3418 hdr->b_flags |= ARC_L2COMPRESS; 3419 if (BP_GET_LEVEL(bp) > 0) 3420 hdr->b_flags |= ARC_INDIRECT; 3421 } else { 3422 /* this block is in the ghost cache */ 3423 ASSERT(GHOST_STATE(hdr->b_state)); 3424 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 3425 ASSERT0(refcount_count(&hdr->b_refcnt)); 3426 ASSERT(hdr->b_buf == NULL); 3427 3428 /* if this is a prefetch, we don't have a reference */ 3429 if (*arc_flags & ARC_PREFETCH) 3430 hdr->b_flags |= ARC_PREFETCH; 3431 else 3432 add_reference(hdr, hash_lock, private); 3433 if (*arc_flags & ARC_L2CACHE) 3434 hdr->b_flags |= ARC_L2CACHE; 3435 if (*arc_flags & ARC_L2COMPRESS) 3436 hdr->b_flags |= ARC_L2COMPRESS; 3437 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 3438 buf->b_hdr = hdr; 3439 buf->b_data = NULL; 3440 buf->b_efunc = NULL; 3441 buf->b_private = NULL; 3442 buf->b_next = NULL; 3443 hdr->b_buf = buf; 3444 ASSERT(hdr->b_datacnt == 0); 3445 hdr->b_datacnt = 1; 3446 arc_get_data_buf(buf); 3447 arc_access(hdr, hash_lock); 3448 } 3449 3450 ASSERT(!GHOST_STATE(hdr->b_state)); 3451 3452 acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP); 3453 acb->acb_done = done; 3454 acb->acb_private = private; 3455 3456 ASSERT(hdr->b_acb == NULL); 3457 hdr->b_acb = acb; 3458 hdr->b_flags |= ARC_IO_IN_PROGRESS; 3459 3460 if (hdr->b_l2hdr != NULL && 3461 (vd = hdr->b_l2hdr->b_dev->l2ad_vdev) != NULL) { 3462 devw = hdr->b_l2hdr->b_dev->l2ad_writing; 3463 addr = hdr->b_l2hdr->b_daddr; 3464 b_compress = hdr->b_l2hdr->b_compress; 3465 b_asize = hdr->b_l2hdr->b_asize; 3466 /* 3467 * Lock out device removal. 3468 */ 3469 if (vdev_is_dead(vd) || 3470 !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER)) 3471 vd = NULL; 3472 } 3473 3474 if (hash_lock != NULL) 3475 mutex_exit(hash_lock); 3476 3477 /* 3478 * At this point, we have a level 1 cache miss. Try again in 3479 * L2ARC if possible. 3480 */ 3481 ASSERT3U(hdr->b_size, ==, size); 3482 DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp, 3483 uint64_t, size, zbookmark_phys_t *, zb); 3484 ARCSTAT_BUMP(arcstat_misses); 3485 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH), 3486 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA, 3487 data, metadata, misses); 3488#ifdef _KERNEL 3489 curthread->td_ru.ru_inblock++; 3490#endif 3491 3492 if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) { 3493 /* 3494 * Read from the L2ARC if the following are true: 3495 * 1. The L2ARC vdev was previously cached. 3496 * 2. This buffer still has L2ARC metadata. 3497 * 3. This buffer isn't currently writing to the L2ARC. 3498 * 4. The L2ARC entry wasn't evicted, which may 3499 * also have invalidated the vdev. 3500 * 5. This isn't prefetch and l2arc_noprefetch is set. 3501 */ 3502 if (hdr->b_l2hdr != NULL && 3503 !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) && 3504 !(l2arc_noprefetch && HDR_PREFETCH(hdr))) { 3505 l2arc_read_callback_t *cb; 3506 3507 DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr); 3508 ARCSTAT_BUMP(arcstat_l2_hits); 3509 3510 cb = kmem_zalloc(sizeof (l2arc_read_callback_t), 3511 KM_SLEEP); 3512 cb->l2rcb_buf = buf; 3513 cb->l2rcb_spa = spa; 3514 cb->l2rcb_bp = *bp; 3515 cb->l2rcb_zb = *zb; 3516 cb->l2rcb_flags = zio_flags; 3517 cb->l2rcb_compress = b_compress; 3518 3519 ASSERT(addr >= VDEV_LABEL_START_SIZE && 3520 addr + size < vd->vdev_psize - 3521 VDEV_LABEL_END_SIZE); 3522 3523 /* 3524 * l2arc read. The SCL_L2ARC lock will be 3525 * released by l2arc_read_done(). 3526 * Issue a null zio if the underlying buffer 3527 * was squashed to zero size by compression. 3528 */ 3529 if (b_compress == ZIO_COMPRESS_EMPTY) { 3530 rzio = zio_null(pio, spa, vd, 3531 l2arc_read_done, cb, 3532 zio_flags | ZIO_FLAG_DONT_CACHE | 3533 ZIO_FLAG_CANFAIL | 3534 ZIO_FLAG_DONT_PROPAGATE | 3535 ZIO_FLAG_DONT_RETRY); 3536 } else { 3537 rzio = zio_read_phys(pio, vd, addr, 3538 b_asize, buf->b_data, 3539 ZIO_CHECKSUM_OFF, 3540 l2arc_read_done, cb, priority, 3541 zio_flags | ZIO_FLAG_DONT_CACHE | 3542 ZIO_FLAG_CANFAIL | 3543 ZIO_FLAG_DONT_PROPAGATE | 3544 ZIO_FLAG_DONT_RETRY, B_FALSE); 3545 } 3546 DTRACE_PROBE2(l2arc__read, vdev_t *, vd, 3547 zio_t *, rzio); 3548 ARCSTAT_INCR(arcstat_l2_read_bytes, b_asize); 3549 3550 if (*arc_flags & ARC_NOWAIT) { 3551 zio_nowait(rzio); 3552 return (0); 3553 } 3554 3555 ASSERT(*arc_flags & ARC_WAIT); 3556 if (zio_wait(rzio) == 0) 3557 return (0); 3558 3559 /* l2arc read error; goto zio_read() */ 3560 } else { 3561 DTRACE_PROBE1(l2arc__miss, 3562 arc_buf_hdr_t *, hdr); 3563 ARCSTAT_BUMP(arcstat_l2_misses); 3564 if (HDR_L2_WRITING(hdr)) 3565 ARCSTAT_BUMP(arcstat_l2_rw_clash); 3566 spa_config_exit(spa, SCL_L2ARC, vd); 3567 } 3568 } else { 3569 if (vd != NULL) 3570 spa_config_exit(spa, SCL_L2ARC, vd); 3571 if (l2arc_ndev != 0) { 3572 DTRACE_PROBE1(l2arc__miss, 3573 arc_buf_hdr_t *, hdr); 3574 ARCSTAT_BUMP(arcstat_l2_misses); 3575 } 3576 } 3577 3578 rzio = zio_read(pio, spa, bp, buf->b_data, size, 3579 arc_read_done, buf, priority, zio_flags, zb); 3580 3581 if (*arc_flags & ARC_WAIT) 3582 return (zio_wait(rzio)); 3583 3584 ASSERT(*arc_flags & ARC_NOWAIT); 3585 zio_nowait(rzio); 3586 } 3587 return (0); 3588} 3589 3590void 3591arc_set_callback(arc_buf_t *buf, arc_evict_func_t *func, void *private) 3592{ 3593 ASSERT(buf->b_hdr != NULL); 3594 ASSERT(buf->b_hdr->b_state != arc_anon); 3595 ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt) || func == NULL); 3596 ASSERT(buf->b_efunc == NULL); 3597 ASSERT(!HDR_BUF_AVAILABLE(buf->b_hdr)); 3598 3599 buf->b_efunc = func; 3600 buf->b_private = private; 3601} 3602 3603/* 3604 * Notify the arc that a block was freed, and thus will never be used again. 3605 */ 3606void 3607arc_freed(spa_t *spa, const blkptr_t *bp) 3608{ 3609 arc_buf_hdr_t *hdr; 3610 kmutex_t *hash_lock; 3611 uint64_t guid = spa_load_guid(spa); 3612 3613 ASSERT(!BP_IS_EMBEDDED(bp)); 3614 3615 hdr = buf_hash_find(guid, bp, &hash_lock); 3616 if (hdr == NULL) 3617 return; 3618 if (HDR_BUF_AVAILABLE(hdr)) { 3619 arc_buf_t *buf = hdr->b_buf; 3620 add_reference(hdr, hash_lock, FTAG); 3621 hdr->b_flags &= ~ARC_BUF_AVAILABLE; 3622 mutex_exit(hash_lock); 3623 3624 arc_release(buf, FTAG); 3625 (void) arc_buf_remove_ref(buf, FTAG); 3626 } else { 3627 mutex_exit(hash_lock); 3628 } 3629 3630} 3631 3632/* 3633 * Clear the user eviction callback set by arc_set_callback(), first calling 3634 * it if it exists. Because the presence of a callback keeps an arc_buf cached 3635 * clearing the callback may result in the arc_buf being destroyed. However, 3636 * it will not result in the *last* arc_buf being destroyed, hence the data 3637 * will remain cached in the ARC. We make a copy of the arc buffer here so 3638 * that we can process the callback without holding any locks. 3639 * 3640 * It's possible that the callback is already in the process of being cleared 3641 * by another thread. In this case we can not clear the callback. 3642 * 3643 * Returns B_TRUE if the callback was successfully called and cleared. 3644 */ 3645boolean_t 3646arc_clear_callback(arc_buf_t *buf) 3647{ 3648 arc_buf_hdr_t *hdr; 3649 kmutex_t *hash_lock; 3650 arc_evict_func_t *efunc = buf->b_efunc; 3651 void *private = buf->b_private; 3652 list_t *list, *evicted_list; 3653 kmutex_t *lock, *evicted_lock; 3654 3655 mutex_enter(&buf->b_evict_lock); 3656 hdr = buf->b_hdr; 3657 if (hdr == NULL) { 3658 /* 3659 * We are in arc_do_user_evicts(). 3660 */ 3661 ASSERT(buf->b_data == NULL); 3662 mutex_exit(&buf->b_evict_lock); 3663 return (B_FALSE); 3664 } else if (buf->b_data == NULL) { 3665 /* 3666 * We are on the eviction list; process this buffer now 3667 * but let arc_do_user_evicts() do the reaping. 3668 */ 3669 buf->b_efunc = NULL; 3670 mutex_exit(&buf->b_evict_lock); 3671 VERIFY0(efunc(private)); 3672 return (B_TRUE); 3673 } 3674 hash_lock = HDR_LOCK(hdr); 3675 mutex_enter(hash_lock); 3676 hdr = buf->b_hdr; 3677 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 3678 3679 ASSERT3U(refcount_count(&hdr->b_refcnt), <, hdr->b_datacnt); 3680 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu); 3681 3682 buf->b_efunc = NULL; 3683 buf->b_private = NULL; 3684 3685 if (hdr->b_datacnt > 1) { 3686 mutex_exit(&buf->b_evict_lock); 3687 arc_buf_destroy(buf, FALSE, TRUE); 3688 } else { 3689 ASSERT(buf == hdr->b_buf); 3690 hdr->b_flags |= ARC_BUF_AVAILABLE; 3691 mutex_exit(&buf->b_evict_lock); 3692 } 3693 3694 mutex_exit(hash_lock); 3695 VERIFY0(efunc(private)); 3696 return (B_TRUE); 3697} 3698 3699/* 3700 * Release this buffer from the cache, making it an anonymous buffer. This 3701 * must be done after a read and prior to modifying the buffer contents. 3702 * If the buffer has more than one reference, we must make 3703 * a new hdr for the buffer. 3704 */ 3705void 3706arc_release(arc_buf_t *buf, void *tag) 3707{ 3708 arc_buf_hdr_t *hdr; 3709 kmutex_t *hash_lock = NULL; 3710 l2arc_buf_hdr_t *l2hdr; 3711 uint64_t buf_size; 3712 3713 /* 3714 * It would be nice to assert that if it's DMU metadata (level > 3715 * 0 || it's the dnode file), then it must be syncing context. 3716 * But we don't know that information at this level. 3717 */ 3718 3719 mutex_enter(&buf->b_evict_lock); 3720 hdr = buf->b_hdr; 3721 3722 /* this buffer is not on any list */ 3723 ASSERT(refcount_count(&hdr->b_refcnt) > 0); 3724 3725 if (hdr->b_state == arc_anon) { 3726 /* this buffer is already released */ 3727 ASSERT(buf->b_efunc == NULL); 3728 } else { 3729 hash_lock = HDR_LOCK(hdr); 3730 mutex_enter(hash_lock); 3731 hdr = buf->b_hdr; 3732 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 3733 } 3734 3735 l2hdr = hdr->b_l2hdr; 3736 if (l2hdr) { 3737 mutex_enter(&l2arc_buflist_mtx); 3738 arc_buf_l2_cdata_free(hdr); 3739 hdr->b_l2hdr = NULL; 3740 list_remove(l2hdr->b_dev->l2ad_buflist, hdr); 3741 } 3742 buf_size = hdr->b_size; 3743 3744 /* 3745 * Do we have more than one buf? 3746 */ 3747 if (hdr->b_datacnt > 1) { 3748 arc_buf_hdr_t *nhdr; 3749 arc_buf_t **bufp; 3750 uint64_t blksz = hdr->b_size; 3751 uint64_t spa = hdr->b_spa; 3752 arc_buf_contents_t type = hdr->b_type; 3753 uint32_t flags = hdr->b_flags; 3754 3755 ASSERT(hdr->b_buf != buf || buf->b_next != NULL); 3756 /* 3757 * Pull the data off of this hdr and attach it to 3758 * a new anonymous hdr. 3759 */ 3760 (void) remove_reference(hdr, hash_lock, tag); 3761 bufp = &hdr->b_buf; 3762 while (*bufp != buf) 3763 bufp = &(*bufp)->b_next; 3764 *bufp = buf->b_next; 3765 buf->b_next = NULL; 3766 3767 ASSERT3U(hdr->b_state->arcs_size, >=, hdr->b_size); 3768 atomic_add_64(&hdr->b_state->arcs_size, -hdr->b_size); 3769 if (refcount_is_zero(&hdr->b_refcnt)) { 3770 uint64_t *size = &hdr->b_state->arcs_lsize[hdr->b_type]; 3771 ASSERT3U(*size, >=, hdr->b_size); 3772 atomic_add_64(size, -hdr->b_size); 3773 } 3774 3775 /* 3776 * We're releasing a duplicate user data buffer, update 3777 * our statistics accordingly. 3778 */ 3779 if (hdr->b_type == ARC_BUFC_DATA) { 3780 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers); 3781 ARCSTAT_INCR(arcstat_duplicate_buffers_size, 3782 -hdr->b_size); 3783 } 3784 hdr->b_datacnt -= 1; 3785 arc_cksum_verify(buf); 3786#ifdef illumos 3787 arc_buf_unwatch(buf); 3788#endif /* illumos */ 3789 3790 mutex_exit(hash_lock); 3791 3792 nhdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE); 3793 nhdr->b_size = blksz; 3794 nhdr->b_spa = spa; 3795 nhdr->b_type = type; 3796 nhdr->b_buf = buf; 3797 nhdr->b_state = arc_anon; 3798 nhdr->b_arc_access = 0; 3799 nhdr->b_flags = flags & ARC_L2_WRITING; 3800 nhdr->b_l2hdr = NULL; 3801 nhdr->b_datacnt = 1; 3802 nhdr->b_freeze_cksum = NULL; 3803 (void) refcount_add(&nhdr->b_refcnt, tag); 3804 buf->b_hdr = nhdr; 3805 mutex_exit(&buf->b_evict_lock); 3806 atomic_add_64(&arc_anon->arcs_size, blksz); 3807 } else { 3808 mutex_exit(&buf->b_evict_lock); 3809 ASSERT(refcount_count(&hdr->b_refcnt) == 1); 3810 ASSERT(!list_link_active(&hdr->b_arc_node)); 3811 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 3812 if (hdr->b_state != arc_anon) 3813 arc_change_state(arc_anon, hdr, hash_lock); 3814 hdr->b_arc_access = 0; 3815 if (hash_lock) 3816 mutex_exit(hash_lock); 3817 3818 buf_discard_identity(hdr); 3819 arc_buf_thaw(buf); 3820 } 3821 buf->b_efunc = NULL; 3822 buf->b_private = NULL; 3823 3824 if (l2hdr) { 3825 ARCSTAT_INCR(arcstat_l2_asize, -l2hdr->b_asize); 3826 vdev_space_update(l2hdr->b_dev->l2ad_vdev, 3827 -l2hdr->b_asize, 0, 0); 3828 trim_map_free(l2hdr->b_dev->l2ad_vdev, l2hdr->b_daddr, 3829 hdr->b_size, 0); 3830 kmem_free(l2hdr, sizeof (l2arc_buf_hdr_t)); 3831 ARCSTAT_INCR(arcstat_l2_size, -buf_size); 3832 mutex_exit(&l2arc_buflist_mtx); 3833 } 3834} 3835 3836int 3837arc_released(arc_buf_t *buf) 3838{ 3839 int released; 3840 3841 mutex_enter(&buf->b_evict_lock); 3842 released = (buf->b_data != NULL && buf->b_hdr->b_state == arc_anon); 3843 mutex_exit(&buf->b_evict_lock); 3844 return (released); 3845} 3846 3847#ifdef ZFS_DEBUG 3848int 3849arc_referenced(arc_buf_t *buf) 3850{ 3851 int referenced; 3852 3853 mutex_enter(&buf->b_evict_lock); 3854 referenced = (refcount_count(&buf->b_hdr->b_refcnt)); 3855 mutex_exit(&buf->b_evict_lock); 3856 return (referenced); 3857} 3858#endif 3859 3860static void 3861arc_write_ready(zio_t *zio) 3862{ 3863 arc_write_callback_t *callback = zio->io_private; 3864 arc_buf_t *buf = callback->awcb_buf; 3865 arc_buf_hdr_t *hdr = buf->b_hdr; 3866 3867 ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt)); 3868 callback->awcb_ready(zio, buf, callback->awcb_private); 3869 3870 /* 3871 * If the IO is already in progress, then this is a re-write 3872 * attempt, so we need to thaw and re-compute the cksum. 3873 * It is the responsibility of the callback to handle the 3874 * accounting for any re-write attempt. 3875 */ 3876 if (HDR_IO_IN_PROGRESS(hdr)) { 3877 mutex_enter(&hdr->b_freeze_lock); 3878 if (hdr->b_freeze_cksum != NULL) { 3879 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 3880 hdr->b_freeze_cksum = NULL; 3881 } 3882 mutex_exit(&hdr->b_freeze_lock); 3883 } 3884 arc_cksum_compute(buf, B_FALSE); 3885 hdr->b_flags |= ARC_IO_IN_PROGRESS; 3886} 3887 3888/* 3889 * The SPA calls this callback for each physical write that happens on behalf 3890 * of a logical write. See the comment in dbuf_write_physdone() for details. 3891 */ 3892static void 3893arc_write_physdone(zio_t *zio) 3894{ 3895 arc_write_callback_t *cb = zio->io_private; 3896 if (cb->awcb_physdone != NULL) 3897 cb->awcb_physdone(zio, cb->awcb_buf, cb->awcb_private); 3898} 3899 3900static void 3901arc_write_done(zio_t *zio) 3902{ 3903 arc_write_callback_t *callback = zio->io_private; 3904 arc_buf_t *buf = callback->awcb_buf; 3905 arc_buf_hdr_t *hdr = buf->b_hdr; 3906 3907 ASSERT(hdr->b_acb == NULL); 3908 3909 if (zio->io_error == 0) { 3910 if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) { 3911 buf_discard_identity(hdr); 3912 } else { 3913 hdr->b_dva = *BP_IDENTITY(zio->io_bp); 3914 hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp); 3915 hdr->b_cksum0 = zio->io_bp->blk_cksum.zc_word[0]; 3916 } 3917 } else { 3918 ASSERT(BUF_EMPTY(hdr)); 3919 } 3920 3921 /* 3922 * If the block to be written was all-zero or compressed enough to be 3923 * embedded in the BP, no write was performed so there will be no 3924 * dva/birth/checksum. The buffer must therefore remain anonymous 3925 * (and uncached). 3926 */ 3927 if (!BUF_EMPTY(hdr)) { 3928 arc_buf_hdr_t *exists; 3929 kmutex_t *hash_lock; 3930 3931 ASSERT(zio->io_error == 0); 3932 3933 arc_cksum_verify(buf); 3934 3935 exists = buf_hash_insert(hdr, &hash_lock); 3936 if (exists) { 3937 /* 3938 * This can only happen if we overwrite for 3939 * sync-to-convergence, because we remove 3940 * buffers from the hash table when we arc_free(). 3941 */ 3942 if (zio->io_flags & ZIO_FLAG_IO_REWRITE) { 3943 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) 3944 panic("bad overwrite, hdr=%p exists=%p", 3945 (void *)hdr, (void *)exists); 3946 ASSERT(refcount_is_zero(&exists->b_refcnt)); 3947 arc_change_state(arc_anon, exists, hash_lock); 3948 mutex_exit(hash_lock); 3949 arc_hdr_destroy(exists); 3950 exists = buf_hash_insert(hdr, &hash_lock); 3951 ASSERT3P(exists, ==, NULL); 3952 } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) { 3953 /* nopwrite */ 3954 ASSERT(zio->io_prop.zp_nopwrite); 3955 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) 3956 panic("bad nopwrite, hdr=%p exists=%p", 3957 (void *)hdr, (void *)exists); 3958 } else { 3959 /* Dedup */ 3960 ASSERT(hdr->b_datacnt == 1); 3961 ASSERT(hdr->b_state == arc_anon); 3962 ASSERT(BP_GET_DEDUP(zio->io_bp)); 3963 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0); 3964 } 3965 } 3966 hdr->b_flags &= ~ARC_IO_IN_PROGRESS; 3967 /* if it's not anon, we are doing a scrub */ 3968 if (!exists && hdr->b_state == arc_anon) 3969 arc_access(hdr, hash_lock); 3970 mutex_exit(hash_lock); 3971 } else { 3972 hdr->b_flags &= ~ARC_IO_IN_PROGRESS; 3973 } 3974 3975 ASSERT(!refcount_is_zero(&hdr->b_refcnt)); 3976 callback->awcb_done(zio, buf, callback->awcb_private); 3977 3978 kmem_free(callback, sizeof (arc_write_callback_t)); 3979} 3980 3981zio_t * 3982arc_write(zio_t *pio, spa_t *spa, uint64_t txg, 3983 blkptr_t *bp, arc_buf_t *buf, boolean_t l2arc, boolean_t l2arc_compress, 3984 const zio_prop_t *zp, arc_done_func_t *ready, arc_done_func_t *physdone, 3985 arc_done_func_t *done, void *private, zio_priority_t priority, 3986 int zio_flags, const zbookmark_phys_t *zb) 3987{ 3988 arc_buf_hdr_t *hdr = buf->b_hdr; 3989 arc_write_callback_t *callback; 3990 zio_t *zio; 3991 3992 ASSERT(ready != NULL); 3993 ASSERT(done != NULL); 3994 ASSERT(!HDR_IO_ERROR(hdr)); 3995 ASSERT((hdr->b_flags & ARC_IO_IN_PROGRESS) == 0); 3996 ASSERT(hdr->b_acb == NULL); 3997 if (l2arc) 3998 hdr->b_flags |= ARC_L2CACHE; 3999 if (l2arc_compress) 4000 hdr->b_flags |= ARC_L2COMPRESS; 4001 callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP); 4002 callback->awcb_ready = ready; 4003 callback->awcb_physdone = physdone; 4004 callback->awcb_done = done; 4005 callback->awcb_private = private; 4006 callback->awcb_buf = buf; 4007 4008 zio = zio_write(pio, spa, txg, bp, buf->b_data, hdr->b_size, zp, 4009 arc_write_ready, arc_write_physdone, arc_write_done, callback, 4010 priority, zio_flags, zb); 4011 4012 return (zio); 4013} 4014 4015static int 4016arc_memory_throttle(uint64_t reserve, uint64_t txg) 4017{ 4018#ifdef _KERNEL 4019 uint64_t available_memory = ptob(freemem); 4020 static uint64_t page_load = 0; 4021 static uint64_t last_txg = 0; 4022 4023#if defined(__i386) || !defined(UMA_MD_SMALL_ALLOC) 4024 available_memory = 4025 MIN(available_memory, ptob(vmem_size(heap_arena, VMEM_FREE))); 4026#endif 4027 4028 if (freemem > (uint64_t)physmem * arc_lotsfree_percent / 100) 4029 return (0); 4030 4031 if (txg > last_txg) { 4032 last_txg = txg; 4033 page_load = 0; 4034 } 4035 /* 4036 * If we are in pageout, we know that memory is already tight, 4037 * the arc is already going to be evicting, so we just want to 4038 * continue to let page writes occur as quickly as possible. 4039 */ 4040 if (curproc == pageproc) { 4041 if (page_load > MAX(ptob(minfree), available_memory) / 4) 4042 return (SET_ERROR(ERESTART)); 4043 /* Note: reserve is inflated, so we deflate */ 4044 page_load += reserve / 8; 4045 return (0); 4046 } else if (page_load > 0 && arc_reclaim_needed()) { 4047 /* memory is low, delay before restarting */ 4048 ARCSTAT_INCR(arcstat_memory_throttle_count, 1); 4049 return (SET_ERROR(EAGAIN)); 4050 } 4051 page_load = 0; 4052#endif 4053 return (0); 4054} 4055 4056void 4057arc_tempreserve_clear(uint64_t reserve) 4058{ 4059 atomic_add_64(&arc_tempreserve, -reserve); 4060 ASSERT((int64_t)arc_tempreserve >= 0); 4061} 4062 4063int 4064arc_tempreserve_space(uint64_t reserve, uint64_t txg) 4065{ 4066 int error; 4067 uint64_t anon_size; 4068 4069 if (reserve > arc_c/4 && !arc_no_grow) { 4070 arc_c = MIN(arc_c_max, reserve * 4); 4071 DTRACE_PROBE1(arc__set_reserve, uint64_t, arc_c); 4072 } 4073 if (reserve > arc_c) 4074 return (SET_ERROR(ENOMEM)); 4075 4076 /* 4077 * Don't count loaned bufs as in flight dirty data to prevent long 4078 * network delays from blocking transactions that are ready to be 4079 * assigned to a txg. 4080 */ 4081 anon_size = MAX((int64_t)(arc_anon->arcs_size - arc_loaned_bytes), 0); 4082 4083 /* 4084 * Writes will, almost always, require additional memory allocations 4085 * in order to compress/encrypt/etc the data. We therefore need to 4086 * make sure that there is sufficient available memory for this. 4087 */ 4088 error = arc_memory_throttle(reserve, txg); 4089 if (error != 0) 4090 return (error); 4091 4092 /* 4093 * Throttle writes when the amount of dirty data in the cache 4094 * gets too large. We try to keep the cache less than half full 4095 * of dirty blocks so that our sync times don't grow too large. 4096 * Note: if two requests come in concurrently, we might let them 4097 * both succeed, when one of them should fail. Not a huge deal. 4098 */ 4099 4100 if (reserve + arc_tempreserve + anon_size > arc_c / 2 && 4101 anon_size > arc_c / 4) { 4102 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK " 4103 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n", 4104 arc_tempreserve>>10, 4105 arc_anon->arcs_lsize[ARC_BUFC_METADATA]>>10, 4106 arc_anon->arcs_lsize[ARC_BUFC_DATA]>>10, 4107 reserve>>10, arc_c>>10); 4108 return (SET_ERROR(ERESTART)); 4109 } 4110 atomic_add_64(&arc_tempreserve, reserve); 4111 return (0); 4112} 4113 4114static kmutex_t arc_lowmem_lock; 4115#ifdef _KERNEL 4116static eventhandler_tag arc_event_lowmem = NULL; 4117 4118static void 4119arc_lowmem(void *arg __unused, int howto __unused) 4120{ 4121 4122 /* Serialize access via arc_lowmem_lock. */ 4123 mutex_enter(&arc_lowmem_lock); 4124 mutex_enter(&arc_reclaim_thr_lock); 4125 needfree = 1; 4126 DTRACE_PROBE(arc__needfree); 4127 cv_signal(&arc_reclaim_thr_cv); 4128 4129 /* 4130 * It is unsafe to block here in arbitrary threads, because we can come 4131 * here from ARC itself and may hold ARC locks and thus risk a deadlock 4132 * with ARC reclaim thread. 4133 */ 4134 if (curproc == pageproc) { 4135 while (needfree) 4136 msleep(&needfree, &arc_reclaim_thr_lock, 0, "zfs:lowmem", 0); 4137 } 4138 mutex_exit(&arc_reclaim_thr_lock); 4139 mutex_exit(&arc_lowmem_lock); 4140} 4141#endif 4142 4143void 4144arc_init(void) 4145{ 4146 int i, prefetch_tunable_set = 0; 4147 4148 mutex_init(&arc_reclaim_thr_lock, NULL, MUTEX_DEFAULT, NULL); 4149 cv_init(&arc_reclaim_thr_cv, NULL, CV_DEFAULT, NULL); 4150 mutex_init(&arc_lowmem_lock, NULL, MUTEX_DEFAULT, NULL); 4151 4152 /* Convert seconds to clock ticks */ 4153 arc_min_prefetch_lifespan = 1 * hz; 4154 4155 /* Start out with 1/8 of all memory */ 4156 arc_c = kmem_size() / 8; 4157 4158#ifdef sun 4159#ifdef _KERNEL 4160 /* 4161 * On architectures where the physical memory can be larger 4162 * than the addressable space (intel in 32-bit mode), we may 4163 * need to limit the cache to 1/8 of VM size. 4164 */ 4165 arc_c = MIN(arc_c, vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 8); 4166#endif 4167#endif /* sun */ 4168 /* set min cache to 1/32 of all memory, or 16MB, whichever is more */ 4169 arc_c_min = MAX(arc_c / 4, 64<<18); 4170 /* set max to 1/2 of all memory, or all but 1GB, whichever is more */ 4171 if (arc_c * 8 >= 1<<30) 4172 arc_c_max = (arc_c * 8) - (1<<30); 4173 else 4174 arc_c_max = arc_c_min; 4175 arc_c_max = MAX(arc_c * 5, arc_c_max); 4176 4177#ifdef _KERNEL 4178 /* 4179 * Allow the tunables to override our calculations if they are 4180 * reasonable (ie. over 16MB) 4181 */ 4182 if (zfs_arc_max > 64<<18 && zfs_arc_max < kmem_size()) 4183 arc_c_max = zfs_arc_max; 4184 if (zfs_arc_min > 64<<18 && zfs_arc_min <= arc_c_max) 4185 arc_c_min = zfs_arc_min; 4186#endif 4187 4188 arc_c = arc_c_max; 4189 arc_p = (arc_c >> 1); 4190 4191 /* limit meta-data to 1/4 of the arc capacity */ 4192 arc_meta_limit = arc_c_max / 4; 4193 4194 /* Allow the tunable to override if it is reasonable */ 4195 if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max) 4196 arc_meta_limit = zfs_arc_meta_limit; 4197 4198 if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0) 4199 arc_c_min = arc_meta_limit / 2; 4200 4201 if (zfs_arc_grow_retry > 0) 4202 arc_grow_retry = zfs_arc_grow_retry; 4203 4204 if (zfs_arc_shrink_shift > 0) 4205 arc_shrink_shift = zfs_arc_shrink_shift; 4206 4207 if (zfs_arc_p_min_shift > 0) 4208 arc_p_min_shift = zfs_arc_p_min_shift; 4209 4210 /* if kmem_flags are set, lets try to use less memory */ 4211 if (kmem_debugging()) 4212 arc_c = arc_c / 2; 4213 if (arc_c < arc_c_min) 4214 arc_c = arc_c_min; 4215 4216 zfs_arc_min = arc_c_min; 4217 zfs_arc_max = arc_c_max; 4218 4219 arc_anon = &ARC_anon; 4220 arc_mru = &ARC_mru; 4221 arc_mru_ghost = &ARC_mru_ghost; 4222 arc_mfu = &ARC_mfu; 4223 arc_mfu_ghost = &ARC_mfu_ghost; 4224 arc_l2c_only = &ARC_l2c_only; 4225 arc_size = 0; 4226 4227 for (i = 0; i < ARC_BUFC_NUMLISTS; i++) { 4228 mutex_init(&arc_anon->arcs_locks[i].arcs_lock, 4229 NULL, MUTEX_DEFAULT, NULL); 4230 mutex_init(&arc_mru->arcs_locks[i].arcs_lock, 4231 NULL, MUTEX_DEFAULT, NULL); 4232 mutex_init(&arc_mru_ghost->arcs_locks[i].arcs_lock, 4233 NULL, MUTEX_DEFAULT, NULL); 4234 mutex_init(&arc_mfu->arcs_locks[i].arcs_lock, 4235 NULL, MUTEX_DEFAULT, NULL); 4236 mutex_init(&arc_mfu_ghost->arcs_locks[i].arcs_lock, 4237 NULL, MUTEX_DEFAULT, NULL); 4238 mutex_init(&arc_l2c_only->arcs_locks[i].arcs_lock, 4239 NULL, MUTEX_DEFAULT, NULL); 4240 4241 list_create(&arc_mru->arcs_lists[i], 4242 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4243 list_create(&arc_mru_ghost->arcs_lists[i], 4244 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4245 list_create(&arc_mfu->arcs_lists[i], 4246 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4247 list_create(&arc_mfu_ghost->arcs_lists[i], 4248 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4249 list_create(&arc_mfu_ghost->arcs_lists[i], 4250 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4251 list_create(&arc_l2c_only->arcs_lists[i], 4252 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4253 } 4254 4255 buf_init(); 4256 4257 arc_thread_exit = 0; 4258 arc_eviction_list = NULL; 4259 mutex_init(&arc_eviction_mtx, NULL, MUTEX_DEFAULT, NULL); 4260 bzero(&arc_eviction_hdr, sizeof (arc_buf_hdr_t)); 4261 4262 arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED, 4263 sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL); 4264 4265 if (arc_ksp != NULL) { 4266 arc_ksp->ks_data = &arc_stats; 4267 kstat_install(arc_ksp); 4268 } 4269 4270 (void) thread_create(NULL, 0, arc_reclaim_thread, NULL, 0, &p0, 4271 TS_RUN, minclsyspri); 4272 4273#ifdef _KERNEL 4274 arc_event_lowmem = EVENTHANDLER_REGISTER(vm_lowmem, arc_lowmem, NULL, 4275 EVENTHANDLER_PRI_FIRST); 4276#endif 4277 4278 arc_dead = FALSE; 4279 arc_warm = B_FALSE; 4280 4281 /* 4282 * Calculate maximum amount of dirty data per pool. 4283 * 4284 * If it has been set by /etc/system, take that. 4285 * Otherwise, use a percentage of physical memory defined by 4286 * zfs_dirty_data_max_percent (default 10%) with a cap at 4287 * zfs_dirty_data_max_max (default 4GB). 4288 */ 4289 if (zfs_dirty_data_max == 0) { 4290 zfs_dirty_data_max = ptob(physmem) * 4291 zfs_dirty_data_max_percent / 100; 4292 zfs_dirty_data_max = MIN(zfs_dirty_data_max, 4293 zfs_dirty_data_max_max); 4294 } 4295 4296#ifdef _KERNEL 4297 if (TUNABLE_INT_FETCH("vfs.zfs.prefetch_disable", &zfs_prefetch_disable)) 4298 prefetch_tunable_set = 1; 4299 4300#ifdef __i386__ 4301 if (prefetch_tunable_set == 0) { 4302 printf("ZFS NOTICE: Prefetch is disabled by default on i386 " 4303 "-- to enable,\n"); 4304 printf(" add \"vfs.zfs.prefetch_disable=0\" " 4305 "to /boot/loader.conf.\n"); 4306 zfs_prefetch_disable = 1; 4307 } 4308#else 4309 if ((((uint64_t)physmem * PAGESIZE) < (1ULL << 32)) && 4310 prefetch_tunable_set == 0) { 4311 printf("ZFS NOTICE: Prefetch is disabled by default if less " 4312 "than 4GB of RAM is present;\n" 4313 " to enable, add \"vfs.zfs.prefetch_disable=0\" " 4314 "to /boot/loader.conf.\n"); 4315 zfs_prefetch_disable = 1; 4316 } 4317#endif 4318 /* Warn about ZFS memory and address space requirements. */ 4319 if (((uint64_t)physmem * PAGESIZE) < (256 + 128 + 64) * (1 << 20)) { 4320 printf("ZFS WARNING: Recommended minimum RAM size is 512MB; " 4321 "expect unstable behavior.\n"); 4322 } 4323 if (kmem_size() < 512 * (1 << 20)) { 4324 printf("ZFS WARNING: Recommended minimum kmem_size is 512MB; " 4325 "expect unstable behavior.\n"); 4326 printf(" Consider tuning vm.kmem_size and " 4327 "vm.kmem_size_max\n"); 4328 printf(" in /boot/loader.conf.\n"); 4329 } 4330#endif 4331} 4332 4333void 4334arc_fini(void) 4335{ 4336 int i; 4337 4338 mutex_enter(&arc_reclaim_thr_lock); 4339 arc_thread_exit = 1; 4340 cv_signal(&arc_reclaim_thr_cv); 4341 while (arc_thread_exit != 0) 4342 cv_wait(&arc_reclaim_thr_cv, &arc_reclaim_thr_lock); 4343 mutex_exit(&arc_reclaim_thr_lock); 4344 4345 arc_flush(NULL); 4346 4347 arc_dead = TRUE; 4348 4349 if (arc_ksp != NULL) { 4350 kstat_delete(arc_ksp); 4351 arc_ksp = NULL; 4352 } 4353 4354 mutex_destroy(&arc_eviction_mtx); 4355 mutex_destroy(&arc_reclaim_thr_lock); 4356 cv_destroy(&arc_reclaim_thr_cv); 4357 4358 for (i = 0; i < ARC_BUFC_NUMLISTS; i++) { 4359 list_destroy(&arc_mru->arcs_lists[i]); 4360 list_destroy(&arc_mru_ghost->arcs_lists[i]); 4361 list_destroy(&arc_mfu->arcs_lists[i]); 4362 list_destroy(&arc_mfu_ghost->arcs_lists[i]); 4363 list_destroy(&arc_l2c_only->arcs_lists[i]); 4364 4365 mutex_destroy(&arc_anon->arcs_locks[i].arcs_lock); 4366 mutex_destroy(&arc_mru->arcs_locks[i].arcs_lock); 4367 mutex_destroy(&arc_mru_ghost->arcs_locks[i].arcs_lock); 4368 mutex_destroy(&arc_mfu->arcs_locks[i].arcs_lock); 4369 mutex_destroy(&arc_mfu_ghost->arcs_locks[i].arcs_lock); 4370 mutex_destroy(&arc_l2c_only->arcs_locks[i].arcs_lock); 4371 } 4372 4373 buf_fini(); 4374 4375 ASSERT(arc_loaned_bytes == 0); 4376 4377 mutex_destroy(&arc_lowmem_lock); 4378#ifdef _KERNEL 4379 if (arc_event_lowmem != NULL) 4380 EVENTHANDLER_DEREGISTER(vm_lowmem, arc_event_lowmem); 4381#endif 4382} 4383 4384/* 4385 * Level 2 ARC 4386 * 4387 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk. 4388 * It uses dedicated storage devices to hold cached data, which are populated 4389 * using large infrequent writes. The main role of this cache is to boost 4390 * the performance of random read workloads. The intended L2ARC devices 4391 * include short-stroked disks, solid state disks, and other media with 4392 * substantially faster read latency than disk. 4393 * 4394 * +-----------------------+ 4395 * | ARC | 4396 * +-----------------------+ 4397 * | ^ ^ 4398 * | | | 4399 * l2arc_feed_thread() arc_read() 4400 * | | | 4401 * | l2arc read | 4402 * V | | 4403 * +---------------+ | 4404 * | L2ARC | | 4405 * +---------------+ | 4406 * | ^ | 4407 * l2arc_write() | | 4408 * | | | 4409 * V | | 4410 * +-------+ +-------+ 4411 * | vdev | | vdev | 4412 * | cache | | cache | 4413 * +-------+ +-------+ 4414 * +=========+ .-----. 4415 * : L2ARC : |-_____-| 4416 * : devices : | Disks | 4417 * +=========+ `-_____-' 4418 * 4419 * Read requests are satisfied from the following sources, in order: 4420 * 4421 * 1) ARC 4422 * 2) vdev cache of L2ARC devices 4423 * 3) L2ARC devices 4424 * 4) vdev cache of disks 4425 * 5) disks 4426 * 4427 * Some L2ARC device types exhibit extremely slow write performance. 4428 * To accommodate for this there are some significant differences between 4429 * the L2ARC and traditional cache design: 4430 * 4431 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from 4432 * the ARC behave as usual, freeing buffers and placing headers on ghost 4433 * lists. The ARC does not send buffers to the L2ARC during eviction as 4434 * this would add inflated write latencies for all ARC memory pressure. 4435 * 4436 * 2. The L2ARC attempts to cache data from the ARC before it is evicted. 4437 * It does this by periodically scanning buffers from the eviction-end of 4438 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are 4439 * not already there. It scans until a headroom of buffers is satisfied, 4440 * which itself is a buffer for ARC eviction. If a compressible buffer is 4441 * found during scanning and selected for writing to an L2ARC device, we 4442 * temporarily boost scanning headroom during the next scan cycle to make 4443 * sure we adapt to compression effects (which might significantly reduce 4444 * the data volume we write to L2ARC). The thread that does this is 4445 * l2arc_feed_thread(), illustrated below; example sizes are included to 4446 * provide a better sense of ratio than this diagram: 4447 * 4448 * head --> tail 4449 * +---------------------+----------+ 4450 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC 4451 * +---------------------+----------+ | o L2ARC eligible 4452 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer 4453 * +---------------------+----------+ | 4454 * 15.9 Gbytes ^ 32 Mbytes | 4455 * headroom | 4456 * l2arc_feed_thread() 4457 * | 4458 * l2arc write hand <--[oooo]--' 4459 * | 8 Mbyte 4460 * | write max 4461 * V 4462 * +==============================+ 4463 * L2ARC dev |####|#|###|###| |####| ... | 4464 * +==============================+ 4465 * 32 Gbytes 4466 * 4467 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of 4468 * evicted, then the L2ARC has cached a buffer much sooner than it probably 4469 * needed to, potentially wasting L2ARC device bandwidth and storage. It is 4470 * safe to say that this is an uncommon case, since buffers at the end of 4471 * the ARC lists have moved there due to inactivity. 4472 * 4473 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom, 4474 * then the L2ARC simply misses copying some buffers. This serves as a 4475 * pressure valve to prevent heavy read workloads from both stalling the ARC 4476 * with waits and clogging the L2ARC with writes. This also helps prevent 4477 * the potential for the L2ARC to churn if it attempts to cache content too 4478 * quickly, such as during backups of the entire pool. 4479 * 4480 * 5. After system boot and before the ARC has filled main memory, there are 4481 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru 4482 * lists can remain mostly static. Instead of searching from tail of these 4483 * lists as pictured, the l2arc_feed_thread() will search from the list heads 4484 * for eligible buffers, greatly increasing its chance of finding them. 4485 * 4486 * The L2ARC device write speed is also boosted during this time so that 4487 * the L2ARC warms up faster. Since there have been no ARC evictions yet, 4488 * there are no L2ARC reads, and no fear of degrading read performance 4489 * through increased writes. 4490 * 4491 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that 4492 * the vdev queue can aggregate them into larger and fewer writes. Each 4493 * device is written to in a rotor fashion, sweeping writes through 4494 * available space then repeating. 4495 * 4496 * 7. The L2ARC does not store dirty content. It never needs to flush 4497 * write buffers back to disk based storage. 4498 * 4499 * 8. If an ARC buffer is written (and dirtied) which also exists in the 4500 * L2ARC, the now stale L2ARC buffer is immediately dropped. 4501 * 4502 * The performance of the L2ARC can be tweaked by a number of tunables, which 4503 * may be necessary for different workloads: 4504 * 4505 * l2arc_write_max max write bytes per interval 4506 * l2arc_write_boost extra write bytes during device warmup 4507 * l2arc_noprefetch skip caching prefetched buffers 4508 * l2arc_headroom number of max device writes to precache 4509 * l2arc_headroom_boost when we find compressed buffers during ARC 4510 * scanning, we multiply headroom by this 4511 * percentage factor for the next scan cycle, 4512 * since more compressed buffers are likely to 4513 * be present 4514 * l2arc_feed_secs seconds between L2ARC writing 4515 * 4516 * Tunables may be removed or added as future performance improvements are 4517 * integrated, and also may become zpool properties. 4518 * 4519 * There are three key functions that control how the L2ARC warms up: 4520 * 4521 * l2arc_write_eligible() check if a buffer is eligible to cache 4522 * l2arc_write_size() calculate how much to write 4523 * l2arc_write_interval() calculate sleep delay between writes 4524 * 4525 * These three functions determine what to write, how much, and how quickly 4526 * to send writes. 4527 */ 4528 4529static boolean_t 4530l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *ab) 4531{ 4532 /* 4533 * A buffer is *not* eligible for the L2ARC if it: 4534 * 1. belongs to a different spa. 4535 * 2. is already cached on the L2ARC. 4536 * 3. has an I/O in progress (it may be an incomplete read). 4537 * 4. is flagged not eligible (zfs property). 4538 */ 4539 if (ab->b_spa != spa_guid) { 4540 ARCSTAT_BUMP(arcstat_l2_write_spa_mismatch); 4541 return (B_FALSE); 4542 } 4543 if (ab->b_l2hdr != NULL) { 4544 ARCSTAT_BUMP(arcstat_l2_write_in_l2); 4545 return (B_FALSE); 4546 } 4547 if (HDR_IO_IN_PROGRESS(ab)) { 4548 ARCSTAT_BUMP(arcstat_l2_write_hdr_io_in_progress); 4549 return (B_FALSE); 4550 } 4551 if (!HDR_L2CACHE(ab)) { 4552 ARCSTAT_BUMP(arcstat_l2_write_not_cacheable); 4553 return (B_FALSE); 4554 } 4555 4556 return (B_TRUE); 4557} 4558 4559static uint64_t 4560l2arc_write_size(void) 4561{ 4562 uint64_t size; 4563 4564 /* 4565 * Make sure our globals have meaningful values in case the user 4566 * altered them. 4567 */ 4568 size = l2arc_write_max; 4569 if (size == 0) { 4570 cmn_err(CE_NOTE, "Bad value for l2arc_write_max, value must " 4571 "be greater than zero, resetting it to the default (%d)", 4572 L2ARC_WRITE_SIZE); 4573 size = l2arc_write_max = L2ARC_WRITE_SIZE; 4574 } 4575 4576 if (arc_warm == B_FALSE) 4577 size += l2arc_write_boost; 4578 4579 return (size); 4580 4581} 4582 4583static clock_t 4584l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote) 4585{ 4586 clock_t interval, next, now; 4587 4588 /* 4589 * If the ARC lists are busy, increase our write rate; if the 4590 * lists are stale, idle back. This is achieved by checking 4591 * how much we previously wrote - if it was more than half of 4592 * what we wanted, schedule the next write much sooner. 4593 */ 4594 if (l2arc_feed_again && wrote > (wanted / 2)) 4595 interval = (hz * l2arc_feed_min_ms) / 1000; 4596 else 4597 interval = hz * l2arc_feed_secs; 4598 4599 now = ddi_get_lbolt(); 4600 next = MAX(now, MIN(now + interval, began + interval)); 4601 4602 return (next); 4603} 4604 4605static void 4606l2arc_hdr_stat_add(void) 4607{ 4608 ARCSTAT_INCR(arcstat_l2_hdr_size, HDR_SIZE + L2HDR_SIZE); 4609 ARCSTAT_INCR(arcstat_hdr_size, -HDR_SIZE); 4610} 4611 4612static void 4613l2arc_hdr_stat_remove(void) 4614{ 4615 ARCSTAT_INCR(arcstat_l2_hdr_size, -(HDR_SIZE + L2HDR_SIZE)); 4616 ARCSTAT_INCR(arcstat_hdr_size, HDR_SIZE); 4617} 4618 4619/* 4620 * Cycle through L2ARC devices. This is how L2ARC load balances. 4621 * If a device is returned, this also returns holding the spa config lock. 4622 */ 4623static l2arc_dev_t * 4624l2arc_dev_get_next(void) 4625{ 4626 l2arc_dev_t *first, *next = NULL; 4627 4628 /* 4629 * Lock out the removal of spas (spa_namespace_lock), then removal 4630 * of cache devices (l2arc_dev_mtx). Once a device has been selected, 4631 * both locks will be dropped and a spa config lock held instead. 4632 */ 4633 mutex_enter(&spa_namespace_lock); 4634 mutex_enter(&l2arc_dev_mtx); 4635 4636 /* if there are no vdevs, there is nothing to do */ 4637 if (l2arc_ndev == 0) 4638 goto out; 4639 4640 first = NULL; 4641 next = l2arc_dev_last; 4642 do { 4643 /* loop around the list looking for a non-faulted vdev */ 4644 if (next == NULL) { 4645 next = list_head(l2arc_dev_list); 4646 } else { 4647 next = list_next(l2arc_dev_list, next); 4648 if (next == NULL) 4649 next = list_head(l2arc_dev_list); 4650 } 4651 4652 /* if we have come back to the start, bail out */ 4653 if (first == NULL) 4654 first = next; 4655 else if (next == first) 4656 break; 4657 4658 } while (vdev_is_dead(next->l2ad_vdev)); 4659 4660 /* if we were unable to find any usable vdevs, return NULL */ 4661 if (vdev_is_dead(next->l2ad_vdev)) 4662 next = NULL; 4663 4664 l2arc_dev_last = next; 4665 4666out: 4667 mutex_exit(&l2arc_dev_mtx); 4668 4669 /* 4670 * Grab the config lock to prevent the 'next' device from being 4671 * removed while we are writing to it. 4672 */ 4673 if (next != NULL) 4674 spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER); 4675 mutex_exit(&spa_namespace_lock); 4676 4677 return (next); 4678} 4679 4680/* 4681 * Free buffers that were tagged for destruction. 4682 */ 4683static void 4684l2arc_do_free_on_write() 4685{ 4686 list_t *buflist; 4687 l2arc_data_free_t *df, *df_prev; 4688 4689 mutex_enter(&l2arc_free_on_write_mtx); 4690 buflist = l2arc_free_on_write; 4691 4692 for (df = list_tail(buflist); df; df = df_prev) { 4693 df_prev = list_prev(buflist, df); 4694 ASSERT(df->l2df_data != NULL); 4695 ASSERT(df->l2df_func != NULL); 4696 df->l2df_func(df->l2df_data, df->l2df_size); 4697 list_remove(buflist, df); 4698 kmem_free(df, sizeof (l2arc_data_free_t)); 4699 } 4700 4701 mutex_exit(&l2arc_free_on_write_mtx); 4702} 4703 4704/* 4705 * A write to a cache device has completed. Update all headers to allow 4706 * reads from these buffers to begin. 4707 */ 4708static void 4709l2arc_write_done(zio_t *zio) 4710{ 4711 l2arc_write_callback_t *cb; 4712 l2arc_dev_t *dev; 4713 list_t *buflist; 4714 arc_buf_hdr_t *head, *ab, *ab_prev; 4715 l2arc_buf_hdr_t *abl2; 4716 kmutex_t *hash_lock; 4717 int64_t bytes_dropped = 0; 4718 4719 cb = zio->io_private; 4720 ASSERT(cb != NULL); 4721 dev = cb->l2wcb_dev; 4722 ASSERT(dev != NULL); 4723 head = cb->l2wcb_head; 4724 ASSERT(head != NULL); 4725 buflist = dev->l2ad_buflist; 4726 ASSERT(buflist != NULL); 4727 DTRACE_PROBE2(l2arc__iodone, zio_t *, zio, 4728 l2arc_write_callback_t *, cb); 4729 4730 if (zio->io_error != 0) 4731 ARCSTAT_BUMP(arcstat_l2_writes_error); 4732 4733 mutex_enter(&l2arc_buflist_mtx); 4734 4735 /* 4736 * All writes completed, or an error was hit. 4737 */ 4738 for (ab = list_prev(buflist, head); ab; ab = ab_prev) { 4739 ab_prev = list_prev(buflist, ab); 4740 abl2 = ab->b_l2hdr; 4741 4742 /* 4743 * Release the temporary compressed buffer as soon as possible. 4744 */ 4745 if (abl2->b_compress != ZIO_COMPRESS_OFF) 4746 l2arc_release_cdata_buf(ab); 4747 4748 hash_lock = HDR_LOCK(ab); 4749 if (!mutex_tryenter(hash_lock)) { 4750 /* 4751 * This buffer misses out. It may be in a stage 4752 * of eviction. Its ARC_L2_WRITING flag will be 4753 * left set, denying reads to this buffer. 4754 */ 4755 ARCSTAT_BUMP(arcstat_l2_writes_hdr_miss); 4756 continue; 4757 } 4758 4759 if (zio->io_error != 0) { 4760 /* 4761 * Error - drop L2ARC entry. 4762 */ 4763 list_remove(buflist, ab); 4764 ARCSTAT_INCR(arcstat_l2_asize, -abl2->b_asize); 4765 bytes_dropped += abl2->b_asize; 4766 ab->b_l2hdr = NULL; 4767 trim_map_free(abl2->b_dev->l2ad_vdev, abl2->b_daddr, 4768 ab->b_size, 0); 4769 kmem_free(abl2, sizeof (l2arc_buf_hdr_t)); 4770 ARCSTAT_INCR(arcstat_l2_size, -ab->b_size); 4771 } 4772 4773 /* 4774 * Allow ARC to begin reads to this L2ARC entry. 4775 */ 4776 ab->b_flags &= ~ARC_L2_WRITING; 4777 4778 mutex_exit(hash_lock); 4779 } 4780 4781 atomic_inc_64(&l2arc_writes_done); 4782 list_remove(buflist, head); 4783 kmem_cache_free(hdr_cache, head); 4784 mutex_exit(&l2arc_buflist_mtx); 4785 4786 vdev_space_update(dev->l2ad_vdev, -bytes_dropped, 0, 0); 4787 4788 l2arc_do_free_on_write(); 4789 4790 kmem_free(cb, sizeof (l2arc_write_callback_t)); 4791} 4792 4793/* 4794 * A read to a cache device completed. Validate buffer contents before 4795 * handing over to the regular ARC routines. 4796 */ 4797static void 4798l2arc_read_done(zio_t *zio) 4799{ 4800 l2arc_read_callback_t *cb; 4801 arc_buf_hdr_t *hdr; 4802 arc_buf_t *buf; 4803 kmutex_t *hash_lock; 4804 int equal; 4805 4806 ASSERT(zio->io_vd != NULL); 4807 ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE); 4808 4809 spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd); 4810 4811 cb = zio->io_private; 4812 ASSERT(cb != NULL); 4813 buf = cb->l2rcb_buf; 4814 ASSERT(buf != NULL); 4815 4816 hash_lock = HDR_LOCK(buf->b_hdr); 4817 mutex_enter(hash_lock); 4818 hdr = buf->b_hdr; 4819 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 4820 4821 /* 4822 * If the buffer was compressed, decompress it first. 4823 */ 4824 if (cb->l2rcb_compress != ZIO_COMPRESS_OFF) 4825 l2arc_decompress_zio(zio, hdr, cb->l2rcb_compress); 4826 ASSERT(zio->io_data != NULL); 4827 4828 /* 4829 * Check this survived the L2ARC journey. 4830 */ 4831 equal = arc_cksum_equal(buf); 4832 if (equal && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) { 4833 mutex_exit(hash_lock); 4834 zio->io_private = buf; 4835 zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */ 4836 zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */ 4837 arc_read_done(zio); 4838 } else { 4839 mutex_exit(hash_lock); 4840 /* 4841 * Buffer didn't survive caching. Increment stats and 4842 * reissue to the original storage device. 4843 */ 4844 if (zio->io_error != 0) { 4845 ARCSTAT_BUMP(arcstat_l2_io_error); 4846 } else { 4847 zio->io_error = SET_ERROR(EIO); 4848 } 4849 if (!equal) 4850 ARCSTAT_BUMP(arcstat_l2_cksum_bad); 4851 4852 /* 4853 * If there's no waiter, issue an async i/o to the primary 4854 * storage now. If there *is* a waiter, the caller must 4855 * issue the i/o in a context where it's OK to block. 4856 */ 4857 if (zio->io_waiter == NULL) { 4858 zio_t *pio = zio_unique_parent(zio); 4859 4860 ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL); 4861 4862 zio_nowait(zio_read(pio, cb->l2rcb_spa, &cb->l2rcb_bp, 4863 buf->b_data, zio->io_size, arc_read_done, buf, 4864 zio->io_priority, cb->l2rcb_flags, &cb->l2rcb_zb)); 4865 } 4866 } 4867 4868 kmem_free(cb, sizeof (l2arc_read_callback_t)); 4869} 4870 4871/* 4872 * This is the list priority from which the L2ARC will search for pages to 4873 * cache. This is used within loops (0..3) to cycle through lists in the 4874 * desired order. This order can have a significant effect on cache 4875 * performance. 4876 * 4877 * Currently the metadata lists are hit first, MFU then MRU, followed by 4878 * the data lists. This function returns a locked list, and also returns 4879 * the lock pointer. 4880 */ 4881static list_t * 4882l2arc_list_locked(int list_num, kmutex_t **lock) 4883{ 4884 list_t *list = NULL; 4885 int idx; 4886 4887 ASSERT(list_num >= 0 && list_num < 2 * ARC_BUFC_NUMLISTS); 4888 4889 if (list_num < ARC_BUFC_NUMMETADATALISTS) { 4890 idx = list_num; 4891 list = &arc_mfu->arcs_lists[idx]; 4892 *lock = ARCS_LOCK(arc_mfu, idx); 4893 } else if (list_num < ARC_BUFC_NUMMETADATALISTS * 2) { 4894 idx = list_num - ARC_BUFC_NUMMETADATALISTS; 4895 list = &arc_mru->arcs_lists[idx]; 4896 *lock = ARCS_LOCK(arc_mru, idx); 4897 } else if (list_num < (ARC_BUFC_NUMMETADATALISTS * 2 + 4898 ARC_BUFC_NUMDATALISTS)) { 4899 idx = list_num - ARC_BUFC_NUMMETADATALISTS; 4900 list = &arc_mfu->arcs_lists[idx]; 4901 *lock = ARCS_LOCK(arc_mfu, idx); 4902 } else { 4903 idx = list_num - ARC_BUFC_NUMLISTS; 4904 list = &arc_mru->arcs_lists[idx]; 4905 *lock = ARCS_LOCK(arc_mru, idx); 4906 } 4907 4908 ASSERT(!(MUTEX_HELD(*lock))); 4909 mutex_enter(*lock); 4910 return (list); 4911} 4912 4913/* 4914 * Evict buffers from the device write hand to the distance specified in 4915 * bytes. This distance may span populated buffers, it may span nothing. 4916 * This is clearing a region on the L2ARC device ready for writing. 4917 * If the 'all' boolean is set, every buffer is evicted. 4918 */ 4919static void 4920l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all) 4921{ 4922 list_t *buflist; 4923 l2arc_buf_hdr_t *abl2; 4924 arc_buf_hdr_t *ab, *ab_prev; 4925 kmutex_t *hash_lock; 4926 uint64_t taddr; 4927 int64_t bytes_evicted = 0; 4928 4929 buflist = dev->l2ad_buflist; 4930 4931 if (buflist == NULL) 4932 return; 4933 4934 if (!all && dev->l2ad_first) { 4935 /* 4936 * This is the first sweep through the device. There is 4937 * nothing to evict. 4938 */ 4939 return; 4940 } 4941 4942 if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) { 4943 /* 4944 * When nearing the end of the device, evict to the end 4945 * before the device write hand jumps to the start. 4946 */ 4947 taddr = dev->l2ad_end; 4948 } else { 4949 taddr = dev->l2ad_hand + distance; 4950 } 4951 DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist, 4952 uint64_t, taddr, boolean_t, all); 4953 4954top: 4955 mutex_enter(&l2arc_buflist_mtx); 4956 for (ab = list_tail(buflist); ab; ab = ab_prev) { 4957 ab_prev = list_prev(buflist, ab); 4958 4959 hash_lock = HDR_LOCK(ab); 4960 if (!mutex_tryenter(hash_lock)) { 4961 /* 4962 * Missed the hash lock. Retry. 4963 */ 4964 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry); 4965 mutex_exit(&l2arc_buflist_mtx); 4966 mutex_enter(hash_lock); 4967 mutex_exit(hash_lock); 4968 goto top; 4969 } 4970 4971 if (HDR_L2_WRITE_HEAD(ab)) { 4972 /* 4973 * We hit a write head node. Leave it for 4974 * l2arc_write_done(). 4975 */ 4976 list_remove(buflist, ab); 4977 mutex_exit(hash_lock); 4978 continue; 4979 } 4980 4981 if (!all && ab->b_l2hdr != NULL && 4982 (ab->b_l2hdr->b_daddr > taddr || 4983 ab->b_l2hdr->b_daddr < dev->l2ad_hand)) { 4984 /* 4985 * We've evicted to the target address, 4986 * or the end of the device. 4987 */ 4988 mutex_exit(hash_lock); 4989 break; 4990 } 4991 4992 if (HDR_FREE_IN_PROGRESS(ab)) { 4993 /* 4994 * Already on the path to destruction. 4995 */ 4996 mutex_exit(hash_lock); 4997 continue; 4998 } 4999 5000 if (ab->b_state == arc_l2c_only) { 5001 ASSERT(!HDR_L2_READING(ab)); 5002 /* 5003 * This doesn't exist in the ARC. Destroy. 5004 * arc_hdr_destroy() will call list_remove() 5005 * and decrement arcstat_l2_size. 5006 */ 5007 arc_change_state(arc_anon, ab, hash_lock); 5008 arc_hdr_destroy(ab); 5009 } else { 5010 /* 5011 * Invalidate issued or about to be issued 5012 * reads, since we may be about to write 5013 * over this location. 5014 */ 5015 if (HDR_L2_READING(ab)) { 5016 ARCSTAT_BUMP(arcstat_l2_evict_reading); 5017 ab->b_flags |= ARC_L2_EVICTED; 5018 } 5019 5020 /* 5021 * Tell ARC this no longer exists in L2ARC. 5022 */ 5023 if (ab->b_l2hdr != NULL) { 5024 abl2 = ab->b_l2hdr; 5025 ARCSTAT_INCR(arcstat_l2_asize, -abl2->b_asize); 5026 bytes_evicted += abl2->b_asize; 5027 ab->b_l2hdr = NULL; 5028 /* 5029 * We are destroying l2hdr, so ensure that 5030 * its compressed buffer, if any, is not leaked. 5031 */ 5032 ASSERT(abl2->b_tmp_cdata == NULL); 5033 kmem_free(abl2, sizeof (l2arc_buf_hdr_t)); 5034 ARCSTAT_INCR(arcstat_l2_size, -ab->b_size); 5035 } 5036 list_remove(buflist, ab); 5037 5038 /* 5039 * This may have been leftover after a 5040 * failed write. 5041 */ 5042 ab->b_flags &= ~ARC_L2_WRITING; 5043 } 5044 mutex_exit(hash_lock); 5045 } 5046 mutex_exit(&l2arc_buflist_mtx); 5047 5048 vdev_space_update(dev->l2ad_vdev, -bytes_evicted, 0, 0); 5049 dev->l2ad_evict = taddr; 5050} 5051 5052/* 5053 * Find and write ARC buffers to the L2ARC device. 5054 * 5055 * An ARC_L2_WRITING flag is set so that the L2ARC buffers are not valid 5056 * for reading until they have completed writing. 5057 * The headroom_boost is an in-out parameter used to maintain headroom boost 5058 * state between calls to this function. 5059 * 5060 * Returns the number of bytes actually written (which may be smaller than 5061 * the delta by which the device hand has changed due to alignment). 5062 */ 5063static uint64_t 5064l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz, 5065 boolean_t *headroom_boost) 5066{ 5067 arc_buf_hdr_t *ab, *ab_prev, *head; 5068 list_t *list; 5069 uint64_t write_asize, write_psize, write_sz, headroom, 5070 buf_compress_minsz; 5071 void *buf_data; 5072 kmutex_t *list_lock; 5073 boolean_t full; 5074 l2arc_write_callback_t *cb; 5075 zio_t *pio, *wzio; 5076 uint64_t guid = spa_load_guid(spa); 5077 const boolean_t do_headroom_boost = *headroom_boost; 5078 int try; 5079 5080 ASSERT(dev->l2ad_vdev != NULL); 5081 5082 /* Lower the flag now, we might want to raise it again later. */ 5083 *headroom_boost = B_FALSE; 5084 5085 pio = NULL; 5086 write_sz = write_asize = write_psize = 0; 5087 full = B_FALSE; 5088 head = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE); 5089 head->b_flags |= ARC_L2_WRITE_HEAD; 5090 5091 ARCSTAT_BUMP(arcstat_l2_write_buffer_iter); 5092 /* 5093 * We will want to try to compress buffers that are at least 2x the 5094 * device sector size. 5095 */ 5096 buf_compress_minsz = 2 << dev->l2ad_vdev->vdev_ashift; 5097 5098 /* 5099 * Copy buffers for L2ARC writing. 5100 */ 5101 mutex_enter(&l2arc_buflist_mtx); 5102 for (try = 0; try < 2 * ARC_BUFC_NUMLISTS; try++) { 5103 uint64_t passed_sz = 0; 5104 5105 list = l2arc_list_locked(try, &list_lock); 5106 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_iter); 5107 5108 /* 5109 * L2ARC fast warmup. 5110 * 5111 * Until the ARC is warm and starts to evict, read from the 5112 * head of the ARC lists rather than the tail. 5113 */ 5114 if (arc_warm == B_FALSE) 5115 ab = list_head(list); 5116 else 5117 ab = list_tail(list); 5118 if (ab == NULL) 5119 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_null_iter); 5120 5121 headroom = target_sz * l2arc_headroom * 2 / ARC_BUFC_NUMLISTS; 5122 if (do_headroom_boost) 5123 headroom = (headroom * l2arc_headroom_boost) / 100; 5124 5125 for (; ab; ab = ab_prev) { 5126 l2arc_buf_hdr_t *l2hdr; 5127 kmutex_t *hash_lock; 5128 uint64_t buf_sz; 5129 5130 if (arc_warm == B_FALSE) 5131 ab_prev = list_next(list, ab); 5132 else 5133 ab_prev = list_prev(list, ab); 5134 ARCSTAT_INCR(arcstat_l2_write_buffer_bytes_scanned, ab->b_size); 5135 5136 hash_lock = HDR_LOCK(ab); 5137 if (!mutex_tryenter(hash_lock)) { 5138 ARCSTAT_BUMP(arcstat_l2_write_trylock_fail); 5139 /* 5140 * Skip this buffer rather than waiting. 5141 */ 5142 continue; 5143 } 5144 5145 passed_sz += ab->b_size; 5146 if (passed_sz > headroom) { 5147 /* 5148 * Searched too far. 5149 */ 5150 mutex_exit(hash_lock); 5151 ARCSTAT_BUMP(arcstat_l2_write_passed_headroom); 5152 break; 5153 } 5154 5155 if (!l2arc_write_eligible(guid, ab)) { 5156 mutex_exit(hash_lock); 5157 continue; 5158 } 5159 5160 if ((write_sz + ab->b_size) > target_sz) { 5161 full = B_TRUE; 5162 mutex_exit(hash_lock); 5163 ARCSTAT_BUMP(arcstat_l2_write_full); 5164 break; 5165 } 5166 5167 if (pio == NULL) { 5168 /* 5169 * Insert a dummy header on the buflist so 5170 * l2arc_write_done() can find where the 5171 * write buffers begin without searching. 5172 */ 5173 list_insert_head(dev->l2ad_buflist, head); 5174 5175 cb = kmem_alloc( 5176 sizeof (l2arc_write_callback_t), KM_SLEEP); 5177 cb->l2wcb_dev = dev; 5178 cb->l2wcb_head = head; 5179 pio = zio_root(spa, l2arc_write_done, cb, 5180 ZIO_FLAG_CANFAIL); 5181 ARCSTAT_BUMP(arcstat_l2_write_pios); 5182 } 5183 5184 /* 5185 * Create and add a new L2ARC header. 5186 */ 5187 l2hdr = kmem_zalloc(sizeof (l2arc_buf_hdr_t), KM_SLEEP); 5188 l2hdr->b_dev = dev; 5189 ab->b_flags |= ARC_L2_WRITING; 5190 5191 /* 5192 * Temporarily stash the data buffer in b_tmp_cdata. 5193 * The subsequent write step will pick it up from 5194 * there. This is because can't access ab->b_buf 5195 * without holding the hash_lock, which we in turn 5196 * can't access without holding the ARC list locks 5197 * (which we want to avoid during compression/writing). 5198 */ 5199 l2hdr->b_compress = ZIO_COMPRESS_OFF; 5200 l2hdr->b_asize = ab->b_size; 5201 l2hdr->b_tmp_cdata = ab->b_buf->b_data; 5202 5203 buf_sz = ab->b_size; 5204 ab->b_l2hdr = l2hdr; 5205 5206 list_insert_head(dev->l2ad_buflist, ab); 5207 5208 /* 5209 * Compute and store the buffer cksum before 5210 * writing. On debug the cksum is verified first. 5211 */ 5212 arc_cksum_verify(ab->b_buf); 5213 arc_cksum_compute(ab->b_buf, B_TRUE); 5214 5215 mutex_exit(hash_lock); 5216 5217 write_sz += buf_sz; 5218 } 5219 5220 mutex_exit(list_lock); 5221 5222 if (full == B_TRUE) 5223 break; 5224 } 5225 5226 /* No buffers selected for writing? */ 5227 if (pio == NULL) { 5228 ASSERT0(write_sz); 5229 mutex_exit(&l2arc_buflist_mtx); 5230 kmem_cache_free(hdr_cache, head); 5231 return (0); 5232 } 5233 5234 /* 5235 * Now start writing the buffers. We're starting at the write head 5236 * and work backwards, retracing the course of the buffer selector 5237 * loop above. 5238 */ 5239 for (ab = list_prev(dev->l2ad_buflist, head); ab; 5240 ab = list_prev(dev->l2ad_buflist, ab)) { 5241 l2arc_buf_hdr_t *l2hdr; 5242 uint64_t buf_sz; 5243 5244 /* 5245 * We shouldn't need to lock the buffer here, since we flagged 5246 * it as ARC_L2_WRITING in the previous step, but we must take 5247 * care to only access its L2 cache parameters. In particular, 5248 * ab->b_buf may be invalid by now due to ARC eviction. 5249 */ 5250 l2hdr = ab->b_l2hdr; 5251 l2hdr->b_daddr = dev->l2ad_hand; 5252 5253 if ((ab->b_flags & ARC_L2COMPRESS) && 5254 l2hdr->b_asize >= buf_compress_minsz) { 5255 if (l2arc_compress_buf(l2hdr)) { 5256 /* 5257 * If compression succeeded, enable headroom 5258 * boost on the next scan cycle. 5259 */ 5260 *headroom_boost = B_TRUE; 5261 } 5262 } 5263 5264 /* 5265 * Pick up the buffer data we had previously stashed away 5266 * (and now potentially also compressed). 5267 */ 5268 buf_data = l2hdr->b_tmp_cdata; 5269 buf_sz = l2hdr->b_asize; 5270 5271 /* 5272 * If the data has not been compressed, then clear b_tmp_cdata 5273 * to make sure that it points only to a temporary compression 5274 * buffer. 5275 */ 5276 if (!L2ARC_IS_VALID_COMPRESS(l2hdr->b_compress)) 5277 l2hdr->b_tmp_cdata = NULL; 5278 5279 /* Compression may have squashed the buffer to zero length. */ 5280 if (buf_sz != 0) { 5281 uint64_t buf_p_sz; 5282 5283 wzio = zio_write_phys(pio, dev->l2ad_vdev, 5284 dev->l2ad_hand, buf_sz, buf_data, ZIO_CHECKSUM_OFF, 5285 NULL, NULL, ZIO_PRIORITY_ASYNC_WRITE, 5286 ZIO_FLAG_CANFAIL, B_FALSE); 5287 5288 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev, 5289 zio_t *, wzio); 5290 (void) zio_nowait(wzio); 5291 5292 write_asize += buf_sz; 5293 /* 5294 * Keep the clock hand suitably device-aligned. 5295 */ 5296 buf_p_sz = vdev_psize_to_asize(dev->l2ad_vdev, buf_sz); 5297 write_psize += buf_p_sz; 5298 dev->l2ad_hand += buf_p_sz; 5299 } 5300 } 5301 5302 mutex_exit(&l2arc_buflist_mtx); 5303 5304 ASSERT3U(write_asize, <=, target_sz); 5305 ARCSTAT_BUMP(arcstat_l2_writes_sent); 5306 ARCSTAT_INCR(arcstat_l2_write_bytes, write_asize); 5307 ARCSTAT_INCR(arcstat_l2_size, write_sz); 5308 ARCSTAT_INCR(arcstat_l2_asize, write_asize); 5309 vdev_space_update(dev->l2ad_vdev, write_asize, 0, 0); 5310 5311 /* 5312 * Bump device hand to the device start if it is approaching the end. 5313 * l2arc_evict() will already have evicted ahead for this case. 5314 */ 5315 if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) { 5316 dev->l2ad_hand = dev->l2ad_start; 5317 dev->l2ad_evict = dev->l2ad_start; 5318 dev->l2ad_first = B_FALSE; 5319 } 5320 5321 dev->l2ad_writing = B_TRUE; 5322 (void) zio_wait(pio); 5323 dev->l2ad_writing = B_FALSE; 5324 5325 return (write_asize); 5326} 5327 5328/* 5329 * Compresses an L2ARC buffer. 5330 * The data to be compressed must be prefilled in l2hdr->b_tmp_cdata and its 5331 * size in l2hdr->b_asize. This routine tries to compress the data and 5332 * depending on the compression result there are three possible outcomes: 5333 * *) The buffer was incompressible. The original l2hdr contents were left 5334 * untouched and are ready for writing to an L2 device. 5335 * *) The buffer was all-zeros, so there is no need to write it to an L2 5336 * device. To indicate this situation b_tmp_cdata is NULL'ed, b_asize is 5337 * set to zero and b_compress is set to ZIO_COMPRESS_EMPTY. 5338 * *) Compression succeeded and b_tmp_cdata was replaced with a temporary 5339 * data buffer which holds the compressed data to be written, and b_asize 5340 * tells us how much data there is. b_compress is set to the appropriate 5341 * compression algorithm. Once writing is done, invoke 5342 * l2arc_release_cdata_buf on this l2hdr to free this temporary buffer. 5343 * 5344 * Returns B_TRUE if compression succeeded, or B_FALSE if it didn't (the 5345 * buffer was incompressible). 5346 */ 5347static boolean_t 5348l2arc_compress_buf(l2arc_buf_hdr_t *l2hdr) 5349{ 5350 void *cdata; 5351 size_t csize, len, rounded; 5352 5353 ASSERT(l2hdr->b_compress == ZIO_COMPRESS_OFF); 5354 ASSERT(l2hdr->b_tmp_cdata != NULL); 5355 5356 len = l2hdr->b_asize; 5357 cdata = zio_data_buf_alloc(len); 5358 csize = zio_compress_data(ZIO_COMPRESS_LZ4, l2hdr->b_tmp_cdata, 5359 cdata, l2hdr->b_asize); 5360 5361 if (csize == 0) { 5362 /* zero block, indicate that there's nothing to write */ 5363 zio_data_buf_free(cdata, len); 5364 l2hdr->b_compress = ZIO_COMPRESS_EMPTY; 5365 l2hdr->b_asize = 0; 5366 l2hdr->b_tmp_cdata = NULL; 5367 ARCSTAT_BUMP(arcstat_l2_compress_zeros); 5368 return (B_TRUE); 5369 } 5370 5371 rounded = P2ROUNDUP(csize, 5372 (size_t)1 << l2hdr->b_dev->l2ad_vdev->vdev_ashift); 5373 if (rounded < len) { 5374 /* 5375 * Compression succeeded, we'll keep the cdata around for 5376 * writing and release it afterwards. 5377 */ 5378 if (rounded > csize) { 5379 bzero((char *)cdata + csize, rounded - csize); 5380 csize = rounded; 5381 } 5382 l2hdr->b_compress = ZIO_COMPRESS_LZ4; 5383 l2hdr->b_asize = csize; 5384 l2hdr->b_tmp_cdata = cdata; 5385 ARCSTAT_BUMP(arcstat_l2_compress_successes); 5386 return (B_TRUE); 5387 } else { 5388 /* 5389 * Compression failed, release the compressed buffer. 5390 * l2hdr will be left unmodified. 5391 */ 5392 zio_data_buf_free(cdata, len); 5393 ARCSTAT_BUMP(arcstat_l2_compress_failures); 5394 return (B_FALSE); 5395 } 5396} 5397 5398/* 5399 * Decompresses a zio read back from an l2arc device. On success, the 5400 * underlying zio's io_data buffer is overwritten by the uncompressed 5401 * version. On decompression error (corrupt compressed stream), the 5402 * zio->io_error value is set to signal an I/O error. 5403 * 5404 * Please note that the compressed data stream is not checksummed, so 5405 * if the underlying device is experiencing data corruption, we may feed 5406 * corrupt data to the decompressor, so the decompressor needs to be 5407 * able to handle this situation (LZ4 does). 5408 */ 5409static void 5410l2arc_decompress_zio(zio_t *zio, arc_buf_hdr_t *hdr, enum zio_compress c) 5411{ 5412 ASSERT(L2ARC_IS_VALID_COMPRESS(c)); 5413 5414 if (zio->io_error != 0) { 5415 /* 5416 * An io error has occured, just restore the original io 5417 * size in preparation for a main pool read. 5418 */ 5419 zio->io_orig_size = zio->io_size = hdr->b_size; 5420 return; 5421 } 5422 5423 if (c == ZIO_COMPRESS_EMPTY) { 5424 /* 5425 * An empty buffer results in a null zio, which means we 5426 * need to fill its io_data after we're done restoring the 5427 * buffer's contents. 5428 */ 5429 ASSERT(hdr->b_buf != NULL); 5430 bzero(hdr->b_buf->b_data, hdr->b_size); 5431 zio->io_data = zio->io_orig_data = hdr->b_buf->b_data; 5432 } else { 5433 ASSERT(zio->io_data != NULL); 5434 /* 5435 * We copy the compressed data from the start of the arc buffer 5436 * (the zio_read will have pulled in only what we need, the 5437 * rest is garbage which we will overwrite at decompression) 5438 * and then decompress back to the ARC data buffer. This way we 5439 * can minimize copying by simply decompressing back over the 5440 * original compressed data (rather than decompressing to an 5441 * aux buffer and then copying back the uncompressed buffer, 5442 * which is likely to be much larger). 5443 */ 5444 uint64_t csize; 5445 void *cdata; 5446 5447 csize = zio->io_size; 5448 cdata = zio_data_buf_alloc(csize); 5449 bcopy(zio->io_data, cdata, csize); 5450 if (zio_decompress_data(c, cdata, zio->io_data, csize, 5451 hdr->b_size) != 0) 5452 zio->io_error = EIO; 5453 zio_data_buf_free(cdata, csize); 5454 } 5455 5456 /* Restore the expected uncompressed IO size. */ 5457 zio->io_orig_size = zio->io_size = hdr->b_size; 5458} 5459 5460/* 5461 * Releases the temporary b_tmp_cdata buffer in an l2arc header structure. 5462 * This buffer serves as a temporary holder of compressed data while 5463 * the buffer entry is being written to an l2arc device. Once that is 5464 * done, we can dispose of it. 5465 */ 5466static void 5467l2arc_release_cdata_buf(arc_buf_hdr_t *ab) 5468{ 5469 l2arc_buf_hdr_t *l2hdr = ab->b_l2hdr; 5470 5471 ASSERT(L2ARC_IS_VALID_COMPRESS(l2hdr->b_compress)); 5472 if (l2hdr->b_compress != ZIO_COMPRESS_EMPTY) { 5473 /* 5474 * If the data was compressed, then we've allocated a 5475 * temporary buffer for it, so now we need to release it. 5476 */ 5477 ASSERT(l2hdr->b_tmp_cdata != NULL); 5478 zio_data_buf_free(l2hdr->b_tmp_cdata, ab->b_size); 5479 l2hdr->b_tmp_cdata = NULL; 5480 } else { 5481 ASSERT(l2hdr->b_tmp_cdata == NULL); 5482 } 5483} 5484 5485/* 5486 * This thread feeds the L2ARC at regular intervals. This is the beating 5487 * heart of the L2ARC. 5488 */ 5489static void 5490l2arc_feed_thread(void *dummy __unused) 5491{ 5492 callb_cpr_t cpr; 5493 l2arc_dev_t *dev; 5494 spa_t *spa; 5495 uint64_t size, wrote; 5496 clock_t begin, next = ddi_get_lbolt(); 5497 boolean_t headroom_boost = B_FALSE; 5498 5499 CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG); 5500 5501 mutex_enter(&l2arc_feed_thr_lock); 5502 5503 while (l2arc_thread_exit == 0) { 5504 CALLB_CPR_SAFE_BEGIN(&cpr); 5505 (void) cv_timedwait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock, 5506 next - ddi_get_lbolt()); 5507 CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock); 5508 next = ddi_get_lbolt() + hz; 5509 5510 /* 5511 * Quick check for L2ARC devices. 5512 */ 5513 mutex_enter(&l2arc_dev_mtx); 5514 if (l2arc_ndev == 0) { 5515 mutex_exit(&l2arc_dev_mtx); 5516 continue; 5517 } 5518 mutex_exit(&l2arc_dev_mtx); 5519 begin = ddi_get_lbolt(); 5520 5521 /* 5522 * This selects the next l2arc device to write to, and in 5523 * doing so the next spa to feed from: dev->l2ad_spa. This 5524 * will return NULL if there are now no l2arc devices or if 5525 * they are all faulted. 5526 * 5527 * If a device is returned, its spa's config lock is also 5528 * held to prevent device removal. l2arc_dev_get_next() 5529 * will grab and release l2arc_dev_mtx. 5530 */ 5531 if ((dev = l2arc_dev_get_next()) == NULL) 5532 continue; 5533 5534 spa = dev->l2ad_spa; 5535 ASSERT(spa != NULL); 5536 5537 /* 5538 * If the pool is read-only then force the feed thread to 5539 * sleep a little longer. 5540 */ 5541 if (!spa_writeable(spa)) { 5542 next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz; 5543 spa_config_exit(spa, SCL_L2ARC, dev); 5544 continue; 5545 } 5546 5547 /* 5548 * Avoid contributing to memory pressure. 5549 */ 5550 if (arc_reclaim_needed()) { 5551 ARCSTAT_BUMP(arcstat_l2_abort_lowmem); 5552 spa_config_exit(spa, SCL_L2ARC, dev); 5553 continue; 5554 } 5555 5556 ARCSTAT_BUMP(arcstat_l2_feeds); 5557 5558 size = l2arc_write_size(); 5559 5560 /* 5561 * Evict L2ARC buffers that will be overwritten. 5562 */ 5563 l2arc_evict(dev, size, B_FALSE); 5564 5565 /* 5566 * Write ARC buffers. 5567 */ 5568 wrote = l2arc_write_buffers(spa, dev, size, &headroom_boost); 5569 5570 /* 5571 * Calculate interval between writes. 5572 */ 5573 next = l2arc_write_interval(begin, size, wrote); 5574 spa_config_exit(spa, SCL_L2ARC, dev); 5575 } 5576 5577 l2arc_thread_exit = 0; 5578 cv_broadcast(&l2arc_feed_thr_cv); 5579 CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */ 5580 thread_exit(); 5581} 5582 5583boolean_t 5584l2arc_vdev_present(vdev_t *vd) 5585{ 5586 l2arc_dev_t *dev; 5587 5588 mutex_enter(&l2arc_dev_mtx); 5589 for (dev = list_head(l2arc_dev_list); dev != NULL; 5590 dev = list_next(l2arc_dev_list, dev)) { 5591 if (dev->l2ad_vdev == vd) 5592 break; 5593 } 5594 mutex_exit(&l2arc_dev_mtx); 5595 5596 return (dev != NULL); 5597} 5598 5599/* 5600 * Add a vdev for use by the L2ARC. By this point the spa has already 5601 * validated the vdev and opened it. 5602 */ 5603void 5604l2arc_add_vdev(spa_t *spa, vdev_t *vd) 5605{ 5606 l2arc_dev_t *adddev; 5607 5608 ASSERT(!l2arc_vdev_present(vd)); 5609 5610 vdev_ashift_optimize(vd); 5611 5612 /* 5613 * Create a new l2arc device entry. 5614 */ 5615 adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP); 5616 adddev->l2ad_spa = spa; 5617 adddev->l2ad_vdev = vd; 5618 adddev->l2ad_start = VDEV_LABEL_START_SIZE; 5619 adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd); 5620 adddev->l2ad_hand = adddev->l2ad_start; 5621 adddev->l2ad_evict = adddev->l2ad_start; 5622 adddev->l2ad_first = B_TRUE; 5623 adddev->l2ad_writing = B_FALSE; 5624 5625 /* 5626 * This is a list of all ARC buffers that are still valid on the 5627 * device. 5628 */ 5629 adddev->l2ad_buflist = kmem_zalloc(sizeof (list_t), KM_SLEEP); 5630 list_create(adddev->l2ad_buflist, sizeof (arc_buf_hdr_t), 5631 offsetof(arc_buf_hdr_t, b_l2node)); 5632 5633 vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand); 5634 5635 /* 5636 * Add device to global list 5637 */ 5638 mutex_enter(&l2arc_dev_mtx); 5639 list_insert_head(l2arc_dev_list, adddev); 5640 atomic_inc_64(&l2arc_ndev); 5641 mutex_exit(&l2arc_dev_mtx); 5642} 5643 5644/* 5645 * Remove a vdev from the L2ARC. 5646 */ 5647void 5648l2arc_remove_vdev(vdev_t *vd) 5649{ 5650 l2arc_dev_t *dev, *nextdev, *remdev = NULL; 5651 5652 /* 5653 * Find the device by vdev 5654 */ 5655 mutex_enter(&l2arc_dev_mtx); 5656 for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) { 5657 nextdev = list_next(l2arc_dev_list, dev); 5658 if (vd == dev->l2ad_vdev) { 5659 remdev = dev; 5660 break; 5661 } 5662 } 5663 ASSERT(remdev != NULL); 5664 5665 /* 5666 * Remove device from global list 5667 */ 5668 list_remove(l2arc_dev_list, remdev); 5669 l2arc_dev_last = NULL; /* may have been invalidated */ 5670 atomic_dec_64(&l2arc_ndev); 5671 mutex_exit(&l2arc_dev_mtx); 5672 5673 /* 5674 * Clear all buflists and ARC references. L2ARC device flush. 5675 */ 5676 l2arc_evict(remdev, 0, B_TRUE); 5677 list_destroy(remdev->l2ad_buflist); 5678 kmem_free(remdev->l2ad_buflist, sizeof (list_t)); 5679 kmem_free(remdev, sizeof (l2arc_dev_t)); 5680} 5681 5682void 5683l2arc_init(void) 5684{ 5685 l2arc_thread_exit = 0; 5686 l2arc_ndev = 0; 5687 l2arc_writes_sent = 0; 5688 l2arc_writes_done = 0; 5689 5690 mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL); 5691 cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL); 5692 mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL); 5693 mutex_init(&l2arc_buflist_mtx, NULL, MUTEX_DEFAULT, NULL); 5694 mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL); 5695 5696 l2arc_dev_list = &L2ARC_dev_list; 5697 l2arc_free_on_write = &L2ARC_free_on_write; 5698 list_create(l2arc_dev_list, sizeof (l2arc_dev_t), 5699 offsetof(l2arc_dev_t, l2ad_node)); 5700 list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t), 5701 offsetof(l2arc_data_free_t, l2df_list_node)); 5702} 5703 5704void 5705l2arc_fini(void) 5706{ 5707 /* 5708 * This is called from dmu_fini(), which is called from spa_fini(); 5709 * Because of this, we can assume that all l2arc devices have 5710 * already been removed when the pools themselves were removed. 5711 */ 5712 5713 l2arc_do_free_on_write(); 5714 5715 mutex_destroy(&l2arc_feed_thr_lock); 5716 cv_destroy(&l2arc_feed_thr_cv); 5717 mutex_destroy(&l2arc_dev_mtx); 5718 mutex_destroy(&l2arc_buflist_mtx); 5719 mutex_destroy(&l2arc_free_on_write_mtx); 5720 5721 list_destroy(l2arc_dev_list); 5722 list_destroy(l2arc_free_on_write); 5723} 5724 5725void 5726l2arc_start(void) 5727{ 5728 if (!(spa_mode_global & FWRITE)) 5729 return; 5730 5731 (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0, 5732 TS_RUN, minclsyspri); 5733} 5734 5735void 5736l2arc_stop(void) 5737{ 5738 if (!(spa_mode_global & FWRITE)) 5739 return; 5740 5741 mutex_enter(&l2arc_feed_thr_lock); 5742 cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */ 5743 l2arc_thread_exit = 1; 5744 while (l2arc_thread_exit != 0) 5745 cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock); 5746 mutex_exit(&l2arc_feed_thr_lock); 5747} 5748