arc.c revision 275492
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); 223TUNABLE_INT("vfs.zfs.arc_shrink_shift", &zfs_arc_shrink_shift); 224SYSCTL_DECL(_vfs_zfs); 225SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_max, CTLFLAG_RDTUN, &zfs_arc_max, 0, 226 "Maximum ARC size"); 227SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_min, CTLFLAG_RDTUN, &zfs_arc_min, 0, 228 "Minimum ARC size"); 229SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_average_blocksize, CTLFLAG_RDTUN, 230 &zfs_arc_average_blocksize, 0, 231 "ARC average blocksize"); 232SYSCTL_INT(_vfs_zfs, OID_AUTO, arc_shrink_shift, CTLFLAG_RW, 233 &arc_shrink_shift, 0, 234 "log2(fraction of arc to reclaim)"); 235 236/* 237 * We don't have a tunable for arc_free_target due to the dependency on 238 * pagedaemon initialisation. 239 */ 240SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_free_target, 241 CTLTYPE_UINT | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, sizeof(u_int), 242 sysctl_vfs_zfs_arc_free_target, "IU", 243 "Desired number of free pages below which ARC triggers reclaim"); 244 245static int 246sysctl_vfs_zfs_arc_free_target(SYSCTL_HANDLER_ARGS) 247{ 248 u_int val; 249 int err; 250 251 val = zfs_arc_free_target; 252 err = sysctl_handle_int(oidp, &val, 0, req); 253 if (err != 0 || req->newptr == NULL) 254 return (err); 255 256 if (val < minfree) 257 return (EINVAL); 258 if (val > cnt.v_page_count) 259 return (EINVAL); 260 261 zfs_arc_free_target = val; 262 263 return (0); 264} 265#endif 266 267/* 268 * Note that buffers can be in one of 6 states: 269 * ARC_anon - anonymous (discussed below) 270 * ARC_mru - recently used, currently cached 271 * ARC_mru_ghost - recentely used, no longer in cache 272 * ARC_mfu - frequently used, currently cached 273 * ARC_mfu_ghost - frequently used, no longer in cache 274 * ARC_l2c_only - exists in L2ARC but not other states 275 * When there are no active references to the buffer, they are 276 * are linked onto a list in one of these arc states. These are 277 * the only buffers that can be evicted or deleted. Within each 278 * state there are multiple lists, one for meta-data and one for 279 * non-meta-data. Meta-data (indirect blocks, blocks of dnodes, 280 * etc.) is tracked separately so that it can be managed more 281 * explicitly: favored over data, limited explicitly. 282 * 283 * Anonymous buffers are buffers that are not associated with 284 * a DVA. These are buffers that hold dirty block copies 285 * before they are written to stable storage. By definition, 286 * they are "ref'd" and are considered part of arc_mru 287 * that cannot be freed. Generally, they will aquire a DVA 288 * as they are written and migrate onto the arc_mru list. 289 * 290 * The ARC_l2c_only state is for buffers that are in the second 291 * level ARC but no longer in any of the ARC_m* lists. The second 292 * level ARC itself may also contain buffers that are in any of 293 * the ARC_m* states - meaning that a buffer can exist in two 294 * places. The reason for the ARC_l2c_only state is to keep the 295 * buffer header in the hash table, so that reads that hit the 296 * second level ARC benefit from these fast lookups. 297 */ 298 299#define ARCS_LOCK_PAD CACHE_LINE_SIZE 300struct arcs_lock { 301 kmutex_t arcs_lock; 302#ifdef _KERNEL 303 unsigned char pad[(ARCS_LOCK_PAD - sizeof (kmutex_t))]; 304#endif 305}; 306 307/* 308 * must be power of two for mask use to work 309 * 310 */ 311#define ARC_BUFC_NUMDATALISTS 16 312#define ARC_BUFC_NUMMETADATALISTS 16 313#define ARC_BUFC_NUMLISTS (ARC_BUFC_NUMMETADATALISTS + ARC_BUFC_NUMDATALISTS) 314 315typedef struct arc_state { 316 uint64_t arcs_lsize[ARC_BUFC_NUMTYPES]; /* amount of evictable data */ 317 uint64_t arcs_size; /* total amount of data in this state */ 318 list_t arcs_lists[ARC_BUFC_NUMLISTS]; /* list of evictable buffers */ 319 struct arcs_lock arcs_locks[ARC_BUFC_NUMLISTS] __aligned(CACHE_LINE_SIZE); 320} arc_state_t; 321 322#define ARCS_LOCK(s, i) (&((s)->arcs_locks[(i)].arcs_lock)) 323 324/* The 6 states: */ 325static arc_state_t ARC_anon; 326static arc_state_t ARC_mru; 327static arc_state_t ARC_mru_ghost; 328static arc_state_t ARC_mfu; 329static arc_state_t ARC_mfu_ghost; 330static arc_state_t ARC_l2c_only; 331 332typedef struct arc_stats { 333 kstat_named_t arcstat_hits; 334 kstat_named_t arcstat_misses; 335 kstat_named_t arcstat_demand_data_hits; 336 kstat_named_t arcstat_demand_data_misses; 337 kstat_named_t arcstat_demand_metadata_hits; 338 kstat_named_t arcstat_demand_metadata_misses; 339 kstat_named_t arcstat_prefetch_data_hits; 340 kstat_named_t arcstat_prefetch_data_misses; 341 kstat_named_t arcstat_prefetch_metadata_hits; 342 kstat_named_t arcstat_prefetch_metadata_misses; 343 kstat_named_t arcstat_mru_hits; 344 kstat_named_t arcstat_mru_ghost_hits; 345 kstat_named_t arcstat_mfu_hits; 346 kstat_named_t arcstat_mfu_ghost_hits; 347 kstat_named_t arcstat_allocated; 348 kstat_named_t arcstat_deleted; 349 kstat_named_t arcstat_stolen; 350 kstat_named_t arcstat_recycle_miss; 351 /* 352 * Number of buffers that could not be evicted because the hash lock 353 * was held by another thread. The lock may not necessarily be held 354 * by something using the same buffer, since hash locks are shared 355 * by multiple buffers. 356 */ 357 kstat_named_t arcstat_mutex_miss; 358 /* 359 * Number of buffers skipped because they have I/O in progress, are 360 * indrect prefetch buffers that have not lived long enough, or are 361 * not from the spa we're trying to evict from. 362 */ 363 kstat_named_t arcstat_evict_skip; 364 kstat_named_t arcstat_evict_l2_cached; 365 kstat_named_t arcstat_evict_l2_eligible; 366 kstat_named_t arcstat_evict_l2_ineligible; 367 kstat_named_t arcstat_hash_elements; 368 kstat_named_t arcstat_hash_elements_max; 369 kstat_named_t arcstat_hash_collisions; 370 kstat_named_t arcstat_hash_chains; 371 kstat_named_t arcstat_hash_chain_max; 372 kstat_named_t arcstat_p; 373 kstat_named_t arcstat_c; 374 kstat_named_t arcstat_c_min; 375 kstat_named_t arcstat_c_max; 376 kstat_named_t arcstat_size; 377 kstat_named_t arcstat_hdr_size; 378 kstat_named_t arcstat_data_size; 379 kstat_named_t arcstat_other_size; 380 kstat_named_t arcstat_l2_hits; 381 kstat_named_t arcstat_l2_misses; 382 kstat_named_t arcstat_l2_feeds; 383 kstat_named_t arcstat_l2_rw_clash; 384 kstat_named_t arcstat_l2_read_bytes; 385 kstat_named_t arcstat_l2_write_bytes; 386 kstat_named_t arcstat_l2_writes_sent; 387 kstat_named_t arcstat_l2_writes_done; 388 kstat_named_t arcstat_l2_writes_error; 389 kstat_named_t arcstat_l2_writes_hdr_miss; 390 kstat_named_t arcstat_l2_evict_lock_retry; 391 kstat_named_t arcstat_l2_evict_reading; 392 kstat_named_t arcstat_l2_free_on_write; 393 kstat_named_t arcstat_l2_cdata_free_on_write; 394 kstat_named_t arcstat_l2_abort_lowmem; 395 kstat_named_t arcstat_l2_cksum_bad; 396 kstat_named_t arcstat_l2_io_error; 397 kstat_named_t arcstat_l2_size; 398 kstat_named_t arcstat_l2_asize; 399 kstat_named_t arcstat_l2_hdr_size; 400 kstat_named_t arcstat_l2_compress_successes; 401 kstat_named_t arcstat_l2_compress_zeros; 402 kstat_named_t arcstat_l2_compress_failures; 403 kstat_named_t arcstat_l2_write_trylock_fail; 404 kstat_named_t arcstat_l2_write_passed_headroom; 405 kstat_named_t arcstat_l2_write_spa_mismatch; 406 kstat_named_t arcstat_l2_write_in_l2; 407 kstat_named_t arcstat_l2_write_hdr_io_in_progress; 408 kstat_named_t arcstat_l2_write_not_cacheable; 409 kstat_named_t arcstat_l2_write_full; 410 kstat_named_t arcstat_l2_write_buffer_iter; 411 kstat_named_t arcstat_l2_write_pios; 412 kstat_named_t arcstat_l2_write_buffer_bytes_scanned; 413 kstat_named_t arcstat_l2_write_buffer_list_iter; 414 kstat_named_t arcstat_l2_write_buffer_list_null_iter; 415 kstat_named_t arcstat_memory_throttle_count; 416 kstat_named_t arcstat_duplicate_buffers; 417 kstat_named_t arcstat_duplicate_buffers_size; 418 kstat_named_t arcstat_duplicate_reads; 419} arc_stats_t; 420 421static arc_stats_t arc_stats = { 422 { "hits", KSTAT_DATA_UINT64 }, 423 { "misses", KSTAT_DATA_UINT64 }, 424 { "demand_data_hits", KSTAT_DATA_UINT64 }, 425 { "demand_data_misses", KSTAT_DATA_UINT64 }, 426 { "demand_metadata_hits", KSTAT_DATA_UINT64 }, 427 { "demand_metadata_misses", KSTAT_DATA_UINT64 }, 428 { "prefetch_data_hits", KSTAT_DATA_UINT64 }, 429 { "prefetch_data_misses", KSTAT_DATA_UINT64 }, 430 { "prefetch_metadata_hits", KSTAT_DATA_UINT64 }, 431 { "prefetch_metadata_misses", KSTAT_DATA_UINT64 }, 432 { "mru_hits", KSTAT_DATA_UINT64 }, 433 { "mru_ghost_hits", KSTAT_DATA_UINT64 }, 434 { "mfu_hits", KSTAT_DATA_UINT64 }, 435 { "mfu_ghost_hits", KSTAT_DATA_UINT64 }, 436 { "allocated", KSTAT_DATA_UINT64 }, 437 { "deleted", KSTAT_DATA_UINT64 }, 438 { "stolen", KSTAT_DATA_UINT64 }, 439 { "recycle_miss", KSTAT_DATA_UINT64 }, 440 { "mutex_miss", KSTAT_DATA_UINT64 }, 441 { "evict_skip", KSTAT_DATA_UINT64 }, 442 { "evict_l2_cached", KSTAT_DATA_UINT64 }, 443 { "evict_l2_eligible", KSTAT_DATA_UINT64 }, 444 { "evict_l2_ineligible", KSTAT_DATA_UINT64 }, 445 { "hash_elements", KSTAT_DATA_UINT64 }, 446 { "hash_elements_max", KSTAT_DATA_UINT64 }, 447 { "hash_collisions", KSTAT_DATA_UINT64 }, 448 { "hash_chains", KSTAT_DATA_UINT64 }, 449 { "hash_chain_max", KSTAT_DATA_UINT64 }, 450 { "p", KSTAT_DATA_UINT64 }, 451 { "c", KSTAT_DATA_UINT64 }, 452 { "c_min", KSTAT_DATA_UINT64 }, 453 { "c_max", KSTAT_DATA_UINT64 }, 454 { "size", KSTAT_DATA_UINT64 }, 455 { "hdr_size", KSTAT_DATA_UINT64 }, 456 { "data_size", KSTAT_DATA_UINT64 }, 457 { "other_size", KSTAT_DATA_UINT64 }, 458 { "l2_hits", KSTAT_DATA_UINT64 }, 459 { "l2_misses", KSTAT_DATA_UINT64 }, 460 { "l2_feeds", KSTAT_DATA_UINT64 }, 461 { "l2_rw_clash", KSTAT_DATA_UINT64 }, 462 { "l2_read_bytes", KSTAT_DATA_UINT64 }, 463 { "l2_write_bytes", KSTAT_DATA_UINT64 }, 464 { "l2_writes_sent", KSTAT_DATA_UINT64 }, 465 { "l2_writes_done", KSTAT_DATA_UINT64 }, 466 { "l2_writes_error", KSTAT_DATA_UINT64 }, 467 { "l2_writes_hdr_miss", KSTAT_DATA_UINT64 }, 468 { "l2_evict_lock_retry", KSTAT_DATA_UINT64 }, 469 { "l2_evict_reading", KSTAT_DATA_UINT64 }, 470 { "l2_free_on_write", KSTAT_DATA_UINT64 }, 471 { "l2_cdata_free_on_write", KSTAT_DATA_UINT64 }, 472 { "l2_abort_lowmem", KSTAT_DATA_UINT64 }, 473 { "l2_cksum_bad", KSTAT_DATA_UINT64 }, 474 { "l2_io_error", KSTAT_DATA_UINT64 }, 475 { "l2_size", KSTAT_DATA_UINT64 }, 476 { "l2_asize", KSTAT_DATA_UINT64 }, 477 { "l2_hdr_size", KSTAT_DATA_UINT64 }, 478 { "l2_compress_successes", KSTAT_DATA_UINT64 }, 479 { "l2_compress_zeros", KSTAT_DATA_UINT64 }, 480 { "l2_compress_failures", KSTAT_DATA_UINT64 }, 481 { "l2_write_trylock_fail", KSTAT_DATA_UINT64 }, 482 { "l2_write_passed_headroom", KSTAT_DATA_UINT64 }, 483 { "l2_write_spa_mismatch", KSTAT_DATA_UINT64 }, 484 { "l2_write_in_l2", KSTAT_DATA_UINT64 }, 485 { "l2_write_io_in_progress", KSTAT_DATA_UINT64 }, 486 { "l2_write_not_cacheable", KSTAT_DATA_UINT64 }, 487 { "l2_write_full", KSTAT_DATA_UINT64 }, 488 { "l2_write_buffer_iter", KSTAT_DATA_UINT64 }, 489 { "l2_write_pios", KSTAT_DATA_UINT64 }, 490 { "l2_write_buffer_bytes_scanned", KSTAT_DATA_UINT64 }, 491 { "l2_write_buffer_list_iter", KSTAT_DATA_UINT64 }, 492 { "l2_write_buffer_list_null_iter", KSTAT_DATA_UINT64 }, 493 { "memory_throttle_count", KSTAT_DATA_UINT64 }, 494 { "duplicate_buffers", KSTAT_DATA_UINT64 }, 495 { "duplicate_buffers_size", KSTAT_DATA_UINT64 }, 496 { "duplicate_reads", KSTAT_DATA_UINT64 } 497}; 498 499#define ARCSTAT(stat) (arc_stats.stat.value.ui64) 500 501#define ARCSTAT_INCR(stat, val) \ 502 atomic_add_64(&arc_stats.stat.value.ui64, (val)) 503 504#define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1) 505#define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1) 506 507#define ARCSTAT_MAX(stat, val) { \ 508 uint64_t m; \ 509 while ((val) > (m = arc_stats.stat.value.ui64) && \ 510 (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \ 511 continue; \ 512} 513 514#define ARCSTAT_MAXSTAT(stat) \ 515 ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64) 516 517/* 518 * We define a macro to allow ARC hits/misses to be easily broken down by 519 * two separate conditions, giving a total of four different subtypes for 520 * each of hits and misses (so eight statistics total). 521 */ 522#define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \ 523 if (cond1) { \ 524 if (cond2) { \ 525 ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \ 526 } else { \ 527 ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \ 528 } \ 529 } else { \ 530 if (cond2) { \ 531 ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \ 532 } else { \ 533 ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\ 534 } \ 535 } 536 537kstat_t *arc_ksp; 538static arc_state_t *arc_anon; 539static arc_state_t *arc_mru; 540static arc_state_t *arc_mru_ghost; 541static arc_state_t *arc_mfu; 542static arc_state_t *arc_mfu_ghost; 543static arc_state_t *arc_l2c_only; 544 545/* 546 * There are several ARC variables that are critical to export as kstats -- 547 * but we don't want to have to grovel around in the kstat whenever we wish to 548 * manipulate them. For these variables, we therefore define them to be in 549 * terms of the statistic variable. This assures that we are not introducing 550 * the possibility of inconsistency by having shadow copies of the variables, 551 * while still allowing the code to be readable. 552 */ 553#define arc_size ARCSTAT(arcstat_size) /* actual total arc size */ 554#define arc_p ARCSTAT(arcstat_p) /* target size of MRU */ 555#define arc_c ARCSTAT(arcstat_c) /* target size of cache */ 556#define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */ 557#define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */ 558 559#define L2ARC_IS_VALID_COMPRESS(_c_) \ 560 ((_c_) == ZIO_COMPRESS_LZ4 || (_c_) == ZIO_COMPRESS_EMPTY) 561 562static int arc_no_grow; /* Don't try to grow cache size */ 563static uint64_t arc_tempreserve; 564static uint64_t arc_loaned_bytes; 565static uint64_t arc_meta_used; 566static uint64_t arc_meta_limit; 567static uint64_t arc_meta_max = 0; 568SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_meta_used, CTLFLAG_RD, &arc_meta_used, 0, 569 "ARC metadata used"); 570SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_meta_limit, CTLFLAG_RW, &arc_meta_limit, 0, 571 "ARC metadata limit"); 572 573typedef struct l2arc_buf_hdr l2arc_buf_hdr_t; 574 575typedef struct arc_callback arc_callback_t; 576 577struct arc_callback { 578 void *acb_private; 579 arc_done_func_t *acb_done; 580 arc_buf_t *acb_buf; 581 zio_t *acb_zio_dummy; 582 arc_callback_t *acb_next; 583}; 584 585typedef struct arc_write_callback arc_write_callback_t; 586 587struct arc_write_callback { 588 void *awcb_private; 589 arc_done_func_t *awcb_ready; 590 arc_done_func_t *awcb_physdone; 591 arc_done_func_t *awcb_done; 592 arc_buf_t *awcb_buf; 593}; 594 595struct arc_buf_hdr { 596 /* protected by hash lock */ 597 dva_t b_dva; 598 uint64_t b_birth; 599 uint64_t b_cksum0; 600 601 kmutex_t b_freeze_lock; 602 zio_cksum_t *b_freeze_cksum; 603 void *b_thawed; 604 605 arc_buf_hdr_t *b_hash_next; 606 arc_buf_t *b_buf; 607 uint32_t b_flags; 608 uint32_t b_datacnt; 609 610 arc_callback_t *b_acb; 611 kcondvar_t b_cv; 612 613 /* immutable */ 614 arc_buf_contents_t b_type; 615 uint64_t b_size; 616 uint64_t b_spa; 617 618 /* protected by arc state mutex */ 619 arc_state_t *b_state; 620 list_node_t b_arc_node; 621 622 /* updated atomically */ 623 clock_t b_arc_access; 624 625 /* self protecting */ 626 refcount_t b_refcnt; 627 628 l2arc_buf_hdr_t *b_l2hdr; 629 list_node_t b_l2node; 630}; 631 632static arc_buf_t *arc_eviction_list; 633static kmutex_t arc_eviction_mtx; 634static arc_buf_hdr_t arc_eviction_hdr; 635static void arc_get_data_buf(arc_buf_t *buf); 636static void arc_access(arc_buf_hdr_t *buf, kmutex_t *hash_lock); 637static int arc_evict_needed(arc_buf_contents_t type); 638static void arc_evict_ghost(arc_state_t *state, uint64_t spa, int64_t bytes); 639#ifdef illumos 640static void arc_buf_watch(arc_buf_t *buf); 641#endif /* illumos */ 642 643static boolean_t l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *ab); 644 645#define GHOST_STATE(state) \ 646 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \ 647 (state) == arc_l2c_only) 648 649/* 650 * Private ARC flags. These flags are private ARC only flags that will show up 651 * in b_flags in the arc_hdr_buf_t. Some flags are publicly declared, and can 652 * be passed in as arc_flags in things like arc_read. However, these flags 653 * should never be passed and should only be set by ARC code. When adding new 654 * public flags, make sure not to smash the private ones. 655 */ 656 657#define ARC_IN_HASH_TABLE (1 << 9) /* this buffer is hashed */ 658#define ARC_IO_IN_PROGRESS (1 << 10) /* I/O in progress for buf */ 659#define ARC_IO_ERROR (1 << 11) /* I/O failed for buf */ 660#define ARC_FREED_IN_READ (1 << 12) /* buf freed while in read */ 661#define ARC_BUF_AVAILABLE (1 << 13) /* block not in active use */ 662#define ARC_INDIRECT (1 << 14) /* this is an indirect block */ 663#define ARC_FREE_IN_PROGRESS (1 << 15) /* hdr about to be freed */ 664#define ARC_L2_WRITING (1 << 16) /* L2ARC write in progress */ 665#define ARC_L2_EVICTED (1 << 17) /* evicted during I/O */ 666#define ARC_L2_WRITE_HEAD (1 << 18) /* head of write list */ 667 668#define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_IN_HASH_TABLE) 669#define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_IO_IN_PROGRESS) 670#define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_IO_ERROR) 671#define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_PREFETCH) 672#define HDR_FREED_IN_READ(hdr) ((hdr)->b_flags & ARC_FREED_IN_READ) 673#define HDR_BUF_AVAILABLE(hdr) ((hdr)->b_flags & ARC_BUF_AVAILABLE) 674#define HDR_FREE_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FREE_IN_PROGRESS) 675#define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_L2CACHE) 676#define HDR_L2_READING(hdr) ((hdr)->b_flags & ARC_IO_IN_PROGRESS && \ 677 (hdr)->b_l2hdr != NULL) 678#define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_L2_WRITING) 679#define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_L2_EVICTED) 680#define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_L2_WRITE_HEAD) 681 682/* 683 * Other sizes 684 */ 685 686#define HDR_SIZE ((int64_t)sizeof (arc_buf_hdr_t)) 687#define L2HDR_SIZE ((int64_t)sizeof (l2arc_buf_hdr_t)) 688 689/* 690 * Hash table routines 691 */ 692 693#define HT_LOCK_PAD CACHE_LINE_SIZE 694 695struct ht_lock { 696 kmutex_t ht_lock; 697#ifdef _KERNEL 698 unsigned char pad[(HT_LOCK_PAD - sizeof (kmutex_t))]; 699#endif 700}; 701 702#define BUF_LOCKS 256 703typedef struct buf_hash_table { 704 uint64_t ht_mask; 705 arc_buf_hdr_t **ht_table; 706 struct ht_lock ht_locks[BUF_LOCKS] __aligned(CACHE_LINE_SIZE); 707} buf_hash_table_t; 708 709static buf_hash_table_t buf_hash_table; 710 711#define BUF_HASH_INDEX(spa, dva, birth) \ 712 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask) 713#define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)]) 714#define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock)) 715#define HDR_LOCK(hdr) \ 716 (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth))) 717 718uint64_t zfs_crc64_table[256]; 719 720/* 721 * Level 2 ARC 722 */ 723 724#define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */ 725#define L2ARC_HEADROOM 2 /* num of writes */ 726/* 727 * If we discover during ARC scan any buffers to be compressed, we boost 728 * our headroom for the next scanning cycle by this percentage multiple. 729 */ 730#define L2ARC_HEADROOM_BOOST 200 731#define L2ARC_FEED_SECS 1 /* caching interval secs */ 732#define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */ 733 734#define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent) 735#define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done) 736 737/* L2ARC Performance Tunables */ 738uint64_t l2arc_write_max = L2ARC_WRITE_SIZE; /* default max write size */ 739uint64_t l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra write during warmup */ 740uint64_t l2arc_headroom = L2ARC_HEADROOM; /* number of dev writes */ 741uint64_t l2arc_headroom_boost = L2ARC_HEADROOM_BOOST; 742uint64_t l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */ 743uint64_t l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval milliseconds */ 744boolean_t l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */ 745boolean_t l2arc_feed_again = B_TRUE; /* turbo warmup */ 746boolean_t l2arc_norw = B_TRUE; /* no reads during writes */ 747 748SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_max, CTLFLAG_RW, 749 &l2arc_write_max, 0, "max write size"); 750SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_boost, CTLFLAG_RW, 751 &l2arc_write_boost, 0, "extra write during warmup"); 752SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_headroom, CTLFLAG_RW, 753 &l2arc_headroom, 0, "number of dev writes"); 754SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_secs, CTLFLAG_RW, 755 &l2arc_feed_secs, 0, "interval seconds"); 756SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_min_ms, CTLFLAG_RW, 757 &l2arc_feed_min_ms, 0, "min interval milliseconds"); 758 759SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_noprefetch, CTLFLAG_RW, 760 &l2arc_noprefetch, 0, "don't cache prefetch bufs"); 761SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_feed_again, CTLFLAG_RW, 762 &l2arc_feed_again, 0, "turbo warmup"); 763SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_norw, CTLFLAG_RW, 764 &l2arc_norw, 0, "no reads during writes"); 765 766SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_size, CTLFLAG_RD, 767 &ARC_anon.arcs_size, 0, "size of anonymous state"); 768SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_metadata_lsize, CTLFLAG_RD, 769 &ARC_anon.arcs_lsize[ARC_BUFC_METADATA], 0, "size of anonymous state"); 770SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_data_lsize, CTLFLAG_RD, 771 &ARC_anon.arcs_lsize[ARC_BUFC_DATA], 0, "size of anonymous state"); 772 773SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_size, CTLFLAG_RD, 774 &ARC_mru.arcs_size, 0, "size of mru state"); 775SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_metadata_lsize, CTLFLAG_RD, 776 &ARC_mru.arcs_lsize[ARC_BUFC_METADATA], 0, "size of metadata in mru state"); 777SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_data_lsize, CTLFLAG_RD, 778 &ARC_mru.arcs_lsize[ARC_BUFC_DATA], 0, "size of data in mru state"); 779 780SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_size, CTLFLAG_RD, 781 &ARC_mru_ghost.arcs_size, 0, "size of mru ghost state"); 782SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_metadata_lsize, CTLFLAG_RD, 783 &ARC_mru_ghost.arcs_lsize[ARC_BUFC_METADATA], 0, 784 "size of metadata in mru ghost state"); 785SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_data_lsize, CTLFLAG_RD, 786 &ARC_mru_ghost.arcs_lsize[ARC_BUFC_DATA], 0, 787 "size of data in mru ghost state"); 788 789SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_size, CTLFLAG_RD, 790 &ARC_mfu.arcs_size, 0, "size of mfu state"); 791SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_metadata_lsize, CTLFLAG_RD, 792 &ARC_mfu.arcs_lsize[ARC_BUFC_METADATA], 0, "size of metadata in mfu state"); 793SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_data_lsize, CTLFLAG_RD, 794 &ARC_mfu.arcs_lsize[ARC_BUFC_DATA], 0, "size of data in mfu state"); 795 796SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_size, CTLFLAG_RD, 797 &ARC_mfu_ghost.arcs_size, 0, "size of mfu ghost state"); 798SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_metadata_lsize, CTLFLAG_RD, 799 &ARC_mfu_ghost.arcs_lsize[ARC_BUFC_METADATA], 0, 800 "size of metadata in mfu ghost state"); 801SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_data_lsize, CTLFLAG_RD, 802 &ARC_mfu_ghost.arcs_lsize[ARC_BUFC_DATA], 0, 803 "size of data in mfu ghost state"); 804 805SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2c_only_size, CTLFLAG_RD, 806 &ARC_l2c_only.arcs_size, 0, "size of mru state"); 807 808/* 809 * L2ARC Internals 810 */ 811typedef struct l2arc_dev { 812 vdev_t *l2ad_vdev; /* vdev */ 813 spa_t *l2ad_spa; /* spa */ 814 uint64_t l2ad_hand; /* next write location */ 815 uint64_t l2ad_start; /* first addr on device */ 816 uint64_t l2ad_end; /* last addr on device */ 817 uint64_t l2ad_evict; /* last addr eviction reached */ 818 boolean_t l2ad_first; /* first sweep through */ 819 boolean_t l2ad_writing; /* currently writing */ 820 list_t *l2ad_buflist; /* buffer list */ 821 list_node_t l2ad_node; /* device list node */ 822} l2arc_dev_t; 823 824static list_t L2ARC_dev_list; /* device list */ 825static list_t *l2arc_dev_list; /* device list pointer */ 826static kmutex_t l2arc_dev_mtx; /* device list mutex */ 827static l2arc_dev_t *l2arc_dev_last; /* last device used */ 828static kmutex_t l2arc_buflist_mtx; /* mutex for all buflists */ 829static list_t L2ARC_free_on_write; /* free after write buf list */ 830static list_t *l2arc_free_on_write; /* free after write list ptr */ 831static kmutex_t l2arc_free_on_write_mtx; /* mutex for list */ 832static uint64_t l2arc_ndev; /* number of devices */ 833 834typedef struct l2arc_read_callback { 835 arc_buf_t *l2rcb_buf; /* read buffer */ 836 spa_t *l2rcb_spa; /* spa */ 837 blkptr_t l2rcb_bp; /* original blkptr */ 838 zbookmark_phys_t l2rcb_zb; /* original bookmark */ 839 int l2rcb_flags; /* original flags */ 840 enum zio_compress l2rcb_compress; /* applied compress */ 841} l2arc_read_callback_t; 842 843typedef struct l2arc_write_callback { 844 l2arc_dev_t *l2wcb_dev; /* device info */ 845 arc_buf_hdr_t *l2wcb_head; /* head of write buflist */ 846} l2arc_write_callback_t; 847 848struct l2arc_buf_hdr { 849 /* protected by arc_buf_hdr mutex */ 850 l2arc_dev_t *b_dev; /* L2ARC device */ 851 uint64_t b_daddr; /* disk address, offset byte */ 852 /* compression applied to buffer data */ 853 enum zio_compress b_compress; 854 /* real alloc'd buffer size depending on b_compress applied */ 855 int b_asize; 856 /* temporary buffer holder for in-flight compressed data */ 857 void *b_tmp_cdata; 858}; 859 860typedef struct l2arc_data_free { 861 /* protected by l2arc_free_on_write_mtx */ 862 void *l2df_data; 863 size_t l2df_size; 864 void (*l2df_func)(void *, size_t); 865 list_node_t l2df_list_node; 866} l2arc_data_free_t; 867 868static kmutex_t l2arc_feed_thr_lock; 869static kcondvar_t l2arc_feed_thr_cv; 870static uint8_t l2arc_thread_exit; 871 872static void l2arc_read_done(zio_t *zio); 873static void l2arc_hdr_stat_add(void); 874static void l2arc_hdr_stat_remove(void); 875 876static boolean_t l2arc_compress_buf(l2arc_buf_hdr_t *l2hdr); 877static void l2arc_decompress_zio(zio_t *zio, arc_buf_hdr_t *hdr, 878 enum zio_compress c); 879static void l2arc_release_cdata_buf(arc_buf_hdr_t *ab); 880 881static uint64_t 882buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth) 883{ 884 uint8_t *vdva = (uint8_t *)dva; 885 uint64_t crc = -1ULL; 886 int i; 887 888 ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY); 889 890 for (i = 0; i < sizeof (dva_t); i++) 891 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ vdva[i]) & 0xFF]; 892 893 crc ^= (spa>>8) ^ birth; 894 895 return (crc); 896} 897 898#define BUF_EMPTY(buf) \ 899 ((buf)->b_dva.dva_word[0] == 0 && \ 900 (buf)->b_dva.dva_word[1] == 0 && \ 901 (buf)->b_cksum0 == 0) 902 903#define BUF_EQUAL(spa, dva, birth, buf) \ 904 ((buf)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \ 905 ((buf)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \ 906 ((buf)->b_birth == birth) && ((buf)->b_spa == spa) 907 908static void 909buf_discard_identity(arc_buf_hdr_t *hdr) 910{ 911 hdr->b_dva.dva_word[0] = 0; 912 hdr->b_dva.dva_word[1] = 0; 913 hdr->b_birth = 0; 914 hdr->b_cksum0 = 0; 915} 916 917static arc_buf_hdr_t * 918buf_hash_find(uint64_t spa, const blkptr_t *bp, kmutex_t **lockp) 919{ 920 const dva_t *dva = BP_IDENTITY(bp); 921 uint64_t birth = BP_PHYSICAL_BIRTH(bp); 922 uint64_t idx = BUF_HASH_INDEX(spa, dva, birth); 923 kmutex_t *hash_lock = BUF_HASH_LOCK(idx); 924 arc_buf_hdr_t *buf; 925 926 mutex_enter(hash_lock); 927 for (buf = buf_hash_table.ht_table[idx]; buf != NULL; 928 buf = buf->b_hash_next) { 929 if (BUF_EQUAL(spa, dva, birth, buf)) { 930 *lockp = hash_lock; 931 return (buf); 932 } 933 } 934 mutex_exit(hash_lock); 935 *lockp = NULL; 936 return (NULL); 937} 938 939/* 940 * Insert an entry into the hash table. If there is already an element 941 * equal to elem in the hash table, then the already existing element 942 * will be returned and the new element will not be inserted. 943 * Otherwise returns NULL. 944 */ 945static arc_buf_hdr_t * 946buf_hash_insert(arc_buf_hdr_t *buf, kmutex_t **lockp) 947{ 948 uint64_t idx = BUF_HASH_INDEX(buf->b_spa, &buf->b_dva, buf->b_birth); 949 kmutex_t *hash_lock = BUF_HASH_LOCK(idx); 950 arc_buf_hdr_t *fbuf; 951 uint32_t i; 952 953 ASSERT(!DVA_IS_EMPTY(&buf->b_dva)); 954 ASSERT(buf->b_birth != 0); 955 ASSERT(!HDR_IN_HASH_TABLE(buf)); 956 *lockp = hash_lock; 957 mutex_enter(hash_lock); 958 for (fbuf = buf_hash_table.ht_table[idx], i = 0; fbuf != NULL; 959 fbuf = fbuf->b_hash_next, i++) { 960 if (BUF_EQUAL(buf->b_spa, &buf->b_dva, buf->b_birth, fbuf)) 961 return (fbuf); 962 } 963 964 buf->b_hash_next = buf_hash_table.ht_table[idx]; 965 buf_hash_table.ht_table[idx] = buf; 966 buf->b_flags |= ARC_IN_HASH_TABLE; 967 968 /* collect some hash table performance data */ 969 if (i > 0) { 970 ARCSTAT_BUMP(arcstat_hash_collisions); 971 if (i == 1) 972 ARCSTAT_BUMP(arcstat_hash_chains); 973 974 ARCSTAT_MAX(arcstat_hash_chain_max, i); 975 } 976 977 ARCSTAT_BUMP(arcstat_hash_elements); 978 ARCSTAT_MAXSTAT(arcstat_hash_elements); 979 980 return (NULL); 981} 982 983static void 984buf_hash_remove(arc_buf_hdr_t *buf) 985{ 986 arc_buf_hdr_t *fbuf, **bufp; 987 uint64_t idx = BUF_HASH_INDEX(buf->b_spa, &buf->b_dva, buf->b_birth); 988 989 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx))); 990 ASSERT(HDR_IN_HASH_TABLE(buf)); 991 992 bufp = &buf_hash_table.ht_table[idx]; 993 while ((fbuf = *bufp) != buf) { 994 ASSERT(fbuf != NULL); 995 bufp = &fbuf->b_hash_next; 996 } 997 *bufp = buf->b_hash_next; 998 buf->b_hash_next = NULL; 999 buf->b_flags &= ~ARC_IN_HASH_TABLE; 1000 1001 /* collect some hash table performance data */ 1002 ARCSTAT_BUMPDOWN(arcstat_hash_elements); 1003 1004 if (buf_hash_table.ht_table[idx] && 1005 buf_hash_table.ht_table[idx]->b_hash_next == NULL) 1006 ARCSTAT_BUMPDOWN(arcstat_hash_chains); 1007} 1008 1009/* 1010 * Global data structures and functions for the buf kmem cache. 1011 */ 1012static kmem_cache_t *hdr_cache; 1013static kmem_cache_t *buf_cache; 1014 1015static void 1016buf_fini(void) 1017{ 1018 int i; 1019 1020 kmem_free(buf_hash_table.ht_table, 1021 (buf_hash_table.ht_mask + 1) * sizeof (void *)); 1022 for (i = 0; i < BUF_LOCKS; i++) 1023 mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock); 1024 kmem_cache_destroy(hdr_cache); 1025 kmem_cache_destroy(buf_cache); 1026} 1027 1028/* 1029 * Constructor callback - called when the cache is empty 1030 * and a new buf is requested. 1031 */ 1032/* ARGSUSED */ 1033static int 1034hdr_cons(void *vbuf, void *unused, int kmflag) 1035{ 1036 arc_buf_hdr_t *buf = vbuf; 1037 1038 bzero(buf, sizeof (arc_buf_hdr_t)); 1039 refcount_create(&buf->b_refcnt); 1040 cv_init(&buf->b_cv, NULL, CV_DEFAULT, NULL); 1041 mutex_init(&buf->b_freeze_lock, NULL, MUTEX_DEFAULT, NULL); 1042 arc_space_consume(sizeof (arc_buf_hdr_t), ARC_SPACE_HDRS); 1043 1044 return (0); 1045} 1046 1047/* ARGSUSED */ 1048static int 1049buf_cons(void *vbuf, void *unused, int kmflag) 1050{ 1051 arc_buf_t *buf = vbuf; 1052 1053 bzero(buf, sizeof (arc_buf_t)); 1054 mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL); 1055 arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS); 1056 1057 return (0); 1058} 1059 1060/* 1061 * Destructor callback - called when a cached buf is 1062 * no longer required. 1063 */ 1064/* ARGSUSED */ 1065static void 1066hdr_dest(void *vbuf, void *unused) 1067{ 1068 arc_buf_hdr_t *buf = vbuf; 1069 1070 ASSERT(BUF_EMPTY(buf)); 1071 refcount_destroy(&buf->b_refcnt); 1072 cv_destroy(&buf->b_cv); 1073 mutex_destroy(&buf->b_freeze_lock); 1074 arc_space_return(sizeof (arc_buf_hdr_t), ARC_SPACE_HDRS); 1075} 1076 1077/* ARGSUSED */ 1078static void 1079buf_dest(void *vbuf, void *unused) 1080{ 1081 arc_buf_t *buf = vbuf; 1082 1083 mutex_destroy(&buf->b_evict_lock); 1084 arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS); 1085} 1086 1087/* 1088 * Reclaim callback -- invoked when memory is low. 1089 */ 1090/* ARGSUSED */ 1091static void 1092hdr_recl(void *unused) 1093{ 1094 dprintf("hdr_recl called\n"); 1095 /* 1096 * umem calls the reclaim func when we destroy the buf cache, 1097 * which is after we do arc_fini(). 1098 */ 1099 if (!arc_dead) 1100 cv_signal(&arc_reclaim_thr_cv); 1101} 1102 1103static void 1104buf_init(void) 1105{ 1106 uint64_t *ct; 1107 uint64_t hsize = 1ULL << 12; 1108 int i, j; 1109 1110 /* 1111 * The hash table is big enough to fill all of physical memory 1112 * with an average block size of zfs_arc_average_blocksize (default 8K). 1113 * By default, the table will take up 1114 * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers). 1115 */ 1116 while (hsize * zfs_arc_average_blocksize < (uint64_t)physmem * PAGESIZE) 1117 hsize <<= 1; 1118retry: 1119 buf_hash_table.ht_mask = hsize - 1; 1120 buf_hash_table.ht_table = 1121 kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP); 1122 if (buf_hash_table.ht_table == NULL) { 1123 ASSERT(hsize > (1ULL << 8)); 1124 hsize >>= 1; 1125 goto retry; 1126 } 1127 1128 hdr_cache = kmem_cache_create("arc_buf_hdr_t", sizeof (arc_buf_hdr_t), 1129 0, hdr_cons, hdr_dest, hdr_recl, NULL, NULL, 0); 1130 buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t), 1131 0, buf_cons, buf_dest, NULL, NULL, NULL, 0); 1132 1133 for (i = 0; i < 256; i++) 1134 for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--) 1135 *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY); 1136 1137 for (i = 0; i < BUF_LOCKS; i++) { 1138 mutex_init(&buf_hash_table.ht_locks[i].ht_lock, 1139 NULL, MUTEX_DEFAULT, NULL); 1140 } 1141} 1142 1143#define ARC_MINTIME (hz>>4) /* 62 ms */ 1144 1145static void 1146arc_cksum_verify(arc_buf_t *buf) 1147{ 1148 zio_cksum_t zc; 1149 1150 if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 1151 return; 1152 1153 mutex_enter(&buf->b_hdr->b_freeze_lock); 1154 if (buf->b_hdr->b_freeze_cksum == NULL || 1155 (buf->b_hdr->b_flags & ARC_IO_ERROR)) { 1156 mutex_exit(&buf->b_hdr->b_freeze_lock); 1157 return; 1158 } 1159 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc); 1160 if (!ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc)) 1161 panic("buffer modified while frozen!"); 1162 mutex_exit(&buf->b_hdr->b_freeze_lock); 1163} 1164 1165static int 1166arc_cksum_equal(arc_buf_t *buf) 1167{ 1168 zio_cksum_t zc; 1169 int equal; 1170 1171 mutex_enter(&buf->b_hdr->b_freeze_lock); 1172 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc); 1173 equal = ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc); 1174 mutex_exit(&buf->b_hdr->b_freeze_lock); 1175 1176 return (equal); 1177} 1178 1179static void 1180arc_cksum_compute(arc_buf_t *buf, boolean_t force) 1181{ 1182 if (!force && !(zfs_flags & ZFS_DEBUG_MODIFY)) 1183 return; 1184 1185 mutex_enter(&buf->b_hdr->b_freeze_lock); 1186 if (buf->b_hdr->b_freeze_cksum != NULL) { 1187 mutex_exit(&buf->b_hdr->b_freeze_lock); 1188 return; 1189 } 1190 buf->b_hdr->b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t), KM_SLEEP); 1191 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, 1192 buf->b_hdr->b_freeze_cksum); 1193 mutex_exit(&buf->b_hdr->b_freeze_lock); 1194#ifdef illumos 1195 arc_buf_watch(buf); 1196#endif /* illumos */ 1197} 1198 1199#ifdef illumos 1200#ifndef _KERNEL 1201typedef struct procctl { 1202 long cmd; 1203 prwatch_t prwatch; 1204} procctl_t; 1205#endif 1206 1207/* ARGSUSED */ 1208static void 1209arc_buf_unwatch(arc_buf_t *buf) 1210{ 1211#ifndef _KERNEL 1212 if (arc_watch) { 1213 int result; 1214 procctl_t ctl; 1215 ctl.cmd = PCWATCH; 1216 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data; 1217 ctl.prwatch.pr_size = 0; 1218 ctl.prwatch.pr_wflags = 0; 1219 result = write(arc_procfd, &ctl, sizeof (ctl)); 1220 ASSERT3U(result, ==, sizeof (ctl)); 1221 } 1222#endif 1223} 1224 1225/* ARGSUSED */ 1226static void 1227arc_buf_watch(arc_buf_t *buf) 1228{ 1229#ifndef _KERNEL 1230 if (arc_watch) { 1231 int result; 1232 procctl_t ctl; 1233 ctl.cmd = PCWATCH; 1234 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data; 1235 ctl.prwatch.pr_size = buf->b_hdr->b_size; 1236 ctl.prwatch.pr_wflags = WA_WRITE; 1237 result = write(arc_procfd, &ctl, sizeof (ctl)); 1238 ASSERT3U(result, ==, sizeof (ctl)); 1239 } 1240#endif 1241} 1242#endif /* illumos */ 1243 1244void 1245arc_buf_thaw(arc_buf_t *buf) 1246{ 1247 if (zfs_flags & ZFS_DEBUG_MODIFY) { 1248 if (buf->b_hdr->b_state != arc_anon) 1249 panic("modifying non-anon buffer!"); 1250 if (buf->b_hdr->b_flags & ARC_IO_IN_PROGRESS) 1251 panic("modifying buffer while i/o in progress!"); 1252 arc_cksum_verify(buf); 1253 } 1254 1255 mutex_enter(&buf->b_hdr->b_freeze_lock); 1256 if (buf->b_hdr->b_freeze_cksum != NULL) { 1257 kmem_free(buf->b_hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 1258 buf->b_hdr->b_freeze_cksum = NULL; 1259 } 1260 1261 if (zfs_flags & ZFS_DEBUG_MODIFY) { 1262 if (buf->b_hdr->b_thawed) 1263 kmem_free(buf->b_hdr->b_thawed, 1); 1264 buf->b_hdr->b_thawed = kmem_alloc(1, KM_SLEEP); 1265 } 1266 1267 mutex_exit(&buf->b_hdr->b_freeze_lock); 1268 1269#ifdef illumos 1270 arc_buf_unwatch(buf); 1271#endif /* illumos */ 1272} 1273 1274void 1275arc_buf_freeze(arc_buf_t *buf) 1276{ 1277 kmutex_t *hash_lock; 1278 1279 if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 1280 return; 1281 1282 hash_lock = HDR_LOCK(buf->b_hdr); 1283 mutex_enter(hash_lock); 1284 1285 ASSERT(buf->b_hdr->b_freeze_cksum != NULL || 1286 buf->b_hdr->b_state == arc_anon); 1287 arc_cksum_compute(buf, B_FALSE); 1288 mutex_exit(hash_lock); 1289 1290} 1291 1292static void 1293get_buf_info(arc_buf_hdr_t *ab, arc_state_t *state, list_t **list, kmutex_t **lock) 1294{ 1295 uint64_t buf_hashid = buf_hash(ab->b_spa, &ab->b_dva, ab->b_birth); 1296 1297 if (ab->b_type == ARC_BUFC_METADATA) 1298 buf_hashid &= (ARC_BUFC_NUMMETADATALISTS - 1); 1299 else { 1300 buf_hashid &= (ARC_BUFC_NUMDATALISTS - 1); 1301 buf_hashid += ARC_BUFC_NUMMETADATALISTS; 1302 } 1303 1304 *list = &state->arcs_lists[buf_hashid]; 1305 *lock = ARCS_LOCK(state, buf_hashid); 1306} 1307 1308 1309static void 1310add_reference(arc_buf_hdr_t *ab, kmutex_t *hash_lock, void *tag) 1311{ 1312 ASSERT(MUTEX_HELD(hash_lock)); 1313 1314 if ((refcount_add(&ab->b_refcnt, tag) == 1) && 1315 (ab->b_state != arc_anon)) { 1316 uint64_t delta = ab->b_size * ab->b_datacnt; 1317 uint64_t *size = &ab->b_state->arcs_lsize[ab->b_type]; 1318 list_t *list; 1319 kmutex_t *lock; 1320 1321 get_buf_info(ab, ab->b_state, &list, &lock); 1322 ASSERT(!MUTEX_HELD(lock)); 1323 mutex_enter(lock); 1324 ASSERT(list_link_active(&ab->b_arc_node)); 1325 list_remove(list, ab); 1326 if (GHOST_STATE(ab->b_state)) { 1327 ASSERT0(ab->b_datacnt); 1328 ASSERT3P(ab->b_buf, ==, NULL); 1329 delta = ab->b_size; 1330 } 1331 ASSERT(delta > 0); 1332 ASSERT3U(*size, >=, delta); 1333 atomic_add_64(size, -delta); 1334 mutex_exit(lock); 1335 /* remove the prefetch flag if we get a reference */ 1336 if (ab->b_flags & ARC_PREFETCH) 1337 ab->b_flags &= ~ARC_PREFETCH; 1338 } 1339} 1340 1341static int 1342remove_reference(arc_buf_hdr_t *ab, kmutex_t *hash_lock, void *tag) 1343{ 1344 int cnt; 1345 arc_state_t *state = ab->b_state; 1346 1347 ASSERT(state == arc_anon || MUTEX_HELD(hash_lock)); 1348 ASSERT(!GHOST_STATE(state)); 1349 1350 if (((cnt = refcount_remove(&ab->b_refcnt, tag)) == 0) && 1351 (state != arc_anon)) { 1352 uint64_t *size = &state->arcs_lsize[ab->b_type]; 1353 list_t *list; 1354 kmutex_t *lock; 1355 1356 get_buf_info(ab, state, &list, &lock); 1357 ASSERT(!MUTEX_HELD(lock)); 1358 mutex_enter(lock); 1359 ASSERT(!list_link_active(&ab->b_arc_node)); 1360 list_insert_head(list, ab); 1361 ASSERT(ab->b_datacnt > 0); 1362 atomic_add_64(size, ab->b_size * ab->b_datacnt); 1363 mutex_exit(lock); 1364 } 1365 return (cnt); 1366} 1367 1368/* 1369 * Move the supplied buffer to the indicated state. The mutex 1370 * for the buffer must be held by the caller. 1371 */ 1372static void 1373arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *ab, kmutex_t *hash_lock) 1374{ 1375 arc_state_t *old_state = ab->b_state; 1376 int64_t refcnt = refcount_count(&ab->b_refcnt); 1377 uint64_t from_delta, to_delta; 1378 list_t *list; 1379 kmutex_t *lock; 1380 1381 ASSERT(MUTEX_HELD(hash_lock)); 1382 ASSERT3P(new_state, !=, old_state); 1383 ASSERT(refcnt == 0 || ab->b_datacnt > 0); 1384 ASSERT(ab->b_datacnt == 0 || !GHOST_STATE(new_state)); 1385 ASSERT(ab->b_datacnt <= 1 || old_state != arc_anon); 1386 1387 from_delta = to_delta = ab->b_datacnt * ab->b_size; 1388 1389 /* 1390 * If this buffer is evictable, transfer it from the 1391 * old state list to the new state list. 1392 */ 1393 if (refcnt == 0) { 1394 if (old_state != arc_anon) { 1395 int use_mutex; 1396 uint64_t *size = &old_state->arcs_lsize[ab->b_type]; 1397 1398 get_buf_info(ab, old_state, &list, &lock); 1399 use_mutex = !MUTEX_HELD(lock); 1400 if (use_mutex) 1401 mutex_enter(lock); 1402 1403 ASSERT(list_link_active(&ab->b_arc_node)); 1404 list_remove(list, ab); 1405 1406 /* 1407 * If prefetching out of the ghost cache, 1408 * we will have a non-zero datacnt. 1409 */ 1410 if (GHOST_STATE(old_state) && ab->b_datacnt == 0) { 1411 /* ghost elements have a ghost size */ 1412 ASSERT(ab->b_buf == NULL); 1413 from_delta = ab->b_size; 1414 } 1415 ASSERT3U(*size, >=, from_delta); 1416 atomic_add_64(size, -from_delta); 1417 1418 if (use_mutex) 1419 mutex_exit(lock); 1420 } 1421 if (new_state != arc_anon) { 1422 int use_mutex; 1423 uint64_t *size = &new_state->arcs_lsize[ab->b_type]; 1424 1425 get_buf_info(ab, new_state, &list, &lock); 1426 use_mutex = !MUTEX_HELD(lock); 1427 if (use_mutex) 1428 mutex_enter(lock); 1429 1430 list_insert_head(list, ab); 1431 1432 /* ghost elements have a ghost size */ 1433 if (GHOST_STATE(new_state)) { 1434 ASSERT(ab->b_datacnt == 0); 1435 ASSERT(ab->b_buf == NULL); 1436 to_delta = ab->b_size; 1437 } 1438 atomic_add_64(size, to_delta); 1439 1440 if (use_mutex) 1441 mutex_exit(lock); 1442 } 1443 } 1444 1445 ASSERT(!BUF_EMPTY(ab)); 1446 if (new_state == arc_anon && HDR_IN_HASH_TABLE(ab)) 1447 buf_hash_remove(ab); 1448 1449 /* adjust state sizes */ 1450 if (to_delta) 1451 atomic_add_64(&new_state->arcs_size, to_delta); 1452 if (from_delta) { 1453 ASSERT3U(old_state->arcs_size, >=, from_delta); 1454 atomic_add_64(&old_state->arcs_size, -from_delta); 1455 } 1456 ab->b_state = new_state; 1457 1458 /* adjust l2arc hdr stats */ 1459 if (new_state == arc_l2c_only) 1460 l2arc_hdr_stat_add(); 1461 else if (old_state == arc_l2c_only) 1462 l2arc_hdr_stat_remove(); 1463} 1464 1465void 1466arc_space_consume(uint64_t space, arc_space_type_t type) 1467{ 1468 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES); 1469 1470 switch (type) { 1471 case ARC_SPACE_DATA: 1472 ARCSTAT_INCR(arcstat_data_size, space); 1473 break; 1474 case ARC_SPACE_OTHER: 1475 ARCSTAT_INCR(arcstat_other_size, space); 1476 break; 1477 case ARC_SPACE_HDRS: 1478 ARCSTAT_INCR(arcstat_hdr_size, space); 1479 break; 1480 case ARC_SPACE_L2HDRS: 1481 ARCSTAT_INCR(arcstat_l2_hdr_size, space); 1482 break; 1483 } 1484 1485 atomic_add_64(&arc_meta_used, space); 1486 atomic_add_64(&arc_size, space); 1487} 1488 1489void 1490arc_space_return(uint64_t space, arc_space_type_t type) 1491{ 1492 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES); 1493 1494 switch (type) { 1495 case ARC_SPACE_DATA: 1496 ARCSTAT_INCR(arcstat_data_size, -space); 1497 break; 1498 case ARC_SPACE_OTHER: 1499 ARCSTAT_INCR(arcstat_other_size, -space); 1500 break; 1501 case ARC_SPACE_HDRS: 1502 ARCSTAT_INCR(arcstat_hdr_size, -space); 1503 break; 1504 case ARC_SPACE_L2HDRS: 1505 ARCSTAT_INCR(arcstat_l2_hdr_size, -space); 1506 break; 1507 } 1508 1509 ASSERT(arc_meta_used >= space); 1510 if (arc_meta_max < arc_meta_used) 1511 arc_meta_max = arc_meta_used; 1512 atomic_add_64(&arc_meta_used, -space); 1513 ASSERT(arc_size >= space); 1514 atomic_add_64(&arc_size, -space); 1515} 1516 1517arc_buf_t * 1518arc_buf_alloc(spa_t *spa, int size, void *tag, arc_buf_contents_t type) 1519{ 1520 arc_buf_hdr_t *hdr; 1521 arc_buf_t *buf; 1522 1523 ASSERT3U(size, >, 0); 1524 hdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE); 1525 ASSERT(BUF_EMPTY(hdr)); 1526 hdr->b_size = size; 1527 hdr->b_type = type; 1528 hdr->b_spa = spa_load_guid(spa); 1529 hdr->b_state = arc_anon; 1530 hdr->b_arc_access = 0; 1531 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 1532 buf->b_hdr = hdr; 1533 buf->b_data = NULL; 1534 buf->b_efunc = NULL; 1535 buf->b_private = NULL; 1536 buf->b_next = NULL; 1537 hdr->b_buf = buf; 1538 arc_get_data_buf(buf); 1539 hdr->b_datacnt = 1; 1540 hdr->b_flags = 0; 1541 ASSERT(refcount_is_zero(&hdr->b_refcnt)); 1542 (void) refcount_add(&hdr->b_refcnt, tag); 1543 1544 return (buf); 1545} 1546 1547static char *arc_onloan_tag = "onloan"; 1548 1549/* 1550 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in 1551 * flight data by arc_tempreserve_space() until they are "returned". Loaned 1552 * buffers must be returned to the arc before they can be used by the DMU or 1553 * freed. 1554 */ 1555arc_buf_t * 1556arc_loan_buf(spa_t *spa, int size) 1557{ 1558 arc_buf_t *buf; 1559 1560 buf = arc_buf_alloc(spa, size, arc_onloan_tag, ARC_BUFC_DATA); 1561 1562 atomic_add_64(&arc_loaned_bytes, size); 1563 return (buf); 1564} 1565 1566/* 1567 * Return a loaned arc buffer to the arc. 1568 */ 1569void 1570arc_return_buf(arc_buf_t *buf, void *tag) 1571{ 1572 arc_buf_hdr_t *hdr = buf->b_hdr; 1573 1574 ASSERT(buf->b_data != NULL); 1575 (void) refcount_add(&hdr->b_refcnt, tag); 1576 (void) refcount_remove(&hdr->b_refcnt, arc_onloan_tag); 1577 1578 atomic_add_64(&arc_loaned_bytes, -hdr->b_size); 1579} 1580 1581/* Detach an arc_buf from a dbuf (tag) */ 1582void 1583arc_loan_inuse_buf(arc_buf_t *buf, void *tag) 1584{ 1585 arc_buf_hdr_t *hdr; 1586 1587 ASSERT(buf->b_data != NULL); 1588 hdr = buf->b_hdr; 1589 (void) refcount_add(&hdr->b_refcnt, arc_onloan_tag); 1590 (void) refcount_remove(&hdr->b_refcnt, tag); 1591 buf->b_efunc = NULL; 1592 buf->b_private = NULL; 1593 1594 atomic_add_64(&arc_loaned_bytes, hdr->b_size); 1595} 1596 1597static arc_buf_t * 1598arc_buf_clone(arc_buf_t *from) 1599{ 1600 arc_buf_t *buf; 1601 arc_buf_hdr_t *hdr = from->b_hdr; 1602 uint64_t size = hdr->b_size; 1603 1604 ASSERT(hdr->b_state != arc_anon); 1605 1606 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 1607 buf->b_hdr = hdr; 1608 buf->b_data = NULL; 1609 buf->b_efunc = NULL; 1610 buf->b_private = NULL; 1611 buf->b_next = hdr->b_buf; 1612 hdr->b_buf = buf; 1613 arc_get_data_buf(buf); 1614 bcopy(from->b_data, buf->b_data, size); 1615 1616 /* 1617 * This buffer already exists in the arc so create a duplicate 1618 * copy for the caller. If the buffer is associated with user data 1619 * then track the size and number of duplicates. These stats will be 1620 * updated as duplicate buffers are created and destroyed. 1621 */ 1622 if (hdr->b_type == ARC_BUFC_DATA) { 1623 ARCSTAT_BUMP(arcstat_duplicate_buffers); 1624 ARCSTAT_INCR(arcstat_duplicate_buffers_size, size); 1625 } 1626 hdr->b_datacnt += 1; 1627 return (buf); 1628} 1629 1630void 1631arc_buf_add_ref(arc_buf_t *buf, void* tag) 1632{ 1633 arc_buf_hdr_t *hdr; 1634 kmutex_t *hash_lock; 1635 1636 /* 1637 * Check to see if this buffer is evicted. Callers 1638 * must verify b_data != NULL to know if the add_ref 1639 * was successful. 1640 */ 1641 mutex_enter(&buf->b_evict_lock); 1642 if (buf->b_data == NULL) { 1643 mutex_exit(&buf->b_evict_lock); 1644 return; 1645 } 1646 hash_lock = HDR_LOCK(buf->b_hdr); 1647 mutex_enter(hash_lock); 1648 hdr = buf->b_hdr; 1649 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 1650 mutex_exit(&buf->b_evict_lock); 1651 1652 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu); 1653 add_reference(hdr, hash_lock, tag); 1654 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr); 1655 arc_access(hdr, hash_lock); 1656 mutex_exit(hash_lock); 1657 ARCSTAT_BUMP(arcstat_hits); 1658 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH), 1659 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA, 1660 data, metadata, hits); 1661} 1662 1663static void 1664arc_buf_free_on_write(void *data, size_t size, 1665 void (*free_func)(void *, size_t)) 1666{ 1667 l2arc_data_free_t *df; 1668 1669 df = kmem_alloc(sizeof (l2arc_data_free_t), KM_SLEEP); 1670 df->l2df_data = data; 1671 df->l2df_size = size; 1672 df->l2df_func = free_func; 1673 mutex_enter(&l2arc_free_on_write_mtx); 1674 list_insert_head(l2arc_free_on_write, df); 1675 mutex_exit(&l2arc_free_on_write_mtx); 1676} 1677 1678/* 1679 * Free the arc data buffer. If it is an l2arc write in progress, 1680 * the buffer is placed on l2arc_free_on_write to be freed later. 1681 */ 1682static void 1683arc_buf_data_free(arc_buf_t *buf, void (*free_func)(void *, size_t)) 1684{ 1685 arc_buf_hdr_t *hdr = buf->b_hdr; 1686 1687 if (HDR_L2_WRITING(hdr)) { 1688 arc_buf_free_on_write(buf->b_data, hdr->b_size, free_func); 1689 ARCSTAT_BUMP(arcstat_l2_free_on_write); 1690 } else { 1691 free_func(buf->b_data, hdr->b_size); 1692 } 1693} 1694 1695/* 1696 * Free up buf->b_data and if 'remove' is set, then pull the 1697 * arc_buf_t off of the the arc_buf_hdr_t's list and free it. 1698 */ 1699static void 1700arc_buf_l2_cdata_free(arc_buf_hdr_t *hdr) 1701{ 1702 l2arc_buf_hdr_t *l2hdr = hdr->b_l2hdr; 1703 1704 ASSERT(MUTEX_HELD(&l2arc_buflist_mtx)); 1705 1706 if (l2hdr->b_tmp_cdata == NULL) 1707 return; 1708 1709 ASSERT(HDR_L2_WRITING(hdr)); 1710 arc_buf_free_on_write(l2hdr->b_tmp_cdata, hdr->b_size, 1711 zio_data_buf_free); 1712 ARCSTAT_BUMP(arcstat_l2_cdata_free_on_write); 1713 l2hdr->b_tmp_cdata = NULL; 1714} 1715 1716static void 1717arc_buf_destroy(arc_buf_t *buf, boolean_t recycle, boolean_t remove) 1718{ 1719 arc_buf_t **bufp; 1720 1721 /* free up data associated with the buf */ 1722 if (buf->b_data) { 1723 arc_state_t *state = buf->b_hdr->b_state; 1724 uint64_t size = buf->b_hdr->b_size; 1725 arc_buf_contents_t type = buf->b_hdr->b_type; 1726 1727 arc_cksum_verify(buf); 1728#ifdef illumos 1729 arc_buf_unwatch(buf); 1730#endif /* illumos */ 1731 1732 if (!recycle) { 1733 if (type == ARC_BUFC_METADATA) { 1734 arc_buf_data_free(buf, zio_buf_free); 1735 arc_space_return(size, ARC_SPACE_DATA); 1736 } else { 1737 ASSERT(type == ARC_BUFC_DATA); 1738 arc_buf_data_free(buf, zio_data_buf_free); 1739 ARCSTAT_INCR(arcstat_data_size, -size); 1740 atomic_add_64(&arc_size, -size); 1741 } 1742 } 1743 if (list_link_active(&buf->b_hdr->b_arc_node)) { 1744 uint64_t *cnt = &state->arcs_lsize[type]; 1745 1746 ASSERT(refcount_is_zero(&buf->b_hdr->b_refcnt)); 1747 ASSERT(state != arc_anon); 1748 1749 ASSERT3U(*cnt, >=, size); 1750 atomic_add_64(cnt, -size); 1751 } 1752 ASSERT3U(state->arcs_size, >=, size); 1753 atomic_add_64(&state->arcs_size, -size); 1754 buf->b_data = NULL; 1755 1756 /* 1757 * If we're destroying a duplicate buffer make sure 1758 * that the appropriate statistics are updated. 1759 */ 1760 if (buf->b_hdr->b_datacnt > 1 && 1761 buf->b_hdr->b_type == ARC_BUFC_DATA) { 1762 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers); 1763 ARCSTAT_INCR(arcstat_duplicate_buffers_size, -size); 1764 } 1765 ASSERT(buf->b_hdr->b_datacnt > 0); 1766 buf->b_hdr->b_datacnt -= 1; 1767 } 1768 1769 /* only remove the buf if requested */ 1770 if (!remove) 1771 return; 1772 1773 /* remove the buf from the hdr list */ 1774 for (bufp = &buf->b_hdr->b_buf; *bufp != buf; bufp = &(*bufp)->b_next) 1775 continue; 1776 *bufp = buf->b_next; 1777 buf->b_next = NULL; 1778 1779 ASSERT(buf->b_efunc == NULL); 1780 1781 /* clean up the buf */ 1782 buf->b_hdr = NULL; 1783 kmem_cache_free(buf_cache, buf); 1784} 1785 1786static void 1787arc_hdr_destroy(arc_buf_hdr_t *hdr) 1788{ 1789 ASSERT(refcount_is_zero(&hdr->b_refcnt)); 1790 ASSERT3P(hdr->b_state, ==, arc_anon); 1791 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 1792 l2arc_buf_hdr_t *l2hdr = hdr->b_l2hdr; 1793 1794 if (l2hdr != NULL) { 1795 boolean_t buflist_held = MUTEX_HELD(&l2arc_buflist_mtx); 1796 /* 1797 * To prevent arc_free() and l2arc_evict() from 1798 * attempting to free the same buffer at the same time, 1799 * a FREE_IN_PROGRESS flag is given to arc_free() to 1800 * give it priority. l2arc_evict() can't destroy this 1801 * header while we are waiting on l2arc_buflist_mtx. 1802 * 1803 * The hdr may be removed from l2ad_buflist before we 1804 * grab l2arc_buflist_mtx, so b_l2hdr is rechecked. 1805 */ 1806 if (!buflist_held) { 1807 mutex_enter(&l2arc_buflist_mtx); 1808 l2hdr = hdr->b_l2hdr; 1809 } 1810 1811 if (l2hdr != NULL) { 1812 trim_map_free(l2hdr->b_dev->l2ad_vdev, l2hdr->b_daddr, 1813 hdr->b_size, 0); 1814 list_remove(l2hdr->b_dev->l2ad_buflist, hdr); 1815 arc_buf_l2_cdata_free(hdr); 1816 ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size); 1817 ARCSTAT_INCR(arcstat_l2_asize, -l2hdr->b_asize); 1818 vdev_space_update(l2hdr->b_dev->l2ad_vdev, 1819 -l2hdr->b_asize, 0, 0); 1820 kmem_free(l2hdr, sizeof (l2arc_buf_hdr_t)); 1821 if (hdr->b_state == arc_l2c_only) 1822 l2arc_hdr_stat_remove(); 1823 hdr->b_l2hdr = NULL; 1824 } 1825 1826 if (!buflist_held) 1827 mutex_exit(&l2arc_buflist_mtx); 1828 } 1829 1830 if (!BUF_EMPTY(hdr)) { 1831 ASSERT(!HDR_IN_HASH_TABLE(hdr)); 1832 buf_discard_identity(hdr); 1833 } 1834 while (hdr->b_buf) { 1835 arc_buf_t *buf = hdr->b_buf; 1836 1837 if (buf->b_efunc) { 1838 mutex_enter(&arc_eviction_mtx); 1839 mutex_enter(&buf->b_evict_lock); 1840 ASSERT(buf->b_hdr != NULL); 1841 arc_buf_destroy(hdr->b_buf, FALSE, FALSE); 1842 hdr->b_buf = buf->b_next; 1843 buf->b_hdr = &arc_eviction_hdr; 1844 buf->b_next = arc_eviction_list; 1845 arc_eviction_list = buf; 1846 mutex_exit(&buf->b_evict_lock); 1847 mutex_exit(&arc_eviction_mtx); 1848 } else { 1849 arc_buf_destroy(hdr->b_buf, FALSE, TRUE); 1850 } 1851 } 1852 if (hdr->b_freeze_cksum != NULL) { 1853 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 1854 hdr->b_freeze_cksum = NULL; 1855 } 1856 if (hdr->b_thawed) { 1857 kmem_free(hdr->b_thawed, 1); 1858 hdr->b_thawed = NULL; 1859 } 1860 1861 ASSERT(!list_link_active(&hdr->b_arc_node)); 1862 ASSERT3P(hdr->b_hash_next, ==, NULL); 1863 ASSERT3P(hdr->b_acb, ==, NULL); 1864 kmem_cache_free(hdr_cache, hdr); 1865} 1866 1867void 1868arc_buf_free(arc_buf_t *buf, void *tag) 1869{ 1870 arc_buf_hdr_t *hdr = buf->b_hdr; 1871 int hashed = hdr->b_state != arc_anon; 1872 1873 ASSERT(buf->b_efunc == NULL); 1874 ASSERT(buf->b_data != NULL); 1875 1876 if (hashed) { 1877 kmutex_t *hash_lock = HDR_LOCK(hdr); 1878 1879 mutex_enter(hash_lock); 1880 hdr = buf->b_hdr; 1881 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 1882 1883 (void) remove_reference(hdr, hash_lock, tag); 1884 if (hdr->b_datacnt > 1) { 1885 arc_buf_destroy(buf, FALSE, TRUE); 1886 } else { 1887 ASSERT(buf == hdr->b_buf); 1888 ASSERT(buf->b_efunc == NULL); 1889 hdr->b_flags |= ARC_BUF_AVAILABLE; 1890 } 1891 mutex_exit(hash_lock); 1892 } else if (HDR_IO_IN_PROGRESS(hdr)) { 1893 int destroy_hdr; 1894 /* 1895 * We are in the middle of an async write. Don't destroy 1896 * this buffer unless the write completes before we finish 1897 * decrementing the reference count. 1898 */ 1899 mutex_enter(&arc_eviction_mtx); 1900 (void) remove_reference(hdr, NULL, tag); 1901 ASSERT(refcount_is_zero(&hdr->b_refcnt)); 1902 destroy_hdr = !HDR_IO_IN_PROGRESS(hdr); 1903 mutex_exit(&arc_eviction_mtx); 1904 if (destroy_hdr) 1905 arc_hdr_destroy(hdr); 1906 } else { 1907 if (remove_reference(hdr, NULL, tag) > 0) 1908 arc_buf_destroy(buf, FALSE, TRUE); 1909 else 1910 arc_hdr_destroy(hdr); 1911 } 1912} 1913 1914boolean_t 1915arc_buf_remove_ref(arc_buf_t *buf, void* tag) 1916{ 1917 arc_buf_hdr_t *hdr = buf->b_hdr; 1918 kmutex_t *hash_lock = HDR_LOCK(hdr); 1919 boolean_t no_callback = (buf->b_efunc == NULL); 1920 1921 if (hdr->b_state == arc_anon) { 1922 ASSERT(hdr->b_datacnt == 1); 1923 arc_buf_free(buf, tag); 1924 return (no_callback); 1925 } 1926 1927 mutex_enter(hash_lock); 1928 hdr = buf->b_hdr; 1929 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 1930 ASSERT(hdr->b_state != arc_anon); 1931 ASSERT(buf->b_data != NULL); 1932 1933 (void) remove_reference(hdr, hash_lock, tag); 1934 if (hdr->b_datacnt > 1) { 1935 if (no_callback) 1936 arc_buf_destroy(buf, FALSE, TRUE); 1937 } else if (no_callback) { 1938 ASSERT(hdr->b_buf == buf && buf->b_next == NULL); 1939 ASSERT(buf->b_efunc == NULL); 1940 hdr->b_flags |= ARC_BUF_AVAILABLE; 1941 } 1942 ASSERT(no_callback || hdr->b_datacnt > 1 || 1943 refcount_is_zero(&hdr->b_refcnt)); 1944 mutex_exit(hash_lock); 1945 return (no_callback); 1946} 1947 1948int 1949arc_buf_size(arc_buf_t *buf) 1950{ 1951 return (buf->b_hdr->b_size); 1952} 1953 1954/* 1955 * Called from the DMU to determine if the current buffer should be 1956 * evicted. In order to ensure proper locking, the eviction must be initiated 1957 * from the DMU. Return true if the buffer is associated with user data and 1958 * duplicate buffers still exist. 1959 */ 1960boolean_t 1961arc_buf_eviction_needed(arc_buf_t *buf) 1962{ 1963 arc_buf_hdr_t *hdr; 1964 boolean_t evict_needed = B_FALSE; 1965 1966 if (zfs_disable_dup_eviction) 1967 return (B_FALSE); 1968 1969 mutex_enter(&buf->b_evict_lock); 1970 hdr = buf->b_hdr; 1971 if (hdr == NULL) { 1972 /* 1973 * We are in arc_do_user_evicts(); let that function 1974 * perform the eviction. 1975 */ 1976 ASSERT(buf->b_data == NULL); 1977 mutex_exit(&buf->b_evict_lock); 1978 return (B_FALSE); 1979 } else if (buf->b_data == NULL) { 1980 /* 1981 * We have already been added to the arc eviction list; 1982 * recommend eviction. 1983 */ 1984 ASSERT3P(hdr, ==, &arc_eviction_hdr); 1985 mutex_exit(&buf->b_evict_lock); 1986 return (B_TRUE); 1987 } 1988 1989 if (hdr->b_datacnt > 1 && hdr->b_type == ARC_BUFC_DATA) 1990 evict_needed = B_TRUE; 1991 1992 mutex_exit(&buf->b_evict_lock); 1993 return (evict_needed); 1994} 1995 1996/* 1997 * Evict buffers from list until we've removed the specified number of 1998 * bytes. Move the removed buffers to the appropriate evict state. 1999 * If the recycle flag is set, then attempt to "recycle" a buffer: 2000 * - look for a buffer to evict that is `bytes' long. 2001 * - return the data block from this buffer rather than freeing it. 2002 * This flag is used by callers that are trying to make space for a 2003 * new buffer in a full arc cache. 2004 * 2005 * This function makes a "best effort". It skips over any buffers 2006 * it can't get a hash_lock on, and so may not catch all candidates. 2007 * It may also return without evicting as much space as requested. 2008 */ 2009static void * 2010arc_evict(arc_state_t *state, uint64_t spa, int64_t bytes, boolean_t recycle, 2011 arc_buf_contents_t type) 2012{ 2013 arc_state_t *evicted_state; 2014 uint64_t bytes_evicted = 0, skipped = 0, missed = 0; 2015 int64_t bytes_remaining; 2016 arc_buf_hdr_t *ab, *ab_prev = NULL; 2017 list_t *evicted_list, *list, *evicted_list_start, *list_start; 2018 kmutex_t *lock, *evicted_lock; 2019 kmutex_t *hash_lock; 2020 boolean_t have_lock; 2021 void *stolen = NULL; 2022 arc_buf_hdr_t marker = { 0 }; 2023 int count = 0; 2024 static int evict_metadata_offset, evict_data_offset; 2025 int i, idx, offset, list_count, lists; 2026 2027 ASSERT(state == arc_mru || state == arc_mfu); 2028 2029 evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost; 2030 2031 if (type == ARC_BUFC_METADATA) { 2032 offset = 0; 2033 list_count = ARC_BUFC_NUMMETADATALISTS; 2034 list_start = &state->arcs_lists[0]; 2035 evicted_list_start = &evicted_state->arcs_lists[0]; 2036 idx = evict_metadata_offset; 2037 } else { 2038 offset = ARC_BUFC_NUMMETADATALISTS; 2039 list_start = &state->arcs_lists[offset]; 2040 evicted_list_start = &evicted_state->arcs_lists[offset]; 2041 list_count = ARC_BUFC_NUMDATALISTS; 2042 idx = evict_data_offset; 2043 } 2044 bytes_remaining = evicted_state->arcs_lsize[type]; 2045 lists = 0; 2046 2047evict_start: 2048 list = &list_start[idx]; 2049 evicted_list = &evicted_list_start[idx]; 2050 lock = ARCS_LOCK(state, (offset + idx)); 2051 evicted_lock = ARCS_LOCK(evicted_state, (offset + idx)); 2052 2053 mutex_enter(lock); 2054 mutex_enter(evicted_lock); 2055 2056 for (ab = list_tail(list); ab; ab = ab_prev) { 2057 ab_prev = list_prev(list, ab); 2058 bytes_remaining -= (ab->b_size * ab->b_datacnt); 2059 /* prefetch buffers have a minimum lifespan */ 2060 if (HDR_IO_IN_PROGRESS(ab) || 2061 (spa && ab->b_spa != spa) || 2062 (ab->b_flags & (ARC_PREFETCH|ARC_INDIRECT) && 2063 ddi_get_lbolt() - ab->b_arc_access < 2064 arc_min_prefetch_lifespan)) { 2065 skipped++; 2066 continue; 2067 } 2068 /* "lookahead" for better eviction candidate */ 2069 if (recycle && ab->b_size != bytes && 2070 ab_prev && ab_prev->b_size == bytes) 2071 continue; 2072 2073 /* ignore markers */ 2074 if (ab->b_spa == 0) 2075 continue; 2076 2077 /* 2078 * It may take a long time to evict all the bufs requested. 2079 * To avoid blocking all arc activity, periodically drop 2080 * the arcs_mtx and give other threads a chance to run 2081 * before reacquiring the lock. 2082 * 2083 * If we are looking for a buffer to recycle, we are in 2084 * the hot code path, so don't sleep. 2085 */ 2086 if (!recycle && count++ > arc_evict_iterations) { 2087 list_insert_after(list, ab, &marker); 2088 mutex_exit(evicted_lock); 2089 mutex_exit(lock); 2090 kpreempt(KPREEMPT_SYNC); 2091 mutex_enter(lock); 2092 mutex_enter(evicted_lock); 2093 ab_prev = list_prev(list, &marker); 2094 list_remove(list, &marker); 2095 count = 0; 2096 continue; 2097 } 2098 2099 hash_lock = HDR_LOCK(ab); 2100 have_lock = MUTEX_HELD(hash_lock); 2101 if (have_lock || mutex_tryenter(hash_lock)) { 2102 ASSERT0(refcount_count(&ab->b_refcnt)); 2103 ASSERT(ab->b_datacnt > 0); 2104 while (ab->b_buf) { 2105 arc_buf_t *buf = ab->b_buf; 2106 if (!mutex_tryenter(&buf->b_evict_lock)) { 2107 missed += 1; 2108 break; 2109 } 2110 if (buf->b_data) { 2111 bytes_evicted += ab->b_size; 2112 if (recycle && ab->b_type == type && 2113 ab->b_size == bytes && 2114 !HDR_L2_WRITING(ab)) { 2115 stolen = buf->b_data; 2116 recycle = FALSE; 2117 } 2118 } 2119 if (buf->b_efunc) { 2120 mutex_enter(&arc_eviction_mtx); 2121 arc_buf_destroy(buf, 2122 buf->b_data == stolen, FALSE); 2123 ab->b_buf = buf->b_next; 2124 buf->b_hdr = &arc_eviction_hdr; 2125 buf->b_next = arc_eviction_list; 2126 arc_eviction_list = buf; 2127 mutex_exit(&arc_eviction_mtx); 2128 mutex_exit(&buf->b_evict_lock); 2129 } else { 2130 mutex_exit(&buf->b_evict_lock); 2131 arc_buf_destroy(buf, 2132 buf->b_data == stolen, TRUE); 2133 } 2134 } 2135 2136 if (ab->b_l2hdr) { 2137 ARCSTAT_INCR(arcstat_evict_l2_cached, 2138 ab->b_size); 2139 } else { 2140 if (l2arc_write_eligible(ab->b_spa, ab)) { 2141 ARCSTAT_INCR(arcstat_evict_l2_eligible, 2142 ab->b_size); 2143 } else { 2144 ARCSTAT_INCR( 2145 arcstat_evict_l2_ineligible, 2146 ab->b_size); 2147 } 2148 } 2149 2150 if (ab->b_datacnt == 0) { 2151 arc_change_state(evicted_state, ab, hash_lock); 2152 ASSERT(HDR_IN_HASH_TABLE(ab)); 2153 ab->b_flags |= ARC_IN_HASH_TABLE; 2154 ab->b_flags &= ~ARC_BUF_AVAILABLE; 2155 DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, ab); 2156 } 2157 if (!have_lock) 2158 mutex_exit(hash_lock); 2159 if (bytes >= 0 && bytes_evicted >= bytes) 2160 break; 2161 if (bytes_remaining > 0) { 2162 mutex_exit(evicted_lock); 2163 mutex_exit(lock); 2164 idx = ((idx + 1) & (list_count - 1)); 2165 lists++; 2166 goto evict_start; 2167 } 2168 } else { 2169 missed += 1; 2170 } 2171 } 2172 2173 mutex_exit(evicted_lock); 2174 mutex_exit(lock); 2175 2176 idx = ((idx + 1) & (list_count - 1)); 2177 lists++; 2178 2179 if (bytes_evicted < bytes) { 2180 if (lists < list_count) 2181 goto evict_start; 2182 else 2183 dprintf("only evicted %lld bytes from %x", 2184 (longlong_t)bytes_evicted, state); 2185 } 2186 if (type == ARC_BUFC_METADATA) 2187 evict_metadata_offset = idx; 2188 else 2189 evict_data_offset = idx; 2190 2191 if (skipped) 2192 ARCSTAT_INCR(arcstat_evict_skip, skipped); 2193 2194 if (missed) 2195 ARCSTAT_INCR(arcstat_mutex_miss, missed); 2196 2197 /* 2198 * Note: we have just evicted some data into the ghost state, 2199 * potentially putting the ghost size over the desired size. Rather 2200 * that evicting from the ghost list in this hot code path, leave 2201 * this chore to the arc_reclaim_thread(). 2202 */ 2203 2204 if (stolen) 2205 ARCSTAT_BUMP(arcstat_stolen); 2206 return (stolen); 2207} 2208 2209/* 2210 * Remove buffers from list until we've removed the specified number of 2211 * bytes. Destroy the buffers that are removed. 2212 */ 2213static void 2214arc_evict_ghost(arc_state_t *state, uint64_t spa, int64_t bytes) 2215{ 2216 arc_buf_hdr_t *ab, *ab_prev; 2217 arc_buf_hdr_t marker = { 0 }; 2218 list_t *list, *list_start; 2219 kmutex_t *hash_lock, *lock; 2220 uint64_t bytes_deleted = 0; 2221 uint64_t bufs_skipped = 0; 2222 int count = 0; 2223 static int evict_offset; 2224 int list_count, idx = evict_offset; 2225 int offset, lists = 0; 2226 2227 ASSERT(GHOST_STATE(state)); 2228 2229 /* 2230 * data lists come after metadata lists 2231 */ 2232 list_start = &state->arcs_lists[ARC_BUFC_NUMMETADATALISTS]; 2233 list_count = ARC_BUFC_NUMDATALISTS; 2234 offset = ARC_BUFC_NUMMETADATALISTS; 2235 2236evict_start: 2237 list = &list_start[idx]; 2238 lock = ARCS_LOCK(state, idx + offset); 2239 2240 mutex_enter(lock); 2241 for (ab = list_tail(list); ab; ab = ab_prev) { 2242 ab_prev = list_prev(list, ab); 2243 if (ab->b_type > ARC_BUFC_NUMTYPES) 2244 panic("invalid ab=%p", (void *)ab); 2245 if (spa && ab->b_spa != spa) 2246 continue; 2247 2248 /* ignore markers */ 2249 if (ab->b_spa == 0) 2250 continue; 2251 2252 hash_lock = HDR_LOCK(ab); 2253 /* caller may be trying to modify this buffer, skip it */ 2254 if (MUTEX_HELD(hash_lock)) 2255 continue; 2256 2257 /* 2258 * It may take a long time to evict all the bufs requested. 2259 * To avoid blocking all arc activity, periodically drop 2260 * the arcs_mtx and give other threads a chance to run 2261 * before reacquiring the lock. 2262 */ 2263 if (count++ > arc_evict_iterations) { 2264 list_insert_after(list, ab, &marker); 2265 mutex_exit(lock); 2266 kpreempt(KPREEMPT_SYNC); 2267 mutex_enter(lock); 2268 ab_prev = list_prev(list, &marker); 2269 list_remove(list, &marker); 2270 count = 0; 2271 continue; 2272 } 2273 if (mutex_tryenter(hash_lock)) { 2274 ASSERT(!HDR_IO_IN_PROGRESS(ab)); 2275 ASSERT(ab->b_buf == NULL); 2276 ARCSTAT_BUMP(arcstat_deleted); 2277 bytes_deleted += ab->b_size; 2278 2279 if (ab->b_l2hdr != NULL) { 2280 /* 2281 * This buffer is cached on the 2nd Level ARC; 2282 * don't destroy the header. 2283 */ 2284 arc_change_state(arc_l2c_only, ab, hash_lock); 2285 mutex_exit(hash_lock); 2286 } else { 2287 arc_change_state(arc_anon, ab, hash_lock); 2288 mutex_exit(hash_lock); 2289 arc_hdr_destroy(ab); 2290 } 2291 2292 DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, ab); 2293 if (bytes >= 0 && bytes_deleted >= bytes) 2294 break; 2295 } else if (bytes < 0) { 2296 /* 2297 * Insert a list marker and then wait for the 2298 * hash lock to become available. Once its 2299 * available, restart from where we left off. 2300 */ 2301 list_insert_after(list, ab, &marker); 2302 mutex_exit(lock); 2303 mutex_enter(hash_lock); 2304 mutex_exit(hash_lock); 2305 mutex_enter(lock); 2306 ab_prev = list_prev(list, &marker); 2307 list_remove(list, &marker); 2308 } else { 2309 bufs_skipped += 1; 2310 } 2311 2312 } 2313 mutex_exit(lock); 2314 idx = ((idx + 1) & (ARC_BUFC_NUMDATALISTS - 1)); 2315 lists++; 2316 2317 if (lists < list_count) 2318 goto evict_start; 2319 2320 evict_offset = idx; 2321 if ((uintptr_t)list > (uintptr_t)&state->arcs_lists[ARC_BUFC_NUMMETADATALISTS] && 2322 (bytes < 0 || bytes_deleted < bytes)) { 2323 list_start = &state->arcs_lists[0]; 2324 list_count = ARC_BUFC_NUMMETADATALISTS; 2325 offset = lists = 0; 2326 goto evict_start; 2327 } 2328 2329 if (bufs_skipped) { 2330 ARCSTAT_INCR(arcstat_mutex_miss, bufs_skipped); 2331 ASSERT(bytes >= 0); 2332 } 2333 2334 if (bytes_deleted < bytes) 2335 dprintf("only deleted %lld bytes from %p", 2336 (longlong_t)bytes_deleted, state); 2337} 2338 2339static void 2340arc_adjust(void) 2341{ 2342 int64_t adjustment, delta; 2343 2344 /* 2345 * Adjust MRU size 2346 */ 2347 2348 adjustment = MIN((int64_t)(arc_size - arc_c), 2349 (int64_t)(arc_anon->arcs_size + arc_mru->arcs_size + arc_meta_used - 2350 arc_p)); 2351 2352 if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_DATA] > 0) { 2353 delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_DATA], adjustment); 2354 (void) arc_evict(arc_mru, 0, delta, FALSE, ARC_BUFC_DATA); 2355 adjustment -= delta; 2356 } 2357 2358 if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_METADATA] > 0) { 2359 delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_METADATA], adjustment); 2360 (void) arc_evict(arc_mru, 0, delta, FALSE, 2361 ARC_BUFC_METADATA); 2362 } 2363 2364 /* 2365 * Adjust MFU size 2366 */ 2367 2368 adjustment = arc_size - arc_c; 2369 2370 if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_DATA] > 0) { 2371 delta = MIN(adjustment, arc_mfu->arcs_lsize[ARC_BUFC_DATA]); 2372 (void) arc_evict(arc_mfu, 0, delta, FALSE, ARC_BUFC_DATA); 2373 adjustment -= delta; 2374 } 2375 2376 if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_METADATA] > 0) { 2377 int64_t delta = MIN(adjustment, 2378 arc_mfu->arcs_lsize[ARC_BUFC_METADATA]); 2379 (void) arc_evict(arc_mfu, 0, delta, FALSE, 2380 ARC_BUFC_METADATA); 2381 } 2382 2383 /* 2384 * Adjust ghost lists 2385 */ 2386 2387 adjustment = arc_mru->arcs_size + arc_mru_ghost->arcs_size - arc_c; 2388 2389 if (adjustment > 0 && arc_mru_ghost->arcs_size > 0) { 2390 delta = MIN(arc_mru_ghost->arcs_size, adjustment); 2391 arc_evict_ghost(arc_mru_ghost, 0, delta); 2392 } 2393 2394 adjustment = 2395 arc_mru_ghost->arcs_size + arc_mfu_ghost->arcs_size - arc_c; 2396 2397 if (adjustment > 0 && arc_mfu_ghost->arcs_size > 0) { 2398 delta = MIN(arc_mfu_ghost->arcs_size, adjustment); 2399 arc_evict_ghost(arc_mfu_ghost, 0, delta); 2400 } 2401} 2402 2403static void 2404arc_do_user_evicts(void) 2405{ 2406 static arc_buf_t *tmp_arc_eviction_list; 2407 2408 /* 2409 * Move list over to avoid LOR 2410 */ 2411restart: 2412 mutex_enter(&arc_eviction_mtx); 2413 tmp_arc_eviction_list = arc_eviction_list; 2414 arc_eviction_list = NULL; 2415 mutex_exit(&arc_eviction_mtx); 2416 2417 while (tmp_arc_eviction_list != NULL) { 2418 arc_buf_t *buf = tmp_arc_eviction_list; 2419 tmp_arc_eviction_list = buf->b_next; 2420 mutex_enter(&buf->b_evict_lock); 2421 buf->b_hdr = NULL; 2422 mutex_exit(&buf->b_evict_lock); 2423 2424 if (buf->b_efunc != NULL) 2425 VERIFY0(buf->b_efunc(buf->b_private)); 2426 2427 buf->b_efunc = NULL; 2428 buf->b_private = NULL; 2429 kmem_cache_free(buf_cache, buf); 2430 } 2431 2432 if (arc_eviction_list != NULL) 2433 goto restart; 2434} 2435 2436/* 2437 * Flush all *evictable* data from the cache for the given spa. 2438 * NOTE: this will not touch "active" (i.e. referenced) data. 2439 */ 2440void 2441arc_flush(spa_t *spa) 2442{ 2443 uint64_t guid = 0; 2444 2445 if (spa) 2446 guid = spa_load_guid(spa); 2447 2448 while (arc_mru->arcs_lsize[ARC_BUFC_DATA]) { 2449 (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_DATA); 2450 if (spa) 2451 break; 2452 } 2453 while (arc_mru->arcs_lsize[ARC_BUFC_METADATA]) { 2454 (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_METADATA); 2455 if (spa) 2456 break; 2457 } 2458 while (arc_mfu->arcs_lsize[ARC_BUFC_DATA]) { 2459 (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_DATA); 2460 if (spa) 2461 break; 2462 } 2463 while (arc_mfu->arcs_lsize[ARC_BUFC_METADATA]) { 2464 (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_METADATA); 2465 if (spa) 2466 break; 2467 } 2468 2469 arc_evict_ghost(arc_mru_ghost, guid, -1); 2470 arc_evict_ghost(arc_mfu_ghost, guid, -1); 2471 2472 mutex_enter(&arc_reclaim_thr_lock); 2473 arc_do_user_evicts(); 2474 mutex_exit(&arc_reclaim_thr_lock); 2475 ASSERT(spa || arc_eviction_list == NULL); 2476} 2477 2478void 2479arc_shrink(void) 2480{ 2481 2482 if (arc_c > arc_c_min) { 2483 uint64_t to_free; 2484 2485 DTRACE_PROBE4(arc__shrink, uint64_t, arc_c, uint64_t, 2486 arc_c_min, uint64_t, arc_p, uint64_t, to_free); 2487#ifdef _KERNEL 2488 to_free = arc_c >> arc_shrink_shift; 2489#else 2490 to_free = arc_c >> arc_shrink_shift; 2491#endif 2492 if (arc_c > arc_c_min + to_free) 2493 atomic_add_64(&arc_c, -to_free); 2494 else 2495 arc_c = arc_c_min; 2496 2497 atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift)); 2498 if (arc_c > arc_size) 2499 arc_c = MAX(arc_size, arc_c_min); 2500 if (arc_p > arc_c) 2501 arc_p = (arc_c >> 1); 2502 2503 DTRACE_PROBE2(arc__shrunk, uint64_t, arc_c, uint64_t, 2504 arc_p); 2505 2506 ASSERT(arc_c >= arc_c_min); 2507 ASSERT((int64_t)arc_p >= 0); 2508 } 2509 2510 if (arc_size > arc_c) { 2511 DTRACE_PROBE2(arc__shrink_adjust, uint64_t, arc_size, 2512 uint64_t, arc_c); 2513 arc_adjust(); 2514 } 2515} 2516 2517static int needfree = 0; 2518 2519static int 2520arc_reclaim_needed(void) 2521{ 2522 2523#ifdef _KERNEL 2524 2525 if (needfree) { 2526 DTRACE_PROBE(arc__reclaim_needfree); 2527 return (1); 2528 } 2529 2530 /* 2531 * Cooperate with pagedaemon when it's time for it to scan 2532 * and reclaim some pages. 2533 */ 2534 if (freemem < zfs_arc_free_target) { 2535 DTRACE_PROBE2(arc__reclaim_freemem, uint64_t, 2536 freemem, uint64_t, zfs_arc_free_target); 2537 return (1); 2538 } 2539 2540#ifdef sun 2541 /* 2542 * take 'desfree' extra pages, so we reclaim sooner, rather than later 2543 */ 2544 extra = desfree; 2545 2546 /* 2547 * check that we're out of range of the pageout scanner. It starts to 2548 * schedule paging if freemem is less than lotsfree and needfree. 2549 * lotsfree is the high-water mark for pageout, and needfree is the 2550 * number of needed free pages. We add extra pages here to make sure 2551 * the scanner doesn't start up while we're freeing memory. 2552 */ 2553 if (freemem < lotsfree + needfree + extra) 2554 return (1); 2555 2556 /* 2557 * check to make sure that swapfs has enough space so that anon 2558 * reservations can still succeed. anon_resvmem() checks that the 2559 * availrmem is greater than swapfs_minfree, and the number of reserved 2560 * swap pages. We also add a bit of extra here just to prevent 2561 * circumstances from getting really dire. 2562 */ 2563 if (availrmem < swapfs_minfree + swapfs_reserve + extra) 2564 return (1); 2565 2566 /* 2567 * Check that we have enough availrmem that memory locking (e.g., via 2568 * mlock(3C) or memcntl(2)) can still succeed. (pages_pp_maximum 2569 * stores the number of pages that cannot be locked; when availrmem 2570 * drops below pages_pp_maximum, page locking mechanisms such as 2571 * page_pp_lock() will fail.) 2572 */ 2573 if (availrmem <= pages_pp_maximum) 2574 return (1); 2575 2576#endif /* sun */ 2577#if defined(__i386) || !defined(UMA_MD_SMALL_ALLOC) 2578 /* 2579 * If we're on an i386 platform, it's possible that we'll exhaust the 2580 * kernel heap space before we ever run out of available physical 2581 * memory. Most checks of the size of the heap_area compare against 2582 * tune.t_minarmem, which is the minimum available real memory that we 2583 * can have in the system. However, this is generally fixed at 25 pages 2584 * which is so low that it's useless. In this comparison, we seek to 2585 * calculate the total heap-size, and reclaim if more than 3/4ths of the 2586 * heap is allocated. (Or, in the calculation, if less than 1/4th is 2587 * free) 2588 */ 2589 if (vmem_size(heap_arena, VMEM_FREE) < 2590 (vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC) >> 2)) { 2591 DTRACE_PROBE2(arc__reclaim_used, uint64_t, 2592 vmem_size(heap_arena, VMEM_FREE), uint64_t, 2593 (vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC)) >> 2); 2594 return (1); 2595 } 2596#endif 2597#ifdef sun 2598 /* 2599 * If zio data pages are being allocated out of a separate heap segment, 2600 * then enforce that the size of available vmem for this arena remains 2601 * above about 1/16th free. 2602 * 2603 * Note: The 1/16th arena free requirement was put in place 2604 * to aggressively evict memory from the arc in order to avoid 2605 * memory fragmentation issues. 2606 */ 2607 if (zio_arena != NULL && 2608 vmem_size(zio_arena, VMEM_FREE) < 2609 (vmem_size(zio_arena, VMEM_ALLOC) >> 4)) 2610 return (1); 2611#endif /* sun */ 2612#else /* _KERNEL */ 2613 if (spa_get_random(100) == 0) 2614 return (1); 2615#endif /* _KERNEL */ 2616 DTRACE_PROBE(arc__reclaim_no); 2617 2618 return (0); 2619} 2620 2621extern kmem_cache_t *zio_buf_cache[]; 2622extern kmem_cache_t *zio_data_buf_cache[]; 2623extern kmem_cache_t *range_seg_cache; 2624 2625static void __noinline 2626arc_kmem_reap_now(arc_reclaim_strategy_t strat) 2627{ 2628 size_t i; 2629 kmem_cache_t *prev_cache = NULL; 2630 kmem_cache_t *prev_data_cache = NULL; 2631 2632 DTRACE_PROBE(arc__kmem_reap_start); 2633#ifdef _KERNEL 2634 if (arc_meta_used >= arc_meta_limit) { 2635 /* 2636 * We are exceeding our meta-data cache limit. 2637 * Purge some DNLC entries to release holds on meta-data. 2638 */ 2639 dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent); 2640 } 2641#if defined(__i386) 2642 /* 2643 * Reclaim unused memory from all kmem caches. 2644 */ 2645 kmem_reap(); 2646#endif 2647#endif 2648 2649 /* 2650 * An aggressive reclamation will shrink the cache size as well as 2651 * reap free buffers from the arc kmem caches. 2652 */ 2653 if (strat == ARC_RECLAIM_AGGR) 2654 arc_shrink(); 2655 2656 for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) { 2657 if (zio_buf_cache[i] != prev_cache) { 2658 prev_cache = zio_buf_cache[i]; 2659 kmem_cache_reap_now(zio_buf_cache[i]); 2660 } 2661 if (zio_data_buf_cache[i] != prev_data_cache) { 2662 prev_data_cache = zio_data_buf_cache[i]; 2663 kmem_cache_reap_now(zio_data_buf_cache[i]); 2664 } 2665 } 2666 kmem_cache_reap_now(buf_cache); 2667 kmem_cache_reap_now(hdr_cache); 2668 kmem_cache_reap_now(range_seg_cache); 2669 2670#ifdef sun 2671 /* 2672 * Ask the vmem arena to reclaim unused memory from its 2673 * quantum caches. 2674 */ 2675 if (zio_arena != NULL && strat == ARC_RECLAIM_AGGR) 2676 vmem_qcache_reap(zio_arena); 2677#endif 2678 DTRACE_PROBE(arc__kmem_reap_end); 2679} 2680 2681static void 2682arc_reclaim_thread(void *dummy __unused) 2683{ 2684 clock_t growtime = 0; 2685 arc_reclaim_strategy_t last_reclaim = ARC_RECLAIM_CONS; 2686 callb_cpr_t cpr; 2687 2688 CALLB_CPR_INIT(&cpr, &arc_reclaim_thr_lock, callb_generic_cpr, FTAG); 2689 2690 mutex_enter(&arc_reclaim_thr_lock); 2691 while (arc_thread_exit == 0) { 2692 if (arc_reclaim_needed()) { 2693 2694 if (arc_no_grow) { 2695 if (last_reclaim == ARC_RECLAIM_CONS) { 2696 DTRACE_PROBE(arc__reclaim_aggr_no_grow); 2697 last_reclaim = ARC_RECLAIM_AGGR; 2698 } else { 2699 last_reclaim = ARC_RECLAIM_CONS; 2700 } 2701 } else { 2702 arc_no_grow = TRUE; 2703 last_reclaim = ARC_RECLAIM_AGGR; 2704 DTRACE_PROBE(arc__reclaim_aggr); 2705 membar_producer(); 2706 } 2707 2708 /* reset the growth delay for every reclaim */ 2709 growtime = ddi_get_lbolt() + (arc_grow_retry * hz); 2710 2711 if (needfree && last_reclaim == ARC_RECLAIM_CONS) { 2712 /* 2713 * If needfree is TRUE our vm_lowmem hook 2714 * was called and in that case we must free some 2715 * memory, so switch to aggressive mode. 2716 */ 2717 arc_no_grow = TRUE; 2718 last_reclaim = ARC_RECLAIM_AGGR; 2719 } 2720 arc_kmem_reap_now(last_reclaim); 2721 arc_warm = B_TRUE; 2722 2723 } else if (arc_no_grow && ddi_get_lbolt() >= growtime) { 2724 arc_no_grow = FALSE; 2725 } 2726 2727 arc_adjust(); 2728 2729 if (arc_eviction_list != NULL) 2730 arc_do_user_evicts(); 2731 2732#ifdef _KERNEL 2733 if (needfree) { 2734 needfree = 0; 2735 wakeup(&needfree); 2736 } 2737#endif 2738 2739 /* block until needed, or one second, whichever is shorter */ 2740 CALLB_CPR_SAFE_BEGIN(&cpr); 2741 (void) cv_timedwait(&arc_reclaim_thr_cv, 2742 &arc_reclaim_thr_lock, hz); 2743 CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_thr_lock); 2744 } 2745 2746 arc_thread_exit = 0; 2747 cv_broadcast(&arc_reclaim_thr_cv); 2748 CALLB_CPR_EXIT(&cpr); /* drops arc_reclaim_thr_lock */ 2749 thread_exit(); 2750} 2751 2752/* 2753 * Adapt arc info given the number of bytes we are trying to add and 2754 * the state that we are comming from. This function is only called 2755 * when we are adding new content to the cache. 2756 */ 2757static void 2758arc_adapt(int bytes, arc_state_t *state) 2759{ 2760 int mult; 2761 uint64_t arc_p_min = (arc_c >> arc_p_min_shift); 2762 2763 if (state == arc_l2c_only) 2764 return; 2765 2766 ASSERT(bytes > 0); 2767 /* 2768 * Adapt the target size of the MRU list: 2769 * - if we just hit in the MRU ghost list, then increase 2770 * the target size of the MRU list. 2771 * - if we just hit in the MFU ghost list, then increase 2772 * the target size of the MFU list by decreasing the 2773 * target size of the MRU list. 2774 */ 2775 if (state == arc_mru_ghost) { 2776 mult = ((arc_mru_ghost->arcs_size >= arc_mfu_ghost->arcs_size) ? 2777 1 : (arc_mfu_ghost->arcs_size/arc_mru_ghost->arcs_size)); 2778 mult = MIN(mult, 10); /* avoid wild arc_p adjustment */ 2779 2780 arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult); 2781 } else if (state == arc_mfu_ghost) { 2782 uint64_t delta; 2783 2784 mult = ((arc_mfu_ghost->arcs_size >= arc_mru_ghost->arcs_size) ? 2785 1 : (arc_mru_ghost->arcs_size/arc_mfu_ghost->arcs_size)); 2786 mult = MIN(mult, 10); 2787 2788 delta = MIN(bytes * mult, arc_p); 2789 arc_p = MAX(arc_p_min, arc_p - delta); 2790 } 2791 ASSERT((int64_t)arc_p >= 0); 2792 2793 if (arc_reclaim_needed()) { 2794 cv_signal(&arc_reclaim_thr_cv); 2795 return; 2796 } 2797 2798 if (arc_no_grow) 2799 return; 2800 2801 if (arc_c >= arc_c_max) 2802 return; 2803 2804 /* 2805 * If we're within (2 * maxblocksize) bytes of the target 2806 * cache size, increment the target cache size 2807 */ 2808 if (arc_size > arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) { 2809 DTRACE_PROBE1(arc__inc_adapt, int, bytes); 2810 atomic_add_64(&arc_c, (int64_t)bytes); 2811 if (arc_c > arc_c_max) 2812 arc_c = arc_c_max; 2813 else if (state == arc_anon) 2814 atomic_add_64(&arc_p, (int64_t)bytes); 2815 if (arc_p > arc_c) 2816 arc_p = arc_c; 2817 } 2818 ASSERT((int64_t)arc_p >= 0); 2819} 2820 2821/* 2822 * Check if the cache has reached its limits and eviction is required 2823 * prior to insert. 2824 */ 2825static int 2826arc_evict_needed(arc_buf_contents_t type) 2827{ 2828 if (type == ARC_BUFC_METADATA && arc_meta_used >= arc_meta_limit) 2829 return (1); 2830 2831 if (arc_reclaim_needed()) 2832 return (1); 2833 2834 return (arc_size > arc_c); 2835} 2836 2837/* 2838 * The buffer, supplied as the first argument, needs a data block. 2839 * So, if we are at cache max, determine which cache should be victimized. 2840 * We have the following cases: 2841 * 2842 * 1. Insert for MRU, p > sizeof(arc_anon + arc_mru) -> 2843 * In this situation if we're out of space, but the resident size of the MFU is 2844 * under the limit, victimize the MFU cache to satisfy this insertion request. 2845 * 2846 * 2. Insert for MRU, p <= sizeof(arc_anon + arc_mru) -> 2847 * Here, we've used up all of the available space for the MRU, so we need to 2848 * evict from our own cache instead. Evict from the set of resident MRU 2849 * entries. 2850 * 2851 * 3. Insert for MFU (c - p) > sizeof(arc_mfu) -> 2852 * c minus p represents the MFU space in the cache, since p is the size of the 2853 * cache that is dedicated to the MRU. In this situation there's still space on 2854 * the MFU side, so the MRU side needs to be victimized. 2855 * 2856 * 4. Insert for MFU (c - p) < sizeof(arc_mfu) -> 2857 * MFU's resident set is consuming more space than it has been allotted. In 2858 * this situation, we must victimize our own cache, the MFU, for this insertion. 2859 */ 2860static void 2861arc_get_data_buf(arc_buf_t *buf) 2862{ 2863 arc_state_t *state = buf->b_hdr->b_state; 2864 uint64_t size = buf->b_hdr->b_size; 2865 arc_buf_contents_t type = buf->b_hdr->b_type; 2866 2867 arc_adapt(size, state); 2868 2869 /* 2870 * We have not yet reached cache maximum size, 2871 * just allocate a new buffer. 2872 */ 2873 if (!arc_evict_needed(type)) { 2874 if (type == ARC_BUFC_METADATA) { 2875 buf->b_data = zio_buf_alloc(size); 2876 arc_space_consume(size, ARC_SPACE_DATA); 2877 } else { 2878 ASSERT(type == ARC_BUFC_DATA); 2879 buf->b_data = zio_data_buf_alloc(size); 2880 ARCSTAT_INCR(arcstat_data_size, size); 2881 atomic_add_64(&arc_size, size); 2882 } 2883 goto out; 2884 } 2885 2886 /* 2887 * If we are prefetching from the mfu ghost list, this buffer 2888 * will end up on the mru list; so steal space from there. 2889 */ 2890 if (state == arc_mfu_ghost) 2891 state = buf->b_hdr->b_flags & ARC_PREFETCH ? arc_mru : arc_mfu; 2892 else if (state == arc_mru_ghost) 2893 state = arc_mru; 2894 2895 if (state == arc_mru || state == arc_anon) { 2896 uint64_t mru_used = arc_anon->arcs_size + arc_mru->arcs_size; 2897 state = (arc_mfu->arcs_lsize[type] >= size && 2898 arc_p > mru_used) ? arc_mfu : arc_mru; 2899 } else { 2900 /* MFU cases */ 2901 uint64_t mfu_space = arc_c - arc_p; 2902 state = (arc_mru->arcs_lsize[type] >= size && 2903 mfu_space > arc_mfu->arcs_size) ? arc_mru : arc_mfu; 2904 } 2905 if ((buf->b_data = arc_evict(state, 0, size, TRUE, type)) == NULL) { 2906 if (type == ARC_BUFC_METADATA) { 2907 buf->b_data = zio_buf_alloc(size); 2908 arc_space_consume(size, ARC_SPACE_DATA); 2909 } else { 2910 ASSERT(type == ARC_BUFC_DATA); 2911 buf->b_data = zio_data_buf_alloc(size); 2912 ARCSTAT_INCR(arcstat_data_size, size); 2913 atomic_add_64(&arc_size, size); 2914 } 2915 ARCSTAT_BUMP(arcstat_recycle_miss); 2916 } 2917 ASSERT(buf->b_data != NULL); 2918out: 2919 /* 2920 * Update the state size. Note that ghost states have a 2921 * "ghost size" and so don't need to be updated. 2922 */ 2923 if (!GHOST_STATE(buf->b_hdr->b_state)) { 2924 arc_buf_hdr_t *hdr = buf->b_hdr; 2925 2926 atomic_add_64(&hdr->b_state->arcs_size, size); 2927 if (list_link_active(&hdr->b_arc_node)) { 2928 ASSERT(refcount_is_zero(&hdr->b_refcnt)); 2929 atomic_add_64(&hdr->b_state->arcs_lsize[type], size); 2930 } 2931 /* 2932 * If we are growing the cache, and we are adding anonymous 2933 * data, and we have outgrown arc_p, update arc_p 2934 */ 2935 if (arc_size < arc_c && hdr->b_state == arc_anon && 2936 arc_anon->arcs_size + arc_mru->arcs_size > arc_p) 2937 arc_p = MIN(arc_c, arc_p + size); 2938 } 2939 ARCSTAT_BUMP(arcstat_allocated); 2940} 2941 2942/* 2943 * This routine is called whenever a buffer is accessed. 2944 * NOTE: the hash lock is dropped in this function. 2945 */ 2946static void 2947arc_access(arc_buf_hdr_t *buf, kmutex_t *hash_lock) 2948{ 2949 clock_t now; 2950 2951 ASSERT(MUTEX_HELD(hash_lock)); 2952 2953 if (buf->b_state == arc_anon) { 2954 /* 2955 * This buffer is not in the cache, and does not 2956 * appear in our "ghost" list. Add the new buffer 2957 * to the MRU state. 2958 */ 2959 2960 ASSERT(buf->b_arc_access == 0); 2961 buf->b_arc_access = ddi_get_lbolt(); 2962 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, buf); 2963 arc_change_state(arc_mru, buf, hash_lock); 2964 2965 } else if (buf->b_state == arc_mru) { 2966 now = ddi_get_lbolt(); 2967 2968 /* 2969 * If this buffer is here because of a prefetch, then either: 2970 * - clear the flag if this is a "referencing" read 2971 * (any subsequent access will bump this into the MFU state). 2972 * or 2973 * - move the buffer to the head of the list if this is 2974 * another prefetch (to make it less likely to be evicted). 2975 */ 2976 if ((buf->b_flags & ARC_PREFETCH) != 0) { 2977 if (refcount_count(&buf->b_refcnt) == 0) { 2978 ASSERT(list_link_active(&buf->b_arc_node)); 2979 } else { 2980 buf->b_flags &= ~ARC_PREFETCH; 2981 ARCSTAT_BUMP(arcstat_mru_hits); 2982 } 2983 buf->b_arc_access = now; 2984 return; 2985 } 2986 2987 /* 2988 * This buffer has been "accessed" only once so far, 2989 * but it is still in the cache. Move it to the MFU 2990 * state. 2991 */ 2992 if (now > buf->b_arc_access + ARC_MINTIME) { 2993 /* 2994 * More than 125ms have passed since we 2995 * instantiated this buffer. Move it to the 2996 * most frequently used state. 2997 */ 2998 buf->b_arc_access = now; 2999 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf); 3000 arc_change_state(arc_mfu, buf, hash_lock); 3001 } 3002 ARCSTAT_BUMP(arcstat_mru_hits); 3003 } else if (buf->b_state == arc_mru_ghost) { 3004 arc_state_t *new_state; 3005 /* 3006 * This buffer has been "accessed" recently, but 3007 * was evicted from the cache. Move it to the 3008 * MFU state. 3009 */ 3010 3011 if (buf->b_flags & ARC_PREFETCH) { 3012 new_state = arc_mru; 3013 if (refcount_count(&buf->b_refcnt) > 0) 3014 buf->b_flags &= ~ARC_PREFETCH; 3015 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, buf); 3016 } else { 3017 new_state = arc_mfu; 3018 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf); 3019 } 3020 3021 buf->b_arc_access = ddi_get_lbolt(); 3022 arc_change_state(new_state, buf, hash_lock); 3023 3024 ARCSTAT_BUMP(arcstat_mru_ghost_hits); 3025 } else if (buf->b_state == arc_mfu) { 3026 /* 3027 * This buffer has been accessed more than once and is 3028 * still in the cache. Keep it in the MFU state. 3029 * 3030 * NOTE: an add_reference() that occurred when we did 3031 * the arc_read() will have kicked this off the list. 3032 * If it was a prefetch, we will explicitly move it to 3033 * the head of the list now. 3034 */ 3035 if ((buf->b_flags & ARC_PREFETCH) != 0) { 3036 ASSERT(refcount_count(&buf->b_refcnt) == 0); 3037 ASSERT(list_link_active(&buf->b_arc_node)); 3038 } 3039 ARCSTAT_BUMP(arcstat_mfu_hits); 3040 buf->b_arc_access = ddi_get_lbolt(); 3041 } else if (buf->b_state == arc_mfu_ghost) { 3042 arc_state_t *new_state = arc_mfu; 3043 /* 3044 * This buffer has been accessed more than once but has 3045 * been evicted from the cache. Move it back to the 3046 * MFU state. 3047 */ 3048 3049 if (buf->b_flags & ARC_PREFETCH) { 3050 /* 3051 * This is a prefetch access... 3052 * move this block back to the MRU state. 3053 */ 3054 ASSERT0(refcount_count(&buf->b_refcnt)); 3055 new_state = arc_mru; 3056 } 3057 3058 buf->b_arc_access = ddi_get_lbolt(); 3059 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf); 3060 arc_change_state(new_state, buf, hash_lock); 3061 3062 ARCSTAT_BUMP(arcstat_mfu_ghost_hits); 3063 } else if (buf->b_state == arc_l2c_only) { 3064 /* 3065 * This buffer is on the 2nd Level ARC. 3066 */ 3067 3068 buf->b_arc_access = ddi_get_lbolt(); 3069 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf); 3070 arc_change_state(arc_mfu, buf, hash_lock); 3071 } else { 3072 ASSERT(!"invalid arc state"); 3073 } 3074} 3075 3076/* a generic arc_done_func_t which you can use */ 3077/* ARGSUSED */ 3078void 3079arc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg) 3080{ 3081 if (zio == NULL || zio->io_error == 0) 3082 bcopy(buf->b_data, arg, buf->b_hdr->b_size); 3083 VERIFY(arc_buf_remove_ref(buf, arg)); 3084} 3085 3086/* a generic arc_done_func_t */ 3087void 3088arc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg) 3089{ 3090 arc_buf_t **bufp = arg; 3091 if (zio && zio->io_error) { 3092 VERIFY(arc_buf_remove_ref(buf, arg)); 3093 *bufp = NULL; 3094 } else { 3095 *bufp = buf; 3096 ASSERT(buf->b_data); 3097 } 3098} 3099 3100static void 3101arc_read_done(zio_t *zio) 3102{ 3103 arc_buf_hdr_t *hdr; 3104 arc_buf_t *buf; 3105 arc_buf_t *abuf; /* buffer we're assigning to callback */ 3106 kmutex_t *hash_lock = NULL; 3107 arc_callback_t *callback_list, *acb; 3108 int freeable = FALSE; 3109 3110 buf = zio->io_private; 3111 hdr = buf->b_hdr; 3112 3113 /* 3114 * The hdr was inserted into hash-table and removed from lists 3115 * prior to starting I/O. We should find this header, since 3116 * it's in the hash table, and it should be legit since it's 3117 * not possible to evict it during the I/O. The only possible 3118 * reason for it not to be found is if we were freed during the 3119 * read. 3120 */ 3121 if (HDR_IN_HASH_TABLE(hdr)) { 3122 ASSERT3U(hdr->b_birth, ==, BP_PHYSICAL_BIRTH(zio->io_bp)); 3123 ASSERT3U(hdr->b_dva.dva_word[0], ==, 3124 BP_IDENTITY(zio->io_bp)->dva_word[0]); 3125 ASSERT3U(hdr->b_dva.dva_word[1], ==, 3126 BP_IDENTITY(zio->io_bp)->dva_word[1]); 3127 3128 arc_buf_hdr_t *found = buf_hash_find(hdr->b_spa, zio->io_bp, 3129 &hash_lock); 3130 3131 ASSERT((found == NULL && HDR_FREED_IN_READ(hdr) && 3132 hash_lock == NULL) || 3133 (found == hdr && 3134 DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) || 3135 (found == hdr && HDR_L2_READING(hdr))); 3136 } 3137 3138 hdr->b_flags &= ~ARC_L2_EVICTED; 3139 if (l2arc_noprefetch && (hdr->b_flags & ARC_PREFETCH)) 3140 hdr->b_flags &= ~ARC_L2CACHE; 3141 3142 /* byteswap if necessary */ 3143 callback_list = hdr->b_acb; 3144 ASSERT(callback_list != NULL); 3145 if (BP_SHOULD_BYTESWAP(zio->io_bp) && zio->io_error == 0) { 3146 dmu_object_byteswap_t bswap = 3147 DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp)); 3148 arc_byteswap_func_t *func = BP_GET_LEVEL(zio->io_bp) > 0 ? 3149 byteswap_uint64_array : 3150 dmu_ot_byteswap[bswap].ob_func; 3151 func(buf->b_data, hdr->b_size); 3152 } 3153 3154 arc_cksum_compute(buf, B_FALSE); 3155#ifdef illumos 3156 arc_buf_watch(buf); 3157#endif /* illumos */ 3158 3159 if (hash_lock && zio->io_error == 0 && hdr->b_state == arc_anon) { 3160 /* 3161 * Only call arc_access on anonymous buffers. This is because 3162 * if we've issued an I/O for an evicted buffer, we've already 3163 * called arc_access (to prevent any simultaneous readers from 3164 * getting confused). 3165 */ 3166 arc_access(hdr, hash_lock); 3167 } 3168 3169 /* create copies of the data buffer for the callers */ 3170 abuf = buf; 3171 for (acb = callback_list; acb; acb = acb->acb_next) { 3172 if (acb->acb_done) { 3173 if (abuf == NULL) { 3174 ARCSTAT_BUMP(arcstat_duplicate_reads); 3175 abuf = arc_buf_clone(buf); 3176 } 3177 acb->acb_buf = abuf; 3178 abuf = NULL; 3179 } 3180 } 3181 hdr->b_acb = NULL; 3182 hdr->b_flags &= ~ARC_IO_IN_PROGRESS; 3183 ASSERT(!HDR_BUF_AVAILABLE(hdr)); 3184 if (abuf == buf) { 3185 ASSERT(buf->b_efunc == NULL); 3186 ASSERT(hdr->b_datacnt == 1); 3187 hdr->b_flags |= ARC_BUF_AVAILABLE; 3188 } 3189 3190 ASSERT(refcount_is_zero(&hdr->b_refcnt) || callback_list != NULL); 3191 3192 if (zio->io_error != 0) { 3193 hdr->b_flags |= ARC_IO_ERROR; 3194 if (hdr->b_state != arc_anon) 3195 arc_change_state(arc_anon, hdr, hash_lock); 3196 if (HDR_IN_HASH_TABLE(hdr)) 3197 buf_hash_remove(hdr); 3198 freeable = refcount_is_zero(&hdr->b_refcnt); 3199 } 3200 3201 /* 3202 * Broadcast before we drop the hash_lock to avoid the possibility 3203 * that the hdr (and hence the cv) might be freed before we get to 3204 * the cv_broadcast(). 3205 */ 3206 cv_broadcast(&hdr->b_cv); 3207 3208 if (hash_lock) { 3209 mutex_exit(hash_lock); 3210 } else { 3211 /* 3212 * This block was freed while we waited for the read to 3213 * complete. It has been removed from the hash table and 3214 * moved to the anonymous state (so that it won't show up 3215 * in the cache). 3216 */ 3217 ASSERT3P(hdr->b_state, ==, arc_anon); 3218 freeable = refcount_is_zero(&hdr->b_refcnt); 3219 } 3220 3221 /* execute each callback and free its structure */ 3222 while ((acb = callback_list) != NULL) { 3223 if (acb->acb_done) 3224 acb->acb_done(zio, acb->acb_buf, acb->acb_private); 3225 3226 if (acb->acb_zio_dummy != NULL) { 3227 acb->acb_zio_dummy->io_error = zio->io_error; 3228 zio_nowait(acb->acb_zio_dummy); 3229 } 3230 3231 callback_list = acb->acb_next; 3232 kmem_free(acb, sizeof (arc_callback_t)); 3233 } 3234 3235 if (freeable) 3236 arc_hdr_destroy(hdr); 3237} 3238 3239/* 3240 * "Read" the block block at the specified DVA (in bp) via the 3241 * cache. If the block is found in the cache, invoke the provided 3242 * callback immediately and return. Note that the `zio' parameter 3243 * in the callback will be NULL in this case, since no IO was 3244 * required. If the block is not in the cache pass the read request 3245 * on to the spa with a substitute callback function, so that the 3246 * requested block will be added to the cache. 3247 * 3248 * If a read request arrives for a block that has a read in-progress, 3249 * either wait for the in-progress read to complete (and return the 3250 * results); or, if this is a read with a "done" func, add a record 3251 * to the read to invoke the "done" func when the read completes, 3252 * and return; or just return. 3253 * 3254 * arc_read_done() will invoke all the requested "done" functions 3255 * for readers of this block. 3256 */ 3257int 3258arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_done_func_t *done, 3259 void *private, zio_priority_t priority, int zio_flags, uint32_t *arc_flags, 3260 const zbookmark_phys_t *zb) 3261{ 3262 arc_buf_hdr_t *hdr = NULL; 3263 arc_buf_t *buf = NULL; 3264 kmutex_t *hash_lock = NULL; 3265 zio_t *rzio; 3266 uint64_t guid = spa_load_guid(spa); 3267 3268 ASSERT(!BP_IS_EMBEDDED(bp) || 3269 BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA); 3270 3271top: 3272 if (!BP_IS_EMBEDDED(bp)) { 3273 /* 3274 * Embedded BP's have no DVA and require no I/O to "read". 3275 * Create an anonymous arc buf to back it. 3276 */ 3277 hdr = buf_hash_find(guid, bp, &hash_lock); 3278 } 3279 3280 if (hdr != NULL && hdr->b_datacnt > 0) { 3281 3282 *arc_flags |= ARC_CACHED; 3283 3284 if (HDR_IO_IN_PROGRESS(hdr)) { 3285 3286 if (*arc_flags & ARC_WAIT) { 3287 cv_wait(&hdr->b_cv, hash_lock); 3288 mutex_exit(hash_lock); 3289 goto top; 3290 } 3291 ASSERT(*arc_flags & ARC_NOWAIT); 3292 3293 if (done) { 3294 arc_callback_t *acb = NULL; 3295 3296 acb = kmem_zalloc(sizeof (arc_callback_t), 3297 KM_SLEEP); 3298 acb->acb_done = done; 3299 acb->acb_private = private; 3300 if (pio != NULL) 3301 acb->acb_zio_dummy = zio_null(pio, 3302 spa, NULL, NULL, NULL, zio_flags); 3303 3304 ASSERT(acb->acb_done != NULL); 3305 acb->acb_next = hdr->b_acb; 3306 hdr->b_acb = acb; 3307 add_reference(hdr, hash_lock, private); 3308 mutex_exit(hash_lock); 3309 return (0); 3310 } 3311 mutex_exit(hash_lock); 3312 return (0); 3313 } 3314 3315 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu); 3316 3317 if (done) { 3318 add_reference(hdr, hash_lock, private); 3319 /* 3320 * If this block is already in use, create a new 3321 * copy of the data so that we will be guaranteed 3322 * that arc_release() will always succeed. 3323 */ 3324 buf = hdr->b_buf; 3325 ASSERT(buf); 3326 ASSERT(buf->b_data); 3327 if (HDR_BUF_AVAILABLE(hdr)) { 3328 ASSERT(buf->b_efunc == NULL); 3329 hdr->b_flags &= ~ARC_BUF_AVAILABLE; 3330 } else { 3331 buf = arc_buf_clone(buf); 3332 } 3333 3334 } else if (*arc_flags & ARC_PREFETCH && 3335 refcount_count(&hdr->b_refcnt) == 0) { 3336 hdr->b_flags |= ARC_PREFETCH; 3337 } 3338 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr); 3339 arc_access(hdr, hash_lock); 3340 if (*arc_flags & ARC_L2CACHE) 3341 hdr->b_flags |= ARC_L2CACHE; 3342 if (*arc_flags & ARC_L2COMPRESS) 3343 hdr->b_flags |= ARC_L2COMPRESS; 3344 mutex_exit(hash_lock); 3345 ARCSTAT_BUMP(arcstat_hits); 3346 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH), 3347 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA, 3348 data, metadata, hits); 3349 3350 if (done) 3351 done(NULL, buf, private); 3352 } else { 3353 uint64_t size = BP_GET_LSIZE(bp); 3354 arc_callback_t *acb; 3355 vdev_t *vd = NULL; 3356 uint64_t addr = 0; 3357 boolean_t devw = B_FALSE; 3358 enum zio_compress b_compress = ZIO_COMPRESS_OFF; 3359 uint64_t b_asize = 0; 3360 3361 if (hdr == NULL) { 3362 /* this block is not in the cache */ 3363 arc_buf_hdr_t *exists = NULL; 3364 arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp); 3365 buf = arc_buf_alloc(spa, size, private, type); 3366 hdr = buf->b_hdr; 3367 if (!BP_IS_EMBEDDED(bp)) { 3368 hdr->b_dva = *BP_IDENTITY(bp); 3369 hdr->b_birth = BP_PHYSICAL_BIRTH(bp); 3370 hdr->b_cksum0 = bp->blk_cksum.zc_word[0]; 3371 exists = buf_hash_insert(hdr, &hash_lock); 3372 } 3373 if (exists != NULL) { 3374 /* somebody beat us to the hash insert */ 3375 mutex_exit(hash_lock); 3376 buf_discard_identity(hdr); 3377 (void) arc_buf_remove_ref(buf, private); 3378 goto top; /* restart the IO request */ 3379 } 3380 /* if this is a prefetch, we don't have a reference */ 3381 if (*arc_flags & ARC_PREFETCH) { 3382 (void) remove_reference(hdr, hash_lock, 3383 private); 3384 hdr->b_flags |= ARC_PREFETCH; 3385 } 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 if (BP_GET_LEVEL(bp) > 0) 3391 hdr->b_flags |= ARC_INDIRECT; 3392 } else { 3393 /* this block is in the ghost cache */ 3394 ASSERT(GHOST_STATE(hdr->b_state)); 3395 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 3396 ASSERT0(refcount_count(&hdr->b_refcnt)); 3397 ASSERT(hdr->b_buf == NULL); 3398 3399 /* if this is a prefetch, we don't have a reference */ 3400 if (*arc_flags & ARC_PREFETCH) 3401 hdr->b_flags |= ARC_PREFETCH; 3402 else 3403 add_reference(hdr, hash_lock, private); 3404 if (*arc_flags & ARC_L2CACHE) 3405 hdr->b_flags |= ARC_L2CACHE; 3406 if (*arc_flags & ARC_L2COMPRESS) 3407 hdr->b_flags |= ARC_L2COMPRESS; 3408 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 3409 buf->b_hdr = hdr; 3410 buf->b_data = NULL; 3411 buf->b_efunc = NULL; 3412 buf->b_private = NULL; 3413 buf->b_next = NULL; 3414 hdr->b_buf = buf; 3415 ASSERT(hdr->b_datacnt == 0); 3416 hdr->b_datacnt = 1; 3417 arc_get_data_buf(buf); 3418 arc_access(hdr, hash_lock); 3419 } 3420 3421 ASSERT(!GHOST_STATE(hdr->b_state)); 3422 3423 acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP); 3424 acb->acb_done = done; 3425 acb->acb_private = private; 3426 3427 ASSERT(hdr->b_acb == NULL); 3428 hdr->b_acb = acb; 3429 hdr->b_flags |= ARC_IO_IN_PROGRESS; 3430 3431 if (hdr->b_l2hdr != NULL && 3432 (vd = hdr->b_l2hdr->b_dev->l2ad_vdev) != NULL) { 3433 devw = hdr->b_l2hdr->b_dev->l2ad_writing; 3434 addr = hdr->b_l2hdr->b_daddr; 3435 b_compress = hdr->b_l2hdr->b_compress; 3436 b_asize = hdr->b_l2hdr->b_asize; 3437 /* 3438 * Lock out device removal. 3439 */ 3440 if (vdev_is_dead(vd) || 3441 !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER)) 3442 vd = NULL; 3443 } 3444 3445 if (hash_lock != NULL) 3446 mutex_exit(hash_lock); 3447 3448 /* 3449 * At this point, we have a level 1 cache miss. Try again in 3450 * L2ARC if possible. 3451 */ 3452 ASSERT3U(hdr->b_size, ==, size); 3453 DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp, 3454 uint64_t, size, zbookmark_phys_t *, zb); 3455 ARCSTAT_BUMP(arcstat_misses); 3456 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH), 3457 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA, 3458 data, metadata, misses); 3459#ifdef _KERNEL 3460 curthread->td_ru.ru_inblock++; 3461#endif 3462 3463 if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) { 3464 /* 3465 * Read from the L2ARC if the following are true: 3466 * 1. The L2ARC vdev was previously cached. 3467 * 2. This buffer still has L2ARC metadata. 3468 * 3. This buffer isn't currently writing to the L2ARC. 3469 * 4. The L2ARC entry wasn't evicted, which may 3470 * also have invalidated the vdev. 3471 * 5. This isn't prefetch and l2arc_noprefetch is set. 3472 */ 3473 if (hdr->b_l2hdr != NULL && 3474 !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) && 3475 !(l2arc_noprefetch && HDR_PREFETCH(hdr))) { 3476 l2arc_read_callback_t *cb; 3477 3478 DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr); 3479 ARCSTAT_BUMP(arcstat_l2_hits); 3480 3481 cb = kmem_zalloc(sizeof (l2arc_read_callback_t), 3482 KM_SLEEP); 3483 cb->l2rcb_buf = buf; 3484 cb->l2rcb_spa = spa; 3485 cb->l2rcb_bp = *bp; 3486 cb->l2rcb_zb = *zb; 3487 cb->l2rcb_flags = zio_flags; 3488 cb->l2rcb_compress = b_compress; 3489 3490 ASSERT(addr >= VDEV_LABEL_START_SIZE && 3491 addr + size < vd->vdev_psize - 3492 VDEV_LABEL_END_SIZE); 3493 3494 /* 3495 * l2arc read. The SCL_L2ARC lock will be 3496 * released by l2arc_read_done(). 3497 * Issue a null zio if the underlying buffer 3498 * was squashed to zero size by compression. 3499 */ 3500 if (b_compress == ZIO_COMPRESS_EMPTY) { 3501 rzio = zio_null(pio, spa, vd, 3502 l2arc_read_done, cb, 3503 zio_flags | ZIO_FLAG_DONT_CACHE | 3504 ZIO_FLAG_CANFAIL | 3505 ZIO_FLAG_DONT_PROPAGATE | 3506 ZIO_FLAG_DONT_RETRY); 3507 } else { 3508 rzio = zio_read_phys(pio, vd, addr, 3509 b_asize, buf->b_data, 3510 ZIO_CHECKSUM_OFF, 3511 l2arc_read_done, cb, priority, 3512 zio_flags | ZIO_FLAG_DONT_CACHE | 3513 ZIO_FLAG_CANFAIL | 3514 ZIO_FLAG_DONT_PROPAGATE | 3515 ZIO_FLAG_DONT_RETRY, B_FALSE); 3516 } 3517 DTRACE_PROBE2(l2arc__read, vdev_t *, vd, 3518 zio_t *, rzio); 3519 ARCSTAT_INCR(arcstat_l2_read_bytes, b_asize); 3520 3521 if (*arc_flags & ARC_NOWAIT) { 3522 zio_nowait(rzio); 3523 return (0); 3524 } 3525 3526 ASSERT(*arc_flags & ARC_WAIT); 3527 if (zio_wait(rzio) == 0) 3528 return (0); 3529 3530 /* l2arc read error; goto zio_read() */ 3531 } else { 3532 DTRACE_PROBE1(l2arc__miss, 3533 arc_buf_hdr_t *, hdr); 3534 ARCSTAT_BUMP(arcstat_l2_misses); 3535 if (HDR_L2_WRITING(hdr)) 3536 ARCSTAT_BUMP(arcstat_l2_rw_clash); 3537 spa_config_exit(spa, SCL_L2ARC, vd); 3538 } 3539 } else { 3540 if (vd != NULL) 3541 spa_config_exit(spa, SCL_L2ARC, vd); 3542 if (l2arc_ndev != 0) { 3543 DTRACE_PROBE1(l2arc__miss, 3544 arc_buf_hdr_t *, hdr); 3545 ARCSTAT_BUMP(arcstat_l2_misses); 3546 } 3547 } 3548 3549 rzio = zio_read(pio, spa, bp, buf->b_data, size, 3550 arc_read_done, buf, priority, zio_flags, zb); 3551 3552 if (*arc_flags & ARC_WAIT) 3553 return (zio_wait(rzio)); 3554 3555 ASSERT(*arc_flags & ARC_NOWAIT); 3556 zio_nowait(rzio); 3557 } 3558 return (0); 3559} 3560 3561void 3562arc_set_callback(arc_buf_t *buf, arc_evict_func_t *func, void *private) 3563{ 3564 ASSERT(buf->b_hdr != NULL); 3565 ASSERT(buf->b_hdr->b_state != arc_anon); 3566 ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt) || func == NULL); 3567 ASSERT(buf->b_efunc == NULL); 3568 ASSERT(!HDR_BUF_AVAILABLE(buf->b_hdr)); 3569 3570 buf->b_efunc = func; 3571 buf->b_private = private; 3572} 3573 3574/* 3575 * Notify the arc that a block was freed, and thus will never be used again. 3576 */ 3577void 3578arc_freed(spa_t *spa, const blkptr_t *bp) 3579{ 3580 arc_buf_hdr_t *hdr; 3581 kmutex_t *hash_lock; 3582 uint64_t guid = spa_load_guid(spa); 3583 3584 ASSERT(!BP_IS_EMBEDDED(bp)); 3585 3586 hdr = buf_hash_find(guid, bp, &hash_lock); 3587 if (hdr == NULL) 3588 return; 3589 if (HDR_BUF_AVAILABLE(hdr)) { 3590 arc_buf_t *buf = hdr->b_buf; 3591 add_reference(hdr, hash_lock, FTAG); 3592 hdr->b_flags &= ~ARC_BUF_AVAILABLE; 3593 mutex_exit(hash_lock); 3594 3595 arc_release(buf, FTAG); 3596 (void) arc_buf_remove_ref(buf, FTAG); 3597 } else { 3598 mutex_exit(hash_lock); 3599 } 3600 3601} 3602 3603/* 3604 * Clear the user eviction callback set by arc_set_callback(), first calling 3605 * it if it exists. Because the presence of a callback keeps an arc_buf cached 3606 * clearing the callback may result in the arc_buf being destroyed. However, 3607 * it will not result in the *last* arc_buf being destroyed, hence the data 3608 * will remain cached in the ARC. We make a copy of the arc buffer here so 3609 * that we can process the callback without holding any locks. 3610 * 3611 * It's possible that the callback is already in the process of being cleared 3612 * by another thread. In this case we can not clear the callback. 3613 * 3614 * Returns B_TRUE if the callback was successfully called and cleared. 3615 */ 3616boolean_t 3617arc_clear_callback(arc_buf_t *buf) 3618{ 3619 arc_buf_hdr_t *hdr; 3620 kmutex_t *hash_lock; 3621 arc_evict_func_t *efunc = buf->b_efunc; 3622 void *private = buf->b_private; 3623 list_t *list, *evicted_list; 3624 kmutex_t *lock, *evicted_lock; 3625 3626 mutex_enter(&buf->b_evict_lock); 3627 hdr = buf->b_hdr; 3628 if (hdr == NULL) { 3629 /* 3630 * We are in arc_do_user_evicts(). 3631 */ 3632 ASSERT(buf->b_data == NULL); 3633 mutex_exit(&buf->b_evict_lock); 3634 return (B_FALSE); 3635 } else if (buf->b_data == NULL) { 3636 /* 3637 * We are on the eviction list; process this buffer now 3638 * but let arc_do_user_evicts() do the reaping. 3639 */ 3640 buf->b_efunc = NULL; 3641 mutex_exit(&buf->b_evict_lock); 3642 VERIFY0(efunc(private)); 3643 return (B_TRUE); 3644 } 3645 hash_lock = HDR_LOCK(hdr); 3646 mutex_enter(hash_lock); 3647 hdr = buf->b_hdr; 3648 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 3649 3650 ASSERT3U(refcount_count(&hdr->b_refcnt), <, hdr->b_datacnt); 3651 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu); 3652 3653 buf->b_efunc = NULL; 3654 buf->b_private = NULL; 3655 3656 if (hdr->b_datacnt > 1) { 3657 mutex_exit(&buf->b_evict_lock); 3658 arc_buf_destroy(buf, FALSE, TRUE); 3659 } else { 3660 ASSERT(buf == hdr->b_buf); 3661 hdr->b_flags |= ARC_BUF_AVAILABLE; 3662 mutex_exit(&buf->b_evict_lock); 3663 } 3664 3665 mutex_exit(hash_lock); 3666 VERIFY0(efunc(private)); 3667 return (B_TRUE); 3668} 3669 3670/* 3671 * Release this buffer from the cache, making it an anonymous buffer. This 3672 * must be done after a read and prior to modifying the buffer contents. 3673 * If the buffer has more than one reference, we must make 3674 * a new hdr for the buffer. 3675 */ 3676void 3677arc_release(arc_buf_t *buf, void *tag) 3678{ 3679 arc_buf_hdr_t *hdr; 3680 kmutex_t *hash_lock = NULL; 3681 l2arc_buf_hdr_t *l2hdr; 3682 uint64_t buf_size; 3683 3684 /* 3685 * It would be nice to assert that if it's DMU metadata (level > 3686 * 0 || it's the dnode file), then it must be syncing context. 3687 * But we don't know that information at this level. 3688 */ 3689 3690 mutex_enter(&buf->b_evict_lock); 3691 hdr = buf->b_hdr; 3692 3693 /* this buffer is not on any list */ 3694 ASSERT(refcount_count(&hdr->b_refcnt) > 0); 3695 3696 if (hdr->b_state == arc_anon) { 3697 /* this buffer is already released */ 3698 ASSERT(buf->b_efunc == NULL); 3699 } else { 3700 hash_lock = HDR_LOCK(hdr); 3701 mutex_enter(hash_lock); 3702 hdr = buf->b_hdr; 3703 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 3704 } 3705 3706 l2hdr = hdr->b_l2hdr; 3707 if (l2hdr) { 3708 mutex_enter(&l2arc_buflist_mtx); 3709 arc_buf_l2_cdata_free(hdr); 3710 hdr->b_l2hdr = NULL; 3711 list_remove(l2hdr->b_dev->l2ad_buflist, hdr); 3712 } 3713 buf_size = hdr->b_size; 3714 3715 /* 3716 * Do we have more than one buf? 3717 */ 3718 if (hdr->b_datacnt > 1) { 3719 arc_buf_hdr_t *nhdr; 3720 arc_buf_t **bufp; 3721 uint64_t blksz = hdr->b_size; 3722 uint64_t spa = hdr->b_spa; 3723 arc_buf_contents_t type = hdr->b_type; 3724 uint32_t flags = hdr->b_flags; 3725 3726 ASSERT(hdr->b_buf != buf || buf->b_next != NULL); 3727 /* 3728 * Pull the data off of this hdr and attach it to 3729 * a new anonymous hdr. 3730 */ 3731 (void) remove_reference(hdr, hash_lock, tag); 3732 bufp = &hdr->b_buf; 3733 while (*bufp != buf) 3734 bufp = &(*bufp)->b_next; 3735 *bufp = buf->b_next; 3736 buf->b_next = NULL; 3737 3738 ASSERT3U(hdr->b_state->arcs_size, >=, hdr->b_size); 3739 atomic_add_64(&hdr->b_state->arcs_size, -hdr->b_size); 3740 if (refcount_is_zero(&hdr->b_refcnt)) { 3741 uint64_t *size = &hdr->b_state->arcs_lsize[hdr->b_type]; 3742 ASSERT3U(*size, >=, hdr->b_size); 3743 atomic_add_64(size, -hdr->b_size); 3744 } 3745 3746 /* 3747 * We're releasing a duplicate user data buffer, update 3748 * our statistics accordingly. 3749 */ 3750 if (hdr->b_type == ARC_BUFC_DATA) { 3751 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers); 3752 ARCSTAT_INCR(arcstat_duplicate_buffers_size, 3753 -hdr->b_size); 3754 } 3755 hdr->b_datacnt -= 1; 3756 arc_cksum_verify(buf); 3757#ifdef illumos 3758 arc_buf_unwatch(buf); 3759#endif /* illumos */ 3760 3761 mutex_exit(hash_lock); 3762 3763 nhdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE); 3764 nhdr->b_size = blksz; 3765 nhdr->b_spa = spa; 3766 nhdr->b_type = type; 3767 nhdr->b_buf = buf; 3768 nhdr->b_state = arc_anon; 3769 nhdr->b_arc_access = 0; 3770 nhdr->b_flags = flags & ARC_L2_WRITING; 3771 nhdr->b_l2hdr = NULL; 3772 nhdr->b_datacnt = 1; 3773 nhdr->b_freeze_cksum = NULL; 3774 (void) refcount_add(&nhdr->b_refcnt, tag); 3775 buf->b_hdr = nhdr; 3776 mutex_exit(&buf->b_evict_lock); 3777 atomic_add_64(&arc_anon->arcs_size, blksz); 3778 } else { 3779 mutex_exit(&buf->b_evict_lock); 3780 ASSERT(refcount_count(&hdr->b_refcnt) == 1); 3781 ASSERT(!list_link_active(&hdr->b_arc_node)); 3782 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 3783 if (hdr->b_state != arc_anon) 3784 arc_change_state(arc_anon, hdr, hash_lock); 3785 hdr->b_arc_access = 0; 3786 if (hash_lock) 3787 mutex_exit(hash_lock); 3788 3789 buf_discard_identity(hdr); 3790 arc_buf_thaw(buf); 3791 } 3792 buf->b_efunc = NULL; 3793 buf->b_private = NULL; 3794 3795 if (l2hdr) { 3796 ARCSTAT_INCR(arcstat_l2_asize, -l2hdr->b_asize); 3797 vdev_space_update(l2hdr->b_dev->l2ad_vdev, 3798 -l2hdr->b_asize, 0, 0); 3799 trim_map_free(l2hdr->b_dev->l2ad_vdev, l2hdr->b_daddr, 3800 hdr->b_size, 0); 3801 kmem_free(l2hdr, sizeof (l2arc_buf_hdr_t)); 3802 ARCSTAT_INCR(arcstat_l2_size, -buf_size); 3803 mutex_exit(&l2arc_buflist_mtx); 3804 } 3805} 3806 3807int 3808arc_released(arc_buf_t *buf) 3809{ 3810 int released; 3811 3812 mutex_enter(&buf->b_evict_lock); 3813 released = (buf->b_data != NULL && buf->b_hdr->b_state == arc_anon); 3814 mutex_exit(&buf->b_evict_lock); 3815 return (released); 3816} 3817 3818#ifdef ZFS_DEBUG 3819int 3820arc_referenced(arc_buf_t *buf) 3821{ 3822 int referenced; 3823 3824 mutex_enter(&buf->b_evict_lock); 3825 referenced = (refcount_count(&buf->b_hdr->b_refcnt)); 3826 mutex_exit(&buf->b_evict_lock); 3827 return (referenced); 3828} 3829#endif 3830 3831static void 3832arc_write_ready(zio_t *zio) 3833{ 3834 arc_write_callback_t *callback = zio->io_private; 3835 arc_buf_t *buf = callback->awcb_buf; 3836 arc_buf_hdr_t *hdr = buf->b_hdr; 3837 3838 ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt)); 3839 callback->awcb_ready(zio, buf, callback->awcb_private); 3840 3841 /* 3842 * If the IO is already in progress, then this is a re-write 3843 * attempt, so we need to thaw and re-compute the cksum. 3844 * It is the responsibility of the callback to handle the 3845 * accounting for any re-write attempt. 3846 */ 3847 if (HDR_IO_IN_PROGRESS(hdr)) { 3848 mutex_enter(&hdr->b_freeze_lock); 3849 if (hdr->b_freeze_cksum != NULL) { 3850 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 3851 hdr->b_freeze_cksum = NULL; 3852 } 3853 mutex_exit(&hdr->b_freeze_lock); 3854 } 3855 arc_cksum_compute(buf, B_FALSE); 3856 hdr->b_flags |= ARC_IO_IN_PROGRESS; 3857} 3858 3859/* 3860 * The SPA calls this callback for each physical write that happens on behalf 3861 * of a logical write. See the comment in dbuf_write_physdone() for details. 3862 */ 3863static void 3864arc_write_physdone(zio_t *zio) 3865{ 3866 arc_write_callback_t *cb = zio->io_private; 3867 if (cb->awcb_physdone != NULL) 3868 cb->awcb_physdone(zio, cb->awcb_buf, cb->awcb_private); 3869} 3870 3871static void 3872arc_write_done(zio_t *zio) 3873{ 3874 arc_write_callback_t *callback = zio->io_private; 3875 arc_buf_t *buf = callback->awcb_buf; 3876 arc_buf_hdr_t *hdr = buf->b_hdr; 3877 3878 ASSERT(hdr->b_acb == NULL); 3879 3880 if (zio->io_error == 0) { 3881 if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) { 3882 buf_discard_identity(hdr); 3883 } else { 3884 hdr->b_dva = *BP_IDENTITY(zio->io_bp); 3885 hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp); 3886 hdr->b_cksum0 = zio->io_bp->blk_cksum.zc_word[0]; 3887 } 3888 } else { 3889 ASSERT(BUF_EMPTY(hdr)); 3890 } 3891 3892 /* 3893 * If the block to be written was all-zero or compressed enough to be 3894 * embedded in the BP, no write was performed so there will be no 3895 * dva/birth/checksum. The buffer must therefore remain anonymous 3896 * (and uncached). 3897 */ 3898 if (!BUF_EMPTY(hdr)) { 3899 arc_buf_hdr_t *exists; 3900 kmutex_t *hash_lock; 3901 3902 ASSERT(zio->io_error == 0); 3903 3904 arc_cksum_verify(buf); 3905 3906 exists = buf_hash_insert(hdr, &hash_lock); 3907 if (exists) { 3908 /* 3909 * This can only happen if we overwrite for 3910 * sync-to-convergence, because we remove 3911 * buffers from the hash table when we arc_free(). 3912 */ 3913 if (zio->io_flags & ZIO_FLAG_IO_REWRITE) { 3914 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) 3915 panic("bad overwrite, hdr=%p exists=%p", 3916 (void *)hdr, (void *)exists); 3917 ASSERT(refcount_is_zero(&exists->b_refcnt)); 3918 arc_change_state(arc_anon, exists, hash_lock); 3919 mutex_exit(hash_lock); 3920 arc_hdr_destroy(exists); 3921 exists = buf_hash_insert(hdr, &hash_lock); 3922 ASSERT3P(exists, ==, NULL); 3923 } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) { 3924 /* nopwrite */ 3925 ASSERT(zio->io_prop.zp_nopwrite); 3926 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) 3927 panic("bad nopwrite, hdr=%p exists=%p", 3928 (void *)hdr, (void *)exists); 3929 } else { 3930 /* Dedup */ 3931 ASSERT(hdr->b_datacnt == 1); 3932 ASSERT(hdr->b_state == arc_anon); 3933 ASSERT(BP_GET_DEDUP(zio->io_bp)); 3934 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0); 3935 } 3936 } 3937 hdr->b_flags &= ~ARC_IO_IN_PROGRESS; 3938 /* if it's not anon, we are doing a scrub */ 3939 if (!exists && hdr->b_state == arc_anon) 3940 arc_access(hdr, hash_lock); 3941 mutex_exit(hash_lock); 3942 } else { 3943 hdr->b_flags &= ~ARC_IO_IN_PROGRESS; 3944 } 3945 3946 ASSERT(!refcount_is_zero(&hdr->b_refcnt)); 3947 callback->awcb_done(zio, buf, callback->awcb_private); 3948 3949 kmem_free(callback, sizeof (arc_write_callback_t)); 3950} 3951 3952zio_t * 3953arc_write(zio_t *pio, spa_t *spa, uint64_t txg, 3954 blkptr_t *bp, arc_buf_t *buf, boolean_t l2arc, boolean_t l2arc_compress, 3955 const zio_prop_t *zp, arc_done_func_t *ready, arc_done_func_t *physdone, 3956 arc_done_func_t *done, void *private, zio_priority_t priority, 3957 int zio_flags, const zbookmark_phys_t *zb) 3958{ 3959 arc_buf_hdr_t *hdr = buf->b_hdr; 3960 arc_write_callback_t *callback; 3961 zio_t *zio; 3962 3963 ASSERT(ready != NULL); 3964 ASSERT(done != NULL); 3965 ASSERT(!HDR_IO_ERROR(hdr)); 3966 ASSERT((hdr->b_flags & ARC_IO_IN_PROGRESS) == 0); 3967 ASSERT(hdr->b_acb == NULL); 3968 if (l2arc) 3969 hdr->b_flags |= ARC_L2CACHE; 3970 if (l2arc_compress) 3971 hdr->b_flags |= ARC_L2COMPRESS; 3972 callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP); 3973 callback->awcb_ready = ready; 3974 callback->awcb_physdone = physdone; 3975 callback->awcb_done = done; 3976 callback->awcb_private = private; 3977 callback->awcb_buf = buf; 3978 3979 zio = zio_write(pio, spa, txg, bp, buf->b_data, hdr->b_size, zp, 3980 arc_write_ready, arc_write_physdone, arc_write_done, callback, 3981 priority, zio_flags, zb); 3982 3983 return (zio); 3984} 3985 3986static int 3987arc_memory_throttle(uint64_t reserve, uint64_t txg) 3988{ 3989#ifdef _KERNEL 3990 uint64_t available_memory = ptob(freemem); 3991 static uint64_t page_load = 0; 3992 static uint64_t last_txg = 0; 3993 3994#if defined(__i386) || !defined(UMA_MD_SMALL_ALLOC) 3995 available_memory = 3996 MIN(available_memory, ptob(vmem_size(heap_arena, VMEM_FREE))); 3997#endif 3998 3999 if (freemem > (uint64_t)physmem * arc_lotsfree_percent / 100) 4000 return (0); 4001 4002 if (txg > last_txg) { 4003 last_txg = txg; 4004 page_load = 0; 4005 } 4006 /* 4007 * If we are in pageout, we know that memory is already tight, 4008 * the arc is already going to be evicting, so we just want to 4009 * continue to let page writes occur as quickly as possible. 4010 */ 4011 if (curproc == pageproc) { 4012 if (page_load > MAX(ptob(minfree), available_memory) / 4) 4013 return (SET_ERROR(ERESTART)); 4014 /* Note: reserve is inflated, so we deflate */ 4015 page_load += reserve / 8; 4016 return (0); 4017 } else if (page_load > 0 && arc_reclaim_needed()) { 4018 /* memory is low, delay before restarting */ 4019 ARCSTAT_INCR(arcstat_memory_throttle_count, 1); 4020 return (SET_ERROR(EAGAIN)); 4021 } 4022 page_load = 0; 4023#endif 4024 return (0); 4025} 4026 4027void 4028arc_tempreserve_clear(uint64_t reserve) 4029{ 4030 atomic_add_64(&arc_tempreserve, -reserve); 4031 ASSERT((int64_t)arc_tempreserve >= 0); 4032} 4033 4034int 4035arc_tempreserve_space(uint64_t reserve, uint64_t txg) 4036{ 4037 int error; 4038 uint64_t anon_size; 4039 4040 if (reserve > arc_c/4 && !arc_no_grow) { 4041 arc_c = MIN(arc_c_max, reserve * 4); 4042 DTRACE_PROBE1(arc__set_reserve, uint64_t, arc_c); 4043 } 4044 if (reserve > arc_c) 4045 return (SET_ERROR(ENOMEM)); 4046 4047 /* 4048 * Don't count loaned bufs as in flight dirty data to prevent long 4049 * network delays from blocking transactions that are ready to be 4050 * assigned to a txg. 4051 */ 4052 anon_size = MAX((int64_t)(arc_anon->arcs_size - arc_loaned_bytes), 0); 4053 4054 /* 4055 * Writes will, almost always, require additional memory allocations 4056 * in order to compress/encrypt/etc the data. We therefore need to 4057 * make sure that there is sufficient available memory for this. 4058 */ 4059 error = arc_memory_throttle(reserve, txg); 4060 if (error != 0) 4061 return (error); 4062 4063 /* 4064 * Throttle writes when the amount of dirty data in the cache 4065 * gets too large. We try to keep the cache less than half full 4066 * of dirty blocks so that our sync times don't grow too large. 4067 * Note: if two requests come in concurrently, we might let them 4068 * both succeed, when one of them should fail. Not a huge deal. 4069 */ 4070 4071 if (reserve + arc_tempreserve + anon_size > arc_c / 2 && 4072 anon_size > arc_c / 4) { 4073 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK " 4074 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n", 4075 arc_tempreserve>>10, 4076 arc_anon->arcs_lsize[ARC_BUFC_METADATA]>>10, 4077 arc_anon->arcs_lsize[ARC_BUFC_DATA]>>10, 4078 reserve>>10, arc_c>>10); 4079 return (SET_ERROR(ERESTART)); 4080 } 4081 atomic_add_64(&arc_tempreserve, reserve); 4082 return (0); 4083} 4084 4085static kmutex_t arc_lowmem_lock; 4086#ifdef _KERNEL 4087static eventhandler_tag arc_event_lowmem = NULL; 4088 4089static void 4090arc_lowmem(void *arg __unused, int howto __unused) 4091{ 4092 4093 /* Serialize access via arc_lowmem_lock. */ 4094 mutex_enter(&arc_lowmem_lock); 4095 mutex_enter(&arc_reclaim_thr_lock); 4096 needfree = 1; 4097 DTRACE_PROBE(arc__needfree); 4098 cv_signal(&arc_reclaim_thr_cv); 4099 4100 /* 4101 * It is unsafe to block here in arbitrary threads, because we can come 4102 * here from ARC itself and may hold ARC locks and thus risk a deadlock 4103 * with ARC reclaim thread. 4104 */ 4105 if (curproc == pageproc) { 4106 while (needfree) 4107 msleep(&needfree, &arc_reclaim_thr_lock, 0, "zfs:lowmem", 0); 4108 } 4109 mutex_exit(&arc_reclaim_thr_lock); 4110 mutex_exit(&arc_lowmem_lock); 4111} 4112#endif 4113 4114void 4115arc_init(void) 4116{ 4117 int i, prefetch_tunable_set = 0; 4118 4119 mutex_init(&arc_reclaim_thr_lock, NULL, MUTEX_DEFAULT, NULL); 4120 cv_init(&arc_reclaim_thr_cv, NULL, CV_DEFAULT, NULL); 4121 mutex_init(&arc_lowmem_lock, NULL, MUTEX_DEFAULT, NULL); 4122 4123 /* Convert seconds to clock ticks */ 4124 arc_min_prefetch_lifespan = 1 * hz; 4125 4126 /* Start out with 1/8 of all memory */ 4127 arc_c = kmem_size() / 8; 4128 4129#ifdef sun 4130#ifdef _KERNEL 4131 /* 4132 * On architectures where the physical memory can be larger 4133 * than the addressable space (intel in 32-bit mode), we may 4134 * need to limit the cache to 1/8 of VM size. 4135 */ 4136 arc_c = MIN(arc_c, vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 8); 4137#endif 4138#endif /* sun */ 4139 /* set min cache to 1/32 of all memory, or 16MB, whichever is more */ 4140 arc_c_min = MAX(arc_c / 4, 64<<18); 4141 /* set max to 1/2 of all memory, or all but 1GB, whichever is more */ 4142 if (arc_c * 8 >= 1<<30) 4143 arc_c_max = (arc_c * 8) - (1<<30); 4144 else 4145 arc_c_max = arc_c_min; 4146 arc_c_max = MAX(arc_c * 5, arc_c_max); 4147 4148#ifdef _KERNEL 4149 /* 4150 * Allow the tunables to override our calculations if they are 4151 * reasonable (ie. over 16MB) 4152 */ 4153 if (zfs_arc_max > 64<<18 && zfs_arc_max < kmem_size()) 4154 arc_c_max = zfs_arc_max; 4155 if (zfs_arc_min > 64<<18 && zfs_arc_min <= arc_c_max) 4156 arc_c_min = zfs_arc_min; 4157#endif 4158 4159 arc_c = arc_c_max; 4160 arc_p = (arc_c >> 1); 4161 4162 /* limit meta-data to 1/4 of the arc capacity */ 4163 arc_meta_limit = arc_c_max / 4; 4164 4165 /* Allow the tunable to override if it is reasonable */ 4166 if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max) 4167 arc_meta_limit = zfs_arc_meta_limit; 4168 4169 if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0) 4170 arc_c_min = arc_meta_limit / 2; 4171 4172 if (zfs_arc_grow_retry > 0) 4173 arc_grow_retry = zfs_arc_grow_retry; 4174 4175 if (zfs_arc_shrink_shift > 0) 4176 arc_shrink_shift = zfs_arc_shrink_shift; 4177 4178 if (zfs_arc_p_min_shift > 0) 4179 arc_p_min_shift = zfs_arc_p_min_shift; 4180 4181 /* if kmem_flags are set, lets try to use less memory */ 4182 if (kmem_debugging()) 4183 arc_c = arc_c / 2; 4184 if (arc_c < arc_c_min) 4185 arc_c = arc_c_min; 4186 4187 zfs_arc_min = arc_c_min; 4188 zfs_arc_max = arc_c_max; 4189 4190 arc_anon = &ARC_anon; 4191 arc_mru = &ARC_mru; 4192 arc_mru_ghost = &ARC_mru_ghost; 4193 arc_mfu = &ARC_mfu; 4194 arc_mfu_ghost = &ARC_mfu_ghost; 4195 arc_l2c_only = &ARC_l2c_only; 4196 arc_size = 0; 4197 4198 for (i = 0; i < ARC_BUFC_NUMLISTS; i++) { 4199 mutex_init(&arc_anon->arcs_locks[i].arcs_lock, 4200 NULL, MUTEX_DEFAULT, NULL); 4201 mutex_init(&arc_mru->arcs_locks[i].arcs_lock, 4202 NULL, MUTEX_DEFAULT, NULL); 4203 mutex_init(&arc_mru_ghost->arcs_locks[i].arcs_lock, 4204 NULL, MUTEX_DEFAULT, NULL); 4205 mutex_init(&arc_mfu->arcs_locks[i].arcs_lock, 4206 NULL, MUTEX_DEFAULT, NULL); 4207 mutex_init(&arc_mfu_ghost->arcs_locks[i].arcs_lock, 4208 NULL, MUTEX_DEFAULT, NULL); 4209 mutex_init(&arc_l2c_only->arcs_locks[i].arcs_lock, 4210 NULL, MUTEX_DEFAULT, NULL); 4211 4212 list_create(&arc_mru->arcs_lists[i], 4213 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4214 list_create(&arc_mru_ghost->arcs_lists[i], 4215 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4216 list_create(&arc_mfu->arcs_lists[i], 4217 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4218 list_create(&arc_mfu_ghost->arcs_lists[i], 4219 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4220 list_create(&arc_mfu_ghost->arcs_lists[i], 4221 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4222 list_create(&arc_l2c_only->arcs_lists[i], 4223 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4224 } 4225 4226 buf_init(); 4227 4228 arc_thread_exit = 0; 4229 arc_eviction_list = NULL; 4230 mutex_init(&arc_eviction_mtx, NULL, MUTEX_DEFAULT, NULL); 4231 bzero(&arc_eviction_hdr, sizeof (arc_buf_hdr_t)); 4232 4233 arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED, 4234 sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL); 4235 4236 if (arc_ksp != NULL) { 4237 arc_ksp->ks_data = &arc_stats; 4238 kstat_install(arc_ksp); 4239 } 4240 4241 (void) thread_create(NULL, 0, arc_reclaim_thread, NULL, 0, &p0, 4242 TS_RUN, minclsyspri); 4243 4244#ifdef _KERNEL 4245 arc_event_lowmem = EVENTHANDLER_REGISTER(vm_lowmem, arc_lowmem, NULL, 4246 EVENTHANDLER_PRI_FIRST); 4247#endif 4248 4249 arc_dead = FALSE; 4250 arc_warm = B_FALSE; 4251 4252 /* 4253 * Calculate maximum amount of dirty data per pool. 4254 * 4255 * If it has been set by /etc/system, take that. 4256 * Otherwise, use a percentage of physical memory defined by 4257 * zfs_dirty_data_max_percent (default 10%) with a cap at 4258 * zfs_dirty_data_max_max (default 4GB). 4259 */ 4260 if (zfs_dirty_data_max == 0) { 4261 zfs_dirty_data_max = ptob(physmem) * 4262 zfs_dirty_data_max_percent / 100; 4263 zfs_dirty_data_max = MIN(zfs_dirty_data_max, 4264 zfs_dirty_data_max_max); 4265 } 4266 4267#ifdef _KERNEL 4268 if (TUNABLE_INT_FETCH("vfs.zfs.prefetch_disable", &zfs_prefetch_disable)) 4269 prefetch_tunable_set = 1; 4270 4271#ifdef __i386__ 4272 if (prefetch_tunable_set == 0) { 4273 printf("ZFS NOTICE: Prefetch is disabled by default on i386 " 4274 "-- to enable,\n"); 4275 printf(" add \"vfs.zfs.prefetch_disable=0\" " 4276 "to /boot/loader.conf.\n"); 4277 zfs_prefetch_disable = 1; 4278 } 4279#else 4280 if ((((uint64_t)physmem * PAGESIZE) < (1ULL << 32)) && 4281 prefetch_tunable_set == 0) { 4282 printf("ZFS NOTICE: Prefetch is disabled by default if less " 4283 "than 4GB of RAM is present;\n" 4284 " to enable, add \"vfs.zfs.prefetch_disable=0\" " 4285 "to /boot/loader.conf.\n"); 4286 zfs_prefetch_disable = 1; 4287 } 4288#endif 4289 /* Warn about ZFS memory and address space requirements. */ 4290 if (((uint64_t)physmem * PAGESIZE) < (256 + 128 + 64) * (1 << 20)) { 4291 printf("ZFS WARNING: Recommended minimum RAM size is 512MB; " 4292 "expect unstable behavior.\n"); 4293 } 4294 if (kmem_size() < 512 * (1 << 20)) { 4295 printf("ZFS WARNING: Recommended minimum kmem_size is 512MB; " 4296 "expect unstable behavior.\n"); 4297 printf(" Consider tuning vm.kmem_size and " 4298 "vm.kmem_size_max\n"); 4299 printf(" in /boot/loader.conf.\n"); 4300 } 4301#endif 4302} 4303 4304void 4305arc_fini(void) 4306{ 4307 int i; 4308 4309 mutex_enter(&arc_reclaim_thr_lock); 4310 arc_thread_exit = 1; 4311 cv_signal(&arc_reclaim_thr_cv); 4312 while (arc_thread_exit != 0) 4313 cv_wait(&arc_reclaim_thr_cv, &arc_reclaim_thr_lock); 4314 mutex_exit(&arc_reclaim_thr_lock); 4315 4316 arc_flush(NULL); 4317 4318 arc_dead = TRUE; 4319 4320 if (arc_ksp != NULL) { 4321 kstat_delete(arc_ksp); 4322 arc_ksp = NULL; 4323 } 4324 4325 mutex_destroy(&arc_eviction_mtx); 4326 mutex_destroy(&arc_reclaim_thr_lock); 4327 cv_destroy(&arc_reclaim_thr_cv); 4328 4329 for (i = 0; i < ARC_BUFC_NUMLISTS; i++) { 4330 list_destroy(&arc_mru->arcs_lists[i]); 4331 list_destroy(&arc_mru_ghost->arcs_lists[i]); 4332 list_destroy(&arc_mfu->arcs_lists[i]); 4333 list_destroy(&arc_mfu_ghost->arcs_lists[i]); 4334 list_destroy(&arc_l2c_only->arcs_lists[i]); 4335 4336 mutex_destroy(&arc_anon->arcs_locks[i].arcs_lock); 4337 mutex_destroy(&arc_mru->arcs_locks[i].arcs_lock); 4338 mutex_destroy(&arc_mru_ghost->arcs_locks[i].arcs_lock); 4339 mutex_destroy(&arc_mfu->arcs_locks[i].arcs_lock); 4340 mutex_destroy(&arc_mfu_ghost->arcs_locks[i].arcs_lock); 4341 mutex_destroy(&arc_l2c_only->arcs_locks[i].arcs_lock); 4342 } 4343 4344 buf_fini(); 4345 4346 ASSERT(arc_loaned_bytes == 0); 4347 4348 mutex_destroy(&arc_lowmem_lock); 4349#ifdef _KERNEL 4350 if (arc_event_lowmem != NULL) 4351 EVENTHANDLER_DEREGISTER(vm_lowmem, arc_event_lowmem); 4352#endif 4353} 4354 4355/* 4356 * Level 2 ARC 4357 * 4358 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk. 4359 * It uses dedicated storage devices to hold cached data, which are populated 4360 * using large infrequent writes. The main role of this cache is to boost 4361 * the performance of random read workloads. The intended L2ARC devices 4362 * include short-stroked disks, solid state disks, and other media with 4363 * substantially faster read latency than disk. 4364 * 4365 * +-----------------------+ 4366 * | ARC | 4367 * +-----------------------+ 4368 * | ^ ^ 4369 * | | | 4370 * l2arc_feed_thread() arc_read() 4371 * | | | 4372 * | l2arc read | 4373 * V | | 4374 * +---------------+ | 4375 * | L2ARC | | 4376 * +---------------+ | 4377 * | ^ | 4378 * l2arc_write() | | 4379 * | | | 4380 * V | | 4381 * +-------+ +-------+ 4382 * | vdev | | vdev | 4383 * | cache | | cache | 4384 * +-------+ +-------+ 4385 * +=========+ .-----. 4386 * : L2ARC : |-_____-| 4387 * : devices : | Disks | 4388 * +=========+ `-_____-' 4389 * 4390 * Read requests are satisfied from the following sources, in order: 4391 * 4392 * 1) ARC 4393 * 2) vdev cache of L2ARC devices 4394 * 3) L2ARC devices 4395 * 4) vdev cache of disks 4396 * 5) disks 4397 * 4398 * Some L2ARC device types exhibit extremely slow write performance. 4399 * To accommodate for this there are some significant differences between 4400 * the L2ARC and traditional cache design: 4401 * 4402 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from 4403 * the ARC behave as usual, freeing buffers and placing headers on ghost 4404 * lists. The ARC does not send buffers to the L2ARC during eviction as 4405 * this would add inflated write latencies for all ARC memory pressure. 4406 * 4407 * 2. The L2ARC attempts to cache data from the ARC before it is evicted. 4408 * It does this by periodically scanning buffers from the eviction-end of 4409 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are 4410 * not already there. It scans until a headroom of buffers is satisfied, 4411 * which itself is a buffer for ARC eviction. If a compressible buffer is 4412 * found during scanning and selected for writing to an L2ARC device, we 4413 * temporarily boost scanning headroom during the next scan cycle to make 4414 * sure we adapt to compression effects (which might significantly reduce 4415 * the data volume we write to L2ARC). The thread that does this is 4416 * l2arc_feed_thread(), illustrated below; example sizes are included to 4417 * provide a better sense of ratio than this diagram: 4418 * 4419 * head --> tail 4420 * +---------------------+----------+ 4421 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC 4422 * +---------------------+----------+ | o L2ARC eligible 4423 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer 4424 * +---------------------+----------+ | 4425 * 15.9 Gbytes ^ 32 Mbytes | 4426 * headroom | 4427 * l2arc_feed_thread() 4428 * | 4429 * l2arc write hand <--[oooo]--' 4430 * | 8 Mbyte 4431 * | write max 4432 * V 4433 * +==============================+ 4434 * L2ARC dev |####|#|###|###| |####| ... | 4435 * +==============================+ 4436 * 32 Gbytes 4437 * 4438 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of 4439 * evicted, then the L2ARC has cached a buffer much sooner than it probably 4440 * needed to, potentially wasting L2ARC device bandwidth and storage. It is 4441 * safe to say that this is an uncommon case, since buffers at the end of 4442 * the ARC lists have moved there due to inactivity. 4443 * 4444 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom, 4445 * then the L2ARC simply misses copying some buffers. This serves as a 4446 * pressure valve to prevent heavy read workloads from both stalling the ARC 4447 * with waits and clogging the L2ARC with writes. This also helps prevent 4448 * the potential for the L2ARC to churn if it attempts to cache content too 4449 * quickly, such as during backups of the entire pool. 4450 * 4451 * 5. After system boot and before the ARC has filled main memory, there are 4452 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru 4453 * lists can remain mostly static. Instead of searching from tail of these 4454 * lists as pictured, the l2arc_feed_thread() will search from the list heads 4455 * for eligible buffers, greatly increasing its chance of finding them. 4456 * 4457 * The L2ARC device write speed is also boosted during this time so that 4458 * the L2ARC warms up faster. Since there have been no ARC evictions yet, 4459 * there are no L2ARC reads, and no fear of degrading read performance 4460 * through increased writes. 4461 * 4462 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that 4463 * the vdev queue can aggregate them into larger and fewer writes. Each 4464 * device is written to in a rotor fashion, sweeping writes through 4465 * available space then repeating. 4466 * 4467 * 7. The L2ARC does not store dirty content. It never needs to flush 4468 * write buffers back to disk based storage. 4469 * 4470 * 8. If an ARC buffer is written (and dirtied) which also exists in the 4471 * L2ARC, the now stale L2ARC buffer is immediately dropped. 4472 * 4473 * The performance of the L2ARC can be tweaked by a number of tunables, which 4474 * may be necessary for different workloads: 4475 * 4476 * l2arc_write_max max write bytes per interval 4477 * l2arc_write_boost extra write bytes during device warmup 4478 * l2arc_noprefetch skip caching prefetched buffers 4479 * l2arc_headroom number of max device writes to precache 4480 * l2arc_headroom_boost when we find compressed buffers during ARC 4481 * scanning, we multiply headroom by this 4482 * percentage factor for the next scan cycle, 4483 * since more compressed buffers are likely to 4484 * be present 4485 * l2arc_feed_secs seconds between L2ARC writing 4486 * 4487 * Tunables may be removed or added as future performance improvements are 4488 * integrated, and also may become zpool properties. 4489 * 4490 * There are three key functions that control how the L2ARC warms up: 4491 * 4492 * l2arc_write_eligible() check if a buffer is eligible to cache 4493 * l2arc_write_size() calculate how much to write 4494 * l2arc_write_interval() calculate sleep delay between writes 4495 * 4496 * These three functions determine what to write, how much, and how quickly 4497 * to send writes. 4498 */ 4499 4500static boolean_t 4501l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *ab) 4502{ 4503 /* 4504 * A buffer is *not* eligible for the L2ARC if it: 4505 * 1. belongs to a different spa. 4506 * 2. is already cached on the L2ARC. 4507 * 3. has an I/O in progress (it may be an incomplete read). 4508 * 4. is flagged not eligible (zfs property). 4509 */ 4510 if (ab->b_spa != spa_guid) { 4511 ARCSTAT_BUMP(arcstat_l2_write_spa_mismatch); 4512 return (B_FALSE); 4513 } 4514 if (ab->b_l2hdr != NULL) { 4515 ARCSTAT_BUMP(arcstat_l2_write_in_l2); 4516 return (B_FALSE); 4517 } 4518 if (HDR_IO_IN_PROGRESS(ab)) { 4519 ARCSTAT_BUMP(arcstat_l2_write_hdr_io_in_progress); 4520 return (B_FALSE); 4521 } 4522 if (!HDR_L2CACHE(ab)) { 4523 ARCSTAT_BUMP(arcstat_l2_write_not_cacheable); 4524 return (B_FALSE); 4525 } 4526 4527 return (B_TRUE); 4528} 4529 4530static uint64_t 4531l2arc_write_size(void) 4532{ 4533 uint64_t size; 4534 4535 /* 4536 * Make sure our globals have meaningful values in case the user 4537 * altered them. 4538 */ 4539 size = l2arc_write_max; 4540 if (size == 0) { 4541 cmn_err(CE_NOTE, "Bad value for l2arc_write_max, value must " 4542 "be greater than zero, resetting it to the default (%d)", 4543 L2ARC_WRITE_SIZE); 4544 size = l2arc_write_max = L2ARC_WRITE_SIZE; 4545 } 4546 4547 if (arc_warm == B_FALSE) 4548 size += l2arc_write_boost; 4549 4550 return (size); 4551 4552} 4553 4554static clock_t 4555l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote) 4556{ 4557 clock_t interval, next, now; 4558 4559 /* 4560 * If the ARC lists are busy, increase our write rate; if the 4561 * lists are stale, idle back. This is achieved by checking 4562 * how much we previously wrote - if it was more than half of 4563 * what we wanted, schedule the next write much sooner. 4564 */ 4565 if (l2arc_feed_again && wrote > (wanted / 2)) 4566 interval = (hz * l2arc_feed_min_ms) / 1000; 4567 else 4568 interval = hz * l2arc_feed_secs; 4569 4570 now = ddi_get_lbolt(); 4571 next = MAX(now, MIN(now + interval, began + interval)); 4572 4573 return (next); 4574} 4575 4576static void 4577l2arc_hdr_stat_add(void) 4578{ 4579 ARCSTAT_INCR(arcstat_l2_hdr_size, HDR_SIZE + L2HDR_SIZE); 4580 ARCSTAT_INCR(arcstat_hdr_size, -HDR_SIZE); 4581} 4582 4583static void 4584l2arc_hdr_stat_remove(void) 4585{ 4586 ARCSTAT_INCR(arcstat_l2_hdr_size, -(HDR_SIZE + L2HDR_SIZE)); 4587 ARCSTAT_INCR(arcstat_hdr_size, HDR_SIZE); 4588} 4589 4590/* 4591 * Cycle through L2ARC devices. This is how L2ARC load balances. 4592 * If a device is returned, this also returns holding the spa config lock. 4593 */ 4594static l2arc_dev_t * 4595l2arc_dev_get_next(void) 4596{ 4597 l2arc_dev_t *first, *next = NULL; 4598 4599 /* 4600 * Lock out the removal of spas (spa_namespace_lock), then removal 4601 * of cache devices (l2arc_dev_mtx). Once a device has been selected, 4602 * both locks will be dropped and a spa config lock held instead. 4603 */ 4604 mutex_enter(&spa_namespace_lock); 4605 mutex_enter(&l2arc_dev_mtx); 4606 4607 /* if there are no vdevs, there is nothing to do */ 4608 if (l2arc_ndev == 0) 4609 goto out; 4610 4611 first = NULL; 4612 next = l2arc_dev_last; 4613 do { 4614 /* loop around the list looking for a non-faulted vdev */ 4615 if (next == NULL) { 4616 next = list_head(l2arc_dev_list); 4617 } else { 4618 next = list_next(l2arc_dev_list, next); 4619 if (next == NULL) 4620 next = list_head(l2arc_dev_list); 4621 } 4622 4623 /* if we have come back to the start, bail out */ 4624 if (first == NULL) 4625 first = next; 4626 else if (next == first) 4627 break; 4628 4629 } while (vdev_is_dead(next->l2ad_vdev)); 4630 4631 /* if we were unable to find any usable vdevs, return NULL */ 4632 if (vdev_is_dead(next->l2ad_vdev)) 4633 next = NULL; 4634 4635 l2arc_dev_last = next; 4636 4637out: 4638 mutex_exit(&l2arc_dev_mtx); 4639 4640 /* 4641 * Grab the config lock to prevent the 'next' device from being 4642 * removed while we are writing to it. 4643 */ 4644 if (next != NULL) 4645 spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER); 4646 mutex_exit(&spa_namespace_lock); 4647 4648 return (next); 4649} 4650 4651/* 4652 * Free buffers that were tagged for destruction. 4653 */ 4654static void 4655l2arc_do_free_on_write() 4656{ 4657 list_t *buflist; 4658 l2arc_data_free_t *df, *df_prev; 4659 4660 mutex_enter(&l2arc_free_on_write_mtx); 4661 buflist = l2arc_free_on_write; 4662 4663 for (df = list_tail(buflist); df; df = df_prev) { 4664 df_prev = list_prev(buflist, df); 4665 ASSERT(df->l2df_data != NULL); 4666 ASSERT(df->l2df_func != NULL); 4667 df->l2df_func(df->l2df_data, df->l2df_size); 4668 list_remove(buflist, df); 4669 kmem_free(df, sizeof (l2arc_data_free_t)); 4670 } 4671 4672 mutex_exit(&l2arc_free_on_write_mtx); 4673} 4674 4675/* 4676 * A write to a cache device has completed. Update all headers to allow 4677 * reads from these buffers to begin. 4678 */ 4679static void 4680l2arc_write_done(zio_t *zio) 4681{ 4682 l2arc_write_callback_t *cb; 4683 l2arc_dev_t *dev; 4684 list_t *buflist; 4685 arc_buf_hdr_t *head, *ab, *ab_prev; 4686 l2arc_buf_hdr_t *abl2; 4687 kmutex_t *hash_lock; 4688 int64_t bytes_dropped = 0; 4689 4690 cb = zio->io_private; 4691 ASSERT(cb != NULL); 4692 dev = cb->l2wcb_dev; 4693 ASSERT(dev != NULL); 4694 head = cb->l2wcb_head; 4695 ASSERT(head != NULL); 4696 buflist = dev->l2ad_buflist; 4697 ASSERT(buflist != NULL); 4698 DTRACE_PROBE2(l2arc__iodone, zio_t *, zio, 4699 l2arc_write_callback_t *, cb); 4700 4701 if (zio->io_error != 0) 4702 ARCSTAT_BUMP(arcstat_l2_writes_error); 4703 4704 mutex_enter(&l2arc_buflist_mtx); 4705 4706 /* 4707 * All writes completed, or an error was hit. 4708 */ 4709 for (ab = list_prev(buflist, head); ab; ab = ab_prev) { 4710 ab_prev = list_prev(buflist, ab); 4711 abl2 = ab->b_l2hdr; 4712 4713 /* 4714 * Release the temporary compressed buffer as soon as possible. 4715 */ 4716 if (abl2->b_compress != ZIO_COMPRESS_OFF) 4717 l2arc_release_cdata_buf(ab); 4718 4719 hash_lock = HDR_LOCK(ab); 4720 if (!mutex_tryenter(hash_lock)) { 4721 /* 4722 * This buffer misses out. It may be in a stage 4723 * of eviction. Its ARC_L2_WRITING flag will be 4724 * left set, denying reads to this buffer. 4725 */ 4726 ARCSTAT_BUMP(arcstat_l2_writes_hdr_miss); 4727 continue; 4728 } 4729 4730 if (zio->io_error != 0) { 4731 /* 4732 * Error - drop L2ARC entry. 4733 */ 4734 list_remove(buflist, ab); 4735 ARCSTAT_INCR(arcstat_l2_asize, -abl2->b_asize); 4736 bytes_dropped += abl2->b_asize; 4737 ab->b_l2hdr = NULL; 4738 trim_map_free(abl2->b_dev->l2ad_vdev, abl2->b_daddr, 4739 ab->b_size, 0); 4740 kmem_free(abl2, sizeof (l2arc_buf_hdr_t)); 4741 ARCSTAT_INCR(arcstat_l2_size, -ab->b_size); 4742 } 4743 4744 /* 4745 * Allow ARC to begin reads to this L2ARC entry. 4746 */ 4747 ab->b_flags &= ~ARC_L2_WRITING; 4748 4749 mutex_exit(hash_lock); 4750 } 4751 4752 atomic_inc_64(&l2arc_writes_done); 4753 list_remove(buflist, head); 4754 kmem_cache_free(hdr_cache, head); 4755 mutex_exit(&l2arc_buflist_mtx); 4756 4757 vdev_space_update(dev->l2ad_vdev, -bytes_dropped, 0, 0); 4758 4759 l2arc_do_free_on_write(); 4760 4761 kmem_free(cb, sizeof (l2arc_write_callback_t)); 4762} 4763 4764/* 4765 * A read to a cache device completed. Validate buffer contents before 4766 * handing over to the regular ARC routines. 4767 */ 4768static void 4769l2arc_read_done(zio_t *zio) 4770{ 4771 l2arc_read_callback_t *cb; 4772 arc_buf_hdr_t *hdr; 4773 arc_buf_t *buf; 4774 kmutex_t *hash_lock; 4775 int equal; 4776 4777 ASSERT(zio->io_vd != NULL); 4778 ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE); 4779 4780 spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd); 4781 4782 cb = zio->io_private; 4783 ASSERT(cb != NULL); 4784 buf = cb->l2rcb_buf; 4785 ASSERT(buf != NULL); 4786 4787 hash_lock = HDR_LOCK(buf->b_hdr); 4788 mutex_enter(hash_lock); 4789 hdr = buf->b_hdr; 4790 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 4791 4792 /* 4793 * If the buffer was compressed, decompress it first. 4794 */ 4795 if (cb->l2rcb_compress != ZIO_COMPRESS_OFF) 4796 l2arc_decompress_zio(zio, hdr, cb->l2rcb_compress); 4797 ASSERT(zio->io_data != NULL); 4798 4799 /* 4800 * Check this survived the L2ARC journey. 4801 */ 4802 equal = arc_cksum_equal(buf); 4803 if (equal && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) { 4804 mutex_exit(hash_lock); 4805 zio->io_private = buf; 4806 zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */ 4807 zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */ 4808 arc_read_done(zio); 4809 } else { 4810 mutex_exit(hash_lock); 4811 /* 4812 * Buffer didn't survive caching. Increment stats and 4813 * reissue to the original storage device. 4814 */ 4815 if (zio->io_error != 0) { 4816 ARCSTAT_BUMP(arcstat_l2_io_error); 4817 } else { 4818 zio->io_error = SET_ERROR(EIO); 4819 } 4820 if (!equal) 4821 ARCSTAT_BUMP(arcstat_l2_cksum_bad); 4822 4823 /* 4824 * If there's no waiter, issue an async i/o to the primary 4825 * storage now. If there *is* a waiter, the caller must 4826 * issue the i/o in a context where it's OK to block. 4827 */ 4828 if (zio->io_waiter == NULL) { 4829 zio_t *pio = zio_unique_parent(zio); 4830 4831 ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL); 4832 4833 zio_nowait(zio_read(pio, cb->l2rcb_spa, &cb->l2rcb_bp, 4834 buf->b_data, zio->io_size, arc_read_done, buf, 4835 zio->io_priority, cb->l2rcb_flags, &cb->l2rcb_zb)); 4836 } 4837 } 4838 4839 kmem_free(cb, sizeof (l2arc_read_callback_t)); 4840} 4841 4842/* 4843 * This is the list priority from which the L2ARC will search for pages to 4844 * cache. This is used within loops (0..3) to cycle through lists in the 4845 * desired order. This order can have a significant effect on cache 4846 * performance. 4847 * 4848 * Currently the metadata lists are hit first, MFU then MRU, followed by 4849 * the data lists. This function returns a locked list, and also returns 4850 * the lock pointer. 4851 */ 4852static list_t * 4853l2arc_list_locked(int list_num, kmutex_t **lock) 4854{ 4855 list_t *list = NULL; 4856 int idx; 4857 4858 ASSERT(list_num >= 0 && list_num < 2 * ARC_BUFC_NUMLISTS); 4859 4860 if (list_num < ARC_BUFC_NUMMETADATALISTS) { 4861 idx = list_num; 4862 list = &arc_mfu->arcs_lists[idx]; 4863 *lock = ARCS_LOCK(arc_mfu, idx); 4864 } else if (list_num < ARC_BUFC_NUMMETADATALISTS * 2) { 4865 idx = list_num - ARC_BUFC_NUMMETADATALISTS; 4866 list = &arc_mru->arcs_lists[idx]; 4867 *lock = ARCS_LOCK(arc_mru, idx); 4868 } else if (list_num < (ARC_BUFC_NUMMETADATALISTS * 2 + 4869 ARC_BUFC_NUMDATALISTS)) { 4870 idx = list_num - ARC_BUFC_NUMMETADATALISTS; 4871 list = &arc_mfu->arcs_lists[idx]; 4872 *lock = ARCS_LOCK(arc_mfu, idx); 4873 } else { 4874 idx = list_num - ARC_BUFC_NUMLISTS; 4875 list = &arc_mru->arcs_lists[idx]; 4876 *lock = ARCS_LOCK(arc_mru, idx); 4877 } 4878 4879 ASSERT(!(MUTEX_HELD(*lock))); 4880 mutex_enter(*lock); 4881 return (list); 4882} 4883 4884/* 4885 * Evict buffers from the device write hand to the distance specified in 4886 * bytes. This distance may span populated buffers, it may span nothing. 4887 * This is clearing a region on the L2ARC device ready for writing. 4888 * If the 'all' boolean is set, every buffer is evicted. 4889 */ 4890static void 4891l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all) 4892{ 4893 list_t *buflist; 4894 l2arc_buf_hdr_t *abl2; 4895 arc_buf_hdr_t *ab, *ab_prev; 4896 kmutex_t *hash_lock; 4897 uint64_t taddr; 4898 int64_t bytes_evicted = 0; 4899 4900 buflist = dev->l2ad_buflist; 4901 4902 if (buflist == NULL) 4903 return; 4904 4905 if (!all && dev->l2ad_first) { 4906 /* 4907 * This is the first sweep through the device. There is 4908 * nothing to evict. 4909 */ 4910 return; 4911 } 4912 4913 if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) { 4914 /* 4915 * When nearing the end of the device, evict to the end 4916 * before the device write hand jumps to the start. 4917 */ 4918 taddr = dev->l2ad_end; 4919 } else { 4920 taddr = dev->l2ad_hand + distance; 4921 } 4922 DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist, 4923 uint64_t, taddr, boolean_t, all); 4924 4925top: 4926 mutex_enter(&l2arc_buflist_mtx); 4927 for (ab = list_tail(buflist); ab; ab = ab_prev) { 4928 ab_prev = list_prev(buflist, ab); 4929 4930 hash_lock = HDR_LOCK(ab); 4931 if (!mutex_tryenter(hash_lock)) { 4932 /* 4933 * Missed the hash lock. Retry. 4934 */ 4935 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry); 4936 mutex_exit(&l2arc_buflist_mtx); 4937 mutex_enter(hash_lock); 4938 mutex_exit(hash_lock); 4939 goto top; 4940 } 4941 4942 if (HDR_L2_WRITE_HEAD(ab)) { 4943 /* 4944 * We hit a write head node. Leave it for 4945 * l2arc_write_done(). 4946 */ 4947 list_remove(buflist, ab); 4948 mutex_exit(hash_lock); 4949 continue; 4950 } 4951 4952 if (!all && ab->b_l2hdr != NULL && 4953 (ab->b_l2hdr->b_daddr > taddr || 4954 ab->b_l2hdr->b_daddr < dev->l2ad_hand)) { 4955 /* 4956 * We've evicted to the target address, 4957 * or the end of the device. 4958 */ 4959 mutex_exit(hash_lock); 4960 break; 4961 } 4962 4963 if (HDR_FREE_IN_PROGRESS(ab)) { 4964 /* 4965 * Already on the path to destruction. 4966 */ 4967 mutex_exit(hash_lock); 4968 continue; 4969 } 4970 4971 if (ab->b_state == arc_l2c_only) { 4972 ASSERT(!HDR_L2_READING(ab)); 4973 /* 4974 * This doesn't exist in the ARC. Destroy. 4975 * arc_hdr_destroy() will call list_remove() 4976 * and decrement arcstat_l2_size. 4977 */ 4978 arc_change_state(arc_anon, ab, hash_lock); 4979 arc_hdr_destroy(ab); 4980 } else { 4981 /* 4982 * Invalidate issued or about to be issued 4983 * reads, since we may be about to write 4984 * over this location. 4985 */ 4986 if (HDR_L2_READING(ab)) { 4987 ARCSTAT_BUMP(arcstat_l2_evict_reading); 4988 ab->b_flags |= ARC_L2_EVICTED; 4989 } 4990 4991 /* 4992 * Tell ARC this no longer exists in L2ARC. 4993 */ 4994 if (ab->b_l2hdr != NULL) { 4995 abl2 = ab->b_l2hdr; 4996 ARCSTAT_INCR(arcstat_l2_asize, -abl2->b_asize); 4997 bytes_evicted += abl2->b_asize; 4998 ab->b_l2hdr = NULL; 4999 /* 5000 * We are destroying l2hdr, so ensure that 5001 * its compressed buffer, if any, is not leaked. 5002 */ 5003 ASSERT(abl2->b_tmp_cdata == NULL); 5004 kmem_free(abl2, sizeof (l2arc_buf_hdr_t)); 5005 ARCSTAT_INCR(arcstat_l2_size, -ab->b_size); 5006 } 5007 list_remove(buflist, ab); 5008 5009 /* 5010 * This may have been leftover after a 5011 * failed write. 5012 */ 5013 ab->b_flags &= ~ARC_L2_WRITING; 5014 } 5015 mutex_exit(hash_lock); 5016 } 5017 mutex_exit(&l2arc_buflist_mtx); 5018 5019 vdev_space_update(dev->l2ad_vdev, -bytes_evicted, 0, 0); 5020 dev->l2ad_evict = taddr; 5021} 5022 5023/* 5024 * Find and write ARC buffers to the L2ARC device. 5025 * 5026 * An ARC_L2_WRITING flag is set so that the L2ARC buffers are not valid 5027 * for reading until they have completed writing. 5028 * The headroom_boost is an in-out parameter used to maintain headroom boost 5029 * state between calls to this function. 5030 * 5031 * Returns the number of bytes actually written (which may be smaller than 5032 * the delta by which the device hand has changed due to alignment). 5033 */ 5034static uint64_t 5035l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz, 5036 boolean_t *headroom_boost) 5037{ 5038 arc_buf_hdr_t *ab, *ab_prev, *head; 5039 list_t *list; 5040 uint64_t write_asize, write_psize, write_sz, headroom, 5041 buf_compress_minsz; 5042 void *buf_data; 5043 kmutex_t *list_lock; 5044 boolean_t full; 5045 l2arc_write_callback_t *cb; 5046 zio_t *pio, *wzio; 5047 uint64_t guid = spa_load_guid(spa); 5048 const boolean_t do_headroom_boost = *headroom_boost; 5049 int try; 5050 5051 ASSERT(dev->l2ad_vdev != NULL); 5052 5053 /* Lower the flag now, we might want to raise it again later. */ 5054 *headroom_boost = B_FALSE; 5055 5056 pio = NULL; 5057 write_sz = write_asize = write_psize = 0; 5058 full = B_FALSE; 5059 head = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE); 5060 head->b_flags |= ARC_L2_WRITE_HEAD; 5061 5062 ARCSTAT_BUMP(arcstat_l2_write_buffer_iter); 5063 /* 5064 * We will want to try to compress buffers that are at least 2x the 5065 * device sector size. 5066 */ 5067 buf_compress_minsz = 2 << dev->l2ad_vdev->vdev_ashift; 5068 5069 /* 5070 * Copy buffers for L2ARC writing. 5071 */ 5072 mutex_enter(&l2arc_buflist_mtx); 5073 for (try = 0; try < 2 * ARC_BUFC_NUMLISTS; try++) { 5074 uint64_t passed_sz = 0; 5075 5076 list = l2arc_list_locked(try, &list_lock); 5077 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_iter); 5078 5079 /* 5080 * L2ARC fast warmup. 5081 * 5082 * Until the ARC is warm and starts to evict, read from the 5083 * head of the ARC lists rather than the tail. 5084 */ 5085 if (arc_warm == B_FALSE) 5086 ab = list_head(list); 5087 else 5088 ab = list_tail(list); 5089 if (ab == NULL) 5090 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_null_iter); 5091 5092 headroom = target_sz * l2arc_headroom * 2 / ARC_BUFC_NUMLISTS; 5093 if (do_headroom_boost) 5094 headroom = (headroom * l2arc_headroom_boost) / 100; 5095 5096 for (; ab; ab = ab_prev) { 5097 l2arc_buf_hdr_t *l2hdr; 5098 kmutex_t *hash_lock; 5099 uint64_t buf_sz; 5100 5101 if (arc_warm == B_FALSE) 5102 ab_prev = list_next(list, ab); 5103 else 5104 ab_prev = list_prev(list, ab); 5105 ARCSTAT_INCR(arcstat_l2_write_buffer_bytes_scanned, ab->b_size); 5106 5107 hash_lock = HDR_LOCK(ab); 5108 if (!mutex_tryenter(hash_lock)) { 5109 ARCSTAT_BUMP(arcstat_l2_write_trylock_fail); 5110 /* 5111 * Skip this buffer rather than waiting. 5112 */ 5113 continue; 5114 } 5115 5116 passed_sz += ab->b_size; 5117 if (passed_sz > headroom) { 5118 /* 5119 * Searched too far. 5120 */ 5121 mutex_exit(hash_lock); 5122 ARCSTAT_BUMP(arcstat_l2_write_passed_headroom); 5123 break; 5124 } 5125 5126 if (!l2arc_write_eligible(guid, ab)) { 5127 mutex_exit(hash_lock); 5128 continue; 5129 } 5130 5131 if ((write_sz + ab->b_size) > target_sz) { 5132 full = B_TRUE; 5133 mutex_exit(hash_lock); 5134 ARCSTAT_BUMP(arcstat_l2_write_full); 5135 break; 5136 } 5137 5138 if (pio == NULL) { 5139 /* 5140 * Insert a dummy header on the buflist so 5141 * l2arc_write_done() can find where the 5142 * write buffers begin without searching. 5143 */ 5144 list_insert_head(dev->l2ad_buflist, head); 5145 5146 cb = kmem_alloc( 5147 sizeof (l2arc_write_callback_t), KM_SLEEP); 5148 cb->l2wcb_dev = dev; 5149 cb->l2wcb_head = head; 5150 pio = zio_root(spa, l2arc_write_done, cb, 5151 ZIO_FLAG_CANFAIL); 5152 ARCSTAT_BUMP(arcstat_l2_write_pios); 5153 } 5154 5155 /* 5156 * Create and add a new L2ARC header. 5157 */ 5158 l2hdr = kmem_zalloc(sizeof (l2arc_buf_hdr_t), KM_SLEEP); 5159 l2hdr->b_dev = dev; 5160 ab->b_flags |= ARC_L2_WRITING; 5161 5162 /* 5163 * Temporarily stash the data buffer in b_tmp_cdata. 5164 * The subsequent write step will pick it up from 5165 * there. This is because can't access ab->b_buf 5166 * without holding the hash_lock, which we in turn 5167 * can't access without holding the ARC list locks 5168 * (which we want to avoid during compression/writing). 5169 */ 5170 l2hdr->b_compress = ZIO_COMPRESS_OFF; 5171 l2hdr->b_asize = ab->b_size; 5172 l2hdr->b_tmp_cdata = ab->b_buf->b_data; 5173 5174 buf_sz = ab->b_size; 5175 ab->b_l2hdr = l2hdr; 5176 5177 list_insert_head(dev->l2ad_buflist, ab); 5178 5179 /* 5180 * Compute and store the buffer cksum before 5181 * writing. On debug the cksum is verified first. 5182 */ 5183 arc_cksum_verify(ab->b_buf); 5184 arc_cksum_compute(ab->b_buf, B_TRUE); 5185 5186 mutex_exit(hash_lock); 5187 5188 write_sz += buf_sz; 5189 } 5190 5191 mutex_exit(list_lock); 5192 5193 if (full == B_TRUE) 5194 break; 5195 } 5196 5197 /* No buffers selected for writing? */ 5198 if (pio == NULL) { 5199 ASSERT0(write_sz); 5200 mutex_exit(&l2arc_buflist_mtx); 5201 kmem_cache_free(hdr_cache, head); 5202 return (0); 5203 } 5204 5205 /* 5206 * Now start writing the buffers. We're starting at the write head 5207 * and work backwards, retracing the course of the buffer selector 5208 * loop above. 5209 */ 5210 for (ab = list_prev(dev->l2ad_buflist, head); ab; 5211 ab = list_prev(dev->l2ad_buflist, ab)) { 5212 l2arc_buf_hdr_t *l2hdr; 5213 uint64_t buf_sz; 5214 5215 /* 5216 * We shouldn't need to lock the buffer here, since we flagged 5217 * it as ARC_L2_WRITING in the previous step, but we must take 5218 * care to only access its L2 cache parameters. In particular, 5219 * ab->b_buf may be invalid by now due to ARC eviction. 5220 */ 5221 l2hdr = ab->b_l2hdr; 5222 l2hdr->b_daddr = dev->l2ad_hand; 5223 5224 if ((ab->b_flags & ARC_L2COMPRESS) && 5225 l2hdr->b_asize >= buf_compress_minsz) { 5226 if (l2arc_compress_buf(l2hdr)) { 5227 /* 5228 * If compression succeeded, enable headroom 5229 * boost on the next scan cycle. 5230 */ 5231 *headroom_boost = B_TRUE; 5232 } 5233 } 5234 5235 /* 5236 * Pick up the buffer data we had previously stashed away 5237 * (and now potentially also compressed). 5238 */ 5239 buf_data = l2hdr->b_tmp_cdata; 5240 buf_sz = l2hdr->b_asize; 5241 5242 /* 5243 * If the data has not been compressed, then clear b_tmp_cdata 5244 * to make sure that it points only to a temporary compression 5245 * buffer. 5246 */ 5247 if (!L2ARC_IS_VALID_COMPRESS(l2hdr->b_compress)) 5248 l2hdr->b_tmp_cdata = NULL; 5249 5250 /* Compression may have squashed the buffer to zero length. */ 5251 if (buf_sz != 0) { 5252 uint64_t buf_p_sz; 5253 5254 wzio = zio_write_phys(pio, dev->l2ad_vdev, 5255 dev->l2ad_hand, buf_sz, buf_data, ZIO_CHECKSUM_OFF, 5256 NULL, NULL, ZIO_PRIORITY_ASYNC_WRITE, 5257 ZIO_FLAG_CANFAIL, B_FALSE); 5258 5259 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev, 5260 zio_t *, wzio); 5261 (void) zio_nowait(wzio); 5262 5263 write_asize += buf_sz; 5264 /* 5265 * Keep the clock hand suitably device-aligned. 5266 */ 5267 buf_p_sz = vdev_psize_to_asize(dev->l2ad_vdev, buf_sz); 5268 write_psize += buf_p_sz; 5269 dev->l2ad_hand += buf_p_sz; 5270 } 5271 } 5272 5273 mutex_exit(&l2arc_buflist_mtx); 5274 5275 ASSERT3U(write_asize, <=, target_sz); 5276 ARCSTAT_BUMP(arcstat_l2_writes_sent); 5277 ARCSTAT_INCR(arcstat_l2_write_bytes, write_asize); 5278 ARCSTAT_INCR(arcstat_l2_size, write_sz); 5279 ARCSTAT_INCR(arcstat_l2_asize, write_asize); 5280 vdev_space_update(dev->l2ad_vdev, write_psize, 0, 0); 5281 5282 /* 5283 * Bump device hand to the device start if it is approaching the end. 5284 * l2arc_evict() will already have evicted ahead for this case. 5285 */ 5286 if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) { 5287 dev->l2ad_hand = dev->l2ad_start; 5288 dev->l2ad_evict = dev->l2ad_start; 5289 dev->l2ad_first = B_FALSE; 5290 } 5291 5292 dev->l2ad_writing = B_TRUE; 5293 (void) zio_wait(pio); 5294 dev->l2ad_writing = B_FALSE; 5295 5296 return (write_asize); 5297} 5298 5299/* 5300 * Compresses an L2ARC buffer. 5301 * The data to be compressed must be prefilled in l2hdr->b_tmp_cdata and its 5302 * size in l2hdr->b_asize. This routine tries to compress the data and 5303 * depending on the compression result there are three possible outcomes: 5304 * *) The buffer was incompressible. The original l2hdr contents were left 5305 * untouched and are ready for writing to an L2 device. 5306 * *) The buffer was all-zeros, so there is no need to write it to an L2 5307 * device. To indicate this situation b_tmp_cdata is NULL'ed, b_asize is 5308 * set to zero and b_compress is set to ZIO_COMPRESS_EMPTY. 5309 * *) Compression succeeded and b_tmp_cdata was replaced with a temporary 5310 * data buffer which holds the compressed data to be written, and b_asize 5311 * tells us how much data there is. b_compress is set to the appropriate 5312 * compression algorithm. Once writing is done, invoke 5313 * l2arc_release_cdata_buf on this l2hdr to free this temporary buffer. 5314 * 5315 * Returns B_TRUE if compression succeeded, or B_FALSE if it didn't (the 5316 * buffer was incompressible). 5317 */ 5318static boolean_t 5319l2arc_compress_buf(l2arc_buf_hdr_t *l2hdr) 5320{ 5321 void *cdata; 5322 size_t csize, len, rounded; 5323 5324 ASSERT(l2hdr->b_compress == ZIO_COMPRESS_OFF); 5325 ASSERT(l2hdr->b_tmp_cdata != NULL); 5326 5327 len = l2hdr->b_asize; 5328 cdata = zio_data_buf_alloc(len); 5329 csize = zio_compress_data(ZIO_COMPRESS_LZ4, l2hdr->b_tmp_cdata, 5330 cdata, l2hdr->b_asize); 5331 5332 rounded = P2ROUNDUP(csize, (size_t)SPA_MINBLOCKSIZE); 5333 if (rounded > csize) { 5334 bzero((char *)cdata + csize, rounded - csize); 5335 csize = rounded; 5336 } 5337 5338 if (csize == 0) { 5339 /* zero block, indicate that there's nothing to write */ 5340 zio_data_buf_free(cdata, len); 5341 l2hdr->b_compress = ZIO_COMPRESS_EMPTY; 5342 l2hdr->b_asize = 0; 5343 l2hdr->b_tmp_cdata = NULL; 5344 ARCSTAT_BUMP(arcstat_l2_compress_zeros); 5345 return (B_TRUE); 5346 } else if (csize > 0 && csize < len) { 5347 /* 5348 * Compression succeeded, we'll keep the cdata around for 5349 * writing and release it afterwards. 5350 */ 5351 l2hdr->b_compress = ZIO_COMPRESS_LZ4; 5352 l2hdr->b_asize = csize; 5353 l2hdr->b_tmp_cdata = cdata; 5354 ARCSTAT_BUMP(arcstat_l2_compress_successes); 5355 return (B_TRUE); 5356 } else { 5357 /* 5358 * Compression failed, release the compressed buffer. 5359 * l2hdr will be left unmodified. 5360 */ 5361 zio_data_buf_free(cdata, len); 5362 ARCSTAT_BUMP(arcstat_l2_compress_failures); 5363 return (B_FALSE); 5364 } 5365} 5366 5367/* 5368 * Decompresses a zio read back from an l2arc device. On success, the 5369 * underlying zio's io_data buffer is overwritten by the uncompressed 5370 * version. On decompression error (corrupt compressed stream), the 5371 * zio->io_error value is set to signal an I/O error. 5372 * 5373 * Please note that the compressed data stream is not checksummed, so 5374 * if the underlying device is experiencing data corruption, we may feed 5375 * corrupt data to the decompressor, so the decompressor needs to be 5376 * able to handle this situation (LZ4 does). 5377 */ 5378static void 5379l2arc_decompress_zio(zio_t *zio, arc_buf_hdr_t *hdr, enum zio_compress c) 5380{ 5381 ASSERT(L2ARC_IS_VALID_COMPRESS(c)); 5382 5383 if (zio->io_error != 0) { 5384 /* 5385 * An io error has occured, just restore the original io 5386 * size in preparation for a main pool read. 5387 */ 5388 zio->io_orig_size = zio->io_size = hdr->b_size; 5389 return; 5390 } 5391 5392 if (c == ZIO_COMPRESS_EMPTY) { 5393 /* 5394 * An empty buffer results in a null zio, which means we 5395 * need to fill its io_data after we're done restoring the 5396 * buffer's contents. 5397 */ 5398 ASSERT(hdr->b_buf != NULL); 5399 bzero(hdr->b_buf->b_data, hdr->b_size); 5400 zio->io_data = zio->io_orig_data = hdr->b_buf->b_data; 5401 } else { 5402 ASSERT(zio->io_data != NULL); 5403 /* 5404 * We copy the compressed data from the start of the arc buffer 5405 * (the zio_read will have pulled in only what we need, the 5406 * rest is garbage which we will overwrite at decompression) 5407 * and then decompress back to the ARC data buffer. This way we 5408 * can minimize copying by simply decompressing back over the 5409 * original compressed data (rather than decompressing to an 5410 * aux buffer and then copying back the uncompressed buffer, 5411 * which is likely to be much larger). 5412 */ 5413 uint64_t csize; 5414 void *cdata; 5415 5416 csize = zio->io_size; 5417 cdata = zio_data_buf_alloc(csize); 5418 bcopy(zio->io_data, cdata, csize); 5419 if (zio_decompress_data(c, cdata, zio->io_data, csize, 5420 hdr->b_size) != 0) 5421 zio->io_error = EIO; 5422 zio_data_buf_free(cdata, csize); 5423 } 5424 5425 /* Restore the expected uncompressed IO size. */ 5426 zio->io_orig_size = zio->io_size = hdr->b_size; 5427} 5428 5429/* 5430 * Releases the temporary b_tmp_cdata buffer in an l2arc header structure. 5431 * This buffer serves as a temporary holder of compressed data while 5432 * the buffer entry is being written to an l2arc device. Once that is 5433 * done, we can dispose of it. 5434 */ 5435static void 5436l2arc_release_cdata_buf(arc_buf_hdr_t *ab) 5437{ 5438 l2arc_buf_hdr_t *l2hdr = ab->b_l2hdr; 5439 5440 ASSERT(L2ARC_IS_VALID_COMPRESS(l2hdr->b_compress)); 5441 if (l2hdr->b_compress != ZIO_COMPRESS_EMPTY) { 5442 /* 5443 * If the data was compressed, then we've allocated a 5444 * temporary buffer for it, so now we need to release it. 5445 */ 5446 ASSERT(l2hdr->b_tmp_cdata != NULL); 5447 zio_data_buf_free(l2hdr->b_tmp_cdata, ab->b_size); 5448 l2hdr->b_tmp_cdata = NULL; 5449 } else { 5450 ASSERT(l2hdr->b_tmp_cdata == NULL); 5451 } 5452} 5453 5454/* 5455 * This thread feeds the L2ARC at regular intervals. This is the beating 5456 * heart of the L2ARC. 5457 */ 5458static void 5459l2arc_feed_thread(void *dummy __unused) 5460{ 5461 callb_cpr_t cpr; 5462 l2arc_dev_t *dev; 5463 spa_t *spa; 5464 uint64_t size, wrote; 5465 clock_t begin, next = ddi_get_lbolt(); 5466 boolean_t headroom_boost = B_FALSE; 5467 5468 CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG); 5469 5470 mutex_enter(&l2arc_feed_thr_lock); 5471 5472 while (l2arc_thread_exit == 0) { 5473 CALLB_CPR_SAFE_BEGIN(&cpr); 5474 (void) cv_timedwait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock, 5475 next - ddi_get_lbolt()); 5476 CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock); 5477 next = ddi_get_lbolt() + hz; 5478 5479 /* 5480 * Quick check for L2ARC devices. 5481 */ 5482 mutex_enter(&l2arc_dev_mtx); 5483 if (l2arc_ndev == 0) { 5484 mutex_exit(&l2arc_dev_mtx); 5485 continue; 5486 } 5487 mutex_exit(&l2arc_dev_mtx); 5488 begin = ddi_get_lbolt(); 5489 5490 /* 5491 * This selects the next l2arc device to write to, and in 5492 * doing so the next spa to feed from: dev->l2ad_spa. This 5493 * will return NULL if there are now no l2arc devices or if 5494 * they are all faulted. 5495 * 5496 * If a device is returned, its spa's config lock is also 5497 * held to prevent device removal. l2arc_dev_get_next() 5498 * will grab and release l2arc_dev_mtx. 5499 */ 5500 if ((dev = l2arc_dev_get_next()) == NULL) 5501 continue; 5502 5503 spa = dev->l2ad_spa; 5504 ASSERT(spa != NULL); 5505 5506 /* 5507 * If the pool is read-only then force the feed thread to 5508 * sleep a little longer. 5509 */ 5510 if (!spa_writeable(spa)) { 5511 next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz; 5512 spa_config_exit(spa, SCL_L2ARC, dev); 5513 continue; 5514 } 5515 5516 /* 5517 * Avoid contributing to memory pressure. 5518 */ 5519 if (arc_reclaim_needed()) { 5520 ARCSTAT_BUMP(arcstat_l2_abort_lowmem); 5521 spa_config_exit(spa, SCL_L2ARC, dev); 5522 continue; 5523 } 5524 5525 ARCSTAT_BUMP(arcstat_l2_feeds); 5526 5527 size = l2arc_write_size(); 5528 5529 /* 5530 * Evict L2ARC buffers that will be overwritten. 5531 */ 5532 l2arc_evict(dev, size, B_FALSE); 5533 5534 /* 5535 * Write ARC buffers. 5536 */ 5537 wrote = l2arc_write_buffers(spa, dev, size, &headroom_boost); 5538 5539 /* 5540 * Calculate interval between writes. 5541 */ 5542 next = l2arc_write_interval(begin, size, wrote); 5543 spa_config_exit(spa, SCL_L2ARC, dev); 5544 } 5545 5546 l2arc_thread_exit = 0; 5547 cv_broadcast(&l2arc_feed_thr_cv); 5548 CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */ 5549 thread_exit(); 5550} 5551 5552boolean_t 5553l2arc_vdev_present(vdev_t *vd) 5554{ 5555 l2arc_dev_t *dev; 5556 5557 mutex_enter(&l2arc_dev_mtx); 5558 for (dev = list_head(l2arc_dev_list); dev != NULL; 5559 dev = list_next(l2arc_dev_list, dev)) { 5560 if (dev->l2ad_vdev == vd) 5561 break; 5562 } 5563 mutex_exit(&l2arc_dev_mtx); 5564 5565 return (dev != NULL); 5566} 5567 5568/* 5569 * Add a vdev for use by the L2ARC. By this point the spa has already 5570 * validated the vdev and opened it. 5571 */ 5572void 5573l2arc_add_vdev(spa_t *spa, vdev_t *vd) 5574{ 5575 l2arc_dev_t *adddev; 5576 5577 ASSERT(!l2arc_vdev_present(vd)); 5578 5579 vdev_ashift_optimize(vd); 5580 5581 /* 5582 * Create a new l2arc device entry. 5583 */ 5584 adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP); 5585 adddev->l2ad_spa = spa; 5586 adddev->l2ad_vdev = vd; 5587 adddev->l2ad_start = VDEV_LABEL_START_SIZE; 5588 adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd); 5589 adddev->l2ad_hand = adddev->l2ad_start; 5590 adddev->l2ad_evict = adddev->l2ad_start; 5591 adddev->l2ad_first = B_TRUE; 5592 adddev->l2ad_writing = B_FALSE; 5593 5594 /* 5595 * This is a list of all ARC buffers that are still valid on the 5596 * device. 5597 */ 5598 adddev->l2ad_buflist = kmem_zalloc(sizeof (list_t), KM_SLEEP); 5599 list_create(adddev->l2ad_buflist, sizeof (arc_buf_hdr_t), 5600 offsetof(arc_buf_hdr_t, b_l2node)); 5601 5602 vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand); 5603 5604 /* 5605 * Add device to global list 5606 */ 5607 mutex_enter(&l2arc_dev_mtx); 5608 list_insert_head(l2arc_dev_list, adddev); 5609 atomic_inc_64(&l2arc_ndev); 5610 mutex_exit(&l2arc_dev_mtx); 5611} 5612 5613/* 5614 * Remove a vdev from the L2ARC. 5615 */ 5616void 5617l2arc_remove_vdev(vdev_t *vd) 5618{ 5619 l2arc_dev_t *dev, *nextdev, *remdev = NULL; 5620 5621 /* 5622 * Find the device by vdev 5623 */ 5624 mutex_enter(&l2arc_dev_mtx); 5625 for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) { 5626 nextdev = list_next(l2arc_dev_list, dev); 5627 if (vd == dev->l2ad_vdev) { 5628 remdev = dev; 5629 break; 5630 } 5631 } 5632 ASSERT(remdev != NULL); 5633 5634 /* 5635 * Remove device from global list 5636 */ 5637 list_remove(l2arc_dev_list, remdev); 5638 l2arc_dev_last = NULL; /* may have been invalidated */ 5639 atomic_dec_64(&l2arc_ndev); 5640 mutex_exit(&l2arc_dev_mtx); 5641 5642 /* 5643 * Clear all buflists and ARC references. L2ARC device flush. 5644 */ 5645 l2arc_evict(remdev, 0, B_TRUE); 5646 list_destroy(remdev->l2ad_buflist); 5647 kmem_free(remdev->l2ad_buflist, sizeof (list_t)); 5648 kmem_free(remdev, sizeof (l2arc_dev_t)); 5649} 5650 5651void 5652l2arc_init(void) 5653{ 5654 l2arc_thread_exit = 0; 5655 l2arc_ndev = 0; 5656 l2arc_writes_sent = 0; 5657 l2arc_writes_done = 0; 5658 5659 mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL); 5660 cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL); 5661 mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL); 5662 mutex_init(&l2arc_buflist_mtx, NULL, MUTEX_DEFAULT, NULL); 5663 mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL); 5664 5665 l2arc_dev_list = &L2ARC_dev_list; 5666 l2arc_free_on_write = &L2ARC_free_on_write; 5667 list_create(l2arc_dev_list, sizeof (l2arc_dev_t), 5668 offsetof(l2arc_dev_t, l2ad_node)); 5669 list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t), 5670 offsetof(l2arc_data_free_t, l2df_list_node)); 5671} 5672 5673void 5674l2arc_fini(void) 5675{ 5676 /* 5677 * This is called from dmu_fini(), which is called from spa_fini(); 5678 * Because of this, we can assume that all l2arc devices have 5679 * already been removed when the pools themselves were removed. 5680 */ 5681 5682 l2arc_do_free_on_write(); 5683 5684 mutex_destroy(&l2arc_feed_thr_lock); 5685 cv_destroy(&l2arc_feed_thr_cv); 5686 mutex_destroy(&l2arc_dev_mtx); 5687 mutex_destroy(&l2arc_buflist_mtx); 5688 mutex_destroy(&l2arc_free_on_write_mtx); 5689 5690 list_destroy(l2arc_dev_list); 5691 list_destroy(l2arc_free_on_write); 5692} 5693 5694void 5695l2arc_start(void) 5696{ 5697 if (!(spa_mode_global & FWRITE)) 5698 return; 5699 5700 (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0, 5701 TS_RUN, minclsyspri); 5702} 5703 5704void 5705l2arc_stop(void) 5706{ 5707 if (!(spa_mode_global & FWRITE)) 5708 return; 5709 5710 mutex_enter(&l2arc_feed_thr_lock); 5711 cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */ 5712 l2arc_thread_exit = 1; 5713 while (l2arc_thread_exit != 0) 5714 cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock); 5715 mutex_exit(&l2arc_feed_thr_lock); 5716} 5717