arc.c revision 288547
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) 2012, Joyent, Inc. All rights reserved. 24 * Copyright (c) 2011, 2014 by Delphix. All rights reserved. 25 * Copyright (c) 2014 by Saso Kiselkov. All rights reserved. 26 * Copyright 2014 Nexenta Systems, Inc. All rights reserved. 27 */ 28 29/* 30 * DVA-based Adjustable Replacement Cache 31 * 32 * While much of the theory of operation used here is 33 * based on the self-tuning, low overhead replacement cache 34 * presented by Megiddo and Modha at FAST 2003, there are some 35 * significant differences: 36 * 37 * 1. The Megiddo and Modha model assumes any page is evictable. 38 * Pages in its cache cannot be "locked" into memory. This makes 39 * the eviction algorithm simple: evict the last page in the list. 40 * This also make the performance characteristics easy to reason 41 * about. Our cache is not so simple. At any given moment, some 42 * subset of the blocks in the cache are un-evictable because we 43 * have handed out a reference to them. Blocks are only evictable 44 * when there are no external references active. This makes 45 * eviction far more problematic: we choose to evict the evictable 46 * blocks that are the "lowest" in the list. 47 * 48 * There are times when it is not possible to evict the requested 49 * space. In these circumstances we are unable to adjust the cache 50 * size. To prevent the cache growing unbounded at these times we 51 * implement a "cache throttle" that slows the flow of new data 52 * into the cache until we can make space available. 53 * 54 * 2. The Megiddo and Modha model assumes a fixed cache size. 55 * Pages are evicted when the cache is full and there is a cache 56 * miss. Our model has a variable sized cache. It grows with 57 * high use, but also tries to react to memory pressure from the 58 * operating system: decreasing its size when system memory is 59 * tight. 60 * 61 * 3. The Megiddo and Modha model assumes a fixed page size. All 62 * elements of the cache are therefore exactly the same size. So 63 * when adjusting the cache size following a cache miss, its simply 64 * a matter of choosing a single page to evict. In our model, we 65 * have variable sized cache blocks (rangeing from 512 bytes to 66 * 128K bytes). We therefore choose a set of blocks to evict to make 67 * space for a cache miss that approximates as closely as possible 68 * the space used by the new block. 69 * 70 * See also: "ARC: A Self-Tuning, Low Overhead Replacement Cache" 71 * by N. Megiddo & D. Modha, FAST 2003 72 */ 73 74/* 75 * The locking model: 76 * 77 * A new reference to a cache buffer can be obtained in two 78 * ways: 1) via a hash table lookup using the DVA as a key, 79 * or 2) via one of the ARC lists. The arc_read() interface 80 * uses method 1, while the internal arc algorithms for 81 * adjusting the cache use method 2. We therefore provide two 82 * types of locks: 1) the hash table lock array, and 2) the 83 * arc list locks. 84 * 85 * Buffers do not have their own mutexs, rather they rely on the 86 * hash table mutexs for the bulk of their protection (i.e. most 87 * fields in the arc_buf_hdr_t are protected by these mutexs). 88 * 89 * buf_hash_find() returns the appropriate mutex (held) when it 90 * locates the requested buffer in the hash table. It returns 91 * NULL for the mutex if the buffer was not in the table. 92 * 93 * buf_hash_remove() expects the appropriate hash mutex to be 94 * already held before it is invoked. 95 * 96 * Each arc state also has a mutex which is used to protect the 97 * buffer list associated with the state. When attempting to 98 * obtain a hash table lock while holding an arc list lock you 99 * must use: mutex_tryenter() to avoid deadlock. Also note that 100 * the active state mutex must be held before the ghost state mutex. 101 * 102 * Arc buffers may have an associated eviction callback function. 103 * This function will be invoked prior to removing the buffer (e.g. 104 * in arc_do_user_evicts()). Note however that the data associated 105 * with the buffer may be evicted prior to the callback. The callback 106 * must be made with *no locks held* (to prevent deadlock). Additionally, 107 * the users of callbacks must ensure that their private data is 108 * protected from simultaneous callbacks from arc_clear_callback() 109 * and arc_do_user_evicts(). 110 * 111 * Note that the majority of the performance stats are manipulated 112 * with atomic operations. 113 * 114 * The L2ARC uses the l2ad_mtx on each vdev for the following: 115 * 116 * - L2ARC buflist creation 117 * - L2ARC buflist eviction 118 * - L2ARC write completion, which walks L2ARC buflists 119 * - ARC header destruction, as it removes from L2ARC buflists 120 * - ARC header release, as it removes from L2ARC buflists 121 */ 122 123#include <sys/spa.h> 124#include <sys/zio.h> 125#include <sys/zio_compress.h> 126#include <sys/zfs_context.h> 127#include <sys/arc.h> 128#include <sys/refcount.h> 129#include <sys/vdev.h> 130#include <sys/vdev_impl.h> 131#include <sys/dsl_pool.h> 132#ifdef _KERNEL 133#include <sys/dnlc.h> 134#endif 135#include <sys/callb.h> 136#include <sys/kstat.h> 137#include <sys/trim_map.h> 138#include <zfs_fletcher.h> 139#include <sys/sdt.h> 140 141#include <vm/vm_pageout.h> 142#include <machine/vmparam.h> 143 144#ifdef illumos 145#ifndef _KERNEL 146/* set with ZFS_DEBUG=watch, to enable watchpoints on frozen buffers */ 147boolean_t arc_watch = B_FALSE; 148int arc_procfd; 149#endif 150#endif /* illumos */ 151 152static kmutex_t arc_reclaim_thr_lock; 153static kcondvar_t arc_reclaim_thr_cv; /* used to signal reclaim thr */ 154static uint8_t arc_thread_exit; 155 156#define ARC_REDUCE_DNLC_PERCENT 3 157uint_t arc_reduce_dnlc_percent = ARC_REDUCE_DNLC_PERCENT; 158 159typedef enum arc_reclaim_strategy { 160 ARC_RECLAIM_AGGR, /* Aggressive reclaim strategy */ 161 ARC_RECLAIM_CONS /* Conservative reclaim strategy */ 162} arc_reclaim_strategy_t; 163 164/* 165 * The number of iterations through arc_evict_*() before we 166 * drop & reacquire the lock. 167 */ 168int arc_evict_iterations = 100; 169 170/* number of seconds before growing cache again */ 171static int arc_grow_retry = 60; 172 173/* shift of arc_c for calculating both min and max arc_p */ 174static int arc_p_min_shift = 4; 175 176/* log2(fraction of arc to reclaim) */ 177static int arc_shrink_shift = 5; 178 179/* 180 * minimum lifespan of a prefetch block in clock ticks 181 * (initialized in arc_init()) 182 */ 183static int arc_min_prefetch_lifespan; 184 185/* 186 * If this percent of memory is free, don't throttle. 187 */ 188int arc_lotsfree_percent = 10; 189 190static int arc_dead; 191extern int zfs_prefetch_disable; 192 193/* 194 * The arc has filled available memory and has now warmed up. 195 */ 196static boolean_t arc_warm; 197 198uint64_t zfs_arc_max; 199uint64_t zfs_arc_min; 200uint64_t zfs_arc_meta_limit = 0; 201uint64_t zfs_arc_meta_min = 0; 202int zfs_arc_grow_retry = 0; 203int zfs_arc_shrink_shift = 0; 204int zfs_arc_p_min_shift = 0; 205int zfs_disable_dup_eviction = 0; 206uint64_t zfs_arc_average_blocksize = 8 * 1024; /* 8KB */ 207u_int zfs_arc_free_target = 0; 208 209static int sysctl_vfs_zfs_arc_free_target(SYSCTL_HANDLER_ARGS); 210static int sysctl_vfs_zfs_arc_meta_limit(SYSCTL_HANDLER_ARGS); 211 212#ifdef _KERNEL 213static void 214arc_free_target_init(void *unused __unused) 215{ 216 217 zfs_arc_free_target = vm_pageout_wakeup_thresh; 218} 219SYSINIT(arc_free_target_init, SI_SUB_KTHREAD_PAGE, SI_ORDER_ANY, 220 arc_free_target_init, NULL); 221 222TUNABLE_QUAD("vfs.zfs.arc_max", &zfs_arc_max); 223TUNABLE_QUAD("vfs.zfs.arc_min", &zfs_arc_min); 224TUNABLE_QUAD("vfs.zfs.arc_meta_limit", &zfs_arc_meta_limit); 225TUNABLE_QUAD("vfs.zfs.arc_meta_min", &zfs_arc_meta_min); 226TUNABLE_QUAD("vfs.zfs.arc_average_blocksize", &zfs_arc_average_blocksize); 227TUNABLE_INT("vfs.zfs.arc_shrink_shift", &zfs_arc_shrink_shift); 228SYSCTL_DECL(_vfs_zfs); 229SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_max, CTLFLAG_RDTUN, &zfs_arc_max, 0, 230 "Maximum ARC size"); 231SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_min, CTLFLAG_RDTUN, &zfs_arc_min, 0, 232 "Minimum ARC size"); 233SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_average_blocksize, CTLFLAG_RDTUN, 234 &zfs_arc_average_blocksize, 0, 235 "ARC average blocksize"); 236SYSCTL_INT(_vfs_zfs, OID_AUTO, arc_shrink_shift, CTLFLAG_RW, 237 &arc_shrink_shift, 0, 238 "log2(fraction of arc to reclaim)"); 239 240/* 241 * We don't have a tunable for arc_free_target due to the dependency on 242 * pagedaemon initialisation. 243 */ 244SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_free_target, 245 CTLTYPE_UINT | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, sizeof(u_int), 246 sysctl_vfs_zfs_arc_free_target, "IU", 247 "Desired number of free pages below which ARC triggers reclaim"); 248 249static int 250sysctl_vfs_zfs_arc_free_target(SYSCTL_HANDLER_ARGS) 251{ 252 u_int val; 253 int err; 254 255 val = zfs_arc_free_target; 256 err = sysctl_handle_int(oidp, &val, 0, req); 257 if (err != 0 || req->newptr == NULL) 258 return (err); 259 260 if (val < minfree) 261 return (EINVAL); 262 if (val > cnt.v_page_count) 263 return (EINVAL); 264 265 zfs_arc_free_target = val; 266 267 return (0); 268} 269 270/* 271 * Must be declared here, before the definition of corresponding kstat 272 * macro which uses the same names will confuse the compiler. 273 */ 274SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_meta_limit, 275 CTLTYPE_U64 | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, sizeof(uint64_t), 276 sysctl_vfs_zfs_arc_meta_limit, "QU", 277 "ARC metadata limit"); 278#endif 279 280/* 281 * Note that buffers can be in one of 6 states: 282 * ARC_anon - anonymous (discussed below) 283 * ARC_mru - recently used, currently cached 284 * ARC_mru_ghost - recentely used, no longer in cache 285 * ARC_mfu - frequently used, currently cached 286 * ARC_mfu_ghost - frequently used, no longer in cache 287 * ARC_l2c_only - exists in L2ARC but not other states 288 * When there are no active references to the buffer, they are 289 * are linked onto a list in one of these arc states. These are 290 * the only buffers that can be evicted or deleted. Within each 291 * state there are multiple lists, one for meta-data and one for 292 * non-meta-data. Meta-data (indirect blocks, blocks of dnodes, 293 * etc.) is tracked separately so that it can be managed more 294 * explicitly: favored over data, limited explicitly. 295 * 296 * Anonymous buffers are buffers that are not associated with 297 * a DVA. These are buffers that hold dirty block copies 298 * before they are written to stable storage. By definition, 299 * they are "ref'd" and are considered part of arc_mru 300 * that cannot be freed. Generally, they will aquire a DVA 301 * as they are written and migrate onto the arc_mru list. 302 * 303 * The ARC_l2c_only state is for buffers that are in the second 304 * level ARC but no longer in any of the ARC_m* lists. The second 305 * level ARC itself may also contain buffers that are in any of 306 * the ARC_m* states - meaning that a buffer can exist in two 307 * places. The reason for the ARC_l2c_only state is to keep the 308 * buffer header in the hash table, so that reads that hit the 309 * second level ARC benefit from these fast lookups. 310 */ 311 312#define ARCS_LOCK_PAD CACHE_LINE_SIZE 313struct arcs_lock { 314 kmutex_t arcs_lock; 315#ifdef _KERNEL 316 unsigned char pad[(ARCS_LOCK_PAD - sizeof (kmutex_t))]; 317#endif 318}; 319 320/* 321 * must be power of two for mask use to work 322 * 323 */ 324#define ARC_BUFC_NUMDATALISTS 16 325#define ARC_BUFC_NUMMETADATALISTS 16 326#define ARC_BUFC_NUMLISTS (ARC_BUFC_NUMMETADATALISTS + ARC_BUFC_NUMDATALISTS) 327 328typedef struct arc_state { 329 uint64_t arcs_lsize[ARC_BUFC_NUMTYPES]; /* amount of evictable data */ 330 uint64_t arcs_size; /* total amount of data in this state */ 331 list_t arcs_lists[ARC_BUFC_NUMLISTS]; /* list of evictable buffers */ 332 struct arcs_lock arcs_locks[ARC_BUFC_NUMLISTS] __aligned(CACHE_LINE_SIZE); 333} arc_state_t; 334 335#define ARCS_LOCK(s, i) (&((s)->arcs_locks[(i)].arcs_lock)) 336 337/* The 6 states: */ 338static arc_state_t ARC_anon; 339static arc_state_t ARC_mru; 340static arc_state_t ARC_mru_ghost; 341static arc_state_t ARC_mfu; 342static arc_state_t ARC_mfu_ghost; 343static arc_state_t ARC_l2c_only; 344 345typedef struct arc_stats { 346 kstat_named_t arcstat_hits; 347 kstat_named_t arcstat_misses; 348 kstat_named_t arcstat_demand_data_hits; 349 kstat_named_t arcstat_demand_data_misses; 350 kstat_named_t arcstat_demand_metadata_hits; 351 kstat_named_t arcstat_demand_metadata_misses; 352 kstat_named_t arcstat_prefetch_data_hits; 353 kstat_named_t arcstat_prefetch_data_misses; 354 kstat_named_t arcstat_prefetch_metadata_hits; 355 kstat_named_t arcstat_prefetch_metadata_misses; 356 kstat_named_t arcstat_mru_hits; 357 kstat_named_t arcstat_mru_ghost_hits; 358 kstat_named_t arcstat_mfu_hits; 359 kstat_named_t arcstat_mfu_ghost_hits; 360 kstat_named_t arcstat_allocated; 361 kstat_named_t arcstat_deleted; 362 kstat_named_t arcstat_stolen; 363 kstat_named_t arcstat_recycle_miss; 364 /* 365 * Number of buffers that could not be evicted because the hash lock 366 * was held by another thread. The lock may not necessarily be held 367 * by something using the same buffer, since hash locks are shared 368 * by multiple buffers. 369 */ 370 kstat_named_t arcstat_mutex_miss; 371 /* 372 * Number of buffers skipped because they have I/O in progress, are 373 * indrect prefetch buffers that have not lived long enough, or are 374 * not from the spa we're trying to evict from. 375 */ 376 kstat_named_t arcstat_evict_skip; 377 kstat_named_t arcstat_evict_l2_cached; 378 kstat_named_t arcstat_evict_l2_eligible; 379 kstat_named_t arcstat_evict_l2_ineligible; 380 kstat_named_t arcstat_hash_elements; 381 kstat_named_t arcstat_hash_elements_max; 382 kstat_named_t arcstat_hash_collisions; 383 kstat_named_t arcstat_hash_chains; 384 kstat_named_t arcstat_hash_chain_max; 385 kstat_named_t arcstat_p; 386 kstat_named_t arcstat_c; 387 kstat_named_t arcstat_c_min; 388 kstat_named_t arcstat_c_max; 389 kstat_named_t arcstat_size; 390 kstat_named_t arcstat_hdr_size; 391 kstat_named_t arcstat_data_size; 392 kstat_named_t arcstat_other_size; 393 kstat_named_t arcstat_l2_hits; 394 kstat_named_t arcstat_l2_misses; 395 kstat_named_t arcstat_l2_feeds; 396 kstat_named_t arcstat_l2_rw_clash; 397 kstat_named_t arcstat_l2_read_bytes; 398 kstat_named_t arcstat_l2_write_bytes; 399 kstat_named_t arcstat_l2_writes_sent; 400 kstat_named_t arcstat_l2_writes_done; 401 kstat_named_t arcstat_l2_writes_error; 402 kstat_named_t arcstat_l2_writes_hdr_miss; 403 kstat_named_t arcstat_l2_evict_lock_retry; 404 kstat_named_t arcstat_l2_evict_reading; 405 kstat_named_t arcstat_l2_evict_l1cached; 406 kstat_named_t arcstat_l2_free_on_write; 407 kstat_named_t arcstat_l2_cdata_free_on_write; 408 kstat_named_t arcstat_l2_abort_lowmem; 409 kstat_named_t arcstat_l2_cksum_bad; 410 kstat_named_t arcstat_l2_io_error; 411 kstat_named_t arcstat_l2_size; 412 kstat_named_t arcstat_l2_asize; 413 kstat_named_t arcstat_l2_hdr_size; 414 kstat_named_t arcstat_l2_compress_successes; 415 kstat_named_t arcstat_l2_compress_zeros; 416 kstat_named_t arcstat_l2_compress_failures; 417 kstat_named_t arcstat_l2_write_trylock_fail; 418 kstat_named_t arcstat_l2_write_passed_headroom; 419 kstat_named_t arcstat_l2_write_spa_mismatch; 420 kstat_named_t arcstat_l2_write_in_l2; 421 kstat_named_t arcstat_l2_write_hdr_io_in_progress; 422 kstat_named_t arcstat_l2_write_not_cacheable; 423 kstat_named_t arcstat_l2_write_full; 424 kstat_named_t arcstat_l2_write_buffer_iter; 425 kstat_named_t arcstat_l2_write_pios; 426 kstat_named_t arcstat_l2_write_buffer_bytes_scanned; 427 kstat_named_t arcstat_l2_write_buffer_list_iter; 428 kstat_named_t arcstat_l2_write_buffer_list_null_iter; 429 kstat_named_t arcstat_memory_throttle_count; 430 kstat_named_t arcstat_duplicate_buffers; 431 kstat_named_t arcstat_duplicate_buffers_size; 432 kstat_named_t arcstat_duplicate_reads; 433 kstat_named_t arcstat_meta_used; 434 kstat_named_t arcstat_meta_limit; 435 kstat_named_t arcstat_meta_max; 436 kstat_named_t arcstat_meta_min; 437} arc_stats_t; 438 439static arc_stats_t arc_stats = { 440 { "hits", KSTAT_DATA_UINT64 }, 441 { "misses", KSTAT_DATA_UINT64 }, 442 { "demand_data_hits", KSTAT_DATA_UINT64 }, 443 { "demand_data_misses", KSTAT_DATA_UINT64 }, 444 { "demand_metadata_hits", KSTAT_DATA_UINT64 }, 445 { "demand_metadata_misses", KSTAT_DATA_UINT64 }, 446 { "prefetch_data_hits", KSTAT_DATA_UINT64 }, 447 { "prefetch_data_misses", KSTAT_DATA_UINT64 }, 448 { "prefetch_metadata_hits", KSTAT_DATA_UINT64 }, 449 { "prefetch_metadata_misses", KSTAT_DATA_UINT64 }, 450 { "mru_hits", KSTAT_DATA_UINT64 }, 451 { "mru_ghost_hits", KSTAT_DATA_UINT64 }, 452 { "mfu_hits", KSTAT_DATA_UINT64 }, 453 { "mfu_ghost_hits", KSTAT_DATA_UINT64 }, 454 { "allocated", KSTAT_DATA_UINT64 }, 455 { "deleted", KSTAT_DATA_UINT64 }, 456 { "stolen", KSTAT_DATA_UINT64 }, 457 { "recycle_miss", KSTAT_DATA_UINT64 }, 458 { "mutex_miss", KSTAT_DATA_UINT64 }, 459 { "evict_skip", KSTAT_DATA_UINT64 }, 460 { "evict_l2_cached", KSTAT_DATA_UINT64 }, 461 { "evict_l2_eligible", KSTAT_DATA_UINT64 }, 462 { "evict_l2_ineligible", KSTAT_DATA_UINT64 }, 463 { "hash_elements", KSTAT_DATA_UINT64 }, 464 { "hash_elements_max", KSTAT_DATA_UINT64 }, 465 { "hash_collisions", KSTAT_DATA_UINT64 }, 466 { "hash_chains", KSTAT_DATA_UINT64 }, 467 { "hash_chain_max", KSTAT_DATA_UINT64 }, 468 { "p", KSTAT_DATA_UINT64 }, 469 { "c", KSTAT_DATA_UINT64 }, 470 { "c_min", KSTAT_DATA_UINT64 }, 471 { "c_max", KSTAT_DATA_UINT64 }, 472 { "size", KSTAT_DATA_UINT64 }, 473 { "hdr_size", KSTAT_DATA_UINT64 }, 474 { "data_size", KSTAT_DATA_UINT64 }, 475 { "other_size", KSTAT_DATA_UINT64 }, 476 { "l2_hits", KSTAT_DATA_UINT64 }, 477 { "l2_misses", KSTAT_DATA_UINT64 }, 478 { "l2_feeds", KSTAT_DATA_UINT64 }, 479 { "l2_rw_clash", KSTAT_DATA_UINT64 }, 480 { "l2_read_bytes", KSTAT_DATA_UINT64 }, 481 { "l2_write_bytes", KSTAT_DATA_UINT64 }, 482 { "l2_writes_sent", KSTAT_DATA_UINT64 }, 483 { "l2_writes_done", KSTAT_DATA_UINT64 }, 484 { "l2_writes_error", KSTAT_DATA_UINT64 }, 485 { "l2_writes_hdr_miss", KSTAT_DATA_UINT64 }, 486 { "l2_evict_lock_retry", KSTAT_DATA_UINT64 }, 487 { "l2_evict_reading", KSTAT_DATA_UINT64 }, 488 { "l2_evict_l1cached", KSTAT_DATA_UINT64 }, 489 { "l2_free_on_write", KSTAT_DATA_UINT64 }, 490 { "l2_cdata_free_on_write", KSTAT_DATA_UINT64 }, 491 { "l2_abort_lowmem", KSTAT_DATA_UINT64 }, 492 { "l2_cksum_bad", KSTAT_DATA_UINT64 }, 493 { "l2_io_error", KSTAT_DATA_UINT64 }, 494 { "l2_size", KSTAT_DATA_UINT64 }, 495 { "l2_asize", KSTAT_DATA_UINT64 }, 496 { "l2_hdr_size", KSTAT_DATA_UINT64 }, 497 { "l2_compress_successes", KSTAT_DATA_UINT64 }, 498 { "l2_compress_zeros", KSTAT_DATA_UINT64 }, 499 { "l2_compress_failures", KSTAT_DATA_UINT64 }, 500 { "l2_write_trylock_fail", KSTAT_DATA_UINT64 }, 501 { "l2_write_passed_headroom", KSTAT_DATA_UINT64 }, 502 { "l2_write_spa_mismatch", KSTAT_DATA_UINT64 }, 503 { "l2_write_in_l2", KSTAT_DATA_UINT64 }, 504 { "l2_write_io_in_progress", KSTAT_DATA_UINT64 }, 505 { "l2_write_not_cacheable", KSTAT_DATA_UINT64 }, 506 { "l2_write_full", KSTAT_DATA_UINT64 }, 507 { "l2_write_buffer_iter", KSTAT_DATA_UINT64 }, 508 { "l2_write_pios", KSTAT_DATA_UINT64 }, 509 { "l2_write_buffer_bytes_scanned", KSTAT_DATA_UINT64 }, 510 { "l2_write_buffer_list_iter", KSTAT_DATA_UINT64 }, 511 { "l2_write_buffer_list_null_iter", KSTAT_DATA_UINT64 }, 512 { "memory_throttle_count", KSTAT_DATA_UINT64 }, 513 { "duplicate_buffers", KSTAT_DATA_UINT64 }, 514 { "duplicate_buffers_size", KSTAT_DATA_UINT64 }, 515 { "duplicate_reads", KSTAT_DATA_UINT64 }, 516 { "arc_meta_used", KSTAT_DATA_UINT64 }, 517 { "arc_meta_limit", KSTAT_DATA_UINT64 }, 518 { "arc_meta_max", KSTAT_DATA_UINT64 }, 519 { "arc_meta_min", KSTAT_DATA_UINT64 } 520}; 521 522#define ARCSTAT(stat) (arc_stats.stat.value.ui64) 523 524#define ARCSTAT_INCR(stat, val) \ 525 atomic_add_64(&arc_stats.stat.value.ui64, (val)) 526 527#define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1) 528#define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1) 529 530#define ARCSTAT_MAX(stat, val) { \ 531 uint64_t m; \ 532 while ((val) > (m = arc_stats.stat.value.ui64) && \ 533 (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \ 534 continue; \ 535} 536 537#define ARCSTAT_MAXSTAT(stat) \ 538 ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64) 539 540/* 541 * We define a macro to allow ARC hits/misses to be easily broken down by 542 * two separate conditions, giving a total of four different subtypes for 543 * each of hits and misses (so eight statistics total). 544 */ 545#define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \ 546 if (cond1) { \ 547 if (cond2) { \ 548 ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \ 549 } else { \ 550 ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \ 551 } \ 552 } else { \ 553 if (cond2) { \ 554 ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \ 555 } else { \ 556 ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\ 557 } \ 558 } 559 560kstat_t *arc_ksp; 561static arc_state_t *arc_anon; 562static arc_state_t *arc_mru; 563static arc_state_t *arc_mru_ghost; 564static arc_state_t *arc_mfu; 565static arc_state_t *arc_mfu_ghost; 566static arc_state_t *arc_l2c_only; 567 568/* 569 * There are several ARC variables that are critical to export as kstats -- 570 * but we don't want to have to grovel around in the kstat whenever we wish to 571 * manipulate them. For these variables, we therefore define them to be in 572 * terms of the statistic variable. This assures that we are not introducing 573 * the possibility of inconsistency by having shadow copies of the variables, 574 * while still allowing the code to be readable. 575 */ 576#define arc_size ARCSTAT(arcstat_size) /* actual total arc size */ 577#define arc_p ARCSTAT(arcstat_p) /* target size of MRU */ 578#define arc_c ARCSTAT(arcstat_c) /* target size of cache */ 579#define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */ 580#define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */ 581#define arc_meta_limit ARCSTAT(arcstat_meta_limit) /* max size for metadata */ 582#define arc_meta_min ARCSTAT(arcstat_meta_min) /* min size for metadata */ 583#define arc_meta_used ARCSTAT(arcstat_meta_used) /* size of metadata */ 584#define arc_meta_max ARCSTAT(arcstat_meta_max) /* max size of metadata */ 585 586#define L2ARC_IS_VALID_COMPRESS(_c_) \ 587 ((_c_) == ZIO_COMPRESS_LZ4 || (_c_) == ZIO_COMPRESS_EMPTY) 588 589static int arc_no_grow; /* Don't try to grow cache size */ 590static uint64_t arc_tempreserve; 591static uint64_t arc_loaned_bytes; 592 593typedef struct arc_callback arc_callback_t; 594 595struct arc_callback { 596 void *acb_private; 597 arc_done_func_t *acb_done; 598 arc_buf_t *acb_buf; 599 zio_t *acb_zio_dummy; 600 arc_callback_t *acb_next; 601}; 602 603typedef struct arc_write_callback arc_write_callback_t; 604 605struct arc_write_callback { 606 void *awcb_private; 607 arc_done_func_t *awcb_ready; 608 arc_done_func_t *awcb_physdone; 609 arc_done_func_t *awcb_done; 610 arc_buf_t *awcb_buf; 611}; 612 613/* 614 * ARC buffers are separated into multiple structs as a memory saving measure: 615 * - Common fields struct, always defined, and embedded within it: 616 * - L2-only fields, always allocated but undefined when not in L2ARC 617 * - L1-only fields, only allocated when in L1ARC 618 * 619 * Buffer in L1 Buffer only in L2 620 * +------------------------+ +------------------------+ 621 * | arc_buf_hdr_t | | arc_buf_hdr_t | 622 * | | | | 623 * | | | | 624 * | | | | 625 * +------------------------+ +------------------------+ 626 * | l2arc_buf_hdr_t | | l2arc_buf_hdr_t | 627 * | (undefined if L1-only) | | | 628 * +------------------------+ +------------------------+ 629 * | l1arc_buf_hdr_t | 630 * | | 631 * | | 632 * | | 633 * | | 634 * +------------------------+ 635 * 636 * Because it's possible for the L2ARC to become extremely large, we can wind 637 * up eating a lot of memory in L2ARC buffer headers, so the size of a header 638 * is minimized by only allocating the fields necessary for an L1-cached buffer 639 * when a header is actually in the L1 cache. The sub-headers (l1arc_buf_hdr and 640 * l2arc_buf_hdr) are embedded rather than allocated separately to save a couple 641 * words in pointers. arc_hdr_realloc() is used to switch a header between 642 * these two allocation states. 643 */ 644typedef struct l1arc_buf_hdr { 645 kmutex_t b_freeze_lock; 646#ifdef ZFS_DEBUG 647 /* 648 * used for debugging wtih kmem_flags - by allocating and freeing 649 * b_thawed when the buffer is thawed, we get a record of the stack 650 * trace that thawed it. 651 */ 652 void *b_thawed; 653#endif 654 655 arc_buf_t *b_buf; 656 uint32_t b_datacnt; 657 /* for waiting on writes to complete */ 658 kcondvar_t b_cv; 659 660 /* protected by arc state mutex */ 661 arc_state_t *b_state; 662 list_node_t b_arc_node; 663 664 /* updated atomically */ 665 clock_t b_arc_access; 666 667 /* self protecting */ 668 refcount_t b_refcnt; 669 670 arc_callback_t *b_acb; 671 /* temporary buffer holder for in-flight compressed data */ 672 void *b_tmp_cdata; 673} l1arc_buf_hdr_t; 674 675typedef struct l2arc_dev l2arc_dev_t; 676 677typedef struct l2arc_buf_hdr { 678 /* protected by arc_buf_hdr mutex */ 679 l2arc_dev_t *b_dev; /* L2ARC device */ 680 uint64_t b_daddr; /* disk address, offset byte */ 681 /* real alloc'd buffer size depending on b_compress applied */ 682 int32_t b_asize; 683 684 list_node_t b_l2node; 685} l2arc_buf_hdr_t; 686 687struct arc_buf_hdr { 688 /* protected by hash lock */ 689 dva_t b_dva; 690 uint64_t b_birth; 691 /* 692 * Even though this checksum is only set/verified when a buffer is in 693 * the L1 cache, it needs to be in the set of common fields because it 694 * must be preserved from the time before a buffer is written out to 695 * L2ARC until after it is read back in. 696 */ 697 zio_cksum_t *b_freeze_cksum; 698 699 arc_buf_hdr_t *b_hash_next; 700 arc_flags_t b_flags; 701 702 /* immutable */ 703 int32_t b_size; 704 uint64_t b_spa; 705 706 /* L2ARC fields. Undefined when not in L2ARC. */ 707 l2arc_buf_hdr_t b_l2hdr; 708 /* L1ARC fields. Undefined when in l2arc_only state */ 709 l1arc_buf_hdr_t b_l1hdr; 710}; 711 712#ifdef _KERNEL 713static int 714sysctl_vfs_zfs_arc_meta_limit(SYSCTL_HANDLER_ARGS) 715{ 716 uint64_t val; 717 int err; 718 719 val = arc_meta_limit; 720 err = sysctl_handle_64(oidp, &val, 0, req); 721 if (err != 0 || req->newptr == NULL) 722 return (err); 723 724 if (val <= 0 || val > arc_c_max) 725 return (EINVAL); 726 727 arc_meta_limit = val; 728 return (0); 729} 730#endif 731 732static arc_buf_t *arc_eviction_list; 733static kmutex_t arc_eviction_mtx; 734static arc_buf_hdr_t arc_eviction_hdr; 735 736#define GHOST_STATE(state) \ 737 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \ 738 (state) == arc_l2c_only) 739 740#define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_FLAG_IN_HASH_TABLE) 741#define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) 742#define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_FLAG_IO_ERROR) 743#define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_FLAG_PREFETCH) 744#define HDR_FREED_IN_READ(hdr) ((hdr)->b_flags & ARC_FLAG_FREED_IN_READ) 745#define HDR_BUF_AVAILABLE(hdr) ((hdr)->b_flags & ARC_FLAG_BUF_AVAILABLE) 746 747#define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_FLAG_L2CACHE) 748#define HDR_L2COMPRESS(hdr) ((hdr)->b_flags & ARC_FLAG_L2COMPRESS) 749#define HDR_L2_READING(hdr) \ 750 (((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) && \ 751 ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR)) 752#define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITING) 753#define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_FLAG_L2_EVICTED) 754#define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITE_HEAD) 755 756#define HDR_ISTYPE_METADATA(hdr) \ 757 ((hdr)->b_flags & ARC_FLAG_BUFC_METADATA) 758#define HDR_ISTYPE_DATA(hdr) (!HDR_ISTYPE_METADATA(hdr)) 759 760#define HDR_HAS_L1HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L1HDR) 761#define HDR_HAS_L2HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR) 762 763/* For storing compression mode in b_flags */ 764#define HDR_COMPRESS_OFFSET 24 765#define HDR_COMPRESS_NBITS 7 766 767#define HDR_GET_COMPRESS(hdr) ((enum zio_compress)BF32_GET(hdr->b_flags, \ 768 HDR_COMPRESS_OFFSET, HDR_COMPRESS_NBITS)) 769#define HDR_SET_COMPRESS(hdr, cmp) BF32_SET(hdr->b_flags, \ 770 HDR_COMPRESS_OFFSET, HDR_COMPRESS_NBITS, (cmp)) 771 772/* 773 * Other sizes 774 */ 775 776#define HDR_FULL_SIZE ((int64_t)sizeof (arc_buf_hdr_t)) 777#define HDR_L2ONLY_SIZE ((int64_t)offsetof(arc_buf_hdr_t, b_l1hdr)) 778 779/* 780 * Hash table routines 781 */ 782 783#define HT_LOCK_PAD CACHE_LINE_SIZE 784 785struct ht_lock { 786 kmutex_t ht_lock; 787#ifdef _KERNEL 788 unsigned char pad[(HT_LOCK_PAD - sizeof (kmutex_t))]; 789#endif 790}; 791 792#define BUF_LOCKS 256 793typedef struct buf_hash_table { 794 uint64_t ht_mask; 795 arc_buf_hdr_t **ht_table; 796 struct ht_lock ht_locks[BUF_LOCKS] __aligned(CACHE_LINE_SIZE); 797} buf_hash_table_t; 798 799static buf_hash_table_t buf_hash_table; 800 801#define BUF_HASH_INDEX(spa, dva, birth) \ 802 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask) 803#define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)]) 804#define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock)) 805#define HDR_LOCK(hdr) \ 806 (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth))) 807 808uint64_t zfs_crc64_table[256]; 809 810/* 811 * Level 2 ARC 812 */ 813 814#define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */ 815#define L2ARC_HEADROOM 2 /* num of writes */ 816/* 817 * If we discover during ARC scan any buffers to be compressed, we boost 818 * our headroom for the next scanning cycle by this percentage multiple. 819 */ 820#define L2ARC_HEADROOM_BOOST 200 821#define L2ARC_FEED_SECS 1 /* caching interval secs */ 822#define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */ 823 824#define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent) 825#define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done) 826 827/* L2ARC Performance Tunables */ 828uint64_t l2arc_write_max = L2ARC_WRITE_SIZE; /* default max write size */ 829uint64_t l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra write during warmup */ 830uint64_t l2arc_headroom = L2ARC_HEADROOM; /* number of dev writes */ 831uint64_t l2arc_headroom_boost = L2ARC_HEADROOM_BOOST; 832uint64_t l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */ 833uint64_t l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval milliseconds */ 834boolean_t l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */ 835boolean_t l2arc_feed_again = B_TRUE; /* turbo warmup */ 836boolean_t l2arc_norw = B_TRUE; /* no reads during writes */ 837 838SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_max, CTLFLAG_RW, 839 &l2arc_write_max, 0, "max write size"); 840SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_boost, CTLFLAG_RW, 841 &l2arc_write_boost, 0, "extra write during warmup"); 842SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_headroom, CTLFLAG_RW, 843 &l2arc_headroom, 0, "number of dev writes"); 844SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_secs, CTLFLAG_RW, 845 &l2arc_feed_secs, 0, "interval seconds"); 846SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_min_ms, CTLFLAG_RW, 847 &l2arc_feed_min_ms, 0, "min interval milliseconds"); 848 849SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_noprefetch, CTLFLAG_RW, 850 &l2arc_noprefetch, 0, "don't cache prefetch bufs"); 851SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_feed_again, CTLFLAG_RW, 852 &l2arc_feed_again, 0, "turbo warmup"); 853SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_norw, CTLFLAG_RW, 854 &l2arc_norw, 0, "no reads during writes"); 855 856SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_size, CTLFLAG_RD, 857 &ARC_anon.arcs_size, 0, "size of anonymous state"); 858SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_metadata_lsize, CTLFLAG_RD, 859 &ARC_anon.arcs_lsize[ARC_BUFC_METADATA], 0, "size of anonymous state"); 860SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_data_lsize, CTLFLAG_RD, 861 &ARC_anon.arcs_lsize[ARC_BUFC_DATA], 0, "size of anonymous state"); 862 863SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_size, CTLFLAG_RD, 864 &ARC_mru.arcs_size, 0, "size of mru state"); 865SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_metadata_lsize, CTLFLAG_RD, 866 &ARC_mru.arcs_lsize[ARC_BUFC_METADATA], 0, "size of metadata in mru state"); 867SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_data_lsize, CTLFLAG_RD, 868 &ARC_mru.arcs_lsize[ARC_BUFC_DATA], 0, "size of data in mru state"); 869 870SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_size, CTLFLAG_RD, 871 &ARC_mru_ghost.arcs_size, 0, "size of mru ghost state"); 872SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_metadata_lsize, CTLFLAG_RD, 873 &ARC_mru_ghost.arcs_lsize[ARC_BUFC_METADATA], 0, 874 "size of metadata in mru ghost state"); 875SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_data_lsize, CTLFLAG_RD, 876 &ARC_mru_ghost.arcs_lsize[ARC_BUFC_DATA], 0, 877 "size of data in mru ghost state"); 878 879SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_size, CTLFLAG_RD, 880 &ARC_mfu.arcs_size, 0, "size of mfu state"); 881SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_metadata_lsize, CTLFLAG_RD, 882 &ARC_mfu.arcs_lsize[ARC_BUFC_METADATA], 0, "size of metadata in mfu state"); 883SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_data_lsize, CTLFLAG_RD, 884 &ARC_mfu.arcs_lsize[ARC_BUFC_DATA], 0, "size of data in mfu state"); 885 886SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_size, CTLFLAG_RD, 887 &ARC_mfu_ghost.arcs_size, 0, "size of mfu ghost state"); 888SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_metadata_lsize, CTLFLAG_RD, 889 &ARC_mfu_ghost.arcs_lsize[ARC_BUFC_METADATA], 0, 890 "size of metadata in mfu ghost state"); 891SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_data_lsize, CTLFLAG_RD, 892 &ARC_mfu_ghost.arcs_lsize[ARC_BUFC_DATA], 0, 893 "size of data in mfu ghost state"); 894 895SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2c_only_size, CTLFLAG_RD, 896 &ARC_l2c_only.arcs_size, 0, "size of mru state"); 897 898/* 899 * L2ARC Internals 900 */ 901struct l2arc_dev { 902 vdev_t *l2ad_vdev; /* vdev */ 903 spa_t *l2ad_spa; /* spa */ 904 uint64_t l2ad_hand; /* next write location */ 905 uint64_t l2ad_start; /* first addr on device */ 906 uint64_t l2ad_end; /* last addr on device */ 907 uint64_t l2ad_evict; /* last addr eviction reached */ 908 boolean_t l2ad_first; /* first sweep through */ 909 boolean_t l2ad_writing; /* currently writing */ 910 kmutex_t l2ad_mtx; /* lock for buffer list */ 911 list_t l2ad_buflist; /* buffer list */ 912 list_node_t l2ad_node; /* device list node */ 913}; 914 915static list_t L2ARC_dev_list; /* device list */ 916static list_t *l2arc_dev_list; /* device list pointer */ 917static kmutex_t l2arc_dev_mtx; /* device list mutex */ 918static l2arc_dev_t *l2arc_dev_last; /* last device used */ 919static list_t L2ARC_free_on_write; /* free after write buf list */ 920static list_t *l2arc_free_on_write; /* free after write list ptr */ 921static kmutex_t l2arc_free_on_write_mtx; /* mutex for list */ 922static uint64_t l2arc_ndev; /* number of devices */ 923 924typedef struct l2arc_read_callback { 925 arc_buf_t *l2rcb_buf; /* read buffer */ 926 spa_t *l2rcb_spa; /* spa */ 927 blkptr_t l2rcb_bp; /* original blkptr */ 928 zbookmark_phys_t l2rcb_zb; /* original bookmark */ 929 int l2rcb_flags; /* original flags */ 930 enum zio_compress l2rcb_compress; /* applied compress */ 931} l2arc_read_callback_t; 932 933typedef struct l2arc_write_callback { 934 l2arc_dev_t *l2wcb_dev; /* device info */ 935 arc_buf_hdr_t *l2wcb_head; /* head of write buflist */ 936} l2arc_write_callback_t; 937 938typedef struct l2arc_data_free { 939 /* protected by l2arc_free_on_write_mtx */ 940 void *l2df_data; 941 size_t l2df_size; 942 void (*l2df_func)(void *, size_t); 943 list_node_t l2df_list_node; 944} l2arc_data_free_t; 945 946static kmutex_t l2arc_feed_thr_lock; 947static kcondvar_t l2arc_feed_thr_cv; 948static uint8_t l2arc_thread_exit; 949 950static void arc_get_data_buf(arc_buf_t *); 951static void arc_access(arc_buf_hdr_t *, kmutex_t *); 952static int arc_evict_needed(arc_buf_contents_t); 953static void arc_evict_ghost(arc_state_t *, uint64_t, int64_t); 954static void arc_buf_watch(arc_buf_t *); 955 956static arc_buf_contents_t arc_buf_type(arc_buf_hdr_t *); 957static uint32_t arc_bufc_to_flags(arc_buf_contents_t); 958 959static boolean_t l2arc_write_eligible(uint64_t, arc_buf_hdr_t *); 960static void l2arc_read_done(zio_t *); 961 962static boolean_t l2arc_compress_buf(arc_buf_hdr_t *); 963static void l2arc_decompress_zio(zio_t *, arc_buf_hdr_t *, enum zio_compress); 964static void l2arc_release_cdata_buf(arc_buf_hdr_t *); 965 966static uint64_t 967buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth) 968{ 969 uint8_t *vdva = (uint8_t *)dva; 970 uint64_t crc = -1ULL; 971 int i; 972 973 ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY); 974 975 for (i = 0; i < sizeof (dva_t); i++) 976 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ vdva[i]) & 0xFF]; 977 978 crc ^= (spa>>8) ^ birth; 979 980 return (crc); 981} 982 983#define BUF_EMPTY(buf) \ 984 ((buf)->b_dva.dva_word[0] == 0 && \ 985 (buf)->b_dva.dva_word[1] == 0) 986 987#define BUF_EQUAL(spa, dva, birth, buf) \ 988 ((buf)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \ 989 ((buf)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \ 990 ((buf)->b_birth == birth) && ((buf)->b_spa == spa) 991 992static void 993buf_discard_identity(arc_buf_hdr_t *hdr) 994{ 995 hdr->b_dva.dva_word[0] = 0; 996 hdr->b_dva.dva_word[1] = 0; 997 hdr->b_birth = 0; 998} 999 1000static arc_buf_hdr_t * 1001buf_hash_find(uint64_t spa, const blkptr_t *bp, kmutex_t **lockp) 1002{ 1003 const dva_t *dva = BP_IDENTITY(bp); 1004 uint64_t birth = BP_PHYSICAL_BIRTH(bp); 1005 uint64_t idx = BUF_HASH_INDEX(spa, dva, birth); 1006 kmutex_t *hash_lock = BUF_HASH_LOCK(idx); 1007 arc_buf_hdr_t *hdr; 1008 1009 mutex_enter(hash_lock); 1010 for (hdr = buf_hash_table.ht_table[idx]; hdr != NULL; 1011 hdr = hdr->b_hash_next) { 1012 if (BUF_EQUAL(spa, dva, birth, hdr)) { 1013 *lockp = hash_lock; 1014 return (hdr); 1015 } 1016 } 1017 mutex_exit(hash_lock); 1018 *lockp = NULL; 1019 return (NULL); 1020} 1021 1022/* 1023 * Insert an entry into the hash table. If there is already an element 1024 * equal to elem in the hash table, then the already existing element 1025 * will be returned and the new element will not be inserted. 1026 * Otherwise returns NULL. 1027 * If lockp == NULL, the caller is assumed to already hold the hash lock. 1028 */ 1029static arc_buf_hdr_t * 1030buf_hash_insert(arc_buf_hdr_t *hdr, kmutex_t **lockp) 1031{ 1032 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth); 1033 kmutex_t *hash_lock = BUF_HASH_LOCK(idx); 1034 arc_buf_hdr_t *fhdr; 1035 uint32_t i; 1036 1037 ASSERT(!DVA_IS_EMPTY(&hdr->b_dva)); 1038 ASSERT(hdr->b_birth != 0); 1039 ASSERT(!HDR_IN_HASH_TABLE(hdr)); 1040 1041 if (lockp != NULL) { 1042 *lockp = hash_lock; 1043 mutex_enter(hash_lock); 1044 } else { 1045 ASSERT(MUTEX_HELD(hash_lock)); 1046 } 1047 1048 for (fhdr = buf_hash_table.ht_table[idx], i = 0; fhdr != NULL; 1049 fhdr = fhdr->b_hash_next, i++) { 1050 if (BUF_EQUAL(hdr->b_spa, &hdr->b_dva, hdr->b_birth, fhdr)) 1051 return (fhdr); 1052 } 1053 1054 hdr->b_hash_next = buf_hash_table.ht_table[idx]; 1055 buf_hash_table.ht_table[idx] = hdr; 1056 hdr->b_flags |= ARC_FLAG_IN_HASH_TABLE; 1057 1058 /* collect some hash table performance data */ 1059 if (i > 0) { 1060 ARCSTAT_BUMP(arcstat_hash_collisions); 1061 if (i == 1) 1062 ARCSTAT_BUMP(arcstat_hash_chains); 1063 1064 ARCSTAT_MAX(arcstat_hash_chain_max, i); 1065 } 1066 1067 ARCSTAT_BUMP(arcstat_hash_elements); 1068 ARCSTAT_MAXSTAT(arcstat_hash_elements); 1069 1070 return (NULL); 1071} 1072 1073static void 1074buf_hash_remove(arc_buf_hdr_t *hdr) 1075{ 1076 arc_buf_hdr_t *fhdr, **hdrp; 1077 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth); 1078 1079 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx))); 1080 ASSERT(HDR_IN_HASH_TABLE(hdr)); 1081 1082 hdrp = &buf_hash_table.ht_table[idx]; 1083 while ((fhdr = *hdrp) != hdr) { 1084 ASSERT(fhdr != NULL); 1085 hdrp = &fhdr->b_hash_next; 1086 } 1087 *hdrp = hdr->b_hash_next; 1088 hdr->b_hash_next = NULL; 1089 hdr->b_flags &= ~ARC_FLAG_IN_HASH_TABLE; 1090 1091 /* collect some hash table performance data */ 1092 ARCSTAT_BUMPDOWN(arcstat_hash_elements); 1093 1094 if (buf_hash_table.ht_table[idx] && 1095 buf_hash_table.ht_table[idx]->b_hash_next == NULL) 1096 ARCSTAT_BUMPDOWN(arcstat_hash_chains); 1097} 1098 1099/* 1100 * Global data structures and functions for the buf kmem cache. 1101 */ 1102static kmem_cache_t *hdr_full_cache; 1103static kmem_cache_t *hdr_l2only_cache; 1104static kmem_cache_t *buf_cache; 1105 1106static void 1107buf_fini(void) 1108{ 1109 int i; 1110 1111 kmem_free(buf_hash_table.ht_table, 1112 (buf_hash_table.ht_mask + 1) * sizeof (void *)); 1113 for (i = 0; i < BUF_LOCKS; i++) 1114 mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock); 1115 kmem_cache_destroy(hdr_full_cache); 1116 kmem_cache_destroy(hdr_l2only_cache); 1117 kmem_cache_destroy(buf_cache); 1118} 1119 1120/* 1121 * Constructor callback - called when the cache is empty 1122 * and a new buf is requested. 1123 */ 1124/* ARGSUSED */ 1125static int 1126hdr_full_cons(void *vbuf, void *unused, int kmflag) 1127{ 1128 arc_buf_hdr_t *hdr = vbuf; 1129 1130 bzero(hdr, HDR_FULL_SIZE); 1131 cv_init(&hdr->b_l1hdr.b_cv, NULL, CV_DEFAULT, NULL); 1132 refcount_create(&hdr->b_l1hdr.b_refcnt); 1133 mutex_init(&hdr->b_l1hdr.b_freeze_lock, NULL, MUTEX_DEFAULT, NULL); 1134 arc_space_consume(HDR_FULL_SIZE, ARC_SPACE_HDRS); 1135 1136 return (0); 1137} 1138 1139/* ARGSUSED */ 1140static int 1141hdr_l2only_cons(void *vbuf, void *unused, int kmflag) 1142{ 1143 arc_buf_hdr_t *hdr = vbuf; 1144 1145 bzero(hdr, HDR_L2ONLY_SIZE); 1146 arc_space_consume(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS); 1147 1148 return (0); 1149} 1150 1151/* ARGSUSED */ 1152static int 1153buf_cons(void *vbuf, void *unused, int kmflag) 1154{ 1155 arc_buf_t *buf = vbuf; 1156 1157 bzero(buf, sizeof (arc_buf_t)); 1158 mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL); 1159 arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS); 1160 1161 return (0); 1162} 1163 1164/* 1165 * Destructor callback - called when a cached buf is 1166 * no longer required. 1167 */ 1168/* ARGSUSED */ 1169static void 1170hdr_full_dest(void *vbuf, void *unused) 1171{ 1172 arc_buf_hdr_t *hdr = vbuf; 1173 1174 ASSERT(BUF_EMPTY(hdr)); 1175 cv_destroy(&hdr->b_l1hdr.b_cv); 1176 refcount_destroy(&hdr->b_l1hdr.b_refcnt); 1177 mutex_destroy(&hdr->b_l1hdr.b_freeze_lock); 1178 arc_space_return(HDR_FULL_SIZE, ARC_SPACE_HDRS); 1179} 1180 1181/* ARGSUSED */ 1182static void 1183hdr_l2only_dest(void *vbuf, void *unused) 1184{ 1185 arc_buf_hdr_t *hdr = vbuf; 1186 1187 ASSERT(BUF_EMPTY(hdr)); 1188 arc_space_return(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS); 1189} 1190 1191/* ARGSUSED */ 1192static void 1193buf_dest(void *vbuf, void *unused) 1194{ 1195 arc_buf_t *buf = vbuf; 1196 1197 mutex_destroy(&buf->b_evict_lock); 1198 arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS); 1199} 1200 1201/* 1202 * Reclaim callback -- invoked when memory is low. 1203 */ 1204/* ARGSUSED */ 1205static void 1206hdr_recl(void *unused) 1207{ 1208 dprintf("hdr_recl called\n"); 1209 /* 1210 * umem calls the reclaim func when we destroy the buf cache, 1211 * which is after we do arc_fini(). 1212 */ 1213 if (!arc_dead) 1214 cv_signal(&arc_reclaim_thr_cv); 1215} 1216 1217static void 1218buf_init(void) 1219{ 1220 uint64_t *ct; 1221 uint64_t hsize = 1ULL << 12; 1222 int i, j; 1223 1224 /* 1225 * The hash table is big enough to fill all of physical memory 1226 * with an average block size of zfs_arc_average_blocksize (default 8K). 1227 * By default, the table will take up 1228 * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers). 1229 */ 1230 while (hsize * zfs_arc_average_blocksize < (uint64_t)physmem * PAGESIZE) 1231 hsize <<= 1; 1232retry: 1233 buf_hash_table.ht_mask = hsize - 1; 1234 buf_hash_table.ht_table = 1235 kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP); 1236 if (buf_hash_table.ht_table == NULL) { 1237 ASSERT(hsize > (1ULL << 8)); 1238 hsize >>= 1; 1239 goto retry; 1240 } 1241 1242 hdr_full_cache = kmem_cache_create("arc_buf_hdr_t_full", HDR_FULL_SIZE, 1243 0, hdr_full_cons, hdr_full_dest, hdr_recl, NULL, NULL, 0); 1244 hdr_l2only_cache = kmem_cache_create("arc_buf_hdr_t_l2only", 1245 HDR_L2ONLY_SIZE, 0, hdr_l2only_cons, hdr_l2only_dest, hdr_recl, 1246 NULL, NULL, 0); 1247 buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t), 1248 0, buf_cons, buf_dest, NULL, NULL, NULL, 0); 1249 1250 for (i = 0; i < 256; i++) 1251 for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--) 1252 *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY); 1253 1254 for (i = 0; i < BUF_LOCKS; i++) { 1255 mutex_init(&buf_hash_table.ht_locks[i].ht_lock, 1256 NULL, MUTEX_DEFAULT, NULL); 1257 } 1258} 1259 1260/* 1261 * Transition between the two allocation states for the arc_buf_hdr struct. 1262 * The arc_buf_hdr struct can be allocated with (hdr_full_cache) or without 1263 * (hdr_l2only_cache) the fields necessary for the L1 cache - the smaller 1264 * version is used when a cache buffer is only in the L2ARC in order to reduce 1265 * memory usage. 1266 */ 1267static arc_buf_hdr_t * 1268arc_hdr_realloc(arc_buf_hdr_t *hdr, kmem_cache_t *old, kmem_cache_t *new) 1269{ 1270 ASSERT(HDR_HAS_L2HDR(hdr)); 1271 1272 arc_buf_hdr_t *nhdr; 1273 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev; 1274 1275 ASSERT((old == hdr_full_cache && new == hdr_l2only_cache) || 1276 (old == hdr_l2only_cache && new == hdr_full_cache)); 1277 1278 nhdr = kmem_cache_alloc(new, KM_PUSHPAGE); 1279 1280 ASSERT(MUTEX_HELD(HDR_LOCK(hdr))); 1281 buf_hash_remove(hdr); 1282 1283 bcopy(hdr, nhdr, HDR_L2ONLY_SIZE); 1284 if (new == hdr_full_cache) { 1285 nhdr->b_flags |= ARC_FLAG_HAS_L1HDR; 1286 /* 1287 * arc_access and arc_change_state need to be aware that a 1288 * header has just come out of L2ARC, so we set its state to 1289 * l2c_only even though it's about to change. 1290 */ 1291 nhdr->b_l1hdr.b_state = arc_l2c_only; 1292 } else { 1293 ASSERT(hdr->b_l1hdr.b_buf == NULL); 1294 ASSERT0(hdr->b_l1hdr.b_datacnt); 1295 ASSERT(!list_link_active(&hdr->b_l1hdr.b_arc_node)); 1296 /* 1297 * We might be removing the L1hdr of a buffer which was just 1298 * written out to L2ARC. If such a buffer is compressed then we 1299 * need to free its b_tmp_cdata before destroying the header. 1300 */ 1301 if (hdr->b_l1hdr.b_tmp_cdata != NULL && 1302 HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF) 1303 l2arc_release_cdata_buf(hdr); 1304 nhdr->b_flags &= ~ARC_FLAG_HAS_L1HDR; 1305 } 1306 /* 1307 * The header has been reallocated so we need to re-insert it into any 1308 * lists it was on. 1309 */ 1310 (void) buf_hash_insert(nhdr, NULL); 1311 1312 ASSERT(list_link_active(&hdr->b_l2hdr.b_l2node)); 1313 1314 mutex_enter(&dev->l2ad_mtx); 1315 1316 /* 1317 * We must place the realloc'ed header back into the list at 1318 * the same spot. Otherwise, if it's placed earlier in the list, 1319 * l2arc_write_buffers() could find it during the function's 1320 * write phase, and try to write it out to the l2arc. 1321 */ 1322 list_insert_after(&dev->l2ad_buflist, hdr, nhdr); 1323 list_remove(&dev->l2ad_buflist, hdr); 1324 1325 mutex_exit(&dev->l2ad_mtx); 1326 1327 buf_discard_identity(hdr); 1328 hdr->b_freeze_cksum = NULL; 1329 kmem_cache_free(old, hdr); 1330 1331 return (nhdr); 1332} 1333 1334 1335#define ARC_MINTIME (hz>>4) /* 62 ms */ 1336 1337static void 1338arc_cksum_verify(arc_buf_t *buf) 1339{ 1340 zio_cksum_t zc; 1341 1342 if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 1343 return; 1344 1345 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1346 if (buf->b_hdr->b_freeze_cksum == NULL || HDR_IO_ERROR(buf->b_hdr)) { 1347 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1348 return; 1349 } 1350 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc); 1351 if (!ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc)) 1352 panic("buffer modified while frozen!"); 1353 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1354} 1355 1356static int 1357arc_cksum_equal(arc_buf_t *buf) 1358{ 1359 zio_cksum_t zc; 1360 int equal; 1361 1362 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1363 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc); 1364 equal = ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc); 1365 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1366 1367 return (equal); 1368} 1369 1370static void 1371arc_cksum_compute(arc_buf_t *buf, boolean_t force) 1372{ 1373 if (!force && !(zfs_flags & ZFS_DEBUG_MODIFY)) 1374 return; 1375 1376 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1377 if (buf->b_hdr->b_freeze_cksum != NULL) { 1378 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1379 return; 1380 } 1381 buf->b_hdr->b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t), KM_SLEEP); 1382 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, 1383 buf->b_hdr->b_freeze_cksum); 1384 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1385#ifdef illumos 1386 arc_buf_watch(buf); 1387#endif /* illumos */ 1388} 1389 1390#ifdef illumos 1391#ifndef _KERNEL 1392typedef struct procctl { 1393 long cmd; 1394 prwatch_t prwatch; 1395} procctl_t; 1396#endif 1397 1398/* ARGSUSED */ 1399static void 1400arc_buf_unwatch(arc_buf_t *buf) 1401{ 1402#ifndef _KERNEL 1403 if (arc_watch) { 1404 int result; 1405 procctl_t ctl; 1406 ctl.cmd = PCWATCH; 1407 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data; 1408 ctl.prwatch.pr_size = 0; 1409 ctl.prwatch.pr_wflags = 0; 1410 result = write(arc_procfd, &ctl, sizeof (ctl)); 1411 ASSERT3U(result, ==, sizeof (ctl)); 1412 } 1413#endif 1414} 1415 1416/* ARGSUSED */ 1417static void 1418arc_buf_watch(arc_buf_t *buf) 1419{ 1420#ifndef _KERNEL 1421 if (arc_watch) { 1422 int result; 1423 procctl_t ctl; 1424 ctl.cmd = PCWATCH; 1425 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data; 1426 ctl.prwatch.pr_size = buf->b_hdr->b_size; 1427 ctl.prwatch.pr_wflags = WA_WRITE; 1428 result = write(arc_procfd, &ctl, sizeof (ctl)); 1429 ASSERT3U(result, ==, sizeof (ctl)); 1430 } 1431#endif 1432} 1433#endif /* illumos */ 1434 1435static arc_buf_contents_t 1436arc_buf_type(arc_buf_hdr_t *hdr) 1437{ 1438 if (HDR_ISTYPE_METADATA(hdr)) { 1439 return (ARC_BUFC_METADATA); 1440 } else { 1441 return (ARC_BUFC_DATA); 1442 } 1443} 1444 1445static uint32_t 1446arc_bufc_to_flags(arc_buf_contents_t type) 1447{ 1448 switch (type) { 1449 case ARC_BUFC_DATA: 1450 /* metadata field is 0 if buffer contains normal data */ 1451 return (0); 1452 case ARC_BUFC_METADATA: 1453 return (ARC_FLAG_BUFC_METADATA); 1454 default: 1455 break; 1456 } 1457 panic("undefined ARC buffer type!"); 1458 return ((uint32_t)-1); 1459} 1460 1461void 1462arc_buf_thaw(arc_buf_t *buf) 1463{ 1464 if (zfs_flags & ZFS_DEBUG_MODIFY) { 1465 if (buf->b_hdr->b_l1hdr.b_state != arc_anon) 1466 panic("modifying non-anon buffer!"); 1467 if (HDR_IO_IN_PROGRESS(buf->b_hdr)) 1468 panic("modifying buffer while i/o in progress!"); 1469 arc_cksum_verify(buf); 1470 } 1471 1472 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1473 if (buf->b_hdr->b_freeze_cksum != NULL) { 1474 kmem_free(buf->b_hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 1475 buf->b_hdr->b_freeze_cksum = NULL; 1476 } 1477 1478#ifdef ZFS_DEBUG 1479 if (zfs_flags & ZFS_DEBUG_MODIFY) { 1480 if (buf->b_hdr->b_l1hdr.b_thawed != NULL) 1481 kmem_free(buf->b_hdr->b_l1hdr.b_thawed, 1); 1482 buf->b_hdr->b_l1hdr.b_thawed = kmem_alloc(1, KM_SLEEP); 1483 } 1484#endif 1485 1486 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1487 1488#ifdef illumos 1489 arc_buf_unwatch(buf); 1490#endif /* illumos */ 1491} 1492 1493void 1494arc_buf_freeze(arc_buf_t *buf) 1495{ 1496 kmutex_t *hash_lock; 1497 1498 if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 1499 return; 1500 1501 hash_lock = HDR_LOCK(buf->b_hdr); 1502 mutex_enter(hash_lock); 1503 1504 ASSERT(buf->b_hdr->b_freeze_cksum != NULL || 1505 buf->b_hdr->b_l1hdr.b_state == arc_anon); 1506 arc_cksum_compute(buf, B_FALSE); 1507 mutex_exit(hash_lock); 1508 1509} 1510 1511static void 1512get_buf_info(arc_buf_hdr_t *hdr, arc_state_t *state, list_t **list, kmutex_t **lock) 1513{ 1514 uint64_t buf_hashid = buf_hash(hdr->b_spa, &hdr->b_dva, hdr->b_birth); 1515 1516 if (arc_buf_type(hdr) == ARC_BUFC_METADATA) 1517 buf_hashid &= (ARC_BUFC_NUMMETADATALISTS - 1); 1518 else { 1519 buf_hashid &= (ARC_BUFC_NUMDATALISTS - 1); 1520 buf_hashid += ARC_BUFC_NUMMETADATALISTS; 1521 } 1522 1523 *list = &state->arcs_lists[buf_hashid]; 1524 *lock = ARCS_LOCK(state, buf_hashid); 1525} 1526 1527 1528static void 1529add_reference(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, void *tag) 1530{ 1531 ASSERT(HDR_HAS_L1HDR(hdr)); 1532 ASSERT(MUTEX_HELD(hash_lock)); 1533 arc_state_t *state = hdr->b_l1hdr.b_state; 1534 1535 if ((refcount_add(&hdr->b_l1hdr.b_refcnt, tag) == 1) && 1536 (state != arc_anon)) { 1537 /* We don't use the L2-only state list. */ 1538 if (state != arc_l2c_only) { 1539 uint64_t delta = hdr->b_size * hdr->b_l1hdr.b_datacnt; 1540 uint64_t *size = &state->arcs_lsize[arc_buf_type(hdr)]; 1541 list_t *list; 1542 kmutex_t *lock; 1543 1544 get_buf_info(hdr, state, &list, &lock); 1545 ASSERT(!MUTEX_HELD(lock)); 1546 mutex_enter(lock); 1547 ASSERT(list_link_active(&hdr->b_l1hdr.b_arc_node)); 1548 list_remove(list, hdr); 1549 if (GHOST_STATE(state)) { 1550 ASSERT0(hdr->b_l1hdr.b_datacnt); 1551 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 1552 delta = hdr->b_size; 1553 } 1554 ASSERT(delta > 0); 1555 ASSERT3U(*size, >=, delta); 1556 atomic_add_64(size, -delta); 1557 mutex_exit(lock); 1558 } 1559 /* remove the prefetch flag if we get a reference */ 1560 hdr->b_flags &= ~ARC_FLAG_PREFETCH; 1561 } 1562} 1563 1564static int 1565remove_reference(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, void *tag) 1566{ 1567 int cnt; 1568 arc_state_t *state = hdr->b_l1hdr.b_state; 1569 1570 ASSERT(HDR_HAS_L1HDR(hdr)); 1571 ASSERT(state == arc_anon || MUTEX_HELD(hash_lock)); 1572 ASSERT(!GHOST_STATE(state)); 1573 1574 /* 1575 * arc_l2c_only counts as a ghost state so we don't need to explicitly 1576 * check to prevent usage of the arc_l2c_only list. 1577 */ 1578 if (((cnt = refcount_remove(&hdr->b_l1hdr.b_refcnt, tag)) == 0) && 1579 (state != arc_anon)) { 1580 uint64_t *size = &state->arcs_lsize[arc_buf_type(hdr)]; 1581 list_t *list; 1582 kmutex_t *lock; 1583 1584 get_buf_info(hdr, state, &list, &lock); 1585 ASSERT(!MUTEX_HELD(lock)); 1586 mutex_enter(lock); 1587 ASSERT(!list_link_active(&hdr->b_l1hdr.b_arc_node)); 1588 list_insert_head(list, hdr); 1589 ASSERT(hdr->b_l1hdr.b_datacnt > 0); 1590 atomic_add_64(size, hdr->b_size * 1591 hdr->b_l1hdr.b_datacnt); 1592 mutex_exit(lock); 1593 } 1594 return (cnt); 1595} 1596 1597/* 1598 * Move the supplied buffer to the indicated state. The mutex 1599 * for the buffer must be held by the caller. 1600 */ 1601static void 1602arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *hdr, 1603 kmutex_t *hash_lock) 1604{ 1605 arc_state_t *old_state; 1606 int64_t refcnt; 1607 uint32_t datacnt; 1608 uint64_t from_delta, to_delta; 1609 arc_buf_contents_t buftype = arc_buf_type(hdr); 1610 list_t *list; 1611 kmutex_t *lock; 1612 1613 /* 1614 * We almost always have an L1 hdr here, since we call arc_hdr_realloc() 1615 * in arc_read() when bringing a buffer out of the L2ARC. However, the 1616 * L1 hdr doesn't always exist when we change state to arc_anon before 1617 * destroying a header, in which case reallocating to add the L1 hdr is 1618 * pointless. 1619 */ 1620 if (HDR_HAS_L1HDR(hdr)) { 1621 old_state = hdr->b_l1hdr.b_state; 1622 refcnt = refcount_count(&hdr->b_l1hdr.b_refcnt); 1623 datacnt = hdr->b_l1hdr.b_datacnt; 1624 } else { 1625 old_state = arc_l2c_only; 1626 refcnt = 0; 1627 datacnt = 0; 1628 } 1629 1630 ASSERT(MUTEX_HELD(hash_lock)); 1631 ASSERT3P(new_state, !=, old_state); 1632 ASSERT(refcnt == 0 || datacnt > 0); 1633 ASSERT(!GHOST_STATE(new_state) || datacnt == 0); 1634 ASSERT(old_state != arc_anon || datacnt <= 1); 1635 1636 from_delta = to_delta = datacnt * hdr->b_size; 1637 1638 /* 1639 * If this buffer is evictable, transfer it from the 1640 * old state list to the new state list. 1641 */ 1642 if (refcnt == 0) { 1643 if (old_state != arc_anon && old_state != arc_l2c_only) { 1644 int use_mutex; 1645 uint64_t *size = &old_state->arcs_lsize[buftype]; 1646 1647 get_buf_info(hdr, old_state, &list, &lock); 1648 use_mutex = !MUTEX_HELD(lock); 1649 if (use_mutex) 1650 mutex_enter(lock); 1651 1652 ASSERT(HDR_HAS_L1HDR(hdr)); 1653 ASSERT(list_link_active(&hdr->b_l1hdr.b_arc_node)); 1654 list_remove(list, hdr); 1655 1656 /* 1657 * If prefetching out of the ghost cache, 1658 * we will have a non-zero datacnt. 1659 */ 1660 if (GHOST_STATE(old_state) && datacnt == 0) { 1661 /* ghost elements have a ghost size */ 1662 ASSERT(hdr->b_l1hdr.b_buf == NULL); 1663 from_delta = hdr->b_size; 1664 } 1665 ASSERT3U(*size, >=, from_delta); 1666 atomic_add_64(size, -from_delta); 1667 1668 if (use_mutex) 1669 mutex_exit(lock); 1670 } 1671 if (new_state != arc_anon && new_state != arc_l2c_only) { 1672 int use_mutex; 1673 uint64_t *size = &new_state->arcs_lsize[buftype]; 1674 1675 /* 1676 * An L1 header always exists here, since if we're 1677 * moving to some L1-cached state (i.e. not l2c_only or 1678 * anonymous), we realloc the header to add an L1hdr 1679 * beforehand. 1680 */ 1681 ASSERT(HDR_HAS_L1HDR(hdr)); 1682 get_buf_info(hdr, new_state, &list, &lock); 1683 use_mutex = !MUTEX_HELD(lock); 1684 if (use_mutex) 1685 mutex_enter(lock); 1686 1687 list_insert_head(list, hdr); 1688 1689 /* ghost elements have a ghost size */ 1690 if (GHOST_STATE(new_state)) { 1691 ASSERT(datacnt == 0); 1692 ASSERT(hdr->b_l1hdr.b_buf == NULL); 1693 to_delta = hdr->b_size; 1694 } 1695 atomic_add_64(size, to_delta); 1696 1697 if (use_mutex) 1698 mutex_exit(lock); 1699 } 1700 } 1701 1702 ASSERT(!BUF_EMPTY(hdr)); 1703 if (new_state == arc_anon && HDR_IN_HASH_TABLE(hdr)) 1704 buf_hash_remove(hdr); 1705 1706 /* adjust state sizes (ignore arc_l2c_only) */ 1707 if (to_delta && new_state != arc_l2c_only) 1708 atomic_add_64(&new_state->arcs_size, to_delta); 1709 if (from_delta && old_state != arc_l2c_only) { 1710 ASSERT3U(old_state->arcs_size, >=, from_delta); 1711 atomic_add_64(&old_state->arcs_size, -from_delta); 1712 } 1713 if (HDR_HAS_L1HDR(hdr)) 1714 hdr->b_l1hdr.b_state = new_state; 1715 1716 /* 1717 * L2 headers should never be on the L2 state list since they don't 1718 * have L1 headers allocated. 1719 */ 1720 ASSERT(list_is_empty(&arc_l2c_only->arcs_list[ARC_BUFC_DATA]) && 1721 list_is_empty(&arc_l2c_only->arcs_list[ARC_BUFC_METADATA])); 1722} 1723 1724void 1725arc_space_consume(uint64_t space, arc_space_type_t type) 1726{ 1727 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES); 1728 1729 switch (type) { 1730 case ARC_SPACE_DATA: 1731 ARCSTAT_INCR(arcstat_data_size, space); 1732 break; 1733 case ARC_SPACE_OTHER: 1734 ARCSTAT_INCR(arcstat_other_size, space); 1735 break; 1736 case ARC_SPACE_HDRS: 1737 ARCSTAT_INCR(arcstat_hdr_size, space); 1738 break; 1739 case ARC_SPACE_L2HDRS: 1740 ARCSTAT_INCR(arcstat_l2_hdr_size, space); 1741 break; 1742 } 1743 1744 ARCSTAT_INCR(arcstat_meta_used, space); 1745 atomic_add_64(&arc_size, space); 1746} 1747 1748void 1749arc_space_return(uint64_t space, arc_space_type_t type) 1750{ 1751 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES); 1752 1753 switch (type) { 1754 case ARC_SPACE_DATA: 1755 ARCSTAT_INCR(arcstat_data_size, -space); 1756 break; 1757 case ARC_SPACE_OTHER: 1758 ARCSTAT_INCR(arcstat_other_size, -space); 1759 break; 1760 case ARC_SPACE_HDRS: 1761 ARCSTAT_INCR(arcstat_hdr_size, -space); 1762 break; 1763 case ARC_SPACE_L2HDRS: 1764 ARCSTAT_INCR(arcstat_l2_hdr_size, -space); 1765 break; 1766 } 1767 1768 ASSERT(arc_meta_used >= space); 1769 if (arc_meta_max < arc_meta_used) 1770 arc_meta_max = arc_meta_used; 1771 ARCSTAT_INCR(arcstat_meta_used, -space); 1772 ASSERT(arc_size >= space); 1773 atomic_add_64(&arc_size, -space); 1774} 1775 1776arc_buf_t * 1777arc_buf_alloc(spa_t *spa, int32_t size, void *tag, arc_buf_contents_t type) 1778{ 1779 arc_buf_hdr_t *hdr; 1780 arc_buf_t *buf; 1781 1782 ASSERT3U(size, >, 0); 1783 hdr = kmem_cache_alloc(hdr_full_cache, KM_PUSHPAGE); 1784 ASSERT(BUF_EMPTY(hdr)); 1785 ASSERT3P(hdr->b_freeze_cksum, ==, NULL); 1786 hdr->b_size = size; 1787 hdr->b_spa = spa_load_guid(spa); 1788 1789 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 1790 buf->b_hdr = hdr; 1791 buf->b_data = NULL; 1792 buf->b_efunc = NULL; 1793 buf->b_private = NULL; 1794 buf->b_next = NULL; 1795 1796 hdr->b_flags = arc_bufc_to_flags(type); 1797 hdr->b_flags |= ARC_FLAG_HAS_L1HDR; 1798 1799 hdr->b_l1hdr.b_buf = buf; 1800 hdr->b_l1hdr.b_state = arc_anon; 1801 hdr->b_l1hdr.b_arc_access = 0; 1802 hdr->b_l1hdr.b_datacnt = 1; 1803 1804 arc_get_data_buf(buf); 1805 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 1806 (void) refcount_add(&hdr->b_l1hdr.b_refcnt, tag); 1807 1808 return (buf); 1809} 1810 1811static char *arc_onloan_tag = "onloan"; 1812 1813/* 1814 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in 1815 * flight data by arc_tempreserve_space() until they are "returned". Loaned 1816 * buffers must be returned to the arc before they can be used by the DMU or 1817 * freed. 1818 */ 1819arc_buf_t * 1820arc_loan_buf(spa_t *spa, int size) 1821{ 1822 arc_buf_t *buf; 1823 1824 buf = arc_buf_alloc(spa, size, arc_onloan_tag, ARC_BUFC_DATA); 1825 1826 atomic_add_64(&arc_loaned_bytes, size); 1827 return (buf); 1828} 1829 1830/* 1831 * Return a loaned arc buffer to the arc. 1832 */ 1833void 1834arc_return_buf(arc_buf_t *buf, void *tag) 1835{ 1836 arc_buf_hdr_t *hdr = buf->b_hdr; 1837 1838 ASSERT(buf->b_data != NULL); 1839 ASSERT(HDR_HAS_L1HDR(hdr)); 1840 (void) refcount_add(&hdr->b_l1hdr.b_refcnt, tag); 1841 (void) refcount_remove(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag); 1842 1843 atomic_add_64(&arc_loaned_bytes, -hdr->b_size); 1844} 1845 1846/* Detach an arc_buf from a dbuf (tag) */ 1847void 1848arc_loan_inuse_buf(arc_buf_t *buf, void *tag) 1849{ 1850 arc_buf_hdr_t *hdr = buf->b_hdr; 1851 1852 ASSERT(buf->b_data != NULL); 1853 ASSERT(HDR_HAS_L1HDR(hdr)); 1854 (void) refcount_add(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag); 1855 (void) refcount_remove(&hdr->b_l1hdr.b_refcnt, tag); 1856 buf->b_efunc = NULL; 1857 buf->b_private = NULL; 1858 1859 atomic_add_64(&arc_loaned_bytes, hdr->b_size); 1860} 1861 1862static arc_buf_t * 1863arc_buf_clone(arc_buf_t *from) 1864{ 1865 arc_buf_t *buf; 1866 arc_buf_hdr_t *hdr = from->b_hdr; 1867 uint64_t size = hdr->b_size; 1868 1869 ASSERT(HDR_HAS_L1HDR(hdr)); 1870 ASSERT(hdr->b_l1hdr.b_state != arc_anon); 1871 1872 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 1873 buf->b_hdr = hdr; 1874 buf->b_data = NULL; 1875 buf->b_efunc = NULL; 1876 buf->b_private = NULL; 1877 buf->b_next = hdr->b_l1hdr.b_buf; 1878 hdr->b_l1hdr.b_buf = buf; 1879 arc_get_data_buf(buf); 1880 bcopy(from->b_data, buf->b_data, size); 1881 1882 /* 1883 * This buffer already exists in the arc so create a duplicate 1884 * copy for the caller. If the buffer is associated with user data 1885 * then track the size and number of duplicates. These stats will be 1886 * updated as duplicate buffers are created and destroyed. 1887 */ 1888 if (HDR_ISTYPE_DATA(hdr)) { 1889 ARCSTAT_BUMP(arcstat_duplicate_buffers); 1890 ARCSTAT_INCR(arcstat_duplicate_buffers_size, size); 1891 } 1892 hdr->b_l1hdr.b_datacnt += 1; 1893 return (buf); 1894} 1895 1896void 1897arc_buf_add_ref(arc_buf_t *buf, void* tag) 1898{ 1899 arc_buf_hdr_t *hdr; 1900 kmutex_t *hash_lock; 1901 1902 /* 1903 * Check to see if this buffer is evicted. Callers 1904 * must verify b_data != NULL to know if the add_ref 1905 * was successful. 1906 */ 1907 mutex_enter(&buf->b_evict_lock); 1908 if (buf->b_data == NULL) { 1909 mutex_exit(&buf->b_evict_lock); 1910 return; 1911 } 1912 hash_lock = HDR_LOCK(buf->b_hdr); 1913 mutex_enter(hash_lock); 1914 hdr = buf->b_hdr; 1915 ASSERT(HDR_HAS_L1HDR(hdr)); 1916 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 1917 mutex_exit(&buf->b_evict_lock); 1918 1919 ASSERT(hdr->b_l1hdr.b_state == arc_mru || 1920 hdr->b_l1hdr.b_state == arc_mfu); 1921 1922 add_reference(hdr, hash_lock, tag); 1923 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr); 1924 arc_access(hdr, hash_lock); 1925 mutex_exit(hash_lock); 1926 ARCSTAT_BUMP(arcstat_hits); 1927 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr), 1928 demand, prefetch, !HDR_ISTYPE_METADATA(hdr), 1929 data, metadata, hits); 1930} 1931 1932static void 1933arc_buf_free_on_write(void *data, size_t size, 1934 void (*free_func)(void *, size_t)) 1935{ 1936 l2arc_data_free_t *df; 1937 1938 df = kmem_alloc(sizeof (l2arc_data_free_t), KM_SLEEP); 1939 df->l2df_data = data; 1940 df->l2df_size = size; 1941 df->l2df_func = free_func; 1942 mutex_enter(&l2arc_free_on_write_mtx); 1943 list_insert_head(l2arc_free_on_write, df); 1944 mutex_exit(&l2arc_free_on_write_mtx); 1945} 1946 1947/* 1948 * Free the arc data buffer. If it is an l2arc write in progress, 1949 * the buffer is placed on l2arc_free_on_write to be freed later. 1950 */ 1951static void 1952arc_buf_data_free(arc_buf_t *buf, void (*free_func)(void *, size_t)) 1953{ 1954 arc_buf_hdr_t *hdr = buf->b_hdr; 1955 1956 if (HDR_L2_WRITING(hdr)) { 1957 arc_buf_free_on_write(buf->b_data, hdr->b_size, free_func); 1958 ARCSTAT_BUMP(arcstat_l2_free_on_write); 1959 } else { 1960 free_func(buf->b_data, hdr->b_size); 1961 } 1962} 1963 1964/* 1965 * Free up buf->b_data and if 'remove' is set, then pull the 1966 * arc_buf_t off of the the arc_buf_hdr_t's list and free it. 1967 */ 1968static void 1969arc_buf_l2_cdata_free(arc_buf_hdr_t *hdr) 1970{ 1971 ASSERT(HDR_HAS_L2HDR(hdr)); 1972 ASSERT(MUTEX_HELD(&hdr->b_l2hdr.b_dev->l2ad_mtx)); 1973 1974 /* 1975 * The b_tmp_cdata field is linked off of the b_l1hdr, so if 1976 * that doesn't exist, the header is in the arc_l2c_only state, 1977 * and there isn't anything to free (it's already been freed). 1978 */ 1979 if (!HDR_HAS_L1HDR(hdr)) 1980 return; 1981 1982 if (hdr->b_l1hdr.b_tmp_cdata == NULL) 1983 return; 1984 1985 ASSERT(HDR_L2_WRITING(hdr)); 1986 arc_buf_free_on_write(hdr->b_l1hdr.b_tmp_cdata, hdr->b_size, 1987 zio_data_buf_free); 1988 1989 ARCSTAT_BUMP(arcstat_l2_cdata_free_on_write); 1990 hdr->b_l1hdr.b_tmp_cdata = NULL; 1991} 1992 1993static void 1994arc_buf_destroy(arc_buf_t *buf, boolean_t recycle, boolean_t remove) 1995{ 1996 arc_buf_t **bufp; 1997 1998 /* free up data associated with the buf */ 1999 if (buf->b_data != NULL) { 2000 arc_state_t *state = buf->b_hdr->b_l1hdr.b_state; 2001 uint64_t size = buf->b_hdr->b_size; 2002 arc_buf_contents_t type = arc_buf_type(buf->b_hdr); 2003 2004 arc_cksum_verify(buf); 2005#ifdef illumos 2006 arc_buf_unwatch(buf); 2007#endif /* illumos */ 2008 2009 if (!recycle) { 2010 if (type == ARC_BUFC_METADATA) { 2011 arc_buf_data_free(buf, zio_buf_free); 2012 arc_space_return(size, ARC_SPACE_DATA); 2013 } else { 2014 ASSERT(type == ARC_BUFC_DATA); 2015 arc_buf_data_free(buf, zio_data_buf_free); 2016 ARCSTAT_INCR(arcstat_data_size, -size); 2017 atomic_add_64(&arc_size, -size); 2018 } 2019 } 2020 if (list_link_active(&buf->b_hdr->b_l1hdr.b_arc_node)) { 2021 uint64_t *cnt = &state->arcs_lsize[type]; 2022 2023 ASSERT(refcount_is_zero( 2024 &buf->b_hdr->b_l1hdr.b_refcnt)); 2025 ASSERT(state != arc_anon && state != arc_l2c_only); 2026 2027 ASSERT3U(*cnt, >=, size); 2028 atomic_add_64(cnt, -size); 2029 } 2030 ASSERT3U(state->arcs_size, >=, size); 2031 atomic_add_64(&state->arcs_size, -size); 2032 buf->b_data = NULL; 2033 2034 /* 2035 * If we're destroying a duplicate buffer make sure 2036 * that the appropriate statistics are updated. 2037 */ 2038 if (buf->b_hdr->b_l1hdr.b_datacnt > 1 && 2039 HDR_ISTYPE_DATA(buf->b_hdr)) { 2040 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers); 2041 ARCSTAT_INCR(arcstat_duplicate_buffers_size, -size); 2042 } 2043 ASSERT(buf->b_hdr->b_l1hdr.b_datacnt > 0); 2044 buf->b_hdr->b_l1hdr.b_datacnt -= 1; 2045 } 2046 2047 /* only remove the buf if requested */ 2048 if (!remove) 2049 return; 2050 2051 /* remove the buf from the hdr list */ 2052 for (bufp = &buf->b_hdr->b_l1hdr.b_buf; *bufp != buf; 2053 bufp = &(*bufp)->b_next) 2054 continue; 2055 *bufp = buf->b_next; 2056 buf->b_next = NULL; 2057 2058 ASSERT(buf->b_efunc == NULL); 2059 2060 /* clean up the buf */ 2061 buf->b_hdr = NULL; 2062 kmem_cache_free(buf_cache, buf); 2063} 2064 2065static void 2066arc_hdr_destroy(arc_buf_hdr_t *hdr) 2067{ 2068 if (HDR_HAS_L1HDR(hdr)) { 2069 ASSERT(hdr->b_l1hdr.b_buf == NULL || 2070 hdr->b_l1hdr.b_datacnt > 0); 2071 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 2072 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon); 2073 } 2074 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 2075 ASSERT(!HDR_IN_HASH_TABLE(hdr)); 2076 2077 if (HDR_HAS_L2HDR(hdr)) { 2078 l2arc_buf_hdr_t *l2hdr = &hdr->b_l2hdr; 2079 boolean_t buflist_held = MUTEX_HELD(&l2hdr->b_dev->l2ad_mtx); 2080 2081 if (!buflist_held) { 2082 mutex_enter(&l2hdr->b_dev->l2ad_mtx); 2083 l2hdr = &hdr->b_l2hdr; 2084 } 2085 2086 trim_map_free(l2hdr->b_dev->l2ad_vdev, l2hdr->b_daddr, 2087 l2hdr->b_asize, 0); 2088 list_remove(&l2hdr->b_dev->l2ad_buflist, hdr); 2089 2090 /* 2091 * We don't want to leak the b_tmp_cdata buffer that was 2092 * allocated in l2arc_write_buffers() 2093 */ 2094 arc_buf_l2_cdata_free(hdr); 2095 2096 ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size); 2097 ARCSTAT_INCR(arcstat_l2_asize, -l2hdr->b_asize); 2098 2099 if (!buflist_held) 2100 mutex_exit(&l2hdr->b_dev->l2ad_mtx); 2101 2102 hdr->b_flags &= ~ARC_FLAG_HAS_L2HDR; 2103 } 2104 2105 if (!BUF_EMPTY(hdr)) 2106 buf_discard_identity(hdr); 2107 if (hdr->b_freeze_cksum != NULL) { 2108 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 2109 hdr->b_freeze_cksum = NULL; 2110 } 2111 2112 if (HDR_HAS_L1HDR(hdr)) { 2113 while (hdr->b_l1hdr.b_buf) { 2114 arc_buf_t *buf = hdr->b_l1hdr.b_buf; 2115 2116 if (buf->b_efunc != NULL) { 2117 mutex_enter(&arc_eviction_mtx); 2118 mutex_enter(&buf->b_evict_lock); 2119 ASSERT(buf->b_hdr != NULL); 2120 arc_buf_destroy(hdr->b_l1hdr.b_buf, FALSE, 2121 FALSE); 2122 hdr->b_l1hdr.b_buf = buf->b_next; 2123 buf->b_hdr = &arc_eviction_hdr; 2124 buf->b_next = arc_eviction_list; 2125 arc_eviction_list = buf; 2126 mutex_exit(&buf->b_evict_lock); 2127 mutex_exit(&arc_eviction_mtx); 2128 } else { 2129 arc_buf_destroy(hdr->b_l1hdr.b_buf, FALSE, 2130 TRUE); 2131 } 2132 } 2133#ifdef ZFS_DEBUG 2134 if (hdr->b_l1hdr.b_thawed != NULL) { 2135 kmem_free(hdr->b_l1hdr.b_thawed, 1); 2136 hdr->b_l1hdr.b_thawed = NULL; 2137 } 2138#endif 2139 } 2140 2141 ASSERT3P(hdr->b_hash_next, ==, NULL); 2142 if (HDR_HAS_L1HDR(hdr)) { 2143 ASSERT(!list_link_active(&hdr->b_l1hdr.b_arc_node)); 2144 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL); 2145 kmem_cache_free(hdr_full_cache, hdr); 2146 } else { 2147 kmem_cache_free(hdr_l2only_cache, hdr); 2148 } 2149} 2150 2151void 2152arc_buf_free(arc_buf_t *buf, void *tag) 2153{ 2154 arc_buf_hdr_t *hdr = buf->b_hdr; 2155 int hashed = hdr->b_l1hdr.b_state != arc_anon; 2156 2157 ASSERT(buf->b_efunc == NULL); 2158 ASSERT(buf->b_data != NULL); 2159 2160 if (hashed) { 2161 kmutex_t *hash_lock = HDR_LOCK(hdr); 2162 2163 mutex_enter(hash_lock); 2164 hdr = buf->b_hdr; 2165 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 2166 2167 (void) remove_reference(hdr, hash_lock, tag); 2168 if (hdr->b_l1hdr.b_datacnt > 1) { 2169 arc_buf_destroy(buf, FALSE, TRUE); 2170 } else { 2171 ASSERT(buf == hdr->b_l1hdr.b_buf); 2172 ASSERT(buf->b_efunc == NULL); 2173 hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE; 2174 } 2175 mutex_exit(hash_lock); 2176 } else if (HDR_IO_IN_PROGRESS(hdr)) { 2177 int destroy_hdr; 2178 /* 2179 * We are in the middle of an async write. Don't destroy 2180 * this buffer unless the write completes before we finish 2181 * decrementing the reference count. 2182 */ 2183 mutex_enter(&arc_eviction_mtx); 2184 (void) remove_reference(hdr, NULL, tag); 2185 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 2186 destroy_hdr = !HDR_IO_IN_PROGRESS(hdr); 2187 mutex_exit(&arc_eviction_mtx); 2188 if (destroy_hdr) 2189 arc_hdr_destroy(hdr); 2190 } else { 2191 if (remove_reference(hdr, NULL, tag) > 0) 2192 arc_buf_destroy(buf, FALSE, TRUE); 2193 else 2194 arc_hdr_destroy(hdr); 2195 } 2196} 2197 2198boolean_t 2199arc_buf_remove_ref(arc_buf_t *buf, void* tag) 2200{ 2201 arc_buf_hdr_t *hdr = buf->b_hdr; 2202 kmutex_t *hash_lock = HDR_LOCK(hdr); 2203 boolean_t no_callback = (buf->b_efunc == NULL); 2204 2205 if (hdr->b_l1hdr.b_state == arc_anon) { 2206 ASSERT(hdr->b_l1hdr.b_datacnt == 1); 2207 arc_buf_free(buf, tag); 2208 return (no_callback); 2209 } 2210 2211 mutex_enter(hash_lock); 2212 hdr = buf->b_hdr; 2213 ASSERT(hdr->b_l1hdr.b_datacnt > 0); 2214 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 2215 ASSERT(hdr->b_l1hdr.b_state != arc_anon); 2216 ASSERT(buf->b_data != NULL); 2217 2218 (void) remove_reference(hdr, hash_lock, tag); 2219 if (hdr->b_l1hdr.b_datacnt > 1) { 2220 if (no_callback) 2221 arc_buf_destroy(buf, FALSE, TRUE); 2222 } else if (no_callback) { 2223 ASSERT(hdr->b_l1hdr.b_buf == buf && buf->b_next == NULL); 2224 ASSERT(buf->b_efunc == NULL); 2225 hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE; 2226 } 2227 ASSERT(no_callback || hdr->b_l1hdr.b_datacnt > 1 || 2228 refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 2229 mutex_exit(hash_lock); 2230 return (no_callback); 2231} 2232 2233int32_t 2234arc_buf_size(arc_buf_t *buf) 2235{ 2236 return (buf->b_hdr->b_size); 2237} 2238 2239/* 2240 * Called from the DMU to determine if the current buffer should be 2241 * evicted. In order to ensure proper locking, the eviction must be initiated 2242 * from the DMU. Return true if the buffer is associated with user data and 2243 * duplicate buffers still exist. 2244 */ 2245boolean_t 2246arc_buf_eviction_needed(arc_buf_t *buf) 2247{ 2248 arc_buf_hdr_t *hdr; 2249 boolean_t evict_needed = B_FALSE; 2250 2251 if (zfs_disable_dup_eviction) 2252 return (B_FALSE); 2253 2254 mutex_enter(&buf->b_evict_lock); 2255 hdr = buf->b_hdr; 2256 if (hdr == NULL) { 2257 /* 2258 * We are in arc_do_user_evicts(); let that function 2259 * perform the eviction. 2260 */ 2261 ASSERT(buf->b_data == NULL); 2262 mutex_exit(&buf->b_evict_lock); 2263 return (B_FALSE); 2264 } else if (buf->b_data == NULL) { 2265 /* 2266 * We have already been added to the arc eviction list; 2267 * recommend eviction. 2268 */ 2269 ASSERT3P(hdr, ==, &arc_eviction_hdr); 2270 mutex_exit(&buf->b_evict_lock); 2271 return (B_TRUE); 2272 } 2273 2274 if (hdr->b_l1hdr.b_datacnt > 1 && HDR_ISTYPE_DATA(hdr)) 2275 evict_needed = B_TRUE; 2276 2277 mutex_exit(&buf->b_evict_lock); 2278 return (evict_needed); 2279} 2280 2281/* 2282 * Evict buffers from list until we've removed the specified number of 2283 * bytes. Move the removed buffers to the appropriate evict state. 2284 * If the recycle flag is set, then attempt to "recycle" a buffer: 2285 * - look for a buffer to evict that is `bytes' long. 2286 * - return the data block from this buffer rather than freeing it. 2287 * This flag is used by callers that are trying to make space for a 2288 * new buffer in a full arc cache. 2289 * 2290 * This function makes a "best effort". It skips over any buffers 2291 * it can't get a hash_lock on, and so may not catch all candidates. 2292 * It may also return without evicting as much space as requested. 2293 */ 2294static void * 2295arc_evict(arc_state_t *state, uint64_t spa, int64_t bytes, boolean_t recycle, 2296 arc_buf_contents_t type) 2297{ 2298 arc_state_t *evicted_state; 2299 uint64_t bytes_evicted = 0, skipped = 0, missed = 0; 2300 int64_t bytes_remaining; 2301 arc_buf_hdr_t *hdr, *hdr_prev = NULL; 2302 list_t *evicted_list, *list, *evicted_list_start, *list_start; 2303 kmutex_t *lock, *evicted_lock; 2304 kmutex_t *hash_lock; 2305 boolean_t have_lock; 2306 void *stolen = NULL; 2307 arc_buf_hdr_t marker = { 0 }; 2308 int count = 0; 2309 static int evict_metadata_offset, evict_data_offset; 2310 int i, idx, offset, list_count, lists; 2311 2312 ASSERT(state == arc_mru || state == arc_mfu); 2313 2314 evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost; 2315 2316 /* 2317 * Decide which "type" (data vs metadata) to recycle from. 2318 * 2319 * If we are over the metadata limit, recycle from metadata. 2320 * If we are under the metadata minimum, recycle from data. 2321 * Otherwise, recycle from whichever type has the oldest (least 2322 * recently accessed) header. This is not yet implemented. 2323 */ 2324 if (recycle) { 2325 arc_buf_contents_t realtype; 2326 if (state->arcs_lsize[ARC_BUFC_DATA] == 0) { 2327 realtype = ARC_BUFC_METADATA; 2328 } else if (state->arcs_lsize[ARC_BUFC_METADATA] == 0) { 2329 realtype = ARC_BUFC_DATA; 2330 } else if (arc_meta_used >= arc_meta_limit) { 2331 realtype = ARC_BUFC_METADATA; 2332 } else if (arc_meta_used <= arc_meta_min) { 2333 realtype = ARC_BUFC_DATA; 2334#ifdef illumos 2335 } else if (HDR_HAS_L1HDR(data_hdr) && 2336 HDR_HAS_L1HDR(metadata_hdr) && 2337 data_hdr->b_l1hdr.b_arc_access < 2338 metadata_hdr->b_l1hdr.b_arc_access) { 2339 realtype = ARC_BUFC_DATA; 2340 } else { 2341 realtype = ARC_BUFC_METADATA; 2342#else 2343 } else { 2344 /* TODO */ 2345 realtype = type; 2346#endif 2347 } 2348 if (realtype != type) { 2349 /* 2350 * If we want to evict from a different list, 2351 * we can not recycle, because DATA vs METADATA 2352 * buffers are segregated into different kmem 2353 * caches (and vmem arenas). 2354 */ 2355 type = realtype; 2356 recycle = B_FALSE; 2357 } 2358 } 2359 2360 if (type == ARC_BUFC_METADATA) { 2361 offset = 0; 2362 list_count = ARC_BUFC_NUMMETADATALISTS; 2363 list_start = &state->arcs_lists[0]; 2364 evicted_list_start = &evicted_state->arcs_lists[0]; 2365 idx = evict_metadata_offset; 2366 } else { 2367 offset = ARC_BUFC_NUMMETADATALISTS; 2368 list_start = &state->arcs_lists[offset]; 2369 evicted_list_start = &evicted_state->arcs_lists[offset]; 2370 list_count = ARC_BUFC_NUMDATALISTS; 2371 idx = evict_data_offset; 2372 } 2373 bytes_remaining = evicted_state->arcs_lsize[type]; 2374 lists = 0; 2375 2376evict_start: 2377 list = &list_start[idx]; 2378 evicted_list = &evicted_list_start[idx]; 2379 lock = ARCS_LOCK(state, (offset + idx)); 2380 evicted_lock = ARCS_LOCK(evicted_state, (offset + idx)); 2381 2382 /* 2383 * The ghost list lock must be acquired first in order to prevent 2384 * a 3 party deadlock: 2385 * 2386 * - arc_evict_ghost acquires arc_*_ghost->arcs_mtx, followed by 2387 * l2ad_mtx in arc_hdr_realloc 2388 * - l2arc_write_buffers acquires l2ad_mtx, followed by arc_*->arcs_mtx 2389 * - arc_evict acquires arc_*_ghost->arcs_mtx, followed by 2390 * arc_*_ghost->arcs_mtx and forms a deadlock cycle. 2391 * 2392 * This situation is avoided by acquiring the ghost list lock first. 2393 */ 2394 mutex_enter(evicted_lock); 2395 mutex_enter(lock); 2396 2397 for (hdr = list_tail(list); hdr; hdr = hdr_prev) { 2398 hdr_prev = list_prev(list, hdr); 2399 if (HDR_HAS_L1HDR(hdr)) { 2400 bytes_remaining -= 2401 (hdr->b_size * hdr->b_l1hdr.b_datacnt); 2402 } 2403 /* prefetch buffers have a minimum lifespan */ 2404 if (HDR_IO_IN_PROGRESS(hdr) || 2405 (spa && hdr->b_spa != spa) || 2406 ((hdr->b_flags & (ARC_FLAG_PREFETCH | ARC_FLAG_INDIRECT)) && 2407 ddi_get_lbolt() - hdr->b_l1hdr.b_arc_access < 2408 arc_min_prefetch_lifespan)) { 2409 skipped++; 2410 continue; 2411 } 2412 /* "lookahead" for better eviction candidate */ 2413 if (recycle && hdr->b_size != bytes && 2414 hdr_prev && hdr_prev->b_size == bytes) 2415 continue; 2416 2417 /* ignore markers */ 2418 if (hdr->b_spa == 0) 2419 continue; 2420 2421 /* 2422 * It may take a long time to evict all the bufs requested. 2423 * To avoid blocking all arc activity, periodically drop 2424 * the arcs_mtx and give other threads a chance to run 2425 * before reacquiring the lock. 2426 * 2427 * If we are looking for a buffer to recycle, we are in 2428 * the hot code path, so don't sleep. 2429 */ 2430 if (!recycle && count++ > arc_evict_iterations) { 2431 list_insert_after(list, hdr, &marker); 2432 mutex_exit(lock); 2433 mutex_exit(evicted_lock); 2434 kpreempt(KPREEMPT_SYNC); 2435 mutex_enter(evicted_lock); 2436 mutex_enter(lock); 2437 hdr_prev = list_prev(list, &marker); 2438 list_remove(list, &marker); 2439 count = 0; 2440 continue; 2441 } 2442 2443 hash_lock = HDR_LOCK(hdr); 2444 have_lock = MUTEX_HELD(hash_lock); 2445 if (have_lock || mutex_tryenter(hash_lock)) { 2446 ASSERT0(refcount_count(&hdr->b_l1hdr.b_refcnt)); 2447 ASSERT3U(hdr->b_l1hdr.b_datacnt, >, 0); 2448 while (hdr->b_l1hdr.b_buf) { 2449 arc_buf_t *buf = hdr->b_l1hdr.b_buf; 2450 if (!mutex_tryenter(&buf->b_evict_lock)) { 2451 missed += 1; 2452 break; 2453 } 2454 if (buf->b_data != NULL) { 2455 bytes_evicted += hdr->b_size; 2456 if (recycle && 2457 arc_buf_type(hdr) == type && 2458 hdr->b_size == bytes && 2459 !HDR_L2_WRITING(hdr)) { 2460 stolen = buf->b_data; 2461 recycle = FALSE; 2462 } 2463 } 2464 if (buf->b_efunc != NULL) { 2465 mutex_enter(&arc_eviction_mtx); 2466 arc_buf_destroy(buf, 2467 buf->b_data == stolen, FALSE); 2468 hdr->b_l1hdr.b_buf = buf->b_next; 2469 buf->b_hdr = &arc_eviction_hdr; 2470 buf->b_next = arc_eviction_list; 2471 arc_eviction_list = buf; 2472 mutex_exit(&arc_eviction_mtx); 2473 mutex_exit(&buf->b_evict_lock); 2474 } else { 2475 mutex_exit(&buf->b_evict_lock); 2476 arc_buf_destroy(buf, 2477 buf->b_data == stolen, TRUE); 2478 } 2479 } 2480 2481 if (HDR_HAS_L2HDR(hdr)) { 2482 ARCSTAT_INCR(arcstat_evict_l2_cached, 2483 hdr->b_size); 2484 } else { 2485 if (l2arc_write_eligible(hdr->b_spa, hdr)) { 2486 ARCSTAT_INCR(arcstat_evict_l2_eligible, 2487 hdr->b_size); 2488 } else { 2489 ARCSTAT_INCR( 2490 arcstat_evict_l2_ineligible, 2491 hdr->b_size); 2492 } 2493 } 2494 2495 if (hdr->b_l1hdr.b_datacnt == 0) { 2496 arc_change_state(evicted_state, hdr, hash_lock); 2497 ASSERT(HDR_IN_HASH_TABLE(hdr)); 2498 hdr->b_flags |= ARC_FLAG_IN_HASH_TABLE; 2499 hdr->b_flags &= ~ARC_FLAG_BUF_AVAILABLE; 2500 DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, hdr); 2501 } 2502 if (!have_lock) 2503 mutex_exit(hash_lock); 2504 if (bytes >= 0 && bytes_evicted >= bytes) 2505 break; 2506 if (bytes_remaining > 0) { 2507 mutex_exit(evicted_lock); 2508 mutex_exit(lock); 2509 idx = ((idx + 1) & (list_count - 1)); 2510 lists++; 2511 goto evict_start; 2512 } 2513 } else { 2514 missed += 1; 2515 } 2516 } 2517 2518 mutex_exit(lock); 2519 mutex_exit(evicted_lock); 2520 2521 idx = ((idx + 1) & (list_count - 1)); 2522 lists++; 2523 2524 if (bytes_evicted < bytes) { 2525 if (lists < list_count) 2526 goto evict_start; 2527 else 2528 dprintf("only evicted %lld bytes from %x", 2529 (longlong_t)bytes_evicted, state); 2530 } 2531 if (type == ARC_BUFC_METADATA) 2532 evict_metadata_offset = idx; 2533 else 2534 evict_data_offset = idx; 2535 2536 if (skipped) 2537 ARCSTAT_INCR(arcstat_evict_skip, skipped); 2538 2539 if (missed) 2540 ARCSTAT_INCR(arcstat_mutex_miss, missed); 2541 2542 /* 2543 * Note: we have just evicted some data into the ghost state, 2544 * potentially putting the ghost size over the desired size. Rather 2545 * that evicting from the ghost list in this hot code path, leave 2546 * this chore to the arc_reclaim_thread(). 2547 */ 2548 2549 if (stolen) 2550 ARCSTAT_BUMP(arcstat_stolen); 2551 return (stolen); 2552} 2553 2554/* 2555 * Remove buffers from list until we've removed the specified number of 2556 * bytes. Destroy the buffers that are removed. 2557 */ 2558static void 2559arc_evict_ghost(arc_state_t *state, uint64_t spa, int64_t bytes) 2560{ 2561 arc_buf_hdr_t *hdr, *hdr_prev; 2562 arc_buf_hdr_t marker = { 0 }; 2563 list_t *list, *list_start; 2564 kmutex_t *hash_lock, *lock; 2565 uint64_t bytes_deleted = 0; 2566 uint64_t bufs_skipped = 0; 2567 int count = 0; 2568 static int evict_offset; 2569 int list_count, idx = evict_offset; 2570 int offset, lists = 0; 2571 2572 ASSERT(GHOST_STATE(state)); 2573 2574 /* 2575 * data lists come after metadata lists 2576 */ 2577 list_start = &state->arcs_lists[ARC_BUFC_NUMMETADATALISTS]; 2578 list_count = ARC_BUFC_NUMDATALISTS; 2579 offset = ARC_BUFC_NUMMETADATALISTS; 2580 2581evict_start: 2582 list = &list_start[idx]; 2583 lock = ARCS_LOCK(state, idx + offset); 2584 2585 mutex_enter(lock); 2586 for (hdr = list_tail(list); hdr; hdr = hdr_prev) { 2587 hdr_prev = list_prev(list, hdr); 2588 if (arc_buf_type(hdr) >= ARC_BUFC_NUMTYPES) 2589 panic("invalid hdr=%p", (void *)hdr); 2590 if (spa && hdr->b_spa != spa) 2591 continue; 2592 2593 /* ignore markers */ 2594 if (hdr->b_spa == 0) 2595 continue; 2596 2597 hash_lock = HDR_LOCK(hdr); 2598 /* caller may be trying to modify this buffer, skip it */ 2599 if (MUTEX_HELD(hash_lock)) 2600 continue; 2601 2602 /* 2603 * It may take a long time to evict all the bufs requested. 2604 * To avoid blocking all arc activity, periodically drop 2605 * the arcs_mtx and give other threads a chance to run 2606 * before reacquiring the lock. 2607 */ 2608 if (count++ > arc_evict_iterations) { 2609 list_insert_after(list, hdr, &marker); 2610 mutex_exit(lock); 2611 kpreempt(KPREEMPT_SYNC); 2612 mutex_enter(lock); 2613 hdr_prev = list_prev(list, &marker); 2614 list_remove(list, &marker); 2615 count = 0; 2616 continue; 2617 } 2618 if (mutex_tryenter(hash_lock)) { 2619 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 2620 ASSERT(!HDR_HAS_L1HDR(hdr) || 2621 hdr->b_l1hdr.b_buf == NULL); 2622 ARCSTAT_BUMP(arcstat_deleted); 2623 bytes_deleted += hdr->b_size; 2624 2625 if (HDR_HAS_L2HDR(hdr)) { 2626 /* 2627 * This buffer is cached on the 2nd Level ARC; 2628 * don't destroy the header. 2629 */ 2630 arc_change_state(arc_l2c_only, hdr, hash_lock); 2631 /* 2632 * dropping from L1+L2 cached to L2-only, 2633 * realloc to remove the L1 header. 2634 */ 2635 hdr = arc_hdr_realloc(hdr, hdr_full_cache, 2636 hdr_l2only_cache); 2637 mutex_exit(hash_lock); 2638 } else { 2639 arc_change_state(arc_anon, hdr, hash_lock); 2640 mutex_exit(hash_lock); 2641 arc_hdr_destroy(hdr); 2642 } 2643 2644 DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, hdr); 2645 if (bytes >= 0 && bytes_deleted >= bytes) 2646 break; 2647 } else if (bytes < 0) { 2648 /* 2649 * Insert a list marker and then wait for the 2650 * hash lock to become available. Once its 2651 * available, restart from where we left off. 2652 */ 2653 list_insert_after(list, hdr, &marker); 2654 mutex_exit(lock); 2655 mutex_enter(hash_lock); 2656 mutex_exit(hash_lock); 2657 mutex_enter(lock); 2658 hdr_prev = list_prev(list, &marker); 2659 list_remove(list, &marker); 2660 } else { 2661 bufs_skipped += 1; 2662 } 2663 2664 } 2665 mutex_exit(lock); 2666 idx = ((idx + 1) & (ARC_BUFC_NUMDATALISTS - 1)); 2667 lists++; 2668 2669 if (lists < list_count) 2670 goto evict_start; 2671 2672 evict_offset = idx; 2673 if ((uintptr_t)list > (uintptr_t)&state->arcs_lists[ARC_BUFC_NUMMETADATALISTS] && 2674 (bytes < 0 || bytes_deleted < bytes)) { 2675 list_start = &state->arcs_lists[0]; 2676 list_count = ARC_BUFC_NUMMETADATALISTS; 2677 offset = lists = 0; 2678 goto evict_start; 2679 } 2680 2681 if (bufs_skipped) { 2682 ARCSTAT_INCR(arcstat_mutex_miss, bufs_skipped); 2683 ASSERT(bytes >= 0); 2684 } 2685 2686 if (bytes_deleted < bytes) 2687 dprintf("only deleted %lld bytes from %p", 2688 (longlong_t)bytes_deleted, state); 2689} 2690 2691static void 2692arc_adjust(void) 2693{ 2694 int64_t adjustment, delta; 2695 2696 /* 2697 * Adjust MRU size 2698 */ 2699 2700 adjustment = MIN((int64_t)(arc_size - arc_c), 2701 (int64_t)(arc_anon->arcs_size + arc_mru->arcs_size + arc_meta_used - 2702 arc_p)); 2703 2704 if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_DATA] > 0) { 2705 delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_DATA], adjustment); 2706 (void) arc_evict(arc_mru, 0, delta, FALSE, ARC_BUFC_DATA); 2707 adjustment -= delta; 2708 } 2709 2710 if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_METADATA] > 0) { 2711 delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_METADATA], adjustment); 2712 (void) arc_evict(arc_mru, 0, delta, FALSE, 2713 ARC_BUFC_METADATA); 2714 } 2715 2716 /* 2717 * Adjust MFU size 2718 */ 2719 2720 adjustment = arc_size - arc_c; 2721 2722 if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_DATA] > 0) { 2723 delta = MIN(adjustment, arc_mfu->arcs_lsize[ARC_BUFC_DATA]); 2724 (void) arc_evict(arc_mfu, 0, delta, FALSE, ARC_BUFC_DATA); 2725 adjustment -= delta; 2726 } 2727 2728 if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_METADATA] > 0) { 2729 int64_t delta = MIN(adjustment, 2730 arc_mfu->arcs_lsize[ARC_BUFC_METADATA]); 2731 (void) arc_evict(arc_mfu, 0, delta, FALSE, 2732 ARC_BUFC_METADATA); 2733 } 2734 2735 /* 2736 * Adjust ghost lists 2737 */ 2738 2739 adjustment = arc_mru->arcs_size + arc_mru_ghost->arcs_size - arc_c; 2740 2741 if (adjustment > 0 && arc_mru_ghost->arcs_size > 0) { 2742 delta = MIN(arc_mru_ghost->arcs_size, adjustment); 2743 arc_evict_ghost(arc_mru_ghost, 0, delta); 2744 } 2745 2746 adjustment = 2747 arc_mru_ghost->arcs_size + arc_mfu_ghost->arcs_size - arc_c; 2748 2749 if (adjustment > 0 && arc_mfu_ghost->arcs_size > 0) { 2750 delta = MIN(arc_mfu_ghost->arcs_size, adjustment); 2751 arc_evict_ghost(arc_mfu_ghost, 0, delta); 2752 } 2753} 2754 2755static void 2756arc_do_user_evicts(void) 2757{ 2758 static arc_buf_t *tmp_arc_eviction_list; 2759 2760 /* 2761 * Move list over to avoid LOR 2762 */ 2763restart: 2764 mutex_enter(&arc_eviction_mtx); 2765 tmp_arc_eviction_list = arc_eviction_list; 2766 arc_eviction_list = NULL; 2767 mutex_exit(&arc_eviction_mtx); 2768 2769 while (tmp_arc_eviction_list != NULL) { 2770 arc_buf_t *buf = tmp_arc_eviction_list; 2771 tmp_arc_eviction_list = buf->b_next; 2772 mutex_enter(&buf->b_evict_lock); 2773 buf->b_hdr = NULL; 2774 mutex_exit(&buf->b_evict_lock); 2775 2776 if (buf->b_efunc != NULL) 2777 VERIFY0(buf->b_efunc(buf->b_private)); 2778 2779 buf->b_efunc = NULL; 2780 buf->b_private = NULL; 2781 kmem_cache_free(buf_cache, buf); 2782 } 2783 2784 if (arc_eviction_list != NULL) 2785 goto restart; 2786} 2787 2788/* 2789 * Flush all *evictable* data from the cache for the given spa. 2790 * NOTE: this will not touch "active" (i.e. referenced) data. 2791 */ 2792void 2793arc_flush(spa_t *spa) 2794{ 2795 uint64_t guid = 0; 2796 2797 if (spa != NULL) 2798 guid = spa_load_guid(spa); 2799 2800 while (arc_mru->arcs_lsize[ARC_BUFC_DATA]) { 2801 (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_DATA); 2802 if (spa != NULL) 2803 break; 2804 } 2805 while (arc_mru->arcs_lsize[ARC_BUFC_METADATA]) { 2806 (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_METADATA); 2807 if (spa != NULL) 2808 break; 2809 } 2810 while (arc_mfu->arcs_lsize[ARC_BUFC_DATA]) { 2811 (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_DATA); 2812 if (spa != NULL) 2813 break; 2814 } 2815 while (arc_mfu->arcs_lsize[ARC_BUFC_METADATA]) { 2816 (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_METADATA); 2817 if (spa != NULL) 2818 break; 2819 } 2820 2821 arc_evict_ghost(arc_mru_ghost, guid, -1); 2822 arc_evict_ghost(arc_mfu_ghost, guid, -1); 2823 2824 mutex_enter(&arc_reclaim_thr_lock); 2825 arc_do_user_evicts(); 2826 mutex_exit(&arc_reclaim_thr_lock); 2827 ASSERT(spa || arc_eviction_list == NULL); 2828} 2829 2830void 2831arc_shrink(void) 2832{ 2833 2834 if (arc_c > arc_c_min) { 2835 uint64_t to_free; 2836 2837 to_free = arc_c >> arc_shrink_shift; 2838 DTRACE_PROBE4(arc__shrink, uint64_t, arc_c, uint64_t, 2839 arc_c_min, uint64_t, arc_p, uint64_t, to_free); 2840 if (arc_c > arc_c_min + to_free) 2841 atomic_add_64(&arc_c, -to_free); 2842 else 2843 arc_c = arc_c_min; 2844 2845 atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift)); 2846 if (arc_c > arc_size) 2847 arc_c = MAX(arc_size, arc_c_min); 2848 if (arc_p > arc_c) 2849 arc_p = (arc_c >> 1); 2850 2851 DTRACE_PROBE2(arc__shrunk, uint64_t, arc_c, uint64_t, 2852 arc_p); 2853 2854 ASSERT(arc_c >= arc_c_min); 2855 ASSERT((int64_t)arc_p >= 0); 2856 } 2857 2858 if (arc_size > arc_c) { 2859 DTRACE_PROBE2(arc__shrink_adjust, uint64_t, arc_size, 2860 uint64_t, arc_c); 2861 arc_adjust(); 2862 } 2863} 2864 2865static int needfree = 0; 2866 2867static int 2868arc_reclaim_needed(void) 2869{ 2870 2871#ifdef _KERNEL 2872 2873 if (needfree) { 2874 DTRACE_PROBE(arc__reclaim_needfree); 2875 return (1); 2876 } 2877 2878 /* 2879 * Cooperate with pagedaemon when it's time for it to scan 2880 * and reclaim some pages. 2881 */ 2882 if (freemem < zfs_arc_free_target) { 2883 DTRACE_PROBE2(arc__reclaim_freemem, uint64_t, 2884 freemem, uint64_t, zfs_arc_free_target); 2885 return (1); 2886 } 2887 2888#ifdef sun 2889 /* 2890 * take 'desfree' extra pages, so we reclaim sooner, rather than later 2891 */ 2892 extra = desfree; 2893 2894 /* 2895 * check that we're out of range of the pageout scanner. It starts to 2896 * schedule paging if freemem is less than lotsfree and needfree. 2897 * lotsfree is the high-water mark for pageout, and needfree is the 2898 * number of needed free pages. We add extra pages here to make sure 2899 * the scanner doesn't start up while we're freeing memory. 2900 */ 2901 if (freemem < lotsfree + needfree + extra) 2902 return (1); 2903 2904 /* 2905 * check to make sure that swapfs has enough space so that anon 2906 * reservations can still succeed. anon_resvmem() checks that the 2907 * availrmem is greater than swapfs_minfree, and the number of reserved 2908 * swap pages. We also add a bit of extra here just to prevent 2909 * circumstances from getting really dire. 2910 */ 2911 if (availrmem < swapfs_minfree + swapfs_reserve + extra) 2912 return (1); 2913 2914 /* 2915 * Check that we have enough availrmem that memory locking (e.g., via 2916 * mlock(3C) or memcntl(2)) can still succeed. (pages_pp_maximum 2917 * stores the number of pages that cannot be locked; when availrmem 2918 * drops below pages_pp_maximum, page locking mechanisms such as 2919 * page_pp_lock() will fail.) 2920 */ 2921 if (availrmem <= pages_pp_maximum) 2922 return (1); 2923 2924#endif /* sun */ 2925#if defined(__i386) || !defined(UMA_MD_SMALL_ALLOC) 2926 /* 2927 * If we're on an i386 platform, it's possible that we'll exhaust the 2928 * kernel heap space before we ever run out of available physical 2929 * memory. Most checks of the size of the heap_area compare against 2930 * tune.t_minarmem, which is the minimum available real memory that we 2931 * can have in the system. However, this is generally fixed at 25 pages 2932 * which is so low that it's useless. In this comparison, we seek to 2933 * calculate the total heap-size, and reclaim if more than 3/4ths of the 2934 * heap is allocated. (Or, in the calculation, if less than 1/4th is 2935 * free) 2936 */ 2937 if (vmem_size(heap_arena, VMEM_FREE) < 2938 (vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC) >> 2)) { 2939 DTRACE_PROBE2(arc__reclaim_used, uint64_t, 2940 vmem_size(heap_arena, VMEM_FREE), uint64_t, 2941 (vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC)) >> 2); 2942 return (1); 2943 } 2944#define zio_arena NULL 2945#else 2946#define zio_arena heap_arena 2947#endif 2948 2949 /* 2950 * If zio data pages are being allocated out of a separate heap segment, 2951 * then enforce that the size of available vmem for this arena remains 2952 * above about 1/16th free. 2953 * 2954 * Note: The 1/16th arena free requirement was put in place 2955 * to aggressively evict memory from the arc in order to avoid 2956 * memory fragmentation issues. 2957 */ 2958 if (zio_arena != NULL && 2959 vmem_size(zio_arena, VMEM_FREE) < 2960 (vmem_size(zio_arena, VMEM_ALLOC) >> 4)) 2961 return (1); 2962 2963 /* 2964 * Above limits know nothing about real level of KVA fragmentation. 2965 * Start aggressive reclamation if too little sequential KVA left. 2966 */ 2967 if (vmem_size(heap_arena, VMEM_MAXFREE) < zfs_max_recordsize) { 2968 DTRACE_PROBE2(arc__reclaim_maxfree, uint64_t, 2969 vmem_size(heap_arena, VMEM_MAXFREE), 2970 uint64_t, zfs_max_recordsize); 2971 return (1); 2972 } 2973 2974#else /* _KERNEL */ 2975 if (spa_get_random(100) == 0) 2976 return (1); 2977#endif /* _KERNEL */ 2978 DTRACE_PROBE(arc__reclaim_no); 2979 2980 return (0); 2981} 2982 2983extern kmem_cache_t *zio_buf_cache[]; 2984extern kmem_cache_t *zio_data_buf_cache[]; 2985extern kmem_cache_t *range_seg_cache; 2986 2987static __noinline void 2988arc_kmem_reap_now(arc_reclaim_strategy_t strat) 2989{ 2990 size_t i; 2991 kmem_cache_t *prev_cache = NULL; 2992 kmem_cache_t *prev_data_cache = NULL; 2993 2994 DTRACE_PROBE(arc__kmem_reap_start); 2995#ifdef _KERNEL 2996 if (arc_meta_used >= arc_meta_limit) { 2997 /* 2998 * We are exceeding our meta-data cache limit. 2999 * Purge some DNLC entries to release holds on meta-data. 3000 */ 3001 dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent); 3002 } 3003#if defined(__i386) 3004 /* 3005 * Reclaim unused memory from all kmem caches. 3006 */ 3007 kmem_reap(); 3008#endif 3009#endif 3010 3011 /* 3012 * An aggressive reclamation will shrink the cache size as well as 3013 * reap free buffers from the arc kmem caches. 3014 */ 3015 if (strat == ARC_RECLAIM_AGGR) 3016 arc_shrink(); 3017 3018 for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) { 3019 if (zio_buf_cache[i] != prev_cache) { 3020 prev_cache = zio_buf_cache[i]; 3021 kmem_cache_reap_now(zio_buf_cache[i]); 3022 } 3023 if (zio_data_buf_cache[i] != prev_data_cache) { 3024 prev_data_cache = zio_data_buf_cache[i]; 3025 kmem_cache_reap_now(zio_data_buf_cache[i]); 3026 } 3027 } 3028 kmem_cache_reap_now(buf_cache); 3029 kmem_cache_reap_now(hdr_full_cache); 3030 kmem_cache_reap_now(hdr_l2only_cache); 3031 kmem_cache_reap_now(range_seg_cache); 3032 3033#ifdef sun 3034 /* 3035 * Ask the vmem arena to reclaim unused memory from its 3036 * quantum caches. 3037 */ 3038 if (zio_arena != NULL && strat == ARC_RECLAIM_AGGR) 3039 vmem_qcache_reap(zio_arena); 3040#endif 3041 DTRACE_PROBE(arc__kmem_reap_end); 3042} 3043 3044static void 3045arc_reclaim_thread(void *dummy __unused) 3046{ 3047 clock_t growtime = 0; 3048 arc_reclaim_strategy_t last_reclaim = ARC_RECLAIM_CONS; 3049 callb_cpr_t cpr; 3050 3051 CALLB_CPR_INIT(&cpr, &arc_reclaim_thr_lock, callb_generic_cpr, FTAG); 3052 3053 mutex_enter(&arc_reclaim_thr_lock); 3054 while (arc_thread_exit == 0) { 3055 if (arc_reclaim_needed()) { 3056 3057 if (arc_no_grow) { 3058 if (last_reclaim == ARC_RECLAIM_CONS) { 3059 DTRACE_PROBE(arc__reclaim_aggr_no_grow); 3060 last_reclaim = ARC_RECLAIM_AGGR; 3061 } else { 3062 last_reclaim = ARC_RECLAIM_CONS; 3063 } 3064 } else { 3065 arc_no_grow = TRUE; 3066 last_reclaim = ARC_RECLAIM_AGGR; 3067 DTRACE_PROBE(arc__reclaim_aggr); 3068 membar_producer(); 3069 } 3070 3071 /* reset the growth delay for every reclaim */ 3072 growtime = ddi_get_lbolt() + (arc_grow_retry * hz); 3073 3074 if (needfree && last_reclaim == ARC_RECLAIM_CONS) { 3075 /* 3076 * If needfree is TRUE our vm_lowmem hook 3077 * was called and in that case we must free some 3078 * memory, so switch to aggressive mode. 3079 */ 3080 arc_no_grow = TRUE; 3081 last_reclaim = ARC_RECLAIM_AGGR; 3082 } 3083 arc_kmem_reap_now(last_reclaim); 3084 arc_warm = B_TRUE; 3085 3086 } else if (arc_no_grow && ddi_get_lbolt() >= growtime) { 3087 arc_no_grow = FALSE; 3088 } 3089 3090 arc_adjust(); 3091 3092 if (arc_eviction_list != NULL) 3093 arc_do_user_evicts(); 3094 3095#ifdef _KERNEL 3096 if (needfree) { 3097 needfree = 0; 3098 wakeup(&needfree); 3099 } 3100#endif 3101 3102 /* block until needed, or one second, whichever is shorter */ 3103 CALLB_CPR_SAFE_BEGIN(&cpr); 3104 (void) cv_timedwait(&arc_reclaim_thr_cv, 3105 &arc_reclaim_thr_lock, hz); 3106 CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_thr_lock); 3107 } 3108 3109 arc_thread_exit = 0; 3110 cv_broadcast(&arc_reclaim_thr_cv); 3111 CALLB_CPR_EXIT(&cpr); /* drops arc_reclaim_thr_lock */ 3112 thread_exit(); 3113} 3114 3115/* 3116 * Adapt arc info given the number of bytes we are trying to add and 3117 * the state that we are comming from. This function is only called 3118 * when we are adding new content to the cache. 3119 */ 3120static void 3121arc_adapt(int bytes, arc_state_t *state) 3122{ 3123 int mult; 3124 uint64_t arc_p_min = (arc_c >> arc_p_min_shift); 3125 3126 if (state == arc_l2c_only) 3127 return; 3128 3129 ASSERT(bytes > 0); 3130 /* 3131 * Adapt the target size of the MRU list: 3132 * - if we just hit in the MRU ghost list, then increase 3133 * the target size of the MRU list. 3134 * - if we just hit in the MFU ghost list, then increase 3135 * the target size of the MFU list by decreasing the 3136 * target size of the MRU list. 3137 */ 3138 if (state == arc_mru_ghost) { 3139 mult = ((arc_mru_ghost->arcs_size >= arc_mfu_ghost->arcs_size) ? 3140 1 : (arc_mfu_ghost->arcs_size/arc_mru_ghost->arcs_size)); 3141 mult = MIN(mult, 10); /* avoid wild arc_p adjustment */ 3142 3143 arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult); 3144 } else if (state == arc_mfu_ghost) { 3145 uint64_t delta; 3146 3147 mult = ((arc_mfu_ghost->arcs_size >= arc_mru_ghost->arcs_size) ? 3148 1 : (arc_mru_ghost->arcs_size/arc_mfu_ghost->arcs_size)); 3149 mult = MIN(mult, 10); 3150 3151 delta = MIN(bytes * mult, arc_p); 3152 arc_p = MAX(arc_p_min, arc_p - delta); 3153 } 3154 ASSERT((int64_t)arc_p >= 0); 3155 3156 if (arc_reclaim_needed()) { 3157 cv_signal(&arc_reclaim_thr_cv); 3158 return; 3159 } 3160 3161 if (arc_no_grow) 3162 return; 3163 3164 if (arc_c >= arc_c_max) 3165 return; 3166 3167 /* 3168 * If we're within (2 * maxblocksize) bytes of the target 3169 * cache size, increment the target cache size 3170 */ 3171 if (arc_size > arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) { 3172 DTRACE_PROBE1(arc__inc_adapt, int, bytes); 3173 atomic_add_64(&arc_c, (int64_t)bytes); 3174 if (arc_c > arc_c_max) 3175 arc_c = arc_c_max; 3176 else if (state == arc_anon) 3177 atomic_add_64(&arc_p, (int64_t)bytes); 3178 if (arc_p > arc_c) 3179 arc_p = arc_c; 3180 } 3181 ASSERT((int64_t)arc_p >= 0); 3182} 3183 3184/* 3185 * Check if the cache has reached its limits and eviction is required 3186 * prior to insert. 3187 */ 3188static int 3189arc_evict_needed(arc_buf_contents_t type) 3190{ 3191 if (type == ARC_BUFC_METADATA && arc_meta_used >= arc_meta_limit) 3192 return (1); 3193 3194 if (arc_reclaim_needed()) 3195 return (1); 3196 3197 return (arc_size > arc_c); 3198} 3199 3200/* 3201 * The buffer, supplied as the first argument, needs a data block. 3202 * So, if we are at cache max, determine which cache should be victimized. 3203 * We have the following cases: 3204 * 3205 * 1. Insert for MRU, p > sizeof(arc_anon + arc_mru) -> 3206 * In this situation if we're out of space, but the resident size of the MFU is 3207 * under the limit, victimize the MFU cache to satisfy this insertion request. 3208 * 3209 * 2. Insert for MRU, p <= sizeof(arc_anon + arc_mru) -> 3210 * Here, we've used up all of the available space for the MRU, so we need to 3211 * evict from our own cache instead. Evict from the set of resident MRU 3212 * entries. 3213 * 3214 * 3. Insert for MFU (c - p) > sizeof(arc_mfu) -> 3215 * c minus p represents the MFU space in the cache, since p is the size of the 3216 * cache that is dedicated to the MRU. In this situation there's still space on 3217 * the MFU side, so the MRU side needs to be victimized. 3218 * 3219 * 4. Insert for MFU (c - p) < sizeof(arc_mfu) -> 3220 * MFU's resident set is consuming more space than it has been allotted. In 3221 * this situation, we must victimize our own cache, the MFU, for this insertion. 3222 */ 3223static void 3224arc_get_data_buf(arc_buf_t *buf) 3225{ 3226 arc_state_t *state = buf->b_hdr->b_l1hdr.b_state; 3227 uint64_t size = buf->b_hdr->b_size; 3228 arc_buf_contents_t type = arc_buf_type(buf->b_hdr); 3229 3230 arc_adapt(size, state); 3231 3232 /* 3233 * We have not yet reached cache maximum size, 3234 * just allocate a new buffer. 3235 */ 3236 if (!arc_evict_needed(type)) { 3237 if (type == ARC_BUFC_METADATA) { 3238 buf->b_data = zio_buf_alloc(size); 3239 arc_space_consume(size, ARC_SPACE_DATA); 3240 } else { 3241 ASSERT(type == ARC_BUFC_DATA); 3242 buf->b_data = zio_data_buf_alloc(size); 3243 ARCSTAT_INCR(arcstat_data_size, size); 3244 atomic_add_64(&arc_size, size); 3245 } 3246 goto out; 3247 } 3248 3249 /* 3250 * If we are prefetching from the mfu ghost list, this buffer 3251 * will end up on the mru list; so steal space from there. 3252 */ 3253 if (state == arc_mfu_ghost) 3254 state = HDR_PREFETCH(buf->b_hdr) ? arc_mru : arc_mfu; 3255 else if (state == arc_mru_ghost) 3256 state = arc_mru; 3257 3258 if (state == arc_mru || state == arc_anon) { 3259 uint64_t mru_used = arc_anon->arcs_size + arc_mru->arcs_size; 3260 state = (arc_mfu->arcs_lsize[type] >= size && 3261 arc_p > mru_used) ? arc_mfu : arc_mru; 3262 } else { 3263 /* MFU cases */ 3264 uint64_t mfu_space = arc_c - arc_p; 3265 state = (arc_mru->arcs_lsize[type] >= size && 3266 mfu_space > arc_mfu->arcs_size) ? arc_mru : arc_mfu; 3267 } 3268 if ((buf->b_data = arc_evict(state, 0, size, TRUE, type)) == NULL) { 3269 if (type == ARC_BUFC_METADATA) { 3270 buf->b_data = zio_buf_alloc(size); 3271 arc_space_consume(size, ARC_SPACE_DATA); 3272 } else { 3273 ASSERT(type == ARC_BUFC_DATA); 3274 buf->b_data = zio_data_buf_alloc(size); 3275 ARCSTAT_INCR(arcstat_data_size, size); 3276 atomic_add_64(&arc_size, size); 3277 } 3278 ARCSTAT_BUMP(arcstat_recycle_miss); 3279 } 3280 ASSERT(buf->b_data != NULL); 3281out: 3282 /* 3283 * Update the state size. Note that ghost states have a 3284 * "ghost size" and so don't need to be updated. 3285 */ 3286 if (!GHOST_STATE(buf->b_hdr->b_l1hdr.b_state)) { 3287 arc_buf_hdr_t *hdr = buf->b_hdr; 3288 3289 atomic_add_64(&hdr->b_l1hdr.b_state->arcs_size, size); 3290 if (list_link_active(&hdr->b_l1hdr.b_arc_node)) { 3291 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 3292 atomic_add_64(&hdr->b_l1hdr.b_state->arcs_lsize[type], 3293 size); 3294 } 3295 /* 3296 * If we are growing the cache, and we are adding anonymous 3297 * data, and we have outgrown arc_p, update arc_p 3298 */ 3299 if (arc_size < arc_c && hdr->b_l1hdr.b_state == arc_anon && 3300 arc_anon->arcs_size + arc_mru->arcs_size > arc_p) 3301 arc_p = MIN(arc_c, arc_p + size); 3302 } 3303 ARCSTAT_BUMP(arcstat_allocated); 3304} 3305 3306/* 3307 * This routine is called whenever a buffer is accessed. 3308 * NOTE: the hash lock is dropped in this function. 3309 */ 3310static void 3311arc_access(arc_buf_hdr_t *hdr, kmutex_t *hash_lock) 3312{ 3313 clock_t now; 3314 3315 ASSERT(MUTEX_HELD(hash_lock)); 3316 ASSERT(HDR_HAS_L1HDR(hdr)); 3317 3318 if (hdr->b_l1hdr.b_state == arc_anon) { 3319 /* 3320 * This buffer is not in the cache, and does not 3321 * appear in our "ghost" list. Add the new buffer 3322 * to the MRU state. 3323 */ 3324 3325 ASSERT0(hdr->b_l1hdr.b_arc_access); 3326 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 3327 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr); 3328 arc_change_state(arc_mru, hdr, hash_lock); 3329 3330 } else if (hdr->b_l1hdr.b_state == arc_mru) { 3331 now = ddi_get_lbolt(); 3332 3333 /* 3334 * If this buffer is here because of a prefetch, then either: 3335 * - clear the flag if this is a "referencing" read 3336 * (any subsequent access will bump this into the MFU state). 3337 * or 3338 * - move the buffer to the head of the list if this is 3339 * another prefetch (to make it less likely to be evicted). 3340 */ 3341 if (HDR_PREFETCH(hdr)) { 3342 if (refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) { 3343 ASSERT(list_link_active( 3344 &hdr->b_l1hdr.b_arc_node)); 3345 } else { 3346 hdr->b_flags &= ~ARC_FLAG_PREFETCH; 3347 ARCSTAT_BUMP(arcstat_mru_hits); 3348 } 3349 hdr->b_l1hdr.b_arc_access = now; 3350 return; 3351 } 3352 3353 /* 3354 * This buffer has been "accessed" only once so far, 3355 * but it is still in the cache. Move it to the MFU 3356 * state. 3357 */ 3358 if (now > hdr->b_l1hdr.b_arc_access + ARC_MINTIME) { 3359 /* 3360 * More than 125ms have passed since we 3361 * instantiated this buffer. Move it to the 3362 * most frequently used state. 3363 */ 3364 hdr->b_l1hdr.b_arc_access = now; 3365 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 3366 arc_change_state(arc_mfu, hdr, hash_lock); 3367 } 3368 ARCSTAT_BUMP(arcstat_mru_hits); 3369 } else if (hdr->b_l1hdr.b_state == arc_mru_ghost) { 3370 arc_state_t *new_state; 3371 /* 3372 * This buffer has been "accessed" recently, but 3373 * was evicted from the cache. Move it to the 3374 * MFU state. 3375 */ 3376 3377 if (HDR_PREFETCH(hdr)) { 3378 new_state = arc_mru; 3379 if (refcount_count(&hdr->b_l1hdr.b_refcnt) > 0) 3380 hdr->b_flags &= ~ARC_FLAG_PREFETCH; 3381 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr); 3382 } else { 3383 new_state = arc_mfu; 3384 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 3385 } 3386 3387 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 3388 arc_change_state(new_state, hdr, hash_lock); 3389 3390 ARCSTAT_BUMP(arcstat_mru_ghost_hits); 3391 } else if (hdr->b_l1hdr.b_state == arc_mfu) { 3392 /* 3393 * This buffer has been accessed more than once and is 3394 * still in the cache. Keep it in the MFU state. 3395 * 3396 * NOTE: an add_reference() that occurred when we did 3397 * the arc_read() will have kicked this off the list. 3398 * If it was a prefetch, we will explicitly move it to 3399 * the head of the list now. 3400 */ 3401 if ((HDR_PREFETCH(hdr)) != 0) { 3402 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 3403 ASSERT(list_link_active(&hdr->b_l1hdr.b_arc_node)); 3404 } 3405 ARCSTAT_BUMP(arcstat_mfu_hits); 3406 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 3407 } else if (hdr->b_l1hdr.b_state == arc_mfu_ghost) { 3408 arc_state_t *new_state = arc_mfu; 3409 /* 3410 * This buffer has been accessed more than once but has 3411 * been evicted from the cache. Move it back to the 3412 * MFU state. 3413 */ 3414 3415 if (HDR_PREFETCH(hdr)) { 3416 /* 3417 * This is a prefetch access... 3418 * move this block back to the MRU state. 3419 */ 3420 ASSERT0(refcount_count(&hdr->b_l1hdr.b_refcnt)); 3421 new_state = arc_mru; 3422 } 3423 3424 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 3425 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 3426 arc_change_state(new_state, hdr, hash_lock); 3427 3428 ARCSTAT_BUMP(arcstat_mfu_ghost_hits); 3429 } else if (hdr->b_l1hdr.b_state == arc_l2c_only) { 3430 /* 3431 * This buffer is on the 2nd Level ARC. 3432 */ 3433 3434 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 3435 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 3436 arc_change_state(arc_mfu, hdr, hash_lock); 3437 } else { 3438 ASSERT(!"invalid arc state"); 3439 } 3440} 3441 3442/* a generic arc_done_func_t which you can use */ 3443/* ARGSUSED */ 3444void 3445arc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg) 3446{ 3447 if (zio == NULL || zio->io_error == 0) 3448 bcopy(buf->b_data, arg, buf->b_hdr->b_size); 3449 VERIFY(arc_buf_remove_ref(buf, arg)); 3450} 3451 3452/* a generic arc_done_func_t */ 3453void 3454arc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg) 3455{ 3456 arc_buf_t **bufp = arg; 3457 if (zio && zio->io_error) { 3458 VERIFY(arc_buf_remove_ref(buf, arg)); 3459 *bufp = NULL; 3460 } else { 3461 *bufp = buf; 3462 ASSERT(buf->b_data); 3463 } 3464} 3465 3466static void 3467arc_read_done(zio_t *zio) 3468{ 3469 arc_buf_hdr_t *hdr; 3470 arc_buf_t *buf; 3471 arc_buf_t *abuf; /* buffer we're assigning to callback */ 3472 kmutex_t *hash_lock = NULL; 3473 arc_callback_t *callback_list, *acb; 3474 int freeable = FALSE; 3475 3476 buf = zio->io_private; 3477 hdr = buf->b_hdr; 3478 3479 /* 3480 * The hdr was inserted into hash-table and removed from lists 3481 * prior to starting I/O. We should find this header, since 3482 * it's in the hash table, and it should be legit since it's 3483 * not possible to evict it during the I/O. The only possible 3484 * reason for it not to be found is if we were freed during the 3485 * read. 3486 */ 3487 if (HDR_IN_HASH_TABLE(hdr)) { 3488 ASSERT3U(hdr->b_birth, ==, BP_PHYSICAL_BIRTH(zio->io_bp)); 3489 ASSERT3U(hdr->b_dva.dva_word[0], ==, 3490 BP_IDENTITY(zio->io_bp)->dva_word[0]); 3491 ASSERT3U(hdr->b_dva.dva_word[1], ==, 3492 BP_IDENTITY(zio->io_bp)->dva_word[1]); 3493 3494 arc_buf_hdr_t *found = buf_hash_find(hdr->b_spa, zio->io_bp, 3495 &hash_lock); 3496 3497 ASSERT((found == NULL && HDR_FREED_IN_READ(hdr) && 3498 hash_lock == NULL) || 3499 (found == hdr && 3500 DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) || 3501 (found == hdr && HDR_L2_READING(hdr))); 3502 } 3503 3504 hdr->b_flags &= ~ARC_FLAG_L2_EVICTED; 3505 if (l2arc_noprefetch && HDR_PREFETCH(hdr)) 3506 hdr->b_flags &= ~ARC_FLAG_L2CACHE; 3507 3508 /* byteswap if necessary */ 3509 callback_list = hdr->b_l1hdr.b_acb; 3510 ASSERT(callback_list != NULL); 3511 if (BP_SHOULD_BYTESWAP(zio->io_bp) && zio->io_error == 0) { 3512 dmu_object_byteswap_t bswap = 3513 DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp)); 3514 arc_byteswap_func_t *func = BP_GET_LEVEL(zio->io_bp) > 0 ? 3515 byteswap_uint64_array : 3516 dmu_ot_byteswap[bswap].ob_func; 3517 func(buf->b_data, hdr->b_size); 3518 } 3519 3520 arc_cksum_compute(buf, B_FALSE); 3521#ifdef illumos 3522 arc_buf_watch(buf); 3523#endif /* illumos */ 3524 3525 if (hash_lock && zio->io_error == 0 && 3526 hdr->b_l1hdr.b_state == arc_anon) { 3527 /* 3528 * Only call arc_access on anonymous buffers. This is because 3529 * if we've issued an I/O for an evicted buffer, we've already 3530 * called arc_access (to prevent any simultaneous readers from 3531 * getting confused). 3532 */ 3533 arc_access(hdr, hash_lock); 3534 } 3535 3536 /* create copies of the data buffer for the callers */ 3537 abuf = buf; 3538 for (acb = callback_list; acb; acb = acb->acb_next) { 3539 if (acb->acb_done) { 3540 if (abuf == NULL) { 3541 ARCSTAT_BUMP(arcstat_duplicate_reads); 3542 abuf = arc_buf_clone(buf); 3543 } 3544 acb->acb_buf = abuf; 3545 abuf = NULL; 3546 } 3547 } 3548 hdr->b_l1hdr.b_acb = NULL; 3549 hdr->b_flags &= ~ARC_FLAG_IO_IN_PROGRESS; 3550 ASSERT(!HDR_BUF_AVAILABLE(hdr)); 3551 if (abuf == buf) { 3552 ASSERT(buf->b_efunc == NULL); 3553 ASSERT(hdr->b_l1hdr.b_datacnt == 1); 3554 hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE; 3555 } 3556 3557 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt) || 3558 callback_list != NULL); 3559 3560 if (zio->io_error != 0) { 3561 hdr->b_flags |= ARC_FLAG_IO_ERROR; 3562 if (hdr->b_l1hdr.b_state != arc_anon) 3563 arc_change_state(arc_anon, hdr, hash_lock); 3564 if (HDR_IN_HASH_TABLE(hdr)) 3565 buf_hash_remove(hdr); 3566 freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt); 3567 } 3568 3569 /* 3570 * Broadcast before we drop the hash_lock to avoid the possibility 3571 * that the hdr (and hence the cv) might be freed before we get to 3572 * the cv_broadcast(). 3573 */ 3574 cv_broadcast(&hdr->b_l1hdr.b_cv); 3575 3576 if (hash_lock != NULL) { 3577 mutex_exit(hash_lock); 3578 } else { 3579 /* 3580 * This block was freed while we waited for the read to 3581 * complete. It has been removed from the hash table and 3582 * moved to the anonymous state (so that it won't show up 3583 * in the cache). 3584 */ 3585 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon); 3586 freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt); 3587 } 3588 3589 /* execute each callback and free its structure */ 3590 while ((acb = callback_list) != NULL) { 3591 if (acb->acb_done) 3592 acb->acb_done(zio, acb->acb_buf, acb->acb_private); 3593 3594 if (acb->acb_zio_dummy != NULL) { 3595 acb->acb_zio_dummy->io_error = zio->io_error; 3596 zio_nowait(acb->acb_zio_dummy); 3597 } 3598 3599 callback_list = acb->acb_next; 3600 kmem_free(acb, sizeof (arc_callback_t)); 3601 } 3602 3603 if (freeable) 3604 arc_hdr_destroy(hdr); 3605} 3606 3607/* 3608 * "Read" the block block at the specified DVA (in bp) via the 3609 * cache. If the block is found in the cache, invoke the provided 3610 * callback immediately and return. Note that the `zio' parameter 3611 * in the callback will be NULL in this case, since no IO was 3612 * required. If the block is not in the cache pass the read request 3613 * on to the spa with a substitute callback function, so that the 3614 * requested block will be added to the cache. 3615 * 3616 * If a read request arrives for a block that has a read in-progress, 3617 * either wait for the in-progress read to complete (and return the 3618 * results); or, if this is a read with a "done" func, add a record 3619 * to the read to invoke the "done" func when the read completes, 3620 * and return; or just return. 3621 * 3622 * arc_read_done() will invoke all the requested "done" functions 3623 * for readers of this block. 3624 */ 3625int 3626arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_done_func_t *done, 3627 void *private, zio_priority_t priority, int zio_flags, 3628 arc_flags_t *arc_flags, const zbookmark_phys_t *zb) 3629{ 3630 arc_buf_hdr_t *hdr = NULL; 3631 arc_buf_t *buf = NULL; 3632 kmutex_t *hash_lock = NULL; 3633 zio_t *rzio; 3634 uint64_t guid = spa_load_guid(spa); 3635 3636 ASSERT(!BP_IS_EMBEDDED(bp) || 3637 BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA); 3638 3639top: 3640 if (!BP_IS_EMBEDDED(bp)) { 3641 /* 3642 * Embedded BP's have no DVA and require no I/O to "read". 3643 * Create an anonymous arc buf to back it. 3644 */ 3645 hdr = buf_hash_find(guid, bp, &hash_lock); 3646 } 3647 3648 if (hdr != NULL && HDR_HAS_L1HDR(hdr) && hdr->b_l1hdr.b_datacnt > 0) { 3649 3650 *arc_flags |= ARC_FLAG_CACHED; 3651 3652 if (HDR_IO_IN_PROGRESS(hdr)) { 3653 3654 if (*arc_flags & ARC_FLAG_WAIT) { 3655 cv_wait(&hdr->b_l1hdr.b_cv, hash_lock); 3656 mutex_exit(hash_lock); 3657 goto top; 3658 } 3659 ASSERT(*arc_flags & ARC_FLAG_NOWAIT); 3660 3661 if (done) { 3662 arc_callback_t *acb = NULL; 3663 3664 acb = kmem_zalloc(sizeof (arc_callback_t), 3665 KM_SLEEP); 3666 acb->acb_done = done; 3667 acb->acb_private = private; 3668 if (pio != NULL) 3669 acb->acb_zio_dummy = zio_null(pio, 3670 spa, NULL, NULL, NULL, zio_flags); 3671 3672 ASSERT(acb->acb_done != NULL); 3673 acb->acb_next = hdr->b_l1hdr.b_acb; 3674 hdr->b_l1hdr.b_acb = acb; 3675 add_reference(hdr, hash_lock, private); 3676 mutex_exit(hash_lock); 3677 return (0); 3678 } 3679 mutex_exit(hash_lock); 3680 return (0); 3681 } 3682 3683 ASSERT(hdr->b_l1hdr.b_state == arc_mru || 3684 hdr->b_l1hdr.b_state == arc_mfu); 3685 3686 if (done) { 3687 add_reference(hdr, hash_lock, private); 3688 /* 3689 * If this block is already in use, create a new 3690 * copy of the data so that we will be guaranteed 3691 * that arc_release() will always succeed. 3692 */ 3693 buf = hdr->b_l1hdr.b_buf; 3694 ASSERT(buf); 3695 ASSERT(buf->b_data); 3696 if (HDR_BUF_AVAILABLE(hdr)) { 3697 ASSERT(buf->b_efunc == NULL); 3698 hdr->b_flags &= ~ARC_FLAG_BUF_AVAILABLE; 3699 } else { 3700 buf = arc_buf_clone(buf); 3701 } 3702 3703 } else if (*arc_flags & ARC_FLAG_PREFETCH && 3704 refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) { 3705 hdr->b_flags |= ARC_FLAG_PREFETCH; 3706 } 3707 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr); 3708 arc_access(hdr, hash_lock); 3709 if (*arc_flags & ARC_FLAG_L2CACHE) 3710 hdr->b_flags |= ARC_FLAG_L2CACHE; 3711 if (*arc_flags & ARC_FLAG_L2COMPRESS) 3712 hdr->b_flags |= ARC_FLAG_L2COMPRESS; 3713 mutex_exit(hash_lock); 3714 ARCSTAT_BUMP(arcstat_hits); 3715 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr), 3716 demand, prefetch, !HDR_ISTYPE_METADATA(hdr), 3717 data, metadata, hits); 3718 3719 if (done) 3720 done(NULL, buf, private); 3721 } else { 3722 uint64_t size = BP_GET_LSIZE(bp); 3723 arc_callback_t *acb; 3724 vdev_t *vd = NULL; 3725 uint64_t addr = 0; 3726 boolean_t devw = B_FALSE; 3727 enum zio_compress b_compress = ZIO_COMPRESS_OFF; 3728 int32_t b_asize = 0; 3729 3730 if (hdr == NULL) { 3731 /* this block is not in the cache */ 3732 arc_buf_hdr_t *exists = NULL; 3733 arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp); 3734 buf = arc_buf_alloc(spa, size, private, type); 3735 hdr = buf->b_hdr; 3736 if (!BP_IS_EMBEDDED(bp)) { 3737 hdr->b_dva = *BP_IDENTITY(bp); 3738 hdr->b_birth = BP_PHYSICAL_BIRTH(bp); 3739 exists = buf_hash_insert(hdr, &hash_lock); 3740 } 3741 if (exists != NULL) { 3742 /* somebody beat us to the hash insert */ 3743 mutex_exit(hash_lock); 3744 buf_discard_identity(hdr); 3745 (void) arc_buf_remove_ref(buf, private); 3746 goto top; /* restart the IO request */ 3747 } 3748 3749 /* if this is a prefetch, we don't have a reference */ 3750 if (*arc_flags & ARC_FLAG_PREFETCH) { 3751 (void) remove_reference(hdr, hash_lock, 3752 private); 3753 hdr->b_flags |= ARC_FLAG_PREFETCH; 3754 } 3755 if (*arc_flags & ARC_FLAG_L2CACHE) 3756 hdr->b_flags |= ARC_FLAG_L2CACHE; 3757 if (*arc_flags & ARC_FLAG_L2COMPRESS) 3758 hdr->b_flags |= ARC_FLAG_L2COMPRESS; 3759 if (BP_GET_LEVEL(bp) > 0) 3760 hdr->b_flags |= ARC_FLAG_INDIRECT; 3761 } else { 3762 /* 3763 * This block is in the ghost cache. If it was L2-only 3764 * (and thus didn't have an L1 hdr), we realloc the 3765 * header to add an L1 hdr. 3766 */ 3767 if (!HDR_HAS_L1HDR(hdr)) { 3768 hdr = arc_hdr_realloc(hdr, hdr_l2only_cache, 3769 hdr_full_cache); 3770 } 3771 3772 ASSERT(GHOST_STATE(hdr->b_l1hdr.b_state)); 3773 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 3774 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 3775 ASSERT(hdr->b_l1hdr.b_buf == NULL); 3776 3777 /* if this is a prefetch, we don't have a reference */ 3778 if (*arc_flags & ARC_FLAG_PREFETCH) 3779 hdr->b_flags |= ARC_FLAG_PREFETCH; 3780 else 3781 add_reference(hdr, hash_lock, private); 3782 if (*arc_flags & ARC_FLAG_L2CACHE) 3783 hdr->b_flags |= ARC_FLAG_L2CACHE; 3784 if (*arc_flags & ARC_FLAG_L2COMPRESS) 3785 hdr->b_flags |= ARC_FLAG_L2COMPRESS; 3786 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 3787 buf->b_hdr = hdr; 3788 buf->b_data = NULL; 3789 buf->b_efunc = NULL; 3790 buf->b_private = NULL; 3791 buf->b_next = NULL; 3792 hdr->b_l1hdr.b_buf = buf; 3793 ASSERT0(hdr->b_l1hdr.b_datacnt); 3794 hdr->b_l1hdr.b_datacnt = 1; 3795 arc_get_data_buf(buf); 3796 arc_access(hdr, hash_lock); 3797 } 3798 3799 ASSERT(!GHOST_STATE(hdr->b_l1hdr.b_state)); 3800 3801 acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP); 3802 acb->acb_done = done; 3803 acb->acb_private = private; 3804 3805 ASSERT(hdr->b_l1hdr.b_acb == NULL); 3806 hdr->b_l1hdr.b_acb = acb; 3807 hdr->b_flags |= ARC_FLAG_IO_IN_PROGRESS; 3808 3809 if (HDR_HAS_L2HDR(hdr) && 3810 (vd = hdr->b_l2hdr.b_dev->l2ad_vdev) != NULL) { 3811 devw = hdr->b_l2hdr.b_dev->l2ad_writing; 3812 addr = hdr->b_l2hdr.b_daddr; 3813 b_compress = HDR_GET_COMPRESS(hdr); 3814 b_asize = hdr->b_l2hdr.b_asize; 3815 /* 3816 * Lock out device removal. 3817 */ 3818 if (vdev_is_dead(vd) || 3819 !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER)) 3820 vd = NULL; 3821 } 3822 3823 if (hash_lock != NULL) 3824 mutex_exit(hash_lock); 3825 3826 /* 3827 * At this point, we have a level 1 cache miss. Try again in 3828 * L2ARC if possible. 3829 */ 3830 ASSERT3U(hdr->b_size, ==, size); 3831 DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp, 3832 uint64_t, size, zbookmark_phys_t *, zb); 3833 ARCSTAT_BUMP(arcstat_misses); 3834 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr), 3835 demand, prefetch, !HDR_ISTYPE_METADATA(hdr), 3836 data, metadata, misses); 3837#ifdef _KERNEL 3838 curthread->td_ru.ru_inblock++; 3839#endif 3840 3841 if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) { 3842 /* 3843 * Read from the L2ARC if the following are true: 3844 * 1. The L2ARC vdev was previously cached. 3845 * 2. This buffer still has L2ARC metadata. 3846 * 3. This buffer isn't currently writing to the L2ARC. 3847 * 4. The L2ARC entry wasn't evicted, which may 3848 * also have invalidated the vdev. 3849 * 5. This isn't prefetch and l2arc_noprefetch is set. 3850 */ 3851 if (HDR_HAS_L2HDR(hdr) && 3852 !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) && 3853 !(l2arc_noprefetch && HDR_PREFETCH(hdr))) { 3854 l2arc_read_callback_t *cb; 3855 3856 DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr); 3857 ARCSTAT_BUMP(arcstat_l2_hits); 3858 3859 cb = kmem_zalloc(sizeof (l2arc_read_callback_t), 3860 KM_SLEEP); 3861 cb->l2rcb_buf = buf; 3862 cb->l2rcb_spa = spa; 3863 cb->l2rcb_bp = *bp; 3864 cb->l2rcb_zb = *zb; 3865 cb->l2rcb_flags = zio_flags; 3866 cb->l2rcb_compress = b_compress; 3867 3868 ASSERT(addr >= VDEV_LABEL_START_SIZE && 3869 addr + size < vd->vdev_psize - 3870 VDEV_LABEL_END_SIZE); 3871 3872 /* 3873 * l2arc read. The SCL_L2ARC lock will be 3874 * released by l2arc_read_done(). 3875 * Issue a null zio if the underlying buffer 3876 * was squashed to zero size by compression. 3877 */ 3878 if (b_compress == ZIO_COMPRESS_EMPTY) { 3879 rzio = zio_null(pio, spa, vd, 3880 l2arc_read_done, cb, 3881 zio_flags | ZIO_FLAG_DONT_CACHE | 3882 ZIO_FLAG_CANFAIL | 3883 ZIO_FLAG_DONT_PROPAGATE | 3884 ZIO_FLAG_DONT_RETRY); 3885 } else { 3886 rzio = zio_read_phys(pio, vd, addr, 3887 b_asize, buf->b_data, 3888 ZIO_CHECKSUM_OFF, 3889 l2arc_read_done, cb, priority, 3890 zio_flags | ZIO_FLAG_DONT_CACHE | 3891 ZIO_FLAG_CANFAIL | 3892 ZIO_FLAG_DONT_PROPAGATE | 3893 ZIO_FLAG_DONT_RETRY, B_FALSE); 3894 } 3895 DTRACE_PROBE2(l2arc__read, vdev_t *, vd, 3896 zio_t *, rzio); 3897 ARCSTAT_INCR(arcstat_l2_read_bytes, b_asize); 3898 3899 if (*arc_flags & ARC_FLAG_NOWAIT) { 3900 zio_nowait(rzio); 3901 return (0); 3902 } 3903 3904 ASSERT(*arc_flags & ARC_FLAG_WAIT); 3905 if (zio_wait(rzio) == 0) 3906 return (0); 3907 3908 /* l2arc read error; goto zio_read() */ 3909 } else { 3910 DTRACE_PROBE1(l2arc__miss, 3911 arc_buf_hdr_t *, hdr); 3912 ARCSTAT_BUMP(arcstat_l2_misses); 3913 if (HDR_L2_WRITING(hdr)) 3914 ARCSTAT_BUMP(arcstat_l2_rw_clash); 3915 spa_config_exit(spa, SCL_L2ARC, vd); 3916 } 3917 } else { 3918 if (vd != NULL) 3919 spa_config_exit(spa, SCL_L2ARC, vd); 3920 if (l2arc_ndev != 0) { 3921 DTRACE_PROBE1(l2arc__miss, 3922 arc_buf_hdr_t *, hdr); 3923 ARCSTAT_BUMP(arcstat_l2_misses); 3924 } 3925 } 3926 3927 rzio = zio_read(pio, spa, bp, buf->b_data, size, 3928 arc_read_done, buf, priority, zio_flags, zb); 3929 3930 if (*arc_flags & ARC_FLAG_WAIT) 3931 return (zio_wait(rzio)); 3932 3933 ASSERT(*arc_flags & ARC_FLAG_NOWAIT); 3934 zio_nowait(rzio); 3935 } 3936 return (0); 3937} 3938 3939void 3940arc_set_callback(arc_buf_t *buf, arc_evict_func_t *func, void *private) 3941{ 3942 ASSERT(buf->b_hdr != NULL); 3943 ASSERT(buf->b_hdr->b_l1hdr.b_state != arc_anon); 3944 ASSERT(!refcount_is_zero(&buf->b_hdr->b_l1hdr.b_refcnt) || 3945 func == NULL); 3946 ASSERT(buf->b_efunc == NULL); 3947 ASSERT(!HDR_BUF_AVAILABLE(buf->b_hdr)); 3948 3949 buf->b_efunc = func; 3950 buf->b_private = private; 3951} 3952 3953/* 3954 * Notify the arc that a block was freed, and thus will never be used again. 3955 */ 3956void 3957arc_freed(spa_t *spa, const blkptr_t *bp) 3958{ 3959 arc_buf_hdr_t *hdr; 3960 kmutex_t *hash_lock; 3961 uint64_t guid = spa_load_guid(spa); 3962 3963 ASSERT(!BP_IS_EMBEDDED(bp)); 3964 3965 hdr = buf_hash_find(guid, bp, &hash_lock); 3966 if (hdr == NULL) 3967 return; 3968 if (HDR_BUF_AVAILABLE(hdr)) { 3969 arc_buf_t *buf = hdr->b_l1hdr.b_buf; 3970 add_reference(hdr, hash_lock, FTAG); 3971 hdr->b_flags &= ~ARC_FLAG_BUF_AVAILABLE; 3972 mutex_exit(hash_lock); 3973 3974 arc_release(buf, FTAG); 3975 (void) arc_buf_remove_ref(buf, FTAG); 3976 } else { 3977 mutex_exit(hash_lock); 3978 } 3979 3980} 3981 3982/* 3983 * Clear the user eviction callback set by arc_set_callback(), first calling 3984 * it if it exists. Because the presence of a callback keeps an arc_buf cached 3985 * clearing the callback may result in the arc_buf being destroyed. However, 3986 * it will not result in the *last* arc_buf being destroyed, hence the data 3987 * will remain cached in the ARC. We make a copy of the arc buffer here so 3988 * that we can process the callback without holding any locks. 3989 * 3990 * It's possible that the callback is already in the process of being cleared 3991 * by another thread. In this case we can not clear the callback. 3992 * 3993 * Returns B_TRUE if the callback was successfully called and cleared. 3994 */ 3995boolean_t 3996arc_clear_callback(arc_buf_t *buf) 3997{ 3998 arc_buf_hdr_t *hdr; 3999 kmutex_t *hash_lock; 4000 arc_evict_func_t *efunc = buf->b_efunc; 4001 void *private = buf->b_private; 4002 list_t *list, *evicted_list; 4003 kmutex_t *lock, *evicted_lock; 4004 4005 mutex_enter(&buf->b_evict_lock); 4006 hdr = buf->b_hdr; 4007 if (hdr == NULL) { 4008 /* 4009 * We are in arc_do_user_evicts(). 4010 */ 4011 ASSERT(buf->b_data == NULL); 4012 mutex_exit(&buf->b_evict_lock); 4013 return (B_FALSE); 4014 } else if (buf->b_data == NULL) { 4015 /* 4016 * We are on the eviction list; process this buffer now 4017 * but let arc_do_user_evicts() do the reaping. 4018 */ 4019 buf->b_efunc = NULL; 4020 mutex_exit(&buf->b_evict_lock); 4021 VERIFY0(efunc(private)); 4022 return (B_TRUE); 4023 } 4024 hash_lock = HDR_LOCK(hdr); 4025 mutex_enter(hash_lock); 4026 hdr = buf->b_hdr; 4027 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 4028 4029 ASSERT3U(refcount_count(&hdr->b_l1hdr.b_refcnt), <, 4030 hdr->b_l1hdr.b_datacnt); 4031 ASSERT(hdr->b_l1hdr.b_state == arc_mru || 4032 hdr->b_l1hdr.b_state == arc_mfu); 4033 4034 buf->b_efunc = NULL; 4035 buf->b_private = NULL; 4036 4037 if (hdr->b_l1hdr.b_datacnt > 1) { 4038 mutex_exit(&buf->b_evict_lock); 4039 arc_buf_destroy(buf, FALSE, TRUE); 4040 } else { 4041 ASSERT(buf == hdr->b_l1hdr.b_buf); 4042 hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE; 4043 mutex_exit(&buf->b_evict_lock); 4044 } 4045 4046 mutex_exit(hash_lock); 4047 VERIFY0(efunc(private)); 4048 return (B_TRUE); 4049} 4050 4051/* 4052 * Release this buffer from the cache, making it an anonymous buffer. This 4053 * must be done after a read and prior to modifying the buffer contents. 4054 * If the buffer has more than one reference, we must make 4055 * a new hdr for the buffer. 4056 */ 4057void 4058arc_release(arc_buf_t *buf, void *tag) 4059{ 4060 arc_buf_hdr_t *hdr = buf->b_hdr; 4061 4062 /* 4063 * It would be nice to assert that if it's DMU metadata (level > 4064 * 0 || it's the dnode file), then it must be syncing context. 4065 * But we don't know that information at this level. 4066 */ 4067 4068 mutex_enter(&buf->b_evict_lock); 4069 /* 4070 * We don't grab the hash lock prior to this check, because if 4071 * the buffer's header is in the arc_anon state, it won't be 4072 * linked into the hash table. 4073 */ 4074 if (hdr->b_l1hdr.b_state == arc_anon) { 4075 mutex_exit(&buf->b_evict_lock); 4076 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 4077 ASSERT(!HDR_IN_HASH_TABLE(hdr)); 4078 ASSERT(!HDR_HAS_L2HDR(hdr)); 4079 ASSERT(BUF_EMPTY(hdr)); 4080 ASSERT3U(hdr->b_l1hdr.b_datacnt, ==, 1); 4081 ASSERT3S(refcount_count(&hdr->b_l1hdr.b_refcnt), ==, 1); 4082 ASSERT(!list_link_active(&hdr->b_l1hdr.b_arc_node)); 4083 4084 ASSERT3P(buf->b_efunc, ==, NULL); 4085 ASSERT3P(buf->b_private, ==, NULL); 4086 4087 hdr->b_l1hdr.b_arc_access = 0; 4088 arc_buf_thaw(buf); 4089 4090 return; 4091 } 4092 4093 kmutex_t *hash_lock = HDR_LOCK(hdr); 4094 mutex_enter(hash_lock); 4095 4096 /* 4097 * This assignment is only valid as long as the hash_lock is 4098 * held, we must be careful not to reference state or the 4099 * b_state field after dropping the lock. 4100 */ 4101 arc_state_t *state = hdr->b_l1hdr.b_state; 4102 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 4103 ASSERT3P(state, !=, arc_anon); 4104 4105 /* this buffer is not on any list */ 4106 ASSERT(refcount_count(&hdr->b_l1hdr.b_refcnt) > 0); 4107 4108 if (HDR_HAS_L2HDR(hdr)) { 4109 ARCSTAT_INCR(arcstat_l2_asize, -hdr->b_l2hdr.b_asize); 4110 ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size); 4111 4112 mutex_enter(&hdr->b_l2hdr.b_dev->l2ad_mtx); 4113 trim_map_free(hdr->b_l2hdr.b_dev->l2ad_vdev, 4114 hdr->b_l2hdr.b_daddr, hdr->b_l2hdr.b_asize, 0); 4115 list_remove(&hdr->b_l2hdr.b_dev->l2ad_buflist, hdr); 4116 4117 /* 4118 * We don't want to leak the b_tmp_cdata buffer that was 4119 * allocated in l2arc_write_buffers() 4120 */ 4121 arc_buf_l2_cdata_free(hdr); 4122 4123 mutex_exit(&hdr->b_l2hdr.b_dev->l2ad_mtx); 4124 4125 hdr->b_flags &= ~ARC_FLAG_HAS_L2HDR; 4126 } 4127 4128 /* 4129 * Do we have more than one buf? 4130 */ 4131 if (hdr->b_l1hdr.b_datacnt > 1) { 4132 arc_buf_hdr_t *nhdr; 4133 arc_buf_t **bufp; 4134 uint64_t blksz = hdr->b_size; 4135 uint64_t spa = hdr->b_spa; 4136 arc_buf_contents_t type = arc_buf_type(hdr); 4137 uint32_t flags = hdr->b_flags; 4138 4139 ASSERT(hdr->b_l1hdr.b_buf != buf || buf->b_next != NULL); 4140 /* 4141 * Pull the data off of this hdr and attach it to 4142 * a new anonymous hdr. 4143 */ 4144 (void) remove_reference(hdr, hash_lock, tag); 4145 bufp = &hdr->b_l1hdr.b_buf; 4146 while (*bufp != buf) 4147 bufp = &(*bufp)->b_next; 4148 *bufp = buf->b_next; 4149 buf->b_next = NULL; 4150 4151 ASSERT3P(state, !=, arc_l2c_only); 4152 ASSERT3U(state->arcs_size, >=, hdr->b_size); 4153 atomic_add_64(&state->arcs_size, -hdr->b_size); 4154 if (refcount_is_zero(&hdr->b_l1hdr.b_refcnt)) { 4155 ASSERT3P(state, !=, arc_l2c_only); 4156 uint64_t *size = &state->arcs_lsize[type]; 4157 ASSERT3U(*size, >=, hdr->b_size); 4158 atomic_add_64(size, -hdr->b_size); 4159 } 4160 4161 /* 4162 * We're releasing a duplicate user data buffer, update 4163 * our statistics accordingly. 4164 */ 4165 if (HDR_ISTYPE_DATA(hdr)) { 4166 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers); 4167 ARCSTAT_INCR(arcstat_duplicate_buffers_size, 4168 -hdr->b_size); 4169 } 4170 hdr->b_l1hdr.b_datacnt -= 1; 4171 arc_cksum_verify(buf); 4172#ifdef illumos 4173 arc_buf_unwatch(buf); 4174#endif /* illumos */ 4175 4176 mutex_exit(hash_lock); 4177 4178 nhdr = kmem_cache_alloc(hdr_full_cache, KM_PUSHPAGE); 4179 nhdr->b_size = blksz; 4180 nhdr->b_spa = spa; 4181 4182 nhdr->b_flags = flags & ARC_FLAG_L2_WRITING; 4183 nhdr->b_flags |= arc_bufc_to_flags(type); 4184 nhdr->b_flags |= ARC_FLAG_HAS_L1HDR; 4185 4186 nhdr->b_l1hdr.b_buf = buf; 4187 nhdr->b_l1hdr.b_datacnt = 1; 4188 nhdr->b_l1hdr.b_state = arc_anon; 4189 nhdr->b_l1hdr.b_arc_access = 0; 4190 nhdr->b_freeze_cksum = NULL; 4191 4192 (void) refcount_add(&nhdr->b_l1hdr.b_refcnt, tag); 4193 buf->b_hdr = nhdr; 4194 mutex_exit(&buf->b_evict_lock); 4195 atomic_add_64(&arc_anon->arcs_size, blksz); 4196 } else { 4197 mutex_exit(&buf->b_evict_lock); 4198 ASSERT(refcount_count(&hdr->b_l1hdr.b_refcnt) == 1); 4199 /* protected by hash lock */ 4200 ASSERT(!list_link_active(&hdr->b_l1hdr.b_arc_node)); 4201 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 4202 arc_change_state(arc_anon, hdr, hash_lock); 4203 hdr->b_l1hdr.b_arc_access = 0; 4204 mutex_exit(hash_lock); 4205 4206 buf_discard_identity(hdr); 4207 arc_buf_thaw(buf); 4208 } 4209 buf->b_efunc = NULL; 4210 buf->b_private = NULL; 4211} 4212 4213int 4214arc_released(arc_buf_t *buf) 4215{ 4216 int released; 4217 4218 mutex_enter(&buf->b_evict_lock); 4219 released = (buf->b_data != NULL && 4220 buf->b_hdr->b_l1hdr.b_state == arc_anon); 4221 mutex_exit(&buf->b_evict_lock); 4222 return (released); 4223} 4224 4225#ifdef ZFS_DEBUG 4226int 4227arc_referenced(arc_buf_t *buf) 4228{ 4229 int referenced; 4230 4231 mutex_enter(&buf->b_evict_lock); 4232 referenced = (refcount_count(&buf->b_hdr->b_l1hdr.b_refcnt)); 4233 mutex_exit(&buf->b_evict_lock); 4234 return (referenced); 4235} 4236#endif 4237 4238static void 4239arc_write_ready(zio_t *zio) 4240{ 4241 arc_write_callback_t *callback = zio->io_private; 4242 arc_buf_t *buf = callback->awcb_buf; 4243 arc_buf_hdr_t *hdr = buf->b_hdr; 4244 4245 ASSERT(HDR_HAS_L1HDR(hdr)); 4246 ASSERT(!refcount_is_zero(&buf->b_hdr->b_l1hdr.b_refcnt)); 4247 ASSERT(hdr->b_l1hdr.b_datacnt > 0); 4248 callback->awcb_ready(zio, buf, callback->awcb_private); 4249 4250 /* 4251 * If the IO is already in progress, then this is a re-write 4252 * attempt, so we need to thaw and re-compute the cksum. 4253 * It is the responsibility of the callback to handle the 4254 * accounting for any re-write attempt. 4255 */ 4256 if (HDR_IO_IN_PROGRESS(hdr)) { 4257 mutex_enter(&hdr->b_l1hdr.b_freeze_lock); 4258 if (hdr->b_freeze_cksum != NULL) { 4259 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 4260 hdr->b_freeze_cksum = NULL; 4261 } 4262 mutex_exit(&hdr->b_l1hdr.b_freeze_lock); 4263 } 4264 arc_cksum_compute(buf, B_FALSE); 4265 hdr->b_flags |= ARC_FLAG_IO_IN_PROGRESS; 4266} 4267 4268/* 4269 * The SPA calls this callback for each physical write that happens on behalf 4270 * of a logical write. See the comment in dbuf_write_physdone() for details. 4271 */ 4272static void 4273arc_write_physdone(zio_t *zio) 4274{ 4275 arc_write_callback_t *cb = zio->io_private; 4276 if (cb->awcb_physdone != NULL) 4277 cb->awcb_physdone(zio, cb->awcb_buf, cb->awcb_private); 4278} 4279 4280static void 4281arc_write_done(zio_t *zio) 4282{ 4283 arc_write_callback_t *callback = zio->io_private; 4284 arc_buf_t *buf = callback->awcb_buf; 4285 arc_buf_hdr_t *hdr = buf->b_hdr; 4286 4287 ASSERT(hdr->b_l1hdr.b_acb == NULL); 4288 4289 if (zio->io_error == 0) { 4290 if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) { 4291 buf_discard_identity(hdr); 4292 } else { 4293 hdr->b_dva = *BP_IDENTITY(zio->io_bp); 4294 hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp); 4295 } 4296 } else { 4297 ASSERT(BUF_EMPTY(hdr)); 4298 } 4299 4300 /* 4301 * If the block to be written was all-zero or compressed enough to be 4302 * embedded in the BP, no write was performed so there will be no 4303 * dva/birth/checksum. The buffer must therefore remain anonymous 4304 * (and uncached). 4305 */ 4306 if (!BUF_EMPTY(hdr)) { 4307 arc_buf_hdr_t *exists; 4308 kmutex_t *hash_lock; 4309 4310 ASSERT(zio->io_error == 0); 4311 4312 arc_cksum_verify(buf); 4313 4314 exists = buf_hash_insert(hdr, &hash_lock); 4315 if (exists != NULL) { 4316 /* 4317 * This can only happen if we overwrite for 4318 * sync-to-convergence, because we remove 4319 * buffers from the hash table when we arc_free(). 4320 */ 4321 if (zio->io_flags & ZIO_FLAG_IO_REWRITE) { 4322 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) 4323 panic("bad overwrite, hdr=%p exists=%p", 4324 (void *)hdr, (void *)exists); 4325 ASSERT(refcount_is_zero( 4326 &exists->b_l1hdr.b_refcnt)); 4327 arc_change_state(arc_anon, exists, hash_lock); 4328 mutex_exit(hash_lock); 4329 arc_hdr_destroy(exists); 4330 exists = buf_hash_insert(hdr, &hash_lock); 4331 ASSERT3P(exists, ==, NULL); 4332 } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) { 4333 /* nopwrite */ 4334 ASSERT(zio->io_prop.zp_nopwrite); 4335 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) 4336 panic("bad nopwrite, hdr=%p exists=%p", 4337 (void *)hdr, (void *)exists); 4338 } else { 4339 /* Dedup */ 4340 ASSERT(hdr->b_l1hdr.b_datacnt == 1); 4341 ASSERT(hdr->b_l1hdr.b_state == arc_anon); 4342 ASSERT(BP_GET_DEDUP(zio->io_bp)); 4343 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0); 4344 } 4345 } 4346 hdr->b_flags &= ~ARC_FLAG_IO_IN_PROGRESS; 4347 /* if it's not anon, we are doing a scrub */ 4348 if (exists == NULL && hdr->b_l1hdr.b_state == arc_anon) 4349 arc_access(hdr, hash_lock); 4350 mutex_exit(hash_lock); 4351 } else { 4352 hdr->b_flags &= ~ARC_FLAG_IO_IN_PROGRESS; 4353 } 4354 4355 ASSERT(!refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 4356 callback->awcb_done(zio, buf, callback->awcb_private); 4357 4358 kmem_free(callback, sizeof (arc_write_callback_t)); 4359} 4360 4361zio_t * 4362arc_write(zio_t *pio, spa_t *spa, uint64_t txg, 4363 blkptr_t *bp, arc_buf_t *buf, boolean_t l2arc, boolean_t l2arc_compress, 4364 const zio_prop_t *zp, arc_done_func_t *ready, arc_done_func_t *physdone, 4365 arc_done_func_t *done, void *private, zio_priority_t priority, 4366 int zio_flags, const zbookmark_phys_t *zb) 4367{ 4368 arc_buf_hdr_t *hdr = buf->b_hdr; 4369 arc_write_callback_t *callback; 4370 zio_t *zio; 4371 4372 ASSERT(ready != NULL); 4373 ASSERT(done != NULL); 4374 ASSERT(!HDR_IO_ERROR(hdr)); 4375 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 4376 ASSERT(hdr->b_l1hdr.b_acb == NULL); 4377 ASSERT(hdr->b_l1hdr.b_datacnt > 0); 4378 if (l2arc) 4379 hdr->b_flags |= ARC_FLAG_L2CACHE; 4380 if (l2arc_compress) 4381 hdr->b_flags |= ARC_FLAG_L2COMPRESS; 4382 callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP); 4383 callback->awcb_ready = ready; 4384 callback->awcb_physdone = physdone; 4385 callback->awcb_done = done; 4386 callback->awcb_private = private; 4387 callback->awcb_buf = buf; 4388 4389 zio = zio_write(pio, spa, txg, bp, buf->b_data, hdr->b_size, zp, 4390 arc_write_ready, arc_write_physdone, arc_write_done, callback, 4391 priority, zio_flags, zb); 4392 4393 return (zio); 4394} 4395 4396static int 4397arc_memory_throttle(uint64_t reserve, uint64_t txg) 4398{ 4399#ifdef _KERNEL 4400 uint64_t available_memory = ptob(freemem); 4401 static uint64_t page_load = 0; 4402 static uint64_t last_txg = 0; 4403 4404#if defined(__i386) || !defined(UMA_MD_SMALL_ALLOC) 4405 available_memory = 4406 MIN(available_memory, ptob(vmem_size(heap_arena, VMEM_FREE))); 4407#endif 4408 4409 if (freemem > (uint64_t)physmem * arc_lotsfree_percent / 100) 4410 return (0); 4411 4412 if (txg > last_txg) { 4413 last_txg = txg; 4414 page_load = 0; 4415 } 4416 /* 4417 * If we are in pageout, we know that memory is already tight, 4418 * the arc is already going to be evicting, so we just want to 4419 * continue to let page writes occur as quickly as possible. 4420 */ 4421 if (curproc == pageproc) { 4422 if (page_load > MAX(ptob(minfree), available_memory) / 4) 4423 return (SET_ERROR(ERESTART)); 4424 /* Note: reserve is inflated, so we deflate */ 4425 page_load += reserve / 8; 4426 return (0); 4427 } else if (page_load > 0 && arc_reclaim_needed()) { 4428 /* memory is low, delay before restarting */ 4429 ARCSTAT_INCR(arcstat_memory_throttle_count, 1); 4430 return (SET_ERROR(EAGAIN)); 4431 } 4432 page_load = 0; 4433#endif 4434 return (0); 4435} 4436 4437void 4438arc_tempreserve_clear(uint64_t reserve) 4439{ 4440 atomic_add_64(&arc_tempreserve, -reserve); 4441 ASSERT((int64_t)arc_tempreserve >= 0); 4442} 4443 4444int 4445arc_tempreserve_space(uint64_t reserve, uint64_t txg) 4446{ 4447 int error; 4448 uint64_t anon_size; 4449 4450 if (reserve > arc_c/4 && !arc_no_grow) { 4451 arc_c = MIN(arc_c_max, reserve * 4); 4452 DTRACE_PROBE1(arc__set_reserve, uint64_t, arc_c); 4453 } 4454 if (reserve > arc_c) 4455 return (SET_ERROR(ENOMEM)); 4456 4457 /* 4458 * Don't count loaned bufs as in flight dirty data to prevent long 4459 * network delays from blocking transactions that are ready to be 4460 * assigned to a txg. 4461 */ 4462 anon_size = MAX((int64_t)(arc_anon->arcs_size - arc_loaned_bytes), 0); 4463 4464 /* 4465 * Writes will, almost always, require additional memory allocations 4466 * in order to compress/encrypt/etc the data. We therefore need to 4467 * make sure that there is sufficient available memory for this. 4468 */ 4469 error = arc_memory_throttle(reserve, txg); 4470 if (error != 0) 4471 return (error); 4472 4473 /* 4474 * Throttle writes when the amount of dirty data in the cache 4475 * gets too large. We try to keep the cache less than half full 4476 * of dirty blocks so that our sync times don't grow too large. 4477 * Note: if two requests come in concurrently, we might let them 4478 * both succeed, when one of them should fail. Not a huge deal. 4479 */ 4480 4481 if (reserve + arc_tempreserve + anon_size > arc_c / 2 && 4482 anon_size > arc_c / 4) { 4483 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK " 4484 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n", 4485 arc_tempreserve>>10, 4486 arc_anon->arcs_lsize[ARC_BUFC_METADATA]>>10, 4487 arc_anon->arcs_lsize[ARC_BUFC_DATA]>>10, 4488 reserve>>10, arc_c>>10); 4489 return (SET_ERROR(ERESTART)); 4490 } 4491 atomic_add_64(&arc_tempreserve, reserve); 4492 return (0); 4493} 4494 4495static kmutex_t arc_lowmem_lock; 4496#ifdef _KERNEL 4497static eventhandler_tag arc_event_lowmem = NULL; 4498 4499static void 4500arc_lowmem(void *arg __unused, int howto __unused) 4501{ 4502 4503 /* Serialize access via arc_lowmem_lock. */ 4504 mutex_enter(&arc_lowmem_lock); 4505 mutex_enter(&arc_reclaim_thr_lock); 4506 needfree = 1; 4507 DTRACE_PROBE(arc__needfree); 4508 cv_signal(&arc_reclaim_thr_cv); 4509 4510 /* 4511 * It is unsafe to block here in arbitrary threads, because we can come 4512 * here from ARC itself and may hold ARC locks and thus risk a deadlock 4513 * with ARC reclaim thread. 4514 */ 4515 if (curproc == pageproc) { 4516 while (needfree) 4517 msleep(&needfree, &arc_reclaim_thr_lock, 0, "zfs:lowmem", 0); 4518 } 4519 mutex_exit(&arc_reclaim_thr_lock); 4520 mutex_exit(&arc_lowmem_lock); 4521} 4522#endif 4523 4524void 4525arc_init(void) 4526{ 4527 int i, prefetch_tunable_set = 0; 4528 4529 mutex_init(&arc_reclaim_thr_lock, NULL, MUTEX_DEFAULT, NULL); 4530 cv_init(&arc_reclaim_thr_cv, NULL, CV_DEFAULT, NULL); 4531 mutex_init(&arc_lowmem_lock, NULL, MUTEX_DEFAULT, NULL); 4532 4533 /* Convert seconds to clock ticks */ 4534 arc_min_prefetch_lifespan = 1 * hz; 4535 4536 /* Start out with 1/8 of all memory */ 4537 arc_c = kmem_size() / 8; 4538 4539#ifdef sun 4540#ifdef _KERNEL 4541 /* 4542 * On architectures where the physical memory can be larger 4543 * than the addressable space (intel in 32-bit mode), we may 4544 * need to limit the cache to 1/8 of VM size. 4545 */ 4546 arc_c = MIN(arc_c, vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 8); 4547#endif 4548#endif /* sun */ 4549 /* set min cache to 1/32 of all memory, or 16MB, whichever is more */ 4550 arc_c_min = MAX(arc_c / 4, 16 << 20); 4551 /* set max to 1/2 of all memory, or all but 1GB, whichever is more */ 4552 if (arc_c * 8 >= 1 << 30) 4553 arc_c_max = (arc_c * 8) - (1 << 30); 4554 else 4555 arc_c_max = arc_c_min; 4556 arc_c_max = MAX(arc_c * 5, arc_c_max); 4557 4558#ifdef _KERNEL 4559 /* 4560 * Allow the tunables to override our calculations if they are 4561 * reasonable (ie. over 16MB) 4562 */ 4563 if (zfs_arc_max > 16 << 20 && zfs_arc_max < kmem_size()) 4564 arc_c_max = zfs_arc_max; 4565 if (zfs_arc_min > 16 << 20 && zfs_arc_min <= arc_c_max) 4566 arc_c_min = zfs_arc_min; 4567#endif 4568 4569 arc_c = arc_c_max; 4570 arc_p = (arc_c >> 1); 4571 4572 /* limit meta-data to 1/4 of the arc capacity */ 4573 arc_meta_limit = arc_c_max / 4; 4574 4575 /* Allow the tunable to override if it is reasonable */ 4576 if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max) 4577 arc_meta_limit = zfs_arc_meta_limit; 4578 4579 if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0) 4580 arc_c_min = arc_meta_limit / 2; 4581 4582 if (zfs_arc_meta_min > 0) { 4583 arc_meta_min = zfs_arc_meta_min; 4584 } else { 4585 arc_meta_min = arc_c_min / 2; 4586 } 4587 4588 if (zfs_arc_grow_retry > 0) 4589 arc_grow_retry = zfs_arc_grow_retry; 4590 4591 if (zfs_arc_shrink_shift > 0) 4592 arc_shrink_shift = zfs_arc_shrink_shift; 4593 4594 if (zfs_arc_p_min_shift > 0) 4595 arc_p_min_shift = zfs_arc_p_min_shift; 4596 4597 /* if kmem_flags are set, lets try to use less memory */ 4598 if (kmem_debugging()) 4599 arc_c = arc_c / 2; 4600 if (arc_c < arc_c_min) 4601 arc_c = arc_c_min; 4602 4603 zfs_arc_min = arc_c_min; 4604 zfs_arc_max = arc_c_max; 4605 4606 arc_anon = &ARC_anon; 4607 arc_mru = &ARC_mru; 4608 arc_mru_ghost = &ARC_mru_ghost; 4609 arc_mfu = &ARC_mfu; 4610 arc_mfu_ghost = &ARC_mfu_ghost; 4611 arc_l2c_only = &ARC_l2c_only; 4612 arc_size = 0; 4613 4614 for (i = 0; i < ARC_BUFC_NUMLISTS; i++) { 4615 mutex_init(&arc_anon->arcs_locks[i].arcs_lock, 4616 NULL, MUTEX_DEFAULT, NULL); 4617 mutex_init(&arc_mru->arcs_locks[i].arcs_lock, 4618 NULL, MUTEX_DEFAULT, NULL); 4619 mutex_init(&arc_mru_ghost->arcs_locks[i].arcs_lock, 4620 NULL, MUTEX_DEFAULT, NULL); 4621 mutex_init(&arc_mfu->arcs_locks[i].arcs_lock, 4622 NULL, MUTEX_DEFAULT, NULL); 4623 mutex_init(&arc_mfu_ghost->arcs_locks[i].arcs_lock, 4624 NULL, MUTEX_DEFAULT, NULL); 4625 mutex_init(&arc_l2c_only->arcs_locks[i].arcs_lock, 4626 NULL, MUTEX_DEFAULT, NULL); 4627 4628 list_create(&arc_mru->arcs_lists[i], 4629 sizeof (arc_buf_hdr_t), 4630 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node)); 4631 list_create(&arc_mru_ghost->arcs_lists[i], 4632 sizeof (arc_buf_hdr_t), 4633 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node)); 4634 list_create(&arc_mfu->arcs_lists[i], 4635 sizeof (arc_buf_hdr_t), 4636 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node)); 4637 list_create(&arc_mfu_ghost->arcs_lists[i], 4638 sizeof (arc_buf_hdr_t), 4639 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node)); 4640 list_create(&arc_mfu_ghost->arcs_lists[i], 4641 sizeof (arc_buf_hdr_t), 4642 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node)); 4643 list_create(&arc_l2c_only->arcs_lists[i], 4644 sizeof (arc_buf_hdr_t), 4645 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node)); 4646 } 4647 4648 buf_init(); 4649 4650 arc_thread_exit = 0; 4651 arc_eviction_list = NULL; 4652 mutex_init(&arc_eviction_mtx, NULL, MUTEX_DEFAULT, NULL); 4653 bzero(&arc_eviction_hdr, sizeof (arc_buf_hdr_t)); 4654 4655 arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED, 4656 sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL); 4657 4658 if (arc_ksp != NULL) { 4659 arc_ksp->ks_data = &arc_stats; 4660 kstat_install(arc_ksp); 4661 } 4662 4663 (void) thread_create(NULL, 0, arc_reclaim_thread, NULL, 0, &p0, 4664 TS_RUN, minclsyspri); 4665 4666#ifdef _KERNEL 4667 arc_event_lowmem = EVENTHANDLER_REGISTER(vm_lowmem, arc_lowmem, NULL, 4668 EVENTHANDLER_PRI_FIRST); 4669#endif 4670 4671 arc_dead = FALSE; 4672 arc_warm = B_FALSE; 4673 4674 /* 4675 * Calculate maximum amount of dirty data per pool. 4676 * 4677 * If it has been set by /etc/system, take that. 4678 * Otherwise, use a percentage of physical memory defined by 4679 * zfs_dirty_data_max_percent (default 10%) with a cap at 4680 * zfs_dirty_data_max_max (default 4GB). 4681 */ 4682 if (zfs_dirty_data_max == 0) { 4683 zfs_dirty_data_max = ptob(physmem) * 4684 zfs_dirty_data_max_percent / 100; 4685 zfs_dirty_data_max = MIN(zfs_dirty_data_max, 4686 zfs_dirty_data_max_max); 4687 } 4688 4689#ifdef _KERNEL 4690 if (TUNABLE_INT_FETCH("vfs.zfs.prefetch_disable", &zfs_prefetch_disable)) 4691 prefetch_tunable_set = 1; 4692 4693#ifdef __i386__ 4694 if (prefetch_tunable_set == 0) { 4695 printf("ZFS NOTICE: Prefetch is disabled by default on i386 " 4696 "-- to enable,\n"); 4697 printf(" add \"vfs.zfs.prefetch_disable=0\" " 4698 "to /boot/loader.conf.\n"); 4699 zfs_prefetch_disable = 1; 4700 } 4701#else 4702 if ((((uint64_t)physmem * PAGESIZE) < (1ULL << 32)) && 4703 prefetch_tunable_set == 0) { 4704 printf("ZFS NOTICE: Prefetch is disabled by default if less " 4705 "than 4GB of RAM is present;\n" 4706 " to enable, add \"vfs.zfs.prefetch_disable=0\" " 4707 "to /boot/loader.conf.\n"); 4708 zfs_prefetch_disable = 1; 4709 } 4710#endif 4711 /* Warn about ZFS memory and address space requirements. */ 4712 if (((uint64_t)physmem * PAGESIZE) < (256 + 128 + 64) * (1 << 20)) { 4713 printf("ZFS WARNING: Recommended minimum RAM size is 512MB; " 4714 "expect unstable behavior.\n"); 4715 } 4716 if (kmem_size() < 512 * (1 << 20)) { 4717 printf("ZFS WARNING: Recommended minimum kmem_size is 512MB; " 4718 "expect unstable behavior.\n"); 4719 printf(" Consider tuning vm.kmem_size and " 4720 "vm.kmem_size_max\n"); 4721 printf(" in /boot/loader.conf.\n"); 4722 } 4723#endif 4724} 4725 4726void 4727arc_fini(void) 4728{ 4729 int i; 4730 4731 mutex_enter(&arc_reclaim_thr_lock); 4732 arc_thread_exit = 1; 4733 cv_signal(&arc_reclaim_thr_cv); 4734 while (arc_thread_exit != 0) 4735 cv_wait(&arc_reclaim_thr_cv, &arc_reclaim_thr_lock); 4736 mutex_exit(&arc_reclaim_thr_lock); 4737 4738 arc_flush(NULL); 4739 4740 arc_dead = TRUE; 4741 4742 if (arc_ksp != NULL) { 4743 kstat_delete(arc_ksp); 4744 arc_ksp = NULL; 4745 } 4746 4747 mutex_destroy(&arc_eviction_mtx); 4748 mutex_destroy(&arc_reclaim_thr_lock); 4749 cv_destroy(&arc_reclaim_thr_cv); 4750 4751 for (i = 0; i < ARC_BUFC_NUMLISTS; i++) { 4752 list_destroy(&arc_mru->arcs_lists[i]); 4753 list_destroy(&arc_mru_ghost->arcs_lists[i]); 4754 list_destroy(&arc_mfu->arcs_lists[i]); 4755 list_destroy(&arc_mfu_ghost->arcs_lists[i]); 4756 list_destroy(&arc_l2c_only->arcs_lists[i]); 4757 4758 mutex_destroy(&arc_anon->arcs_locks[i].arcs_lock); 4759 mutex_destroy(&arc_mru->arcs_locks[i].arcs_lock); 4760 mutex_destroy(&arc_mru_ghost->arcs_locks[i].arcs_lock); 4761 mutex_destroy(&arc_mfu->arcs_locks[i].arcs_lock); 4762 mutex_destroy(&arc_mfu_ghost->arcs_locks[i].arcs_lock); 4763 mutex_destroy(&arc_l2c_only->arcs_locks[i].arcs_lock); 4764 } 4765 4766 buf_fini(); 4767 4768 ASSERT0(arc_loaned_bytes); 4769 4770 mutex_destroy(&arc_lowmem_lock); 4771#ifdef _KERNEL 4772 if (arc_event_lowmem != NULL) 4773 EVENTHANDLER_DEREGISTER(vm_lowmem, arc_event_lowmem); 4774#endif 4775} 4776 4777/* 4778 * Level 2 ARC 4779 * 4780 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk. 4781 * It uses dedicated storage devices to hold cached data, which are populated 4782 * using large infrequent writes. The main role of this cache is to boost 4783 * the performance of random read workloads. The intended L2ARC devices 4784 * include short-stroked disks, solid state disks, and other media with 4785 * substantially faster read latency than disk. 4786 * 4787 * +-----------------------+ 4788 * | ARC | 4789 * +-----------------------+ 4790 * | ^ ^ 4791 * | | | 4792 * l2arc_feed_thread() arc_read() 4793 * | | | 4794 * | l2arc read | 4795 * V | | 4796 * +---------------+ | 4797 * | L2ARC | | 4798 * +---------------+ | 4799 * | ^ | 4800 * l2arc_write() | | 4801 * | | | 4802 * V | | 4803 * +-------+ +-------+ 4804 * | vdev | | vdev | 4805 * | cache | | cache | 4806 * +-------+ +-------+ 4807 * +=========+ .-----. 4808 * : L2ARC : |-_____-| 4809 * : devices : | Disks | 4810 * +=========+ `-_____-' 4811 * 4812 * Read requests are satisfied from the following sources, in order: 4813 * 4814 * 1) ARC 4815 * 2) vdev cache of L2ARC devices 4816 * 3) L2ARC devices 4817 * 4) vdev cache of disks 4818 * 5) disks 4819 * 4820 * Some L2ARC device types exhibit extremely slow write performance. 4821 * To accommodate for this there are some significant differences between 4822 * the L2ARC and traditional cache design: 4823 * 4824 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from 4825 * the ARC behave as usual, freeing buffers and placing headers on ghost 4826 * lists. The ARC does not send buffers to the L2ARC during eviction as 4827 * this would add inflated write latencies for all ARC memory pressure. 4828 * 4829 * 2. The L2ARC attempts to cache data from the ARC before it is evicted. 4830 * It does this by periodically scanning buffers from the eviction-end of 4831 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are 4832 * not already there. It scans until a headroom of buffers is satisfied, 4833 * which itself is a buffer for ARC eviction. If a compressible buffer is 4834 * found during scanning and selected for writing to an L2ARC device, we 4835 * temporarily boost scanning headroom during the next scan cycle to make 4836 * sure we adapt to compression effects (which might significantly reduce 4837 * the data volume we write to L2ARC). The thread that does this is 4838 * l2arc_feed_thread(), illustrated below; example sizes are included to 4839 * provide a better sense of ratio than this diagram: 4840 * 4841 * head --> tail 4842 * +---------------------+----------+ 4843 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC 4844 * +---------------------+----------+ | o L2ARC eligible 4845 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer 4846 * +---------------------+----------+ | 4847 * 15.9 Gbytes ^ 32 Mbytes | 4848 * headroom | 4849 * l2arc_feed_thread() 4850 * | 4851 * l2arc write hand <--[oooo]--' 4852 * | 8 Mbyte 4853 * | write max 4854 * V 4855 * +==============================+ 4856 * L2ARC dev |####|#|###|###| |####| ... | 4857 * +==============================+ 4858 * 32 Gbytes 4859 * 4860 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of 4861 * evicted, then the L2ARC has cached a buffer much sooner than it probably 4862 * needed to, potentially wasting L2ARC device bandwidth and storage. It is 4863 * safe to say that this is an uncommon case, since buffers at the end of 4864 * the ARC lists have moved there due to inactivity. 4865 * 4866 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom, 4867 * then the L2ARC simply misses copying some buffers. This serves as a 4868 * pressure valve to prevent heavy read workloads from both stalling the ARC 4869 * with waits and clogging the L2ARC with writes. This also helps prevent 4870 * the potential for the L2ARC to churn if it attempts to cache content too 4871 * quickly, such as during backups of the entire pool. 4872 * 4873 * 5. After system boot and before the ARC has filled main memory, there are 4874 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru 4875 * lists can remain mostly static. Instead of searching from tail of these 4876 * lists as pictured, the l2arc_feed_thread() will search from the list heads 4877 * for eligible buffers, greatly increasing its chance of finding them. 4878 * 4879 * The L2ARC device write speed is also boosted during this time so that 4880 * the L2ARC warms up faster. Since there have been no ARC evictions yet, 4881 * there are no L2ARC reads, and no fear of degrading read performance 4882 * through increased writes. 4883 * 4884 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that 4885 * the vdev queue can aggregate them into larger and fewer writes. Each 4886 * device is written to in a rotor fashion, sweeping writes through 4887 * available space then repeating. 4888 * 4889 * 7. The L2ARC does not store dirty content. It never needs to flush 4890 * write buffers back to disk based storage. 4891 * 4892 * 8. If an ARC buffer is written (and dirtied) which also exists in the 4893 * L2ARC, the now stale L2ARC buffer is immediately dropped. 4894 * 4895 * The performance of the L2ARC can be tweaked by a number of tunables, which 4896 * may be necessary for different workloads: 4897 * 4898 * l2arc_write_max max write bytes per interval 4899 * l2arc_write_boost extra write bytes during device warmup 4900 * l2arc_noprefetch skip caching prefetched buffers 4901 * l2arc_headroom number of max device writes to precache 4902 * l2arc_headroom_boost when we find compressed buffers during ARC 4903 * scanning, we multiply headroom by this 4904 * percentage factor for the next scan cycle, 4905 * since more compressed buffers are likely to 4906 * be present 4907 * l2arc_feed_secs seconds between L2ARC writing 4908 * 4909 * Tunables may be removed or added as future performance improvements are 4910 * integrated, and also may become zpool properties. 4911 * 4912 * There are three key functions that control how the L2ARC warms up: 4913 * 4914 * l2arc_write_eligible() check if a buffer is eligible to cache 4915 * l2arc_write_size() calculate how much to write 4916 * l2arc_write_interval() calculate sleep delay between writes 4917 * 4918 * These three functions determine what to write, how much, and how quickly 4919 * to send writes. 4920 */ 4921 4922static boolean_t 4923l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *hdr) 4924{ 4925 /* 4926 * A buffer is *not* eligible for the L2ARC if it: 4927 * 1. belongs to a different spa. 4928 * 2. is already cached on the L2ARC. 4929 * 3. has an I/O in progress (it may be an incomplete read). 4930 * 4. is flagged not eligible (zfs property). 4931 */ 4932 if (hdr->b_spa != spa_guid) { 4933 ARCSTAT_BUMP(arcstat_l2_write_spa_mismatch); 4934 return (B_FALSE); 4935 } 4936 if (HDR_HAS_L2HDR(hdr)) { 4937 ARCSTAT_BUMP(arcstat_l2_write_in_l2); 4938 return (B_FALSE); 4939 } 4940 if (HDR_IO_IN_PROGRESS(hdr)) { 4941 ARCSTAT_BUMP(arcstat_l2_write_hdr_io_in_progress); 4942 return (B_FALSE); 4943 } 4944 if (!HDR_L2CACHE(hdr)) { 4945 ARCSTAT_BUMP(arcstat_l2_write_not_cacheable); 4946 return (B_FALSE); 4947 } 4948 4949 return (B_TRUE); 4950} 4951 4952static uint64_t 4953l2arc_write_size(void) 4954{ 4955 uint64_t size; 4956 4957 /* 4958 * Make sure our globals have meaningful values in case the user 4959 * altered them. 4960 */ 4961 size = l2arc_write_max; 4962 if (size == 0) { 4963 cmn_err(CE_NOTE, "Bad value for l2arc_write_max, value must " 4964 "be greater than zero, resetting it to the default (%d)", 4965 L2ARC_WRITE_SIZE); 4966 size = l2arc_write_max = L2ARC_WRITE_SIZE; 4967 } 4968 4969 if (arc_warm == B_FALSE) 4970 size += l2arc_write_boost; 4971 4972 return (size); 4973 4974} 4975 4976static clock_t 4977l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote) 4978{ 4979 clock_t interval, next, now; 4980 4981 /* 4982 * If the ARC lists are busy, increase our write rate; if the 4983 * lists are stale, idle back. This is achieved by checking 4984 * how much we previously wrote - if it was more than half of 4985 * what we wanted, schedule the next write much sooner. 4986 */ 4987 if (l2arc_feed_again && wrote > (wanted / 2)) 4988 interval = (hz * l2arc_feed_min_ms) / 1000; 4989 else 4990 interval = hz * l2arc_feed_secs; 4991 4992 now = ddi_get_lbolt(); 4993 next = MAX(now, MIN(now + interval, began + interval)); 4994 4995 return (next); 4996} 4997 4998/* 4999 * Cycle through L2ARC devices. This is how L2ARC load balances. 5000 * If a device is returned, this also returns holding the spa config lock. 5001 */ 5002static l2arc_dev_t * 5003l2arc_dev_get_next(void) 5004{ 5005 l2arc_dev_t *first, *next = NULL; 5006 5007 /* 5008 * Lock out the removal of spas (spa_namespace_lock), then removal 5009 * of cache devices (l2arc_dev_mtx). Once a device has been selected, 5010 * both locks will be dropped and a spa config lock held instead. 5011 */ 5012 mutex_enter(&spa_namespace_lock); 5013 mutex_enter(&l2arc_dev_mtx); 5014 5015 /* if there are no vdevs, there is nothing to do */ 5016 if (l2arc_ndev == 0) 5017 goto out; 5018 5019 first = NULL; 5020 next = l2arc_dev_last; 5021 do { 5022 /* loop around the list looking for a non-faulted vdev */ 5023 if (next == NULL) { 5024 next = list_head(l2arc_dev_list); 5025 } else { 5026 next = list_next(l2arc_dev_list, next); 5027 if (next == NULL) 5028 next = list_head(l2arc_dev_list); 5029 } 5030 5031 /* if we have come back to the start, bail out */ 5032 if (first == NULL) 5033 first = next; 5034 else if (next == first) 5035 break; 5036 5037 } while (vdev_is_dead(next->l2ad_vdev)); 5038 5039 /* if we were unable to find any usable vdevs, return NULL */ 5040 if (vdev_is_dead(next->l2ad_vdev)) 5041 next = NULL; 5042 5043 l2arc_dev_last = next; 5044 5045out: 5046 mutex_exit(&l2arc_dev_mtx); 5047 5048 /* 5049 * Grab the config lock to prevent the 'next' device from being 5050 * removed while we are writing to it. 5051 */ 5052 if (next != NULL) 5053 spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER); 5054 mutex_exit(&spa_namespace_lock); 5055 5056 return (next); 5057} 5058 5059/* 5060 * Free buffers that were tagged for destruction. 5061 */ 5062static void 5063l2arc_do_free_on_write() 5064{ 5065 list_t *buflist; 5066 l2arc_data_free_t *df, *df_prev; 5067 5068 mutex_enter(&l2arc_free_on_write_mtx); 5069 buflist = l2arc_free_on_write; 5070 5071 for (df = list_tail(buflist); df; df = df_prev) { 5072 df_prev = list_prev(buflist, df); 5073 ASSERT(df->l2df_data != NULL); 5074 ASSERT(df->l2df_func != NULL); 5075 df->l2df_func(df->l2df_data, df->l2df_size); 5076 list_remove(buflist, df); 5077 kmem_free(df, sizeof (l2arc_data_free_t)); 5078 } 5079 5080 mutex_exit(&l2arc_free_on_write_mtx); 5081} 5082 5083/* 5084 * A write to a cache device has completed. Update all headers to allow 5085 * reads from these buffers to begin. 5086 */ 5087static void 5088l2arc_write_done(zio_t *zio) 5089{ 5090 l2arc_write_callback_t *cb; 5091 l2arc_dev_t *dev; 5092 list_t *buflist; 5093 arc_buf_hdr_t *head, *hdr, *hdr_prev; 5094 kmutex_t *hash_lock; 5095 int64_t bytes_dropped = 0; 5096 5097 cb = zio->io_private; 5098 ASSERT(cb != NULL); 5099 dev = cb->l2wcb_dev; 5100 ASSERT(dev != NULL); 5101 head = cb->l2wcb_head; 5102 ASSERT(head != NULL); 5103 buflist = &dev->l2ad_buflist; 5104 ASSERT(buflist != NULL); 5105 DTRACE_PROBE2(l2arc__iodone, zio_t *, zio, 5106 l2arc_write_callback_t *, cb); 5107 5108 if (zio->io_error != 0) 5109 ARCSTAT_BUMP(arcstat_l2_writes_error); 5110 5111 mutex_enter(&dev->l2ad_mtx); 5112 5113 /* 5114 * All writes completed, or an error was hit. 5115 */ 5116 for (hdr = list_prev(buflist, head); hdr; hdr = hdr_prev) { 5117 hdr_prev = list_prev(buflist, hdr); 5118 5119 hash_lock = HDR_LOCK(hdr); 5120 if (!mutex_tryenter(hash_lock)) { 5121 /* 5122 * This buffer misses out. It may be in a stage 5123 * of eviction. Its ARC_FLAG_L2_WRITING flag will be 5124 * left set, denying reads to this buffer. 5125 */ 5126 ARCSTAT_BUMP(arcstat_l2_writes_hdr_miss); 5127 continue; 5128 } 5129 5130 /* 5131 * It's possible that this buffer got evicted from the L1 cache 5132 * before we grabbed the vdev + hash locks, in which case 5133 * arc_hdr_realloc freed b_tmp_cdata for us if it was allocated. 5134 * Only free the buffer if we still have an L1 hdr. 5135 */ 5136 if (HDR_HAS_L1HDR(hdr) && hdr->b_l1hdr.b_tmp_cdata != NULL && 5137 HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF) 5138 l2arc_release_cdata_buf(hdr); 5139 5140 if (zio->io_error != 0) { 5141 /* 5142 * Error - drop L2ARC entry. 5143 */ 5144 trim_map_free(hdr->b_l2hdr.b_dev->l2ad_vdev, 5145 hdr->b_l2hdr.b_daddr, hdr->b_l2hdr.b_asize, 0); 5146 hdr->b_flags &= ~ARC_FLAG_HAS_L2HDR; 5147 5148 ARCSTAT_INCR(arcstat_l2_asize, -hdr->b_l2hdr.b_asize); 5149 ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size); 5150 } 5151 5152 /* 5153 * Allow ARC to begin reads to this L2ARC entry. 5154 */ 5155 hdr->b_flags &= ~ARC_FLAG_L2_WRITING; 5156 5157 mutex_exit(hash_lock); 5158 } 5159 5160 atomic_inc_64(&l2arc_writes_done); 5161 list_remove(buflist, head); 5162 ASSERT(!HDR_HAS_L1HDR(head)); 5163 kmem_cache_free(hdr_l2only_cache, head); 5164 mutex_exit(&dev->l2ad_mtx); 5165 5166 vdev_space_update(dev->l2ad_vdev, -bytes_dropped, 0, 0); 5167 5168 l2arc_do_free_on_write(); 5169 5170 kmem_free(cb, sizeof (l2arc_write_callback_t)); 5171} 5172 5173/* 5174 * A read to a cache device completed. Validate buffer contents before 5175 * handing over to the regular ARC routines. 5176 */ 5177static void 5178l2arc_read_done(zio_t *zio) 5179{ 5180 l2arc_read_callback_t *cb; 5181 arc_buf_hdr_t *hdr; 5182 arc_buf_t *buf; 5183 kmutex_t *hash_lock; 5184 int equal; 5185 5186 ASSERT(zio->io_vd != NULL); 5187 ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE); 5188 5189 spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd); 5190 5191 cb = zio->io_private; 5192 ASSERT(cb != NULL); 5193 buf = cb->l2rcb_buf; 5194 ASSERT(buf != NULL); 5195 5196 hash_lock = HDR_LOCK(buf->b_hdr); 5197 mutex_enter(hash_lock); 5198 hdr = buf->b_hdr; 5199 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 5200 5201 /* 5202 * If the buffer was compressed, decompress it first. 5203 */ 5204 if (cb->l2rcb_compress != ZIO_COMPRESS_OFF) 5205 l2arc_decompress_zio(zio, hdr, cb->l2rcb_compress); 5206 ASSERT(zio->io_data != NULL); 5207 5208 /* 5209 * Check this survived the L2ARC journey. 5210 */ 5211 equal = arc_cksum_equal(buf); 5212 if (equal && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) { 5213 mutex_exit(hash_lock); 5214 zio->io_private = buf; 5215 zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */ 5216 zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */ 5217 arc_read_done(zio); 5218 } else { 5219 mutex_exit(hash_lock); 5220 /* 5221 * Buffer didn't survive caching. Increment stats and 5222 * reissue to the original storage device. 5223 */ 5224 if (zio->io_error != 0) { 5225 ARCSTAT_BUMP(arcstat_l2_io_error); 5226 } else { 5227 zio->io_error = SET_ERROR(EIO); 5228 } 5229 if (!equal) 5230 ARCSTAT_BUMP(arcstat_l2_cksum_bad); 5231 5232 /* 5233 * If there's no waiter, issue an async i/o to the primary 5234 * storage now. If there *is* a waiter, the caller must 5235 * issue the i/o in a context where it's OK to block. 5236 */ 5237 if (zio->io_waiter == NULL) { 5238 zio_t *pio = zio_unique_parent(zio); 5239 5240 ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL); 5241 5242 zio_nowait(zio_read(pio, cb->l2rcb_spa, &cb->l2rcb_bp, 5243 buf->b_data, zio->io_size, arc_read_done, buf, 5244 zio->io_priority, cb->l2rcb_flags, &cb->l2rcb_zb)); 5245 } 5246 } 5247 5248 kmem_free(cb, sizeof (l2arc_read_callback_t)); 5249} 5250 5251/* 5252 * This is the list priority from which the L2ARC will search for pages to 5253 * cache. This is used within loops (0..3) to cycle through lists in the 5254 * desired order. This order can have a significant effect on cache 5255 * performance. 5256 * 5257 * Currently the metadata lists are hit first, MFU then MRU, followed by 5258 * the data lists. This function returns a locked list, and also returns 5259 * the lock pointer. 5260 */ 5261static list_t * 5262l2arc_list_locked(int list_num, kmutex_t **lock) 5263{ 5264 list_t *list = NULL; 5265 int idx; 5266 5267 ASSERT(list_num >= 0 && list_num < 2 * ARC_BUFC_NUMLISTS); 5268 5269 if (list_num < ARC_BUFC_NUMMETADATALISTS) { 5270 idx = list_num; 5271 list = &arc_mfu->arcs_lists[idx]; 5272 *lock = ARCS_LOCK(arc_mfu, idx); 5273 } else if (list_num < ARC_BUFC_NUMMETADATALISTS * 2) { 5274 idx = list_num - ARC_BUFC_NUMMETADATALISTS; 5275 list = &arc_mru->arcs_lists[idx]; 5276 *lock = ARCS_LOCK(arc_mru, idx); 5277 } else if (list_num < (ARC_BUFC_NUMMETADATALISTS * 2 + 5278 ARC_BUFC_NUMDATALISTS)) { 5279 idx = list_num - ARC_BUFC_NUMMETADATALISTS; 5280 list = &arc_mfu->arcs_lists[idx]; 5281 *lock = ARCS_LOCK(arc_mfu, idx); 5282 } else { 5283 idx = list_num - ARC_BUFC_NUMLISTS; 5284 list = &arc_mru->arcs_lists[idx]; 5285 *lock = ARCS_LOCK(arc_mru, idx); 5286 } 5287 5288 ASSERT(!(MUTEX_HELD(*lock))); 5289 mutex_enter(*lock); 5290 return (list); 5291} 5292 5293/* 5294 * Evict buffers from the device write hand to the distance specified in 5295 * bytes. This distance may span populated buffers, it may span nothing. 5296 * This is clearing a region on the L2ARC device ready for writing. 5297 * If the 'all' boolean is set, every buffer is evicted. 5298 */ 5299static void 5300l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all) 5301{ 5302 list_t *buflist; 5303 arc_buf_hdr_t *hdr, *hdr_prev; 5304 kmutex_t *hash_lock; 5305 uint64_t taddr; 5306 int64_t bytes_evicted = 0; 5307 5308 buflist = &dev->l2ad_buflist; 5309 5310 if (!all && dev->l2ad_first) { 5311 /* 5312 * This is the first sweep through the device. There is 5313 * nothing to evict. 5314 */ 5315 return; 5316 } 5317 5318 if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) { 5319 /* 5320 * When nearing the end of the device, evict to the end 5321 * before the device write hand jumps to the start. 5322 */ 5323 taddr = dev->l2ad_end; 5324 } else { 5325 taddr = dev->l2ad_hand + distance; 5326 } 5327 DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist, 5328 uint64_t, taddr, boolean_t, all); 5329 5330top: 5331 mutex_enter(&dev->l2ad_mtx); 5332 for (hdr = list_tail(buflist); hdr; hdr = hdr_prev) { 5333 hdr_prev = list_prev(buflist, hdr); 5334 5335 hash_lock = HDR_LOCK(hdr); 5336 if (!mutex_tryenter(hash_lock)) { 5337 /* 5338 * Missed the hash lock. Retry. 5339 */ 5340 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry); 5341 mutex_exit(&dev->l2ad_mtx); 5342 mutex_enter(hash_lock); 5343 mutex_exit(hash_lock); 5344 goto top; 5345 } 5346 5347 if (HDR_L2_WRITE_HEAD(hdr)) { 5348 /* 5349 * We hit a write head node. Leave it for 5350 * l2arc_write_done(). 5351 */ 5352 list_remove(buflist, hdr); 5353 mutex_exit(hash_lock); 5354 continue; 5355 } 5356 5357 if (!all && HDR_HAS_L2HDR(hdr) && 5358 (hdr->b_l2hdr.b_daddr > taddr || 5359 hdr->b_l2hdr.b_daddr < dev->l2ad_hand)) { 5360 /* 5361 * We've evicted to the target address, 5362 * or the end of the device. 5363 */ 5364 mutex_exit(hash_lock); 5365 break; 5366 } 5367 5368 ASSERT(HDR_HAS_L2HDR(hdr)); 5369 if (!HDR_HAS_L1HDR(hdr)) { 5370 ASSERT(!HDR_L2_READING(hdr)); 5371 /* 5372 * This doesn't exist in the ARC. Destroy. 5373 * arc_hdr_destroy() will call list_remove() 5374 * and decrement arcstat_l2_size. 5375 */ 5376 arc_change_state(arc_anon, hdr, hash_lock); 5377 arc_hdr_destroy(hdr); 5378 } else { 5379 ASSERT(hdr->b_l1hdr.b_state != arc_l2c_only); 5380 ARCSTAT_BUMP(arcstat_l2_evict_l1cached); 5381 /* 5382 * Invalidate issued or about to be issued 5383 * reads, since we may be about to write 5384 * over this location. 5385 */ 5386 if (HDR_L2_READING(hdr)) { 5387 ARCSTAT_BUMP(arcstat_l2_evict_reading); 5388 hdr->b_flags |= ARC_FLAG_L2_EVICTED; 5389 } 5390 5391 /* Tell ARC this no longer exists in L2ARC. */ 5392 ARCSTAT_INCR(arcstat_l2_asize, -hdr->b_l2hdr.b_asize); 5393 ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size); 5394 hdr->b_flags &= ~ARC_FLAG_HAS_L2HDR; 5395 list_remove(buflist, hdr); 5396 5397 /* This may have been leftover after a failed write. */ 5398 hdr->b_flags &= ~ARC_FLAG_L2_WRITING; 5399 } 5400 mutex_exit(hash_lock); 5401 } 5402 mutex_exit(&dev->l2ad_mtx); 5403 5404 vdev_space_update(dev->l2ad_vdev, -bytes_evicted, 0, 0); 5405 dev->l2ad_evict = taddr; 5406} 5407 5408/* 5409 * Find and write ARC buffers to the L2ARC device. 5410 * 5411 * An ARC_FLAG_L2_WRITING flag is set so that the L2ARC buffers are not valid 5412 * for reading until they have completed writing. 5413 * The headroom_boost is an in-out parameter used to maintain headroom boost 5414 * state between calls to this function. 5415 * 5416 * Returns the number of bytes actually written (which may be smaller than 5417 * the delta by which the device hand has changed due to alignment). 5418 */ 5419static uint64_t 5420l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz, 5421 boolean_t *headroom_boost) 5422{ 5423 arc_buf_hdr_t *hdr, *hdr_prev, *head; 5424 list_t *list; 5425 uint64_t write_asize, write_sz, headroom, buf_compress_minsz; 5426 void *buf_data; 5427 kmutex_t *list_lock; 5428 boolean_t full; 5429 l2arc_write_callback_t *cb; 5430 zio_t *pio, *wzio; 5431 uint64_t guid = spa_load_guid(spa); 5432 const boolean_t do_headroom_boost = *headroom_boost; 5433 int try; 5434 5435 ASSERT(dev->l2ad_vdev != NULL); 5436 5437 /* Lower the flag now, we might want to raise it again later. */ 5438 *headroom_boost = B_FALSE; 5439 5440 pio = NULL; 5441 write_sz = write_asize = 0; 5442 full = B_FALSE; 5443 head = kmem_cache_alloc(hdr_l2only_cache, KM_PUSHPAGE); 5444 head->b_flags |= ARC_FLAG_L2_WRITE_HEAD; 5445 head->b_flags |= ARC_FLAG_HAS_L2HDR; 5446 5447 ARCSTAT_BUMP(arcstat_l2_write_buffer_iter); 5448 /* 5449 * We will want to try to compress buffers that are at least 2x the 5450 * device sector size. 5451 */ 5452 buf_compress_minsz = 2 << dev->l2ad_vdev->vdev_ashift; 5453 5454 /* 5455 * Copy buffers for L2ARC writing. 5456 */ 5457 mutex_enter(&dev->l2ad_mtx); 5458 for (try = 0; try < 2 * ARC_BUFC_NUMLISTS; try++) { 5459 uint64_t passed_sz = 0; 5460 5461 list = l2arc_list_locked(try, &list_lock); 5462 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_iter); 5463 5464 /* 5465 * L2ARC fast warmup. 5466 * 5467 * Until the ARC is warm and starts to evict, read from the 5468 * head of the ARC lists rather than the tail. 5469 */ 5470 if (arc_warm == B_FALSE) 5471 hdr = list_head(list); 5472 else 5473 hdr = list_tail(list); 5474 if (hdr == NULL) 5475 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_null_iter); 5476 5477 headroom = target_sz * l2arc_headroom * 2 / ARC_BUFC_NUMLISTS; 5478 if (do_headroom_boost) 5479 headroom = (headroom * l2arc_headroom_boost) / 100; 5480 5481 for (; hdr; hdr = hdr_prev) { 5482 kmutex_t *hash_lock; 5483 uint64_t buf_sz; 5484 uint64_t buf_a_sz; 5485 5486 if (arc_warm == B_FALSE) 5487 hdr_prev = list_next(list, hdr); 5488 else 5489 hdr_prev = list_prev(list, hdr); 5490 ARCSTAT_INCR(arcstat_l2_write_buffer_bytes_scanned, hdr->b_size); 5491 5492 hash_lock = HDR_LOCK(hdr); 5493 if (!mutex_tryenter(hash_lock)) { 5494 ARCSTAT_BUMP(arcstat_l2_write_trylock_fail); 5495 /* 5496 * Skip this buffer rather than waiting. 5497 */ 5498 continue; 5499 } 5500 5501 passed_sz += hdr->b_size; 5502 if (passed_sz > headroom) { 5503 /* 5504 * Searched too far. 5505 */ 5506 mutex_exit(hash_lock); 5507 ARCSTAT_BUMP(arcstat_l2_write_passed_headroom); 5508 break; 5509 } 5510 5511 if (!l2arc_write_eligible(guid, hdr)) { 5512 mutex_exit(hash_lock); 5513 continue; 5514 } 5515 5516 /* 5517 * Assume that the buffer is not going to be compressed 5518 * and could take more space on disk because of a larger 5519 * disk block size. 5520 */ 5521 buf_sz = hdr->b_size; 5522 buf_a_sz = vdev_psize_to_asize(dev->l2ad_vdev, buf_sz); 5523 5524 if ((write_asize + buf_a_sz) > target_sz) { 5525 full = B_TRUE; 5526 mutex_exit(hash_lock); 5527 ARCSTAT_BUMP(arcstat_l2_write_full); 5528 break; 5529 } 5530 5531 if (pio == NULL) { 5532 /* 5533 * Insert a dummy header on the buflist so 5534 * l2arc_write_done() can find where the 5535 * write buffers begin without searching. 5536 */ 5537 list_insert_head(&dev->l2ad_buflist, head); 5538 5539 cb = kmem_alloc( 5540 sizeof (l2arc_write_callback_t), KM_SLEEP); 5541 cb->l2wcb_dev = dev; 5542 cb->l2wcb_head = head; 5543 pio = zio_root(spa, l2arc_write_done, cb, 5544 ZIO_FLAG_CANFAIL); 5545 ARCSTAT_BUMP(arcstat_l2_write_pios); 5546 } 5547 5548 /* 5549 * Create and add a new L2ARC header. 5550 */ 5551 hdr->b_l2hdr.b_dev = dev; 5552 hdr->b_flags |= ARC_FLAG_L2_WRITING; 5553 /* 5554 * Temporarily stash the data buffer in b_tmp_cdata. 5555 * The subsequent write step will pick it up from 5556 * there. This is because can't access b_l1hdr.b_buf 5557 * without holding the hash_lock, which we in turn 5558 * can't access without holding the ARC list locks 5559 * (which we want to avoid during compression/writing). 5560 */ 5561 HDR_SET_COMPRESS(hdr, ZIO_COMPRESS_OFF); 5562 hdr->b_l2hdr.b_asize = hdr->b_size; 5563 hdr->b_l1hdr.b_tmp_cdata = hdr->b_l1hdr.b_buf->b_data; 5564 5565 hdr->b_flags |= ARC_FLAG_HAS_L2HDR; 5566 5567 list_insert_head(&dev->l2ad_buflist, hdr); 5568 5569 /* 5570 * Compute and store the buffer cksum before 5571 * writing. On debug the cksum is verified first. 5572 */ 5573 arc_cksum_verify(hdr->b_l1hdr.b_buf); 5574 arc_cksum_compute(hdr->b_l1hdr.b_buf, B_TRUE); 5575 5576 mutex_exit(hash_lock); 5577 5578 write_sz += buf_sz; 5579 write_asize += buf_a_sz; 5580 } 5581 5582 mutex_exit(list_lock); 5583 5584 if (full == B_TRUE) 5585 break; 5586 } 5587 5588 /* No buffers selected for writing? */ 5589 if (pio == NULL) { 5590 ASSERT0(write_sz); 5591 mutex_exit(&dev->l2ad_mtx); 5592 ASSERT(!HDR_HAS_L1HDR(head)); 5593 kmem_cache_free(hdr_l2only_cache, head); 5594 return (0); 5595 } 5596 5597 /* 5598 * Note that elsewhere in this file arcstat_l2_asize 5599 * and the used space on l2ad_vdev are updated using b_asize, 5600 * which is not necessarily rounded up to the device block size. 5601 * Too keep accounting consistent we do the same here as well: 5602 * stats_size accumulates the sum of b_asize of the written buffers, 5603 * while write_asize accumulates the sum of b_asize rounded up 5604 * to the device block size. 5605 * The latter sum is used only to validate the corectness of the code. 5606 */ 5607 uint64_t stats_size = 0; 5608 write_asize = 0; 5609 5610 /* 5611 * Now start writing the buffers. We're starting at the write head 5612 * and work backwards, retracing the course of the buffer selector 5613 * loop above. 5614 */ 5615 for (hdr = list_prev(&dev->l2ad_buflist, head); hdr; 5616 hdr = list_prev(&dev->l2ad_buflist, hdr)) { 5617 uint64_t buf_sz; 5618 5619 /* 5620 * We shouldn't need to lock the buffer here, since we flagged 5621 * it as ARC_FLAG_L2_WRITING in the previous step, but we must 5622 * take care to only access its L2 cache parameters. In 5623 * particular, hdr->l1hdr.b_buf may be invalid by now due to 5624 * ARC eviction. 5625 */ 5626 hdr->b_l2hdr.b_daddr = dev->l2ad_hand; 5627 5628 if ((HDR_L2COMPRESS(hdr)) && 5629 hdr->b_l2hdr.b_asize >= buf_compress_minsz) { 5630 if (l2arc_compress_buf(hdr)) { 5631 /* 5632 * If compression succeeded, enable headroom 5633 * boost on the next scan cycle. 5634 */ 5635 *headroom_boost = B_TRUE; 5636 } 5637 } 5638 5639 /* 5640 * Pick up the buffer data we had previously stashed away 5641 * (and now potentially also compressed). 5642 */ 5643 buf_data = hdr->b_l1hdr.b_tmp_cdata; 5644 buf_sz = hdr->b_l2hdr.b_asize; 5645 5646 /* 5647 * If the data has not been compressed, then clear b_tmp_cdata 5648 * to make sure that it points only to a temporary compression 5649 * buffer. 5650 */ 5651 if (!L2ARC_IS_VALID_COMPRESS(HDR_GET_COMPRESS(hdr))) 5652 hdr->b_l1hdr.b_tmp_cdata = NULL; 5653 5654 /* Compression may have squashed the buffer to zero length. */ 5655 if (buf_sz != 0) { 5656 uint64_t buf_a_sz; 5657 5658 wzio = zio_write_phys(pio, dev->l2ad_vdev, 5659 dev->l2ad_hand, buf_sz, buf_data, ZIO_CHECKSUM_OFF, 5660 NULL, NULL, ZIO_PRIORITY_ASYNC_WRITE, 5661 ZIO_FLAG_CANFAIL, B_FALSE); 5662 5663 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev, 5664 zio_t *, wzio); 5665 (void) zio_nowait(wzio); 5666 5667 stats_size += buf_sz; 5668 /* 5669 * Keep the clock hand suitably device-aligned. 5670 */ 5671 buf_a_sz = vdev_psize_to_asize(dev->l2ad_vdev, buf_sz); 5672 write_asize += buf_a_sz; 5673 dev->l2ad_hand += buf_a_sz; 5674 } 5675 } 5676 5677 mutex_exit(&dev->l2ad_mtx); 5678 5679 ASSERT3U(write_asize, <=, target_sz); 5680 ARCSTAT_BUMP(arcstat_l2_writes_sent); 5681 ARCSTAT_INCR(arcstat_l2_write_bytes, write_asize); 5682 ARCSTAT_INCR(arcstat_l2_size, write_sz); 5683 ARCSTAT_INCR(arcstat_l2_asize, stats_size); 5684 vdev_space_update(dev->l2ad_vdev, stats_size, 0, 0); 5685 5686 /* 5687 * Bump device hand to the device start if it is approaching the end. 5688 * l2arc_evict() will already have evicted ahead for this case. 5689 */ 5690 if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) { 5691 dev->l2ad_hand = dev->l2ad_start; 5692 dev->l2ad_evict = dev->l2ad_start; 5693 dev->l2ad_first = B_FALSE; 5694 } 5695 5696 dev->l2ad_writing = B_TRUE; 5697 (void) zio_wait(pio); 5698 dev->l2ad_writing = B_FALSE; 5699 5700 return (write_asize); 5701} 5702 5703/* 5704 * Compresses an L2ARC buffer. 5705 * The data to be compressed must be prefilled in l1hdr.b_tmp_cdata and its 5706 * size in l2hdr->b_asize. This routine tries to compress the data and 5707 * depending on the compression result there are three possible outcomes: 5708 * *) The buffer was incompressible. The original l2hdr contents were left 5709 * untouched and are ready for writing to an L2 device. 5710 * *) The buffer was all-zeros, so there is no need to write it to an L2 5711 * device. To indicate this situation b_tmp_cdata is NULL'ed, b_asize is 5712 * set to zero and b_compress is set to ZIO_COMPRESS_EMPTY. 5713 * *) Compression succeeded and b_tmp_cdata was replaced with a temporary 5714 * data buffer which holds the compressed data to be written, and b_asize 5715 * tells us how much data there is. b_compress is set to the appropriate 5716 * compression algorithm. Once writing is done, invoke 5717 * l2arc_release_cdata_buf on this l2hdr to free this temporary buffer. 5718 * 5719 * Returns B_TRUE if compression succeeded, or B_FALSE if it didn't (the 5720 * buffer was incompressible). 5721 */ 5722static boolean_t 5723l2arc_compress_buf(arc_buf_hdr_t *hdr) 5724{ 5725 void *cdata; 5726 size_t csize, len, rounded; 5727 ASSERT(HDR_HAS_L2HDR(hdr)); 5728 l2arc_buf_hdr_t *l2hdr = &hdr->b_l2hdr; 5729 5730 ASSERT(HDR_HAS_L1HDR(hdr)); 5731 ASSERT(HDR_GET_COMPRESS(hdr) == ZIO_COMPRESS_OFF); 5732 ASSERT(hdr->b_l1hdr.b_tmp_cdata != NULL); 5733 5734 len = l2hdr->b_asize; 5735 cdata = zio_data_buf_alloc(len); 5736 ASSERT3P(cdata, !=, NULL); 5737 csize = zio_compress_data(ZIO_COMPRESS_LZ4, hdr->b_l1hdr.b_tmp_cdata, 5738 cdata, l2hdr->b_asize); 5739 5740 if (csize == 0) { 5741 /* zero block, indicate that there's nothing to write */ 5742 zio_data_buf_free(cdata, len); 5743 HDR_SET_COMPRESS(hdr, ZIO_COMPRESS_EMPTY); 5744 l2hdr->b_asize = 0; 5745 hdr->b_l1hdr.b_tmp_cdata = NULL; 5746 ARCSTAT_BUMP(arcstat_l2_compress_zeros); 5747 return (B_TRUE); 5748 } 5749 5750 rounded = P2ROUNDUP(csize, 5751 (size_t)1 << l2hdr->b_dev->l2ad_vdev->vdev_ashift); 5752 if (rounded < len) { 5753 /* 5754 * Compression succeeded, we'll keep the cdata around for 5755 * writing and release it afterwards. 5756 */ 5757 if (rounded > csize) { 5758 bzero((char *)cdata + csize, rounded - csize); 5759 csize = rounded; 5760 } 5761 HDR_SET_COMPRESS(hdr, ZIO_COMPRESS_LZ4); 5762 l2hdr->b_asize = csize; 5763 hdr->b_l1hdr.b_tmp_cdata = cdata; 5764 ARCSTAT_BUMP(arcstat_l2_compress_successes); 5765 return (B_TRUE); 5766 } else { 5767 /* 5768 * Compression failed, release the compressed buffer. 5769 * l2hdr will be left unmodified. 5770 */ 5771 zio_data_buf_free(cdata, len); 5772 ARCSTAT_BUMP(arcstat_l2_compress_failures); 5773 return (B_FALSE); 5774 } 5775} 5776 5777/* 5778 * Decompresses a zio read back from an l2arc device. On success, the 5779 * underlying zio's io_data buffer is overwritten by the uncompressed 5780 * version. On decompression error (corrupt compressed stream), the 5781 * zio->io_error value is set to signal an I/O error. 5782 * 5783 * Please note that the compressed data stream is not checksummed, so 5784 * if the underlying device is experiencing data corruption, we may feed 5785 * corrupt data to the decompressor, so the decompressor needs to be 5786 * able to handle this situation (LZ4 does). 5787 */ 5788static void 5789l2arc_decompress_zio(zio_t *zio, arc_buf_hdr_t *hdr, enum zio_compress c) 5790{ 5791 ASSERT(L2ARC_IS_VALID_COMPRESS(c)); 5792 5793 if (zio->io_error != 0) { 5794 /* 5795 * An io error has occured, just restore the original io 5796 * size in preparation for a main pool read. 5797 */ 5798 zio->io_orig_size = zio->io_size = hdr->b_size; 5799 return; 5800 } 5801 5802 if (c == ZIO_COMPRESS_EMPTY) { 5803 /* 5804 * An empty buffer results in a null zio, which means we 5805 * need to fill its io_data after we're done restoring the 5806 * buffer's contents. 5807 */ 5808 ASSERT(hdr->b_l1hdr.b_buf != NULL); 5809 bzero(hdr->b_l1hdr.b_buf->b_data, hdr->b_size); 5810 zio->io_data = zio->io_orig_data = hdr->b_l1hdr.b_buf->b_data; 5811 } else { 5812 ASSERT(zio->io_data != NULL); 5813 /* 5814 * We copy the compressed data from the start of the arc buffer 5815 * (the zio_read will have pulled in only what we need, the 5816 * rest is garbage which we will overwrite at decompression) 5817 * and then decompress back to the ARC data buffer. This way we 5818 * can minimize copying by simply decompressing back over the 5819 * original compressed data (rather than decompressing to an 5820 * aux buffer and then copying back the uncompressed buffer, 5821 * which is likely to be much larger). 5822 */ 5823 uint64_t csize; 5824 void *cdata; 5825 5826 csize = zio->io_size; 5827 cdata = zio_data_buf_alloc(csize); 5828 bcopy(zio->io_data, cdata, csize); 5829 if (zio_decompress_data(c, cdata, zio->io_data, csize, 5830 hdr->b_size) != 0) 5831 zio->io_error = EIO; 5832 zio_data_buf_free(cdata, csize); 5833 } 5834 5835 /* Restore the expected uncompressed IO size. */ 5836 zio->io_orig_size = zio->io_size = hdr->b_size; 5837} 5838 5839/* 5840 * Releases the temporary b_tmp_cdata buffer in an l2arc header structure. 5841 * This buffer serves as a temporary holder of compressed data while 5842 * the buffer entry is being written to an l2arc device. Once that is 5843 * done, we can dispose of it. 5844 */ 5845static void 5846l2arc_release_cdata_buf(arc_buf_hdr_t *hdr) 5847{ 5848 ASSERT(HDR_HAS_L1HDR(hdr)); 5849 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_EMPTY) { 5850 /* 5851 * If the data was compressed, then we've allocated a 5852 * temporary buffer for it, so now we need to release it. 5853 */ 5854 ASSERT(hdr->b_l1hdr.b_tmp_cdata != NULL); 5855 zio_data_buf_free(hdr->b_l1hdr.b_tmp_cdata, 5856 hdr->b_size); 5857 hdr->b_l1hdr.b_tmp_cdata = NULL; 5858 } else { 5859 ASSERT(hdr->b_l1hdr.b_tmp_cdata == NULL); 5860 } 5861} 5862 5863/* 5864 * This thread feeds the L2ARC at regular intervals. This is the beating 5865 * heart of the L2ARC. 5866 */ 5867static void 5868l2arc_feed_thread(void *dummy __unused) 5869{ 5870 callb_cpr_t cpr; 5871 l2arc_dev_t *dev; 5872 spa_t *spa; 5873 uint64_t size, wrote; 5874 clock_t begin, next = ddi_get_lbolt(); 5875 boolean_t headroom_boost = B_FALSE; 5876 5877 CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG); 5878 5879 mutex_enter(&l2arc_feed_thr_lock); 5880 5881 while (l2arc_thread_exit == 0) { 5882 CALLB_CPR_SAFE_BEGIN(&cpr); 5883 (void) cv_timedwait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock, 5884 next - ddi_get_lbolt()); 5885 CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock); 5886 next = ddi_get_lbolt() + hz; 5887 5888 /* 5889 * Quick check for L2ARC devices. 5890 */ 5891 mutex_enter(&l2arc_dev_mtx); 5892 if (l2arc_ndev == 0) { 5893 mutex_exit(&l2arc_dev_mtx); 5894 continue; 5895 } 5896 mutex_exit(&l2arc_dev_mtx); 5897 begin = ddi_get_lbolt(); 5898 5899 /* 5900 * This selects the next l2arc device to write to, and in 5901 * doing so the next spa to feed from: dev->l2ad_spa. This 5902 * will return NULL if there are now no l2arc devices or if 5903 * they are all faulted. 5904 * 5905 * If a device is returned, its spa's config lock is also 5906 * held to prevent device removal. l2arc_dev_get_next() 5907 * will grab and release l2arc_dev_mtx. 5908 */ 5909 if ((dev = l2arc_dev_get_next()) == NULL) 5910 continue; 5911 5912 spa = dev->l2ad_spa; 5913 ASSERT(spa != NULL); 5914 5915 /* 5916 * If the pool is read-only then force the feed thread to 5917 * sleep a little longer. 5918 */ 5919 if (!spa_writeable(spa)) { 5920 next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz; 5921 spa_config_exit(spa, SCL_L2ARC, dev); 5922 continue; 5923 } 5924 5925 /* 5926 * Avoid contributing to memory pressure. 5927 */ 5928 if (arc_reclaim_needed()) { 5929 ARCSTAT_BUMP(arcstat_l2_abort_lowmem); 5930 spa_config_exit(spa, SCL_L2ARC, dev); 5931 continue; 5932 } 5933 5934 ARCSTAT_BUMP(arcstat_l2_feeds); 5935 5936 size = l2arc_write_size(); 5937 5938 /* 5939 * Evict L2ARC buffers that will be overwritten. 5940 */ 5941 l2arc_evict(dev, size, B_FALSE); 5942 5943 /* 5944 * Write ARC buffers. 5945 */ 5946 wrote = l2arc_write_buffers(spa, dev, size, &headroom_boost); 5947 5948 /* 5949 * Calculate interval between writes. 5950 */ 5951 next = l2arc_write_interval(begin, size, wrote); 5952 spa_config_exit(spa, SCL_L2ARC, dev); 5953 } 5954 5955 l2arc_thread_exit = 0; 5956 cv_broadcast(&l2arc_feed_thr_cv); 5957 CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */ 5958 thread_exit(); 5959} 5960 5961boolean_t 5962l2arc_vdev_present(vdev_t *vd) 5963{ 5964 l2arc_dev_t *dev; 5965 5966 mutex_enter(&l2arc_dev_mtx); 5967 for (dev = list_head(l2arc_dev_list); dev != NULL; 5968 dev = list_next(l2arc_dev_list, dev)) { 5969 if (dev->l2ad_vdev == vd) 5970 break; 5971 } 5972 mutex_exit(&l2arc_dev_mtx); 5973 5974 return (dev != NULL); 5975} 5976 5977/* 5978 * Add a vdev for use by the L2ARC. By this point the spa has already 5979 * validated the vdev and opened it. 5980 */ 5981void 5982l2arc_add_vdev(spa_t *spa, vdev_t *vd) 5983{ 5984 l2arc_dev_t *adddev; 5985 5986 ASSERT(!l2arc_vdev_present(vd)); 5987 5988 vdev_ashift_optimize(vd); 5989 5990 /* 5991 * Create a new l2arc device entry. 5992 */ 5993 adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP); 5994 adddev->l2ad_spa = spa; 5995 adddev->l2ad_vdev = vd; 5996 adddev->l2ad_start = VDEV_LABEL_START_SIZE; 5997 adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd); 5998 adddev->l2ad_hand = adddev->l2ad_start; 5999 adddev->l2ad_evict = adddev->l2ad_start; 6000 adddev->l2ad_first = B_TRUE; 6001 adddev->l2ad_writing = B_FALSE; 6002 6003 mutex_init(&adddev->l2ad_mtx, NULL, MUTEX_DEFAULT, NULL); 6004 /* 6005 * This is a list of all ARC buffers that are still valid on the 6006 * device. 6007 */ 6008 list_create(&adddev->l2ad_buflist, sizeof (arc_buf_hdr_t), 6009 offsetof(arc_buf_hdr_t, b_l2hdr.b_l2node)); 6010 6011 vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand); 6012 6013 /* 6014 * Add device to global list 6015 */ 6016 mutex_enter(&l2arc_dev_mtx); 6017 list_insert_head(l2arc_dev_list, adddev); 6018 atomic_inc_64(&l2arc_ndev); 6019 mutex_exit(&l2arc_dev_mtx); 6020} 6021 6022/* 6023 * Remove a vdev from the L2ARC. 6024 */ 6025void 6026l2arc_remove_vdev(vdev_t *vd) 6027{ 6028 l2arc_dev_t *dev, *nextdev, *remdev = NULL; 6029 6030 /* 6031 * Find the device by vdev 6032 */ 6033 mutex_enter(&l2arc_dev_mtx); 6034 for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) { 6035 nextdev = list_next(l2arc_dev_list, dev); 6036 if (vd == dev->l2ad_vdev) { 6037 remdev = dev; 6038 break; 6039 } 6040 } 6041 ASSERT(remdev != NULL); 6042 6043 /* 6044 * Remove device from global list 6045 */ 6046 list_remove(l2arc_dev_list, remdev); 6047 l2arc_dev_last = NULL; /* may have been invalidated */ 6048 atomic_dec_64(&l2arc_ndev); 6049 mutex_exit(&l2arc_dev_mtx); 6050 6051 /* 6052 * Clear all buflists and ARC references. L2ARC device flush. 6053 */ 6054 l2arc_evict(remdev, 0, B_TRUE); 6055 list_destroy(&remdev->l2ad_buflist); 6056 mutex_destroy(&remdev->l2ad_mtx); 6057 kmem_free(remdev, sizeof (l2arc_dev_t)); 6058} 6059 6060void 6061l2arc_init(void) 6062{ 6063 l2arc_thread_exit = 0; 6064 l2arc_ndev = 0; 6065 l2arc_writes_sent = 0; 6066 l2arc_writes_done = 0; 6067 6068 mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL); 6069 cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL); 6070 mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL); 6071 mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL); 6072 6073 l2arc_dev_list = &L2ARC_dev_list; 6074 l2arc_free_on_write = &L2ARC_free_on_write; 6075 list_create(l2arc_dev_list, sizeof (l2arc_dev_t), 6076 offsetof(l2arc_dev_t, l2ad_node)); 6077 list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t), 6078 offsetof(l2arc_data_free_t, l2df_list_node)); 6079} 6080 6081void 6082l2arc_fini(void) 6083{ 6084 /* 6085 * This is called from dmu_fini(), which is called from spa_fini(); 6086 * Because of this, we can assume that all l2arc devices have 6087 * already been removed when the pools themselves were removed. 6088 */ 6089 6090 l2arc_do_free_on_write(); 6091 6092 mutex_destroy(&l2arc_feed_thr_lock); 6093 cv_destroy(&l2arc_feed_thr_cv); 6094 mutex_destroy(&l2arc_dev_mtx); 6095 mutex_destroy(&l2arc_free_on_write_mtx); 6096 6097 list_destroy(l2arc_dev_list); 6098 list_destroy(l2arc_free_on_write); 6099} 6100 6101void 6102l2arc_start(void) 6103{ 6104 if (!(spa_mode_global & FWRITE)) 6105 return; 6106 6107 (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0, 6108 TS_RUN, minclsyspri); 6109} 6110 6111void 6112l2arc_stop(void) 6113{ 6114 if (!(spa_mode_global & FWRITE)) 6115 return; 6116 6117 mutex_enter(&l2arc_feed_thr_lock); 6118 cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */ 6119 l2arc_thread_exit = 1; 6120 while (l2arc_thread_exit != 0) 6121 cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock); 6122 mutex_exit(&l2arc_feed_thr_lock); 6123} 6124