arc.c revision 273191
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); 208 209#ifdef _KERNEL 210static void 211arc_free_target_init(void *unused __unused) 212{ 213 214 zfs_arc_free_target = vm_pageout_wakeup_thresh; 215} 216SYSINIT(arc_free_target_init, SI_SUB_KTHREAD_PAGE, SI_ORDER_ANY, 217 arc_free_target_init, NULL); 218 219TUNABLE_QUAD("vfs.zfs.arc_max", &zfs_arc_max); 220TUNABLE_QUAD("vfs.zfs.arc_min", &zfs_arc_min); 221TUNABLE_QUAD("vfs.zfs.arc_meta_limit", &zfs_arc_meta_limit); 222TUNABLE_QUAD("vfs.zfs.arc_average_blocksize", &zfs_arc_average_blocksize); 223SYSCTL_DECL(_vfs_zfs); 224SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_max, CTLFLAG_RDTUN, &zfs_arc_max, 0, 225 "Maximum ARC size"); 226SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_min, CTLFLAG_RDTUN, &zfs_arc_min, 0, 227 "Minimum ARC size"); 228SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_average_blocksize, CTLFLAG_RDTUN, 229 &zfs_arc_average_blocksize, 0, 230 "ARC average blocksize"); 231/* 232 * We don't have a tunable for arc_free_target due to the dependency on 233 * pagedaemon initialisation. 234 */ 235SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_free_target, 236 CTLTYPE_UINT | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, sizeof(u_int), 237 sysctl_vfs_zfs_arc_free_target, "IU", 238 "Desired number of free pages below which ARC triggers reclaim"); 239 240static int 241sysctl_vfs_zfs_arc_free_target(SYSCTL_HANDLER_ARGS) 242{ 243 u_int val; 244 int err; 245 246 val = zfs_arc_free_target; 247 err = sysctl_handle_int(oidp, &val, 0, req); 248 if (err != 0 || req->newptr == NULL) 249 return (err); 250 251 if (val < minfree) 252 return (EINVAL); 253 if (val > cnt.v_page_count) 254 return (EINVAL); 255 256 zfs_arc_free_target = val; 257 258 return (0); 259} 260#endif 261 262/* 263 * Note that buffers can be in one of 6 states: 264 * ARC_anon - anonymous (discussed below) 265 * ARC_mru - recently used, currently cached 266 * ARC_mru_ghost - recentely used, no longer in cache 267 * ARC_mfu - frequently used, currently cached 268 * ARC_mfu_ghost - frequently used, no longer in cache 269 * ARC_l2c_only - exists in L2ARC but not other states 270 * When there are no active references to the buffer, they are 271 * are linked onto a list in one of these arc states. These are 272 * the only buffers that can be evicted or deleted. Within each 273 * state there are multiple lists, one for meta-data and one for 274 * non-meta-data. Meta-data (indirect blocks, blocks of dnodes, 275 * etc.) is tracked separately so that it can be managed more 276 * explicitly: favored over data, limited explicitly. 277 * 278 * Anonymous buffers are buffers that are not associated with 279 * a DVA. These are buffers that hold dirty block copies 280 * before they are written to stable storage. By definition, 281 * they are "ref'd" and are considered part of arc_mru 282 * that cannot be freed. Generally, they will aquire a DVA 283 * as they are written and migrate onto the arc_mru list. 284 * 285 * The ARC_l2c_only state is for buffers that are in the second 286 * level ARC but no longer in any of the ARC_m* lists. The second 287 * level ARC itself may also contain buffers that are in any of 288 * the ARC_m* states - meaning that a buffer can exist in two 289 * places. The reason for the ARC_l2c_only state is to keep the 290 * buffer header in the hash table, so that reads that hit the 291 * second level ARC benefit from these fast lookups. 292 */ 293 294#define ARCS_LOCK_PAD CACHE_LINE_SIZE 295struct arcs_lock { 296 kmutex_t arcs_lock; 297#ifdef _KERNEL 298 unsigned char pad[(ARCS_LOCK_PAD - sizeof (kmutex_t))]; 299#endif 300}; 301 302/* 303 * must be power of two for mask use to work 304 * 305 */ 306#define ARC_BUFC_NUMDATALISTS 16 307#define ARC_BUFC_NUMMETADATALISTS 16 308#define ARC_BUFC_NUMLISTS (ARC_BUFC_NUMMETADATALISTS + ARC_BUFC_NUMDATALISTS) 309 310typedef struct arc_state { 311 uint64_t arcs_lsize[ARC_BUFC_NUMTYPES]; /* amount of evictable data */ 312 uint64_t arcs_size; /* total amount of data in this state */ 313 list_t arcs_lists[ARC_BUFC_NUMLISTS]; /* list of evictable buffers */ 314 struct arcs_lock arcs_locks[ARC_BUFC_NUMLISTS] __aligned(CACHE_LINE_SIZE); 315} arc_state_t; 316 317#define ARCS_LOCK(s, i) (&((s)->arcs_locks[(i)].arcs_lock)) 318 319/* The 6 states: */ 320static arc_state_t ARC_anon; 321static arc_state_t ARC_mru; 322static arc_state_t ARC_mru_ghost; 323static arc_state_t ARC_mfu; 324static arc_state_t ARC_mfu_ghost; 325static arc_state_t ARC_l2c_only; 326 327typedef struct arc_stats { 328 kstat_named_t arcstat_hits; 329 kstat_named_t arcstat_misses; 330 kstat_named_t arcstat_demand_data_hits; 331 kstat_named_t arcstat_demand_data_misses; 332 kstat_named_t arcstat_demand_metadata_hits; 333 kstat_named_t arcstat_demand_metadata_misses; 334 kstat_named_t arcstat_prefetch_data_hits; 335 kstat_named_t arcstat_prefetch_data_misses; 336 kstat_named_t arcstat_prefetch_metadata_hits; 337 kstat_named_t arcstat_prefetch_metadata_misses; 338 kstat_named_t arcstat_mru_hits; 339 kstat_named_t arcstat_mru_ghost_hits; 340 kstat_named_t arcstat_mfu_hits; 341 kstat_named_t arcstat_mfu_ghost_hits; 342 kstat_named_t arcstat_allocated; 343 kstat_named_t arcstat_deleted; 344 kstat_named_t arcstat_stolen; 345 kstat_named_t arcstat_recycle_miss; 346 /* 347 * Number of buffers that could not be evicted because the hash lock 348 * was held by another thread. The lock may not necessarily be held 349 * by something using the same buffer, since hash locks are shared 350 * by multiple buffers. 351 */ 352 kstat_named_t arcstat_mutex_miss; 353 /* 354 * Number of buffers skipped because they have I/O in progress, are 355 * indrect prefetch buffers that have not lived long enough, or are 356 * not from the spa we're trying to evict from. 357 */ 358 kstat_named_t arcstat_evict_skip; 359 kstat_named_t arcstat_evict_l2_cached; 360 kstat_named_t arcstat_evict_l2_eligible; 361 kstat_named_t arcstat_evict_l2_ineligible; 362 kstat_named_t arcstat_hash_elements; 363 kstat_named_t arcstat_hash_elements_max; 364 kstat_named_t arcstat_hash_collisions; 365 kstat_named_t arcstat_hash_chains; 366 kstat_named_t arcstat_hash_chain_max; 367 kstat_named_t arcstat_p; 368 kstat_named_t arcstat_c; 369 kstat_named_t arcstat_c_min; 370 kstat_named_t arcstat_c_max; 371 kstat_named_t arcstat_size; 372 kstat_named_t arcstat_hdr_size; 373 kstat_named_t arcstat_data_size; 374 kstat_named_t arcstat_other_size; 375 kstat_named_t arcstat_l2_hits; 376 kstat_named_t arcstat_l2_misses; 377 kstat_named_t arcstat_l2_feeds; 378 kstat_named_t arcstat_l2_rw_clash; 379 kstat_named_t arcstat_l2_read_bytes; 380 kstat_named_t arcstat_l2_write_bytes; 381 kstat_named_t arcstat_l2_writes_sent; 382 kstat_named_t arcstat_l2_writes_done; 383 kstat_named_t arcstat_l2_writes_error; 384 kstat_named_t arcstat_l2_writes_hdr_miss; 385 kstat_named_t arcstat_l2_evict_lock_retry; 386 kstat_named_t arcstat_l2_evict_reading; 387 kstat_named_t arcstat_l2_free_on_write; 388 kstat_named_t arcstat_l2_abort_lowmem; 389 kstat_named_t arcstat_l2_cksum_bad; 390 kstat_named_t arcstat_l2_io_error; 391 kstat_named_t arcstat_l2_size; 392 kstat_named_t arcstat_l2_asize; 393 kstat_named_t arcstat_l2_hdr_size; 394 kstat_named_t arcstat_l2_compress_successes; 395 kstat_named_t arcstat_l2_compress_zeros; 396 kstat_named_t arcstat_l2_compress_failures; 397 kstat_named_t arcstat_l2_write_trylock_fail; 398 kstat_named_t arcstat_l2_write_passed_headroom; 399 kstat_named_t arcstat_l2_write_spa_mismatch; 400 kstat_named_t arcstat_l2_write_in_l2; 401 kstat_named_t arcstat_l2_write_hdr_io_in_progress; 402 kstat_named_t arcstat_l2_write_not_cacheable; 403 kstat_named_t arcstat_l2_write_full; 404 kstat_named_t arcstat_l2_write_buffer_iter; 405 kstat_named_t arcstat_l2_write_pios; 406 kstat_named_t arcstat_l2_write_buffer_bytes_scanned; 407 kstat_named_t arcstat_l2_write_buffer_list_iter; 408 kstat_named_t arcstat_l2_write_buffer_list_null_iter; 409 kstat_named_t arcstat_memory_throttle_count; 410 kstat_named_t arcstat_duplicate_buffers; 411 kstat_named_t arcstat_duplicate_buffers_size; 412 kstat_named_t arcstat_duplicate_reads; 413} arc_stats_t; 414 415static arc_stats_t arc_stats = { 416 { "hits", KSTAT_DATA_UINT64 }, 417 { "misses", KSTAT_DATA_UINT64 }, 418 { "demand_data_hits", KSTAT_DATA_UINT64 }, 419 { "demand_data_misses", KSTAT_DATA_UINT64 }, 420 { "demand_metadata_hits", KSTAT_DATA_UINT64 }, 421 { "demand_metadata_misses", KSTAT_DATA_UINT64 }, 422 { "prefetch_data_hits", KSTAT_DATA_UINT64 }, 423 { "prefetch_data_misses", KSTAT_DATA_UINT64 }, 424 { "prefetch_metadata_hits", KSTAT_DATA_UINT64 }, 425 { "prefetch_metadata_misses", KSTAT_DATA_UINT64 }, 426 { "mru_hits", KSTAT_DATA_UINT64 }, 427 { "mru_ghost_hits", KSTAT_DATA_UINT64 }, 428 { "mfu_hits", KSTAT_DATA_UINT64 }, 429 { "mfu_ghost_hits", KSTAT_DATA_UINT64 }, 430 { "allocated", KSTAT_DATA_UINT64 }, 431 { "deleted", KSTAT_DATA_UINT64 }, 432 { "stolen", KSTAT_DATA_UINT64 }, 433 { "recycle_miss", KSTAT_DATA_UINT64 }, 434 { "mutex_miss", KSTAT_DATA_UINT64 }, 435 { "evict_skip", KSTAT_DATA_UINT64 }, 436 { "evict_l2_cached", KSTAT_DATA_UINT64 }, 437 { "evict_l2_eligible", KSTAT_DATA_UINT64 }, 438 { "evict_l2_ineligible", KSTAT_DATA_UINT64 }, 439 { "hash_elements", KSTAT_DATA_UINT64 }, 440 { "hash_elements_max", KSTAT_DATA_UINT64 }, 441 { "hash_collisions", KSTAT_DATA_UINT64 }, 442 { "hash_chains", KSTAT_DATA_UINT64 }, 443 { "hash_chain_max", KSTAT_DATA_UINT64 }, 444 { "p", KSTAT_DATA_UINT64 }, 445 { "c", KSTAT_DATA_UINT64 }, 446 { "c_min", KSTAT_DATA_UINT64 }, 447 { "c_max", KSTAT_DATA_UINT64 }, 448 { "size", KSTAT_DATA_UINT64 }, 449 { "hdr_size", KSTAT_DATA_UINT64 }, 450 { "data_size", KSTAT_DATA_UINT64 }, 451 { "other_size", KSTAT_DATA_UINT64 }, 452 { "l2_hits", KSTAT_DATA_UINT64 }, 453 { "l2_misses", KSTAT_DATA_UINT64 }, 454 { "l2_feeds", KSTAT_DATA_UINT64 }, 455 { "l2_rw_clash", KSTAT_DATA_UINT64 }, 456 { "l2_read_bytes", KSTAT_DATA_UINT64 }, 457 { "l2_write_bytes", KSTAT_DATA_UINT64 }, 458 { "l2_writes_sent", KSTAT_DATA_UINT64 }, 459 { "l2_writes_done", KSTAT_DATA_UINT64 }, 460 { "l2_writes_error", KSTAT_DATA_UINT64 }, 461 { "l2_writes_hdr_miss", KSTAT_DATA_UINT64 }, 462 { "l2_evict_lock_retry", KSTAT_DATA_UINT64 }, 463 { "l2_evict_reading", KSTAT_DATA_UINT64 }, 464 { "l2_free_on_write", KSTAT_DATA_UINT64 }, 465 { "l2_abort_lowmem", KSTAT_DATA_UINT64 }, 466 { "l2_cksum_bad", KSTAT_DATA_UINT64 }, 467 { "l2_io_error", KSTAT_DATA_UINT64 }, 468 { "l2_size", KSTAT_DATA_UINT64 }, 469 { "l2_asize", KSTAT_DATA_UINT64 }, 470 { "l2_hdr_size", KSTAT_DATA_UINT64 }, 471 { "l2_compress_successes", KSTAT_DATA_UINT64 }, 472 { "l2_compress_zeros", KSTAT_DATA_UINT64 }, 473 { "l2_compress_failures", KSTAT_DATA_UINT64 }, 474 { "l2_write_trylock_fail", KSTAT_DATA_UINT64 }, 475 { "l2_write_passed_headroom", KSTAT_DATA_UINT64 }, 476 { "l2_write_spa_mismatch", KSTAT_DATA_UINT64 }, 477 { "l2_write_in_l2", KSTAT_DATA_UINT64 }, 478 { "l2_write_io_in_progress", KSTAT_DATA_UINT64 }, 479 { "l2_write_not_cacheable", KSTAT_DATA_UINT64 }, 480 { "l2_write_full", KSTAT_DATA_UINT64 }, 481 { "l2_write_buffer_iter", KSTAT_DATA_UINT64 }, 482 { "l2_write_pios", KSTAT_DATA_UINT64 }, 483 { "l2_write_buffer_bytes_scanned", KSTAT_DATA_UINT64 }, 484 { "l2_write_buffer_list_iter", KSTAT_DATA_UINT64 }, 485 { "l2_write_buffer_list_null_iter", KSTAT_DATA_UINT64 }, 486 { "memory_throttle_count", KSTAT_DATA_UINT64 }, 487 { "duplicate_buffers", KSTAT_DATA_UINT64 }, 488 { "duplicate_buffers_size", KSTAT_DATA_UINT64 }, 489 { "duplicate_reads", KSTAT_DATA_UINT64 } 490}; 491 492#define ARCSTAT(stat) (arc_stats.stat.value.ui64) 493 494#define ARCSTAT_INCR(stat, val) \ 495 atomic_add_64(&arc_stats.stat.value.ui64, (val)) 496 497#define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1) 498#define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1) 499 500#define ARCSTAT_MAX(stat, val) { \ 501 uint64_t m; \ 502 while ((val) > (m = arc_stats.stat.value.ui64) && \ 503 (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \ 504 continue; \ 505} 506 507#define ARCSTAT_MAXSTAT(stat) \ 508 ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64) 509 510/* 511 * We define a macro to allow ARC hits/misses to be easily broken down by 512 * two separate conditions, giving a total of four different subtypes for 513 * each of hits and misses (so eight statistics total). 514 */ 515#define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \ 516 if (cond1) { \ 517 if (cond2) { \ 518 ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \ 519 } else { \ 520 ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \ 521 } \ 522 } else { \ 523 if (cond2) { \ 524 ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \ 525 } else { \ 526 ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\ 527 } \ 528 } 529 530kstat_t *arc_ksp; 531static arc_state_t *arc_anon; 532static arc_state_t *arc_mru; 533static arc_state_t *arc_mru_ghost; 534static arc_state_t *arc_mfu; 535static arc_state_t *arc_mfu_ghost; 536static arc_state_t *arc_l2c_only; 537 538/* 539 * There are several ARC variables that are critical to export as kstats -- 540 * but we don't want to have to grovel around in the kstat whenever we wish to 541 * manipulate them. For these variables, we therefore define them to be in 542 * terms of the statistic variable. This assures that we are not introducing 543 * the possibility of inconsistency by having shadow copies of the variables, 544 * while still allowing the code to be readable. 545 */ 546#define arc_size ARCSTAT(arcstat_size) /* actual total arc size */ 547#define arc_p ARCSTAT(arcstat_p) /* target size of MRU */ 548#define arc_c ARCSTAT(arcstat_c) /* target size of cache */ 549#define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */ 550#define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */ 551 552#define L2ARC_IS_VALID_COMPRESS(_c_) \ 553 ((_c_) == ZIO_COMPRESS_LZ4 || (_c_) == ZIO_COMPRESS_EMPTY) 554 555static int arc_no_grow; /* Don't try to grow cache size */ 556static uint64_t arc_tempreserve; 557static uint64_t arc_loaned_bytes; 558static uint64_t arc_meta_used; 559static uint64_t arc_meta_limit; 560static uint64_t arc_meta_max = 0; 561SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_meta_used, CTLFLAG_RD, &arc_meta_used, 0, 562 "ARC metadata used"); 563SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_meta_limit, CTLFLAG_RW, &arc_meta_limit, 0, 564 "ARC metadata limit"); 565 566typedef struct l2arc_buf_hdr l2arc_buf_hdr_t; 567 568typedef struct arc_callback arc_callback_t; 569 570struct arc_callback { 571 void *acb_private; 572 arc_done_func_t *acb_done; 573 arc_buf_t *acb_buf; 574 zio_t *acb_zio_dummy; 575 arc_callback_t *acb_next; 576}; 577 578typedef struct arc_write_callback arc_write_callback_t; 579 580struct arc_write_callback { 581 void *awcb_private; 582 arc_done_func_t *awcb_ready; 583 arc_done_func_t *awcb_physdone; 584 arc_done_func_t *awcb_done; 585 arc_buf_t *awcb_buf; 586}; 587 588struct arc_buf_hdr { 589 /* protected by hash lock */ 590 dva_t b_dva; 591 uint64_t b_birth; 592 uint64_t b_cksum0; 593 594 kmutex_t b_freeze_lock; 595 zio_cksum_t *b_freeze_cksum; 596 void *b_thawed; 597 598 arc_buf_hdr_t *b_hash_next; 599 arc_buf_t *b_buf; 600 uint32_t b_flags; 601 uint32_t b_datacnt; 602 603 arc_callback_t *b_acb; 604 kcondvar_t b_cv; 605 606 /* immutable */ 607 arc_buf_contents_t b_type; 608 uint64_t b_size; 609 uint64_t b_spa; 610 611 /* protected by arc state mutex */ 612 arc_state_t *b_state; 613 list_node_t b_arc_node; 614 615 /* updated atomically */ 616 clock_t b_arc_access; 617 618 /* self protecting */ 619 refcount_t b_refcnt; 620 621 l2arc_buf_hdr_t *b_l2hdr; 622 list_node_t b_l2node; 623}; 624 625static arc_buf_t *arc_eviction_list; 626static kmutex_t arc_eviction_mtx; 627static arc_buf_hdr_t arc_eviction_hdr; 628static void arc_get_data_buf(arc_buf_t *buf); 629static void arc_access(arc_buf_hdr_t *buf, kmutex_t *hash_lock); 630static int arc_evict_needed(arc_buf_contents_t type); 631static void arc_evict_ghost(arc_state_t *state, uint64_t spa, int64_t bytes); 632#ifdef illumos 633static void arc_buf_watch(arc_buf_t *buf); 634#endif /* illumos */ 635 636static boolean_t l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *ab); 637 638#define GHOST_STATE(state) \ 639 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \ 640 (state) == arc_l2c_only) 641 642/* 643 * Private ARC flags. These flags are private ARC only flags that will show up 644 * in b_flags in the arc_hdr_buf_t. Some flags are publicly declared, and can 645 * be passed in as arc_flags in things like arc_read. However, these flags 646 * should never be passed and should only be set by ARC code. When adding new 647 * public flags, make sure not to smash the private ones. 648 */ 649 650#define ARC_IN_HASH_TABLE (1 << 9) /* this buffer is hashed */ 651#define ARC_IO_IN_PROGRESS (1 << 10) /* I/O in progress for buf */ 652#define ARC_IO_ERROR (1 << 11) /* I/O failed for buf */ 653#define ARC_FREED_IN_READ (1 << 12) /* buf freed while in read */ 654#define ARC_BUF_AVAILABLE (1 << 13) /* block not in active use */ 655#define ARC_INDIRECT (1 << 14) /* this is an indirect block */ 656#define ARC_FREE_IN_PROGRESS (1 << 15) /* hdr about to be freed */ 657#define ARC_L2_WRITING (1 << 16) /* L2ARC write in progress */ 658#define ARC_L2_EVICTED (1 << 17) /* evicted during I/O */ 659#define ARC_L2_WRITE_HEAD (1 << 18) /* head of write list */ 660 661#define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_IN_HASH_TABLE) 662#define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_IO_IN_PROGRESS) 663#define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_IO_ERROR) 664#define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_PREFETCH) 665#define HDR_FREED_IN_READ(hdr) ((hdr)->b_flags & ARC_FREED_IN_READ) 666#define HDR_BUF_AVAILABLE(hdr) ((hdr)->b_flags & ARC_BUF_AVAILABLE) 667#define HDR_FREE_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FREE_IN_PROGRESS) 668#define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_L2CACHE) 669#define HDR_L2_READING(hdr) ((hdr)->b_flags & ARC_IO_IN_PROGRESS && \ 670 (hdr)->b_l2hdr != NULL) 671#define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_L2_WRITING) 672#define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_L2_EVICTED) 673#define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_L2_WRITE_HEAD) 674 675/* 676 * Other sizes 677 */ 678 679#define HDR_SIZE ((int64_t)sizeof (arc_buf_hdr_t)) 680#define L2HDR_SIZE ((int64_t)sizeof (l2arc_buf_hdr_t)) 681 682/* 683 * Hash table routines 684 */ 685 686#define HT_LOCK_PAD CACHE_LINE_SIZE 687 688struct ht_lock { 689 kmutex_t ht_lock; 690#ifdef _KERNEL 691 unsigned char pad[(HT_LOCK_PAD - sizeof (kmutex_t))]; 692#endif 693}; 694 695#define BUF_LOCKS 256 696typedef struct buf_hash_table { 697 uint64_t ht_mask; 698 arc_buf_hdr_t **ht_table; 699 struct ht_lock ht_locks[BUF_LOCKS] __aligned(CACHE_LINE_SIZE); 700} buf_hash_table_t; 701 702static buf_hash_table_t buf_hash_table; 703 704#define BUF_HASH_INDEX(spa, dva, birth) \ 705 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask) 706#define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)]) 707#define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock)) 708#define HDR_LOCK(hdr) \ 709 (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth))) 710 711uint64_t zfs_crc64_table[256]; 712 713/* 714 * Level 2 ARC 715 */ 716 717#define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */ 718#define L2ARC_HEADROOM 2 /* num of writes */ 719/* 720 * If we discover during ARC scan any buffers to be compressed, we boost 721 * our headroom for the next scanning cycle by this percentage multiple. 722 */ 723#define L2ARC_HEADROOM_BOOST 200 724#define L2ARC_FEED_SECS 1 /* caching interval secs */ 725#define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */ 726 727#define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent) 728#define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done) 729 730/* L2ARC Performance Tunables */ 731uint64_t l2arc_write_max = L2ARC_WRITE_SIZE; /* default max write size */ 732uint64_t l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra write during warmup */ 733uint64_t l2arc_headroom = L2ARC_HEADROOM; /* number of dev writes */ 734uint64_t l2arc_headroom_boost = L2ARC_HEADROOM_BOOST; 735uint64_t l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */ 736uint64_t l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval milliseconds */ 737boolean_t l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */ 738boolean_t l2arc_feed_again = B_TRUE; /* turbo warmup */ 739boolean_t l2arc_norw = B_TRUE; /* no reads during writes */ 740 741SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_max, CTLFLAG_RW, 742 &l2arc_write_max, 0, "max write size"); 743SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_boost, CTLFLAG_RW, 744 &l2arc_write_boost, 0, "extra write during warmup"); 745SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_headroom, CTLFLAG_RW, 746 &l2arc_headroom, 0, "number of dev writes"); 747SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_secs, CTLFLAG_RW, 748 &l2arc_feed_secs, 0, "interval seconds"); 749SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_min_ms, CTLFLAG_RW, 750 &l2arc_feed_min_ms, 0, "min interval milliseconds"); 751 752SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_noprefetch, CTLFLAG_RW, 753 &l2arc_noprefetch, 0, "don't cache prefetch bufs"); 754SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_feed_again, CTLFLAG_RW, 755 &l2arc_feed_again, 0, "turbo warmup"); 756SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_norw, CTLFLAG_RW, 757 &l2arc_norw, 0, "no reads during writes"); 758 759SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_size, CTLFLAG_RD, 760 &ARC_anon.arcs_size, 0, "size of anonymous state"); 761SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_metadata_lsize, CTLFLAG_RD, 762 &ARC_anon.arcs_lsize[ARC_BUFC_METADATA], 0, "size of anonymous state"); 763SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_data_lsize, CTLFLAG_RD, 764 &ARC_anon.arcs_lsize[ARC_BUFC_DATA], 0, "size of anonymous state"); 765 766SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_size, CTLFLAG_RD, 767 &ARC_mru.arcs_size, 0, "size of mru state"); 768SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_metadata_lsize, CTLFLAG_RD, 769 &ARC_mru.arcs_lsize[ARC_BUFC_METADATA], 0, "size of metadata in mru state"); 770SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_data_lsize, CTLFLAG_RD, 771 &ARC_mru.arcs_lsize[ARC_BUFC_DATA], 0, "size of data in mru state"); 772 773SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_size, CTLFLAG_RD, 774 &ARC_mru_ghost.arcs_size, 0, "size of mru ghost state"); 775SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_metadata_lsize, CTLFLAG_RD, 776 &ARC_mru_ghost.arcs_lsize[ARC_BUFC_METADATA], 0, 777 "size of metadata in mru ghost state"); 778SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_data_lsize, CTLFLAG_RD, 779 &ARC_mru_ghost.arcs_lsize[ARC_BUFC_DATA], 0, 780 "size of data in mru ghost state"); 781 782SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_size, CTLFLAG_RD, 783 &ARC_mfu.arcs_size, 0, "size of mfu state"); 784SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_metadata_lsize, CTLFLAG_RD, 785 &ARC_mfu.arcs_lsize[ARC_BUFC_METADATA], 0, "size of metadata in mfu state"); 786SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_data_lsize, CTLFLAG_RD, 787 &ARC_mfu.arcs_lsize[ARC_BUFC_DATA], 0, "size of data in mfu state"); 788 789SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_size, CTLFLAG_RD, 790 &ARC_mfu_ghost.arcs_size, 0, "size of mfu ghost state"); 791SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_metadata_lsize, CTLFLAG_RD, 792 &ARC_mfu_ghost.arcs_lsize[ARC_BUFC_METADATA], 0, 793 "size of metadata in mfu ghost state"); 794SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_data_lsize, CTLFLAG_RD, 795 &ARC_mfu_ghost.arcs_lsize[ARC_BUFC_DATA], 0, 796 "size of data in mfu ghost state"); 797 798SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2c_only_size, CTLFLAG_RD, 799 &ARC_l2c_only.arcs_size, 0, "size of mru state"); 800 801/* 802 * L2ARC Internals 803 */ 804typedef struct l2arc_dev { 805 vdev_t *l2ad_vdev; /* vdev */ 806 spa_t *l2ad_spa; /* spa */ 807 uint64_t l2ad_hand; /* next write location */ 808 uint64_t l2ad_start; /* first addr on device */ 809 uint64_t l2ad_end; /* last addr on device */ 810 uint64_t l2ad_evict; /* last addr eviction reached */ 811 boolean_t l2ad_first; /* first sweep through */ 812 boolean_t l2ad_writing; /* currently writing */ 813 list_t *l2ad_buflist; /* buffer list */ 814 list_node_t l2ad_node; /* device list node */ 815} l2arc_dev_t; 816 817static list_t L2ARC_dev_list; /* device list */ 818static list_t *l2arc_dev_list; /* device list pointer */ 819static kmutex_t l2arc_dev_mtx; /* device list mutex */ 820static l2arc_dev_t *l2arc_dev_last; /* last device used */ 821static kmutex_t l2arc_buflist_mtx; /* mutex for all buflists */ 822static list_t L2ARC_free_on_write; /* free after write buf list */ 823static list_t *l2arc_free_on_write; /* free after write list ptr */ 824static kmutex_t l2arc_free_on_write_mtx; /* mutex for list */ 825static uint64_t l2arc_ndev; /* number of devices */ 826 827typedef struct l2arc_read_callback { 828 arc_buf_t *l2rcb_buf; /* read buffer */ 829 spa_t *l2rcb_spa; /* spa */ 830 blkptr_t l2rcb_bp; /* original blkptr */ 831 zbookmark_phys_t l2rcb_zb; /* original bookmark */ 832 int l2rcb_flags; /* original flags */ 833 enum zio_compress l2rcb_compress; /* applied compress */ 834} l2arc_read_callback_t; 835 836typedef struct l2arc_write_callback { 837 l2arc_dev_t *l2wcb_dev; /* device info */ 838 arc_buf_hdr_t *l2wcb_head; /* head of write buflist */ 839} l2arc_write_callback_t; 840 841struct l2arc_buf_hdr { 842 /* protected by arc_buf_hdr mutex */ 843 l2arc_dev_t *b_dev; /* L2ARC device */ 844 uint64_t b_daddr; /* disk address, offset byte */ 845 /* compression applied to buffer data */ 846 enum zio_compress b_compress; 847 /* real alloc'd buffer size depending on b_compress applied */ 848 int b_asize; 849 /* temporary buffer holder for in-flight compressed data */ 850 void *b_tmp_cdata; 851}; 852 853typedef struct l2arc_data_free { 854 /* protected by l2arc_free_on_write_mtx */ 855 void *l2df_data; 856 size_t l2df_size; 857 void (*l2df_func)(void *, size_t); 858 list_node_t l2df_list_node; 859} l2arc_data_free_t; 860 861static kmutex_t l2arc_feed_thr_lock; 862static kcondvar_t l2arc_feed_thr_cv; 863static uint8_t l2arc_thread_exit; 864 865static void l2arc_read_done(zio_t *zio); 866static void l2arc_hdr_stat_add(void); 867static void l2arc_hdr_stat_remove(void); 868 869static boolean_t l2arc_compress_buf(l2arc_buf_hdr_t *l2hdr); 870static void l2arc_decompress_zio(zio_t *zio, arc_buf_hdr_t *hdr, 871 enum zio_compress c); 872static void l2arc_release_cdata_buf(arc_buf_hdr_t *ab); 873 874static uint64_t 875buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth) 876{ 877 uint8_t *vdva = (uint8_t *)dva; 878 uint64_t crc = -1ULL; 879 int i; 880 881 ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY); 882 883 for (i = 0; i < sizeof (dva_t); i++) 884 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ vdva[i]) & 0xFF]; 885 886 crc ^= (spa>>8) ^ birth; 887 888 return (crc); 889} 890 891#define BUF_EMPTY(buf) \ 892 ((buf)->b_dva.dva_word[0] == 0 && \ 893 (buf)->b_dva.dva_word[1] == 0 && \ 894 (buf)->b_cksum0 == 0) 895 896#define BUF_EQUAL(spa, dva, birth, buf) \ 897 ((buf)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \ 898 ((buf)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \ 899 ((buf)->b_birth == birth) && ((buf)->b_spa == spa) 900 901static void 902buf_discard_identity(arc_buf_hdr_t *hdr) 903{ 904 hdr->b_dva.dva_word[0] = 0; 905 hdr->b_dva.dva_word[1] = 0; 906 hdr->b_birth = 0; 907 hdr->b_cksum0 = 0; 908} 909 910static arc_buf_hdr_t * 911buf_hash_find(uint64_t spa, const blkptr_t *bp, kmutex_t **lockp) 912{ 913 const dva_t *dva = BP_IDENTITY(bp); 914 uint64_t birth = BP_PHYSICAL_BIRTH(bp); 915 uint64_t idx = BUF_HASH_INDEX(spa, dva, birth); 916 kmutex_t *hash_lock = BUF_HASH_LOCK(idx); 917 arc_buf_hdr_t *buf; 918 919 mutex_enter(hash_lock); 920 for (buf = buf_hash_table.ht_table[idx]; buf != NULL; 921 buf = buf->b_hash_next) { 922 if (BUF_EQUAL(spa, dva, birth, buf)) { 923 *lockp = hash_lock; 924 return (buf); 925 } 926 } 927 mutex_exit(hash_lock); 928 *lockp = NULL; 929 return (NULL); 930} 931 932/* 933 * Insert an entry into the hash table. If there is already an element 934 * equal to elem in the hash table, then the already existing element 935 * will be returned and the new element will not be inserted. 936 * Otherwise returns NULL. 937 */ 938static arc_buf_hdr_t * 939buf_hash_insert(arc_buf_hdr_t *buf, kmutex_t **lockp) 940{ 941 uint64_t idx = BUF_HASH_INDEX(buf->b_spa, &buf->b_dva, buf->b_birth); 942 kmutex_t *hash_lock = BUF_HASH_LOCK(idx); 943 arc_buf_hdr_t *fbuf; 944 uint32_t i; 945 946 ASSERT(!DVA_IS_EMPTY(&buf->b_dva)); 947 ASSERT(buf->b_birth != 0); 948 ASSERT(!HDR_IN_HASH_TABLE(buf)); 949 *lockp = hash_lock; 950 mutex_enter(hash_lock); 951 for (fbuf = buf_hash_table.ht_table[idx], i = 0; fbuf != NULL; 952 fbuf = fbuf->b_hash_next, i++) { 953 if (BUF_EQUAL(buf->b_spa, &buf->b_dva, buf->b_birth, fbuf)) 954 return (fbuf); 955 } 956 957 buf->b_hash_next = buf_hash_table.ht_table[idx]; 958 buf_hash_table.ht_table[idx] = buf; 959 buf->b_flags |= ARC_IN_HASH_TABLE; 960 961 /* collect some hash table performance data */ 962 if (i > 0) { 963 ARCSTAT_BUMP(arcstat_hash_collisions); 964 if (i == 1) 965 ARCSTAT_BUMP(arcstat_hash_chains); 966 967 ARCSTAT_MAX(arcstat_hash_chain_max, i); 968 } 969 970 ARCSTAT_BUMP(arcstat_hash_elements); 971 ARCSTAT_MAXSTAT(arcstat_hash_elements); 972 973 return (NULL); 974} 975 976static void 977buf_hash_remove(arc_buf_hdr_t *buf) 978{ 979 arc_buf_hdr_t *fbuf, **bufp; 980 uint64_t idx = BUF_HASH_INDEX(buf->b_spa, &buf->b_dva, buf->b_birth); 981 982 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx))); 983 ASSERT(HDR_IN_HASH_TABLE(buf)); 984 985 bufp = &buf_hash_table.ht_table[idx]; 986 while ((fbuf = *bufp) != buf) { 987 ASSERT(fbuf != NULL); 988 bufp = &fbuf->b_hash_next; 989 } 990 *bufp = buf->b_hash_next; 991 buf->b_hash_next = NULL; 992 buf->b_flags &= ~ARC_IN_HASH_TABLE; 993 994 /* collect some hash table performance data */ 995 ARCSTAT_BUMPDOWN(arcstat_hash_elements); 996 997 if (buf_hash_table.ht_table[idx] && 998 buf_hash_table.ht_table[idx]->b_hash_next == NULL) 999 ARCSTAT_BUMPDOWN(arcstat_hash_chains); 1000} 1001 1002/* 1003 * Global data structures and functions for the buf kmem cache. 1004 */ 1005static kmem_cache_t *hdr_cache; 1006static kmem_cache_t *buf_cache; 1007 1008static void 1009buf_fini(void) 1010{ 1011 int i; 1012 1013 kmem_free(buf_hash_table.ht_table, 1014 (buf_hash_table.ht_mask + 1) * sizeof (void *)); 1015 for (i = 0; i < BUF_LOCKS; i++) 1016 mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock); 1017 kmem_cache_destroy(hdr_cache); 1018 kmem_cache_destroy(buf_cache); 1019} 1020 1021/* 1022 * Constructor callback - called when the cache is empty 1023 * and a new buf is requested. 1024 */ 1025/* ARGSUSED */ 1026static int 1027hdr_cons(void *vbuf, void *unused, int kmflag) 1028{ 1029 arc_buf_hdr_t *buf = vbuf; 1030 1031 bzero(buf, sizeof (arc_buf_hdr_t)); 1032 refcount_create(&buf->b_refcnt); 1033 cv_init(&buf->b_cv, NULL, CV_DEFAULT, NULL); 1034 mutex_init(&buf->b_freeze_lock, NULL, MUTEX_DEFAULT, NULL); 1035 arc_space_consume(sizeof (arc_buf_hdr_t), ARC_SPACE_HDRS); 1036 1037 return (0); 1038} 1039 1040/* ARGSUSED */ 1041static int 1042buf_cons(void *vbuf, void *unused, int kmflag) 1043{ 1044 arc_buf_t *buf = vbuf; 1045 1046 bzero(buf, sizeof (arc_buf_t)); 1047 mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL); 1048 arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS); 1049 1050 return (0); 1051} 1052 1053/* 1054 * Destructor callback - called when a cached buf is 1055 * no longer required. 1056 */ 1057/* ARGSUSED */ 1058static void 1059hdr_dest(void *vbuf, void *unused) 1060{ 1061 arc_buf_hdr_t *buf = vbuf; 1062 1063 ASSERT(BUF_EMPTY(buf)); 1064 refcount_destroy(&buf->b_refcnt); 1065 cv_destroy(&buf->b_cv); 1066 mutex_destroy(&buf->b_freeze_lock); 1067 arc_space_return(sizeof (arc_buf_hdr_t), ARC_SPACE_HDRS); 1068} 1069 1070/* ARGSUSED */ 1071static void 1072buf_dest(void *vbuf, void *unused) 1073{ 1074 arc_buf_t *buf = vbuf; 1075 1076 mutex_destroy(&buf->b_evict_lock); 1077 arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS); 1078} 1079 1080/* 1081 * Reclaim callback -- invoked when memory is low. 1082 */ 1083/* ARGSUSED */ 1084static void 1085hdr_recl(void *unused) 1086{ 1087 dprintf("hdr_recl called\n"); 1088 /* 1089 * umem calls the reclaim func when we destroy the buf cache, 1090 * which is after we do arc_fini(). 1091 */ 1092 if (!arc_dead) 1093 cv_signal(&arc_reclaim_thr_cv); 1094} 1095 1096static void 1097buf_init(void) 1098{ 1099 uint64_t *ct; 1100 uint64_t hsize = 1ULL << 12; 1101 int i, j; 1102 1103 /* 1104 * The hash table is big enough to fill all of physical memory 1105 * with an average block size of zfs_arc_average_blocksize (default 8K). 1106 * By default, the table will take up 1107 * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers). 1108 */ 1109 while (hsize * zfs_arc_average_blocksize < (uint64_t)physmem * PAGESIZE) 1110 hsize <<= 1; 1111retry: 1112 buf_hash_table.ht_mask = hsize - 1; 1113 buf_hash_table.ht_table = 1114 kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP); 1115 if (buf_hash_table.ht_table == NULL) { 1116 ASSERT(hsize > (1ULL << 8)); 1117 hsize >>= 1; 1118 goto retry; 1119 } 1120 1121 hdr_cache = kmem_cache_create("arc_buf_hdr_t", sizeof (arc_buf_hdr_t), 1122 0, hdr_cons, hdr_dest, hdr_recl, NULL, NULL, 0); 1123 buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t), 1124 0, buf_cons, buf_dest, NULL, NULL, NULL, 0); 1125 1126 for (i = 0; i < 256; i++) 1127 for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--) 1128 *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY); 1129 1130 for (i = 0; i < BUF_LOCKS; i++) { 1131 mutex_init(&buf_hash_table.ht_locks[i].ht_lock, 1132 NULL, MUTEX_DEFAULT, NULL); 1133 } 1134} 1135 1136#define ARC_MINTIME (hz>>4) /* 62 ms */ 1137 1138static void 1139arc_cksum_verify(arc_buf_t *buf) 1140{ 1141 zio_cksum_t zc; 1142 1143 if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 1144 return; 1145 1146 mutex_enter(&buf->b_hdr->b_freeze_lock); 1147 if (buf->b_hdr->b_freeze_cksum == NULL || 1148 (buf->b_hdr->b_flags & ARC_IO_ERROR)) { 1149 mutex_exit(&buf->b_hdr->b_freeze_lock); 1150 return; 1151 } 1152 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc); 1153 if (!ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc)) 1154 panic("buffer modified while frozen!"); 1155 mutex_exit(&buf->b_hdr->b_freeze_lock); 1156} 1157 1158static int 1159arc_cksum_equal(arc_buf_t *buf) 1160{ 1161 zio_cksum_t zc; 1162 int equal; 1163 1164 mutex_enter(&buf->b_hdr->b_freeze_lock); 1165 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc); 1166 equal = ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc); 1167 mutex_exit(&buf->b_hdr->b_freeze_lock); 1168 1169 return (equal); 1170} 1171 1172static void 1173arc_cksum_compute(arc_buf_t *buf, boolean_t force) 1174{ 1175 if (!force && !(zfs_flags & ZFS_DEBUG_MODIFY)) 1176 return; 1177 1178 mutex_enter(&buf->b_hdr->b_freeze_lock); 1179 if (buf->b_hdr->b_freeze_cksum != NULL) { 1180 mutex_exit(&buf->b_hdr->b_freeze_lock); 1181 return; 1182 } 1183 buf->b_hdr->b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t), KM_SLEEP); 1184 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, 1185 buf->b_hdr->b_freeze_cksum); 1186 mutex_exit(&buf->b_hdr->b_freeze_lock); 1187#ifdef illumos 1188 arc_buf_watch(buf); 1189#endif /* illumos */ 1190} 1191 1192#ifdef illumos 1193#ifndef _KERNEL 1194typedef struct procctl { 1195 long cmd; 1196 prwatch_t prwatch; 1197} procctl_t; 1198#endif 1199 1200/* ARGSUSED */ 1201static void 1202arc_buf_unwatch(arc_buf_t *buf) 1203{ 1204#ifndef _KERNEL 1205 if (arc_watch) { 1206 int result; 1207 procctl_t ctl; 1208 ctl.cmd = PCWATCH; 1209 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data; 1210 ctl.prwatch.pr_size = 0; 1211 ctl.prwatch.pr_wflags = 0; 1212 result = write(arc_procfd, &ctl, sizeof (ctl)); 1213 ASSERT3U(result, ==, sizeof (ctl)); 1214 } 1215#endif 1216} 1217 1218/* ARGSUSED */ 1219static void 1220arc_buf_watch(arc_buf_t *buf) 1221{ 1222#ifndef _KERNEL 1223 if (arc_watch) { 1224 int result; 1225 procctl_t ctl; 1226 ctl.cmd = PCWATCH; 1227 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data; 1228 ctl.prwatch.pr_size = buf->b_hdr->b_size; 1229 ctl.prwatch.pr_wflags = WA_WRITE; 1230 result = write(arc_procfd, &ctl, sizeof (ctl)); 1231 ASSERT3U(result, ==, sizeof (ctl)); 1232 } 1233#endif 1234} 1235#endif /* illumos */ 1236 1237void 1238arc_buf_thaw(arc_buf_t *buf) 1239{ 1240 if (zfs_flags & ZFS_DEBUG_MODIFY) { 1241 if (buf->b_hdr->b_state != arc_anon) 1242 panic("modifying non-anon buffer!"); 1243 if (buf->b_hdr->b_flags & ARC_IO_IN_PROGRESS) 1244 panic("modifying buffer while i/o in progress!"); 1245 arc_cksum_verify(buf); 1246 } 1247 1248 mutex_enter(&buf->b_hdr->b_freeze_lock); 1249 if (buf->b_hdr->b_freeze_cksum != NULL) { 1250 kmem_free(buf->b_hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 1251 buf->b_hdr->b_freeze_cksum = NULL; 1252 } 1253 1254 if (zfs_flags & ZFS_DEBUG_MODIFY) { 1255 if (buf->b_hdr->b_thawed) 1256 kmem_free(buf->b_hdr->b_thawed, 1); 1257 buf->b_hdr->b_thawed = kmem_alloc(1, KM_SLEEP); 1258 } 1259 1260 mutex_exit(&buf->b_hdr->b_freeze_lock); 1261 1262#ifdef illumos 1263 arc_buf_unwatch(buf); 1264#endif /* illumos */ 1265} 1266 1267void 1268arc_buf_freeze(arc_buf_t *buf) 1269{ 1270 kmutex_t *hash_lock; 1271 1272 if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 1273 return; 1274 1275 hash_lock = HDR_LOCK(buf->b_hdr); 1276 mutex_enter(hash_lock); 1277 1278 ASSERT(buf->b_hdr->b_freeze_cksum != NULL || 1279 buf->b_hdr->b_state == arc_anon); 1280 arc_cksum_compute(buf, B_FALSE); 1281 mutex_exit(hash_lock); 1282 1283} 1284 1285static void 1286get_buf_info(arc_buf_hdr_t *ab, arc_state_t *state, list_t **list, kmutex_t **lock) 1287{ 1288 uint64_t buf_hashid = buf_hash(ab->b_spa, &ab->b_dva, ab->b_birth); 1289 1290 if (ab->b_type == ARC_BUFC_METADATA) 1291 buf_hashid &= (ARC_BUFC_NUMMETADATALISTS - 1); 1292 else { 1293 buf_hashid &= (ARC_BUFC_NUMDATALISTS - 1); 1294 buf_hashid += ARC_BUFC_NUMMETADATALISTS; 1295 } 1296 1297 *list = &state->arcs_lists[buf_hashid]; 1298 *lock = ARCS_LOCK(state, buf_hashid); 1299} 1300 1301 1302static void 1303add_reference(arc_buf_hdr_t *ab, kmutex_t *hash_lock, void *tag) 1304{ 1305 ASSERT(MUTEX_HELD(hash_lock)); 1306 1307 if ((refcount_add(&ab->b_refcnt, tag) == 1) && 1308 (ab->b_state != arc_anon)) { 1309 uint64_t delta = ab->b_size * ab->b_datacnt; 1310 uint64_t *size = &ab->b_state->arcs_lsize[ab->b_type]; 1311 list_t *list; 1312 kmutex_t *lock; 1313 1314 get_buf_info(ab, ab->b_state, &list, &lock); 1315 ASSERT(!MUTEX_HELD(lock)); 1316 mutex_enter(lock); 1317 ASSERT(list_link_active(&ab->b_arc_node)); 1318 list_remove(list, ab); 1319 if (GHOST_STATE(ab->b_state)) { 1320 ASSERT0(ab->b_datacnt); 1321 ASSERT3P(ab->b_buf, ==, NULL); 1322 delta = ab->b_size; 1323 } 1324 ASSERT(delta > 0); 1325 ASSERT3U(*size, >=, delta); 1326 atomic_add_64(size, -delta); 1327 mutex_exit(lock); 1328 /* remove the prefetch flag if we get a reference */ 1329 if (ab->b_flags & ARC_PREFETCH) 1330 ab->b_flags &= ~ARC_PREFETCH; 1331 } 1332} 1333 1334static int 1335remove_reference(arc_buf_hdr_t *ab, kmutex_t *hash_lock, void *tag) 1336{ 1337 int cnt; 1338 arc_state_t *state = ab->b_state; 1339 1340 ASSERT(state == arc_anon || MUTEX_HELD(hash_lock)); 1341 ASSERT(!GHOST_STATE(state)); 1342 1343 if (((cnt = refcount_remove(&ab->b_refcnt, tag)) == 0) && 1344 (state != arc_anon)) { 1345 uint64_t *size = &state->arcs_lsize[ab->b_type]; 1346 list_t *list; 1347 kmutex_t *lock; 1348 1349 get_buf_info(ab, state, &list, &lock); 1350 ASSERT(!MUTEX_HELD(lock)); 1351 mutex_enter(lock); 1352 ASSERT(!list_link_active(&ab->b_arc_node)); 1353 list_insert_head(list, ab); 1354 ASSERT(ab->b_datacnt > 0); 1355 atomic_add_64(size, ab->b_size * ab->b_datacnt); 1356 mutex_exit(lock); 1357 } 1358 return (cnt); 1359} 1360 1361/* 1362 * Move the supplied buffer to the indicated state. The mutex 1363 * for the buffer must be held by the caller. 1364 */ 1365static void 1366arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *ab, kmutex_t *hash_lock) 1367{ 1368 arc_state_t *old_state = ab->b_state; 1369 int64_t refcnt = refcount_count(&ab->b_refcnt); 1370 uint64_t from_delta, to_delta; 1371 list_t *list; 1372 kmutex_t *lock; 1373 1374 ASSERT(MUTEX_HELD(hash_lock)); 1375 ASSERT3P(new_state, !=, old_state); 1376 ASSERT(refcnt == 0 || ab->b_datacnt > 0); 1377 ASSERT(ab->b_datacnt == 0 || !GHOST_STATE(new_state)); 1378 ASSERT(ab->b_datacnt <= 1 || old_state != arc_anon); 1379 1380 from_delta = to_delta = ab->b_datacnt * ab->b_size; 1381 1382 /* 1383 * If this buffer is evictable, transfer it from the 1384 * old state list to the new state list. 1385 */ 1386 if (refcnt == 0) { 1387 if (old_state != arc_anon) { 1388 int use_mutex; 1389 uint64_t *size = &old_state->arcs_lsize[ab->b_type]; 1390 1391 get_buf_info(ab, old_state, &list, &lock); 1392 use_mutex = !MUTEX_HELD(lock); 1393 if (use_mutex) 1394 mutex_enter(lock); 1395 1396 ASSERT(list_link_active(&ab->b_arc_node)); 1397 list_remove(list, ab); 1398 1399 /* 1400 * If prefetching out of the ghost cache, 1401 * we will have a non-zero datacnt. 1402 */ 1403 if (GHOST_STATE(old_state) && ab->b_datacnt == 0) { 1404 /* ghost elements have a ghost size */ 1405 ASSERT(ab->b_buf == NULL); 1406 from_delta = ab->b_size; 1407 } 1408 ASSERT3U(*size, >=, from_delta); 1409 atomic_add_64(size, -from_delta); 1410 1411 if (use_mutex) 1412 mutex_exit(lock); 1413 } 1414 if (new_state != arc_anon) { 1415 int use_mutex; 1416 uint64_t *size = &new_state->arcs_lsize[ab->b_type]; 1417 1418 get_buf_info(ab, new_state, &list, &lock); 1419 use_mutex = !MUTEX_HELD(lock); 1420 if (use_mutex) 1421 mutex_enter(lock); 1422 1423 list_insert_head(list, ab); 1424 1425 /* ghost elements have a ghost size */ 1426 if (GHOST_STATE(new_state)) { 1427 ASSERT(ab->b_datacnt == 0); 1428 ASSERT(ab->b_buf == NULL); 1429 to_delta = ab->b_size; 1430 } 1431 atomic_add_64(size, to_delta); 1432 1433 if (use_mutex) 1434 mutex_exit(lock); 1435 } 1436 } 1437 1438 ASSERT(!BUF_EMPTY(ab)); 1439 if (new_state == arc_anon && HDR_IN_HASH_TABLE(ab)) 1440 buf_hash_remove(ab); 1441 1442 /* adjust state sizes */ 1443 if (to_delta) 1444 atomic_add_64(&new_state->arcs_size, to_delta); 1445 if (from_delta) { 1446 ASSERT3U(old_state->arcs_size, >=, from_delta); 1447 atomic_add_64(&old_state->arcs_size, -from_delta); 1448 } 1449 ab->b_state = new_state; 1450 1451 /* adjust l2arc hdr stats */ 1452 if (new_state == arc_l2c_only) 1453 l2arc_hdr_stat_add(); 1454 else if (old_state == arc_l2c_only) 1455 l2arc_hdr_stat_remove(); 1456} 1457 1458void 1459arc_space_consume(uint64_t space, arc_space_type_t type) 1460{ 1461 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES); 1462 1463 switch (type) { 1464 case ARC_SPACE_DATA: 1465 ARCSTAT_INCR(arcstat_data_size, space); 1466 break; 1467 case ARC_SPACE_OTHER: 1468 ARCSTAT_INCR(arcstat_other_size, space); 1469 break; 1470 case ARC_SPACE_HDRS: 1471 ARCSTAT_INCR(arcstat_hdr_size, space); 1472 break; 1473 case ARC_SPACE_L2HDRS: 1474 ARCSTAT_INCR(arcstat_l2_hdr_size, space); 1475 break; 1476 } 1477 1478 atomic_add_64(&arc_meta_used, space); 1479 atomic_add_64(&arc_size, space); 1480} 1481 1482void 1483arc_space_return(uint64_t space, arc_space_type_t type) 1484{ 1485 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES); 1486 1487 switch (type) { 1488 case ARC_SPACE_DATA: 1489 ARCSTAT_INCR(arcstat_data_size, -space); 1490 break; 1491 case ARC_SPACE_OTHER: 1492 ARCSTAT_INCR(arcstat_other_size, -space); 1493 break; 1494 case ARC_SPACE_HDRS: 1495 ARCSTAT_INCR(arcstat_hdr_size, -space); 1496 break; 1497 case ARC_SPACE_L2HDRS: 1498 ARCSTAT_INCR(arcstat_l2_hdr_size, -space); 1499 break; 1500 } 1501 1502 ASSERT(arc_meta_used >= space); 1503 if (arc_meta_max < arc_meta_used) 1504 arc_meta_max = arc_meta_used; 1505 atomic_add_64(&arc_meta_used, -space); 1506 ASSERT(arc_size >= space); 1507 atomic_add_64(&arc_size, -space); 1508} 1509 1510arc_buf_t * 1511arc_buf_alloc(spa_t *spa, int size, void *tag, arc_buf_contents_t type) 1512{ 1513 arc_buf_hdr_t *hdr; 1514 arc_buf_t *buf; 1515 1516 ASSERT3U(size, >, 0); 1517 hdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE); 1518 ASSERT(BUF_EMPTY(hdr)); 1519 hdr->b_size = size; 1520 hdr->b_type = type; 1521 hdr->b_spa = spa_load_guid(spa); 1522 hdr->b_state = arc_anon; 1523 hdr->b_arc_access = 0; 1524 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 1525 buf->b_hdr = hdr; 1526 buf->b_data = NULL; 1527 buf->b_efunc = NULL; 1528 buf->b_private = NULL; 1529 buf->b_next = NULL; 1530 hdr->b_buf = buf; 1531 arc_get_data_buf(buf); 1532 hdr->b_datacnt = 1; 1533 hdr->b_flags = 0; 1534 ASSERT(refcount_is_zero(&hdr->b_refcnt)); 1535 (void) refcount_add(&hdr->b_refcnt, tag); 1536 1537 return (buf); 1538} 1539 1540static char *arc_onloan_tag = "onloan"; 1541 1542/* 1543 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in 1544 * flight data by arc_tempreserve_space() until they are "returned". Loaned 1545 * buffers must be returned to the arc before they can be used by the DMU or 1546 * freed. 1547 */ 1548arc_buf_t * 1549arc_loan_buf(spa_t *spa, int size) 1550{ 1551 arc_buf_t *buf; 1552 1553 buf = arc_buf_alloc(spa, size, arc_onloan_tag, ARC_BUFC_DATA); 1554 1555 atomic_add_64(&arc_loaned_bytes, size); 1556 return (buf); 1557} 1558 1559/* 1560 * Return a loaned arc buffer to the arc. 1561 */ 1562void 1563arc_return_buf(arc_buf_t *buf, void *tag) 1564{ 1565 arc_buf_hdr_t *hdr = buf->b_hdr; 1566 1567 ASSERT(buf->b_data != NULL); 1568 (void) refcount_add(&hdr->b_refcnt, tag); 1569 (void) refcount_remove(&hdr->b_refcnt, arc_onloan_tag); 1570 1571 atomic_add_64(&arc_loaned_bytes, -hdr->b_size); 1572} 1573 1574/* Detach an arc_buf from a dbuf (tag) */ 1575void 1576arc_loan_inuse_buf(arc_buf_t *buf, void *tag) 1577{ 1578 arc_buf_hdr_t *hdr; 1579 1580 ASSERT(buf->b_data != NULL); 1581 hdr = buf->b_hdr; 1582 (void) refcount_add(&hdr->b_refcnt, arc_onloan_tag); 1583 (void) refcount_remove(&hdr->b_refcnt, tag); 1584 buf->b_efunc = NULL; 1585 buf->b_private = NULL; 1586 1587 atomic_add_64(&arc_loaned_bytes, hdr->b_size); 1588} 1589 1590static arc_buf_t * 1591arc_buf_clone(arc_buf_t *from) 1592{ 1593 arc_buf_t *buf; 1594 arc_buf_hdr_t *hdr = from->b_hdr; 1595 uint64_t size = hdr->b_size; 1596 1597 ASSERT(hdr->b_state != arc_anon); 1598 1599 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 1600 buf->b_hdr = hdr; 1601 buf->b_data = NULL; 1602 buf->b_efunc = NULL; 1603 buf->b_private = NULL; 1604 buf->b_next = hdr->b_buf; 1605 hdr->b_buf = buf; 1606 arc_get_data_buf(buf); 1607 bcopy(from->b_data, buf->b_data, size); 1608 1609 /* 1610 * This buffer already exists in the arc so create a duplicate 1611 * copy for the caller. If the buffer is associated with user data 1612 * then track the size and number of duplicates. These stats will be 1613 * updated as duplicate buffers are created and destroyed. 1614 */ 1615 if (hdr->b_type == ARC_BUFC_DATA) { 1616 ARCSTAT_BUMP(arcstat_duplicate_buffers); 1617 ARCSTAT_INCR(arcstat_duplicate_buffers_size, size); 1618 } 1619 hdr->b_datacnt += 1; 1620 return (buf); 1621} 1622 1623void 1624arc_buf_add_ref(arc_buf_t *buf, void* tag) 1625{ 1626 arc_buf_hdr_t *hdr; 1627 kmutex_t *hash_lock; 1628 1629 /* 1630 * Check to see if this buffer is evicted. Callers 1631 * must verify b_data != NULL to know if the add_ref 1632 * was successful. 1633 */ 1634 mutex_enter(&buf->b_evict_lock); 1635 if (buf->b_data == NULL) { 1636 mutex_exit(&buf->b_evict_lock); 1637 return; 1638 } 1639 hash_lock = HDR_LOCK(buf->b_hdr); 1640 mutex_enter(hash_lock); 1641 hdr = buf->b_hdr; 1642 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 1643 mutex_exit(&buf->b_evict_lock); 1644 1645 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu); 1646 add_reference(hdr, hash_lock, tag); 1647 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr); 1648 arc_access(hdr, hash_lock); 1649 mutex_exit(hash_lock); 1650 ARCSTAT_BUMP(arcstat_hits); 1651 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH), 1652 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA, 1653 data, metadata, hits); 1654} 1655 1656/* 1657 * Free the arc data buffer. If it is an l2arc write in progress, 1658 * the buffer is placed on l2arc_free_on_write to be freed later. 1659 */ 1660static void 1661arc_buf_data_free(arc_buf_t *buf, void (*free_func)(void *, size_t)) 1662{ 1663 arc_buf_hdr_t *hdr = buf->b_hdr; 1664 1665 if (HDR_L2_WRITING(hdr)) { 1666 l2arc_data_free_t *df; 1667 df = kmem_alloc(sizeof (l2arc_data_free_t), KM_SLEEP); 1668 df->l2df_data = buf->b_data; 1669 df->l2df_size = hdr->b_size; 1670 df->l2df_func = free_func; 1671 mutex_enter(&l2arc_free_on_write_mtx); 1672 list_insert_head(l2arc_free_on_write, df); 1673 mutex_exit(&l2arc_free_on_write_mtx); 1674 ARCSTAT_BUMP(arcstat_l2_free_on_write); 1675 } else { 1676 free_func(buf->b_data, hdr->b_size); 1677 } 1678} 1679 1680/* 1681 * Free up buf->b_data and if 'remove' is set, then pull the 1682 * arc_buf_t off of the the arc_buf_hdr_t's list and free it. 1683 */ 1684static void 1685arc_buf_destroy(arc_buf_t *buf, boolean_t recycle, boolean_t remove) 1686{ 1687 arc_buf_t **bufp; 1688 1689 /* free up data associated with the buf */ 1690 if (buf->b_data) { 1691 arc_state_t *state = buf->b_hdr->b_state; 1692 uint64_t size = buf->b_hdr->b_size; 1693 arc_buf_contents_t type = buf->b_hdr->b_type; 1694 1695 arc_cksum_verify(buf); 1696#ifdef illumos 1697 arc_buf_unwatch(buf); 1698#endif /* illumos */ 1699 1700 if (!recycle) { 1701 if (type == ARC_BUFC_METADATA) { 1702 arc_buf_data_free(buf, zio_buf_free); 1703 arc_space_return(size, ARC_SPACE_DATA); 1704 } else { 1705 ASSERT(type == ARC_BUFC_DATA); 1706 arc_buf_data_free(buf, zio_data_buf_free); 1707 ARCSTAT_INCR(arcstat_data_size, -size); 1708 atomic_add_64(&arc_size, -size); 1709 } 1710 } 1711 if (list_link_active(&buf->b_hdr->b_arc_node)) { 1712 uint64_t *cnt = &state->arcs_lsize[type]; 1713 1714 ASSERT(refcount_is_zero(&buf->b_hdr->b_refcnt)); 1715 ASSERT(state != arc_anon); 1716 1717 ASSERT3U(*cnt, >=, size); 1718 atomic_add_64(cnt, -size); 1719 } 1720 ASSERT3U(state->arcs_size, >=, size); 1721 atomic_add_64(&state->arcs_size, -size); 1722 buf->b_data = NULL; 1723 1724 /* 1725 * If we're destroying a duplicate buffer make sure 1726 * that the appropriate statistics are updated. 1727 */ 1728 if (buf->b_hdr->b_datacnt > 1 && 1729 buf->b_hdr->b_type == ARC_BUFC_DATA) { 1730 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers); 1731 ARCSTAT_INCR(arcstat_duplicate_buffers_size, -size); 1732 } 1733 ASSERT(buf->b_hdr->b_datacnt > 0); 1734 buf->b_hdr->b_datacnt -= 1; 1735 } 1736 1737 /* only remove the buf if requested */ 1738 if (!remove) 1739 return; 1740 1741 /* remove the buf from the hdr list */ 1742 for (bufp = &buf->b_hdr->b_buf; *bufp != buf; bufp = &(*bufp)->b_next) 1743 continue; 1744 *bufp = buf->b_next; 1745 buf->b_next = NULL; 1746 1747 ASSERT(buf->b_efunc == NULL); 1748 1749 /* clean up the buf */ 1750 buf->b_hdr = NULL; 1751 kmem_cache_free(buf_cache, buf); 1752} 1753 1754static void 1755arc_hdr_destroy(arc_buf_hdr_t *hdr) 1756{ 1757 ASSERT(refcount_is_zero(&hdr->b_refcnt)); 1758 ASSERT3P(hdr->b_state, ==, arc_anon); 1759 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 1760 l2arc_buf_hdr_t *l2hdr = hdr->b_l2hdr; 1761 1762 if (l2hdr != NULL) { 1763 boolean_t buflist_held = MUTEX_HELD(&l2arc_buflist_mtx); 1764 /* 1765 * To prevent arc_free() and l2arc_evict() from 1766 * attempting to free the same buffer at the same time, 1767 * a FREE_IN_PROGRESS flag is given to arc_free() to 1768 * give it priority. l2arc_evict() can't destroy this 1769 * header while we are waiting on l2arc_buflist_mtx. 1770 * 1771 * The hdr may be removed from l2ad_buflist before we 1772 * grab l2arc_buflist_mtx, so b_l2hdr is rechecked. 1773 */ 1774 if (!buflist_held) { 1775 mutex_enter(&l2arc_buflist_mtx); 1776 l2hdr = hdr->b_l2hdr; 1777 } 1778 1779 if (l2hdr != NULL) { 1780 trim_map_free(l2hdr->b_dev->l2ad_vdev, l2hdr->b_daddr, 1781 hdr->b_size, 0); 1782 list_remove(l2hdr->b_dev->l2ad_buflist, hdr); 1783 ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size); 1784 ARCSTAT_INCR(arcstat_l2_asize, -l2hdr->b_asize); 1785 vdev_space_update(l2hdr->b_dev->l2ad_vdev, 1786 -l2hdr->b_asize, 0, 0); 1787 kmem_free(l2hdr, sizeof (l2arc_buf_hdr_t)); 1788 if (hdr->b_state == arc_l2c_only) 1789 l2arc_hdr_stat_remove(); 1790 hdr->b_l2hdr = NULL; 1791 } 1792 1793 if (!buflist_held) 1794 mutex_exit(&l2arc_buflist_mtx); 1795 } 1796 1797 if (!BUF_EMPTY(hdr)) { 1798 ASSERT(!HDR_IN_HASH_TABLE(hdr)); 1799 buf_discard_identity(hdr); 1800 } 1801 while (hdr->b_buf) { 1802 arc_buf_t *buf = hdr->b_buf; 1803 1804 if (buf->b_efunc) { 1805 mutex_enter(&arc_eviction_mtx); 1806 mutex_enter(&buf->b_evict_lock); 1807 ASSERT(buf->b_hdr != NULL); 1808 arc_buf_destroy(hdr->b_buf, FALSE, FALSE); 1809 hdr->b_buf = buf->b_next; 1810 buf->b_hdr = &arc_eviction_hdr; 1811 buf->b_next = arc_eviction_list; 1812 arc_eviction_list = buf; 1813 mutex_exit(&buf->b_evict_lock); 1814 mutex_exit(&arc_eviction_mtx); 1815 } else { 1816 arc_buf_destroy(hdr->b_buf, FALSE, TRUE); 1817 } 1818 } 1819 if (hdr->b_freeze_cksum != NULL) { 1820 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 1821 hdr->b_freeze_cksum = NULL; 1822 } 1823 if (hdr->b_thawed) { 1824 kmem_free(hdr->b_thawed, 1); 1825 hdr->b_thawed = NULL; 1826 } 1827 1828 ASSERT(!list_link_active(&hdr->b_arc_node)); 1829 ASSERT3P(hdr->b_hash_next, ==, NULL); 1830 ASSERT3P(hdr->b_acb, ==, NULL); 1831 kmem_cache_free(hdr_cache, hdr); 1832} 1833 1834void 1835arc_buf_free(arc_buf_t *buf, void *tag) 1836{ 1837 arc_buf_hdr_t *hdr = buf->b_hdr; 1838 int hashed = hdr->b_state != arc_anon; 1839 1840 ASSERT(buf->b_efunc == NULL); 1841 ASSERT(buf->b_data != NULL); 1842 1843 if (hashed) { 1844 kmutex_t *hash_lock = HDR_LOCK(hdr); 1845 1846 mutex_enter(hash_lock); 1847 hdr = buf->b_hdr; 1848 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 1849 1850 (void) remove_reference(hdr, hash_lock, tag); 1851 if (hdr->b_datacnt > 1) { 1852 arc_buf_destroy(buf, FALSE, TRUE); 1853 } else { 1854 ASSERT(buf == hdr->b_buf); 1855 ASSERT(buf->b_efunc == NULL); 1856 hdr->b_flags |= ARC_BUF_AVAILABLE; 1857 } 1858 mutex_exit(hash_lock); 1859 } else if (HDR_IO_IN_PROGRESS(hdr)) { 1860 int destroy_hdr; 1861 /* 1862 * We are in the middle of an async write. Don't destroy 1863 * this buffer unless the write completes before we finish 1864 * decrementing the reference count. 1865 */ 1866 mutex_enter(&arc_eviction_mtx); 1867 (void) remove_reference(hdr, NULL, tag); 1868 ASSERT(refcount_is_zero(&hdr->b_refcnt)); 1869 destroy_hdr = !HDR_IO_IN_PROGRESS(hdr); 1870 mutex_exit(&arc_eviction_mtx); 1871 if (destroy_hdr) 1872 arc_hdr_destroy(hdr); 1873 } else { 1874 if (remove_reference(hdr, NULL, tag) > 0) 1875 arc_buf_destroy(buf, FALSE, TRUE); 1876 else 1877 arc_hdr_destroy(hdr); 1878 } 1879} 1880 1881boolean_t 1882arc_buf_remove_ref(arc_buf_t *buf, void* tag) 1883{ 1884 arc_buf_hdr_t *hdr = buf->b_hdr; 1885 kmutex_t *hash_lock = HDR_LOCK(hdr); 1886 boolean_t no_callback = (buf->b_efunc == NULL); 1887 1888 if (hdr->b_state == arc_anon) { 1889 ASSERT(hdr->b_datacnt == 1); 1890 arc_buf_free(buf, tag); 1891 return (no_callback); 1892 } 1893 1894 mutex_enter(hash_lock); 1895 hdr = buf->b_hdr; 1896 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 1897 ASSERT(hdr->b_state != arc_anon); 1898 ASSERT(buf->b_data != NULL); 1899 1900 (void) remove_reference(hdr, hash_lock, tag); 1901 if (hdr->b_datacnt > 1) { 1902 if (no_callback) 1903 arc_buf_destroy(buf, FALSE, TRUE); 1904 } else if (no_callback) { 1905 ASSERT(hdr->b_buf == buf && buf->b_next == NULL); 1906 ASSERT(buf->b_efunc == NULL); 1907 hdr->b_flags |= ARC_BUF_AVAILABLE; 1908 } 1909 ASSERT(no_callback || hdr->b_datacnt > 1 || 1910 refcount_is_zero(&hdr->b_refcnt)); 1911 mutex_exit(hash_lock); 1912 return (no_callback); 1913} 1914 1915int 1916arc_buf_size(arc_buf_t *buf) 1917{ 1918 return (buf->b_hdr->b_size); 1919} 1920 1921/* 1922 * Called from the DMU to determine if the current buffer should be 1923 * evicted. In order to ensure proper locking, the eviction must be initiated 1924 * from the DMU. Return true if the buffer is associated with user data and 1925 * duplicate buffers still exist. 1926 */ 1927boolean_t 1928arc_buf_eviction_needed(arc_buf_t *buf) 1929{ 1930 arc_buf_hdr_t *hdr; 1931 boolean_t evict_needed = B_FALSE; 1932 1933 if (zfs_disable_dup_eviction) 1934 return (B_FALSE); 1935 1936 mutex_enter(&buf->b_evict_lock); 1937 hdr = buf->b_hdr; 1938 if (hdr == NULL) { 1939 /* 1940 * We are in arc_do_user_evicts(); let that function 1941 * perform the eviction. 1942 */ 1943 ASSERT(buf->b_data == NULL); 1944 mutex_exit(&buf->b_evict_lock); 1945 return (B_FALSE); 1946 } else if (buf->b_data == NULL) { 1947 /* 1948 * We have already been added to the arc eviction list; 1949 * recommend eviction. 1950 */ 1951 ASSERT3P(hdr, ==, &arc_eviction_hdr); 1952 mutex_exit(&buf->b_evict_lock); 1953 return (B_TRUE); 1954 } 1955 1956 if (hdr->b_datacnt > 1 && hdr->b_type == ARC_BUFC_DATA) 1957 evict_needed = B_TRUE; 1958 1959 mutex_exit(&buf->b_evict_lock); 1960 return (evict_needed); 1961} 1962 1963/* 1964 * Evict buffers from list until we've removed the specified number of 1965 * bytes. Move the removed buffers to the appropriate evict state. 1966 * If the recycle flag is set, then attempt to "recycle" a buffer: 1967 * - look for a buffer to evict that is `bytes' long. 1968 * - return the data block from this buffer rather than freeing it. 1969 * This flag is used by callers that are trying to make space for a 1970 * new buffer in a full arc cache. 1971 * 1972 * This function makes a "best effort". It skips over any buffers 1973 * it can't get a hash_lock on, and so may not catch all candidates. 1974 * It may also return without evicting as much space as requested. 1975 */ 1976static void * 1977arc_evict(arc_state_t *state, uint64_t spa, int64_t bytes, boolean_t recycle, 1978 arc_buf_contents_t type) 1979{ 1980 arc_state_t *evicted_state; 1981 uint64_t bytes_evicted = 0, skipped = 0, missed = 0; 1982 int64_t bytes_remaining; 1983 arc_buf_hdr_t *ab, *ab_prev = NULL; 1984 list_t *evicted_list, *list, *evicted_list_start, *list_start; 1985 kmutex_t *lock, *evicted_lock; 1986 kmutex_t *hash_lock; 1987 boolean_t have_lock; 1988 void *stolen = NULL; 1989 arc_buf_hdr_t marker = { 0 }; 1990 int count = 0; 1991 static int evict_metadata_offset, evict_data_offset; 1992 int i, idx, offset, list_count, lists; 1993 1994 ASSERT(state == arc_mru || state == arc_mfu); 1995 1996 evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost; 1997 1998 if (type == ARC_BUFC_METADATA) { 1999 offset = 0; 2000 list_count = ARC_BUFC_NUMMETADATALISTS; 2001 list_start = &state->arcs_lists[0]; 2002 evicted_list_start = &evicted_state->arcs_lists[0]; 2003 idx = evict_metadata_offset; 2004 } else { 2005 offset = ARC_BUFC_NUMMETADATALISTS; 2006 list_start = &state->arcs_lists[offset]; 2007 evicted_list_start = &evicted_state->arcs_lists[offset]; 2008 list_count = ARC_BUFC_NUMDATALISTS; 2009 idx = evict_data_offset; 2010 } 2011 bytes_remaining = evicted_state->arcs_lsize[type]; 2012 lists = 0; 2013 2014evict_start: 2015 list = &list_start[idx]; 2016 evicted_list = &evicted_list_start[idx]; 2017 lock = ARCS_LOCK(state, (offset + idx)); 2018 evicted_lock = ARCS_LOCK(evicted_state, (offset + idx)); 2019 2020 mutex_enter(lock); 2021 mutex_enter(evicted_lock); 2022 2023 for (ab = list_tail(list); ab; ab = ab_prev) { 2024 ab_prev = list_prev(list, ab); 2025 bytes_remaining -= (ab->b_size * ab->b_datacnt); 2026 /* prefetch buffers have a minimum lifespan */ 2027 if (HDR_IO_IN_PROGRESS(ab) || 2028 (spa && ab->b_spa != spa) || 2029 (ab->b_flags & (ARC_PREFETCH|ARC_INDIRECT) && 2030 ddi_get_lbolt() - ab->b_arc_access < 2031 arc_min_prefetch_lifespan)) { 2032 skipped++; 2033 continue; 2034 } 2035 /* "lookahead" for better eviction candidate */ 2036 if (recycle && ab->b_size != bytes && 2037 ab_prev && ab_prev->b_size == bytes) 2038 continue; 2039 2040 /* ignore markers */ 2041 if (ab->b_spa == 0) 2042 continue; 2043 2044 /* 2045 * It may take a long time to evict all the bufs requested. 2046 * To avoid blocking all arc activity, periodically drop 2047 * the arcs_mtx and give other threads a chance to run 2048 * before reacquiring the lock. 2049 * 2050 * If we are looking for a buffer to recycle, we are in 2051 * the hot code path, so don't sleep. 2052 */ 2053 if (!recycle && count++ > arc_evict_iterations) { 2054 list_insert_after(list, ab, &marker); 2055 mutex_exit(evicted_lock); 2056 mutex_exit(lock); 2057 kpreempt(KPREEMPT_SYNC); 2058 mutex_enter(lock); 2059 mutex_enter(evicted_lock); 2060 ab_prev = list_prev(list, &marker); 2061 list_remove(list, &marker); 2062 count = 0; 2063 continue; 2064 } 2065 2066 hash_lock = HDR_LOCK(ab); 2067 have_lock = MUTEX_HELD(hash_lock); 2068 if (have_lock || mutex_tryenter(hash_lock)) { 2069 ASSERT0(refcount_count(&ab->b_refcnt)); 2070 ASSERT(ab->b_datacnt > 0); 2071 while (ab->b_buf) { 2072 arc_buf_t *buf = ab->b_buf; 2073 if (!mutex_tryenter(&buf->b_evict_lock)) { 2074 missed += 1; 2075 break; 2076 } 2077 if (buf->b_data) { 2078 bytes_evicted += ab->b_size; 2079 if (recycle && ab->b_type == type && 2080 ab->b_size == bytes && 2081 !HDR_L2_WRITING(ab)) { 2082 stolen = buf->b_data; 2083 recycle = FALSE; 2084 } 2085 } 2086 if (buf->b_efunc) { 2087 mutex_enter(&arc_eviction_mtx); 2088 arc_buf_destroy(buf, 2089 buf->b_data == stolen, FALSE); 2090 ab->b_buf = buf->b_next; 2091 buf->b_hdr = &arc_eviction_hdr; 2092 buf->b_next = arc_eviction_list; 2093 arc_eviction_list = buf; 2094 mutex_exit(&arc_eviction_mtx); 2095 mutex_exit(&buf->b_evict_lock); 2096 } else { 2097 mutex_exit(&buf->b_evict_lock); 2098 arc_buf_destroy(buf, 2099 buf->b_data == stolen, TRUE); 2100 } 2101 } 2102 2103 if (ab->b_l2hdr) { 2104 ARCSTAT_INCR(arcstat_evict_l2_cached, 2105 ab->b_size); 2106 } else { 2107 if (l2arc_write_eligible(ab->b_spa, ab)) { 2108 ARCSTAT_INCR(arcstat_evict_l2_eligible, 2109 ab->b_size); 2110 } else { 2111 ARCSTAT_INCR( 2112 arcstat_evict_l2_ineligible, 2113 ab->b_size); 2114 } 2115 } 2116 2117 if (ab->b_datacnt == 0) { 2118 arc_change_state(evicted_state, ab, hash_lock); 2119 ASSERT(HDR_IN_HASH_TABLE(ab)); 2120 ab->b_flags |= ARC_IN_HASH_TABLE; 2121 ab->b_flags &= ~ARC_BUF_AVAILABLE; 2122 DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, ab); 2123 } 2124 if (!have_lock) 2125 mutex_exit(hash_lock); 2126 if (bytes >= 0 && bytes_evicted >= bytes) 2127 break; 2128 if (bytes_remaining > 0) { 2129 mutex_exit(evicted_lock); 2130 mutex_exit(lock); 2131 idx = ((idx + 1) & (list_count - 1)); 2132 lists++; 2133 goto evict_start; 2134 } 2135 } else { 2136 missed += 1; 2137 } 2138 } 2139 2140 mutex_exit(evicted_lock); 2141 mutex_exit(lock); 2142 2143 idx = ((idx + 1) & (list_count - 1)); 2144 lists++; 2145 2146 if (bytes_evicted < bytes) { 2147 if (lists < list_count) 2148 goto evict_start; 2149 else 2150 dprintf("only evicted %lld bytes from %x", 2151 (longlong_t)bytes_evicted, state); 2152 } 2153 if (type == ARC_BUFC_METADATA) 2154 evict_metadata_offset = idx; 2155 else 2156 evict_data_offset = idx; 2157 2158 if (skipped) 2159 ARCSTAT_INCR(arcstat_evict_skip, skipped); 2160 2161 if (missed) 2162 ARCSTAT_INCR(arcstat_mutex_miss, missed); 2163 2164 /* 2165 * Note: we have just evicted some data into the ghost state, 2166 * potentially putting the ghost size over the desired size. Rather 2167 * that evicting from the ghost list in this hot code path, leave 2168 * this chore to the arc_reclaim_thread(). 2169 */ 2170 2171 if (stolen) 2172 ARCSTAT_BUMP(arcstat_stolen); 2173 return (stolen); 2174} 2175 2176/* 2177 * Remove buffers from list until we've removed the specified number of 2178 * bytes. Destroy the buffers that are removed. 2179 */ 2180static void 2181arc_evict_ghost(arc_state_t *state, uint64_t spa, int64_t bytes) 2182{ 2183 arc_buf_hdr_t *ab, *ab_prev; 2184 arc_buf_hdr_t marker = { 0 }; 2185 list_t *list, *list_start; 2186 kmutex_t *hash_lock, *lock; 2187 uint64_t bytes_deleted = 0; 2188 uint64_t bufs_skipped = 0; 2189 int count = 0; 2190 static int evict_offset; 2191 int list_count, idx = evict_offset; 2192 int offset, lists = 0; 2193 2194 ASSERT(GHOST_STATE(state)); 2195 2196 /* 2197 * data lists come after metadata lists 2198 */ 2199 list_start = &state->arcs_lists[ARC_BUFC_NUMMETADATALISTS]; 2200 list_count = ARC_BUFC_NUMDATALISTS; 2201 offset = ARC_BUFC_NUMMETADATALISTS; 2202 2203evict_start: 2204 list = &list_start[idx]; 2205 lock = ARCS_LOCK(state, idx + offset); 2206 2207 mutex_enter(lock); 2208 for (ab = list_tail(list); ab; ab = ab_prev) { 2209 ab_prev = list_prev(list, ab); 2210 if (ab->b_type > ARC_BUFC_NUMTYPES) 2211 panic("invalid ab=%p", (void *)ab); 2212 if (spa && ab->b_spa != spa) 2213 continue; 2214 2215 /* ignore markers */ 2216 if (ab->b_spa == 0) 2217 continue; 2218 2219 hash_lock = HDR_LOCK(ab); 2220 /* caller may be trying to modify this buffer, skip it */ 2221 if (MUTEX_HELD(hash_lock)) 2222 continue; 2223 2224 /* 2225 * It may take a long time to evict all the bufs requested. 2226 * To avoid blocking all arc activity, periodically drop 2227 * the arcs_mtx and give other threads a chance to run 2228 * before reacquiring the lock. 2229 */ 2230 if (count++ > arc_evict_iterations) { 2231 list_insert_after(list, ab, &marker); 2232 mutex_exit(lock); 2233 kpreempt(KPREEMPT_SYNC); 2234 mutex_enter(lock); 2235 ab_prev = list_prev(list, &marker); 2236 list_remove(list, &marker); 2237 count = 0; 2238 continue; 2239 } 2240 if (mutex_tryenter(hash_lock)) { 2241 ASSERT(!HDR_IO_IN_PROGRESS(ab)); 2242 ASSERT(ab->b_buf == NULL); 2243 ARCSTAT_BUMP(arcstat_deleted); 2244 bytes_deleted += ab->b_size; 2245 2246 if (ab->b_l2hdr != NULL) { 2247 /* 2248 * This buffer is cached on the 2nd Level ARC; 2249 * don't destroy the header. 2250 */ 2251 arc_change_state(arc_l2c_only, ab, hash_lock); 2252 mutex_exit(hash_lock); 2253 } else { 2254 arc_change_state(arc_anon, ab, hash_lock); 2255 mutex_exit(hash_lock); 2256 arc_hdr_destroy(ab); 2257 } 2258 2259 DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, ab); 2260 if (bytes >= 0 && bytes_deleted >= bytes) 2261 break; 2262 } else if (bytes < 0) { 2263 /* 2264 * Insert a list marker and then wait for the 2265 * hash lock to become available. Once its 2266 * available, restart from where we left off. 2267 */ 2268 list_insert_after(list, ab, &marker); 2269 mutex_exit(lock); 2270 mutex_enter(hash_lock); 2271 mutex_exit(hash_lock); 2272 mutex_enter(lock); 2273 ab_prev = list_prev(list, &marker); 2274 list_remove(list, &marker); 2275 } else { 2276 bufs_skipped += 1; 2277 } 2278 2279 } 2280 mutex_exit(lock); 2281 idx = ((idx + 1) & (ARC_BUFC_NUMDATALISTS - 1)); 2282 lists++; 2283 2284 if (lists < list_count) 2285 goto evict_start; 2286 2287 evict_offset = idx; 2288 if ((uintptr_t)list > (uintptr_t)&state->arcs_lists[ARC_BUFC_NUMMETADATALISTS] && 2289 (bytes < 0 || bytes_deleted < bytes)) { 2290 list_start = &state->arcs_lists[0]; 2291 list_count = ARC_BUFC_NUMMETADATALISTS; 2292 offset = lists = 0; 2293 goto evict_start; 2294 } 2295 2296 if (bufs_skipped) { 2297 ARCSTAT_INCR(arcstat_mutex_miss, bufs_skipped); 2298 ASSERT(bytes >= 0); 2299 } 2300 2301 if (bytes_deleted < bytes) 2302 dprintf("only deleted %lld bytes from %p", 2303 (longlong_t)bytes_deleted, state); 2304} 2305 2306static void 2307arc_adjust(void) 2308{ 2309 int64_t adjustment, delta; 2310 2311 /* 2312 * Adjust MRU size 2313 */ 2314 2315 adjustment = MIN((int64_t)(arc_size - arc_c), 2316 (int64_t)(arc_anon->arcs_size + arc_mru->arcs_size + arc_meta_used - 2317 arc_p)); 2318 2319 if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_DATA] > 0) { 2320 delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_DATA], adjustment); 2321 (void) arc_evict(arc_mru, 0, delta, FALSE, ARC_BUFC_DATA); 2322 adjustment -= delta; 2323 } 2324 2325 if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_METADATA] > 0) { 2326 delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_METADATA], adjustment); 2327 (void) arc_evict(arc_mru, 0, delta, FALSE, 2328 ARC_BUFC_METADATA); 2329 } 2330 2331 /* 2332 * Adjust MFU size 2333 */ 2334 2335 adjustment = arc_size - arc_c; 2336 2337 if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_DATA] > 0) { 2338 delta = MIN(adjustment, arc_mfu->arcs_lsize[ARC_BUFC_DATA]); 2339 (void) arc_evict(arc_mfu, 0, delta, FALSE, ARC_BUFC_DATA); 2340 adjustment -= delta; 2341 } 2342 2343 if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_METADATA] > 0) { 2344 int64_t delta = MIN(adjustment, 2345 arc_mfu->arcs_lsize[ARC_BUFC_METADATA]); 2346 (void) arc_evict(arc_mfu, 0, delta, FALSE, 2347 ARC_BUFC_METADATA); 2348 } 2349 2350 /* 2351 * Adjust ghost lists 2352 */ 2353 2354 adjustment = arc_mru->arcs_size + arc_mru_ghost->arcs_size - arc_c; 2355 2356 if (adjustment > 0 && arc_mru_ghost->arcs_size > 0) { 2357 delta = MIN(arc_mru_ghost->arcs_size, adjustment); 2358 arc_evict_ghost(arc_mru_ghost, 0, delta); 2359 } 2360 2361 adjustment = 2362 arc_mru_ghost->arcs_size + arc_mfu_ghost->arcs_size - arc_c; 2363 2364 if (adjustment > 0 && arc_mfu_ghost->arcs_size > 0) { 2365 delta = MIN(arc_mfu_ghost->arcs_size, adjustment); 2366 arc_evict_ghost(arc_mfu_ghost, 0, delta); 2367 } 2368} 2369 2370static void 2371arc_do_user_evicts(void) 2372{ 2373 static arc_buf_t *tmp_arc_eviction_list; 2374 2375 /* 2376 * Move list over to avoid LOR 2377 */ 2378restart: 2379 mutex_enter(&arc_eviction_mtx); 2380 tmp_arc_eviction_list = arc_eviction_list; 2381 arc_eviction_list = NULL; 2382 mutex_exit(&arc_eviction_mtx); 2383 2384 while (tmp_arc_eviction_list != NULL) { 2385 arc_buf_t *buf = tmp_arc_eviction_list; 2386 tmp_arc_eviction_list = buf->b_next; 2387 mutex_enter(&buf->b_evict_lock); 2388 buf->b_hdr = NULL; 2389 mutex_exit(&buf->b_evict_lock); 2390 2391 if (buf->b_efunc != NULL) 2392 VERIFY0(buf->b_efunc(buf->b_private)); 2393 2394 buf->b_efunc = NULL; 2395 buf->b_private = NULL; 2396 kmem_cache_free(buf_cache, buf); 2397 } 2398 2399 if (arc_eviction_list != NULL) 2400 goto restart; 2401} 2402 2403/* 2404 * Flush all *evictable* data from the cache for the given spa. 2405 * NOTE: this will not touch "active" (i.e. referenced) data. 2406 */ 2407void 2408arc_flush(spa_t *spa) 2409{ 2410 uint64_t guid = 0; 2411 2412 if (spa) 2413 guid = spa_load_guid(spa); 2414 2415 while (arc_mru->arcs_lsize[ARC_BUFC_DATA]) { 2416 (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_DATA); 2417 if (spa) 2418 break; 2419 } 2420 while (arc_mru->arcs_lsize[ARC_BUFC_METADATA]) { 2421 (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_METADATA); 2422 if (spa) 2423 break; 2424 } 2425 while (arc_mfu->arcs_lsize[ARC_BUFC_DATA]) { 2426 (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_DATA); 2427 if (spa) 2428 break; 2429 } 2430 while (arc_mfu->arcs_lsize[ARC_BUFC_METADATA]) { 2431 (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_METADATA); 2432 if (spa) 2433 break; 2434 } 2435 2436 arc_evict_ghost(arc_mru_ghost, guid, -1); 2437 arc_evict_ghost(arc_mfu_ghost, guid, -1); 2438 2439 mutex_enter(&arc_reclaim_thr_lock); 2440 arc_do_user_evicts(); 2441 mutex_exit(&arc_reclaim_thr_lock); 2442 ASSERT(spa || arc_eviction_list == NULL); 2443} 2444 2445void 2446arc_shrink(void) 2447{ 2448 2449 if (arc_c > arc_c_min) { 2450 uint64_t to_free; 2451 2452 DTRACE_PROBE4(arc__shrink, uint64_t, arc_c, uint64_t, 2453 arc_c_min, uint64_t, arc_p, uint64_t, to_free); 2454#ifdef _KERNEL 2455 to_free = arc_c >> arc_shrink_shift; 2456#else 2457 to_free = arc_c >> arc_shrink_shift; 2458#endif 2459 if (arc_c > arc_c_min + to_free) 2460 atomic_add_64(&arc_c, -to_free); 2461 else 2462 arc_c = arc_c_min; 2463 2464 atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift)); 2465 if (arc_c > arc_size) 2466 arc_c = MAX(arc_size, arc_c_min); 2467 if (arc_p > arc_c) 2468 arc_p = (arc_c >> 1); 2469 2470 DTRACE_PROBE2(arc__shrunk, uint64_t, arc_c, uint64_t, 2471 arc_p); 2472 2473 ASSERT(arc_c >= arc_c_min); 2474 ASSERT((int64_t)arc_p >= 0); 2475 } 2476 2477 if (arc_size > arc_c) { 2478 DTRACE_PROBE2(arc__shrink_adjust, uint64_t, arc_size, 2479 uint64_t, arc_c); 2480 arc_adjust(); 2481 } 2482} 2483 2484static int needfree = 0; 2485 2486static int 2487arc_reclaim_needed(void) 2488{ 2489 2490#ifdef _KERNEL 2491 2492 if (needfree) { 2493 DTRACE_PROBE(arc__reclaim_needfree); 2494 return (1); 2495 } 2496 2497 /* 2498 * Cooperate with pagedaemon when it's time for it to scan 2499 * and reclaim some pages. 2500 */ 2501 if (freemem < zfs_arc_free_target) { 2502 DTRACE_PROBE2(arc__reclaim_freemem, uint64_t, 2503 freemem, uint64_t, zfs_arc_free_target); 2504 return (1); 2505 } 2506 2507#ifdef sun 2508 /* 2509 * take 'desfree' extra pages, so we reclaim sooner, rather than later 2510 */ 2511 extra = desfree; 2512 2513 /* 2514 * check that we're out of range of the pageout scanner. It starts to 2515 * schedule paging if freemem is less than lotsfree and needfree. 2516 * lotsfree is the high-water mark for pageout, and needfree is the 2517 * number of needed free pages. We add extra pages here to make sure 2518 * the scanner doesn't start up while we're freeing memory. 2519 */ 2520 if (freemem < lotsfree + needfree + extra) 2521 return (1); 2522 2523 /* 2524 * check to make sure that swapfs has enough space so that anon 2525 * reservations can still succeed. anon_resvmem() checks that the 2526 * availrmem is greater than swapfs_minfree, and the number of reserved 2527 * swap pages. We also add a bit of extra here just to prevent 2528 * circumstances from getting really dire. 2529 */ 2530 if (availrmem < swapfs_minfree + swapfs_reserve + extra) 2531 return (1); 2532 2533 /* 2534 * Check that we have enough availrmem that memory locking (e.g., via 2535 * mlock(3C) or memcntl(2)) can still succeed. (pages_pp_maximum 2536 * stores the number of pages that cannot be locked; when availrmem 2537 * drops below pages_pp_maximum, page locking mechanisms such as 2538 * page_pp_lock() will fail.) 2539 */ 2540 if (availrmem <= pages_pp_maximum) 2541 return (1); 2542 2543#endif /* sun */ 2544#if defined(__i386) || !defined(UMA_MD_SMALL_ALLOC) 2545 /* 2546 * If we're on an i386 platform, it's possible that we'll exhaust the 2547 * kernel heap space before we ever run out of available physical 2548 * memory. Most checks of the size of the heap_area compare against 2549 * tune.t_minarmem, which is the minimum available real memory that we 2550 * can have in the system. However, this is generally fixed at 25 pages 2551 * which is so low that it's useless. In this comparison, we seek to 2552 * calculate the total heap-size, and reclaim if more than 3/4ths of the 2553 * heap is allocated. (Or, in the calculation, if less than 1/4th is 2554 * free) 2555 */ 2556 if (vmem_size(heap_arena, VMEM_FREE) < 2557 (vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC) >> 2)) { 2558 DTRACE_PROBE2(arc__reclaim_used, uint64_t, 2559 vmem_size(heap_arena, VMEM_FREE), uint64_t, 2560 (vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC)) >> 2); 2561 return (1); 2562 } 2563#endif 2564#ifdef sun 2565 /* 2566 * If zio data pages are being allocated out of a separate heap segment, 2567 * then enforce that the size of available vmem for this arena remains 2568 * above about 1/16th free. 2569 * 2570 * Note: The 1/16th arena free requirement was put in place 2571 * to aggressively evict memory from the arc in order to avoid 2572 * memory fragmentation issues. 2573 */ 2574 if (zio_arena != NULL && 2575 vmem_size(zio_arena, VMEM_FREE) < 2576 (vmem_size(zio_arena, VMEM_ALLOC) >> 4)) 2577 return (1); 2578#endif /* sun */ 2579#else /* _KERNEL */ 2580 if (spa_get_random(100) == 0) 2581 return (1); 2582#endif /* _KERNEL */ 2583 DTRACE_PROBE(arc__reclaim_no); 2584 2585 return (0); 2586} 2587 2588extern kmem_cache_t *zio_buf_cache[]; 2589extern kmem_cache_t *zio_data_buf_cache[]; 2590 2591static void __noinline 2592arc_kmem_reap_now(arc_reclaim_strategy_t strat) 2593{ 2594 size_t i; 2595 kmem_cache_t *prev_cache = NULL; 2596 kmem_cache_t *prev_data_cache = NULL; 2597 2598 DTRACE_PROBE(arc__kmem_reap_start); 2599#ifdef _KERNEL 2600 if (arc_meta_used >= arc_meta_limit) { 2601 /* 2602 * We are exceeding our meta-data cache limit. 2603 * Purge some DNLC entries to release holds on meta-data. 2604 */ 2605 dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent); 2606 } 2607#if defined(__i386) 2608 /* 2609 * Reclaim unused memory from all kmem caches. 2610 */ 2611 kmem_reap(); 2612#endif 2613#endif 2614 2615 /* 2616 * An aggressive reclamation will shrink the cache size as well as 2617 * reap free buffers from the arc kmem caches. 2618 */ 2619 if (strat == ARC_RECLAIM_AGGR) 2620 arc_shrink(); 2621 2622 for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) { 2623 if (zio_buf_cache[i] != prev_cache) { 2624 prev_cache = zio_buf_cache[i]; 2625 kmem_cache_reap_now(zio_buf_cache[i]); 2626 } 2627 if (zio_data_buf_cache[i] != prev_data_cache) { 2628 prev_data_cache = zio_data_buf_cache[i]; 2629 kmem_cache_reap_now(zio_data_buf_cache[i]); 2630 } 2631 } 2632 kmem_cache_reap_now(buf_cache); 2633 kmem_cache_reap_now(hdr_cache); 2634 2635#ifdef sun 2636 /* 2637 * Ask the vmem arena to reclaim unused memory from its 2638 * quantum caches. 2639 */ 2640 if (zio_arena != NULL && strat == ARC_RECLAIM_AGGR) 2641 vmem_qcache_reap(zio_arena); 2642#endif 2643 DTRACE_PROBE(arc__kmem_reap_end); 2644} 2645 2646static void 2647arc_reclaim_thread(void *dummy __unused) 2648{ 2649 clock_t growtime = 0; 2650 arc_reclaim_strategy_t last_reclaim = ARC_RECLAIM_CONS; 2651 callb_cpr_t cpr; 2652 2653 CALLB_CPR_INIT(&cpr, &arc_reclaim_thr_lock, callb_generic_cpr, FTAG); 2654 2655 mutex_enter(&arc_reclaim_thr_lock); 2656 while (arc_thread_exit == 0) { 2657 if (arc_reclaim_needed()) { 2658 2659 if (arc_no_grow) { 2660 if (last_reclaim == ARC_RECLAIM_CONS) { 2661 DTRACE_PROBE(arc__reclaim_aggr_no_grow); 2662 last_reclaim = ARC_RECLAIM_AGGR; 2663 } else { 2664 last_reclaim = ARC_RECLAIM_CONS; 2665 } 2666 } else { 2667 arc_no_grow = TRUE; 2668 last_reclaim = ARC_RECLAIM_AGGR; 2669 DTRACE_PROBE(arc__reclaim_aggr); 2670 membar_producer(); 2671 } 2672 2673 /* reset the growth delay for every reclaim */ 2674 growtime = ddi_get_lbolt() + (arc_grow_retry * hz); 2675 2676 if (needfree && last_reclaim == ARC_RECLAIM_CONS) { 2677 /* 2678 * If needfree is TRUE our vm_lowmem hook 2679 * was called and in that case we must free some 2680 * memory, so switch to aggressive mode. 2681 */ 2682 arc_no_grow = TRUE; 2683 last_reclaim = ARC_RECLAIM_AGGR; 2684 } 2685 arc_kmem_reap_now(last_reclaim); 2686 arc_warm = B_TRUE; 2687 2688 } else if (arc_no_grow && ddi_get_lbolt() >= growtime) { 2689 arc_no_grow = FALSE; 2690 } 2691 2692 arc_adjust(); 2693 2694 if (arc_eviction_list != NULL) 2695 arc_do_user_evicts(); 2696 2697#ifdef _KERNEL 2698 if (needfree) { 2699 needfree = 0; 2700 wakeup(&needfree); 2701 } 2702#endif 2703 2704 /* block until needed, or one second, whichever is shorter */ 2705 CALLB_CPR_SAFE_BEGIN(&cpr); 2706 (void) cv_timedwait(&arc_reclaim_thr_cv, 2707 &arc_reclaim_thr_lock, hz); 2708 CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_thr_lock); 2709 } 2710 2711 arc_thread_exit = 0; 2712 cv_broadcast(&arc_reclaim_thr_cv); 2713 CALLB_CPR_EXIT(&cpr); /* drops arc_reclaim_thr_lock */ 2714 thread_exit(); 2715} 2716 2717/* 2718 * Adapt arc info given the number of bytes we are trying to add and 2719 * the state that we are comming from. This function is only called 2720 * when we are adding new content to the cache. 2721 */ 2722static void 2723arc_adapt(int bytes, arc_state_t *state) 2724{ 2725 int mult; 2726 uint64_t arc_p_min = (arc_c >> arc_p_min_shift); 2727 2728 if (state == arc_l2c_only) 2729 return; 2730 2731 ASSERT(bytes > 0); 2732 /* 2733 * Adapt the target size of the MRU list: 2734 * - if we just hit in the MRU ghost list, then increase 2735 * the target size of the MRU list. 2736 * - if we just hit in the MFU ghost list, then increase 2737 * the target size of the MFU list by decreasing the 2738 * target size of the MRU list. 2739 */ 2740 if (state == arc_mru_ghost) { 2741 mult = ((arc_mru_ghost->arcs_size >= arc_mfu_ghost->arcs_size) ? 2742 1 : (arc_mfu_ghost->arcs_size/arc_mru_ghost->arcs_size)); 2743 mult = MIN(mult, 10); /* avoid wild arc_p adjustment */ 2744 2745 arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult); 2746 } else if (state == arc_mfu_ghost) { 2747 uint64_t delta; 2748 2749 mult = ((arc_mfu_ghost->arcs_size >= arc_mru_ghost->arcs_size) ? 2750 1 : (arc_mru_ghost->arcs_size/arc_mfu_ghost->arcs_size)); 2751 mult = MIN(mult, 10); 2752 2753 delta = MIN(bytes * mult, arc_p); 2754 arc_p = MAX(arc_p_min, arc_p - delta); 2755 } 2756 ASSERT((int64_t)arc_p >= 0); 2757 2758 if (arc_reclaim_needed()) { 2759 cv_signal(&arc_reclaim_thr_cv); 2760 return; 2761 } 2762 2763 if (arc_no_grow) 2764 return; 2765 2766 if (arc_c >= arc_c_max) 2767 return; 2768 2769 /* 2770 * If we're within (2 * maxblocksize) bytes of the target 2771 * cache size, increment the target cache size 2772 */ 2773 if (arc_size > arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) { 2774 DTRACE_PROBE1(arc__inc_adapt, int, bytes); 2775 atomic_add_64(&arc_c, (int64_t)bytes); 2776 if (arc_c > arc_c_max) 2777 arc_c = arc_c_max; 2778 else if (state == arc_anon) 2779 atomic_add_64(&arc_p, (int64_t)bytes); 2780 if (arc_p > arc_c) 2781 arc_p = arc_c; 2782 } 2783 ASSERT((int64_t)arc_p >= 0); 2784} 2785 2786/* 2787 * Check if the cache has reached its limits and eviction is required 2788 * prior to insert. 2789 */ 2790static int 2791arc_evict_needed(arc_buf_contents_t type) 2792{ 2793 if (type == ARC_BUFC_METADATA && arc_meta_used >= arc_meta_limit) 2794 return (1); 2795 2796 if (arc_reclaim_needed()) 2797 return (1); 2798 2799 return (arc_size > arc_c); 2800} 2801 2802/* 2803 * The buffer, supplied as the first argument, needs a data block. 2804 * So, if we are at cache max, determine which cache should be victimized. 2805 * We have the following cases: 2806 * 2807 * 1. Insert for MRU, p > sizeof(arc_anon + arc_mru) -> 2808 * In this situation if we're out of space, but the resident size of the MFU is 2809 * under the limit, victimize the MFU cache to satisfy this insertion request. 2810 * 2811 * 2. Insert for MRU, p <= sizeof(arc_anon + arc_mru) -> 2812 * Here, we've used up all of the available space for the MRU, so we need to 2813 * evict from our own cache instead. Evict from the set of resident MRU 2814 * entries. 2815 * 2816 * 3. Insert for MFU (c - p) > sizeof(arc_mfu) -> 2817 * c minus p represents the MFU space in the cache, since p is the size of the 2818 * cache that is dedicated to the MRU. In this situation there's still space on 2819 * the MFU side, so the MRU side needs to be victimized. 2820 * 2821 * 4. Insert for MFU (c - p) < sizeof(arc_mfu) -> 2822 * MFU's resident set is consuming more space than it has been allotted. In 2823 * this situation, we must victimize our own cache, the MFU, for this insertion. 2824 */ 2825static void 2826arc_get_data_buf(arc_buf_t *buf) 2827{ 2828 arc_state_t *state = buf->b_hdr->b_state; 2829 uint64_t size = buf->b_hdr->b_size; 2830 arc_buf_contents_t type = buf->b_hdr->b_type; 2831 2832 arc_adapt(size, state); 2833 2834 /* 2835 * We have not yet reached cache maximum size, 2836 * just allocate a new buffer. 2837 */ 2838 if (!arc_evict_needed(type)) { 2839 if (type == ARC_BUFC_METADATA) { 2840 buf->b_data = zio_buf_alloc(size); 2841 arc_space_consume(size, ARC_SPACE_DATA); 2842 } else { 2843 ASSERT(type == ARC_BUFC_DATA); 2844 buf->b_data = zio_data_buf_alloc(size); 2845 ARCSTAT_INCR(arcstat_data_size, size); 2846 atomic_add_64(&arc_size, size); 2847 } 2848 goto out; 2849 } 2850 2851 /* 2852 * If we are prefetching from the mfu ghost list, this buffer 2853 * will end up on the mru list; so steal space from there. 2854 */ 2855 if (state == arc_mfu_ghost) 2856 state = buf->b_hdr->b_flags & ARC_PREFETCH ? arc_mru : arc_mfu; 2857 else if (state == arc_mru_ghost) 2858 state = arc_mru; 2859 2860 if (state == arc_mru || state == arc_anon) { 2861 uint64_t mru_used = arc_anon->arcs_size + arc_mru->arcs_size; 2862 state = (arc_mfu->arcs_lsize[type] >= size && 2863 arc_p > mru_used) ? arc_mfu : arc_mru; 2864 } else { 2865 /* MFU cases */ 2866 uint64_t mfu_space = arc_c - arc_p; 2867 state = (arc_mru->arcs_lsize[type] >= size && 2868 mfu_space > arc_mfu->arcs_size) ? arc_mru : arc_mfu; 2869 } 2870 if ((buf->b_data = arc_evict(state, 0, size, TRUE, type)) == NULL) { 2871 if (type == ARC_BUFC_METADATA) { 2872 buf->b_data = zio_buf_alloc(size); 2873 arc_space_consume(size, ARC_SPACE_DATA); 2874 } else { 2875 ASSERT(type == ARC_BUFC_DATA); 2876 buf->b_data = zio_data_buf_alloc(size); 2877 ARCSTAT_INCR(arcstat_data_size, size); 2878 atomic_add_64(&arc_size, size); 2879 } 2880 ARCSTAT_BUMP(arcstat_recycle_miss); 2881 } 2882 ASSERT(buf->b_data != NULL); 2883out: 2884 /* 2885 * Update the state size. Note that ghost states have a 2886 * "ghost size" and so don't need to be updated. 2887 */ 2888 if (!GHOST_STATE(buf->b_hdr->b_state)) { 2889 arc_buf_hdr_t *hdr = buf->b_hdr; 2890 2891 atomic_add_64(&hdr->b_state->arcs_size, size); 2892 if (list_link_active(&hdr->b_arc_node)) { 2893 ASSERT(refcount_is_zero(&hdr->b_refcnt)); 2894 atomic_add_64(&hdr->b_state->arcs_lsize[type], size); 2895 } 2896 /* 2897 * If we are growing the cache, and we are adding anonymous 2898 * data, and we have outgrown arc_p, update arc_p 2899 */ 2900 if (arc_size < arc_c && hdr->b_state == arc_anon && 2901 arc_anon->arcs_size + arc_mru->arcs_size > arc_p) 2902 arc_p = MIN(arc_c, arc_p + size); 2903 } 2904 ARCSTAT_BUMP(arcstat_allocated); 2905} 2906 2907/* 2908 * This routine is called whenever a buffer is accessed. 2909 * NOTE: the hash lock is dropped in this function. 2910 */ 2911static void 2912arc_access(arc_buf_hdr_t *buf, kmutex_t *hash_lock) 2913{ 2914 clock_t now; 2915 2916 ASSERT(MUTEX_HELD(hash_lock)); 2917 2918 if (buf->b_state == arc_anon) { 2919 /* 2920 * This buffer is not in the cache, and does not 2921 * appear in our "ghost" list. Add the new buffer 2922 * to the MRU state. 2923 */ 2924 2925 ASSERT(buf->b_arc_access == 0); 2926 buf->b_arc_access = ddi_get_lbolt(); 2927 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, buf); 2928 arc_change_state(arc_mru, buf, hash_lock); 2929 2930 } else if (buf->b_state == arc_mru) { 2931 now = ddi_get_lbolt(); 2932 2933 /* 2934 * If this buffer is here because of a prefetch, then either: 2935 * - clear the flag if this is a "referencing" read 2936 * (any subsequent access will bump this into the MFU state). 2937 * or 2938 * - move the buffer to the head of the list if this is 2939 * another prefetch (to make it less likely to be evicted). 2940 */ 2941 if ((buf->b_flags & ARC_PREFETCH) != 0) { 2942 if (refcount_count(&buf->b_refcnt) == 0) { 2943 ASSERT(list_link_active(&buf->b_arc_node)); 2944 } else { 2945 buf->b_flags &= ~ARC_PREFETCH; 2946 ARCSTAT_BUMP(arcstat_mru_hits); 2947 } 2948 buf->b_arc_access = now; 2949 return; 2950 } 2951 2952 /* 2953 * This buffer has been "accessed" only once so far, 2954 * but it is still in the cache. Move it to the MFU 2955 * state. 2956 */ 2957 if (now > buf->b_arc_access + ARC_MINTIME) { 2958 /* 2959 * More than 125ms have passed since we 2960 * instantiated this buffer. Move it to the 2961 * most frequently used state. 2962 */ 2963 buf->b_arc_access = now; 2964 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf); 2965 arc_change_state(arc_mfu, buf, hash_lock); 2966 } 2967 ARCSTAT_BUMP(arcstat_mru_hits); 2968 } else if (buf->b_state == arc_mru_ghost) { 2969 arc_state_t *new_state; 2970 /* 2971 * This buffer has been "accessed" recently, but 2972 * was evicted from the cache. Move it to the 2973 * MFU state. 2974 */ 2975 2976 if (buf->b_flags & ARC_PREFETCH) { 2977 new_state = arc_mru; 2978 if (refcount_count(&buf->b_refcnt) > 0) 2979 buf->b_flags &= ~ARC_PREFETCH; 2980 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, buf); 2981 } else { 2982 new_state = arc_mfu; 2983 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf); 2984 } 2985 2986 buf->b_arc_access = ddi_get_lbolt(); 2987 arc_change_state(new_state, buf, hash_lock); 2988 2989 ARCSTAT_BUMP(arcstat_mru_ghost_hits); 2990 } else if (buf->b_state == arc_mfu) { 2991 /* 2992 * This buffer has been accessed more than once and is 2993 * still in the cache. Keep it in the MFU state. 2994 * 2995 * NOTE: an add_reference() that occurred when we did 2996 * the arc_read() will have kicked this off the list. 2997 * If it was a prefetch, we will explicitly move it to 2998 * the head of the list now. 2999 */ 3000 if ((buf->b_flags & ARC_PREFETCH) != 0) { 3001 ASSERT(refcount_count(&buf->b_refcnt) == 0); 3002 ASSERT(list_link_active(&buf->b_arc_node)); 3003 } 3004 ARCSTAT_BUMP(arcstat_mfu_hits); 3005 buf->b_arc_access = ddi_get_lbolt(); 3006 } else if (buf->b_state == arc_mfu_ghost) { 3007 arc_state_t *new_state = arc_mfu; 3008 /* 3009 * This buffer has been accessed more than once but has 3010 * been evicted from the cache. Move it back to the 3011 * MFU state. 3012 */ 3013 3014 if (buf->b_flags & ARC_PREFETCH) { 3015 /* 3016 * This is a prefetch access... 3017 * move this block back to the MRU state. 3018 */ 3019 ASSERT0(refcount_count(&buf->b_refcnt)); 3020 new_state = arc_mru; 3021 } 3022 3023 buf->b_arc_access = ddi_get_lbolt(); 3024 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf); 3025 arc_change_state(new_state, buf, hash_lock); 3026 3027 ARCSTAT_BUMP(arcstat_mfu_ghost_hits); 3028 } else if (buf->b_state == arc_l2c_only) { 3029 /* 3030 * This buffer is on the 2nd Level ARC. 3031 */ 3032 3033 buf->b_arc_access = ddi_get_lbolt(); 3034 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf); 3035 arc_change_state(arc_mfu, buf, hash_lock); 3036 } else { 3037 ASSERT(!"invalid arc state"); 3038 } 3039} 3040 3041/* a generic arc_done_func_t which you can use */ 3042/* ARGSUSED */ 3043void 3044arc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg) 3045{ 3046 if (zio == NULL || zio->io_error == 0) 3047 bcopy(buf->b_data, arg, buf->b_hdr->b_size); 3048 VERIFY(arc_buf_remove_ref(buf, arg)); 3049} 3050 3051/* a generic arc_done_func_t */ 3052void 3053arc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg) 3054{ 3055 arc_buf_t **bufp = arg; 3056 if (zio && zio->io_error) { 3057 VERIFY(arc_buf_remove_ref(buf, arg)); 3058 *bufp = NULL; 3059 } else { 3060 *bufp = buf; 3061 ASSERT(buf->b_data); 3062 } 3063} 3064 3065static void 3066arc_read_done(zio_t *zio) 3067{ 3068 arc_buf_hdr_t *hdr; 3069 arc_buf_t *buf; 3070 arc_buf_t *abuf; /* buffer we're assigning to callback */ 3071 kmutex_t *hash_lock = NULL; 3072 arc_callback_t *callback_list, *acb; 3073 int freeable = FALSE; 3074 3075 buf = zio->io_private; 3076 hdr = buf->b_hdr; 3077 3078 /* 3079 * The hdr was inserted into hash-table and removed from lists 3080 * prior to starting I/O. We should find this header, since 3081 * it's in the hash table, and it should be legit since it's 3082 * not possible to evict it during the I/O. The only possible 3083 * reason for it not to be found is if we were freed during the 3084 * read. 3085 */ 3086 if (HDR_IN_HASH_TABLE(hdr)) { 3087 ASSERT3U(hdr->b_birth, ==, BP_PHYSICAL_BIRTH(zio->io_bp)); 3088 ASSERT3U(hdr->b_dva.dva_word[0], ==, 3089 BP_IDENTITY(zio->io_bp)->dva_word[0]); 3090 ASSERT3U(hdr->b_dva.dva_word[1], ==, 3091 BP_IDENTITY(zio->io_bp)->dva_word[1]); 3092 3093 arc_buf_hdr_t *found = buf_hash_find(hdr->b_spa, zio->io_bp, 3094 &hash_lock); 3095 3096 ASSERT((found == NULL && HDR_FREED_IN_READ(hdr) && 3097 hash_lock == NULL) || 3098 (found == hdr && 3099 DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) || 3100 (found == hdr && HDR_L2_READING(hdr))); 3101 } 3102 3103 hdr->b_flags &= ~ARC_L2_EVICTED; 3104 if (l2arc_noprefetch && (hdr->b_flags & ARC_PREFETCH)) 3105 hdr->b_flags &= ~ARC_L2CACHE; 3106 3107 /* byteswap if necessary */ 3108 callback_list = hdr->b_acb; 3109 ASSERT(callback_list != NULL); 3110 if (BP_SHOULD_BYTESWAP(zio->io_bp) && zio->io_error == 0) { 3111 dmu_object_byteswap_t bswap = 3112 DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp)); 3113 arc_byteswap_func_t *func = BP_GET_LEVEL(zio->io_bp) > 0 ? 3114 byteswap_uint64_array : 3115 dmu_ot_byteswap[bswap].ob_func; 3116 func(buf->b_data, hdr->b_size); 3117 } 3118 3119 arc_cksum_compute(buf, B_FALSE); 3120#ifdef illumos 3121 arc_buf_watch(buf); 3122#endif /* illumos */ 3123 3124 if (hash_lock && zio->io_error == 0 && hdr->b_state == arc_anon) { 3125 /* 3126 * Only call arc_access on anonymous buffers. This is because 3127 * if we've issued an I/O for an evicted buffer, we've already 3128 * called arc_access (to prevent any simultaneous readers from 3129 * getting confused). 3130 */ 3131 arc_access(hdr, hash_lock); 3132 } 3133 3134 /* create copies of the data buffer for the callers */ 3135 abuf = buf; 3136 for (acb = callback_list; acb; acb = acb->acb_next) { 3137 if (acb->acb_done) { 3138 if (abuf == NULL) { 3139 ARCSTAT_BUMP(arcstat_duplicate_reads); 3140 abuf = arc_buf_clone(buf); 3141 } 3142 acb->acb_buf = abuf; 3143 abuf = NULL; 3144 } 3145 } 3146 hdr->b_acb = NULL; 3147 hdr->b_flags &= ~ARC_IO_IN_PROGRESS; 3148 ASSERT(!HDR_BUF_AVAILABLE(hdr)); 3149 if (abuf == buf) { 3150 ASSERT(buf->b_efunc == NULL); 3151 ASSERT(hdr->b_datacnt == 1); 3152 hdr->b_flags |= ARC_BUF_AVAILABLE; 3153 } 3154 3155 ASSERT(refcount_is_zero(&hdr->b_refcnt) || callback_list != NULL); 3156 3157 if (zio->io_error != 0) { 3158 hdr->b_flags |= ARC_IO_ERROR; 3159 if (hdr->b_state != arc_anon) 3160 arc_change_state(arc_anon, hdr, hash_lock); 3161 if (HDR_IN_HASH_TABLE(hdr)) 3162 buf_hash_remove(hdr); 3163 freeable = refcount_is_zero(&hdr->b_refcnt); 3164 } 3165 3166 /* 3167 * Broadcast before we drop the hash_lock to avoid the possibility 3168 * that the hdr (and hence the cv) might be freed before we get to 3169 * the cv_broadcast(). 3170 */ 3171 cv_broadcast(&hdr->b_cv); 3172 3173 if (hash_lock) { 3174 mutex_exit(hash_lock); 3175 } else { 3176 /* 3177 * This block was freed while we waited for the read to 3178 * complete. It has been removed from the hash table and 3179 * moved to the anonymous state (so that it won't show up 3180 * in the cache). 3181 */ 3182 ASSERT3P(hdr->b_state, ==, arc_anon); 3183 freeable = refcount_is_zero(&hdr->b_refcnt); 3184 } 3185 3186 /* execute each callback and free its structure */ 3187 while ((acb = callback_list) != NULL) { 3188 if (acb->acb_done) 3189 acb->acb_done(zio, acb->acb_buf, acb->acb_private); 3190 3191 if (acb->acb_zio_dummy != NULL) { 3192 acb->acb_zio_dummy->io_error = zio->io_error; 3193 zio_nowait(acb->acb_zio_dummy); 3194 } 3195 3196 callback_list = acb->acb_next; 3197 kmem_free(acb, sizeof (arc_callback_t)); 3198 } 3199 3200 if (freeable) 3201 arc_hdr_destroy(hdr); 3202} 3203 3204/* 3205 * "Read" the block block at the specified DVA (in bp) via the 3206 * cache. If the block is found in the cache, invoke the provided 3207 * callback immediately and return. Note that the `zio' parameter 3208 * in the callback will be NULL in this case, since no IO was 3209 * required. If the block is not in the cache pass the read request 3210 * on to the spa with a substitute callback function, so that the 3211 * requested block will be added to the cache. 3212 * 3213 * If a read request arrives for a block that has a read in-progress, 3214 * either wait for the in-progress read to complete (and return the 3215 * results); or, if this is a read with a "done" func, add a record 3216 * to the read to invoke the "done" func when the read completes, 3217 * and return; or just return. 3218 * 3219 * arc_read_done() will invoke all the requested "done" functions 3220 * for readers of this block. 3221 */ 3222int 3223arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_done_func_t *done, 3224 void *private, zio_priority_t priority, int zio_flags, uint32_t *arc_flags, 3225 const zbookmark_phys_t *zb) 3226{ 3227 arc_buf_hdr_t *hdr = NULL; 3228 arc_buf_t *buf = NULL; 3229 kmutex_t *hash_lock = NULL; 3230 zio_t *rzio; 3231 uint64_t guid = spa_load_guid(spa); 3232 3233 ASSERT(!BP_IS_EMBEDDED(bp) || 3234 BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA); 3235 3236top: 3237 if (!BP_IS_EMBEDDED(bp)) { 3238 /* 3239 * Embedded BP's have no DVA and require no I/O to "read". 3240 * Create an anonymous arc buf to back it. 3241 */ 3242 hdr = buf_hash_find(guid, bp, &hash_lock); 3243 } 3244 3245 if (hdr != NULL && hdr->b_datacnt > 0) { 3246 3247 *arc_flags |= ARC_CACHED; 3248 3249 if (HDR_IO_IN_PROGRESS(hdr)) { 3250 3251 if (*arc_flags & ARC_WAIT) { 3252 cv_wait(&hdr->b_cv, hash_lock); 3253 mutex_exit(hash_lock); 3254 goto top; 3255 } 3256 ASSERT(*arc_flags & ARC_NOWAIT); 3257 3258 if (done) { 3259 arc_callback_t *acb = NULL; 3260 3261 acb = kmem_zalloc(sizeof (arc_callback_t), 3262 KM_SLEEP); 3263 acb->acb_done = done; 3264 acb->acb_private = private; 3265 if (pio != NULL) 3266 acb->acb_zio_dummy = zio_null(pio, 3267 spa, NULL, NULL, NULL, zio_flags); 3268 3269 ASSERT(acb->acb_done != NULL); 3270 acb->acb_next = hdr->b_acb; 3271 hdr->b_acb = acb; 3272 add_reference(hdr, hash_lock, private); 3273 mutex_exit(hash_lock); 3274 return (0); 3275 } 3276 mutex_exit(hash_lock); 3277 return (0); 3278 } 3279 3280 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu); 3281 3282 if (done) { 3283 add_reference(hdr, hash_lock, private); 3284 /* 3285 * If this block is already in use, create a new 3286 * copy of the data so that we will be guaranteed 3287 * that arc_release() will always succeed. 3288 */ 3289 buf = hdr->b_buf; 3290 ASSERT(buf); 3291 ASSERT(buf->b_data); 3292 if (HDR_BUF_AVAILABLE(hdr)) { 3293 ASSERT(buf->b_efunc == NULL); 3294 hdr->b_flags &= ~ARC_BUF_AVAILABLE; 3295 } else { 3296 buf = arc_buf_clone(buf); 3297 } 3298 3299 } else if (*arc_flags & ARC_PREFETCH && 3300 refcount_count(&hdr->b_refcnt) == 0) { 3301 hdr->b_flags |= ARC_PREFETCH; 3302 } 3303 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr); 3304 arc_access(hdr, hash_lock); 3305 if (*arc_flags & ARC_L2CACHE) 3306 hdr->b_flags |= ARC_L2CACHE; 3307 if (*arc_flags & ARC_L2COMPRESS) 3308 hdr->b_flags |= ARC_L2COMPRESS; 3309 mutex_exit(hash_lock); 3310 ARCSTAT_BUMP(arcstat_hits); 3311 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH), 3312 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA, 3313 data, metadata, hits); 3314 3315 if (done) 3316 done(NULL, buf, private); 3317 } else { 3318 uint64_t size = BP_GET_LSIZE(bp); 3319 arc_callback_t *acb; 3320 vdev_t *vd = NULL; 3321 uint64_t addr = 0; 3322 boolean_t devw = B_FALSE; 3323 enum zio_compress b_compress = ZIO_COMPRESS_OFF; 3324 uint64_t b_asize = 0; 3325 3326 if (hdr == NULL) { 3327 /* this block is not in the cache */ 3328 arc_buf_hdr_t *exists = NULL; 3329 arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp); 3330 buf = arc_buf_alloc(spa, size, private, type); 3331 hdr = buf->b_hdr; 3332 if (!BP_IS_EMBEDDED(bp)) { 3333 hdr->b_dva = *BP_IDENTITY(bp); 3334 hdr->b_birth = BP_PHYSICAL_BIRTH(bp); 3335 hdr->b_cksum0 = bp->blk_cksum.zc_word[0]; 3336 exists = buf_hash_insert(hdr, &hash_lock); 3337 } 3338 if (exists != NULL) { 3339 /* somebody beat us to the hash insert */ 3340 mutex_exit(hash_lock); 3341 buf_discard_identity(hdr); 3342 (void) arc_buf_remove_ref(buf, private); 3343 goto top; /* restart the IO request */ 3344 } 3345 /* if this is a prefetch, we don't have a reference */ 3346 if (*arc_flags & ARC_PREFETCH) { 3347 (void) remove_reference(hdr, hash_lock, 3348 private); 3349 hdr->b_flags |= ARC_PREFETCH; 3350 } 3351 if (*arc_flags & ARC_L2CACHE) 3352 hdr->b_flags |= ARC_L2CACHE; 3353 if (*arc_flags & ARC_L2COMPRESS) 3354 hdr->b_flags |= ARC_L2COMPRESS; 3355 if (BP_GET_LEVEL(bp) > 0) 3356 hdr->b_flags |= ARC_INDIRECT; 3357 } else { 3358 /* this block is in the ghost cache */ 3359 ASSERT(GHOST_STATE(hdr->b_state)); 3360 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 3361 ASSERT0(refcount_count(&hdr->b_refcnt)); 3362 ASSERT(hdr->b_buf == NULL); 3363 3364 /* if this is a prefetch, we don't have a reference */ 3365 if (*arc_flags & ARC_PREFETCH) 3366 hdr->b_flags |= ARC_PREFETCH; 3367 else 3368 add_reference(hdr, hash_lock, private); 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 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 3374 buf->b_hdr = hdr; 3375 buf->b_data = NULL; 3376 buf->b_efunc = NULL; 3377 buf->b_private = NULL; 3378 buf->b_next = NULL; 3379 hdr->b_buf = buf; 3380 ASSERT(hdr->b_datacnt == 0); 3381 hdr->b_datacnt = 1; 3382 arc_get_data_buf(buf); 3383 arc_access(hdr, hash_lock); 3384 } 3385 3386 ASSERT(!GHOST_STATE(hdr->b_state)); 3387 3388 acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP); 3389 acb->acb_done = done; 3390 acb->acb_private = private; 3391 3392 ASSERT(hdr->b_acb == NULL); 3393 hdr->b_acb = acb; 3394 hdr->b_flags |= ARC_IO_IN_PROGRESS; 3395 3396 if (hdr->b_l2hdr != NULL && 3397 (vd = hdr->b_l2hdr->b_dev->l2ad_vdev) != NULL) { 3398 devw = hdr->b_l2hdr->b_dev->l2ad_writing; 3399 addr = hdr->b_l2hdr->b_daddr; 3400 b_compress = hdr->b_l2hdr->b_compress; 3401 b_asize = hdr->b_l2hdr->b_asize; 3402 /* 3403 * Lock out device removal. 3404 */ 3405 if (vdev_is_dead(vd) || 3406 !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER)) 3407 vd = NULL; 3408 } 3409 3410 if (hash_lock != NULL) 3411 mutex_exit(hash_lock); 3412 3413 /* 3414 * At this point, we have a level 1 cache miss. Try again in 3415 * L2ARC if possible. 3416 */ 3417 ASSERT3U(hdr->b_size, ==, size); 3418 DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp, 3419 uint64_t, size, zbookmark_phys_t *, zb); 3420 ARCSTAT_BUMP(arcstat_misses); 3421 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH), 3422 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA, 3423 data, metadata, misses); 3424#ifdef _KERNEL 3425 curthread->td_ru.ru_inblock++; 3426#endif 3427 3428 if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) { 3429 /* 3430 * Read from the L2ARC if the following are true: 3431 * 1. The L2ARC vdev was previously cached. 3432 * 2. This buffer still has L2ARC metadata. 3433 * 3. This buffer isn't currently writing to the L2ARC. 3434 * 4. The L2ARC entry wasn't evicted, which may 3435 * also have invalidated the vdev. 3436 * 5. This isn't prefetch and l2arc_noprefetch is set. 3437 */ 3438 if (hdr->b_l2hdr != NULL && 3439 !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) && 3440 !(l2arc_noprefetch && HDR_PREFETCH(hdr))) { 3441 l2arc_read_callback_t *cb; 3442 3443 DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr); 3444 ARCSTAT_BUMP(arcstat_l2_hits); 3445 3446 cb = kmem_zalloc(sizeof (l2arc_read_callback_t), 3447 KM_SLEEP); 3448 cb->l2rcb_buf = buf; 3449 cb->l2rcb_spa = spa; 3450 cb->l2rcb_bp = *bp; 3451 cb->l2rcb_zb = *zb; 3452 cb->l2rcb_flags = zio_flags; 3453 cb->l2rcb_compress = b_compress; 3454 3455 ASSERT(addr >= VDEV_LABEL_START_SIZE && 3456 addr + size < vd->vdev_psize - 3457 VDEV_LABEL_END_SIZE); 3458 3459 /* 3460 * l2arc read. The SCL_L2ARC lock will be 3461 * released by l2arc_read_done(). 3462 * Issue a null zio if the underlying buffer 3463 * was squashed to zero size by compression. 3464 */ 3465 if (b_compress == ZIO_COMPRESS_EMPTY) { 3466 rzio = zio_null(pio, spa, vd, 3467 l2arc_read_done, cb, 3468 zio_flags | ZIO_FLAG_DONT_CACHE | 3469 ZIO_FLAG_CANFAIL | 3470 ZIO_FLAG_DONT_PROPAGATE | 3471 ZIO_FLAG_DONT_RETRY); 3472 } else { 3473 rzio = zio_read_phys(pio, vd, addr, 3474 b_asize, buf->b_data, 3475 ZIO_CHECKSUM_OFF, 3476 l2arc_read_done, cb, priority, 3477 zio_flags | ZIO_FLAG_DONT_CACHE | 3478 ZIO_FLAG_CANFAIL | 3479 ZIO_FLAG_DONT_PROPAGATE | 3480 ZIO_FLAG_DONT_RETRY, B_FALSE); 3481 } 3482 DTRACE_PROBE2(l2arc__read, vdev_t *, vd, 3483 zio_t *, rzio); 3484 ARCSTAT_INCR(arcstat_l2_read_bytes, b_asize); 3485 3486 if (*arc_flags & ARC_NOWAIT) { 3487 zio_nowait(rzio); 3488 return (0); 3489 } 3490 3491 ASSERT(*arc_flags & ARC_WAIT); 3492 if (zio_wait(rzio) == 0) 3493 return (0); 3494 3495 /* l2arc read error; goto zio_read() */ 3496 } else { 3497 DTRACE_PROBE1(l2arc__miss, 3498 arc_buf_hdr_t *, hdr); 3499 ARCSTAT_BUMP(arcstat_l2_misses); 3500 if (HDR_L2_WRITING(hdr)) 3501 ARCSTAT_BUMP(arcstat_l2_rw_clash); 3502 spa_config_exit(spa, SCL_L2ARC, vd); 3503 } 3504 } else { 3505 if (vd != NULL) 3506 spa_config_exit(spa, SCL_L2ARC, vd); 3507 if (l2arc_ndev != 0) { 3508 DTRACE_PROBE1(l2arc__miss, 3509 arc_buf_hdr_t *, hdr); 3510 ARCSTAT_BUMP(arcstat_l2_misses); 3511 } 3512 } 3513 3514 rzio = zio_read(pio, spa, bp, buf->b_data, size, 3515 arc_read_done, buf, priority, zio_flags, zb); 3516 3517 if (*arc_flags & ARC_WAIT) 3518 return (zio_wait(rzio)); 3519 3520 ASSERT(*arc_flags & ARC_NOWAIT); 3521 zio_nowait(rzio); 3522 } 3523 return (0); 3524} 3525 3526void 3527arc_set_callback(arc_buf_t *buf, arc_evict_func_t *func, void *private) 3528{ 3529 ASSERT(buf->b_hdr != NULL); 3530 ASSERT(buf->b_hdr->b_state != arc_anon); 3531 ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt) || func == NULL); 3532 ASSERT(buf->b_efunc == NULL); 3533 ASSERT(!HDR_BUF_AVAILABLE(buf->b_hdr)); 3534 3535 buf->b_efunc = func; 3536 buf->b_private = private; 3537} 3538 3539/* 3540 * Notify the arc that a block was freed, and thus will never be used again. 3541 */ 3542void 3543arc_freed(spa_t *spa, const blkptr_t *bp) 3544{ 3545 arc_buf_hdr_t *hdr; 3546 kmutex_t *hash_lock; 3547 uint64_t guid = spa_load_guid(spa); 3548 3549 ASSERT(!BP_IS_EMBEDDED(bp)); 3550 3551 hdr = buf_hash_find(guid, bp, &hash_lock); 3552 if (hdr == NULL) 3553 return; 3554 if (HDR_BUF_AVAILABLE(hdr)) { 3555 arc_buf_t *buf = hdr->b_buf; 3556 add_reference(hdr, hash_lock, FTAG); 3557 hdr->b_flags &= ~ARC_BUF_AVAILABLE; 3558 mutex_exit(hash_lock); 3559 3560 arc_release(buf, FTAG); 3561 (void) arc_buf_remove_ref(buf, FTAG); 3562 } else { 3563 mutex_exit(hash_lock); 3564 } 3565 3566} 3567 3568/* 3569 * Clear the user eviction callback set by arc_set_callback(), first calling 3570 * it if it exists. Because the presence of a callback keeps an arc_buf cached 3571 * clearing the callback may result in the arc_buf being destroyed. However, 3572 * it will not result in the *last* arc_buf being destroyed, hence the data 3573 * will remain cached in the ARC. We make a copy of the arc buffer here so 3574 * that we can process the callback without holding any locks. 3575 * 3576 * It's possible that the callback is already in the process of being cleared 3577 * by another thread. In this case we can not clear the callback. 3578 * 3579 * Returns B_TRUE if the callback was successfully called and cleared. 3580 */ 3581boolean_t 3582arc_clear_callback(arc_buf_t *buf) 3583{ 3584 arc_buf_hdr_t *hdr; 3585 kmutex_t *hash_lock; 3586 arc_evict_func_t *efunc = buf->b_efunc; 3587 void *private = buf->b_private; 3588 list_t *list, *evicted_list; 3589 kmutex_t *lock, *evicted_lock; 3590 3591 mutex_enter(&buf->b_evict_lock); 3592 hdr = buf->b_hdr; 3593 if (hdr == NULL) { 3594 /* 3595 * We are in arc_do_user_evicts(). 3596 */ 3597 ASSERT(buf->b_data == NULL); 3598 mutex_exit(&buf->b_evict_lock); 3599 return (B_FALSE); 3600 } else if (buf->b_data == NULL) { 3601 /* 3602 * We are on the eviction list; process this buffer now 3603 * but let arc_do_user_evicts() do the reaping. 3604 */ 3605 buf->b_efunc = NULL; 3606 mutex_exit(&buf->b_evict_lock); 3607 VERIFY0(efunc(private)); 3608 return (B_TRUE); 3609 } 3610 hash_lock = HDR_LOCK(hdr); 3611 mutex_enter(hash_lock); 3612 hdr = buf->b_hdr; 3613 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 3614 3615 ASSERT3U(refcount_count(&hdr->b_refcnt), <, hdr->b_datacnt); 3616 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu); 3617 3618 buf->b_efunc = NULL; 3619 buf->b_private = NULL; 3620 3621 if (hdr->b_datacnt > 1) { 3622 mutex_exit(&buf->b_evict_lock); 3623 arc_buf_destroy(buf, FALSE, TRUE); 3624 } else { 3625 ASSERT(buf == hdr->b_buf); 3626 hdr->b_flags |= ARC_BUF_AVAILABLE; 3627 mutex_exit(&buf->b_evict_lock); 3628 } 3629 3630 mutex_exit(hash_lock); 3631 VERIFY0(efunc(private)); 3632 return (B_TRUE); 3633} 3634 3635/* 3636 * Release this buffer from the cache, making it an anonymous buffer. This 3637 * must be done after a read and prior to modifying the buffer contents. 3638 * If the buffer has more than one reference, we must make 3639 * a new hdr for the buffer. 3640 */ 3641void 3642arc_release(arc_buf_t *buf, void *tag) 3643{ 3644 arc_buf_hdr_t *hdr; 3645 kmutex_t *hash_lock = NULL; 3646 l2arc_buf_hdr_t *l2hdr; 3647 uint64_t buf_size; 3648 3649 /* 3650 * It would be nice to assert that if it's DMU metadata (level > 3651 * 0 || it's the dnode file), then it must be syncing context. 3652 * But we don't know that information at this level. 3653 */ 3654 3655 mutex_enter(&buf->b_evict_lock); 3656 hdr = buf->b_hdr; 3657 3658 /* this buffer is not on any list */ 3659 ASSERT(refcount_count(&hdr->b_refcnt) > 0); 3660 3661 if (hdr->b_state == arc_anon) { 3662 /* this buffer is already released */ 3663 ASSERT(buf->b_efunc == NULL); 3664 } else { 3665 hash_lock = HDR_LOCK(hdr); 3666 mutex_enter(hash_lock); 3667 hdr = buf->b_hdr; 3668 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 3669 } 3670 3671 l2hdr = hdr->b_l2hdr; 3672 if (l2hdr) { 3673 mutex_enter(&l2arc_buflist_mtx); 3674 hdr->b_l2hdr = NULL; 3675 list_remove(l2hdr->b_dev->l2ad_buflist, hdr); 3676 } 3677 buf_size = hdr->b_size; 3678 3679 /* 3680 * Do we have more than one buf? 3681 */ 3682 if (hdr->b_datacnt > 1) { 3683 arc_buf_hdr_t *nhdr; 3684 arc_buf_t **bufp; 3685 uint64_t blksz = hdr->b_size; 3686 uint64_t spa = hdr->b_spa; 3687 arc_buf_contents_t type = hdr->b_type; 3688 uint32_t flags = hdr->b_flags; 3689 3690 ASSERT(hdr->b_buf != buf || buf->b_next != NULL); 3691 /* 3692 * Pull the data off of this hdr and attach it to 3693 * a new anonymous hdr. 3694 */ 3695 (void) remove_reference(hdr, hash_lock, tag); 3696 bufp = &hdr->b_buf; 3697 while (*bufp != buf) 3698 bufp = &(*bufp)->b_next; 3699 *bufp = buf->b_next; 3700 buf->b_next = NULL; 3701 3702 ASSERT3U(hdr->b_state->arcs_size, >=, hdr->b_size); 3703 atomic_add_64(&hdr->b_state->arcs_size, -hdr->b_size); 3704 if (refcount_is_zero(&hdr->b_refcnt)) { 3705 uint64_t *size = &hdr->b_state->arcs_lsize[hdr->b_type]; 3706 ASSERT3U(*size, >=, hdr->b_size); 3707 atomic_add_64(size, -hdr->b_size); 3708 } 3709 3710 /* 3711 * We're releasing a duplicate user data buffer, update 3712 * our statistics accordingly. 3713 */ 3714 if (hdr->b_type == ARC_BUFC_DATA) { 3715 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers); 3716 ARCSTAT_INCR(arcstat_duplicate_buffers_size, 3717 -hdr->b_size); 3718 } 3719 hdr->b_datacnt -= 1; 3720 arc_cksum_verify(buf); 3721#ifdef illumos 3722 arc_buf_unwatch(buf); 3723#endif /* illumos */ 3724 3725 mutex_exit(hash_lock); 3726 3727 nhdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE); 3728 nhdr->b_size = blksz; 3729 nhdr->b_spa = spa; 3730 nhdr->b_type = type; 3731 nhdr->b_buf = buf; 3732 nhdr->b_state = arc_anon; 3733 nhdr->b_arc_access = 0; 3734 nhdr->b_flags = flags & ARC_L2_WRITING; 3735 nhdr->b_l2hdr = NULL; 3736 nhdr->b_datacnt = 1; 3737 nhdr->b_freeze_cksum = NULL; 3738 (void) refcount_add(&nhdr->b_refcnt, tag); 3739 buf->b_hdr = nhdr; 3740 mutex_exit(&buf->b_evict_lock); 3741 atomic_add_64(&arc_anon->arcs_size, blksz); 3742 } else { 3743 mutex_exit(&buf->b_evict_lock); 3744 ASSERT(refcount_count(&hdr->b_refcnt) == 1); 3745 ASSERT(!list_link_active(&hdr->b_arc_node)); 3746 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 3747 if (hdr->b_state != arc_anon) 3748 arc_change_state(arc_anon, hdr, hash_lock); 3749 hdr->b_arc_access = 0; 3750 if (hash_lock) 3751 mutex_exit(hash_lock); 3752 3753 buf_discard_identity(hdr); 3754 arc_buf_thaw(buf); 3755 } 3756 buf->b_efunc = NULL; 3757 buf->b_private = NULL; 3758 3759 if (l2hdr) { 3760 ARCSTAT_INCR(arcstat_l2_asize, -l2hdr->b_asize); 3761 vdev_space_update(l2hdr->b_dev->l2ad_vdev, 3762 -l2hdr->b_asize, 0, 0); 3763 trim_map_free(l2hdr->b_dev->l2ad_vdev, l2hdr->b_daddr, 3764 hdr->b_size, 0); 3765 kmem_free(l2hdr, sizeof (l2arc_buf_hdr_t)); 3766 ARCSTAT_INCR(arcstat_l2_size, -buf_size); 3767 mutex_exit(&l2arc_buflist_mtx); 3768 } 3769} 3770 3771int 3772arc_released(arc_buf_t *buf) 3773{ 3774 int released; 3775 3776 mutex_enter(&buf->b_evict_lock); 3777 released = (buf->b_data != NULL && buf->b_hdr->b_state == arc_anon); 3778 mutex_exit(&buf->b_evict_lock); 3779 return (released); 3780} 3781 3782#ifdef ZFS_DEBUG 3783int 3784arc_referenced(arc_buf_t *buf) 3785{ 3786 int referenced; 3787 3788 mutex_enter(&buf->b_evict_lock); 3789 referenced = (refcount_count(&buf->b_hdr->b_refcnt)); 3790 mutex_exit(&buf->b_evict_lock); 3791 return (referenced); 3792} 3793#endif 3794 3795static void 3796arc_write_ready(zio_t *zio) 3797{ 3798 arc_write_callback_t *callback = zio->io_private; 3799 arc_buf_t *buf = callback->awcb_buf; 3800 arc_buf_hdr_t *hdr = buf->b_hdr; 3801 3802 ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt)); 3803 callback->awcb_ready(zio, buf, callback->awcb_private); 3804 3805 /* 3806 * If the IO is already in progress, then this is a re-write 3807 * attempt, so we need to thaw and re-compute the cksum. 3808 * It is the responsibility of the callback to handle the 3809 * accounting for any re-write attempt. 3810 */ 3811 if (HDR_IO_IN_PROGRESS(hdr)) { 3812 mutex_enter(&hdr->b_freeze_lock); 3813 if (hdr->b_freeze_cksum != NULL) { 3814 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 3815 hdr->b_freeze_cksum = NULL; 3816 } 3817 mutex_exit(&hdr->b_freeze_lock); 3818 } 3819 arc_cksum_compute(buf, B_FALSE); 3820 hdr->b_flags |= ARC_IO_IN_PROGRESS; 3821} 3822 3823/* 3824 * The SPA calls this callback for each physical write that happens on behalf 3825 * of a logical write. See the comment in dbuf_write_physdone() for details. 3826 */ 3827static void 3828arc_write_physdone(zio_t *zio) 3829{ 3830 arc_write_callback_t *cb = zio->io_private; 3831 if (cb->awcb_physdone != NULL) 3832 cb->awcb_physdone(zio, cb->awcb_buf, cb->awcb_private); 3833} 3834 3835static void 3836arc_write_done(zio_t *zio) 3837{ 3838 arc_write_callback_t *callback = zio->io_private; 3839 arc_buf_t *buf = callback->awcb_buf; 3840 arc_buf_hdr_t *hdr = buf->b_hdr; 3841 3842 ASSERT(hdr->b_acb == NULL); 3843 3844 if (zio->io_error == 0) { 3845 if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) { 3846 buf_discard_identity(hdr); 3847 } else { 3848 hdr->b_dva = *BP_IDENTITY(zio->io_bp); 3849 hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp); 3850 hdr->b_cksum0 = zio->io_bp->blk_cksum.zc_word[0]; 3851 } 3852 } else { 3853 ASSERT(BUF_EMPTY(hdr)); 3854 } 3855 3856 /* 3857 * If the block to be written was all-zero or compressed enough to be 3858 * embedded in the BP, no write was performed so there will be no 3859 * dva/birth/checksum. The buffer must therefore remain anonymous 3860 * (and uncached). 3861 */ 3862 if (!BUF_EMPTY(hdr)) { 3863 arc_buf_hdr_t *exists; 3864 kmutex_t *hash_lock; 3865 3866 ASSERT(zio->io_error == 0); 3867 3868 arc_cksum_verify(buf); 3869 3870 exists = buf_hash_insert(hdr, &hash_lock); 3871 if (exists) { 3872 /* 3873 * This can only happen if we overwrite for 3874 * sync-to-convergence, because we remove 3875 * buffers from the hash table when we arc_free(). 3876 */ 3877 if (zio->io_flags & ZIO_FLAG_IO_REWRITE) { 3878 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) 3879 panic("bad overwrite, hdr=%p exists=%p", 3880 (void *)hdr, (void *)exists); 3881 ASSERT(refcount_is_zero(&exists->b_refcnt)); 3882 arc_change_state(arc_anon, exists, hash_lock); 3883 mutex_exit(hash_lock); 3884 arc_hdr_destroy(exists); 3885 exists = buf_hash_insert(hdr, &hash_lock); 3886 ASSERT3P(exists, ==, NULL); 3887 } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) { 3888 /* nopwrite */ 3889 ASSERT(zio->io_prop.zp_nopwrite); 3890 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) 3891 panic("bad nopwrite, hdr=%p exists=%p", 3892 (void *)hdr, (void *)exists); 3893 } else { 3894 /* Dedup */ 3895 ASSERT(hdr->b_datacnt == 1); 3896 ASSERT(hdr->b_state == arc_anon); 3897 ASSERT(BP_GET_DEDUP(zio->io_bp)); 3898 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0); 3899 } 3900 } 3901 hdr->b_flags &= ~ARC_IO_IN_PROGRESS; 3902 /* if it's not anon, we are doing a scrub */ 3903 if (!exists && hdr->b_state == arc_anon) 3904 arc_access(hdr, hash_lock); 3905 mutex_exit(hash_lock); 3906 } else { 3907 hdr->b_flags &= ~ARC_IO_IN_PROGRESS; 3908 } 3909 3910 ASSERT(!refcount_is_zero(&hdr->b_refcnt)); 3911 callback->awcb_done(zio, buf, callback->awcb_private); 3912 3913 kmem_free(callback, sizeof (arc_write_callback_t)); 3914} 3915 3916zio_t * 3917arc_write(zio_t *pio, spa_t *spa, uint64_t txg, 3918 blkptr_t *bp, arc_buf_t *buf, boolean_t l2arc, boolean_t l2arc_compress, 3919 const zio_prop_t *zp, arc_done_func_t *ready, arc_done_func_t *physdone, 3920 arc_done_func_t *done, void *private, zio_priority_t priority, 3921 int zio_flags, const zbookmark_phys_t *zb) 3922{ 3923 arc_buf_hdr_t *hdr = buf->b_hdr; 3924 arc_write_callback_t *callback; 3925 zio_t *zio; 3926 3927 ASSERT(ready != NULL); 3928 ASSERT(done != NULL); 3929 ASSERT(!HDR_IO_ERROR(hdr)); 3930 ASSERT((hdr->b_flags & ARC_IO_IN_PROGRESS) == 0); 3931 ASSERT(hdr->b_acb == NULL); 3932 if (l2arc) 3933 hdr->b_flags |= ARC_L2CACHE; 3934 if (l2arc_compress) 3935 hdr->b_flags |= ARC_L2COMPRESS; 3936 callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP); 3937 callback->awcb_ready = ready; 3938 callback->awcb_physdone = physdone; 3939 callback->awcb_done = done; 3940 callback->awcb_private = private; 3941 callback->awcb_buf = buf; 3942 3943 zio = zio_write(pio, spa, txg, bp, buf->b_data, hdr->b_size, zp, 3944 arc_write_ready, arc_write_physdone, arc_write_done, callback, 3945 priority, zio_flags, zb); 3946 3947 return (zio); 3948} 3949 3950static int 3951arc_memory_throttle(uint64_t reserve, uint64_t txg) 3952{ 3953#ifdef _KERNEL 3954 uint64_t available_memory = ptob(freemem); 3955 static uint64_t page_load = 0; 3956 static uint64_t last_txg = 0; 3957 3958#if defined(__i386) || !defined(UMA_MD_SMALL_ALLOC) 3959 available_memory = 3960 MIN(available_memory, ptob(vmem_size(heap_arena, VMEM_FREE))); 3961#endif 3962 3963 if (freemem > (uint64_t)physmem * arc_lotsfree_percent / 100) 3964 return (0); 3965 3966 if (txg > last_txg) { 3967 last_txg = txg; 3968 page_load = 0; 3969 } 3970 /* 3971 * If we are in pageout, we know that memory is already tight, 3972 * the arc is already going to be evicting, so we just want to 3973 * continue to let page writes occur as quickly as possible. 3974 */ 3975 if (curproc == pageproc) { 3976 if (page_load > MAX(ptob(minfree), available_memory) / 4) 3977 return (SET_ERROR(ERESTART)); 3978 /* Note: reserve is inflated, so we deflate */ 3979 page_load += reserve / 8; 3980 return (0); 3981 } else if (page_load > 0 && arc_reclaim_needed()) { 3982 /* memory is low, delay before restarting */ 3983 ARCSTAT_INCR(arcstat_memory_throttle_count, 1); 3984 return (SET_ERROR(EAGAIN)); 3985 } 3986 page_load = 0; 3987#endif 3988 return (0); 3989} 3990 3991void 3992arc_tempreserve_clear(uint64_t reserve) 3993{ 3994 atomic_add_64(&arc_tempreserve, -reserve); 3995 ASSERT((int64_t)arc_tempreserve >= 0); 3996} 3997 3998int 3999arc_tempreserve_space(uint64_t reserve, uint64_t txg) 4000{ 4001 int error; 4002 uint64_t anon_size; 4003 4004 if (reserve > arc_c/4 && !arc_no_grow) { 4005 arc_c = MIN(arc_c_max, reserve * 4); 4006 DTRACE_PROBE1(arc__set_reserve, uint64_t, arc_c); 4007 } 4008 if (reserve > arc_c) 4009 return (SET_ERROR(ENOMEM)); 4010 4011 /* 4012 * Don't count loaned bufs as in flight dirty data to prevent long 4013 * network delays from blocking transactions that are ready to be 4014 * assigned to a txg. 4015 */ 4016 anon_size = MAX((int64_t)(arc_anon->arcs_size - arc_loaned_bytes), 0); 4017 4018 /* 4019 * Writes will, almost always, require additional memory allocations 4020 * in order to compress/encrypt/etc the data. We therefore need to 4021 * make sure that there is sufficient available memory for this. 4022 */ 4023 error = arc_memory_throttle(reserve, txg); 4024 if (error != 0) 4025 return (error); 4026 4027 /* 4028 * Throttle writes when the amount of dirty data in the cache 4029 * gets too large. We try to keep the cache less than half full 4030 * of dirty blocks so that our sync times don't grow too large. 4031 * Note: if two requests come in concurrently, we might let them 4032 * both succeed, when one of them should fail. Not a huge deal. 4033 */ 4034 4035 if (reserve + arc_tempreserve + anon_size > arc_c / 2 && 4036 anon_size > arc_c / 4) { 4037 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK " 4038 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n", 4039 arc_tempreserve>>10, 4040 arc_anon->arcs_lsize[ARC_BUFC_METADATA]>>10, 4041 arc_anon->arcs_lsize[ARC_BUFC_DATA]>>10, 4042 reserve>>10, arc_c>>10); 4043 return (SET_ERROR(ERESTART)); 4044 } 4045 atomic_add_64(&arc_tempreserve, reserve); 4046 return (0); 4047} 4048 4049static kmutex_t arc_lowmem_lock; 4050#ifdef _KERNEL 4051static eventhandler_tag arc_event_lowmem = NULL; 4052 4053static void 4054arc_lowmem(void *arg __unused, int howto __unused) 4055{ 4056 4057 /* Serialize access via arc_lowmem_lock. */ 4058 mutex_enter(&arc_lowmem_lock); 4059 mutex_enter(&arc_reclaim_thr_lock); 4060 needfree = 1; 4061 DTRACE_PROBE(arc__needfree); 4062 cv_signal(&arc_reclaim_thr_cv); 4063 4064 /* 4065 * It is unsafe to block here in arbitrary threads, because we can come 4066 * here from ARC itself and may hold ARC locks and thus risk a deadlock 4067 * with ARC reclaim thread. 4068 */ 4069 if (curproc == pageproc) { 4070 while (needfree) 4071 msleep(&needfree, &arc_reclaim_thr_lock, 0, "zfs:lowmem", 0); 4072 } 4073 mutex_exit(&arc_reclaim_thr_lock); 4074 mutex_exit(&arc_lowmem_lock); 4075} 4076#endif 4077 4078void 4079arc_init(void) 4080{ 4081 int i, prefetch_tunable_set = 0; 4082 4083 mutex_init(&arc_reclaim_thr_lock, NULL, MUTEX_DEFAULT, NULL); 4084 cv_init(&arc_reclaim_thr_cv, NULL, CV_DEFAULT, NULL); 4085 mutex_init(&arc_lowmem_lock, NULL, MUTEX_DEFAULT, NULL); 4086 4087 /* Convert seconds to clock ticks */ 4088 arc_min_prefetch_lifespan = 1 * hz; 4089 4090 /* Start out with 1/8 of all memory */ 4091 arc_c = kmem_size() / 8; 4092 4093#ifdef sun 4094#ifdef _KERNEL 4095 /* 4096 * On architectures where the physical memory can be larger 4097 * than the addressable space (intel in 32-bit mode), we may 4098 * need to limit the cache to 1/8 of VM size. 4099 */ 4100 arc_c = MIN(arc_c, vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 8); 4101#endif 4102#endif /* sun */ 4103 /* set min cache to 1/32 of all memory, or 16MB, whichever is more */ 4104 arc_c_min = MAX(arc_c / 4, 64<<18); 4105 /* set max to 1/2 of all memory, or all but 1GB, whichever is more */ 4106 if (arc_c * 8 >= 1<<30) 4107 arc_c_max = (arc_c * 8) - (1<<30); 4108 else 4109 arc_c_max = arc_c_min; 4110 arc_c_max = MAX(arc_c * 5, arc_c_max); 4111 4112#ifdef _KERNEL 4113 /* 4114 * Allow the tunables to override our calculations if they are 4115 * reasonable (ie. over 16MB) 4116 */ 4117 if (zfs_arc_max > 64<<18 && zfs_arc_max < kmem_size()) 4118 arc_c_max = zfs_arc_max; 4119 if (zfs_arc_min > 64<<18 && zfs_arc_min <= arc_c_max) 4120 arc_c_min = zfs_arc_min; 4121#endif 4122 4123 arc_c = arc_c_max; 4124 arc_p = (arc_c >> 1); 4125 4126 /* limit meta-data to 1/4 of the arc capacity */ 4127 arc_meta_limit = arc_c_max / 4; 4128 4129 /* Allow the tunable to override if it is reasonable */ 4130 if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max) 4131 arc_meta_limit = zfs_arc_meta_limit; 4132 4133 if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0) 4134 arc_c_min = arc_meta_limit / 2; 4135 4136 if (zfs_arc_grow_retry > 0) 4137 arc_grow_retry = zfs_arc_grow_retry; 4138 4139 if (zfs_arc_shrink_shift > 0) 4140 arc_shrink_shift = zfs_arc_shrink_shift; 4141 4142 if (zfs_arc_p_min_shift > 0) 4143 arc_p_min_shift = zfs_arc_p_min_shift; 4144 4145 /* if kmem_flags are set, lets try to use less memory */ 4146 if (kmem_debugging()) 4147 arc_c = arc_c / 2; 4148 if (arc_c < arc_c_min) 4149 arc_c = arc_c_min; 4150 4151 zfs_arc_min = arc_c_min; 4152 zfs_arc_max = arc_c_max; 4153 4154 arc_anon = &ARC_anon; 4155 arc_mru = &ARC_mru; 4156 arc_mru_ghost = &ARC_mru_ghost; 4157 arc_mfu = &ARC_mfu; 4158 arc_mfu_ghost = &ARC_mfu_ghost; 4159 arc_l2c_only = &ARC_l2c_only; 4160 arc_size = 0; 4161 4162 for (i = 0; i < ARC_BUFC_NUMLISTS; i++) { 4163 mutex_init(&arc_anon->arcs_locks[i].arcs_lock, 4164 NULL, MUTEX_DEFAULT, NULL); 4165 mutex_init(&arc_mru->arcs_locks[i].arcs_lock, 4166 NULL, MUTEX_DEFAULT, NULL); 4167 mutex_init(&arc_mru_ghost->arcs_locks[i].arcs_lock, 4168 NULL, MUTEX_DEFAULT, NULL); 4169 mutex_init(&arc_mfu->arcs_locks[i].arcs_lock, 4170 NULL, MUTEX_DEFAULT, NULL); 4171 mutex_init(&arc_mfu_ghost->arcs_locks[i].arcs_lock, 4172 NULL, MUTEX_DEFAULT, NULL); 4173 mutex_init(&arc_l2c_only->arcs_locks[i].arcs_lock, 4174 NULL, MUTEX_DEFAULT, NULL); 4175 4176 list_create(&arc_mru->arcs_lists[i], 4177 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4178 list_create(&arc_mru_ghost->arcs_lists[i], 4179 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4180 list_create(&arc_mfu->arcs_lists[i], 4181 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4182 list_create(&arc_mfu_ghost->arcs_lists[i], 4183 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4184 list_create(&arc_mfu_ghost->arcs_lists[i], 4185 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4186 list_create(&arc_l2c_only->arcs_lists[i], 4187 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4188 } 4189 4190 buf_init(); 4191 4192 arc_thread_exit = 0; 4193 arc_eviction_list = NULL; 4194 mutex_init(&arc_eviction_mtx, NULL, MUTEX_DEFAULT, NULL); 4195 bzero(&arc_eviction_hdr, sizeof (arc_buf_hdr_t)); 4196 4197 arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED, 4198 sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL); 4199 4200 if (arc_ksp != NULL) { 4201 arc_ksp->ks_data = &arc_stats; 4202 kstat_install(arc_ksp); 4203 } 4204 4205 (void) thread_create(NULL, 0, arc_reclaim_thread, NULL, 0, &p0, 4206 TS_RUN, minclsyspri); 4207 4208#ifdef _KERNEL 4209 arc_event_lowmem = EVENTHANDLER_REGISTER(vm_lowmem, arc_lowmem, NULL, 4210 EVENTHANDLER_PRI_FIRST); 4211#endif 4212 4213 arc_dead = FALSE; 4214 arc_warm = B_FALSE; 4215 4216 /* 4217 * Calculate maximum amount of dirty data per pool. 4218 * 4219 * If it has been set by /etc/system, take that. 4220 * Otherwise, use a percentage of physical memory defined by 4221 * zfs_dirty_data_max_percent (default 10%) with a cap at 4222 * zfs_dirty_data_max_max (default 4GB). 4223 */ 4224 if (zfs_dirty_data_max == 0) { 4225 zfs_dirty_data_max = ptob(physmem) * 4226 zfs_dirty_data_max_percent / 100; 4227 zfs_dirty_data_max = MIN(zfs_dirty_data_max, 4228 zfs_dirty_data_max_max); 4229 } 4230 4231#ifdef _KERNEL 4232 if (TUNABLE_INT_FETCH("vfs.zfs.prefetch_disable", &zfs_prefetch_disable)) 4233 prefetch_tunable_set = 1; 4234 4235#ifdef __i386__ 4236 if (prefetch_tunable_set == 0) { 4237 printf("ZFS NOTICE: Prefetch is disabled by default on i386 " 4238 "-- to enable,\n"); 4239 printf(" add \"vfs.zfs.prefetch_disable=0\" " 4240 "to /boot/loader.conf.\n"); 4241 zfs_prefetch_disable = 1; 4242 } 4243#else 4244 if ((((uint64_t)physmem * PAGESIZE) < (1ULL << 32)) && 4245 prefetch_tunable_set == 0) { 4246 printf("ZFS NOTICE: Prefetch is disabled by default if less " 4247 "than 4GB of RAM is present;\n" 4248 " to enable, add \"vfs.zfs.prefetch_disable=0\" " 4249 "to /boot/loader.conf.\n"); 4250 zfs_prefetch_disable = 1; 4251 } 4252#endif 4253 /* Warn about ZFS memory and address space requirements. */ 4254 if (((uint64_t)physmem * PAGESIZE) < (256 + 128 + 64) * (1 << 20)) { 4255 printf("ZFS WARNING: Recommended minimum RAM size is 512MB; " 4256 "expect unstable behavior.\n"); 4257 } 4258 if (kmem_size() < 512 * (1 << 20)) { 4259 printf("ZFS WARNING: Recommended minimum kmem_size is 512MB; " 4260 "expect unstable behavior.\n"); 4261 printf(" Consider tuning vm.kmem_size and " 4262 "vm.kmem_size_max\n"); 4263 printf(" in /boot/loader.conf.\n"); 4264 } 4265#endif 4266} 4267 4268void 4269arc_fini(void) 4270{ 4271 int i; 4272 4273 mutex_enter(&arc_reclaim_thr_lock); 4274 arc_thread_exit = 1; 4275 cv_signal(&arc_reclaim_thr_cv); 4276 while (arc_thread_exit != 0) 4277 cv_wait(&arc_reclaim_thr_cv, &arc_reclaim_thr_lock); 4278 mutex_exit(&arc_reclaim_thr_lock); 4279 4280 arc_flush(NULL); 4281 4282 arc_dead = TRUE; 4283 4284 if (arc_ksp != NULL) { 4285 kstat_delete(arc_ksp); 4286 arc_ksp = NULL; 4287 } 4288 4289 mutex_destroy(&arc_eviction_mtx); 4290 mutex_destroy(&arc_reclaim_thr_lock); 4291 cv_destroy(&arc_reclaim_thr_cv); 4292 4293 for (i = 0; i < ARC_BUFC_NUMLISTS; i++) { 4294 list_destroy(&arc_mru->arcs_lists[i]); 4295 list_destroy(&arc_mru_ghost->arcs_lists[i]); 4296 list_destroy(&arc_mfu->arcs_lists[i]); 4297 list_destroy(&arc_mfu_ghost->arcs_lists[i]); 4298 list_destroy(&arc_l2c_only->arcs_lists[i]); 4299 4300 mutex_destroy(&arc_anon->arcs_locks[i].arcs_lock); 4301 mutex_destroy(&arc_mru->arcs_locks[i].arcs_lock); 4302 mutex_destroy(&arc_mru_ghost->arcs_locks[i].arcs_lock); 4303 mutex_destroy(&arc_mfu->arcs_locks[i].arcs_lock); 4304 mutex_destroy(&arc_mfu_ghost->arcs_locks[i].arcs_lock); 4305 mutex_destroy(&arc_l2c_only->arcs_locks[i].arcs_lock); 4306 } 4307 4308 buf_fini(); 4309 4310 ASSERT(arc_loaned_bytes == 0); 4311 4312 mutex_destroy(&arc_lowmem_lock); 4313#ifdef _KERNEL 4314 if (arc_event_lowmem != NULL) 4315 EVENTHANDLER_DEREGISTER(vm_lowmem, arc_event_lowmem); 4316#endif 4317} 4318 4319/* 4320 * Level 2 ARC 4321 * 4322 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk. 4323 * It uses dedicated storage devices to hold cached data, which are populated 4324 * using large infrequent writes. The main role of this cache is to boost 4325 * the performance of random read workloads. The intended L2ARC devices 4326 * include short-stroked disks, solid state disks, and other media with 4327 * substantially faster read latency than disk. 4328 * 4329 * +-----------------------+ 4330 * | ARC | 4331 * +-----------------------+ 4332 * | ^ ^ 4333 * | | | 4334 * l2arc_feed_thread() arc_read() 4335 * | | | 4336 * | l2arc read | 4337 * V | | 4338 * +---------------+ | 4339 * | L2ARC | | 4340 * +---------------+ | 4341 * | ^ | 4342 * l2arc_write() | | 4343 * | | | 4344 * V | | 4345 * +-------+ +-------+ 4346 * | vdev | | vdev | 4347 * | cache | | cache | 4348 * +-------+ +-------+ 4349 * +=========+ .-----. 4350 * : L2ARC : |-_____-| 4351 * : devices : | Disks | 4352 * +=========+ `-_____-' 4353 * 4354 * Read requests are satisfied from the following sources, in order: 4355 * 4356 * 1) ARC 4357 * 2) vdev cache of L2ARC devices 4358 * 3) L2ARC devices 4359 * 4) vdev cache of disks 4360 * 5) disks 4361 * 4362 * Some L2ARC device types exhibit extremely slow write performance. 4363 * To accommodate for this there are some significant differences between 4364 * the L2ARC and traditional cache design: 4365 * 4366 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from 4367 * the ARC behave as usual, freeing buffers and placing headers on ghost 4368 * lists. The ARC does not send buffers to the L2ARC during eviction as 4369 * this would add inflated write latencies for all ARC memory pressure. 4370 * 4371 * 2. The L2ARC attempts to cache data from the ARC before it is evicted. 4372 * It does this by periodically scanning buffers from the eviction-end of 4373 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are 4374 * not already there. It scans until a headroom of buffers is satisfied, 4375 * which itself is a buffer for ARC eviction. If a compressible buffer is 4376 * found during scanning and selected for writing to an L2ARC device, we 4377 * temporarily boost scanning headroom during the next scan cycle to make 4378 * sure we adapt to compression effects (which might significantly reduce 4379 * the data volume we write to L2ARC). The thread that does this is 4380 * l2arc_feed_thread(), illustrated below; example sizes are included to 4381 * provide a better sense of ratio than this diagram: 4382 * 4383 * head --> tail 4384 * +---------------------+----------+ 4385 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC 4386 * +---------------------+----------+ | o L2ARC eligible 4387 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer 4388 * +---------------------+----------+ | 4389 * 15.9 Gbytes ^ 32 Mbytes | 4390 * headroom | 4391 * l2arc_feed_thread() 4392 * | 4393 * l2arc write hand <--[oooo]--' 4394 * | 8 Mbyte 4395 * | write max 4396 * V 4397 * +==============================+ 4398 * L2ARC dev |####|#|###|###| |####| ... | 4399 * +==============================+ 4400 * 32 Gbytes 4401 * 4402 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of 4403 * evicted, then the L2ARC has cached a buffer much sooner than it probably 4404 * needed to, potentially wasting L2ARC device bandwidth and storage. It is 4405 * safe to say that this is an uncommon case, since buffers at the end of 4406 * the ARC lists have moved there due to inactivity. 4407 * 4408 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom, 4409 * then the L2ARC simply misses copying some buffers. This serves as a 4410 * pressure valve to prevent heavy read workloads from both stalling the ARC 4411 * with waits and clogging the L2ARC with writes. This also helps prevent 4412 * the potential for the L2ARC to churn if it attempts to cache content too 4413 * quickly, such as during backups of the entire pool. 4414 * 4415 * 5. After system boot and before the ARC has filled main memory, there are 4416 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru 4417 * lists can remain mostly static. Instead of searching from tail of these 4418 * lists as pictured, the l2arc_feed_thread() will search from the list heads 4419 * for eligible buffers, greatly increasing its chance of finding them. 4420 * 4421 * The L2ARC device write speed is also boosted during this time so that 4422 * the L2ARC warms up faster. Since there have been no ARC evictions yet, 4423 * there are no L2ARC reads, and no fear of degrading read performance 4424 * through increased writes. 4425 * 4426 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that 4427 * the vdev queue can aggregate them into larger and fewer writes. Each 4428 * device is written to in a rotor fashion, sweeping writes through 4429 * available space then repeating. 4430 * 4431 * 7. The L2ARC does not store dirty content. It never needs to flush 4432 * write buffers back to disk based storage. 4433 * 4434 * 8. If an ARC buffer is written (and dirtied) which also exists in the 4435 * L2ARC, the now stale L2ARC buffer is immediately dropped. 4436 * 4437 * The performance of the L2ARC can be tweaked by a number of tunables, which 4438 * may be necessary for different workloads: 4439 * 4440 * l2arc_write_max max write bytes per interval 4441 * l2arc_write_boost extra write bytes during device warmup 4442 * l2arc_noprefetch skip caching prefetched buffers 4443 * l2arc_headroom number of max device writes to precache 4444 * l2arc_headroom_boost when we find compressed buffers during ARC 4445 * scanning, we multiply headroom by this 4446 * percentage factor for the next scan cycle, 4447 * since more compressed buffers are likely to 4448 * be present 4449 * l2arc_feed_secs seconds between L2ARC writing 4450 * 4451 * Tunables may be removed or added as future performance improvements are 4452 * integrated, and also may become zpool properties. 4453 * 4454 * There are three key functions that control how the L2ARC warms up: 4455 * 4456 * l2arc_write_eligible() check if a buffer is eligible to cache 4457 * l2arc_write_size() calculate how much to write 4458 * l2arc_write_interval() calculate sleep delay between writes 4459 * 4460 * These three functions determine what to write, how much, and how quickly 4461 * to send writes. 4462 */ 4463 4464static boolean_t 4465l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *ab) 4466{ 4467 /* 4468 * A buffer is *not* eligible for the L2ARC if it: 4469 * 1. belongs to a different spa. 4470 * 2. is already cached on the L2ARC. 4471 * 3. has an I/O in progress (it may be an incomplete read). 4472 * 4. is flagged not eligible (zfs property). 4473 */ 4474 if (ab->b_spa != spa_guid) { 4475 ARCSTAT_BUMP(arcstat_l2_write_spa_mismatch); 4476 return (B_FALSE); 4477 } 4478 if (ab->b_l2hdr != NULL) { 4479 ARCSTAT_BUMP(arcstat_l2_write_in_l2); 4480 return (B_FALSE); 4481 } 4482 if (HDR_IO_IN_PROGRESS(ab)) { 4483 ARCSTAT_BUMP(arcstat_l2_write_hdr_io_in_progress); 4484 return (B_FALSE); 4485 } 4486 if (!HDR_L2CACHE(ab)) { 4487 ARCSTAT_BUMP(arcstat_l2_write_not_cacheable); 4488 return (B_FALSE); 4489 } 4490 4491 return (B_TRUE); 4492} 4493 4494static uint64_t 4495l2arc_write_size(void) 4496{ 4497 uint64_t size; 4498 4499 /* 4500 * Make sure our globals have meaningful values in case the user 4501 * altered them. 4502 */ 4503 size = l2arc_write_max; 4504 if (size == 0) { 4505 cmn_err(CE_NOTE, "Bad value for l2arc_write_max, value must " 4506 "be greater than zero, resetting it to the default (%d)", 4507 L2ARC_WRITE_SIZE); 4508 size = l2arc_write_max = L2ARC_WRITE_SIZE; 4509 } 4510 4511 if (arc_warm == B_FALSE) 4512 size += l2arc_write_boost; 4513 4514 return (size); 4515 4516} 4517 4518static clock_t 4519l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote) 4520{ 4521 clock_t interval, next, now; 4522 4523 /* 4524 * If the ARC lists are busy, increase our write rate; if the 4525 * lists are stale, idle back. This is achieved by checking 4526 * how much we previously wrote - if it was more than half of 4527 * what we wanted, schedule the next write much sooner. 4528 */ 4529 if (l2arc_feed_again && wrote > (wanted / 2)) 4530 interval = (hz * l2arc_feed_min_ms) / 1000; 4531 else 4532 interval = hz * l2arc_feed_secs; 4533 4534 now = ddi_get_lbolt(); 4535 next = MAX(now, MIN(now + interval, began + interval)); 4536 4537 return (next); 4538} 4539 4540static void 4541l2arc_hdr_stat_add(void) 4542{ 4543 ARCSTAT_INCR(arcstat_l2_hdr_size, HDR_SIZE + L2HDR_SIZE); 4544 ARCSTAT_INCR(arcstat_hdr_size, -HDR_SIZE); 4545} 4546 4547static void 4548l2arc_hdr_stat_remove(void) 4549{ 4550 ARCSTAT_INCR(arcstat_l2_hdr_size, -(HDR_SIZE + L2HDR_SIZE)); 4551 ARCSTAT_INCR(arcstat_hdr_size, HDR_SIZE); 4552} 4553 4554/* 4555 * Cycle through L2ARC devices. This is how L2ARC load balances. 4556 * If a device is returned, this also returns holding the spa config lock. 4557 */ 4558static l2arc_dev_t * 4559l2arc_dev_get_next(void) 4560{ 4561 l2arc_dev_t *first, *next = NULL; 4562 4563 /* 4564 * Lock out the removal of spas (spa_namespace_lock), then removal 4565 * of cache devices (l2arc_dev_mtx). Once a device has been selected, 4566 * both locks will be dropped and a spa config lock held instead. 4567 */ 4568 mutex_enter(&spa_namespace_lock); 4569 mutex_enter(&l2arc_dev_mtx); 4570 4571 /* if there are no vdevs, there is nothing to do */ 4572 if (l2arc_ndev == 0) 4573 goto out; 4574 4575 first = NULL; 4576 next = l2arc_dev_last; 4577 do { 4578 /* loop around the list looking for a non-faulted vdev */ 4579 if (next == NULL) { 4580 next = list_head(l2arc_dev_list); 4581 } else { 4582 next = list_next(l2arc_dev_list, next); 4583 if (next == NULL) 4584 next = list_head(l2arc_dev_list); 4585 } 4586 4587 /* if we have come back to the start, bail out */ 4588 if (first == NULL) 4589 first = next; 4590 else if (next == first) 4591 break; 4592 4593 } while (vdev_is_dead(next->l2ad_vdev)); 4594 4595 /* if we were unable to find any usable vdevs, return NULL */ 4596 if (vdev_is_dead(next->l2ad_vdev)) 4597 next = NULL; 4598 4599 l2arc_dev_last = next; 4600 4601out: 4602 mutex_exit(&l2arc_dev_mtx); 4603 4604 /* 4605 * Grab the config lock to prevent the 'next' device from being 4606 * removed while we are writing to it. 4607 */ 4608 if (next != NULL) 4609 spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER); 4610 mutex_exit(&spa_namespace_lock); 4611 4612 return (next); 4613} 4614 4615/* 4616 * Free buffers that were tagged for destruction. 4617 */ 4618static void 4619l2arc_do_free_on_write() 4620{ 4621 list_t *buflist; 4622 l2arc_data_free_t *df, *df_prev; 4623 4624 mutex_enter(&l2arc_free_on_write_mtx); 4625 buflist = l2arc_free_on_write; 4626 4627 for (df = list_tail(buflist); df; df = df_prev) { 4628 df_prev = list_prev(buflist, df); 4629 ASSERT(df->l2df_data != NULL); 4630 ASSERT(df->l2df_func != NULL); 4631 df->l2df_func(df->l2df_data, df->l2df_size); 4632 list_remove(buflist, df); 4633 kmem_free(df, sizeof (l2arc_data_free_t)); 4634 } 4635 4636 mutex_exit(&l2arc_free_on_write_mtx); 4637} 4638 4639/* 4640 * A write to a cache device has completed. Update all headers to allow 4641 * reads from these buffers to begin. 4642 */ 4643static void 4644l2arc_write_done(zio_t *zio) 4645{ 4646 l2arc_write_callback_t *cb; 4647 l2arc_dev_t *dev; 4648 list_t *buflist; 4649 arc_buf_hdr_t *head, *ab, *ab_prev; 4650 l2arc_buf_hdr_t *abl2; 4651 kmutex_t *hash_lock; 4652 int64_t bytes_dropped = 0; 4653 4654 cb = zio->io_private; 4655 ASSERT(cb != NULL); 4656 dev = cb->l2wcb_dev; 4657 ASSERT(dev != NULL); 4658 head = cb->l2wcb_head; 4659 ASSERT(head != NULL); 4660 buflist = dev->l2ad_buflist; 4661 ASSERT(buflist != NULL); 4662 DTRACE_PROBE2(l2arc__iodone, zio_t *, zio, 4663 l2arc_write_callback_t *, cb); 4664 4665 if (zio->io_error != 0) 4666 ARCSTAT_BUMP(arcstat_l2_writes_error); 4667 4668 mutex_enter(&l2arc_buflist_mtx); 4669 4670 /* 4671 * All writes completed, or an error was hit. 4672 */ 4673 for (ab = list_prev(buflist, head); ab; ab = ab_prev) { 4674 ab_prev = list_prev(buflist, ab); 4675 abl2 = ab->b_l2hdr; 4676 4677 /* 4678 * Release the temporary compressed buffer as soon as possible. 4679 */ 4680 if (abl2->b_compress != ZIO_COMPRESS_OFF) 4681 l2arc_release_cdata_buf(ab); 4682 4683 hash_lock = HDR_LOCK(ab); 4684 if (!mutex_tryenter(hash_lock)) { 4685 /* 4686 * This buffer misses out. It may be in a stage 4687 * of eviction. Its ARC_L2_WRITING flag will be 4688 * left set, denying reads to this buffer. 4689 */ 4690 ARCSTAT_BUMP(arcstat_l2_writes_hdr_miss); 4691 continue; 4692 } 4693 4694 if (zio->io_error != 0) { 4695 /* 4696 * Error - drop L2ARC entry. 4697 */ 4698 list_remove(buflist, ab); 4699 ARCSTAT_INCR(arcstat_l2_asize, -abl2->b_asize); 4700 bytes_dropped += abl2->b_asize; 4701 ab->b_l2hdr = NULL; 4702 trim_map_free(abl2->b_dev->l2ad_vdev, abl2->b_daddr, 4703 ab->b_size, 0); 4704 kmem_free(abl2, sizeof (l2arc_buf_hdr_t)); 4705 ARCSTAT_INCR(arcstat_l2_size, -ab->b_size); 4706 } 4707 4708 /* 4709 * Allow ARC to begin reads to this L2ARC entry. 4710 */ 4711 ab->b_flags &= ~ARC_L2_WRITING; 4712 4713 mutex_exit(hash_lock); 4714 } 4715 4716 atomic_inc_64(&l2arc_writes_done); 4717 list_remove(buflist, head); 4718 kmem_cache_free(hdr_cache, head); 4719 mutex_exit(&l2arc_buflist_mtx); 4720 4721 vdev_space_update(dev->l2ad_vdev, -bytes_dropped, 0, 0); 4722 4723 l2arc_do_free_on_write(); 4724 4725 kmem_free(cb, sizeof (l2arc_write_callback_t)); 4726} 4727 4728/* 4729 * A read to a cache device completed. Validate buffer contents before 4730 * handing over to the regular ARC routines. 4731 */ 4732static void 4733l2arc_read_done(zio_t *zio) 4734{ 4735 l2arc_read_callback_t *cb; 4736 arc_buf_hdr_t *hdr; 4737 arc_buf_t *buf; 4738 kmutex_t *hash_lock; 4739 int equal; 4740 4741 ASSERT(zio->io_vd != NULL); 4742 ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE); 4743 4744 spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd); 4745 4746 cb = zio->io_private; 4747 ASSERT(cb != NULL); 4748 buf = cb->l2rcb_buf; 4749 ASSERT(buf != NULL); 4750 4751 hash_lock = HDR_LOCK(buf->b_hdr); 4752 mutex_enter(hash_lock); 4753 hdr = buf->b_hdr; 4754 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 4755 4756 /* 4757 * If the buffer was compressed, decompress it first. 4758 */ 4759 if (cb->l2rcb_compress != ZIO_COMPRESS_OFF) 4760 l2arc_decompress_zio(zio, hdr, cb->l2rcb_compress); 4761 ASSERT(zio->io_data != NULL); 4762 4763 /* 4764 * Check this survived the L2ARC journey. 4765 */ 4766 equal = arc_cksum_equal(buf); 4767 if (equal && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) { 4768 mutex_exit(hash_lock); 4769 zio->io_private = buf; 4770 zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */ 4771 zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */ 4772 arc_read_done(zio); 4773 } else { 4774 mutex_exit(hash_lock); 4775 /* 4776 * Buffer didn't survive caching. Increment stats and 4777 * reissue to the original storage device. 4778 */ 4779 if (zio->io_error != 0) { 4780 ARCSTAT_BUMP(arcstat_l2_io_error); 4781 } else { 4782 zio->io_error = SET_ERROR(EIO); 4783 } 4784 if (!equal) 4785 ARCSTAT_BUMP(arcstat_l2_cksum_bad); 4786 4787 /* 4788 * If there's no waiter, issue an async i/o to the primary 4789 * storage now. If there *is* a waiter, the caller must 4790 * issue the i/o in a context where it's OK to block. 4791 */ 4792 if (zio->io_waiter == NULL) { 4793 zio_t *pio = zio_unique_parent(zio); 4794 4795 ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL); 4796 4797 zio_nowait(zio_read(pio, cb->l2rcb_spa, &cb->l2rcb_bp, 4798 buf->b_data, zio->io_size, arc_read_done, buf, 4799 zio->io_priority, cb->l2rcb_flags, &cb->l2rcb_zb)); 4800 } 4801 } 4802 4803 kmem_free(cb, sizeof (l2arc_read_callback_t)); 4804} 4805 4806/* 4807 * This is the list priority from which the L2ARC will search for pages to 4808 * cache. This is used within loops (0..3) to cycle through lists in the 4809 * desired order. This order can have a significant effect on cache 4810 * performance. 4811 * 4812 * Currently the metadata lists are hit first, MFU then MRU, followed by 4813 * the data lists. This function returns a locked list, and also returns 4814 * the lock pointer. 4815 */ 4816static list_t * 4817l2arc_list_locked(int list_num, kmutex_t **lock) 4818{ 4819 list_t *list = NULL; 4820 int idx; 4821 4822 ASSERT(list_num >= 0 && list_num < 2 * ARC_BUFC_NUMLISTS); 4823 4824 if (list_num < ARC_BUFC_NUMMETADATALISTS) { 4825 idx = list_num; 4826 list = &arc_mfu->arcs_lists[idx]; 4827 *lock = ARCS_LOCK(arc_mfu, idx); 4828 } else if (list_num < ARC_BUFC_NUMMETADATALISTS * 2) { 4829 idx = list_num - ARC_BUFC_NUMMETADATALISTS; 4830 list = &arc_mru->arcs_lists[idx]; 4831 *lock = ARCS_LOCK(arc_mru, idx); 4832 } else if (list_num < (ARC_BUFC_NUMMETADATALISTS * 2 + 4833 ARC_BUFC_NUMDATALISTS)) { 4834 idx = list_num - ARC_BUFC_NUMMETADATALISTS; 4835 list = &arc_mfu->arcs_lists[idx]; 4836 *lock = ARCS_LOCK(arc_mfu, idx); 4837 } else { 4838 idx = list_num - ARC_BUFC_NUMLISTS; 4839 list = &arc_mru->arcs_lists[idx]; 4840 *lock = ARCS_LOCK(arc_mru, idx); 4841 } 4842 4843 ASSERT(!(MUTEX_HELD(*lock))); 4844 mutex_enter(*lock); 4845 return (list); 4846} 4847 4848/* 4849 * Evict buffers from the device write hand to the distance specified in 4850 * bytes. This distance may span populated buffers, it may span nothing. 4851 * This is clearing a region on the L2ARC device ready for writing. 4852 * If the 'all' boolean is set, every buffer is evicted. 4853 */ 4854static void 4855l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all) 4856{ 4857 list_t *buflist; 4858 l2arc_buf_hdr_t *abl2; 4859 arc_buf_hdr_t *ab, *ab_prev; 4860 kmutex_t *hash_lock; 4861 uint64_t taddr; 4862 int64_t bytes_evicted = 0; 4863 4864 buflist = dev->l2ad_buflist; 4865 4866 if (buflist == NULL) 4867 return; 4868 4869 if (!all && dev->l2ad_first) { 4870 /* 4871 * This is the first sweep through the device. There is 4872 * nothing to evict. 4873 */ 4874 return; 4875 } 4876 4877 if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) { 4878 /* 4879 * When nearing the end of the device, evict to the end 4880 * before the device write hand jumps to the start. 4881 */ 4882 taddr = dev->l2ad_end; 4883 } else { 4884 taddr = dev->l2ad_hand + distance; 4885 } 4886 DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist, 4887 uint64_t, taddr, boolean_t, all); 4888 4889top: 4890 mutex_enter(&l2arc_buflist_mtx); 4891 for (ab = list_tail(buflist); ab; ab = ab_prev) { 4892 ab_prev = list_prev(buflist, ab); 4893 4894 hash_lock = HDR_LOCK(ab); 4895 if (!mutex_tryenter(hash_lock)) { 4896 /* 4897 * Missed the hash lock. Retry. 4898 */ 4899 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry); 4900 mutex_exit(&l2arc_buflist_mtx); 4901 mutex_enter(hash_lock); 4902 mutex_exit(hash_lock); 4903 goto top; 4904 } 4905 4906 if (HDR_L2_WRITE_HEAD(ab)) { 4907 /* 4908 * We hit a write head node. Leave it for 4909 * l2arc_write_done(). 4910 */ 4911 list_remove(buflist, ab); 4912 mutex_exit(hash_lock); 4913 continue; 4914 } 4915 4916 if (!all && ab->b_l2hdr != NULL && 4917 (ab->b_l2hdr->b_daddr > taddr || 4918 ab->b_l2hdr->b_daddr < dev->l2ad_hand)) { 4919 /* 4920 * We've evicted to the target address, 4921 * or the end of the device. 4922 */ 4923 mutex_exit(hash_lock); 4924 break; 4925 } 4926 4927 if (HDR_FREE_IN_PROGRESS(ab)) { 4928 /* 4929 * Already on the path to destruction. 4930 */ 4931 mutex_exit(hash_lock); 4932 continue; 4933 } 4934 4935 if (ab->b_state == arc_l2c_only) { 4936 ASSERT(!HDR_L2_READING(ab)); 4937 /* 4938 * This doesn't exist in the ARC. Destroy. 4939 * arc_hdr_destroy() will call list_remove() 4940 * and decrement arcstat_l2_size. 4941 */ 4942 arc_change_state(arc_anon, ab, hash_lock); 4943 arc_hdr_destroy(ab); 4944 } else { 4945 /* 4946 * Invalidate issued or about to be issued 4947 * reads, since we may be about to write 4948 * over this location. 4949 */ 4950 if (HDR_L2_READING(ab)) { 4951 ARCSTAT_BUMP(arcstat_l2_evict_reading); 4952 ab->b_flags |= ARC_L2_EVICTED; 4953 } 4954 4955 /* 4956 * Tell ARC this no longer exists in L2ARC. 4957 */ 4958 if (ab->b_l2hdr != NULL) { 4959 abl2 = ab->b_l2hdr; 4960 ARCSTAT_INCR(arcstat_l2_asize, -abl2->b_asize); 4961 bytes_evicted += abl2->b_asize; 4962 ab->b_l2hdr = NULL; 4963 kmem_free(abl2, sizeof (l2arc_buf_hdr_t)); 4964 ARCSTAT_INCR(arcstat_l2_size, -ab->b_size); 4965 } 4966 list_remove(buflist, ab); 4967 4968 /* 4969 * This may have been leftover after a 4970 * failed write. 4971 */ 4972 ab->b_flags &= ~ARC_L2_WRITING; 4973 } 4974 mutex_exit(hash_lock); 4975 } 4976 mutex_exit(&l2arc_buflist_mtx); 4977 4978 vdev_space_update(dev->l2ad_vdev, -bytes_evicted, 0, 0); 4979 dev->l2ad_evict = taddr; 4980} 4981 4982/* 4983 * Find and write ARC buffers to the L2ARC device. 4984 * 4985 * An ARC_L2_WRITING flag is set so that the L2ARC buffers are not valid 4986 * for reading until they have completed writing. 4987 * The headroom_boost is an in-out parameter used to maintain headroom boost 4988 * state between calls to this function. 4989 * 4990 * Returns the number of bytes actually written (which may be smaller than 4991 * the delta by which the device hand has changed due to alignment). 4992 */ 4993static uint64_t 4994l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz, 4995 boolean_t *headroom_boost) 4996{ 4997 arc_buf_hdr_t *ab, *ab_prev, *head; 4998 list_t *list; 4999 uint64_t write_asize, write_psize, write_sz, headroom, 5000 buf_compress_minsz; 5001 void *buf_data; 5002 kmutex_t *list_lock; 5003 boolean_t full; 5004 l2arc_write_callback_t *cb; 5005 zio_t *pio, *wzio; 5006 uint64_t guid = spa_load_guid(spa); 5007 const boolean_t do_headroom_boost = *headroom_boost; 5008 int try; 5009 5010 ASSERT(dev->l2ad_vdev != NULL); 5011 5012 /* Lower the flag now, we might want to raise it again later. */ 5013 *headroom_boost = B_FALSE; 5014 5015 pio = NULL; 5016 write_sz = write_asize = write_psize = 0; 5017 full = B_FALSE; 5018 head = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE); 5019 head->b_flags |= ARC_L2_WRITE_HEAD; 5020 5021 ARCSTAT_BUMP(arcstat_l2_write_buffer_iter); 5022 /* 5023 * We will want to try to compress buffers that are at least 2x the 5024 * device sector size. 5025 */ 5026 buf_compress_minsz = 2 << dev->l2ad_vdev->vdev_ashift; 5027 5028 /* 5029 * Copy buffers for L2ARC writing. 5030 */ 5031 mutex_enter(&l2arc_buflist_mtx); 5032 for (try = 0; try < 2 * ARC_BUFC_NUMLISTS; try++) { 5033 uint64_t passed_sz = 0; 5034 5035 list = l2arc_list_locked(try, &list_lock); 5036 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_iter); 5037 5038 /* 5039 * L2ARC fast warmup. 5040 * 5041 * Until the ARC is warm and starts to evict, read from the 5042 * head of the ARC lists rather than the tail. 5043 */ 5044 if (arc_warm == B_FALSE) 5045 ab = list_head(list); 5046 else 5047 ab = list_tail(list); 5048 if (ab == NULL) 5049 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_null_iter); 5050 5051 headroom = target_sz * l2arc_headroom; 5052 if (do_headroom_boost) 5053 headroom = (headroom * l2arc_headroom_boost) / 100; 5054 5055 for (; ab; ab = ab_prev) { 5056 l2arc_buf_hdr_t *l2hdr; 5057 kmutex_t *hash_lock; 5058 uint64_t buf_sz; 5059 5060 if (arc_warm == B_FALSE) 5061 ab_prev = list_next(list, ab); 5062 else 5063 ab_prev = list_prev(list, ab); 5064 ARCSTAT_INCR(arcstat_l2_write_buffer_bytes_scanned, ab->b_size); 5065 5066 hash_lock = HDR_LOCK(ab); 5067 if (!mutex_tryenter(hash_lock)) { 5068 ARCSTAT_BUMP(arcstat_l2_write_trylock_fail); 5069 /* 5070 * Skip this buffer rather than waiting. 5071 */ 5072 continue; 5073 } 5074 5075 passed_sz += ab->b_size; 5076 if (passed_sz > headroom) { 5077 /* 5078 * Searched too far. 5079 */ 5080 mutex_exit(hash_lock); 5081 ARCSTAT_BUMP(arcstat_l2_write_passed_headroom); 5082 break; 5083 } 5084 5085 if (!l2arc_write_eligible(guid, ab)) { 5086 mutex_exit(hash_lock); 5087 continue; 5088 } 5089 5090 if ((write_sz + ab->b_size) > target_sz) { 5091 full = B_TRUE; 5092 mutex_exit(hash_lock); 5093 ARCSTAT_BUMP(arcstat_l2_write_full); 5094 break; 5095 } 5096 5097 if (pio == NULL) { 5098 /* 5099 * Insert a dummy header on the buflist so 5100 * l2arc_write_done() can find where the 5101 * write buffers begin without searching. 5102 */ 5103 list_insert_head(dev->l2ad_buflist, head); 5104 5105 cb = kmem_alloc( 5106 sizeof (l2arc_write_callback_t), KM_SLEEP); 5107 cb->l2wcb_dev = dev; 5108 cb->l2wcb_head = head; 5109 pio = zio_root(spa, l2arc_write_done, cb, 5110 ZIO_FLAG_CANFAIL); 5111 ARCSTAT_BUMP(arcstat_l2_write_pios); 5112 } 5113 5114 /* 5115 * Create and add a new L2ARC header. 5116 */ 5117 l2hdr = kmem_zalloc(sizeof (l2arc_buf_hdr_t), KM_SLEEP); 5118 l2hdr->b_dev = dev; 5119 ab->b_flags |= ARC_L2_WRITING; 5120 5121 /* 5122 * Temporarily stash the data buffer in b_tmp_cdata. 5123 * The subsequent write step will pick it up from 5124 * there. This is because can't access ab->b_buf 5125 * without holding the hash_lock, which we in turn 5126 * can't access without holding the ARC list locks 5127 * (which we want to avoid during compression/writing). 5128 */ 5129 l2hdr->b_compress = ZIO_COMPRESS_OFF; 5130 l2hdr->b_asize = ab->b_size; 5131 l2hdr->b_tmp_cdata = ab->b_buf->b_data; 5132 5133 buf_sz = ab->b_size; 5134 ab->b_l2hdr = l2hdr; 5135 5136 list_insert_head(dev->l2ad_buflist, ab); 5137 5138 /* 5139 * Compute and store the buffer cksum before 5140 * writing. On debug the cksum is verified first. 5141 */ 5142 arc_cksum_verify(ab->b_buf); 5143 arc_cksum_compute(ab->b_buf, B_TRUE); 5144 5145 mutex_exit(hash_lock); 5146 5147 write_sz += buf_sz; 5148 } 5149 5150 mutex_exit(list_lock); 5151 5152 if (full == B_TRUE) 5153 break; 5154 } 5155 5156 /* No buffers selected for writing? */ 5157 if (pio == NULL) { 5158 ASSERT0(write_sz); 5159 mutex_exit(&l2arc_buflist_mtx); 5160 kmem_cache_free(hdr_cache, head); 5161 return (0); 5162 } 5163 5164 /* 5165 * Now start writing the buffers. We're starting at the write head 5166 * and work backwards, retracing the course of the buffer selector 5167 * loop above. 5168 */ 5169 for (ab = list_prev(dev->l2ad_buflist, head); ab; 5170 ab = list_prev(dev->l2ad_buflist, ab)) { 5171 l2arc_buf_hdr_t *l2hdr; 5172 uint64_t buf_sz; 5173 5174 /* 5175 * We shouldn't need to lock the buffer here, since we flagged 5176 * it as ARC_L2_WRITING in the previous step, but we must take 5177 * care to only access its L2 cache parameters. In particular, 5178 * ab->b_buf may be invalid by now due to ARC eviction. 5179 */ 5180 l2hdr = ab->b_l2hdr; 5181 l2hdr->b_daddr = dev->l2ad_hand; 5182 5183 if ((ab->b_flags & ARC_L2COMPRESS) && 5184 l2hdr->b_asize >= buf_compress_minsz) { 5185 if (l2arc_compress_buf(l2hdr)) { 5186 /* 5187 * If compression succeeded, enable headroom 5188 * boost on the next scan cycle. 5189 */ 5190 *headroom_boost = B_TRUE; 5191 } 5192 } 5193 5194 /* 5195 * Pick up the buffer data we had previously stashed away 5196 * (and now potentially also compressed). 5197 */ 5198 buf_data = l2hdr->b_tmp_cdata; 5199 buf_sz = l2hdr->b_asize; 5200 5201 /* Compression may have squashed the buffer to zero length. */ 5202 if (buf_sz != 0) { 5203 uint64_t buf_p_sz; 5204 5205 wzio = zio_write_phys(pio, dev->l2ad_vdev, 5206 dev->l2ad_hand, buf_sz, buf_data, ZIO_CHECKSUM_OFF, 5207 NULL, NULL, ZIO_PRIORITY_ASYNC_WRITE, 5208 ZIO_FLAG_CANFAIL, B_FALSE); 5209 5210 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev, 5211 zio_t *, wzio); 5212 (void) zio_nowait(wzio); 5213 5214 write_asize += buf_sz; 5215 /* 5216 * Keep the clock hand suitably device-aligned. 5217 */ 5218 buf_p_sz = vdev_psize_to_asize(dev->l2ad_vdev, buf_sz); 5219 write_psize += buf_p_sz; 5220 dev->l2ad_hand += buf_p_sz; 5221 } 5222 } 5223 5224 mutex_exit(&l2arc_buflist_mtx); 5225 5226 ASSERT3U(write_asize, <=, target_sz); 5227 ARCSTAT_BUMP(arcstat_l2_writes_sent); 5228 ARCSTAT_INCR(arcstat_l2_write_bytes, write_asize); 5229 ARCSTAT_INCR(arcstat_l2_size, write_sz); 5230 ARCSTAT_INCR(arcstat_l2_asize, write_asize); 5231 vdev_space_update(dev->l2ad_vdev, write_psize, 0, 0); 5232 5233 /* 5234 * Bump device hand to the device start if it is approaching the end. 5235 * l2arc_evict() will already have evicted ahead for this case. 5236 */ 5237 if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) { 5238 dev->l2ad_hand = dev->l2ad_start; 5239 dev->l2ad_evict = dev->l2ad_start; 5240 dev->l2ad_first = B_FALSE; 5241 } 5242 5243 dev->l2ad_writing = B_TRUE; 5244 (void) zio_wait(pio); 5245 dev->l2ad_writing = B_FALSE; 5246 5247 return (write_asize); 5248} 5249 5250/* 5251 * Compresses an L2ARC buffer. 5252 * The data to be compressed must be prefilled in l2hdr->b_tmp_cdata and its 5253 * size in l2hdr->b_asize. This routine tries to compress the data and 5254 * depending on the compression result there are three possible outcomes: 5255 * *) The buffer was incompressible. The original l2hdr contents were left 5256 * untouched and are ready for writing to an L2 device. 5257 * *) The buffer was all-zeros, so there is no need to write it to an L2 5258 * device. To indicate this situation b_tmp_cdata is NULL'ed, b_asize is 5259 * set to zero and b_compress is set to ZIO_COMPRESS_EMPTY. 5260 * *) Compression succeeded and b_tmp_cdata was replaced with a temporary 5261 * data buffer which holds the compressed data to be written, and b_asize 5262 * tells us how much data there is. b_compress is set to the appropriate 5263 * compression algorithm. Once writing is done, invoke 5264 * l2arc_release_cdata_buf on this l2hdr to free this temporary buffer. 5265 * 5266 * Returns B_TRUE if compression succeeded, or B_FALSE if it didn't (the 5267 * buffer was incompressible). 5268 */ 5269static boolean_t 5270l2arc_compress_buf(l2arc_buf_hdr_t *l2hdr) 5271{ 5272 void *cdata; 5273 size_t csize, len, rounded; 5274 5275 ASSERT(l2hdr->b_compress == ZIO_COMPRESS_OFF); 5276 ASSERT(l2hdr->b_tmp_cdata != NULL); 5277 5278 len = l2hdr->b_asize; 5279 cdata = zio_data_buf_alloc(len); 5280 csize = zio_compress_data(ZIO_COMPRESS_LZ4, l2hdr->b_tmp_cdata, 5281 cdata, l2hdr->b_asize); 5282 5283 rounded = P2ROUNDUP(csize, (size_t)SPA_MINBLOCKSIZE); 5284 if (rounded > csize) { 5285 bzero((char *)cdata + csize, rounded - csize); 5286 csize = rounded; 5287 } 5288 5289 if (csize == 0) { 5290 /* zero block, indicate that there's nothing to write */ 5291 zio_data_buf_free(cdata, len); 5292 l2hdr->b_compress = ZIO_COMPRESS_EMPTY; 5293 l2hdr->b_asize = 0; 5294 l2hdr->b_tmp_cdata = NULL; 5295 ARCSTAT_BUMP(arcstat_l2_compress_zeros); 5296 return (B_TRUE); 5297 } else if (csize > 0 && csize < len) { 5298 /* 5299 * Compression succeeded, we'll keep the cdata around for 5300 * writing and release it afterwards. 5301 */ 5302 l2hdr->b_compress = ZIO_COMPRESS_LZ4; 5303 l2hdr->b_asize = csize; 5304 l2hdr->b_tmp_cdata = cdata; 5305 ARCSTAT_BUMP(arcstat_l2_compress_successes); 5306 return (B_TRUE); 5307 } else { 5308 /* 5309 * Compression failed, release the compressed buffer. 5310 * l2hdr will be left unmodified. 5311 */ 5312 zio_data_buf_free(cdata, len); 5313 ARCSTAT_BUMP(arcstat_l2_compress_failures); 5314 return (B_FALSE); 5315 } 5316} 5317 5318/* 5319 * Decompresses a zio read back from an l2arc device. On success, the 5320 * underlying zio's io_data buffer is overwritten by the uncompressed 5321 * version. On decompression error (corrupt compressed stream), the 5322 * zio->io_error value is set to signal an I/O error. 5323 * 5324 * Please note that the compressed data stream is not checksummed, so 5325 * if the underlying device is experiencing data corruption, we may feed 5326 * corrupt data to the decompressor, so the decompressor needs to be 5327 * able to handle this situation (LZ4 does). 5328 */ 5329static void 5330l2arc_decompress_zio(zio_t *zio, arc_buf_hdr_t *hdr, enum zio_compress c) 5331{ 5332 ASSERT(L2ARC_IS_VALID_COMPRESS(c)); 5333 5334 if (zio->io_error != 0) { 5335 /* 5336 * An io error has occured, just restore the original io 5337 * size in preparation for a main pool read. 5338 */ 5339 zio->io_orig_size = zio->io_size = hdr->b_size; 5340 return; 5341 } 5342 5343 if (c == ZIO_COMPRESS_EMPTY) { 5344 /* 5345 * An empty buffer results in a null zio, which means we 5346 * need to fill its io_data after we're done restoring the 5347 * buffer's contents. 5348 */ 5349 ASSERT(hdr->b_buf != NULL); 5350 bzero(hdr->b_buf->b_data, hdr->b_size); 5351 zio->io_data = zio->io_orig_data = hdr->b_buf->b_data; 5352 } else { 5353 ASSERT(zio->io_data != NULL); 5354 /* 5355 * We copy the compressed data from the start of the arc buffer 5356 * (the zio_read will have pulled in only what we need, the 5357 * rest is garbage which we will overwrite at decompression) 5358 * and then decompress back to the ARC data buffer. This way we 5359 * can minimize copying by simply decompressing back over the 5360 * original compressed data (rather than decompressing to an 5361 * aux buffer and then copying back the uncompressed buffer, 5362 * which is likely to be much larger). 5363 */ 5364 uint64_t csize; 5365 void *cdata; 5366 5367 csize = zio->io_size; 5368 cdata = zio_data_buf_alloc(csize); 5369 bcopy(zio->io_data, cdata, csize); 5370 if (zio_decompress_data(c, cdata, zio->io_data, csize, 5371 hdr->b_size) != 0) 5372 zio->io_error = EIO; 5373 zio_data_buf_free(cdata, csize); 5374 } 5375 5376 /* Restore the expected uncompressed IO size. */ 5377 zio->io_orig_size = zio->io_size = hdr->b_size; 5378} 5379 5380/* 5381 * Releases the temporary b_tmp_cdata buffer in an l2arc header structure. 5382 * This buffer serves as a temporary holder of compressed data while 5383 * the buffer entry is being written to an l2arc device. Once that is 5384 * done, we can dispose of it. 5385 */ 5386static void 5387l2arc_release_cdata_buf(arc_buf_hdr_t *ab) 5388{ 5389 l2arc_buf_hdr_t *l2hdr = ab->b_l2hdr; 5390 5391 if (l2hdr->b_compress == ZIO_COMPRESS_LZ4) { 5392 /* 5393 * If the data was compressed, then we've allocated a 5394 * temporary buffer for it, so now we need to release it. 5395 */ 5396 ASSERT(l2hdr->b_tmp_cdata != NULL); 5397 zio_data_buf_free(l2hdr->b_tmp_cdata, ab->b_size); 5398 } 5399 l2hdr->b_tmp_cdata = NULL; 5400} 5401 5402/* 5403 * This thread feeds the L2ARC at regular intervals. This is the beating 5404 * heart of the L2ARC. 5405 */ 5406static void 5407l2arc_feed_thread(void *dummy __unused) 5408{ 5409 callb_cpr_t cpr; 5410 l2arc_dev_t *dev; 5411 spa_t *spa; 5412 uint64_t size, wrote; 5413 clock_t begin, next = ddi_get_lbolt(); 5414 boolean_t headroom_boost = B_FALSE; 5415 5416 CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG); 5417 5418 mutex_enter(&l2arc_feed_thr_lock); 5419 5420 while (l2arc_thread_exit == 0) { 5421 CALLB_CPR_SAFE_BEGIN(&cpr); 5422 (void) cv_timedwait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock, 5423 next - ddi_get_lbolt()); 5424 CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock); 5425 next = ddi_get_lbolt() + hz; 5426 5427 /* 5428 * Quick check for L2ARC devices. 5429 */ 5430 mutex_enter(&l2arc_dev_mtx); 5431 if (l2arc_ndev == 0) { 5432 mutex_exit(&l2arc_dev_mtx); 5433 continue; 5434 } 5435 mutex_exit(&l2arc_dev_mtx); 5436 begin = ddi_get_lbolt(); 5437 5438 /* 5439 * This selects the next l2arc device to write to, and in 5440 * doing so the next spa to feed from: dev->l2ad_spa. This 5441 * will return NULL if there are now no l2arc devices or if 5442 * they are all faulted. 5443 * 5444 * If a device is returned, its spa's config lock is also 5445 * held to prevent device removal. l2arc_dev_get_next() 5446 * will grab and release l2arc_dev_mtx. 5447 */ 5448 if ((dev = l2arc_dev_get_next()) == NULL) 5449 continue; 5450 5451 spa = dev->l2ad_spa; 5452 ASSERT(spa != NULL); 5453 5454 /* 5455 * If the pool is read-only then force the feed thread to 5456 * sleep a little longer. 5457 */ 5458 if (!spa_writeable(spa)) { 5459 next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz; 5460 spa_config_exit(spa, SCL_L2ARC, dev); 5461 continue; 5462 } 5463 5464 /* 5465 * Avoid contributing to memory pressure. 5466 */ 5467 if (arc_reclaim_needed()) { 5468 ARCSTAT_BUMP(arcstat_l2_abort_lowmem); 5469 spa_config_exit(spa, SCL_L2ARC, dev); 5470 continue; 5471 } 5472 5473 ARCSTAT_BUMP(arcstat_l2_feeds); 5474 5475 size = l2arc_write_size(); 5476 5477 /* 5478 * Evict L2ARC buffers that will be overwritten. 5479 */ 5480 l2arc_evict(dev, size, B_FALSE); 5481 5482 /* 5483 * Write ARC buffers. 5484 */ 5485 wrote = l2arc_write_buffers(spa, dev, size, &headroom_boost); 5486 5487 /* 5488 * Calculate interval between writes. 5489 */ 5490 next = l2arc_write_interval(begin, size, wrote); 5491 spa_config_exit(spa, SCL_L2ARC, dev); 5492 } 5493 5494 l2arc_thread_exit = 0; 5495 cv_broadcast(&l2arc_feed_thr_cv); 5496 CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */ 5497 thread_exit(); 5498} 5499 5500boolean_t 5501l2arc_vdev_present(vdev_t *vd) 5502{ 5503 l2arc_dev_t *dev; 5504 5505 mutex_enter(&l2arc_dev_mtx); 5506 for (dev = list_head(l2arc_dev_list); dev != NULL; 5507 dev = list_next(l2arc_dev_list, dev)) { 5508 if (dev->l2ad_vdev == vd) 5509 break; 5510 } 5511 mutex_exit(&l2arc_dev_mtx); 5512 5513 return (dev != NULL); 5514} 5515 5516/* 5517 * Add a vdev for use by the L2ARC. By this point the spa has already 5518 * validated the vdev and opened it. 5519 */ 5520void 5521l2arc_add_vdev(spa_t *spa, vdev_t *vd) 5522{ 5523 l2arc_dev_t *adddev; 5524 5525 ASSERT(!l2arc_vdev_present(vd)); 5526 5527 vdev_ashift_optimize(vd); 5528 5529 /* 5530 * Create a new l2arc device entry. 5531 */ 5532 adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP); 5533 adddev->l2ad_spa = spa; 5534 adddev->l2ad_vdev = vd; 5535 adddev->l2ad_start = VDEV_LABEL_START_SIZE; 5536 adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd); 5537 adddev->l2ad_hand = adddev->l2ad_start; 5538 adddev->l2ad_evict = adddev->l2ad_start; 5539 adddev->l2ad_first = B_TRUE; 5540 adddev->l2ad_writing = B_FALSE; 5541 5542 /* 5543 * This is a list of all ARC buffers that are still valid on the 5544 * device. 5545 */ 5546 adddev->l2ad_buflist = kmem_zalloc(sizeof (list_t), KM_SLEEP); 5547 list_create(adddev->l2ad_buflist, sizeof (arc_buf_hdr_t), 5548 offsetof(arc_buf_hdr_t, b_l2node)); 5549 5550 vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand); 5551 5552 /* 5553 * Add device to global list 5554 */ 5555 mutex_enter(&l2arc_dev_mtx); 5556 list_insert_head(l2arc_dev_list, adddev); 5557 atomic_inc_64(&l2arc_ndev); 5558 mutex_exit(&l2arc_dev_mtx); 5559} 5560 5561/* 5562 * Remove a vdev from the L2ARC. 5563 */ 5564void 5565l2arc_remove_vdev(vdev_t *vd) 5566{ 5567 l2arc_dev_t *dev, *nextdev, *remdev = NULL; 5568 5569 /* 5570 * Find the device by vdev 5571 */ 5572 mutex_enter(&l2arc_dev_mtx); 5573 for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) { 5574 nextdev = list_next(l2arc_dev_list, dev); 5575 if (vd == dev->l2ad_vdev) { 5576 remdev = dev; 5577 break; 5578 } 5579 } 5580 ASSERT(remdev != NULL); 5581 5582 /* 5583 * Remove device from global list 5584 */ 5585 list_remove(l2arc_dev_list, remdev); 5586 l2arc_dev_last = NULL; /* may have been invalidated */ 5587 atomic_dec_64(&l2arc_ndev); 5588 mutex_exit(&l2arc_dev_mtx); 5589 5590 /* 5591 * Clear all buflists and ARC references. L2ARC device flush. 5592 */ 5593 l2arc_evict(remdev, 0, B_TRUE); 5594 list_destroy(remdev->l2ad_buflist); 5595 kmem_free(remdev->l2ad_buflist, sizeof (list_t)); 5596 kmem_free(remdev, sizeof (l2arc_dev_t)); 5597} 5598 5599void 5600l2arc_init(void) 5601{ 5602 l2arc_thread_exit = 0; 5603 l2arc_ndev = 0; 5604 l2arc_writes_sent = 0; 5605 l2arc_writes_done = 0; 5606 5607 mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL); 5608 cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL); 5609 mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL); 5610 mutex_init(&l2arc_buflist_mtx, NULL, MUTEX_DEFAULT, NULL); 5611 mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL); 5612 5613 l2arc_dev_list = &L2ARC_dev_list; 5614 l2arc_free_on_write = &L2ARC_free_on_write; 5615 list_create(l2arc_dev_list, sizeof (l2arc_dev_t), 5616 offsetof(l2arc_dev_t, l2ad_node)); 5617 list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t), 5618 offsetof(l2arc_data_free_t, l2df_list_node)); 5619} 5620 5621void 5622l2arc_fini(void) 5623{ 5624 /* 5625 * This is called from dmu_fini(), which is called from spa_fini(); 5626 * Because of this, we can assume that all l2arc devices have 5627 * already been removed when the pools themselves were removed. 5628 */ 5629 5630 l2arc_do_free_on_write(); 5631 5632 mutex_destroy(&l2arc_feed_thr_lock); 5633 cv_destroy(&l2arc_feed_thr_cv); 5634 mutex_destroy(&l2arc_dev_mtx); 5635 mutex_destroy(&l2arc_buflist_mtx); 5636 mutex_destroy(&l2arc_free_on_write_mtx); 5637 5638 list_destroy(l2arc_dev_list); 5639 list_destroy(l2arc_free_on_write); 5640} 5641 5642void 5643l2arc_start(void) 5644{ 5645 if (!(spa_mode_global & FWRITE)) 5646 return; 5647 5648 (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0, 5649 TS_RUN, minclsyspri); 5650} 5651 5652void 5653l2arc_stop(void) 5654{ 5655 if (!(spa_mode_global & FWRITE)) 5656 return; 5657 5658 mutex_enter(&l2arc_feed_thr_lock); 5659 cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */ 5660 l2arc_thread_exit = 1; 5661 while (l2arc_thread_exit != 0) 5662 cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock); 5663 mutex_exit(&l2arc_feed_thr_lock); 5664} 5665