arc.c revision 272875
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 1510void * 1511arc_data_buf_alloc(uint64_t size) 1512{ 1513 if (arc_evict_needed(ARC_BUFC_DATA)) 1514 cv_signal(&arc_reclaim_thr_cv); 1515 atomic_add_64(&arc_size, size); 1516 return (zio_data_buf_alloc(size)); 1517} 1518 1519void 1520arc_data_buf_free(void *buf, uint64_t size) 1521{ 1522 zio_data_buf_free(buf, size); 1523 ASSERT(arc_size >= size); 1524 atomic_add_64(&arc_size, -size); 1525} 1526 1527arc_buf_t * 1528arc_buf_alloc(spa_t *spa, int size, void *tag, arc_buf_contents_t type) 1529{ 1530 arc_buf_hdr_t *hdr; 1531 arc_buf_t *buf; 1532 1533 ASSERT3U(size, >, 0); 1534 hdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE); 1535 ASSERT(BUF_EMPTY(hdr)); 1536 hdr->b_size = size; 1537 hdr->b_type = type; 1538 hdr->b_spa = spa_load_guid(spa); 1539 hdr->b_state = arc_anon; 1540 hdr->b_arc_access = 0; 1541 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 1542 buf->b_hdr = hdr; 1543 buf->b_data = NULL; 1544 buf->b_efunc = NULL; 1545 buf->b_private = NULL; 1546 buf->b_next = NULL; 1547 hdr->b_buf = buf; 1548 arc_get_data_buf(buf); 1549 hdr->b_datacnt = 1; 1550 hdr->b_flags = 0; 1551 ASSERT(refcount_is_zero(&hdr->b_refcnt)); 1552 (void) refcount_add(&hdr->b_refcnt, tag); 1553 1554 return (buf); 1555} 1556 1557static char *arc_onloan_tag = "onloan"; 1558 1559/* 1560 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in 1561 * flight data by arc_tempreserve_space() until they are "returned". Loaned 1562 * buffers must be returned to the arc before they can be used by the DMU or 1563 * freed. 1564 */ 1565arc_buf_t * 1566arc_loan_buf(spa_t *spa, int size) 1567{ 1568 arc_buf_t *buf; 1569 1570 buf = arc_buf_alloc(spa, size, arc_onloan_tag, ARC_BUFC_DATA); 1571 1572 atomic_add_64(&arc_loaned_bytes, size); 1573 return (buf); 1574} 1575 1576/* 1577 * Return a loaned arc buffer to the arc. 1578 */ 1579void 1580arc_return_buf(arc_buf_t *buf, void *tag) 1581{ 1582 arc_buf_hdr_t *hdr = buf->b_hdr; 1583 1584 ASSERT(buf->b_data != NULL); 1585 (void) refcount_add(&hdr->b_refcnt, tag); 1586 (void) refcount_remove(&hdr->b_refcnt, arc_onloan_tag); 1587 1588 atomic_add_64(&arc_loaned_bytes, -hdr->b_size); 1589} 1590 1591/* Detach an arc_buf from a dbuf (tag) */ 1592void 1593arc_loan_inuse_buf(arc_buf_t *buf, void *tag) 1594{ 1595 arc_buf_hdr_t *hdr; 1596 1597 ASSERT(buf->b_data != NULL); 1598 hdr = buf->b_hdr; 1599 (void) refcount_add(&hdr->b_refcnt, arc_onloan_tag); 1600 (void) refcount_remove(&hdr->b_refcnt, tag); 1601 buf->b_efunc = NULL; 1602 buf->b_private = NULL; 1603 1604 atomic_add_64(&arc_loaned_bytes, hdr->b_size); 1605} 1606 1607static arc_buf_t * 1608arc_buf_clone(arc_buf_t *from) 1609{ 1610 arc_buf_t *buf; 1611 arc_buf_hdr_t *hdr = from->b_hdr; 1612 uint64_t size = hdr->b_size; 1613 1614 ASSERT(hdr->b_state != arc_anon); 1615 1616 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 1617 buf->b_hdr = hdr; 1618 buf->b_data = NULL; 1619 buf->b_efunc = NULL; 1620 buf->b_private = NULL; 1621 buf->b_next = hdr->b_buf; 1622 hdr->b_buf = buf; 1623 arc_get_data_buf(buf); 1624 bcopy(from->b_data, buf->b_data, size); 1625 1626 /* 1627 * This buffer already exists in the arc so create a duplicate 1628 * copy for the caller. If the buffer is associated with user data 1629 * then track the size and number of duplicates. These stats will be 1630 * updated as duplicate buffers are created and destroyed. 1631 */ 1632 if (hdr->b_type == ARC_BUFC_DATA) { 1633 ARCSTAT_BUMP(arcstat_duplicate_buffers); 1634 ARCSTAT_INCR(arcstat_duplicate_buffers_size, size); 1635 } 1636 hdr->b_datacnt += 1; 1637 return (buf); 1638} 1639 1640void 1641arc_buf_add_ref(arc_buf_t *buf, void* tag) 1642{ 1643 arc_buf_hdr_t *hdr; 1644 kmutex_t *hash_lock; 1645 1646 /* 1647 * Check to see if this buffer is evicted. Callers 1648 * must verify b_data != NULL to know if the add_ref 1649 * was successful. 1650 */ 1651 mutex_enter(&buf->b_evict_lock); 1652 if (buf->b_data == NULL) { 1653 mutex_exit(&buf->b_evict_lock); 1654 return; 1655 } 1656 hash_lock = HDR_LOCK(buf->b_hdr); 1657 mutex_enter(hash_lock); 1658 hdr = buf->b_hdr; 1659 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 1660 mutex_exit(&buf->b_evict_lock); 1661 1662 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu); 1663 add_reference(hdr, hash_lock, tag); 1664 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr); 1665 arc_access(hdr, hash_lock); 1666 mutex_exit(hash_lock); 1667 ARCSTAT_BUMP(arcstat_hits); 1668 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH), 1669 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA, 1670 data, metadata, hits); 1671} 1672 1673/* 1674 * Free the arc data buffer. If it is an l2arc write in progress, 1675 * the buffer is placed on l2arc_free_on_write to be freed later. 1676 */ 1677static void 1678arc_buf_data_free(arc_buf_t *buf, void (*free_func)(void *, size_t)) 1679{ 1680 arc_buf_hdr_t *hdr = buf->b_hdr; 1681 1682 if (HDR_L2_WRITING(hdr)) { 1683 l2arc_data_free_t *df; 1684 df = kmem_alloc(sizeof (l2arc_data_free_t), KM_SLEEP); 1685 df->l2df_data = buf->b_data; 1686 df->l2df_size = hdr->b_size; 1687 df->l2df_func = free_func; 1688 mutex_enter(&l2arc_free_on_write_mtx); 1689 list_insert_head(l2arc_free_on_write, df); 1690 mutex_exit(&l2arc_free_on_write_mtx); 1691 ARCSTAT_BUMP(arcstat_l2_free_on_write); 1692 } else { 1693 free_func(buf->b_data, hdr->b_size); 1694 } 1695} 1696 1697/* 1698 * Free up buf->b_data and if 'remove' is set, then pull the 1699 * arc_buf_t off of the the arc_buf_hdr_t's list and free it. 1700 */ 1701static void 1702arc_buf_destroy(arc_buf_t *buf, boolean_t recycle, boolean_t remove) 1703{ 1704 arc_buf_t **bufp; 1705 1706 /* free up data associated with the buf */ 1707 if (buf->b_data) { 1708 arc_state_t *state = buf->b_hdr->b_state; 1709 uint64_t size = buf->b_hdr->b_size; 1710 arc_buf_contents_t type = buf->b_hdr->b_type; 1711 1712 arc_cksum_verify(buf); 1713#ifdef illumos 1714 arc_buf_unwatch(buf); 1715#endif /* illumos */ 1716 1717 if (!recycle) { 1718 if (type == ARC_BUFC_METADATA) { 1719 arc_buf_data_free(buf, zio_buf_free); 1720 arc_space_return(size, ARC_SPACE_DATA); 1721 } else { 1722 ASSERT(type == ARC_BUFC_DATA); 1723 arc_buf_data_free(buf, zio_data_buf_free); 1724 ARCSTAT_INCR(arcstat_data_size, -size); 1725 atomic_add_64(&arc_size, -size); 1726 } 1727 } 1728 if (list_link_active(&buf->b_hdr->b_arc_node)) { 1729 uint64_t *cnt = &state->arcs_lsize[type]; 1730 1731 ASSERT(refcount_is_zero(&buf->b_hdr->b_refcnt)); 1732 ASSERT(state != arc_anon); 1733 1734 ASSERT3U(*cnt, >=, size); 1735 atomic_add_64(cnt, -size); 1736 } 1737 ASSERT3U(state->arcs_size, >=, size); 1738 atomic_add_64(&state->arcs_size, -size); 1739 buf->b_data = NULL; 1740 1741 /* 1742 * If we're destroying a duplicate buffer make sure 1743 * that the appropriate statistics are updated. 1744 */ 1745 if (buf->b_hdr->b_datacnt > 1 && 1746 buf->b_hdr->b_type == ARC_BUFC_DATA) { 1747 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers); 1748 ARCSTAT_INCR(arcstat_duplicate_buffers_size, -size); 1749 } 1750 ASSERT(buf->b_hdr->b_datacnt > 0); 1751 buf->b_hdr->b_datacnt -= 1; 1752 } 1753 1754 /* only remove the buf if requested */ 1755 if (!remove) 1756 return; 1757 1758 /* remove the buf from the hdr list */ 1759 for (bufp = &buf->b_hdr->b_buf; *bufp != buf; bufp = &(*bufp)->b_next) 1760 continue; 1761 *bufp = buf->b_next; 1762 buf->b_next = NULL; 1763 1764 ASSERT(buf->b_efunc == NULL); 1765 1766 /* clean up the buf */ 1767 buf->b_hdr = NULL; 1768 kmem_cache_free(buf_cache, buf); 1769} 1770 1771static void 1772arc_hdr_destroy(arc_buf_hdr_t *hdr) 1773{ 1774 ASSERT(refcount_is_zero(&hdr->b_refcnt)); 1775 ASSERT3P(hdr->b_state, ==, arc_anon); 1776 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 1777 l2arc_buf_hdr_t *l2hdr = hdr->b_l2hdr; 1778 1779 if (l2hdr != NULL) { 1780 boolean_t buflist_held = MUTEX_HELD(&l2arc_buflist_mtx); 1781 /* 1782 * To prevent arc_free() and l2arc_evict() from 1783 * attempting to free the same buffer at the same time, 1784 * a FREE_IN_PROGRESS flag is given to arc_free() to 1785 * give it priority. l2arc_evict() can't destroy this 1786 * header while we are waiting on l2arc_buflist_mtx. 1787 * 1788 * The hdr may be removed from l2ad_buflist before we 1789 * grab l2arc_buflist_mtx, so b_l2hdr is rechecked. 1790 */ 1791 if (!buflist_held) { 1792 mutex_enter(&l2arc_buflist_mtx); 1793 l2hdr = hdr->b_l2hdr; 1794 } 1795 1796 if (l2hdr != NULL) { 1797 trim_map_free(l2hdr->b_dev->l2ad_vdev, l2hdr->b_daddr, 1798 hdr->b_size, 0); 1799 list_remove(l2hdr->b_dev->l2ad_buflist, hdr); 1800 ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size); 1801 ARCSTAT_INCR(arcstat_l2_asize, -l2hdr->b_asize); 1802 vdev_space_update(l2hdr->b_dev->l2ad_vdev, 1803 -l2hdr->b_asize, 0, 0); 1804 kmem_free(l2hdr, sizeof (l2arc_buf_hdr_t)); 1805 if (hdr->b_state == arc_l2c_only) 1806 l2arc_hdr_stat_remove(); 1807 hdr->b_l2hdr = NULL; 1808 } 1809 1810 if (!buflist_held) 1811 mutex_exit(&l2arc_buflist_mtx); 1812 } 1813 1814 if (!BUF_EMPTY(hdr)) { 1815 ASSERT(!HDR_IN_HASH_TABLE(hdr)); 1816 buf_discard_identity(hdr); 1817 } 1818 while (hdr->b_buf) { 1819 arc_buf_t *buf = hdr->b_buf; 1820 1821 if (buf->b_efunc) { 1822 mutex_enter(&arc_eviction_mtx); 1823 mutex_enter(&buf->b_evict_lock); 1824 ASSERT(buf->b_hdr != NULL); 1825 arc_buf_destroy(hdr->b_buf, FALSE, FALSE); 1826 hdr->b_buf = buf->b_next; 1827 buf->b_hdr = &arc_eviction_hdr; 1828 buf->b_next = arc_eviction_list; 1829 arc_eviction_list = buf; 1830 mutex_exit(&buf->b_evict_lock); 1831 mutex_exit(&arc_eviction_mtx); 1832 } else { 1833 arc_buf_destroy(hdr->b_buf, FALSE, TRUE); 1834 } 1835 } 1836 if (hdr->b_freeze_cksum != NULL) { 1837 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 1838 hdr->b_freeze_cksum = NULL; 1839 } 1840 if (hdr->b_thawed) { 1841 kmem_free(hdr->b_thawed, 1); 1842 hdr->b_thawed = NULL; 1843 } 1844 1845 ASSERT(!list_link_active(&hdr->b_arc_node)); 1846 ASSERT3P(hdr->b_hash_next, ==, NULL); 1847 ASSERT3P(hdr->b_acb, ==, NULL); 1848 kmem_cache_free(hdr_cache, hdr); 1849} 1850 1851void 1852arc_buf_free(arc_buf_t *buf, void *tag) 1853{ 1854 arc_buf_hdr_t *hdr = buf->b_hdr; 1855 int hashed = hdr->b_state != arc_anon; 1856 1857 ASSERT(buf->b_efunc == NULL); 1858 ASSERT(buf->b_data != NULL); 1859 1860 if (hashed) { 1861 kmutex_t *hash_lock = HDR_LOCK(hdr); 1862 1863 mutex_enter(hash_lock); 1864 hdr = buf->b_hdr; 1865 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 1866 1867 (void) remove_reference(hdr, hash_lock, tag); 1868 if (hdr->b_datacnt > 1) { 1869 arc_buf_destroy(buf, FALSE, TRUE); 1870 } else { 1871 ASSERT(buf == hdr->b_buf); 1872 ASSERT(buf->b_efunc == NULL); 1873 hdr->b_flags |= ARC_BUF_AVAILABLE; 1874 } 1875 mutex_exit(hash_lock); 1876 } else if (HDR_IO_IN_PROGRESS(hdr)) { 1877 int destroy_hdr; 1878 /* 1879 * We are in the middle of an async write. Don't destroy 1880 * this buffer unless the write completes before we finish 1881 * decrementing the reference count. 1882 */ 1883 mutex_enter(&arc_eviction_mtx); 1884 (void) remove_reference(hdr, NULL, tag); 1885 ASSERT(refcount_is_zero(&hdr->b_refcnt)); 1886 destroy_hdr = !HDR_IO_IN_PROGRESS(hdr); 1887 mutex_exit(&arc_eviction_mtx); 1888 if (destroy_hdr) 1889 arc_hdr_destroy(hdr); 1890 } else { 1891 if (remove_reference(hdr, NULL, tag) > 0) 1892 arc_buf_destroy(buf, FALSE, TRUE); 1893 else 1894 arc_hdr_destroy(hdr); 1895 } 1896} 1897 1898boolean_t 1899arc_buf_remove_ref(arc_buf_t *buf, void* tag) 1900{ 1901 arc_buf_hdr_t *hdr = buf->b_hdr; 1902 kmutex_t *hash_lock = HDR_LOCK(hdr); 1903 boolean_t no_callback = (buf->b_efunc == NULL); 1904 1905 if (hdr->b_state == arc_anon) { 1906 ASSERT(hdr->b_datacnt == 1); 1907 arc_buf_free(buf, tag); 1908 return (no_callback); 1909 } 1910 1911 mutex_enter(hash_lock); 1912 hdr = buf->b_hdr; 1913 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 1914 ASSERT(hdr->b_state != arc_anon); 1915 ASSERT(buf->b_data != NULL); 1916 1917 (void) remove_reference(hdr, hash_lock, tag); 1918 if (hdr->b_datacnt > 1) { 1919 if (no_callback) 1920 arc_buf_destroy(buf, FALSE, TRUE); 1921 } else if (no_callback) { 1922 ASSERT(hdr->b_buf == buf && buf->b_next == NULL); 1923 ASSERT(buf->b_efunc == NULL); 1924 hdr->b_flags |= ARC_BUF_AVAILABLE; 1925 } 1926 ASSERT(no_callback || hdr->b_datacnt > 1 || 1927 refcount_is_zero(&hdr->b_refcnt)); 1928 mutex_exit(hash_lock); 1929 return (no_callback); 1930} 1931 1932int 1933arc_buf_size(arc_buf_t *buf) 1934{ 1935 return (buf->b_hdr->b_size); 1936} 1937 1938/* 1939 * Called from the DMU to determine if the current buffer should be 1940 * evicted. In order to ensure proper locking, the eviction must be initiated 1941 * from the DMU. Return true if the buffer is associated with user data and 1942 * duplicate buffers still exist. 1943 */ 1944boolean_t 1945arc_buf_eviction_needed(arc_buf_t *buf) 1946{ 1947 arc_buf_hdr_t *hdr; 1948 boolean_t evict_needed = B_FALSE; 1949 1950 if (zfs_disable_dup_eviction) 1951 return (B_FALSE); 1952 1953 mutex_enter(&buf->b_evict_lock); 1954 hdr = buf->b_hdr; 1955 if (hdr == NULL) { 1956 /* 1957 * We are in arc_do_user_evicts(); let that function 1958 * perform the eviction. 1959 */ 1960 ASSERT(buf->b_data == NULL); 1961 mutex_exit(&buf->b_evict_lock); 1962 return (B_FALSE); 1963 } else if (buf->b_data == NULL) { 1964 /* 1965 * We have already been added to the arc eviction list; 1966 * recommend eviction. 1967 */ 1968 ASSERT3P(hdr, ==, &arc_eviction_hdr); 1969 mutex_exit(&buf->b_evict_lock); 1970 return (B_TRUE); 1971 } 1972 1973 if (hdr->b_datacnt > 1 && hdr->b_type == ARC_BUFC_DATA) 1974 evict_needed = B_TRUE; 1975 1976 mutex_exit(&buf->b_evict_lock); 1977 return (evict_needed); 1978} 1979 1980/* 1981 * Evict buffers from list until we've removed the specified number of 1982 * bytes. Move the removed buffers to the appropriate evict state. 1983 * If the recycle flag is set, then attempt to "recycle" a buffer: 1984 * - look for a buffer to evict that is `bytes' long. 1985 * - return the data block from this buffer rather than freeing it. 1986 * This flag is used by callers that are trying to make space for a 1987 * new buffer in a full arc cache. 1988 * 1989 * This function makes a "best effort". It skips over any buffers 1990 * it can't get a hash_lock on, and so may not catch all candidates. 1991 * It may also return without evicting as much space as requested. 1992 */ 1993static void * 1994arc_evict(arc_state_t *state, uint64_t spa, int64_t bytes, boolean_t recycle, 1995 arc_buf_contents_t type) 1996{ 1997 arc_state_t *evicted_state; 1998 uint64_t bytes_evicted = 0, skipped = 0, missed = 0; 1999 int64_t bytes_remaining; 2000 arc_buf_hdr_t *ab, *ab_prev = NULL; 2001 list_t *evicted_list, *list, *evicted_list_start, *list_start; 2002 kmutex_t *lock, *evicted_lock; 2003 kmutex_t *hash_lock; 2004 boolean_t have_lock; 2005 void *stolen = NULL; 2006 arc_buf_hdr_t marker = { 0 }; 2007 int count = 0; 2008 static int evict_metadata_offset, evict_data_offset; 2009 int i, idx, offset, list_count, lists; 2010 2011 ASSERT(state == arc_mru || state == arc_mfu); 2012 2013 evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost; 2014 2015 if (type == ARC_BUFC_METADATA) { 2016 offset = 0; 2017 list_count = ARC_BUFC_NUMMETADATALISTS; 2018 list_start = &state->arcs_lists[0]; 2019 evicted_list_start = &evicted_state->arcs_lists[0]; 2020 idx = evict_metadata_offset; 2021 } else { 2022 offset = ARC_BUFC_NUMMETADATALISTS; 2023 list_start = &state->arcs_lists[offset]; 2024 evicted_list_start = &evicted_state->arcs_lists[offset]; 2025 list_count = ARC_BUFC_NUMDATALISTS; 2026 idx = evict_data_offset; 2027 } 2028 bytes_remaining = evicted_state->arcs_lsize[type]; 2029 lists = 0; 2030 2031evict_start: 2032 list = &list_start[idx]; 2033 evicted_list = &evicted_list_start[idx]; 2034 lock = ARCS_LOCK(state, (offset + idx)); 2035 evicted_lock = ARCS_LOCK(evicted_state, (offset + idx)); 2036 2037 mutex_enter(lock); 2038 mutex_enter(evicted_lock); 2039 2040 for (ab = list_tail(list); ab; ab = ab_prev) { 2041 ab_prev = list_prev(list, ab); 2042 bytes_remaining -= (ab->b_size * ab->b_datacnt); 2043 /* prefetch buffers have a minimum lifespan */ 2044 if (HDR_IO_IN_PROGRESS(ab) || 2045 (spa && ab->b_spa != spa) || 2046 (ab->b_flags & (ARC_PREFETCH|ARC_INDIRECT) && 2047 ddi_get_lbolt() - ab->b_arc_access < 2048 arc_min_prefetch_lifespan)) { 2049 skipped++; 2050 continue; 2051 } 2052 /* "lookahead" for better eviction candidate */ 2053 if (recycle && ab->b_size != bytes && 2054 ab_prev && ab_prev->b_size == bytes) 2055 continue; 2056 2057 /* ignore markers */ 2058 if (ab->b_spa == 0) 2059 continue; 2060 2061 /* 2062 * It may take a long time to evict all the bufs requested. 2063 * To avoid blocking all arc activity, periodically drop 2064 * the arcs_mtx and give other threads a chance to run 2065 * before reacquiring the lock. 2066 * 2067 * If we are looking for a buffer to recycle, we are in 2068 * the hot code path, so don't sleep. 2069 */ 2070 if (!recycle && count++ > arc_evict_iterations) { 2071 list_insert_after(list, ab, &marker); 2072 mutex_exit(evicted_lock); 2073 mutex_exit(lock); 2074 kpreempt(KPREEMPT_SYNC); 2075 mutex_enter(lock); 2076 mutex_enter(evicted_lock); 2077 ab_prev = list_prev(list, &marker); 2078 list_remove(list, &marker); 2079 count = 0; 2080 continue; 2081 } 2082 2083 hash_lock = HDR_LOCK(ab); 2084 have_lock = MUTEX_HELD(hash_lock); 2085 if (have_lock || mutex_tryenter(hash_lock)) { 2086 ASSERT0(refcount_count(&ab->b_refcnt)); 2087 ASSERT(ab->b_datacnt > 0); 2088 while (ab->b_buf) { 2089 arc_buf_t *buf = ab->b_buf; 2090 if (!mutex_tryenter(&buf->b_evict_lock)) { 2091 missed += 1; 2092 break; 2093 } 2094 if (buf->b_data) { 2095 bytes_evicted += ab->b_size; 2096 if (recycle && ab->b_type == type && 2097 ab->b_size == bytes && 2098 !HDR_L2_WRITING(ab)) { 2099 stolen = buf->b_data; 2100 recycle = FALSE; 2101 } 2102 } 2103 if (buf->b_efunc) { 2104 mutex_enter(&arc_eviction_mtx); 2105 arc_buf_destroy(buf, 2106 buf->b_data == stolen, FALSE); 2107 ab->b_buf = buf->b_next; 2108 buf->b_hdr = &arc_eviction_hdr; 2109 buf->b_next = arc_eviction_list; 2110 arc_eviction_list = buf; 2111 mutex_exit(&arc_eviction_mtx); 2112 mutex_exit(&buf->b_evict_lock); 2113 } else { 2114 mutex_exit(&buf->b_evict_lock); 2115 arc_buf_destroy(buf, 2116 buf->b_data == stolen, TRUE); 2117 } 2118 } 2119 2120 if (ab->b_l2hdr) { 2121 ARCSTAT_INCR(arcstat_evict_l2_cached, 2122 ab->b_size); 2123 } else { 2124 if (l2arc_write_eligible(ab->b_spa, ab)) { 2125 ARCSTAT_INCR(arcstat_evict_l2_eligible, 2126 ab->b_size); 2127 } else { 2128 ARCSTAT_INCR( 2129 arcstat_evict_l2_ineligible, 2130 ab->b_size); 2131 } 2132 } 2133 2134 if (ab->b_datacnt == 0) { 2135 arc_change_state(evicted_state, ab, hash_lock); 2136 ASSERT(HDR_IN_HASH_TABLE(ab)); 2137 ab->b_flags |= ARC_IN_HASH_TABLE; 2138 ab->b_flags &= ~ARC_BUF_AVAILABLE; 2139 DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, ab); 2140 } 2141 if (!have_lock) 2142 mutex_exit(hash_lock); 2143 if (bytes >= 0 && bytes_evicted >= bytes) 2144 break; 2145 if (bytes_remaining > 0) { 2146 mutex_exit(evicted_lock); 2147 mutex_exit(lock); 2148 idx = ((idx + 1) & (list_count - 1)); 2149 lists++; 2150 goto evict_start; 2151 } 2152 } else { 2153 missed += 1; 2154 } 2155 } 2156 2157 mutex_exit(evicted_lock); 2158 mutex_exit(lock); 2159 2160 idx = ((idx + 1) & (list_count - 1)); 2161 lists++; 2162 2163 if (bytes_evicted < bytes) { 2164 if (lists < list_count) 2165 goto evict_start; 2166 else 2167 dprintf("only evicted %lld bytes from %x", 2168 (longlong_t)bytes_evicted, state); 2169 } 2170 if (type == ARC_BUFC_METADATA) 2171 evict_metadata_offset = idx; 2172 else 2173 evict_data_offset = idx; 2174 2175 if (skipped) 2176 ARCSTAT_INCR(arcstat_evict_skip, skipped); 2177 2178 if (missed) 2179 ARCSTAT_INCR(arcstat_mutex_miss, missed); 2180 2181 /* 2182 * Note: we have just evicted some data into the ghost state, 2183 * potentially putting the ghost size over the desired size. Rather 2184 * that evicting from the ghost list in this hot code path, leave 2185 * this chore to the arc_reclaim_thread(). 2186 */ 2187 2188 if (stolen) 2189 ARCSTAT_BUMP(arcstat_stolen); 2190 return (stolen); 2191} 2192 2193/* 2194 * Remove buffers from list until we've removed the specified number of 2195 * bytes. Destroy the buffers that are removed. 2196 */ 2197static void 2198arc_evict_ghost(arc_state_t *state, uint64_t spa, int64_t bytes) 2199{ 2200 arc_buf_hdr_t *ab, *ab_prev; 2201 arc_buf_hdr_t marker = { 0 }; 2202 list_t *list, *list_start; 2203 kmutex_t *hash_lock, *lock; 2204 uint64_t bytes_deleted = 0; 2205 uint64_t bufs_skipped = 0; 2206 int count = 0; 2207 static int evict_offset; 2208 int list_count, idx = evict_offset; 2209 int offset, lists = 0; 2210 2211 ASSERT(GHOST_STATE(state)); 2212 2213 /* 2214 * data lists come after metadata lists 2215 */ 2216 list_start = &state->arcs_lists[ARC_BUFC_NUMMETADATALISTS]; 2217 list_count = ARC_BUFC_NUMDATALISTS; 2218 offset = ARC_BUFC_NUMMETADATALISTS; 2219 2220evict_start: 2221 list = &list_start[idx]; 2222 lock = ARCS_LOCK(state, idx + offset); 2223 2224 mutex_enter(lock); 2225 for (ab = list_tail(list); ab; ab = ab_prev) { 2226 ab_prev = list_prev(list, ab); 2227 if (ab->b_type > ARC_BUFC_NUMTYPES) 2228 panic("invalid ab=%p", (void *)ab); 2229 if (spa && ab->b_spa != spa) 2230 continue; 2231 2232 /* ignore markers */ 2233 if (ab->b_spa == 0) 2234 continue; 2235 2236 hash_lock = HDR_LOCK(ab); 2237 /* caller may be trying to modify this buffer, skip it */ 2238 if (MUTEX_HELD(hash_lock)) 2239 continue; 2240 2241 /* 2242 * It may take a long time to evict all the bufs requested. 2243 * To avoid blocking all arc activity, periodically drop 2244 * the arcs_mtx and give other threads a chance to run 2245 * before reacquiring the lock. 2246 */ 2247 if (count++ > arc_evict_iterations) { 2248 list_insert_after(list, ab, &marker); 2249 mutex_exit(lock); 2250 kpreempt(KPREEMPT_SYNC); 2251 mutex_enter(lock); 2252 ab_prev = list_prev(list, &marker); 2253 list_remove(list, &marker); 2254 count = 0; 2255 continue; 2256 } 2257 if (mutex_tryenter(hash_lock)) { 2258 ASSERT(!HDR_IO_IN_PROGRESS(ab)); 2259 ASSERT(ab->b_buf == NULL); 2260 ARCSTAT_BUMP(arcstat_deleted); 2261 bytes_deleted += ab->b_size; 2262 2263 if (ab->b_l2hdr != NULL) { 2264 /* 2265 * This buffer is cached on the 2nd Level ARC; 2266 * don't destroy the header. 2267 */ 2268 arc_change_state(arc_l2c_only, ab, hash_lock); 2269 mutex_exit(hash_lock); 2270 } else { 2271 arc_change_state(arc_anon, ab, hash_lock); 2272 mutex_exit(hash_lock); 2273 arc_hdr_destroy(ab); 2274 } 2275 2276 DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, ab); 2277 if (bytes >= 0 && bytes_deleted >= bytes) 2278 break; 2279 } else if (bytes < 0) { 2280 /* 2281 * Insert a list marker and then wait for the 2282 * hash lock to become available. Once its 2283 * available, restart from where we left off. 2284 */ 2285 list_insert_after(list, ab, &marker); 2286 mutex_exit(lock); 2287 mutex_enter(hash_lock); 2288 mutex_exit(hash_lock); 2289 mutex_enter(lock); 2290 ab_prev = list_prev(list, &marker); 2291 list_remove(list, &marker); 2292 } else { 2293 bufs_skipped += 1; 2294 } 2295 2296 } 2297 mutex_exit(lock); 2298 idx = ((idx + 1) & (ARC_BUFC_NUMDATALISTS - 1)); 2299 lists++; 2300 2301 if (lists < list_count) 2302 goto evict_start; 2303 2304 evict_offset = idx; 2305 if ((uintptr_t)list > (uintptr_t)&state->arcs_lists[ARC_BUFC_NUMMETADATALISTS] && 2306 (bytes < 0 || bytes_deleted < bytes)) { 2307 list_start = &state->arcs_lists[0]; 2308 list_count = ARC_BUFC_NUMMETADATALISTS; 2309 offset = lists = 0; 2310 goto evict_start; 2311 } 2312 2313 if (bufs_skipped) { 2314 ARCSTAT_INCR(arcstat_mutex_miss, bufs_skipped); 2315 ASSERT(bytes >= 0); 2316 } 2317 2318 if (bytes_deleted < bytes) 2319 dprintf("only deleted %lld bytes from %p", 2320 (longlong_t)bytes_deleted, state); 2321} 2322 2323static void 2324arc_adjust(void) 2325{ 2326 int64_t adjustment, delta; 2327 2328 /* 2329 * Adjust MRU size 2330 */ 2331 2332 adjustment = MIN((int64_t)(arc_size - arc_c), 2333 (int64_t)(arc_anon->arcs_size + arc_mru->arcs_size + arc_meta_used - 2334 arc_p)); 2335 2336 if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_DATA] > 0) { 2337 delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_DATA], adjustment); 2338 (void) arc_evict(arc_mru, 0, delta, FALSE, ARC_BUFC_DATA); 2339 adjustment -= delta; 2340 } 2341 2342 if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_METADATA] > 0) { 2343 delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_METADATA], adjustment); 2344 (void) arc_evict(arc_mru, 0, delta, FALSE, 2345 ARC_BUFC_METADATA); 2346 } 2347 2348 /* 2349 * Adjust MFU size 2350 */ 2351 2352 adjustment = arc_size - arc_c; 2353 2354 if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_DATA] > 0) { 2355 delta = MIN(adjustment, arc_mfu->arcs_lsize[ARC_BUFC_DATA]); 2356 (void) arc_evict(arc_mfu, 0, delta, FALSE, ARC_BUFC_DATA); 2357 adjustment -= delta; 2358 } 2359 2360 if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_METADATA] > 0) { 2361 int64_t delta = MIN(adjustment, 2362 arc_mfu->arcs_lsize[ARC_BUFC_METADATA]); 2363 (void) arc_evict(arc_mfu, 0, delta, FALSE, 2364 ARC_BUFC_METADATA); 2365 } 2366 2367 /* 2368 * Adjust ghost lists 2369 */ 2370 2371 adjustment = arc_mru->arcs_size + arc_mru_ghost->arcs_size - arc_c; 2372 2373 if (adjustment > 0 && arc_mru_ghost->arcs_size > 0) { 2374 delta = MIN(arc_mru_ghost->arcs_size, adjustment); 2375 arc_evict_ghost(arc_mru_ghost, 0, delta); 2376 } 2377 2378 adjustment = 2379 arc_mru_ghost->arcs_size + arc_mfu_ghost->arcs_size - arc_c; 2380 2381 if (adjustment > 0 && arc_mfu_ghost->arcs_size > 0) { 2382 delta = MIN(arc_mfu_ghost->arcs_size, adjustment); 2383 arc_evict_ghost(arc_mfu_ghost, 0, delta); 2384 } 2385} 2386 2387static void 2388arc_do_user_evicts(void) 2389{ 2390 static arc_buf_t *tmp_arc_eviction_list; 2391 2392 /* 2393 * Move list over to avoid LOR 2394 */ 2395restart: 2396 mutex_enter(&arc_eviction_mtx); 2397 tmp_arc_eviction_list = arc_eviction_list; 2398 arc_eviction_list = NULL; 2399 mutex_exit(&arc_eviction_mtx); 2400 2401 while (tmp_arc_eviction_list != NULL) { 2402 arc_buf_t *buf = tmp_arc_eviction_list; 2403 tmp_arc_eviction_list = buf->b_next; 2404 mutex_enter(&buf->b_evict_lock); 2405 buf->b_hdr = NULL; 2406 mutex_exit(&buf->b_evict_lock); 2407 2408 if (buf->b_efunc != NULL) 2409 VERIFY0(buf->b_efunc(buf->b_private)); 2410 2411 buf->b_efunc = NULL; 2412 buf->b_private = NULL; 2413 kmem_cache_free(buf_cache, buf); 2414 } 2415 2416 if (arc_eviction_list != NULL) 2417 goto restart; 2418} 2419 2420/* 2421 * Flush all *evictable* data from the cache for the given spa. 2422 * NOTE: this will not touch "active" (i.e. referenced) data. 2423 */ 2424void 2425arc_flush(spa_t *spa) 2426{ 2427 uint64_t guid = 0; 2428 2429 if (spa) 2430 guid = spa_load_guid(spa); 2431 2432 while (arc_mru->arcs_lsize[ARC_BUFC_DATA]) { 2433 (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_DATA); 2434 if (spa) 2435 break; 2436 } 2437 while (arc_mru->arcs_lsize[ARC_BUFC_METADATA]) { 2438 (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_METADATA); 2439 if (spa) 2440 break; 2441 } 2442 while (arc_mfu->arcs_lsize[ARC_BUFC_DATA]) { 2443 (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_DATA); 2444 if (spa) 2445 break; 2446 } 2447 while (arc_mfu->arcs_lsize[ARC_BUFC_METADATA]) { 2448 (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_METADATA); 2449 if (spa) 2450 break; 2451 } 2452 2453 arc_evict_ghost(arc_mru_ghost, guid, -1); 2454 arc_evict_ghost(arc_mfu_ghost, guid, -1); 2455 2456 mutex_enter(&arc_reclaim_thr_lock); 2457 arc_do_user_evicts(); 2458 mutex_exit(&arc_reclaim_thr_lock); 2459 ASSERT(spa || arc_eviction_list == NULL); 2460} 2461 2462void 2463arc_shrink(void) 2464{ 2465 2466 if (arc_c > arc_c_min) { 2467 uint64_t to_free; 2468 2469 DTRACE_PROBE4(arc__shrink, uint64_t, arc_c, uint64_t, 2470 arc_c_min, uint64_t, arc_p, uint64_t, to_free); 2471#ifdef _KERNEL 2472 to_free = arc_c >> arc_shrink_shift; 2473#else 2474 to_free = arc_c >> arc_shrink_shift; 2475#endif 2476 if (arc_c > arc_c_min + to_free) 2477 atomic_add_64(&arc_c, -to_free); 2478 else 2479 arc_c = arc_c_min; 2480 2481 atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift)); 2482 if (arc_c > arc_size) 2483 arc_c = MAX(arc_size, arc_c_min); 2484 if (arc_p > arc_c) 2485 arc_p = (arc_c >> 1); 2486 2487 DTRACE_PROBE2(arc__shrunk, uint64_t, arc_c, uint64_t, 2488 arc_p); 2489 2490 ASSERT(arc_c >= arc_c_min); 2491 ASSERT((int64_t)arc_p >= 0); 2492 } 2493 2494 if (arc_size > arc_c) { 2495 DTRACE_PROBE2(arc__shrink_adjust, uint64_t, arc_size, 2496 uint64_t, arc_c); 2497 arc_adjust(); 2498 } 2499} 2500 2501static int needfree = 0; 2502 2503static int 2504arc_reclaim_needed(void) 2505{ 2506 2507#ifdef _KERNEL 2508 2509 if (needfree) { 2510 DTRACE_PROBE(arc__reclaim_needfree); 2511 return (1); 2512 } 2513 2514 /* 2515 * Cooperate with pagedaemon when it's time for it to scan 2516 * and reclaim some pages. 2517 */ 2518 if (freemem < zfs_arc_free_target) { 2519 DTRACE_PROBE2(arc__reclaim_freemem, uint64_t, 2520 freemem, uint64_t, zfs_arc_free_target); 2521 return (1); 2522 } 2523 2524#ifdef sun 2525 /* 2526 * take 'desfree' extra pages, so we reclaim sooner, rather than later 2527 */ 2528 extra = desfree; 2529 2530 /* 2531 * check that we're out of range of the pageout scanner. It starts to 2532 * schedule paging if freemem is less than lotsfree and needfree. 2533 * lotsfree is the high-water mark for pageout, and needfree is the 2534 * number of needed free pages. We add extra pages here to make sure 2535 * the scanner doesn't start up while we're freeing memory. 2536 */ 2537 if (freemem < lotsfree + needfree + extra) 2538 return (1); 2539 2540 /* 2541 * check to make sure that swapfs has enough space so that anon 2542 * reservations can still succeed. anon_resvmem() checks that the 2543 * availrmem is greater than swapfs_minfree, and the number of reserved 2544 * swap pages. We also add a bit of extra here just to prevent 2545 * circumstances from getting really dire. 2546 */ 2547 if (availrmem < swapfs_minfree + swapfs_reserve + extra) 2548 return (1); 2549 2550 /* 2551 * Check that we have enough availrmem that memory locking (e.g., via 2552 * mlock(3C) or memcntl(2)) can still succeed. (pages_pp_maximum 2553 * stores the number of pages that cannot be locked; when availrmem 2554 * drops below pages_pp_maximum, page locking mechanisms such as 2555 * page_pp_lock() will fail.) 2556 */ 2557 if (availrmem <= pages_pp_maximum) 2558 return (1); 2559 2560#endif /* sun */ 2561#if defined(__i386) || !defined(UMA_MD_SMALL_ALLOC) 2562 /* 2563 * If we're on an i386 platform, it's possible that we'll exhaust the 2564 * kernel heap space before we ever run out of available physical 2565 * memory. Most checks of the size of the heap_area compare against 2566 * tune.t_minarmem, which is the minimum available real memory that we 2567 * can have in the system. However, this is generally fixed at 25 pages 2568 * which is so low that it's useless. In this comparison, we seek to 2569 * calculate the total heap-size, and reclaim if more than 3/4ths of the 2570 * heap is allocated. (Or, in the calculation, if less than 1/4th is 2571 * free) 2572 */ 2573 if (vmem_size(heap_arena, VMEM_FREE) < 2574 (vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC) >> 2)) { 2575 DTRACE_PROBE2(arc__reclaim_used, uint64_t, 2576 vmem_size(heap_arena, VMEM_FREE), uint64_t, 2577 (vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC)) >> 2); 2578 return (1); 2579 } 2580#endif 2581#ifdef sun 2582 /* 2583 * If zio data pages are being allocated out of a separate heap segment, 2584 * then enforce that the size of available vmem for this arena remains 2585 * above about 1/16th free. 2586 * 2587 * Note: The 1/16th arena free requirement was put in place 2588 * to aggressively evict memory from the arc in order to avoid 2589 * memory fragmentation issues. 2590 */ 2591 if (zio_arena != NULL && 2592 vmem_size(zio_arena, VMEM_FREE) < 2593 (vmem_size(zio_arena, VMEM_ALLOC) >> 4)) 2594 return (1); 2595#endif /* sun */ 2596#else /* _KERNEL */ 2597 if (spa_get_random(100) == 0) 2598 return (1); 2599#endif /* _KERNEL */ 2600 DTRACE_PROBE(arc__reclaim_no); 2601 2602 return (0); 2603} 2604 2605extern kmem_cache_t *zio_buf_cache[]; 2606extern kmem_cache_t *zio_data_buf_cache[]; 2607 2608static void __noinline 2609arc_kmem_reap_now(arc_reclaim_strategy_t strat) 2610{ 2611 size_t i; 2612 kmem_cache_t *prev_cache = NULL; 2613 kmem_cache_t *prev_data_cache = NULL; 2614 2615 DTRACE_PROBE(arc__kmem_reap_start); 2616#ifdef _KERNEL 2617 if (arc_meta_used >= arc_meta_limit) { 2618 /* 2619 * We are exceeding our meta-data cache limit. 2620 * Purge some DNLC entries to release holds on meta-data. 2621 */ 2622 dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent); 2623 } 2624#if defined(__i386) 2625 /* 2626 * Reclaim unused memory from all kmem caches. 2627 */ 2628 kmem_reap(); 2629#endif 2630#endif 2631 2632 /* 2633 * An aggressive reclamation will shrink the cache size as well as 2634 * reap free buffers from the arc kmem caches. 2635 */ 2636 if (strat == ARC_RECLAIM_AGGR) 2637 arc_shrink(); 2638 2639 for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) { 2640 if (zio_buf_cache[i] != prev_cache) { 2641 prev_cache = zio_buf_cache[i]; 2642 kmem_cache_reap_now(zio_buf_cache[i]); 2643 } 2644 if (zio_data_buf_cache[i] != prev_data_cache) { 2645 prev_data_cache = zio_data_buf_cache[i]; 2646 kmem_cache_reap_now(zio_data_buf_cache[i]); 2647 } 2648 } 2649 kmem_cache_reap_now(buf_cache); 2650 kmem_cache_reap_now(hdr_cache); 2651 2652#ifdef sun 2653 /* 2654 * Ask the vmem arena to reclaim unused memory from its 2655 * quantum caches. 2656 */ 2657 if (zio_arena != NULL && strat == ARC_RECLAIM_AGGR) 2658 vmem_qcache_reap(zio_arena); 2659#endif 2660 DTRACE_PROBE(arc__kmem_reap_end); 2661} 2662 2663static void 2664arc_reclaim_thread(void *dummy __unused) 2665{ 2666 clock_t growtime = 0; 2667 arc_reclaim_strategy_t last_reclaim = ARC_RECLAIM_CONS; 2668 callb_cpr_t cpr; 2669 2670 CALLB_CPR_INIT(&cpr, &arc_reclaim_thr_lock, callb_generic_cpr, FTAG); 2671 2672 mutex_enter(&arc_reclaim_thr_lock); 2673 while (arc_thread_exit == 0) { 2674 if (arc_reclaim_needed()) { 2675 2676 if (arc_no_grow) { 2677 if (last_reclaim == ARC_RECLAIM_CONS) { 2678 DTRACE_PROBE(arc__reclaim_aggr_no_grow); 2679 last_reclaim = ARC_RECLAIM_AGGR; 2680 } else { 2681 last_reclaim = ARC_RECLAIM_CONS; 2682 } 2683 } else { 2684 arc_no_grow = TRUE; 2685 last_reclaim = ARC_RECLAIM_AGGR; 2686 DTRACE_PROBE(arc__reclaim_aggr); 2687 membar_producer(); 2688 } 2689 2690 /* reset the growth delay for every reclaim */ 2691 growtime = ddi_get_lbolt() + (arc_grow_retry * hz); 2692 2693 if (needfree && last_reclaim == ARC_RECLAIM_CONS) { 2694 /* 2695 * If needfree is TRUE our vm_lowmem hook 2696 * was called and in that case we must free some 2697 * memory, so switch to aggressive mode. 2698 */ 2699 arc_no_grow = TRUE; 2700 last_reclaim = ARC_RECLAIM_AGGR; 2701 } 2702 arc_kmem_reap_now(last_reclaim); 2703 arc_warm = B_TRUE; 2704 2705 } else if (arc_no_grow && ddi_get_lbolt() >= growtime) { 2706 arc_no_grow = FALSE; 2707 } 2708 2709 arc_adjust(); 2710 2711 if (arc_eviction_list != NULL) 2712 arc_do_user_evicts(); 2713 2714#ifdef _KERNEL 2715 if (needfree) { 2716 needfree = 0; 2717 wakeup(&needfree); 2718 } 2719#endif 2720 2721 /* block until needed, or one second, whichever is shorter */ 2722 CALLB_CPR_SAFE_BEGIN(&cpr); 2723 (void) cv_timedwait(&arc_reclaim_thr_cv, 2724 &arc_reclaim_thr_lock, hz); 2725 CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_thr_lock); 2726 } 2727 2728 arc_thread_exit = 0; 2729 cv_broadcast(&arc_reclaim_thr_cv); 2730 CALLB_CPR_EXIT(&cpr); /* drops arc_reclaim_thr_lock */ 2731 thread_exit(); 2732} 2733 2734/* 2735 * Adapt arc info given the number of bytes we are trying to add and 2736 * the state that we are comming from. This function is only called 2737 * when we are adding new content to the cache. 2738 */ 2739static void 2740arc_adapt(int bytes, arc_state_t *state) 2741{ 2742 int mult; 2743 uint64_t arc_p_min = (arc_c >> arc_p_min_shift); 2744 2745 if (state == arc_l2c_only) 2746 return; 2747 2748 ASSERT(bytes > 0); 2749 /* 2750 * Adapt the target size of the MRU list: 2751 * - if we just hit in the MRU ghost list, then increase 2752 * the target size of the MRU list. 2753 * - if we just hit in the MFU ghost list, then increase 2754 * the target size of the MFU list by decreasing the 2755 * target size of the MRU list. 2756 */ 2757 if (state == arc_mru_ghost) { 2758 mult = ((arc_mru_ghost->arcs_size >= arc_mfu_ghost->arcs_size) ? 2759 1 : (arc_mfu_ghost->arcs_size/arc_mru_ghost->arcs_size)); 2760 mult = MIN(mult, 10); /* avoid wild arc_p adjustment */ 2761 2762 arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult); 2763 } else if (state == arc_mfu_ghost) { 2764 uint64_t delta; 2765 2766 mult = ((arc_mfu_ghost->arcs_size >= arc_mru_ghost->arcs_size) ? 2767 1 : (arc_mru_ghost->arcs_size/arc_mfu_ghost->arcs_size)); 2768 mult = MIN(mult, 10); 2769 2770 delta = MIN(bytes * mult, arc_p); 2771 arc_p = MAX(arc_p_min, arc_p - delta); 2772 } 2773 ASSERT((int64_t)arc_p >= 0); 2774 2775 if (arc_reclaim_needed()) { 2776 cv_signal(&arc_reclaim_thr_cv); 2777 return; 2778 } 2779 2780 if (arc_no_grow) 2781 return; 2782 2783 if (arc_c >= arc_c_max) 2784 return; 2785 2786 /* 2787 * If we're within (2 * maxblocksize) bytes of the target 2788 * cache size, increment the target cache size 2789 */ 2790 if (arc_size > arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) { 2791 DTRACE_PROBE1(arc__inc_adapt, int, bytes); 2792 atomic_add_64(&arc_c, (int64_t)bytes); 2793 if (arc_c > arc_c_max) 2794 arc_c = arc_c_max; 2795 else if (state == arc_anon) 2796 atomic_add_64(&arc_p, (int64_t)bytes); 2797 if (arc_p > arc_c) 2798 arc_p = arc_c; 2799 } 2800 ASSERT((int64_t)arc_p >= 0); 2801} 2802 2803/* 2804 * Check if the cache has reached its limits and eviction is required 2805 * prior to insert. 2806 */ 2807static int 2808arc_evict_needed(arc_buf_contents_t type) 2809{ 2810 if (type == ARC_BUFC_METADATA && arc_meta_used >= arc_meta_limit) 2811 return (1); 2812 2813 if (arc_reclaim_needed()) 2814 return (1); 2815 2816 return (arc_size > arc_c); 2817} 2818 2819/* 2820 * The buffer, supplied as the first argument, needs a data block. 2821 * So, if we are at cache max, determine which cache should be victimized. 2822 * We have the following cases: 2823 * 2824 * 1. Insert for MRU, p > sizeof(arc_anon + arc_mru) -> 2825 * In this situation if we're out of space, but the resident size of the MFU is 2826 * under the limit, victimize the MFU cache to satisfy this insertion request. 2827 * 2828 * 2. Insert for MRU, p <= sizeof(arc_anon + arc_mru) -> 2829 * Here, we've used up all of the available space for the MRU, so we need to 2830 * evict from our own cache instead. Evict from the set of resident MRU 2831 * entries. 2832 * 2833 * 3. Insert for MFU (c - p) > sizeof(arc_mfu) -> 2834 * c minus p represents the MFU space in the cache, since p is the size of the 2835 * cache that is dedicated to the MRU. In this situation there's still space on 2836 * the MFU side, so the MRU side needs to be victimized. 2837 * 2838 * 4. Insert for MFU (c - p) < sizeof(arc_mfu) -> 2839 * MFU's resident set is consuming more space than it has been allotted. In 2840 * this situation, we must victimize our own cache, the MFU, for this insertion. 2841 */ 2842static void 2843arc_get_data_buf(arc_buf_t *buf) 2844{ 2845 arc_state_t *state = buf->b_hdr->b_state; 2846 uint64_t size = buf->b_hdr->b_size; 2847 arc_buf_contents_t type = buf->b_hdr->b_type; 2848 2849 arc_adapt(size, state); 2850 2851 /* 2852 * We have not yet reached cache maximum size, 2853 * just allocate a new buffer. 2854 */ 2855 if (!arc_evict_needed(type)) { 2856 if (type == ARC_BUFC_METADATA) { 2857 buf->b_data = zio_buf_alloc(size); 2858 arc_space_consume(size, ARC_SPACE_DATA); 2859 } else { 2860 ASSERT(type == ARC_BUFC_DATA); 2861 buf->b_data = zio_data_buf_alloc(size); 2862 ARCSTAT_INCR(arcstat_data_size, size); 2863 atomic_add_64(&arc_size, size); 2864 } 2865 goto out; 2866 } 2867 2868 /* 2869 * If we are prefetching from the mfu ghost list, this buffer 2870 * will end up on the mru list; so steal space from there. 2871 */ 2872 if (state == arc_mfu_ghost) 2873 state = buf->b_hdr->b_flags & ARC_PREFETCH ? arc_mru : arc_mfu; 2874 else if (state == arc_mru_ghost) 2875 state = arc_mru; 2876 2877 if (state == arc_mru || state == arc_anon) { 2878 uint64_t mru_used = arc_anon->arcs_size + arc_mru->arcs_size; 2879 state = (arc_mfu->arcs_lsize[type] >= size && 2880 arc_p > mru_used) ? arc_mfu : arc_mru; 2881 } else { 2882 /* MFU cases */ 2883 uint64_t mfu_space = arc_c - arc_p; 2884 state = (arc_mru->arcs_lsize[type] >= size && 2885 mfu_space > arc_mfu->arcs_size) ? arc_mru : arc_mfu; 2886 } 2887 if ((buf->b_data = arc_evict(state, 0, size, TRUE, type)) == NULL) { 2888 if (type == ARC_BUFC_METADATA) { 2889 buf->b_data = zio_buf_alloc(size); 2890 arc_space_consume(size, ARC_SPACE_DATA); 2891 } else { 2892 ASSERT(type == ARC_BUFC_DATA); 2893 buf->b_data = zio_data_buf_alloc(size); 2894 ARCSTAT_INCR(arcstat_data_size, size); 2895 atomic_add_64(&arc_size, size); 2896 } 2897 ARCSTAT_BUMP(arcstat_recycle_miss); 2898 } 2899 ASSERT(buf->b_data != NULL); 2900out: 2901 /* 2902 * Update the state size. Note that ghost states have a 2903 * "ghost size" and so don't need to be updated. 2904 */ 2905 if (!GHOST_STATE(buf->b_hdr->b_state)) { 2906 arc_buf_hdr_t *hdr = buf->b_hdr; 2907 2908 atomic_add_64(&hdr->b_state->arcs_size, size); 2909 if (list_link_active(&hdr->b_arc_node)) { 2910 ASSERT(refcount_is_zero(&hdr->b_refcnt)); 2911 atomic_add_64(&hdr->b_state->arcs_lsize[type], size); 2912 } 2913 /* 2914 * If we are growing the cache, and we are adding anonymous 2915 * data, and we have outgrown arc_p, update arc_p 2916 */ 2917 if (arc_size < arc_c && hdr->b_state == arc_anon && 2918 arc_anon->arcs_size + arc_mru->arcs_size > arc_p) 2919 arc_p = MIN(arc_c, arc_p + size); 2920 } 2921 ARCSTAT_BUMP(arcstat_allocated); 2922} 2923 2924/* 2925 * This routine is called whenever a buffer is accessed. 2926 * NOTE: the hash lock is dropped in this function. 2927 */ 2928static void 2929arc_access(arc_buf_hdr_t *buf, kmutex_t *hash_lock) 2930{ 2931 clock_t now; 2932 2933 ASSERT(MUTEX_HELD(hash_lock)); 2934 2935 if (buf->b_state == arc_anon) { 2936 /* 2937 * This buffer is not in the cache, and does not 2938 * appear in our "ghost" list. Add the new buffer 2939 * to the MRU state. 2940 */ 2941 2942 ASSERT(buf->b_arc_access == 0); 2943 buf->b_arc_access = ddi_get_lbolt(); 2944 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, buf); 2945 arc_change_state(arc_mru, buf, hash_lock); 2946 2947 } else if (buf->b_state == arc_mru) { 2948 now = ddi_get_lbolt(); 2949 2950 /* 2951 * If this buffer is here because of a prefetch, then either: 2952 * - clear the flag if this is a "referencing" read 2953 * (any subsequent access will bump this into the MFU state). 2954 * or 2955 * - move the buffer to the head of the list if this is 2956 * another prefetch (to make it less likely to be evicted). 2957 */ 2958 if ((buf->b_flags & ARC_PREFETCH) != 0) { 2959 if (refcount_count(&buf->b_refcnt) == 0) { 2960 ASSERT(list_link_active(&buf->b_arc_node)); 2961 } else { 2962 buf->b_flags &= ~ARC_PREFETCH; 2963 ARCSTAT_BUMP(arcstat_mru_hits); 2964 } 2965 buf->b_arc_access = now; 2966 return; 2967 } 2968 2969 /* 2970 * This buffer has been "accessed" only once so far, 2971 * but it is still in the cache. Move it to the MFU 2972 * state. 2973 */ 2974 if (now > buf->b_arc_access + ARC_MINTIME) { 2975 /* 2976 * More than 125ms have passed since we 2977 * instantiated this buffer. Move it to the 2978 * most frequently used state. 2979 */ 2980 buf->b_arc_access = now; 2981 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf); 2982 arc_change_state(arc_mfu, buf, hash_lock); 2983 } 2984 ARCSTAT_BUMP(arcstat_mru_hits); 2985 } else if (buf->b_state == arc_mru_ghost) { 2986 arc_state_t *new_state; 2987 /* 2988 * This buffer has been "accessed" recently, but 2989 * was evicted from the cache. Move it to the 2990 * MFU state. 2991 */ 2992 2993 if (buf->b_flags & ARC_PREFETCH) { 2994 new_state = arc_mru; 2995 if (refcount_count(&buf->b_refcnt) > 0) 2996 buf->b_flags &= ~ARC_PREFETCH; 2997 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, buf); 2998 } else { 2999 new_state = arc_mfu; 3000 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf); 3001 } 3002 3003 buf->b_arc_access = ddi_get_lbolt(); 3004 arc_change_state(new_state, buf, hash_lock); 3005 3006 ARCSTAT_BUMP(arcstat_mru_ghost_hits); 3007 } else if (buf->b_state == arc_mfu) { 3008 /* 3009 * This buffer has been accessed more than once and is 3010 * still in the cache. Keep it in the MFU state. 3011 * 3012 * NOTE: an add_reference() that occurred when we did 3013 * the arc_read() will have kicked this off the list. 3014 * If it was a prefetch, we will explicitly move it to 3015 * the head of the list now. 3016 */ 3017 if ((buf->b_flags & ARC_PREFETCH) != 0) { 3018 ASSERT(refcount_count(&buf->b_refcnt) == 0); 3019 ASSERT(list_link_active(&buf->b_arc_node)); 3020 } 3021 ARCSTAT_BUMP(arcstat_mfu_hits); 3022 buf->b_arc_access = ddi_get_lbolt(); 3023 } else if (buf->b_state == arc_mfu_ghost) { 3024 arc_state_t *new_state = arc_mfu; 3025 /* 3026 * This buffer has been accessed more than once but has 3027 * been evicted from the cache. Move it back to the 3028 * MFU state. 3029 */ 3030 3031 if (buf->b_flags & ARC_PREFETCH) { 3032 /* 3033 * This is a prefetch access... 3034 * move this block back to the MRU state. 3035 */ 3036 ASSERT0(refcount_count(&buf->b_refcnt)); 3037 new_state = arc_mru; 3038 } 3039 3040 buf->b_arc_access = ddi_get_lbolt(); 3041 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf); 3042 arc_change_state(new_state, buf, hash_lock); 3043 3044 ARCSTAT_BUMP(arcstat_mfu_ghost_hits); 3045 } else if (buf->b_state == arc_l2c_only) { 3046 /* 3047 * This buffer is on the 2nd Level ARC. 3048 */ 3049 3050 buf->b_arc_access = ddi_get_lbolt(); 3051 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf); 3052 arc_change_state(arc_mfu, buf, hash_lock); 3053 } else { 3054 ASSERT(!"invalid arc state"); 3055 } 3056} 3057 3058/* a generic arc_done_func_t which you can use */ 3059/* ARGSUSED */ 3060void 3061arc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg) 3062{ 3063 if (zio == NULL || zio->io_error == 0) 3064 bcopy(buf->b_data, arg, buf->b_hdr->b_size); 3065 VERIFY(arc_buf_remove_ref(buf, arg)); 3066} 3067 3068/* a generic arc_done_func_t */ 3069void 3070arc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg) 3071{ 3072 arc_buf_t **bufp = arg; 3073 if (zio && zio->io_error) { 3074 VERIFY(arc_buf_remove_ref(buf, arg)); 3075 *bufp = NULL; 3076 } else { 3077 *bufp = buf; 3078 ASSERT(buf->b_data); 3079 } 3080} 3081 3082static void 3083arc_read_done(zio_t *zio) 3084{ 3085 arc_buf_hdr_t *hdr; 3086 arc_buf_t *buf; 3087 arc_buf_t *abuf; /* buffer we're assigning to callback */ 3088 kmutex_t *hash_lock = NULL; 3089 arc_callback_t *callback_list, *acb; 3090 int freeable = FALSE; 3091 3092 buf = zio->io_private; 3093 hdr = buf->b_hdr; 3094 3095 /* 3096 * The hdr was inserted into hash-table and removed from lists 3097 * prior to starting I/O. We should find this header, since 3098 * it's in the hash table, and it should be legit since it's 3099 * not possible to evict it during the I/O. The only possible 3100 * reason for it not to be found is if we were freed during the 3101 * read. 3102 */ 3103 if (HDR_IN_HASH_TABLE(hdr)) { 3104 ASSERT3U(hdr->b_birth, ==, BP_PHYSICAL_BIRTH(zio->io_bp)); 3105 ASSERT3U(hdr->b_dva.dva_word[0], ==, 3106 BP_IDENTITY(zio->io_bp)->dva_word[0]); 3107 ASSERT3U(hdr->b_dva.dva_word[1], ==, 3108 BP_IDENTITY(zio->io_bp)->dva_word[1]); 3109 3110 arc_buf_hdr_t *found = buf_hash_find(hdr->b_spa, zio->io_bp, 3111 &hash_lock); 3112 3113 ASSERT((found == NULL && HDR_FREED_IN_READ(hdr) && 3114 hash_lock == NULL) || 3115 (found == hdr && 3116 DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) || 3117 (found == hdr && HDR_L2_READING(hdr))); 3118 } 3119 3120 hdr->b_flags &= ~ARC_L2_EVICTED; 3121 if (l2arc_noprefetch && (hdr->b_flags & ARC_PREFETCH)) 3122 hdr->b_flags &= ~ARC_L2CACHE; 3123 3124 /* byteswap if necessary */ 3125 callback_list = hdr->b_acb; 3126 ASSERT(callback_list != NULL); 3127 if (BP_SHOULD_BYTESWAP(zio->io_bp) && zio->io_error == 0) { 3128 dmu_object_byteswap_t bswap = 3129 DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp)); 3130 arc_byteswap_func_t *func = BP_GET_LEVEL(zio->io_bp) > 0 ? 3131 byteswap_uint64_array : 3132 dmu_ot_byteswap[bswap].ob_func; 3133 func(buf->b_data, hdr->b_size); 3134 } 3135 3136 arc_cksum_compute(buf, B_FALSE); 3137#ifdef illumos 3138 arc_buf_watch(buf); 3139#endif /* illumos */ 3140 3141 if (hash_lock && zio->io_error == 0 && hdr->b_state == arc_anon) { 3142 /* 3143 * Only call arc_access on anonymous buffers. This is because 3144 * if we've issued an I/O for an evicted buffer, we've already 3145 * called arc_access (to prevent any simultaneous readers from 3146 * getting confused). 3147 */ 3148 arc_access(hdr, hash_lock); 3149 } 3150 3151 /* create copies of the data buffer for the callers */ 3152 abuf = buf; 3153 for (acb = callback_list; acb; acb = acb->acb_next) { 3154 if (acb->acb_done) { 3155 if (abuf == NULL) { 3156 ARCSTAT_BUMP(arcstat_duplicate_reads); 3157 abuf = arc_buf_clone(buf); 3158 } 3159 acb->acb_buf = abuf; 3160 abuf = NULL; 3161 } 3162 } 3163 hdr->b_acb = NULL; 3164 hdr->b_flags &= ~ARC_IO_IN_PROGRESS; 3165 ASSERT(!HDR_BUF_AVAILABLE(hdr)); 3166 if (abuf == buf) { 3167 ASSERT(buf->b_efunc == NULL); 3168 ASSERT(hdr->b_datacnt == 1); 3169 hdr->b_flags |= ARC_BUF_AVAILABLE; 3170 } 3171 3172 ASSERT(refcount_is_zero(&hdr->b_refcnt) || callback_list != NULL); 3173 3174 if (zio->io_error != 0) { 3175 hdr->b_flags |= ARC_IO_ERROR; 3176 if (hdr->b_state != arc_anon) 3177 arc_change_state(arc_anon, hdr, hash_lock); 3178 if (HDR_IN_HASH_TABLE(hdr)) 3179 buf_hash_remove(hdr); 3180 freeable = refcount_is_zero(&hdr->b_refcnt); 3181 } 3182 3183 /* 3184 * Broadcast before we drop the hash_lock to avoid the possibility 3185 * that the hdr (and hence the cv) might be freed before we get to 3186 * the cv_broadcast(). 3187 */ 3188 cv_broadcast(&hdr->b_cv); 3189 3190 if (hash_lock) { 3191 mutex_exit(hash_lock); 3192 } else { 3193 /* 3194 * This block was freed while we waited for the read to 3195 * complete. It has been removed from the hash table and 3196 * moved to the anonymous state (so that it won't show up 3197 * in the cache). 3198 */ 3199 ASSERT3P(hdr->b_state, ==, arc_anon); 3200 freeable = refcount_is_zero(&hdr->b_refcnt); 3201 } 3202 3203 /* execute each callback and free its structure */ 3204 while ((acb = callback_list) != NULL) { 3205 if (acb->acb_done) 3206 acb->acb_done(zio, acb->acb_buf, acb->acb_private); 3207 3208 if (acb->acb_zio_dummy != NULL) { 3209 acb->acb_zio_dummy->io_error = zio->io_error; 3210 zio_nowait(acb->acb_zio_dummy); 3211 } 3212 3213 callback_list = acb->acb_next; 3214 kmem_free(acb, sizeof (arc_callback_t)); 3215 } 3216 3217 if (freeable) 3218 arc_hdr_destroy(hdr); 3219} 3220 3221/* 3222 * "Read" the block block at the specified DVA (in bp) via the 3223 * cache. If the block is found in the cache, invoke the provided 3224 * callback immediately and return. Note that the `zio' parameter 3225 * in the callback will be NULL in this case, since no IO was 3226 * required. If the block is not in the cache pass the read request 3227 * on to the spa with a substitute callback function, so that the 3228 * requested block will be added to the cache. 3229 * 3230 * If a read request arrives for a block that has a read in-progress, 3231 * either wait for the in-progress read to complete (and return the 3232 * results); or, if this is a read with a "done" func, add a record 3233 * to the read to invoke the "done" func when the read completes, 3234 * and return; or just return. 3235 * 3236 * arc_read_done() will invoke all the requested "done" functions 3237 * for readers of this block. 3238 */ 3239int 3240arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_done_func_t *done, 3241 void *private, zio_priority_t priority, int zio_flags, uint32_t *arc_flags, 3242 const zbookmark_phys_t *zb) 3243{ 3244 arc_buf_hdr_t *hdr = NULL; 3245 arc_buf_t *buf = NULL; 3246 kmutex_t *hash_lock = NULL; 3247 zio_t *rzio; 3248 uint64_t guid = spa_load_guid(spa); 3249 3250 ASSERT(!BP_IS_EMBEDDED(bp) || 3251 BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA); 3252 3253top: 3254 if (!BP_IS_EMBEDDED(bp)) { 3255 /* 3256 * Embedded BP's have no DVA and require no I/O to "read". 3257 * Create an anonymous arc buf to back it. 3258 */ 3259 hdr = buf_hash_find(guid, bp, &hash_lock); 3260 } 3261 3262 if (hdr != NULL && hdr->b_datacnt > 0) { 3263 3264 *arc_flags |= ARC_CACHED; 3265 3266 if (HDR_IO_IN_PROGRESS(hdr)) { 3267 3268 if (*arc_flags & ARC_WAIT) { 3269 cv_wait(&hdr->b_cv, hash_lock); 3270 mutex_exit(hash_lock); 3271 goto top; 3272 } 3273 ASSERT(*arc_flags & ARC_NOWAIT); 3274 3275 if (done) { 3276 arc_callback_t *acb = NULL; 3277 3278 acb = kmem_zalloc(sizeof (arc_callback_t), 3279 KM_SLEEP); 3280 acb->acb_done = done; 3281 acb->acb_private = private; 3282 if (pio != NULL) 3283 acb->acb_zio_dummy = zio_null(pio, 3284 spa, NULL, NULL, NULL, zio_flags); 3285 3286 ASSERT(acb->acb_done != NULL); 3287 acb->acb_next = hdr->b_acb; 3288 hdr->b_acb = acb; 3289 add_reference(hdr, hash_lock, private); 3290 mutex_exit(hash_lock); 3291 return (0); 3292 } 3293 mutex_exit(hash_lock); 3294 return (0); 3295 } 3296 3297 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu); 3298 3299 if (done) { 3300 add_reference(hdr, hash_lock, private); 3301 /* 3302 * If this block is already in use, create a new 3303 * copy of the data so that we will be guaranteed 3304 * that arc_release() will always succeed. 3305 */ 3306 buf = hdr->b_buf; 3307 ASSERT(buf); 3308 ASSERT(buf->b_data); 3309 if (HDR_BUF_AVAILABLE(hdr)) { 3310 ASSERT(buf->b_efunc == NULL); 3311 hdr->b_flags &= ~ARC_BUF_AVAILABLE; 3312 } else { 3313 buf = arc_buf_clone(buf); 3314 } 3315 3316 } else if (*arc_flags & ARC_PREFETCH && 3317 refcount_count(&hdr->b_refcnt) == 0) { 3318 hdr->b_flags |= ARC_PREFETCH; 3319 } 3320 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr); 3321 arc_access(hdr, hash_lock); 3322 if (*arc_flags & ARC_L2CACHE) 3323 hdr->b_flags |= ARC_L2CACHE; 3324 if (*arc_flags & ARC_L2COMPRESS) 3325 hdr->b_flags |= ARC_L2COMPRESS; 3326 mutex_exit(hash_lock); 3327 ARCSTAT_BUMP(arcstat_hits); 3328 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH), 3329 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA, 3330 data, metadata, hits); 3331 3332 if (done) 3333 done(NULL, buf, private); 3334 } else { 3335 uint64_t size = BP_GET_LSIZE(bp); 3336 arc_callback_t *acb; 3337 vdev_t *vd = NULL; 3338 uint64_t addr = 0; 3339 boolean_t devw = B_FALSE; 3340 enum zio_compress b_compress = ZIO_COMPRESS_OFF; 3341 uint64_t b_asize = 0; 3342 3343 if (hdr == NULL) { 3344 /* this block is not in the cache */ 3345 arc_buf_hdr_t *exists = NULL; 3346 arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp); 3347 buf = arc_buf_alloc(spa, size, private, type); 3348 hdr = buf->b_hdr; 3349 if (!BP_IS_EMBEDDED(bp)) { 3350 hdr->b_dva = *BP_IDENTITY(bp); 3351 hdr->b_birth = BP_PHYSICAL_BIRTH(bp); 3352 hdr->b_cksum0 = bp->blk_cksum.zc_word[0]; 3353 exists = buf_hash_insert(hdr, &hash_lock); 3354 } 3355 if (exists != NULL) { 3356 /* somebody beat us to the hash insert */ 3357 mutex_exit(hash_lock); 3358 buf_discard_identity(hdr); 3359 (void) arc_buf_remove_ref(buf, private); 3360 goto top; /* restart the IO request */ 3361 } 3362 /* if this is a prefetch, we don't have a reference */ 3363 if (*arc_flags & ARC_PREFETCH) { 3364 (void) remove_reference(hdr, hash_lock, 3365 private); 3366 hdr->b_flags |= ARC_PREFETCH; 3367 } 3368 if (*arc_flags & ARC_L2CACHE) 3369 hdr->b_flags |= ARC_L2CACHE; 3370 if (*arc_flags & ARC_L2COMPRESS) 3371 hdr->b_flags |= ARC_L2COMPRESS; 3372 if (BP_GET_LEVEL(bp) > 0) 3373 hdr->b_flags |= ARC_INDIRECT; 3374 } else { 3375 /* this block is in the ghost cache */ 3376 ASSERT(GHOST_STATE(hdr->b_state)); 3377 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 3378 ASSERT0(refcount_count(&hdr->b_refcnt)); 3379 ASSERT(hdr->b_buf == NULL); 3380 3381 /* if this is a prefetch, we don't have a reference */ 3382 if (*arc_flags & ARC_PREFETCH) 3383 hdr->b_flags |= ARC_PREFETCH; 3384 else 3385 add_reference(hdr, hash_lock, private); 3386 if (*arc_flags & ARC_L2CACHE) 3387 hdr->b_flags |= ARC_L2CACHE; 3388 if (*arc_flags & ARC_L2COMPRESS) 3389 hdr->b_flags |= ARC_L2COMPRESS; 3390 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 3391 buf->b_hdr = hdr; 3392 buf->b_data = NULL; 3393 buf->b_efunc = NULL; 3394 buf->b_private = NULL; 3395 buf->b_next = NULL; 3396 hdr->b_buf = buf; 3397 ASSERT(hdr->b_datacnt == 0); 3398 hdr->b_datacnt = 1; 3399 arc_get_data_buf(buf); 3400 arc_access(hdr, hash_lock); 3401 } 3402 3403 ASSERT(!GHOST_STATE(hdr->b_state)); 3404 3405 acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP); 3406 acb->acb_done = done; 3407 acb->acb_private = private; 3408 3409 ASSERT(hdr->b_acb == NULL); 3410 hdr->b_acb = acb; 3411 hdr->b_flags |= ARC_IO_IN_PROGRESS; 3412 3413 if (hdr->b_l2hdr != NULL && 3414 (vd = hdr->b_l2hdr->b_dev->l2ad_vdev) != NULL) { 3415 devw = hdr->b_l2hdr->b_dev->l2ad_writing; 3416 addr = hdr->b_l2hdr->b_daddr; 3417 b_compress = hdr->b_l2hdr->b_compress; 3418 b_asize = hdr->b_l2hdr->b_asize; 3419 /* 3420 * Lock out device removal. 3421 */ 3422 if (vdev_is_dead(vd) || 3423 !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER)) 3424 vd = NULL; 3425 } 3426 3427 if (hash_lock != NULL) 3428 mutex_exit(hash_lock); 3429 3430 /* 3431 * At this point, we have a level 1 cache miss. Try again in 3432 * L2ARC if possible. 3433 */ 3434 ASSERT3U(hdr->b_size, ==, size); 3435 DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp, 3436 uint64_t, size, zbookmark_phys_t *, zb); 3437 ARCSTAT_BUMP(arcstat_misses); 3438 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH), 3439 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA, 3440 data, metadata, misses); 3441#ifdef _KERNEL 3442 curthread->td_ru.ru_inblock++; 3443#endif 3444 3445 if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) { 3446 /* 3447 * Read from the L2ARC if the following are true: 3448 * 1. The L2ARC vdev was previously cached. 3449 * 2. This buffer still has L2ARC metadata. 3450 * 3. This buffer isn't currently writing to the L2ARC. 3451 * 4. The L2ARC entry wasn't evicted, which may 3452 * also have invalidated the vdev. 3453 * 5. This isn't prefetch and l2arc_noprefetch is set. 3454 */ 3455 if (hdr->b_l2hdr != NULL && 3456 !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) && 3457 !(l2arc_noprefetch && HDR_PREFETCH(hdr))) { 3458 l2arc_read_callback_t *cb; 3459 3460 DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr); 3461 ARCSTAT_BUMP(arcstat_l2_hits); 3462 3463 cb = kmem_zalloc(sizeof (l2arc_read_callback_t), 3464 KM_SLEEP); 3465 cb->l2rcb_buf = buf; 3466 cb->l2rcb_spa = spa; 3467 cb->l2rcb_bp = *bp; 3468 cb->l2rcb_zb = *zb; 3469 cb->l2rcb_flags = zio_flags; 3470 cb->l2rcb_compress = b_compress; 3471 3472 ASSERT(addr >= VDEV_LABEL_START_SIZE && 3473 addr + size < vd->vdev_psize - 3474 VDEV_LABEL_END_SIZE); 3475 3476 /* 3477 * l2arc read. The SCL_L2ARC lock will be 3478 * released by l2arc_read_done(). 3479 * Issue a null zio if the underlying buffer 3480 * was squashed to zero size by compression. 3481 */ 3482 if (b_compress == ZIO_COMPRESS_EMPTY) { 3483 rzio = zio_null(pio, spa, vd, 3484 l2arc_read_done, cb, 3485 zio_flags | ZIO_FLAG_DONT_CACHE | 3486 ZIO_FLAG_CANFAIL | 3487 ZIO_FLAG_DONT_PROPAGATE | 3488 ZIO_FLAG_DONT_RETRY); 3489 } else { 3490 rzio = zio_read_phys(pio, vd, addr, 3491 b_asize, buf->b_data, 3492 ZIO_CHECKSUM_OFF, 3493 l2arc_read_done, cb, priority, 3494 zio_flags | ZIO_FLAG_DONT_CACHE | 3495 ZIO_FLAG_CANFAIL | 3496 ZIO_FLAG_DONT_PROPAGATE | 3497 ZIO_FLAG_DONT_RETRY, B_FALSE); 3498 } 3499 DTRACE_PROBE2(l2arc__read, vdev_t *, vd, 3500 zio_t *, rzio); 3501 ARCSTAT_INCR(arcstat_l2_read_bytes, b_asize); 3502 3503 if (*arc_flags & ARC_NOWAIT) { 3504 zio_nowait(rzio); 3505 return (0); 3506 } 3507 3508 ASSERT(*arc_flags & ARC_WAIT); 3509 if (zio_wait(rzio) == 0) 3510 return (0); 3511 3512 /* l2arc read error; goto zio_read() */ 3513 } else { 3514 DTRACE_PROBE1(l2arc__miss, 3515 arc_buf_hdr_t *, hdr); 3516 ARCSTAT_BUMP(arcstat_l2_misses); 3517 if (HDR_L2_WRITING(hdr)) 3518 ARCSTAT_BUMP(arcstat_l2_rw_clash); 3519 spa_config_exit(spa, SCL_L2ARC, vd); 3520 } 3521 } else { 3522 if (vd != NULL) 3523 spa_config_exit(spa, SCL_L2ARC, vd); 3524 if (l2arc_ndev != 0) { 3525 DTRACE_PROBE1(l2arc__miss, 3526 arc_buf_hdr_t *, hdr); 3527 ARCSTAT_BUMP(arcstat_l2_misses); 3528 } 3529 } 3530 3531 rzio = zio_read(pio, spa, bp, buf->b_data, size, 3532 arc_read_done, buf, priority, zio_flags, zb); 3533 3534 if (*arc_flags & ARC_WAIT) 3535 return (zio_wait(rzio)); 3536 3537 ASSERT(*arc_flags & ARC_NOWAIT); 3538 zio_nowait(rzio); 3539 } 3540 return (0); 3541} 3542 3543void 3544arc_set_callback(arc_buf_t *buf, arc_evict_func_t *func, void *private) 3545{ 3546 ASSERT(buf->b_hdr != NULL); 3547 ASSERT(buf->b_hdr->b_state != arc_anon); 3548 ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt) || func == NULL); 3549 ASSERT(buf->b_efunc == NULL); 3550 ASSERT(!HDR_BUF_AVAILABLE(buf->b_hdr)); 3551 3552 buf->b_efunc = func; 3553 buf->b_private = private; 3554} 3555 3556/* 3557 * Notify the arc that a block was freed, and thus will never be used again. 3558 */ 3559void 3560arc_freed(spa_t *spa, const blkptr_t *bp) 3561{ 3562 arc_buf_hdr_t *hdr; 3563 kmutex_t *hash_lock; 3564 uint64_t guid = spa_load_guid(spa); 3565 3566 ASSERT(!BP_IS_EMBEDDED(bp)); 3567 3568 hdr = buf_hash_find(guid, bp, &hash_lock); 3569 if (hdr == NULL) 3570 return; 3571 if (HDR_BUF_AVAILABLE(hdr)) { 3572 arc_buf_t *buf = hdr->b_buf; 3573 add_reference(hdr, hash_lock, FTAG); 3574 hdr->b_flags &= ~ARC_BUF_AVAILABLE; 3575 mutex_exit(hash_lock); 3576 3577 arc_release(buf, FTAG); 3578 (void) arc_buf_remove_ref(buf, FTAG); 3579 } else { 3580 mutex_exit(hash_lock); 3581 } 3582 3583} 3584 3585/* 3586 * Clear the user eviction callback set by arc_set_callback(), first calling 3587 * it if it exists. Because the presence of a callback keeps an arc_buf cached 3588 * clearing the callback may result in the arc_buf being destroyed. However, 3589 * it will not result in the *last* arc_buf being destroyed, hence the data 3590 * will remain cached in the ARC. We make a copy of the arc buffer here so 3591 * that we can process the callback without holding any locks. 3592 * 3593 * It's possible that the callback is already in the process of being cleared 3594 * by another thread. In this case we can not clear the callback. 3595 * 3596 * Returns B_TRUE if the callback was successfully called and cleared. 3597 */ 3598boolean_t 3599arc_clear_callback(arc_buf_t *buf) 3600{ 3601 arc_buf_hdr_t *hdr; 3602 kmutex_t *hash_lock; 3603 arc_evict_func_t *efunc = buf->b_efunc; 3604 void *private = buf->b_private; 3605 list_t *list, *evicted_list; 3606 kmutex_t *lock, *evicted_lock; 3607 3608 mutex_enter(&buf->b_evict_lock); 3609 hdr = buf->b_hdr; 3610 if (hdr == NULL) { 3611 /* 3612 * We are in arc_do_user_evicts(). 3613 */ 3614 ASSERT(buf->b_data == NULL); 3615 mutex_exit(&buf->b_evict_lock); 3616 return (B_FALSE); 3617 } else if (buf->b_data == NULL) { 3618 /* 3619 * We are on the eviction list; process this buffer now 3620 * but let arc_do_user_evicts() do the reaping. 3621 */ 3622 buf->b_efunc = NULL; 3623 mutex_exit(&buf->b_evict_lock); 3624 VERIFY0(efunc(private)); 3625 return (B_TRUE); 3626 } 3627 hash_lock = HDR_LOCK(hdr); 3628 mutex_enter(hash_lock); 3629 hdr = buf->b_hdr; 3630 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 3631 3632 ASSERT3U(refcount_count(&hdr->b_refcnt), <, hdr->b_datacnt); 3633 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu); 3634 3635 buf->b_efunc = NULL; 3636 buf->b_private = NULL; 3637 3638 if (hdr->b_datacnt > 1) { 3639 mutex_exit(&buf->b_evict_lock); 3640 arc_buf_destroy(buf, FALSE, TRUE); 3641 } else { 3642 ASSERT(buf == hdr->b_buf); 3643 hdr->b_flags |= ARC_BUF_AVAILABLE; 3644 mutex_exit(&buf->b_evict_lock); 3645 } 3646 3647 mutex_exit(hash_lock); 3648 VERIFY0(efunc(private)); 3649 return (B_TRUE); 3650} 3651 3652/* 3653 * Release this buffer from the cache, making it an anonymous buffer. This 3654 * must be done after a read and prior to modifying the buffer contents. 3655 * If the buffer has more than one reference, we must make 3656 * a new hdr for the buffer. 3657 */ 3658void 3659arc_release(arc_buf_t *buf, void *tag) 3660{ 3661 arc_buf_hdr_t *hdr; 3662 kmutex_t *hash_lock = NULL; 3663 l2arc_buf_hdr_t *l2hdr; 3664 uint64_t buf_size; 3665 3666 /* 3667 * It would be nice to assert that if it's DMU metadata (level > 3668 * 0 || it's the dnode file), then it must be syncing context. 3669 * But we don't know that information at this level. 3670 */ 3671 3672 mutex_enter(&buf->b_evict_lock); 3673 hdr = buf->b_hdr; 3674 3675 /* this buffer is not on any list */ 3676 ASSERT(refcount_count(&hdr->b_refcnt) > 0); 3677 3678 if (hdr->b_state == arc_anon) { 3679 /* this buffer is already released */ 3680 ASSERT(buf->b_efunc == NULL); 3681 } else { 3682 hash_lock = HDR_LOCK(hdr); 3683 mutex_enter(hash_lock); 3684 hdr = buf->b_hdr; 3685 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 3686 } 3687 3688 l2hdr = hdr->b_l2hdr; 3689 if (l2hdr) { 3690 mutex_enter(&l2arc_buflist_mtx); 3691 hdr->b_l2hdr = NULL; 3692 list_remove(l2hdr->b_dev->l2ad_buflist, hdr); 3693 } 3694 buf_size = hdr->b_size; 3695 3696 /* 3697 * Do we have more than one buf? 3698 */ 3699 if (hdr->b_datacnt > 1) { 3700 arc_buf_hdr_t *nhdr; 3701 arc_buf_t **bufp; 3702 uint64_t blksz = hdr->b_size; 3703 uint64_t spa = hdr->b_spa; 3704 arc_buf_contents_t type = hdr->b_type; 3705 uint32_t flags = hdr->b_flags; 3706 3707 ASSERT(hdr->b_buf != buf || buf->b_next != NULL); 3708 /* 3709 * Pull the data off of this hdr and attach it to 3710 * a new anonymous hdr. 3711 */ 3712 (void) remove_reference(hdr, hash_lock, tag); 3713 bufp = &hdr->b_buf; 3714 while (*bufp != buf) 3715 bufp = &(*bufp)->b_next; 3716 *bufp = buf->b_next; 3717 buf->b_next = NULL; 3718 3719 ASSERT3U(hdr->b_state->arcs_size, >=, hdr->b_size); 3720 atomic_add_64(&hdr->b_state->arcs_size, -hdr->b_size); 3721 if (refcount_is_zero(&hdr->b_refcnt)) { 3722 uint64_t *size = &hdr->b_state->arcs_lsize[hdr->b_type]; 3723 ASSERT3U(*size, >=, hdr->b_size); 3724 atomic_add_64(size, -hdr->b_size); 3725 } 3726 3727 /* 3728 * We're releasing a duplicate user data buffer, update 3729 * our statistics accordingly. 3730 */ 3731 if (hdr->b_type == ARC_BUFC_DATA) { 3732 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers); 3733 ARCSTAT_INCR(arcstat_duplicate_buffers_size, 3734 -hdr->b_size); 3735 } 3736 hdr->b_datacnt -= 1; 3737 arc_cksum_verify(buf); 3738#ifdef illumos 3739 arc_buf_unwatch(buf); 3740#endif /* illumos */ 3741 3742 mutex_exit(hash_lock); 3743 3744 nhdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE); 3745 nhdr->b_size = blksz; 3746 nhdr->b_spa = spa; 3747 nhdr->b_type = type; 3748 nhdr->b_buf = buf; 3749 nhdr->b_state = arc_anon; 3750 nhdr->b_arc_access = 0; 3751 nhdr->b_flags = flags & ARC_L2_WRITING; 3752 nhdr->b_l2hdr = NULL; 3753 nhdr->b_datacnt = 1; 3754 nhdr->b_freeze_cksum = NULL; 3755 (void) refcount_add(&nhdr->b_refcnt, tag); 3756 buf->b_hdr = nhdr; 3757 mutex_exit(&buf->b_evict_lock); 3758 atomic_add_64(&arc_anon->arcs_size, blksz); 3759 } else { 3760 mutex_exit(&buf->b_evict_lock); 3761 ASSERT(refcount_count(&hdr->b_refcnt) == 1); 3762 ASSERT(!list_link_active(&hdr->b_arc_node)); 3763 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 3764 if (hdr->b_state != arc_anon) 3765 arc_change_state(arc_anon, hdr, hash_lock); 3766 hdr->b_arc_access = 0; 3767 if (hash_lock) 3768 mutex_exit(hash_lock); 3769 3770 buf_discard_identity(hdr); 3771 arc_buf_thaw(buf); 3772 } 3773 buf->b_efunc = NULL; 3774 buf->b_private = NULL; 3775 3776 if (l2hdr) { 3777 ARCSTAT_INCR(arcstat_l2_asize, -l2hdr->b_asize); 3778 vdev_space_update(l2hdr->b_dev->l2ad_vdev, 3779 -l2hdr->b_asize, 0, 0); 3780 trim_map_free(l2hdr->b_dev->l2ad_vdev, l2hdr->b_daddr, 3781 hdr->b_size, 0); 3782 kmem_free(l2hdr, sizeof (l2arc_buf_hdr_t)); 3783 ARCSTAT_INCR(arcstat_l2_size, -buf_size); 3784 mutex_exit(&l2arc_buflist_mtx); 3785 } 3786} 3787 3788int 3789arc_released(arc_buf_t *buf) 3790{ 3791 int released; 3792 3793 mutex_enter(&buf->b_evict_lock); 3794 released = (buf->b_data != NULL && buf->b_hdr->b_state == arc_anon); 3795 mutex_exit(&buf->b_evict_lock); 3796 return (released); 3797} 3798 3799#ifdef ZFS_DEBUG 3800int 3801arc_referenced(arc_buf_t *buf) 3802{ 3803 int referenced; 3804 3805 mutex_enter(&buf->b_evict_lock); 3806 referenced = (refcount_count(&buf->b_hdr->b_refcnt)); 3807 mutex_exit(&buf->b_evict_lock); 3808 return (referenced); 3809} 3810#endif 3811 3812static void 3813arc_write_ready(zio_t *zio) 3814{ 3815 arc_write_callback_t *callback = zio->io_private; 3816 arc_buf_t *buf = callback->awcb_buf; 3817 arc_buf_hdr_t *hdr = buf->b_hdr; 3818 3819 ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt)); 3820 callback->awcb_ready(zio, buf, callback->awcb_private); 3821 3822 /* 3823 * If the IO is already in progress, then this is a re-write 3824 * attempt, so we need to thaw and re-compute the cksum. 3825 * It is the responsibility of the callback to handle the 3826 * accounting for any re-write attempt. 3827 */ 3828 if (HDR_IO_IN_PROGRESS(hdr)) { 3829 mutex_enter(&hdr->b_freeze_lock); 3830 if (hdr->b_freeze_cksum != NULL) { 3831 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 3832 hdr->b_freeze_cksum = NULL; 3833 } 3834 mutex_exit(&hdr->b_freeze_lock); 3835 } 3836 arc_cksum_compute(buf, B_FALSE); 3837 hdr->b_flags |= ARC_IO_IN_PROGRESS; 3838} 3839 3840/* 3841 * The SPA calls this callback for each physical write that happens on behalf 3842 * of a logical write. See the comment in dbuf_write_physdone() for details. 3843 */ 3844static void 3845arc_write_physdone(zio_t *zio) 3846{ 3847 arc_write_callback_t *cb = zio->io_private; 3848 if (cb->awcb_physdone != NULL) 3849 cb->awcb_physdone(zio, cb->awcb_buf, cb->awcb_private); 3850} 3851 3852static void 3853arc_write_done(zio_t *zio) 3854{ 3855 arc_write_callback_t *callback = zio->io_private; 3856 arc_buf_t *buf = callback->awcb_buf; 3857 arc_buf_hdr_t *hdr = buf->b_hdr; 3858 3859 ASSERT(hdr->b_acb == NULL); 3860 3861 if (zio->io_error == 0) { 3862 if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) { 3863 buf_discard_identity(hdr); 3864 } else { 3865 hdr->b_dva = *BP_IDENTITY(zio->io_bp); 3866 hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp); 3867 hdr->b_cksum0 = zio->io_bp->blk_cksum.zc_word[0]; 3868 } 3869 } else { 3870 ASSERT(BUF_EMPTY(hdr)); 3871 } 3872 3873 /* 3874 * If the block to be written was all-zero or compressed enough to be 3875 * embedded in the BP, no write was performed so there will be no 3876 * dva/birth/checksum. The buffer must therefore remain anonymous 3877 * (and uncached). 3878 */ 3879 if (!BUF_EMPTY(hdr)) { 3880 arc_buf_hdr_t *exists; 3881 kmutex_t *hash_lock; 3882 3883 ASSERT(zio->io_error == 0); 3884 3885 arc_cksum_verify(buf); 3886 3887 exists = buf_hash_insert(hdr, &hash_lock); 3888 if (exists) { 3889 /* 3890 * This can only happen if we overwrite for 3891 * sync-to-convergence, because we remove 3892 * buffers from the hash table when we arc_free(). 3893 */ 3894 if (zio->io_flags & ZIO_FLAG_IO_REWRITE) { 3895 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) 3896 panic("bad overwrite, hdr=%p exists=%p", 3897 (void *)hdr, (void *)exists); 3898 ASSERT(refcount_is_zero(&exists->b_refcnt)); 3899 arc_change_state(arc_anon, exists, hash_lock); 3900 mutex_exit(hash_lock); 3901 arc_hdr_destroy(exists); 3902 exists = buf_hash_insert(hdr, &hash_lock); 3903 ASSERT3P(exists, ==, NULL); 3904 } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) { 3905 /* nopwrite */ 3906 ASSERT(zio->io_prop.zp_nopwrite); 3907 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) 3908 panic("bad nopwrite, hdr=%p exists=%p", 3909 (void *)hdr, (void *)exists); 3910 } else { 3911 /* Dedup */ 3912 ASSERT(hdr->b_datacnt == 1); 3913 ASSERT(hdr->b_state == arc_anon); 3914 ASSERT(BP_GET_DEDUP(zio->io_bp)); 3915 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0); 3916 } 3917 } 3918 hdr->b_flags &= ~ARC_IO_IN_PROGRESS; 3919 /* if it's not anon, we are doing a scrub */ 3920 if (!exists && hdr->b_state == arc_anon) 3921 arc_access(hdr, hash_lock); 3922 mutex_exit(hash_lock); 3923 } else { 3924 hdr->b_flags &= ~ARC_IO_IN_PROGRESS; 3925 } 3926 3927 ASSERT(!refcount_is_zero(&hdr->b_refcnt)); 3928 callback->awcb_done(zio, buf, callback->awcb_private); 3929 3930 kmem_free(callback, sizeof (arc_write_callback_t)); 3931} 3932 3933zio_t * 3934arc_write(zio_t *pio, spa_t *spa, uint64_t txg, 3935 blkptr_t *bp, arc_buf_t *buf, boolean_t l2arc, boolean_t l2arc_compress, 3936 const zio_prop_t *zp, arc_done_func_t *ready, arc_done_func_t *physdone, 3937 arc_done_func_t *done, void *private, zio_priority_t priority, 3938 int zio_flags, const zbookmark_phys_t *zb) 3939{ 3940 arc_buf_hdr_t *hdr = buf->b_hdr; 3941 arc_write_callback_t *callback; 3942 zio_t *zio; 3943 3944 ASSERT(ready != NULL); 3945 ASSERT(done != NULL); 3946 ASSERT(!HDR_IO_ERROR(hdr)); 3947 ASSERT((hdr->b_flags & ARC_IO_IN_PROGRESS) == 0); 3948 ASSERT(hdr->b_acb == NULL); 3949 if (l2arc) 3950 hdr->b_flags |= ARC_L2CACHE; 3951 if (l2arc_compress) 3952 hdr->b_flags |= ARC_L2COMPRESS; 3953 callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP); 3954 callback->awcb_ready = ready; 3955 callback->awcb_physdone = physdone; 3956 callback->awcb_done = done; 3957 callback->awcb_private = private; 3958 callback->awcb_buf = buf; 3959 3960 zio = zio_write(pio, spa, txg, bp, buf->b_data, hdr->b_size, zp, 3961 arc_write_ready, arc_write_physdone, arc_write_done, callback, 3962 priority, zio_flags, zb); 3963 3964 return (zio); 3965} 3966 3967static int 3968arc_memory_throttle(uint64_t reserve, uint64_t txg) 3969{ 3970#ifdef _KERNEL 3971 uint64_t available_memory = ptob(freemem); 3972 static uint64_t page_load = 0; 3973 static uint64_t last_txg = 0; 3974 3975#if defined(__i386) || !defined(UMA_MD_SMALL_ALLOC) 3976 available_memory = 3977 MIN(available_memory, ptob(vmem_size(heap_arena, VMEM_FREE))); 3978#endif 3979 3980 if (freemem > (uint64_t)physmem * arc_lotsfree_percent / 100) 3981 return (0); 3982 3983 if (txg > last_txg) { 3984 last_txg = txg; 3985 page_load = 0; 3986 } 3987 /* 3988 * If we are in pageout, we know that memory is already tight, 3989 * the arc is already going to be evicting, so we just want to 3990 * continue to let page writes occur as quickly as possible. 3991 */ 3992 if (curproc == pageproc) { 3993 if (page_load > MAX(ptob(minfree), available_memory) / 4) 3994 return (SET_ERROR(ERESTART)); 3995 /* Note: reserve is inflated, so we deflate */ 3996 page_load += reserve / 8; 3997 return (0); 3998 } else if (page_load > 0 && arc_reclaim_needed()) { 3999 /* memory is low, delay before restarting */ 4000 ARCSTAT_INCR(arcstat_memory_throttle_count, 1); 4001 return (SET_ERROR(EAGAIN)); 4002 } 4003 page_load = 0; 4004#endif 4005 return (0); 4006} 4007 4008void 4009arc_tempreserve_clear(uint64_t reserve) 4010{ 4011 atomic_add_64(&arc_tempreserve, -reserve); 4012 ASSERT((int64_t)arc_tempreserve >= 0); 4013} 4014 4015int 4016arc_tempreserve_space(uint64_t reserve, uint64_t txg) 4017{ 4018 int error; 4019 uint64_t anon_size; 4020 4021 if (reserve > arc_c/4 && !arc_no_grow) { 4022 arc_c = MIN(arc_c_max, reserve * 4); 4023 DTRACE_PROBE1(arc__set_reserve, uint64_t, arc_c); 4024 } 4025 if (reserve > arc_c) 4026 return (SET_ERROR(ENOMEM)); 4027 4028 /* 4029 * Don't count loaned bufs as in flight dirty data to prevent long 4030 * network delays from blocking transactions that are ready to be 4031 * assigned to a txg. 4032 */ 4033 anon_size = MAX((int64_t)(arc_anon->arcs_size - arc_loaned_bytes), 0); 4034 4035 /* 4036 * Writes will, almost always, require additional memory allocations 4037 * in order to compress/encrypt/etc the data. We therefore need to 4038 * make sure that there is sufficient available memory for this. 4039 */ 4040 error = arc_memory_throttle(reserve, txg); 4041 if (error != 0) 4042 return (error); 4043 4044 /* 4045 * Throttle writes when the amount of dirty data in the cache 4046 * gets too large. We try to keep the cache less than half full 4047 * of dirty blocks so that our sync times don't grow too large. 4048 * Note: if two requests come in concurrently, we might let them 4049 * both succeed, when one of them should fail. Not a huge deal. 4050 */ 4051 4052 if (reserve + arc_tempreserve + anon_size > arc_c / 2 && 4053 anon_size > arc_c / 4) { 4054 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK " 4055 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n", 4056 arc_tempreserve>>10, 4057 arc_anon->arcs_lsize[ARC_BUFC_METADATA]>>10, 4058 arc_anon->arcs_lsize[ARC_BUFC_DATA]>>10, 4059 reserve>>10, arc_c>>10); 4060 return (SET_ERROR(ERESTART)); 4061 } 4062 atomic_add_64(&arc_tempreserve, reserve); 4063 return (0); 4064} 4065 4066static kmutex_t arc_lowmem_lock; 4067#ifdef _KERNEL 4068static eventhandler_tag arc_event_lowmem = NULL; 4069 4070static void 4071arc_lowmem(void *arg __unused, int howto __unused) 4072{ 4073 4074 /* Serialize access via arc_lowmem_lock. */ 4075 mutex_enter(&arc_lowmem_lock); 4076 mutex_enter(&arc_reclaim_thr_lock); 4077 needfree = 1; 4078 DTRACE_PROBE(arc__needfree); 4079 cv_signal(&arc_reclaim_thr_cv); 4080 4081 /* 4082 * It is unsafe to block here in arbitrary threads, because we can come 4083 * here from ARC itself and may hold ARC locks and thus risk a deadlock 4084 * with ARC reclaim thread. 4085 */ 4086 if (curproc == pageproc) { 4087 while (needfree) 4088 msleep(&needfree, &arc_reclaim_thr_lock, 0, "zfs:lowmem", 0); 4089 } 4090 mutex_exit(&arc_reclaim_thr_lock); 4091 mutex_exit(&arc_lowmem_lock); 4092} 4093#endif 4094 4095void 4096arc_init(void) 4097{ 4098 int i, prefetch_tunable_set = 0; 4099 4100 mutex_init(&arc_reclaim_thr_lock, NULL, MUTEX_DEFAULT, NULL); 4101 cv_init(&arc_reclaim_thr_cv, NULL, CV_DEFAULT, NULL); 4102 mutex_init(&arc_lowmem_lock, NULL, MUTEX_DEFAULT, NULL); 4103 4104 /* Convert seconds to clock ticks */ 4105 arc_min_prefetch_lifespan = 1 * hz; 4106 4107 /* Start out with 1/8 of all memory */ 4108 arc_c = kmem_size() / 8; 4109 4110#ifdef sun 4111#ifdef _KERNEL 4112 /* 4113 * On architectures where the physical memory can be larger 4114 * than the addressable space (intel in 32-bit mode), we may 4115 * need to limit the cache to 1/8 of VM size. 4116 */ 4117 arc_c = MIN(arc_c, vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 8); 4118#endif 4119#endif /* sun */ 4120 /* set min cache to 1/32 of all memory, or 16MB, whichever is more */ 4121 arc_c_min = MAX(arc_c / 4, 64<<18); 4122 /* set max to 1/2 of all memory, or all but 1GB, whichever is more */ 4123 if (arc_c * 8 >= 1<<30) 4124 arc_c_max = (arc_c * 8) - (1<<30); 4125 else 4126 arc_c_max = arc_c_min; 4127 arc_c_max = MAX(arc_c * 5, arc_c_max); 4128 4129#ifdef _KERNEL 4130 /* 4131 * Allow the tunables to override our calculations if they are 4132 * reasonable (ie. over 16MB) 4133 */ 4134 if (zfs_arc_max > 64<<18 && zfs_arc_max < kmem_size()) 4135 arc_c_max = zfs_arc_max; 4136 if (zfs_arc_min > 64<<18 && zfs_arc_min <= arc_c_max) 4137 arc_c_min = zfs_arc_min; 4138#endif 4139 4140 arc_c = arc_c_max; 4141 arc_p = (arc_c >> 1); 4142 4143 /* limit meta-data to 1/4 of the arc capacity */ 4144 arc_meta_limit = arc_c_max / 4; 4145 4146 /* Allow the tunable to override if it is reasonable */ 4147 if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max) 4148 arc_meta_limit = zfs_arc_meta_limit; 4149 4150 if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0) 4151 arc_c_min = arc_meta_limit / 2; 4152 4153 if (zfs_arc_grow_retry > 0) 4154 arc_grow_retry = zfs_arc_grow_retry; 4155 4156 if (zfs_arc_shrink_shift > 0) 4157 arc_shrink_shift = zfs_arc_shrink_shift; 4158 4159 if (zfs_arc_p_min_shift > 0) 4160 arc_p_min_shift = zfs_arc_p_min_shift; 4161 4162 /* if kmem_flags are set, lets try to use less memory */ 4163 if (kmem_debugging()) 4164 arc_c = arc_c / 2; 4165 if (arc_c < arc_c_min) 4166 arc_c = arc_c_min; 4167 4168 zfs_arc_min = arc_c_min; 4169 zfs_arc_max = arc_c_max; 4170 4171 arc_anon = &ARC_anon; 4172 arc_mru = &ARC_mru; 4173 arc_mru_ghost = &ARC_mru_ghost; 4174 arc_mfu = &ARC_mfu; 4175 arc_mfu_ghost = &ARC_mfu_ghost; 4176 arc_l2c_only = &ARC_l2c_only; 4177 arc_size = 0; 4178 4179 for (i = 0; i < ARC_BUFC_NUMLISTS; i++) { 4180 mutex_init(&arc_anon->arcs_locks[i].arcs_lock, 4181 NULL, MUTEX_DEFAULT, NULL); 4182 mutex_init(&arc_mru->arcs_locks[i].arcs_lock, 4183 NULL, MUTEX_DEFAULT, NULL); 4184 mutex_init(&arc_mru_ghost->arcs_locks[i].arcs_lock, 4185 NULL, MUTEX_DEFAULT, NULL); 4186 mutex_init(&arc_mfu->arcs_locks[i].arcs_lock, 4187 NULL, MUTEX_DEFAULT, NULL); 4188 mutex_init(&arc_mfu_ghost->arcs_locks[i].arcs_lock, 4189 NULL, MUTEX_DEFAULT, NULL); 4190 mutex_init(&arc_l2c_only->arcs_locks[i].arcs_lock, 4191 NULL, MUTEX_DEFAULT, NULL); 4192 4193 list_create(&arc_mru->arcs_lists[i], 4194 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4195 list_create(&arc_mru_ghost->arcs_lists[i], 4196 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4197 list_create(&arc_mfu->arcs_lists[i], 4198 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4199 list_create(&arc_mfu_ghost->arcs_lists[i], 4200 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4201 list_create(&arc_mfu_ghost->arcs_lists[i], 4202 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4203 list_create(&arc_l2c_only->arcs_lists[i], 4204 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4205 } 4206 4207 buf_init(); 4208 4209 arc_thread_exit = 0; 4210 arc_eviction_list = NULL; 4211 mutex_init(&arc_eviction_mtx, NULL, MUTEX_DEFAULT, NULL); 4212 bzero(&arc_eviction_hdr, sizeof (arc_buf_hdr_t)); 4213 4214 arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED, 4215 sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL); 4216 4217 if (arc_ksp != NULL) { 4218 arc_ksp->ks_data = &arc_stats; 4219 kstat_install(arc_ksp); 4220 } 4221 4222 (void) thread_create(NULL, 0, arc_reclaim_thread, NULL, 0, &p0, 4223 TS_RUN, minclsyspri); 4224 4225#ifdef _KERNEL 4226 arc_event_lowmem = EVENTHANDLER_REGISTER(vm_lowmem, arc_lowmem, NULL, 4227 EVENTHANDLER_PRI_FIRST); 4228#endif 4229 4230 arc_dead = FALSE; 4231 arc_warm = B_FALSE; 4232 4233 /* 4234 * Calculate maximum amount of dirty data per pool. 4235 * 4236 * If it has been set by /etc/system, take that. 4237 * Otherwise, use a percentage of physical memory defined by 4238 * zfs_dirty_data_max_percent (default 10%) with a cap at 4239 * zfs_dirty_data_max_max (default 4GB). 4240 */ 4241 if (zfs_dirty_data_max == 0) { 4242 zfs_dirty_data_max = ptob(physmem) * 4243 zfs_dirty_data_max_percent / 100; 4244 zfs_dirty_data_max = MIN(zfs_dirty_data_max, 4245 zfs_dirty_data_max_max); 4246 } 4247 4248#ifdef _KERNEL 4249 if (TUNABLE_INT_FETCH("vfs.zfs.prefetch_disable", &zfs_prefetch_disable)) 4250 prefetch_tunable_set = 1; 4251 4252#ifdef __i386__ 4253 if (prefetch_tunable_set == 0) { 4254 printf("ZFS NOTICE: Prefetch is disabled by default on i386 " 4255 "-- to enable,\n"); 4256 printf(" add \"vfs.zfs.prefetch_disable=0\" " 4257 "to /boot/loader.conf.\n"); 4258 zfs_prefetch_disable = 1; 4259 } 4260#else 4261 if ((((uint64_t)physmem * PAGESIZE) < (1ULL << 32)) && 4262 prefetch_tunable_set == 0) { 4263 printf("ZFS NOTICE: Prefetch is disabled by default if less " 4264 "than 4GB of RAM is present;\n" 4265 " to enable, add \"vfs.zfs.prefetch_disable=0\" " 4266 "to /boot/loader.conf.\n"); 4267 zfs_prefetch_disable = 1; 4268 } 4269#endif 4270 /* Warn about ZFS memory and address space requirements. */ 4271 if (((uint64_t)physmem * PAGESIZE) < (256 + 128 + 64) * (1 << 20)) { 4272 printf("ZFS WARNING: Recommended minimum RAM size is 512MB; " 4273 "expect unstable behavior.\n"); 4274 } 4275 if (kmem_size() < 512 * (1 << 20)) { 4276 printf("ZFS WARNING: Recommended minimum kmem_size is 512MB; " 4277 "expect unstable behavior.\n"); 4278 printf(" Consider tuning vm.kmem_size and " 4279 "vm.kmem_size_max\n"); 4280 printf(" in /boot/loader.conf.\n"); 4281 } 4282#endif 4283} 4284 4285void 4286arc_fini(void) 4287{ 4288 int i; 4289 4290 mutex_enter(&arc_reclaim_thr_lock); 4291 arc_thread_exit = 1; 4292 cv_signal(&arc_reclaim_thr_cv); 4293 while (arc_thread_exit != 0) 4294 cv_wait(&arc_reclaim_thr_cv, &arc_reclaim_thr_lock); 4295 mutex_exit(&arc_reclaim_thr_lock); 4296 4297 arc_flush(NULL); 4298 4299 arc_dead = TRUE; 4300 4301 if (arc_ksp != NULL) { 4302 kstat_delete(arc_ksp); 4303 arc_ksp = NULL; 4304 } 4305 4306 mutex_destroy(&arc_eviction_mtx); 4307 mutex_destroy(&arc_reclaim_thr_lock); 4308 cv_destroy(&arc_reclaim_thr_cv); 4309 4310 for (i = 0; i < ARC_BUFC_NUMLISTS; i++) { 4311 list_destroy(&arc_mru->arcs_lists[i]); 4312 list_destroy(&arc_mru_ghost->arcs_lists[i]); 4313 list_destroy(&arc_mfu->arcs_lists[i]); 4314 list_destroy(&arc_mfu_ghost->arcs_lists[i]); 4315 list_destroy(&arc_l2c_only->arcs_lists[i]); 4316 4317 mutex_destroy(&arc_anon->arcs_locks[i].arcs_lock); 4318 mutex_destroy(&arc_mru->arcs_locks[i].arcs_lock); 4319 mutex_destroy(&arc_mru_ghost->arcs_locks[i].arcs_lock); 4320 mutex_destroy(&arc_mfu->arcs_locks[i].arcs_lock); 4321 mutex_destroy(&arc_mfu_ghost->arcs_locks[i].arcs_lock); 4322 mutex_destroy(&arc_l2c_only->arcs_locks[i].arcs_lock); 4323 } 4324 4325 buf_fini(); 4326 4327 ASSERT(arc_loaned_bytes == 0); 4328 4329 mutex_destroy(&arc_lowmem_lock); 4330#ifdef _KERNEL 4331 if (arc_event_lowmem != NULL) 4332 EVENTHANDLER_DEREGISTER(vm_lowmem, arc_event_lowmem); 4333#endif 4334} 4335 4336/* 4337 * Level 2 ARC 4338 * 4339 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk. 4340 * It uses dedicated storage devices to hold cached data, which are populated 4341 * using large infrequent writes. The main role of this cache is to boost 4342 * the performance of random read workloads. The intended L2ARC devices 4343 * include short-stroked disks, solid state disks, and other media with 4344 * substantially faster read latency than disk. 4345 * 4346 * +-----------------------+ 4347 * | ARC | 4348 * +-----------------------+ 4349 * | ^ ^ 4350 * | | | 4351 * l2arc_feed_thread() arc_read() 4352 * | | | 4353 * | l2arc read | 4354 * V | | 4355 * +---------------+ | 4356 * | L2ARC | | 4357 * +---------------+ | 4358 * | ^ | 4359 * l2arc_write() | | 4360 * | | | 4361 * V | | 4362 * +-------+ +-------+ 4363 * | vdev | | vdev | 4364 * | cache | | cache | 4365 * +-------+ +-------+ 4366 * +=========+ .-----. 4367 * : L2ARC : |-_____-| 4368 * : devices : | Disks | 4369 * +=========+ `-_____-' 4370 * 4371 * Read requests are satisfied from the following sources, in order: 4372 * 4373 * 1) ARC 4374 * 2) vdev cache of L2ARC devices 4375 * 3) L2ARC devices 4376 * 4) vdev cache of disks 4377 * 5) disks 4378 * 4379 * Some L2ARC device types exhibit extremely slow write performance. 4380 * To accommodate for this there are some significant differences between 4381 * the L2ARC and traditional cache design: 4382 * 4383 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from 4384 * the ARC behave as usual, freeing buffers and placing headers on ghost 4385 * lists. The ARC does not send buffers to the L2ARC during eviction as 4386 * this would add inflated write latencies for all ARC memory pressure. 4387 * 4388 * 2. The L2ARC attempts to cache data from the ARC before it is evicted. 4389 * It does this by periodically scanning buffers from the eviction-end of 4390 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are 4391 * not already there. It scans until a headroom of buffers is satisfied, 4392 * which itself is a buffer for ARC eviction. If a compressible buffer is 4393 * found during scanning and selected for writing to an L2ARC device, we 4394 * temporarily boost scanning headroom during the next scan cycle to make 4395 * sure we adapt to compression effects (which might significantly reduce 4396 * the data volume we write to L2ARC). The thread that does this is 4397 * l2arc_feed_thread(), illustrated below; example sizes are included to 4398 * provide a better sense of ratio than this diagram: 4399 * 4400 * head --> tail 4401 * +---------------------+----------+ 4402 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC 4403 * +---------------------+----------+ | o L2ARC eligible 4404 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer 4405 * +---------------------+----------+ | 4406 * 15.9 Gbytes ^ 32 Mbytes | 4407 * headroom | 4408 * l2arc_feed_thread() 4409 * | 4410 * l2arc write hand <--[oooo]--' 4411 * | 8 Mbyte 4412 * | write max 4413 * V 4414 * +==============================+ 4415 * L2ARC dev |####|#|###|###| |####| ... | 4416 * +==============================+ 4417 * 32 Gbytes 4418 * 4419 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of 4420 * evicted, then the L2ARC has cached a buffer much sooner than it probably 4421 * needed to, potentially wasting L2ARC device bandwidth and storage. It is 4422 * safe to say that this is an uncommon case, since buffers at the end of 4423 * the ARC lists have moved there due to inactivity. 4424 * 4425 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom, 4426 * then the L2ARC simply misses copying some buffers. This serves as a 4427 * pressure valve to prevent heavy read workloads from both stalling the ARC 4428 * with waits and clogging the L2ARC with writes. This also helps prevent 4429 * the potential for the L2ARC to churn if it attempts to cache content too 4430 * quickly, such as during backups of the entire pool. 4431 * 4432 * 5. After system boot and before the ARC has filled main memory, there are 4433 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru 4434 * lists can remain mostly static. Instead of searching from tail of these 4435 * lists as pictured, the l2arc_feed_thread() will search from the list heads 4436 * for eligible buffers, greatly increasing its chance of finding them. 4437 * 4438 * The L2ARC device write speed is also boosted during this time so that 4439 * the L2ARC warms up faster. Since there have been no ARC evictions yet, 4440 * there are no L2ARC reads, and no fear of degrading read performance 4441 * through increased writes. 4442 * 4443 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that 4444 * the vdev queue can aggregate them into larger and fewer writes. Each 4445 * device is written to in a rotor fashion, sweeping writes through 4446 * available space then repeating. 4447 * 4448 * 7. The L2ARC does not store dirty content. It never needs to flush 4449 * write buffers back to disk based storage. 4450 * 4451 * 8. If an ARC buffer is written (and dirtied) which also exists in the 4452 * L2ARC, the now stale L2ARC buffer is immediately dropped. 4453 * 4454 * The performance of the L2ARC can be tweaked by a number of tunables, which 4455 * may be necessary for different workloads: 4456 * 4457 * l2arc_write_max max write bytes per interval 4458 * l2arc_write_boost extra write bytes during device warmup 4459 * l2arc_noprefetch skip caching prefetched buffers 4460 * l2arc_headroom number of max device writes to precache 4461 * l2arc_headroom_boost when we find compressed buffers during ARC 4462 * scanning, we multiply headroom by this 4463 * percentage factor for the next scan cycle, 4464 * since more compressed buffers are likely to 4465 * be present 4466 * l2arc_feed_secs seconds between L2ARC writing 4467 * 4468 * Tunables may be removed or added as future performance improvements are 4469 * integrated, and also may become zpool properties. 4470 * 4471 * There are three key functions that control how the L2ARC warms up: 4472 * 4473 * l2arc_write_eligible() check if a buffer is eligible to cache 4474 * l2arc_write_size() calculate how much to write 4475 * l2arc_write_interval() calculate sleep delay between writes 4476 * 4477 * These three functions determine what to write, how much, and how quickly 4478 * to send writes. 4479 */ 4480 4481static boolean_t 4482l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *ab) 4483{ 4484 /* 4485 * A buffer is *not* eligible for the L2ARC if it: 4486 * 1. belongs to a different spa. 4487 * 2. is already cached on the L2ARC. 4488 * 3. has an I/O in progress (it may be an incomplete read). 4489 * 4. is flagged not eligible (zfs property). 4490 */ 4491 if (ab->b_spa != spa_guid) { 4492 ARCSTAT_BUMP(arcstat_l2_write_spa_mismatch); 4493 return (B_FALSE); 4494 } 4495 if (ab->b_l2hdr != NULL) { 4496 ARCSTAT_BUMP(arcstat_l2_write_in_l2); 4497 return (B_FALSE); 4498 } 4499 if (HDR_IO_IN_PROGRESS(ab)) { 4500 ARCSTAT_BUMP(arcstat_l2_write_hdr_io_in_progress); 4501 return (B_FALSE); 4502 } 4503 if (!HDR_L2CACHE(ab)) { 4504 ARCSTAT_BUMP(arcstat_l2_write_not_cacheable); 4505 return (B_FALSE); 4506 } 4507 4508 return (B_TRUE); 4509} 4510 4511static uint64_t 4512l2arc_write_size(void) 4513{ 4514 uint64_t size; 4515 4516 /* 4517 * Make sure our globals have meaningful values in case the user 4518 * altered them. 4519 */ 4520 size = l2arc_write_max; 4521 if (size == 0) { 4522 cmn_err(CE_NOTE, "Bad value for l2arc_write_max, value must " 4523 "be greater than zero, resetting it to the default (%d)", 4524 L2ARC_WRITE_SIZE); 4525 size = l2arc_write_max = L2ARC_WRITE_SIZE; 4526 } 4527 4528 if (arc_warm == B_FALSE) 4529 size += l2arc_write_boost; 4530 4531 return (size); 4532 4533} 4534 4535static clock_t 4536l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote) 4537{ 4538 clock_t interval, next, now; 4539 4540 /* 4541 * If the ARC lists are busy, increase our write rate; if the 4542 * lists are stale, idle back. This is achieved by checking 4543 * how much we previously wrote - if it was more than half of 4544 * what we wanted, schedule the next write much sooner. 4545 */ 4546 if (l2arc_feed_again && wrote > (wanted / 2)) 4547 interval = (hz * l2arc_feed_min_ms) / 1000; 4548 else 4549 interval = hz * l2arc_feed_secs; 4550 4551 now = ddi_get_lbolt(); 4552 next = MAX(now, MIN(now + interval, began + interval)); 4553 4554 return (next); 4555} 4556 4557static void 4558l2arc_hdr_stat_add(void) 4559{ 4560 ARCSTAT_INCR(arcstat_l2_hdr_size, HDR_SIZE + L2HDR_SIZE); 4561 ARCSTAT_INCR(arcstat_hdr_size, -HDR_SIZE); 4562} 4563 4564static void 4565l2arc_hdr_stat_remove(void) 4566{ 4567 ARCSTAT_INCR(arcstat_l2_hdr_size, -(HDR_SIZE + L2HDR_SIZE)); 4568 ARCSTAT_INCR(arcstat_hdr_size, HDR_SIZE); 4569} 4570 4571/* 4572 * Cycle through L2ARC devices. This is how L2ARC load balances. 4573 * If a device is returned, this also returns holding the spa config lock. 4574 */ 4575static l2arc_dev_t * 4576l2arc_dev_get_next(void) 4577{ 4578 l2arc_dev_t *first, *next = NULL; 4579 4580 /* 4581 * Lock out the removal of spas (spa_namespace_lock), then removal 4582 * of cache devices (l2arc_dev_mtx). Once a device has been selected, 4583 * both locks will be dropped and a spa config lock held instead. 4584 */ 4585 mutex_enter(&spa_namespace_lock); 4586 mutex_enter(&l2arc_dev_mtx); 4587 4588 /* if there are no vdevs, there is nothing to do */ 4589 if (l2arc_ndev == 0) 4590 goto out; 4591 4592 first = NULL; 4593 next = l2arc_dev_last; 4594 do { 4595 /* loop around the list looking for a non-faulted vdev */ 4596 if (next == NULL) { 4597 next = list_head(l2arc_dev_list); 4598 } else { 4599 next = list_next(l2arc_dev_list, next); 4600 if (next == NULL) 4601 next = list_head(l2arc_dev_list); 4602 } 4603 4604 /* if we have come back to the start, bail out */ 4605 if (first == NULL) 4606 first = next; 4607 else if (next == first) 4608 break; 4609 4610 } while (vdev_is_dead(next->l2ad_vdev)); 4611 4612 /* if we were unable to find any usable vdevs, return NULL */ 4613 if (vdev_is_dead(next->l2ad_vdev)) 4614 next = NULL; 4615 4616 l2arc_dev_last = next; 4617 4618out: 4619 mutex_exit(&l2arc_dev_mtx); 4620 4621 /* 4622 * Grab the config lock to prevent the 'next' device from being 4623 * removed while we are writing to it. 4624 */ 4625 if (next != NULL) 4626 spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER); 4627 mutex_exit(&spa_namespace_lock); 4628 4629 return (next); 4630} 4631 4632/* 4633 * Free buffers that were tagged for destruction. 4634 */ 4635static void 4636l2arc_do_free_on_write() 4637{ 4638 list_t *buflist; 4639 l2arc_data_free_t *df, *df_prev; 4640 4641 mutex_enter(&l2arc_free_on_write_mtx); 4642 buflist = l2arc_free_on_write; 4643 4644 for (df = list_tail(buflist); df; df = df_prev) { 4645 df_prev = list_prev(buflist, df); 4646 ASSERT(df->l2df_data != NULL); 4647 ASSERT(df->l2df_func != NULL); 4648 df->l2df_func(df->l2df_data, df->l2df_size); 4649 list_remove(buflist, df); 4650 kmem_free(df, sizeof (l2arc_data_free_t)); 4651 } 4652 4653 mutex_exit(&l2arc_free_on_write_mtx); 4654} 4655 4656/* 4657 * A write to a cache device has completed. Update all headers to allow 4658 * reads from these buffers to begin. 4659 */ 4660static void 4661l2arc_write_done(zio_t *zio) 4662{ 4663 l2arc_write_callback_t *cb; 4664 l2arc_dev_t *dev; 4665 list_t *buflist; 4666 arc_buf_hdr_t *head, *ab, *ab_prev; 4667 l2arc_buf_hdr_t *abl2; 4668 kmutex_t *hash_lock; 4669 int64_t bytes_dropped = 0; 4670 4671 cb = zio->io_private; 4672 ASSERT(cb != NULL); 4673 dev = cb->l2wcb_dev; 4674 ASSERT(dev != NULL); 4675 head = cb->l2wcb_head; 4676 ASSERT(head != NULL); 4677 buflist = dev->l2ad_buflist; 4678 ASSERT(buflist != NULL); 4679 DTRACE_PROBE2(l2arc__iodone, zio_t *, zio, 4680 l2arc_write_callback_t *, cb); 4681 4682 if (zio->io_error != 0) 4683 ARCSTAT_BUMP(arcstat_l2_writes_error); 4684 4685 mutex_enter(&l2arc_buflist_mtx); 4686 4687 /* 4688 * All writes completed, or an error was hit. 4689 */ 4690 for (ab = list_prev(buflist, head); ab; ab = ab_prev) { 4691 ab_prev = list_prev(buflist, ab); 4692 abl2 = ab->b_l2hdr; 4693 4694 /* 4695 * Release the temporary compressed buffer as soon as possible. 4696 */ 4697 if (abl2->b_compress != ZIO_COMPRESS_OFF) 4698 l2arc_release_cdata_buf(ab); 4699 4700 hash_lock = HDR_LOCK(ab); 4701 if (!mutex_tryenter(hash_lock)) { 4702 /* 4703 * This buffer misses out. It may be in a stage 4704 * of eviction. Its ARC_L2_WRITING flag will be 4705 * left set, denying reads to this buffer. 4706 */ 4707 ARCSTAT_BUMP(arcstat_l2_writes_hdr_miss); 4708 continue; 4709 } 4710 4711 if (zio->io_error != 0) { 4712 /* 4713 * Error - drop L2ARC entry. 4714 */ 4715 list_remove(buflist, ab); 4716 ARCSTAT_INCR(arcstat_l2_asize, -abl2->b_asize); 4717 bytes_dropped += abl2->b_asize; 4718 ab->b_l2hdr = NULL; 4719 trim_map_free(abl2->b_dev->l2ad_vdev, abl2->b_daddr, 4720 ab->b_size, 0); 4721 kmem_free(abl2, sizeof (l2arc_buf_hdr_t)); 4722 ARCSTAT_INCR(arcstat_l2_size, -ab->b_size); 4723 } 4724 4725 /* 4726 * Allow ARC to begin reads to this L2ARC entry. 4727 */ 4728 ab->b_flags &= ~ARC_L2_WRITING; 4729 4730 mutex_exit(hash_lock); 4731 } 4732 4733 atomic_inc_64(&l2arc_writes_done); 4734 list_remove(buflist, head); 4735 kmem_cache_free(hdr_cache, head); 4736 mutex_exit(&l2arc_buflist_mtx); 4737 4738 vdev_space_update(dev->l2ad_vdev, -bytes_dropped, 0, 0); 4739 4740 l2arc_do_free_on_write(); 4741 4742 kmem_free(cb, sizeof (l2arc_write_callback_t)); 4743} 4744 4745/* 4746 * A read to a cache device completed. Validate buffer contents before 4747 * handing over to the regular ARC routines. 4748 */ 4749static void 4750l2arc_read_done(zio_t *zio) 4751{ 4752 l2arc_read_callback_t *cb; 4753 arc_buf_hdr_t *hdr; 4754 arc_buf_t *buf; 4755 kmutex_t *hash_lock; 4756 int equal; 4757 4758 ASSERT(zio->io_vd != NULL); 4759 ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE); 4760 4761 spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd); 4762 4763 cb = zio->io_private; 4764 ASSERT(cb != NULL); 4765 buf = cb->l2rcb_buf; 4766 ASSERT(buf != NULL); 4767 4768 hash_lock = HDR_LOCK(buf->b_hdr); 4769 mutex_enter(hash_lock); 4770 hdr = buf->b_hdr; 4771 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 4772 4773 /* 4774 * If the buffer was compressed, decompress it first. 4775 */ 4776 if (cb->l2rcb_compress != ZIO_COMPRESS_OFF) 4777 l2arc_decompress_zio(zio, hdr, cb->l2rcb_compress); 4778 ASSERT(zio->io_data != NULL); 4779 4780 /* 4781 * Check this survived the L2ARC journey. 4782 */ 4783 equal = arc_cksum_equal(buf); 4784 if (equal && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) { 4785 mutex_exit(hash_lock); 4786 zio->io_private = buf; 4787 zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */ 4788 zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */ 4789 arc_read_done(zio); 4790 } else { 4791 mutex_exit(hash_lock); 4792 /* 4793 * Buffer didn't survive caching. Increment stats and 4794 * reissue to the original storage device. 4795 */ 4796 if (zio->io_error != 0) { 4797 ARCSTAT_BUMP(arcstat_l2_io_error); 4798 } else { 4799 zio->io_error = SET_ERROR(EIO); 4800 } 4801 if (!equal) 4802 ARCSTAT_BUMP(arcstat_l2_cksum_bad); 4803 4804 /* 4805 * If there's no waiter, issue an async i/o to the primary 4806 * storage now. If there *is* a waiter, the caller must 4807 * issue the i/o in a context where it's OK to block. 4808 */ 4809 if (zio->io_waiter == NULL) { 4810 zio_t *pio = zio_unique_parent(zio); 4811 4812 ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL); 4813 4814 zio_nowait(zio_read(pio, cb->l2rcb_spa, &cb->l2rcb_bp, 4815 buf->b_data, zio->io_size, arc_read_done, buf, 4816 zio->io_priority, cb->l2rcb_flags, &cb->l2rcb_zb)); 4817 } 4818 } 4819 4820 kmem_free(cb, sizeof (l2arc_read_callback_t)); 4821} 4822 4823/* 4824 * This is the list priority from which the L2ARC will search for pages to 4825 * cache. This is used within loops (0..3) to cycle through lists in the 4826 * desired order. This order can have a significant effect on cache 4827 * performance. 4828 * 4829 * Currently the metadata lists are hit first, MFU then MRU, followed by 4830 * the data lists. This function returns a locked list, and also returns 4831 * the lock pointer. 4832 */ 4833static list_t * 4834l2arc_list_locked(int list_num, kmutex_t **lock) 4835{ 4836 list_t *list = NULL; 4837 int idx; 4838 4839 ASSERT(list_num >= 0 && list_num < 2 * ARC_BUFC_NUMLISTS); 4840 4841 if (list_num < ARC_BUFC_NUMMETADATALISTS) { 4842 idx = list_num; 4843 list = &arc_mfu->arcs_lists[idx]; 4844 *lock = ARCS_LOCK(arc_mfu, idx); 4845 } else if (list_num < ARC_BUFC_NUMMETADATALISTS * 2) { 4846 idx = list_num - ARC_BUFC_NUMMETADATALISTS; 4847 list = &arc_mru->arcs_lists[idx]; 4848 *lock = ARCS_LOCK(arc_mru, idx); 4849 } else if (list_num < (ARC_BUFC_NUMMETADATALISTS * 2 + 4850 ARC_BUFC_NUMDATALISTS)) { 4851 idx = list_num - ARC_BUFC_NUMMETADATALISTS; 4852 list = &arc_mfu->arcs_lists[idx]; 4853 *lock = ARCS_LOCK(arc_mfu, idx); 4854 } else { 4855 idx = list_num - ARC_BUFC_NUMLISTS; 4856 list = &arc_mru->arcs_lists[idx]; 4857 *lock = ARCS_LOCK(arc_mru, idx); 4858 } 4859 4860 ASSERT(!(MUTEX_HELD(*lock))); 4861 mutex_enter(*lock); 4862 return (list); 4863} 4864 4865/* 4866 * Evict buffers from the device write hand to the distance specified in 4867 * bytes. This distance may span populated buffers, it may span nothing. 4868 * This is clearing a region on the L2ARC device ready for writing. 4869 * If the 'all' boolean is set, every buffer is evicted. 4870 */ 4871static void 4872l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all) 4873{ 4874 list_t *buflist; 4875 l2arc_buf_hdr_t *abl2; 4876 arc_buf_hdr_t *ab, *ab_prev; 4877 kmutex_t *hash_lock; 4878 uint64_t taddr; 4879 int64_t bytes_evicted = 0; 4880 4881 buflist = dev->l2ad_buflist; 4882 4883 if (buflist == NULL) 4884 return; 4885 4886 if (!all && dev->l2ad_first) { 4887 /* 4888 * This is the first sweep through the device. There is 4889 * nothing to evict. 4890 */ 4891 return; 4892 } 4893 4894 if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) { 4895 /* 4896 * When nearing the end of the device, evict to the end 4897 * before the device write hand jumps to the start. 4898 */ 4899 taddr = dev->l2ad_end; 4900 } else { 4901 taddr = dev->l2ad_hand + distance; 4902 } 4903 DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist, 4904 uint64_t, taddr, boolean_t, all); 4905 4906top: 4907 mutex_enter(&l2arc_buflist_mtx); 4908 for (ab = list_tail(buflist); ab; ab = ab_prev) { 4909 ab_prev = list_prev(buflist, ab); 4910 4911 hash_lock = HDR_LOCK(ab); 4912 if (!mutex_tryenter(hash_lock)) { 4913 /* 4914 * Missed the hash lock. Retry. 4915 */ 4916 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry); 4917 mutex_exit(&l2arc_buflist_mtx); 4918 mutex_enter(hash_lock); 4919 mutex_exit(hash_lock); 4920 goto top; 4921 } 4922 4923 if (HDR_L2_WRITE_HEAD(ab)) { 4924 /* 4925 * We hit a write head node. Leave it for 4926 * l2arc_write_done(). 4927 */ 4928 list_remove(buflist, ab); 4929 mutex_exit(hash_lock); 4930 continue; 4931 } 4932 4933 if (!all && ab->b_l2hdr != NULL && 4934 (ab->b_l2hdr->b_daddr > taddr || 4935 ab->b_l2hdr->b_daddr < dev->l2ad_hand)) { 4936 /* 4937 * We've evicted to the target address, 4938 * or the end of the device. 4939 */ 4940 mutex_exit(hash_lock); 4941 break; 4942 } 4943 4944 if (HDR_FREE_IN_PROGRESS(ab)) { 4945 /* 4946 * Already on the path to destruction. 4947 */ 4948 mutex_exit(hash_lock); 4949 continue; 4950 } 4951 4952 if (ab->b_state == arc_l2c_only) { 4953 ASSERT(!HDR_L2_READING(ab)); 4954 /* 4955 * This doesn't exist in the ARC. Destroy. 4956 * arc_hdr_destroy() will call list_remove() 4957 * and decrement arcstat_l2_size. 4958 */ 4959 arc_change_state(arc_anon, ab, hash_lock); 4960 arc_hdr_destroy(ab); 4961 } else { 4962 /* 4963 * Invalidate issued or about to be issued 4964 * reads, since we may be about to write 4965 * over this location. 4966 */ 4967 if (HDR_L2_READING(ab)) { 4968 ARCSTAT_BUMP(arcstat_l2_evict_reading); 4969 ab->b_flags |= ARC_L2_EVICTED; 4970 } 4971 4972 /* 4973 * Tell ARC this no longer exists in L2ARC. 4974 */ 4975 if (ab->b_l2hdr != NULL) { 4976 abl2 = ab->b_l2hdr; 4977 ARCSTAT_INCR(arcstat_l2_asize, -abl2->b_asize); 4978 bytes_evicted += abl2->b_asize; 4979 ab->b_l2hdr = NULL; 4980 kmem_free(abl2, sizeof (l2arc_buf_hdr_t)); 4981 ARCSTAT_INCR(arcstat_l2_size, -ab->b_size); 4982 } 4983 list_remove(buflist, ab); 4984 4985 /* 4986 * This may have been leftover after a 4987 * failed write. 4988 */ 4989 ab->b_flags &= ~ARC_L2_WRITING; 4990 } 4991 mutex_exit(hash_lock); 4992 } 4993 mutex_exit(&l2arc_buflist_mtx); 4994 4995 vdev_space_update(dev->l2ad_vdev, -bytes_evicted, 0, 0); 4996 dev->l2ad_evict = taddr; 4997} 4998 4999/* 5000 * Find and write ARC buffers to the L2ARC device. 5001 * 5002 * An ARC_L2_WRITING flag is set so that the L2ARC buffers are not valid 5003 * for reading until they have completed writing. 5004 * The headroom_boost is an in-out parameter used to maintain headroom boost 5005 * state between calls to this function. 5006 * 5007 * Returns the number of bytes actually written (which may be smaller than 5008 * the delta by which the device hand has changed due to alignment). 5009 */ 5010static uint64_t 5011l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz, 5012 boolean_t *headroom_boost) 5013{ 5014 arc_buf_hdr_t *ab, *ab_prev, *head; 5015 list_t *list; 5016 uint64_t write_asize, write_psize, write_sz, headroom, 5017 buf_compress_minsz; 5018 void *buf_data; 5019 kmutex_t *list_lock; 5020 boolean_t full; 5021 l2arc_write_callback_t *cb; 5022 zio_t *pio, *wzio; 5023 uint64_t guid = spa_load_guid(spa); 5024 const boolean_t do_headroom_boost = *headroom_boost; 5025 int try; 5026 5027 ASSERT(dev->l2ad_vdev != NULL); 5028 5029 /* Lower the flag now, we might want to raise it again later. */ 5030 *headroom_boost = B_FALSE; 5031 5032 pio = NULL; 5033 write_sz = write_asize = write_psize = 0; 5034 full = B_FALSE; 5035 head = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE); 5036 head->b_flags |= ARC_L2_WRITE_HEAD; 5037 5038 ARCSTAT_BUMP(arcstat_l2_write_buffer_iter); 5039 /* 5040 * We will want to try to compress buffers that are at least 2x the 5041 * device sector size. 5042 */ 5043 buf_compress_minsz = 2 << dev->l2ad_vdev->vdev_ashift; 5044 5045 /* 5046 * Copy buffers for L2ARC writing. 5047 */ 5048 mutex_enter(&l2arc_buflist_mtx); 5049 for (try = 0; try < 2 * ARC_BUFC_NUMLISTS; try++) { 5050 uint64_t passed_sz = 0; 5051 5052 list = l2arc_list_locked(try, &list_lock); 5053 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_iter); 5054 5055 /* 5056 * L2ARC fast warmup. 5057 * 5058 * Until the ARC is warm and starts to evict, read from the 5059 * head of the ARC lists rather than the tail. 5060 */ 5061 if (arc_warm == B_FALSE) 5062 ab = list_head(list); 5063 else 5064 ab = list_tail(list); 5065 if (ab == NULL) 5066 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_null_iter); 5067 5068 headroom = target_sz * l2arc_headroom; 5069 if (do_headroom_boost) 5070 headroom = (headroom * l2arc_headroom_boost) / 100; 5071 5072 for (; ab; ab = ab_prev) { 5073 l2arc_buf_hdr_t *l2hdr; 5074 kmutex_t *hash_lock; 5075 uint64_t buf_sz; 5076 5077 if (arc_warm == B_FALSE) 5078 ab_prev = list_next(list, ab); 5079 else 5080 ab_prev = list_prev(list, ab); 5081 ARCSTAT_INCR(arcstat_l2_write_buffer_bytes_scanned, ab->b_size); 5082 5083 hash_lock = HDR_LOCK(ab); 5084 if (!mutex_tryenter(hash_lock)) { 5085 ARCSTAT_BUMP(arcstat_l2_write_trylock_fail); 5086 /* 5087 * Skip this buffer rather than waiting. 5088 */ 5089 continue; 5090 } 5091 5092 passed_sz += ab->b_size; 5093 if (passed_sz > headroom) { 5094 /* 5095 * Searched too far. 5096 */ 5097 mutex_exit(hash_lock); 5098 ARCSTAT_BUMP(arcstat_l2_write_passed_headroom); 5099 break; 5100 } 5101 5102 if (!l2arc_write_eligible(guid, ab)) { 5103 mutex_exit(hash_lock); 5104 continue; 5105 } 5106 5107 if ((write_sz + ab->b_size) > target_sz) { 5108 full = B_TRUE; 5109 mutex_exit(hash_lock); 5110 ARCSTAT_BUMP(arcstat_l2_write_full); 5111 break; 5112 } 5113 5114 if (pio == NULL) { 5115 /* 5116 * Insert a dummy header on the buflist so 5117 * l2arc_write_done() can find where the 5118 * write buffers begin without searching. 5119 */ 5120 list_insert_head(dev->l2ad_buflist, head); 5121 5122 cb = kmem_alloc( 5123 sizeof (l2arc_write_callback_t), KM_SLEEP); 5124 cb->l2wcb_dev = dev; 5125 cb->l2wcb_head = head; 5126 pio = zio_root(spa, l2arc_write_done, cb, 5127 ZIO_FLAG_CANFAIL); 5128 ARCSTAT_BUMP(arcstat_l2_write_pios); 5129 } 5130 5131 /* 5132 * Create and add a new L2ARC header. 5133 */ 5134 l2hdr = kmem_zalloc(sizeof (l2arc_buf_hdr_t), KM_SLEEP); 5135 l2hdr->b_dev = dev; 5136 ab->b_flags |= ARC_L2_WRITING; 5137 5138 /* 5139 * Temporarily stash the data buffer in b_tmp_cdata. 5140 * The subsequent write step will pick it up from 5141 * there. This is because can't access ab->b_buf 5142 * without holding the hash_lock, which we in turn 5143 * can't access without holding the ARC list locks 5144 * (which we want to avoid during compression/writing). 5145 */ 5146 l2hdr->b_compress = ZIO_COMPRESS_OFF; 5147 l2hdr->b_asize = ab->b_size; 5148 l2hdr->b_tmp_cdata = ab->b_buf->b_data; 5149 5150 buf_sz = ab->b_size; 5151 ab->b_l2hdr = l2hdr; 5152 5153 list_insert_head(dev->l2ad_buflist, ab); 5154 5155 /* 5156 * Compute and store the buffer cksum before 5157 * writing. On debug the cksum is verified first. 5158 */ 5159 arc_cksum_verify(ab->b_buf); 5160 arc_cksum_compute(ab->b_buf, B_TRUE); 5161 5162 mutex_exit(hash_lock); 5163 5164 write_sz += buf_sz; 5165 } 5166 5167 mutex_exit(list_lock); 5168 5169 if (full == B_TRUE) 5170 break; 5171 } 5172 5173 /* No buffers selected for writing? */ 5174 if (pio == NULL) { 5175 ASSERT0(write_sz); 5176 mutex_exit(&l2arc_buflist_mtx); 5177 kmem_cache_free(hdr_cache, head); 5178 return (0); 5179 } 5180 5181 /* 5182 * Now start writing the buffers. We're starting at the write head 5183 * and work backwards, retracing the course of the buffer selector 5184 * loop above. 5185 */ 5186 for (ab = list_prev(dev->l2ad_buflist, head); ab; 5187 ab = list_prev(dev->l2ad_buflist, ab)) { 5188 l2arc_buf_hdr_t *l2hdr; 5189 uint64_t buf_sz; 5190 5191 /* 5192 * We shouldn't need to lock the buffer here, since we flagged 5193 * it as ARC_L2_WRITING in the previous step, but we must take 5194 * care to only access its L2 cache parameters. In particular, 5195 * ab->b_buf may be invalid by now due to ARC eviction. 5196 */ 5197 l2hdr = ab->b_l2hdr; 5198 l2hdr->b_daddr = dev->l2ad_hand; 5199 5200 if ((ab->b_flags & ARC_L2COMPRESS) && 5201 l2hdr->b_asize >= buf_compress_minsz) { 5202 if (l2arc_compress_buf(l2hdr)) { 5203 /* 5204 * If compression succeeded, enable headroom 5205 * boost on the next scan cycle. 5206 */ 5207 *headroom_boost = B_TRUE; 5208 } 5209 } 5210 5211 /* 5212 * Pick up the buffer data we had previously stashed away 5213 * (and now potentially also compressed). 5214 */ 5215 buf_data = l2hdr->b_tmp_cdata; 5216 buf_sz = l2hdr->b_asize; 5217 5218 /* Compression may have squashed the buffer to zero length. */ 5219 if (buf_sz != 0) { 5220 uint64_t buf_p_sz; 5221 5222 wzio = zio_write_phys(pio, dev->l2ad_vdev, 5223 dev->l2ad_hand, buf_sz, buf_data, ZIO_CHECKSUM_OFF, 5224 NULL, NULL, ZIO_PRIORITY_ASYNC_WRITE, 5225 ZIO_FLAG_CANFAIL, B_FALSE); 5226 5227 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev, 5228 zio_t *, wzio); 5229 (void) zio_nowait(wzio); 5230 5231 write_asize += buf_sz; 5232 /* 5233 * Keep the clock hand suitably device-aligned. 5234 */ 5235 buf_p_sz = vdev_psize_to_asize(dev->l2ad_vdev, buf_sz); 5236 write_psize += buf_p_sz; 5237 dev->l2ad_hand += buf_p_sz; 5238 } 5239 } 5240 5241 mutex_exit(&l2arc_buflist_mtx); 5242 5243 ASSERT3U(write_asize, <=, target_sz); 5244 ARCSTAT_BUMP(arcstat_l2_writes_sent); 5245 ARCSTAT_INCR(arcstat_l2_write_bytes, write_asize); 5246 ARCSTAT_INCR(arcstat_l2_size, write_sz); 5247 ARCSTAT_INCR(arcstat_l2_asize, write_asize); 5248 vdev_space_update(dev->l2ad_vdev, write_asize, 0, 0); 5249 5250 /* 5251 * Bump device hand to the device start if it is approaching the end. 5252 * l2arc_evict() will already have evicted ahead for this case. 5253 */ 5254 if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) { 5255 dev->l2ad_hand = dev->l2ad_start; 5256 dev->l2ad_evict = dev->l2ad_start; 5257 dev->l2ad_first = B_FALSE; 5258 } 5259 5260 dev->l2ad_writing = B_TRUE; 5261 (void) zio_wait(pio); 5262 dev->l2ad_writing = B_FALSE; 5263 5264 return (write_asize); 5265} 5266 5267/* 5268 * Compresses an L2ARC buffer. 5269 * The data to be compressed must be prefilled in l2hdr->b_tmp_cdata and its 5270 * size in l2hdr->b_asize. This routine tries to compress the data and 5271 * depending on the compression result there are three possible outcomes: 5272 * *) The buffer was incompressible. The original l2hdr contents were left 5273 * untouched and are ready for writing to an L2 device. 5274 * *) The buffer was all-zeros, so there is no need to write it to an L2 5275 * device. To indicate this situation b_tmp_cdata is NULL'ed, b_asize is 5276 * set to zero and b_compress is set to ZIO_COMPRESS_EMPTY. 5277 * *) Compression succeeded and b_tmp_cdata was replaced with a temporary 5278 * data buffer which holds the compressed data to be written, and b_asize 5279 * tells us how much data there is. b_compress is set to the appropriate 5280 * compression algorithm. Once writing is done, invoke 5281 * l2arc_release_cdata_buf on this l2hdr to free this temporary buffer. 5282 * 5283 * Returns B_TRUE if compression succeeded, or B_FALSE if it didn't (the 5284 * buffer was incompressible). 5285 */ 5286static boolean_t 5287l2arc_compress_buf(l2arc_buf_hdr_t *l2hdr) 5288{ 5289 void *cdata; 5290 size_t csize, len, rounded; 5291 5292 ASSERT(l2hdr->b_compress == ZIO_COMPRESS_OFF); 5293 ASSERT(l2hdr->b_tmp_cdata != NULL); 5294 5295 len = l2hdr->b_asize; 5296 cdata = zio_data_buf_alloc(len); 5297 csize = zio_compress_data(ZIO_COMPRESS_LZ4, l2hdr->b_tmp_cdata, 5298 cdata, l2hdr->b_asize); 5299 5300 rounded = P2ROUNDUP(csize, (size_t)SPA_MINBLOCKSIZE); 5301 if (rounded > csize) { 5302 bzero((char *)cdata + csize, rounded - csize); 5303 csize = rounded; 5304 } 5305 5306 if (csize == 0) { 5307 /* zero block, indicate that there's nothing to write */ 5308 zio_data_buf_free(cdata, len); 5309 l2hdr->b_compress = ZIO_COMPRESS_EMPTY; 5310 l2hdr->b_asize = 0; 5311 l2hdr->b_tmp_cdata = NULL; 5312 ARCSTAT_BUMP(arcstat_l2_compress_zeros); 5313 return (B_TRUE); 5314 } else if (csize > 0 && csize < len) { 5315 /* 5316 * Compression succeeded, we'll keep the cdata around for 5317 * writing and release it afterwards. 5318 */ 5319 l2hdr->b_compress = ZIO_COMPRESS_LZ4; 5320 l2hdr->b_asize = csize; 5321 l2hdr->b_tmp_cdata = cdata; 5322 ARCSTAT_BUMP(arcstat_l2_compress_successes); 5323 return (B_TRUE); 5324 } else { 5325 /* 5326 * Compression failed, release the compressed buffer. 5327 * l2hdr will be left unmodified. 5328 */ 5329 zio_data_buf_free(cdata, len); 5330 ARCSTAT_BUMP(arcstat_l2_compress_failures); 5331 return (B_FALSE); 5332 } 5333} 5334 5335/* 5336 * Decompresses a zio read back from an l2arc device. On success, the 5337 * underlying zio's io_data buffer is overwritten by the uncompressed 5338 * version. On decompression error (corrupt compressed stream), the 5339 * zio->io_error value is set to signal an I/O error. 5340 * 5341 * Please note that the compressed data stream is not checksummed, so 5342 * if the underlying device is experiencing data corruption, we may feed 5343 * corrupt data to the decompressor, so the decompressor needs to be 5344 * able to handle this situation (LZ4 does). 5345 */ 5346static void 5347l2arc_decompress_zio(zio_t *zio, arc_buf_hdr_t *hdr, enum zio_compress c) 5348{ 5349 ASSERT(L2ARC_IS_VALID_COMPRESS(c)); 5350 5351 if (zio->io_error != 0) { 5352 /* 5353 * An io error has occured, just restore the original io 5354 * size in preparation for a main pool read. 5355 */ 5356 zio->io_orig_size = zio->io_size = hdr->b_size; 5357 return; 5358 } 5359 5360 if (c == ZIO_COMPRESS_EMPTY) { 5361 /* 5362 * An empty buffer results in a null zio, which means we 5363 * need to fill its io_data after we're done restoring the 5364 * buffer's contents. 5365 */ 5366 ASSERT(hdr->b_buf != NULL); 5367 bzero(hdr->b_buf->b_data, hdr->b_size); 5368 zio->io_data = zio->io_orig_data = hdr->b_buf->b_data; 5369 } else { 5370 ASSERT(zio->io_data != NULL); 5371 /* 5372 * We copy the compressed data from the start of the arc buffer 5373 * (the zio_read will have pulled in only what we need, the 5374 * rest is garbage which we will overwrite at decompression) 5375 * and then decompress back to the ARC data buffer. This way we 5376 * can minimize copying by simply decompressing back over the 5377 * original compressed data (rather than decompressing to an 5378 * aux buffer and then copying back the uncompressed buffer, 5379 * which is likely to be much larger). 5380 */ 5381 uint64_t csize; 5382 void *cdata; 5383 5384 csize = zio->io_size; 5385 cdata = zio_data_buf_alloc(csize); 5386 bcopy(zio->io_data, cdata, csize); 5387 if (zio_decompress_data(c, cdata, zio->io_data, csize, 5388 hdr->b_size) != 0) 5389 zio->io_error = EIO; 5390 zio_data_buf_free(cdata, csize); 5391 } 5392 5393 /* Restore the expected uncompressed IO size. */ 5394 zio->io_orig_size = zio->io_size = hdr->b_size; 5395} 5396 5397/* 5398 * Releases the temporary b_tmp_cdata buffer in an l2arc header structure. 5399 * This buffer serves as a temporary holder of compressed data while 5400 * the buffer entry is being written to an l2arc device. Once that is 5401 * done, we can dispose of it. 5402 */ 5403static void 5404l2arc_release_cdata_buf(arc_buf_hdr_t *ab) 5405{ 5406 l2arc_buf_hdr_t *l2hdr = ab->b_l2hdr; 5407 5408 if (l2hdr->b_compress == ZIO_COMPRESS_LZ4) { 5409 /* 5410 * If the data was compressed, then we've allocated a 5411 * temporary buffer for it, so now we need to release it. 5412 */ 5413 ASSERT(l2hdr->b_tmp_cdata != NULL); 5414 zio_data_buf_free(l2hdr->b_tmp_cdata, ab->b_size); 5415 } 5416 l2hdr->b_tmp_cdata = NULL; 5417} 5418 5419/* 5420 * This thread feeds the L2ARC at regular intervals. This is the beating 5421 * heart of the L2ARC. 5422 */ 5423static void 5424l2arc_feed_thread(void *dummy __unused) 5425{ 5426 callb_cpr_t cpr; 5427 l2arc_dev_t *dev; 5428 spa_t *spa; 5429 uint64_t size, wrote; 5430 clock_t begin, next = ddi_get_lbolt(); 5431 boolean_t headroom_boost = B_FALSE; 5432 5433 CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG); 5434 5435 mutex_enter(&l2arc_feed_thr_lock); 5436 5437 while (l2arc_thread_exit == 0) { 5438 CALLB_CPR_SAFE_BEGIN(&cpr); 5439 (void) cv_timedwait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock, 5440 next - ddi_get_lbolt()); 5441 CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock); 5442 next = ddi_get_lbolt() + hz; 5443 5444 /* 5445 * Quick check for L2ARC devices. 5446 */ 5447 mutex_enter(&l2arc_dev_mtx); 5448 if (l2arc_ndev == 0) { 5449 mutex_exit(&l2arc_dev_mtx); 5450 continue; 5451 } 5452 mutex_exit(&l2arc_dev_mtx); 5453 begin = ddi_get_lbolt(); 5454 5455 /* 5456 * This selects the next l2arc device to write to, and in 5457 * doing so the next spa to feed from: dev->l2ad_spa. This 5458 * will return NULL if there are now no l2arc devices or if 5459 * they are all faulted. 5460 * 5461 * If a device is returned, its spa's config lock is also 5462 * held to prevent device removal. l2arc_dev_get_next() 5463 * will grab and release l2arc_dev_mtx. 5464 */ 5465 if ((dev = l2arc_dev_get_next()) == NULL) 5466 continue; 5467 5468 spa = dev->l2ad_spa; 5469 ASSERT(spa != NULL); 5470 5471 /* 5472 * If the pool is read-only then force the feed thread to 5473 * sleep a little longer. 5474 */ 5475 if (!spa_writeable(spa)) { 5476 next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz; 5477 spa_config_exit(spa, SCL_L2ARC, dev); 5478 continue; 5479 } 5480 5481 /* 5482 * Avoid contributing to memory pressure. 5483 */ 5484 if (arc_reclaim_needed()) { 5485 ARCSTAT_BUMP(arcstat_l2_abort_lowmem); 5486 spa_config_exit(spa, SCL_L2ARC, dev); 5487 continue; 5488 } 5489 5490 ARCSTAT_BUMP(arcstat_l2_feeds); 5491 5492 size = l2arc_write_size(); 5493 5494 /* 5495 * Evict L2ARC buffers that will be overwritten. 5496 */ 5497 l2arc_evict(dev, size, B_FALSE); 5498 5499 /* 5500 * Write ARC buffers. 5501 */ 5502 wrote = l2arc_write_buffers(spa, dev, size, &headroom_boost); 5503 5504 /* 5505 * Calculate interval between writes. 5506 */ 5507 next = l2arc_write_interval(begin, size, wrote); 5508 spa_config_exit(spa, SCL_L2ARC, dev); 5509 } 5510 5511 l2arc_thread_exit = 0; 5512 cv_broadcast(&l2arc_feed_thr_cv); 5513 CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */ 5514 thread_exit(); 5515} 5516 5517boolean_t 5518l2arc_vdev_present(vdev_t *vd) 5519{ 5520 l2arc_dev_t *dev; 5521 5522 mutex_enter(&l2arc_dev_mtx); 5523 for (dev = list_head(l2arc_dev_list); dev != NULL; 5524 dev = list_next(l2arc_dev_list, dev)) { 5525 if (dev->l2ad_vdev == vd) 5526 break; 5527 } 5528 mutex_exit(&l2arc_dev_mtx); 5529 5530 return (dev != NULL); 5531} 5532 5533/* 5534 * Add a vdev for use by the L2ARC. By this point the spa has already 5535 * validated the vdev and opened it. 5536 */ 5537void 5538l2arc_add_vdev(spa_t *spa, vdev_t *vd) 5539{ 5540 l2arc_dev_t *adddev; 5541 5542 ASSERT(!l2arc_vdev_present(vd)); 5543 5544 vdev_ashift_optimize(vd); 5545 5546 /* 5547 * Create a new l2arc device entry. 5548 */ 5549 adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP); 5550 adddev->l2ad_spa = spa; 5551 adddev->l2ad_vdev = vd; 5552 adddev->l2ad_start = VDEV_LABEL_START_SIZE; 5553 adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd); 5554 adddev->l2ad_hand = adddev->l2ad_start; 5555 adddev->l2ad_evict = adddev->l2ad_start; 5556 adddev->l2ad_first = B_TRUE; 5557 adddev->l2ad_writing = B_FALSE; 5558 5559 /* 5560 * This is a list of all ARC buffers that are still valid on the 5561 * device. 5562 */ 5563 adddev->l2ad_buflist = kmem_zalloc(sizeof (list_t), KM_SLEEP); 5564 list_create(adddev->l2ad_buflist, sizeof (arc_buf_hdr_t), 5565 offsetof(arc_buf_hdr_t, b_l2node)); 5566 5567 vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand); 5568 5569 /* 5570 * Add device to global list 5571 */ 5572 mutex_enter(&l2arc_dev_mtx); 5573 list_insert_head(l2arc_dev_list, adddev); 5574 atomic_inc_64(&l2arc_ndev); 5575 mutex_exit(&l2arc_dev_mtx); 5576} 5577 5578/* 5579 * Remove a vdev from the L2ARC. 5580 */ 5581void 5582l2arc_remove_vdev(vdev_t *vd) 5583{ 5584 l2arc_dev_t *dev, *nextdev, *remdev = NULL; 5585 5586 /* 5587 * Find the device by vdev 5588 */ 5589 mutex_enter(&l2arc_dev_mtx); 5590 for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) { 5591 nextdev = list_next(l2arc_dev_list, dev); 5592 if (vd == dev->l2ad_vdev) { 5593 remdev = dev; 5594 break; 5595 } 5596 } 5597 ASSERT(remdev != NULL); 5598 5599 /* 5600 * Remove device from global list 5601 */ 5602 list_remove(l2arc_dev_list, remdev); 5603 l2arc_dev_last = NULL; /* may have been invalidated */ 5604 atomic_dec_64(&l2arc_ndev); 5605 mutex_exit(&l2arc_dev_mtx); 5606 5607 /* 5608 * Clear all buflists and ARC references. L2ARC device flush. 5609 */ 5610 l2arc_evict(remdev, 0, B_TRUE); 5611 list_destroy(remdev->l2ad_buflist); 5612 kmem_free(remdev->l2ad_buflist, sizeof (list_t)); 5613 kmem_free(remdev, sizeof (l2arc_dev_t)); 5614} 5615 5616void 5617l2arc_init(void) 5618{ 5619 l2arc_thread_exit = 0; 5620 l2arc_ndev = 0; 5621 l2arc_writes_sent = 0; 5622 l2arc_writes_done = 0; 5623 5624 mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL); 5625 cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL); 5626 mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL); 5627 mutex_init(&l2arc_buflist_mtx, NULL, MUTEX_DEFAULT, NULL); 5628 mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL); 5629 5630 l2arc_dev_list = &L2ARC_dev_list; 5631 l2arc_free_on_write = &L2ARC_free_on_write; 5632 list_create(l2arc_dev_list, sizeof (l2arc_dev_t), 5633 offsetof(l2arc_dev_t, l2ad_node)); 5634 list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t), 5635 offsetof(l2arc_data_free_t, l2df_list_node)); 5636} 5637 5638void 5639l2arc_fini(void) 5640{ 5641 /* 5642 * This is called from dmu_fini(), which is called from spa_fini(); 5643 * Because of this, we can assume that all l2arc devices have 5644 * already been removed when the pools themselves were removed. 5645 */ 5646 5647 l2arc_do_free_on_write(); 5648 5649 mutex_destroy(&l2arc_feed_thr_lock); 5650 cv_destroy(&l2arc_feed_thr_cv); 5651 mutex_destroy(&l2arc_dev_mtx); 5652 mutex_destroy(&l2arc_buflist_mtx); 5653 mutex_destroy(&l2arc_free_on_write_mtx); 5654 5655 list_destroy(l2arc_dev_list); 5656 list_destroy(l2arc_free_on_write); 5657} 5658 5659void 5660l2arc_start(void) 5661{ 5662 if (!(spa_mode_global & FWRITE)) 5663 return; 5664 5665 (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0, 5666 TS_RUN, minclsyspri); 5667} 5668 5669void 5670l2arc_stop(void) 5671{ 5672 if (!(spa_mode_global & FWRITE)) 5673 return; 5674 5675 mutex_enter(&l2arc_feed_thr_lock); 5676 cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */ 5677 l2arc_thread_exit = 1; 5678 while (l2arc_thread_exit != 0) 5679 cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock); 5680 mutex_exit(&l2arc_feed_thr_lock); 5681} 5682