arc.c revision 274625
1/* 2 * CDDL HEADER START 3 * 4 * The contents of this file are subject to the terms of the 5 * Common Development and Distribution License (the "License"). 6 * You may not use this file except in compliance with the License. 7 * 8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE 9 * or http://www.opensolaris.org/os/licensing. 10 * See the License for the specific language governing permissions 11 * and limitations under the License. 12 * 13 * When distributing Covered Code, include this CDDL HEADER in each 14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE. 15 * If applicable, add the following below this CDDL HEADER, with the 16 * fields enclosed by brackets "[]" replaced with your own identifying 17 * information: Portions Copyright [yyyy] [name of copyright owner] 18 * 19 * CDDL HEADER END 20 */ 21/* 22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved. 23 * Copyright (c) 2011, 2014 by Delphix. All rights reserved. 24 * Copyright (c) 2014 by Saso Kiselkov. All rights reserved. 25 * Copyright 2014 Nexenta Systems, Inc. All rights reserved. 26 */ 27 28/* 29 * DVA-based Adjustable Replacement Cache 30 * 31 * While much of the theory of operation used here is 32 * based on the self-tuning, low overhead replacement cache 33 * presented by Megiddo and Modha at FAST 2003, there are some 34 * significant differences: 35 * 36 * 1. The Megiddo and Modha model assumes any page is evictable. 37 * Pages in its cache cannot be "locked" into memory. This makes 38 * the eviction algorithm simple: evict the last page in the list. 39 * This also make the performance characteristics easy to reason 40 * about. Our cache is not so simple. At any given moment, some 41 * subset of the blocks in the cache are un-evictable because we 42 * have handed out a reference to them. Blocks are only evictable 43 * when there are no external references active. This makes 44 * eviction far more problematic: we choose to evict the evictable 45 * blocks that are the "lowest" in the list. 46 * 47 * There are times when it is not possible to evict the requested 48 * space. In these circumstances we are unable to adjust the cache 49 * size. To prevent the cache growing unbounded at these times we 50 * implement a "cache throttle" that slows the flow of new data 51 * into the cache until we can make space available. 52 * 53 * 2. The Megiddo and Modha model assumes a fixed cache size. 54 * Pages are evicted when the cache is full and there is a cache 55 * miss. Our model has a variable sized cache. It grows with 56 * high use, but also tries to react to memory pressure from the 57 * operating system: decreasing its size when system memory is 58 * tight. 59 * 60 * 3. The Megiddo and Modha model assumes a fixed page size. All 61 * elements of the cache are therefore exactly the same size. So 62 * when adjusting the cache size following a cache miss, its simply 63 * a matter of choosing a single page to evict. In our model, we 64 * have variable sized cache blocks (rangeing from 512 bytes to 65 * 128K bytes). We therefore choose a set of blocks to evict to make 66 * space for a cache miss that approximates as closely as possible 67 * the space used by the new block. 68 * 69 * See also: "ARC: A Self-Tuning, Low Overhead Replacement Cache" 70 * by N. Megiddo & D. Modha, FAST 2003 71 */ 72 73/* 74 * The locking model: 75 * 76 * A new reference to a cache buffer can be obtained in two 77 * ways: 1) via a hash table lookup using the DVA as a key, 78 * or 2) via one of the ARC lists. The arc_read() interface 79 * uses method 1, while the internal arc algorithms for 80 * adjusting the cache use method 2. We therefore provide two 81 * types of locks: 1) the hash table lock array, and 2) the 82 * arc list locks. 83 * 84 * Buffers do not have their own mutexs, rather they rely on the 85 * hash table mutexs for the bulk of their protection (i.e. most 86 * fields in the arc_buf_hdr_t are protected by these mutexs). 87 * 88 * buf_hash_find() returns the appropriate mutex (held) when it 89 * locates the requested buffer in the hash table. It returns 90 * NULL for the mutex if the buffer was not in the table. 91 * 92 * buf_hash_remove() expects the appropriate hash mutex to be 93 * already held before it is invoked. 94 * 95 * Each arc state also has a mutex which is used to protect the 96 * buffer list associated with the state. When attempting to 97 * obtain a hash table lock while holding an arc list lock you 98 * must use: mutex_tryenter() to avoid deadlock. Also note that 99 * the active state mutex must be held before the ghost state mutex. 100 * 101 * Arc buffers may have an associated eviction callback function. 102 * This function will be invoked prior to removing the buffer (e.g. 103 * in arc_do_user_evicts()). Note however that the data associated 104 * with the buffer may be evicted prior to the callback. The callback 105 * must be made with *no locks held* (to prevent deadlock). Additionally, 106 * the users of callbacks must ensure that their private data is 107 * protected from simultaneous callbacks from arc_clear_callback() 108 * and arc_do_user_evicts(). 109 * 110 * Note that the majority of the performance stats are manipulated 111 * with atomic operations. 112 * 113 * The L2ARC uses the l2arc_buflist_mtx global mutex for the following: 114 * 115 * - L2ARC buflist creation 116 * - L2ARC buflist eviction 117 * - L2ARC write completion, which walks L2ARC buflists 118 * - ARC header destruction, as it removes from L2ARC buflists 119 * - ARC header release, as it removes from L2ARC buflists 120 */ 121 122#include <sys/spa.h> 123#include <sys/zio.h> 124#include <sys/zio_compress.h> 125#include <sys/zfs_context.h> 126#include <sys/arc.h> 127#include <sys/refcount.h> 128#include <sys/vdev.h> 129#include <sys/vdev_impl.h> 130#include <sys/dsl_pool.h> 131#ifdef _KERNEL 132#include <sys/dnlc.h> 133#endif 134#include <sys/callb.h> 135#include <sys/kstat.h> 136#include <sys/trim_map.h> 137#include <zfs_fletcher.h> 138#include <sys/sdt.h> 139 140#include <vm/vm_pageout.h> 141#include <machine/vmparam.h> 142 143#ifdef illumos 144#ifndef _KERNEL 145/* set with ZFS_DEBUG=watch, to enable watchpoints on frozen buffers */ 146boolean_t arc_watch = B_FALSE; 147int arc_procfd; 148#endif 149#endif /* illumos */ 150 151static kmutex_t arc_reclaim_thr_lock; 152static kcondvar_t arc_reclaim_thr_cv; /* used to signal reclaim thr */ 153static uint8_t arc_thread_exit; 154 155#define ARC_REDUCE_DNLC_PERCENT 3 156uint_t arc_reduce_dnlc_percent = ARC_REDUCE_DNLC_PERCENT; 157 158typedef enum arc_reclaim_strategy { 159 ARC_RECLAIM_AGGR, /* Aggressive reclaim strategy */ 160 ARC_RECLAIM_CONS /* Conservative reclaim strategy */ 161} arc_reclaim_strategy_t; 162 163/* 164 * The number of iterations through arc_evict_*() before we 165 * drop & reacquire the lock. 166 */ 167int arc_evict_iterations = 100; 168 169/* number of seconds before growing cache again */ 170static int arc_grow_retry = 60; 171 172/* shift of arc_c for calculating both min and max arc_p */ 173static int arc_p_min_shift = 4; 174 175/* log2(fraction of arc to reclaim) */ 176static int arc_shrink_shift = 5; 177 178/* 179 * minimum lifespan of a prefetch block in clock ticks 180 * (initialized in arc_init()) 181 */ 182static int arc_min_prefetch_lifespan; 183 184/* 185 * If this percent of memory is free, don't throttle. 186 */ 187int arc_lotsfree_percent = 10; 188 189static int arc_dead; 190extern int zfs_prefetch_disable; 191 192/* 193 * The arc has filled available memory and has now warmed up. 194 */ 195static boolean_t arc_warm; 196 197uint64_t zfs_arc_max; 198uint64_t zfs_arc_min; 199uint64_t zfs_arc_meta_limit = 0; 200int zfs_arc_grow_retry = 0; 201int zfs_arc_shrink_shift = 0; 202int zfs_arc_p_min_shift = 0; 203int zfs_disable_dup_eviction = 0; 204uint64_t zfs_arc_average_blocksize = 8 * 1024; /* 8KB */ 205u_int zfs_arc_free_target = 0; 206 207static int sysctl_vfs_zfs_arc_free_target(SYSCTL_HANDLER_ARGS); 208 209#ifdef _KERNEL 210static void 211arc_free_target_init(void *unused __unused) 212{ 213 214 zfs_arc_free_target = vm_pageout_wakeup_thresh; 215} 216SYSINIT(arc_free_target_init, SI_SUB_KTHREAD_PAGE, SI_ORDER_ANY, 217 arc_free_target_init, NULL); 218 219TUNABLE_QUAD("vfs.zfs.arc_max", &zfs_arc_max); 220TUNABLE_QUAD("vfs.zfs.arc_min", &zfs_arc_min); 221TUNABLE_QUAD("vfs.zfs.arc_meta_limit", &zfs_arc_meta_limit); 222TUNABLE_QUAD("vfs.zfs.arc_average_blocksize", &zfs_arc_average_blocksize); 223TUNABLE_INT("vfs.zfs.arc_shrink_shift", &zfs_arc_shrink_shift); 224SYSCTL_DECL(_vfs_zfs); 225SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_max, CTLFLAG_RDTUN, &zfs_arc_max, 0, 226 "Maximum ARC size"); 227SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_min, CTLFLAG_RDTUN, &zfs_arc_min, 0, 228 "Minimum ARC size"); 229SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_average_blocksize, CTLFLAG_RDTUN, 230 &zfs_arc_average_blocksize, 0, 231 "ARC average blocksize"); 232SYSCTL_INT(_vfs_zfs, OID_AUTO, arc_shrink_shift, CTLFLAG_RW, 233 &arc_shrink_shift, 0, 234 "log2(fraction of arc to reclaim)"); 235 236/* 237 * We don't have a tunable for arc_free_target due to the dependency on 238 * pagedaemon initialisation. 239 */ 240SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_free_target, 241 CTLTYPE_UINT | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, sizeof(u_int), 242 sysctl_vfs_zfs_arc_free_target, "IU", 243 "Desired number of free pages below which ARC triggers reclaim"); 244 245static int 246sysctl_vfs_zfs_arc_free_target(SYSCTL_HANDLER_ARGS) 247{ 248 u_int val; 249 int err; 250 251 val = zfs_arc_free_target; 252 err = sysctl_handle_int(oidp, &val, 0, req); 253 if (err != 0 || req->newptr == NULL) 254 return (err); 255 256 if (val < minfree) 257 return (EINVAL); 258 if (val > cnt.v_page_count) 259 return (EINVAL); 260 261 zfs_arc_free_target = val; 262 263 return (0); 264} 265#endif 266 267/* 268 * Note that buffers can be in one of 6 states: 269 * ARC_anon - anonymous (discussed below) 270 * ARC_mru - recently used, currently cached 271 * ARC_mru_ghost - recentely used, no longer in cache 272 * ARC_mfu - frequently used, currently cached 273 * ARC_mfu_ghost - frequently used, no longer in cache 274 * ARC_l2c_only - exists in L2ARC but not other states 275 * When there are no active references to the buffer, they are 276 * are linked onto a list in one of these arc states. These are 277 * the only buffers that can be evicted or deleted. Within each 278 * state there are multiple lists, one for meta-data and one for 279 * non-meta-data. Meta-data (indirect blocks, blocks of dnodes, 280 * etc.) is tracked separately so that it can be managed more 281 * explicitly: favored over data, limited explicitly. 282 * 283 * Anonymous buffers are buffers that are not associated with 284 * a DVA. These are buffers that hold dirty block copies 285 * before they are written to stable storage. By definition, 286 * they are "ref'd" and are considered part of arc_mru 287 * that cannot be freed. Generally, they will aquire a DVA 288 * as they are written and migrate onto the arc_mru list. 289 * 290 * The ARC_l2c_only state is for buffers that are in the second 291 * level ARC but no longer in any of the ARC_m* lists. The second 292 * level ARC itself may also contain buffers that are in any of 293 * the ARC_m* states - meaning that a buffer can exist in two 294 * places. The reason for the ARC_l2c_only state is to keep the 295 * buffer header in the hash table, so that reads that hit the 296 * second level ARC benefit from these fast lookups. 297 */ 298 299#define ARCS_LOCK_PAD CACHE_LINE_SIZE 300struct arcs_lock { 301 kmutex_t arcs_lock; 302#ifdef _KERNEL 303 unsigned char pad[(ARCS_LOCK_PAD - sizeof (kmutex_t))]; 304#endif 305}; 306 307/* 308 * must be power of two for mask use to work 309 * 310 */ 311#define ARC_BUFC_NUMDATALISTS 16 312#define ARC_BUFC_NUMMETADATALISTS 16 313#define ARC_BUFC_NUMLISTS (ARC_BUFC_NUMMETADATALISTS + ARC_BUFC_NUMDATALISTS) 314 315typedef struct arc_state { 316 uint64_t arcs_lsize[ARC_BUFC_NUMTYPES]; /* amount of evictable data */ 317 uint64_t arcs_size; /* total amount of data in this state */ 318 list_t arcs_lists[ARC_BUFC_NUMLISTS]; /* list of evictable buffers */ 319 struct arcs_lock arcs_locks[ARC_BUFC_NUMLISTS] __aligned(CACHE_LINE_SIZE); 320} arc_state_t; 321 322#define ARCS_LOCK(s, i) (&((s)->arcs_locks[(i)].arcs_lock)) 323 324/* The 6 states: */ 325static arc_state_t ARC_anon; 326static arc_state_t ARC_mru; 327static arc_state_t ARC_mru_ghost; 328static arc_state_t ARC_mfu; 329static arc_state_t ARC_mfu_ghost; 330static arc_state_t ARC_l2c_only; 331 332typedef struct arc_stats { 333 kstat_named_t arcstat_hits; 334 kstat_named_t arcstat_misses; 335 kstat_named_t arcstat_demand_data_hits; 336 kstat_named_t arcstat_demand_data_misses; 337 kstat_named_t arcstat_demand_metadata_hits; 338 kstat_named_t arcstat_demand_metadata_misses; 339 kstat_named_t arcstat_prefetch_data_hits; 340 kstat_named_t arcstat_prefetch_data_misses; 341 kstat_named_t arcstat_prefetch_metadata_hits; 342 kstat_named_t arcstat_prefetch_metadata_misses; 343 kstat_named_t arcstat_mru_hits; 344 kstat_named_t arcstat_mru_ghost_hits; 345 kstat_named_t arcstat_mfu_hits; 346 kstat_named_t arcstat_mfu_ghost_hits; 347 kstat_named_t arcstat_allocated; 348 kstat_named_t arcstat_deleted; 349 kstat_named_t arcstat_stolen; 350 kstat_named_t arcstat_recycle_miss; 351 /* 352 * Number of buffers that could not be evicted because the hash lock 353 * was held by another thread. The lock may not necessarily be held 354 * by something using the same buffer, since hash locks are shared 355 * by multiple buffers. 356 */ 357 kstat_named_t arcstat_mutex_miss; 358 /* 359 * Number of buffers skipped because they have I/O in progress, are 360 * indrect prefetch buffers that have not lived long enough, or are 361 * not from the spa we're trying to evict from. 362 */ 363 kstat_named_t arcstat_evict_skip; 364 kstat_named_t arcstat_evict_l2_cached; 365 kstat_named_t arcstat_evict_l2_eligible; 366 kstat_named_t arcstat_evict_l2_ineligible; 367 kstat_named_t arcstat_hash_elements; 368 kstat_named_t arcstat_hash_elements_max; 369 kstat_named_t arcstat_hash_collisions; 370 kstat_named_t arcstat_hash_chains; 371 kstat_named_t arcstat_hash_chain_max; 372 kstat_named_t arcstat_p; 373 kstat_named_t arcstat_c; 374 kstat_named_t arcstat_c_min; 375 kstat_named_t arcstat_c_max; 376 kstat_named_t arcstat_size; 377 kstat_named_t arcstat_hdr_size; 378 kstat_named_t arcstat_data_size; 379 kstat_named_t arcstat_other_size; 380 kstat_named_t arcstat_l2_hits; 381 kstat_named_t arcstat_l2_misses; 382 kstat_named_t arcstat_l2_feeds; 383 kstat_named_t arcstat_l2_rw_clash; 384 kstat_named_t arcstat_l2_read_bytes; 385 kstat_named_t arcstat_l2_write_bytes; 386 kstat_named_t arcstat_l2_writes_sent; 387 kstat_named_t arcstat_l2_writes_done; 388 kstat_named_t arcstat_l2_writes_error; 389 kstat_named_t arcstat_l2_writes_hdr_miss; 390 kstat_named_t arcstat_l2_evict_lock_retry; 391 kstat_named_t arcstat_l2_evict_reading; 392 kstat_named_t arcstat_l2_free_on_write; 393 kstat_named_t arcstat_l2_abort_lowmem; 394 kstat_named_t arcstat_l2_cksum_bad; 395 kstat_named_t arcstat_l2_io_error; 396 kstat_named_t arcstat_l2_size; 397 kstat_named_t arcstat_l2_asize; 398 kstat_named_t arcstat_l2_hdr_size; 399 kstat_named_t arcstat_l2_compress_successes; 400 kstat_named_t arcstat_l2_compress_zeros; 401 kstat_named_t arcstat_l2_compress_failures; 402 kstat_named_t arcstat_l2_write_trylock_fail; 403 kstat_named_t arcstat_l2_write_passed_headroom; 404 kstat_named_t arcstat_l2_write_spa_mismatch; 405 kstat_named_t arcstat_l2_write_in_l2; 406 kstat_named_t arcstat_l2_write_hdr_io_in_progress; 407 kstat_named_t arcstat_l2_write_not_cacheable; 408 kstat_named_t arcstat_l2_write_full; 409 kstat_named_t arcstat_l2_write_buffer_iter; 410 kstat_named_t arcstat_l2_write_pios; 411 kstat_named_t arcstat_l2_write_buffer_bytes_scanned; 412 kstat_named_t arcstat_l2_write_buffer_list_iter; 413 kstat_named_t arcstat_l2_write_buffer_list_null_iter; 414 kstat_named_t arcstat_memory_throttle_count; 415 kstat_named_t arcstat_duplicate_buffers; 416 kstat_named_t arcstat_duplicate_buffers_size; 417 kstat_named_t arcstat_duplicate_reads; 418} arc_stats_t; 419 420static arc_stats_t arc_stats = { 421 { "hits", KSTAT_DATA_UINT64 }, 422 { "misses", KSTAT_DATA_UINT64 }, 423 { "demand_data_hits", KSTAT_DATA_UINT64 }, 424 { "demand_data_misses", KSTAT_DATA_UINT64 }, 425 { "demand_metadata_hits", KSTAT_DATA_UINT64 }, 426 { "demand_metadata_misses", KSTAT_DATA_UINT64 }, 427 { "prefetch_data_hits", KSTAT_DATA_UINT64 }, 428 { "prefetch_data_misses", KSTAT_DATA_UINT64 }, 429 { "prefetch_metadata_hits", KSTAT_DATA_UINT64 }, 430 { "prefetch_metadata_misses", KSTAT_DATA_UINT64 }, 431 { "mru_hits", KSTAT_DATA_UINT64 }, 432 { "mru_ghost_hits", KSTAT_DATA_UINT64 }, 433 { "mfu_hits", KSTAT_DATA_UINT64 }, 434 { "mfu_ghost_hits", KSTAT_DATA_UINT64 }, 435 { "allocated", KSTAT_DATA_UINT64 }, 436 { "deleted", KSTAT_DATA_UINT64 }, 437 { "stolen", KSTAT_DATA_UINT64 }, 438 { "recycle_miss", KSTAT_DATA_UINT64 }, 439 { "mutex_miss", KSTAT_DATA_UINT64 }, 440 { "evict_skip", KSTAT_DATA_UINT64 }, 441 { "evict_l2_cached", KSTAT_DATA_UINT64 }, 442 { "evict_l2_eligible", KSTAT_DATA_UINT64 }, 443 { "evict_l2_ineligible", KSTAT_DATA_UINT64 }, 444 { "hash_elements", KSTAT_DATA_UINT64 }, 445 { "hash_elements_max", KSTAT_DATA_UINT64 }, 446 { "hash_collisions", KSTAT_DATA_UINT64 }, 447 { "hash_chains", KSTAT_DATA_UINT64 }, 448 { "hash_chain_max", KSTAT_DATA_UINT64 }, 449 { "p", KSTAT_DATA_UINT64 }, 450 { "c", KSTAT_DATA_UINT64 }, 451 { "c_min", KSTAT_DATA_UINT64 }, 452 { "c_max", KSTAT_DATA_UINT64 }, 453 { "size", KSTAT_DATA_UINT64 }, 454 { "hdr_size", KSTAT_DATA_UINT64 }, 455 { "data_size", KSTAT_DATA_UINT64 }, 456 { "other_size", KSTAT_DATA_UINT64 }, 457 { "l2_hits", KSTAT_DATA_UINT64 }, 458 { "l2_misses", KSTAT_DATA_UINT64 }, 459 { "l2_feeds", KSTAT_DATA_UINT64 }, 460 { "l2_rw_clash", KSTAT_DATA_UINT64 }, 461 { "l2_read_bytes", KSTAT_DATA_UINT64 }, 462 { "l2_write_bytes", KSTAT_DATA_UINT64 }, 463 { "l2_writes_sent", KSTAT_DATA_UINT64 }, 464 { "l2_writes_done", KSTAT_DATA_UINT64 }, 465 { "l2_writes_error", KSTAT_DATA_UINT64 }, 466 { "l2_writes_hdr_miss", KSTAT_DATA_UINT64 }, 467 { "l2_evict_lock_retry", KSTAT_DATA_UINT64 }, 468 { "l2_evict_reading", KSTAT_DATA_UINT64 }, 469 { "l2_free_on_write", KSTAT_DATA_UINT64 }, 470 { "l2_abort_lowmem", KSTAT_DATA_UINT64 }, 471 { "l2_cksum_bad", KSTAT_DATA_UINT64 }, 472 { "l2_io_error", KSTAT_DATA_UINT64 }, 473 { "l2_size", KSTAT_DATA_UINT64 }, 474 { "l2_asize", KSTAT_DATA_UINT64 }, 475 { "l2_hdr_size", KSTAT_DATA_UINT64 }, 476 { "l2_compress_successes", KSTAT_DATA_UINT64 }, 477 { "l2_compress_zeros", KSTAT_DATA_UINT64 }, 478 { "l2_compress_failures", KSTAT_DATA_UINT64 }, 479 { "l2_write_trylock_fail", KSTAT_DATA_UINT64 }, 480 { "l2_write_passed_headroom", KSTAT_DATA_UINT64 }, 481 { "l2_write_spa_mismatch", KSTAT_DATA_UINT64 }, 482 { "l2_write_in_l2", KSTAT_DATA_UINT64 }, 483 { "l2_write_io_in_progress", KSTAT_DATA_UINT64 }, 484 { "l2_write_not_cacheable", KSTAT_DATA_UINT64 }, 485 { "l2_write_full", KSTAT_DATA_UINT64 }, 486 { "l2_write_buffer_iter", KSTAT_DATA_UINT64 }, 487 { "l2_write_pios", KSTAT_DATA_UINT64 }, 488 { "l2_write_buffer_bytes_scanned", KSTAT_DATA_UINT64 }, 489 { "l2_write_buffer_list_iter", KSTAT_DATA_UINT64 }, 490 { "l2_write_buffer_list_null_iter", KSTAT_DATA_UINT64 }, 491 { "memory_throttle_count", KSTAT_DATA_UINT64 }, 492 { "duplicate_buffers", KSTAT_DATA_UINT64 }, 493 { "duplicate_buffers_size", KSTAT_DATA_UINT64 }, 494 { "duplicate_reads", KSTAT_DATA_UINT64 } 495}; 496 497#define ARCSTAT(stat) (arc_stats.stat.value.ui64) 498 499#define ARCSTAT_INCR(stat, val) \ 500 atomic_add_64(&arc_stats.stat.value.ui64, (val)) 501 502#define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1) 503#define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1) 504 505#define ARCSTAT_MAX(stat, val) { \ 506 uint64_t m; \ 507 while ((val) > (m = arc_stats.stat.value.ui64) && \ 508 (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \ 509 continue; \ 510} 511 512#define ARCSTAT_MAXSTAT(stat) \ 513 ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64) 514 515/* 516 * We define a macro to allow ARC hits/misses to be easily broken down by 517 * two separate conditions, giving a total of four different subtypes for 518 * each of hits and misses (so eight statistics total). 519 */ 520#define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \ 521 if (cond1) { \ 522 if (cond2) { \ 523 ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \ 524 } else { \ 525 ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \ 526 } \ 527 } else { \ 528 if (cond2) { \ 529 ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \ 530 } else { \ 531 ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\ 532 } \ 533 } 534 535kstat_t *arc_ksp; 536static arc_state_t *arc_anon; 537static arc_state_t *arc_mru; 538static arc_state_t *arc_mru_ghost; 539static arc_state_t *arc_mfu; 540static arc_state_t *arc_mfu_ghost; 541static arc_state_t *arc_l2c_only; 542 543/* 544 * There are several ARC variables that are critical to export as kstats -- 545 * but we don't want to have to grovel around in the kstat whenever we wish to 546 * manipulate them. For these variables, we therefore define them to be in 547 * terms of the statistic variable. This assures that we are not introducing 548 * the possibility of inconsistency by having shadow copies of the variables, 549 * while still allowing the code to be readable. 550 */ 551#define arc_size ARCSTAT(arcstat_size) /* actual total arc size */ 552#define arc_p ARCSTAT(arcstat_p) /* target size of MRU */ 553#define arc_c ARCSTAT(arcstat_c) /* target size of cache */ 554#define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */ 555#define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */ 556 557#define L2ARC_IS_VALID_COMPRESS(_c_) \ 558 ((_c_) == ZIO_COMPRESS_LZ4 || (_c_) == ZIO_COMPRESS_EMPTY) 559 560static int arc_no_grow; /* Don't try to grow cache size */ 561static uint64_t arc_tempreserve; 562static uint64_t arc_loaned_bytes; 563static uint64_t arc_meta_used; 564static uint64_t arc_meta_limit; 565static uint64_t arc_meta_max = 0; 566SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_meta_used, CTLFLAG_RD, &arc_meta_used, 0, 567 "ARC metadata used"); 568SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_meta_limit, CTLFLAG_RW, &arc_meta_limit, 0, 569 "ARC metadata limit"); 570 571typedef struct l2arc_buf_hdr l2arc_buf_hdr_t; 572 573typedef struct arc_callback arc_callback_t; 574 575struct arc_callback { 576 void *acb_private; 577 arc_done_func_t *acb_done; 578 arc_buf_t *acb_buf; 579 zio_t *acb_zio_dummy; 580 arc_callback_t *acb_next; 581}; 582 583typedef struct arc_write_callback arc_write_callback_t; 584 585struct arc_write_callback { 586 void *awcb_private; 587 arc_done_func_t *awcb_ready; 588 arc_done_func_t *awcb_physdone; 589 arc_done_func_t *awcb_done; 590 arc_buf_t *awcb_buf; 591}; 592 593struct arc_buf_hdr { 594 /* protected by hash lock */ 595 dva_t b_dva; 596 uint64_t b_birth; 597 uint64_t b_cksum0; 598 599 kmutex_t b_freeze_lock; 600 zio_cksum_t *b_freeze_cksum; 601 void *b_thawed; 602 603 arc_buf_hdr_t *b_hash_next; 604 arc_buf_t *b_buf; 605 uint32_t b_flags; 606 uint32_t b_datacnt; 607 608 arc_callback_t *b_acb; 609 kcondvar_t b_cv; 610 611 /* immutable */ 612 arc_buf_contents_t b_type; 613 uint64_t b_size; 614 uint64_t b_spa; 615 616 /* protected by arc state mutex */ 617 arc_state_t *b_state; 618 list_node_t b_arc_node; 619 620 /* updated atomically */ 621 clock_t b_arc_access; 622 623 /* self protecting */ 624 refcount_t b_refcnt; 625 626 l2arc_buf_hdr_t *b_l2hdr; 627 list_node_t b_l2node; 628}; 629 630static arc_buf_t *arc_eviction_list; 631static kmutex_t arc_eviction_mtx; 632static arc_buf_hdr_t arc_eviction_hdr; 633static void arc_get_data_buf(arc_buf_t *buf); 634static void arc_access(arc_buf_hdr_t *buf, kmutex_t *hash_lock); 635static int arc_evict_needed(arc_buf_contents_t type); 636static void arc_evict_ghost(arc_state_t *state, uint64_t spa, int64_t bytes); 637#ifdef illumos 638static void arc_buf_watch(arc_buf_t *buf); 639#endif /* illumos */ 640 641static boolean_t l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *ab); 642 643#define GHOST_STATE(state) \ 644 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \ 645 (state) == arc_l2c_only) 646 647/* 648 * Private ARC flags. These flags are private ARC only flags that will show up 649 * in b_flags in the arc_hdr_buf_t. Some flags are publicly declared, and can 650 * be passed in as arc_flags in things like arc_read. However, these flags 651 * should never be passed and should only be set by ARC code. When adding new 652 * public flags, make sure not to smash the private ones. 653 */ 654 655#define ARC_IN_HASH_TABLE (1 << 9) /* this buffer is hashed */ 656#define ARC_IO_IN_PROGRESS (1 << 10) /* I/O in progress for buf */ 657#define ARC_IO_ERROR (1 << 11) /* I/O failed for buf */ 658#define ARC_FREED_IN_READ (1 << 12) /* buf freed while in read */ 659#define ARC_BUF_AVAILABLE (1 << 13) /* block not in active use */ 660#define ARC_INDIRECT (1 << 14) /* this is an indirect block */ 661#define ARC_FREE_IN_PROGRESS (1 << 15) /* hdr about to be freed */ 662#define ARC_L2_WRITING (1 << 16) /* L2ARC write in progress */ 663#define ARC_L2_EVICTED (1 << 17) /* evicted during I/O */ 664#define ARC_L2_WRITE_HEAD (1 << 18) /* head of write list */ 665 666#define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_IN_HASH_TABLE) 667#define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_IO_IN_PROGRESS) 668#define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_IO_ERROR) 669#define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_PREFETCH) 670#define HDR_FREED_IN_READ(hdr) ((hdr)->b_flags & ARC_FREED_IN_READ) 671#define HDR_BUF_AVAILABLE(hdr) ((hdr)->b_flags & ARC_BUF_AVAILABLE) 672#define HDR_FREE_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FREE_IN_PROGRESS) 673#define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_L2CACHE) 674#define HDR_L2_READING(hdr) ((hdr)->b_flags & ARC_IO_IN_PROGRESS && \ 675 (hdr)->b_l2hdr != NULL) 676#define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_L2_WRITING) 677#define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_L2_EVICTED) 678#define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_L2_WRITE_HEAD) 679 680/* 681 * Other sizes 682 */ 683 684#define HDR_SIZE ((int64_t)sizeof (arc_buf_hdr_t)) 685#define L2HDR_SIZE ((int64_t)sizeof (l2arc_buf_hdr_t)) 686 687/* 688 * Hash table routines 689 */ 690 691#define HT_LOCK_PAD CACHE_LINE_SIZE 692 693struct ht_lock { 694 kmutex_t ht_lock; 695#ifdef _KERNEL 696 unsigned char pad[(HT_LOCK_PAD - sizeof (kmutex_t))]; 697#endif 698}; 699 700#define BUF_LOCKS 256 701typedef struct buf_hash_table { 702 uint64_t ht_mask; 703 arc_buf_hdr_t **ht_table; 704 struct ht_lock ht_locks[BUF_LOCKS] __aligned(CACHE_LINE_SIZE); 705} buf_hash_table_t; 706 707static buf_hash_table_t buf_hash_table; 708 709#define BUF_HASH_INDEX(spa, dva, birth) \ 710 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask) 711#define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)]) 712#define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock)) 713#define HDR_LOCK(hdr) \ 714 (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth))) 715 716uint64_t zfs_crc64_table[256]; 717 718/* 719 * Level 2 ARC 720 */ 721 722#define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */ 723#define L2ARC_HEADROOM 2 /* num of writes */ 724/* 725 * If we discover during ARC scan any buffers to be compressed, we boost 726 * our headroom for the next scanning cycle by this percentage multiple. 727 */ 728#define L2ARC_HEADROOM_BOOST 200 729#define L2ARC_FEED_SECS 1 /* caching interval secs */ 730#define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */ 731 732#define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent) 733#define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done) 734 735/* L2ARC Performance Tunables */ 736uint64_t l2arc_write_max = L2ARC_WRITE_SIZE; /* default max write size */ 737uint64_t l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra write during warmup */ 738uint64_t l2arc_headroom = L2ARC_HEADROOM; /* number of dev writes */ 739uint64_t l2arc_headroom_boost = L2ARC_HEADROOM_BOOST; 740uint64_t l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */ 741uint64_t l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval milliseconds */ 742boolean_t l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */ 743boolean_t l2arc_feed_again = B_TRUE; /* turbo warmup */ 744boolean_t l2arc_norw = B_TRUE; /* no reads during writes */ 745 746SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_max, CTLFLAG_RW, 747 &l2arc_write_max, 0, "max write size"); 748SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_boost, CTLFLAG_RW, 749 &l2arc_write_boost, 0, "extra write during warmup"); 750SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_headroom, CTLFLAG_RW, 751 &l2arc_headroom, 0, "number of dev writes"); 752SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_secs, CTLFLAG_RW, 753 &l2arc_feed_secs, 0, "interval seconds"); 754SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_min_ms, CTLFLAG_RW, 755 &l2arc_feed_min_ms, 0, "min interval milliseconds"); 756 757SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_noprefetch, CTLFLAG_RW, 758 &l2arc_noprefetch, 0, "don't cache prefetch bufs"); 759SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_feed_again, CTLFLAG_RW, 760 &l2arc_feed_again, 0, "turbo warmup"); 761SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_norw, CTLFLAG_RW, 762 &l2arc_norw, 0, "no reads during writes"); 763 764SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_size, CTLFLAG_RD, 765 &ARC_anon.arcs_size, 0, "size of anonymous state"); 766SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_metadata_lsize, CTLFLAG_RD, 767 &ARC_anon.arcs_lsize[ARC_BUFC_METADATA], 0, "size of anonymous state"); 768SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_data_lsize, CTLFLAG_RD, 769 &ARC_anon.arcs_lsize[ARC_BUFC_DATA], 0, "size of anonymous state"); 770 771SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_size, CTLFLAG_RD, 772 &ARC_mru.arcs_size, 0, "size of mru state"); 773SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_metadata_lsize, CTLFLAG_RD, 774 &ARC_mru.arcs_lsize[ARC_BUFC_METADATA], 0, "size of metadata in mru state"); 775SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_data_lsize, CTLFLAG_RD, 776 &ARC_mru.arcs_lsize[ARC_BUFC_DATA], 0, "size of data in mru state"); 777 778SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_size, CTLFLAG_RD, 779 &ARC_mru_ghost.arcs_size, 0, "size of mru ghost state"); 780SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_metadata_lsize, CTLFLAG_RD, 781 &ARC_mru_ghost.arcs_lsize[ARC_BUFC_METADATA], 0, 782 "size of metadata in mru ghost state"); 783SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_data_lsize, CTLFLAG_RD, 784 &ARC_mru_ghost.arcs_lsize[ARC_BUFC_DATA], 0, 785 "size of data in mru ghost state"); 786 787SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_size, CTLFLAG_RD, 788 &ARC_mfu.arcs_size, 0, "size of mfu state"); 789SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_metadata_lsize, CTLFLAG_RD, 790 &ARC_mfu.arcs_lsize[ARC_BUFC_METADATA], 0, "size of metadata in mfu state"); 791SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_data_lsize, CTLFLAG_RD, 792 &ARC_mfu.arcs_lsize[ARC_BUFC_DATA], 0, "size of data in mfu state"); 793 794SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_size, CTLFLAG_RD, 795 &ARC_mfu_ghost.arcs_size, 0, "size of mfu ghost state"); 796SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_metadata_lsize, CTLFLAG_RD, 797 &ARC_mfu_ghost.arcs_lsize[ARC_BUFC_METADATA], 0, 798 "size of metadata in mfu ghost state"); 799SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_data_lsize, CTLFLAG_RD, 800 &ARC_mfu_ghost.arcs_lsize[ARC_BUFC_DATA], 0, 801 "size of data in mfu ghost state"); 802 803SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2c_only_size, CTLFLAG_RD, 804 &ARC_l2c_only.arcs_size, 0, "size of mru state"); 805 806/* 807 * L2ARC Internals 808 */ 809typedef struct l2arc_dev { 810 vdev_t *l2ad_vdev; /* vdev */ 811 spa_t *l2ad_spa; /* spa */ 812 uint64_t l2ad_hand; /* next write location */ 813 uint64_t l2ad_start; /* first addr on device */ 814 uint64_t l2ad_end; /* last addr on device */ 815 uint64_t l2ad_evict; /* last addr eviction reached */ 816 boolean_t l2ad_first; /* first sweep through */ 817 boolean_t l2ad_writing; /* currently writing */ 818 list_t *l2ad_buflist; /* buffer list */ 819 list_node_t l2ad_node; /* device list node */ 820} l2arc_dev_t; 821 822static list_t L2ARC_dev_list; /* device list */ 823static list_t *l2arc_dev_list; /* device list pointer */ 824static kmutex_t l2arc_dev_mtx; /* device list mutex */ 825static l2arc_dev_t *l2arc_dev_last; /* last device used */ 826static kmutex_t l2arc_buflist_mtx; /* mutex for all buflists */ 827static list_t L2ARC_free_on_write; /* free after write buf list */ 828static list_t *l2arc_free_on_write; /* free after write list ptr */ 829static kmutex_t l2arc_free_on_write_mtx; /* mutex for list */ 830static uint64_t l2arc_ndev; /* number of devices */ 831 832typedef struct l2arc_read_callback { 833 arc_buf_t *l2rcb_buf; /* read buffer */ 834 spa_t *l2rcb_spa; /* spa */ 835 blkptr_t l2rcb_bp; /* original blkptr */ 836 zbookmark_phys_t l2rcb_zb; /* original bookmark */ 837 int l2rcb_flags; /* original flags */ 838 enum zio_compress l2rcb_compress; /* applied compress */ 839} l2arc_read_callback_t; 840 841typedef struct l2arc_write_callback { 842 l2arc_dev_t *l2wcb_dev; /* device info */ 843 arc_buf_hdr_t *l2wcb_head; /* head of write buflist */ 844} l2arc_write_callback_t; 845 846struct l2arc_buf_hdr { 847 /* protected by arc_buf_hdr mutex */ 848 l2arc_dev_t *b_dev; /* L2ARC device */ 849 uint64_t b_daddr; /* disk address, offset byte */ 850 /* compression applied to buffer data */ 851 enum zio_compress b_compress; 852 /* real alloc'd buffer size depending on b_compress applied */ 853 int b_asize; 854 /* temporary buffer holder for in-flight compressed data */ 855 void *b_tmp_cdata; 856}; 857 858typedef struct l2arc_data_free { 859 /* protected by l2arc_free_on_write_mtx */ 860 void *l2df_data; 861 size_t l2df_size; 862 void (*l2df_func)(void *, size_t); 863 list_node_t l2df_list_node; 864} l2arc_data_free_t; 865 866static kmutex_t l2arc_feed_thr_lock; 867static kcondvar_t l2arc_feed_thr_cv; 868static uint8_t l2arc_thread_exit; 869 870static void l2arc_read_done(zio_t *zio); 871static void l2arc_hdr_stat_add(void); 872static void l2arc_hdr_stat_remove(void); 873 874static boolean_t l2arc_compress_buf(l2arc_buf_hdr_t *l2hdr); 875static void l2arc_decompress_zio(zio_t *zio, arc_buf_hdr_t *hdr, 876 enum zio_compress c); 877static void l2arc_release_cdata_buf(arc_buf_hdr_t *ab); 878 879static uint64_t 880buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth) 881{ 882 uint8_t *vdva = (uint8_t *)dva; 883 uint64_t crc = -1ULL; 884 int i; 885 886 ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY); 887 888 for (i = 0; i < sizeof (dva_t); i++) 889 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ vdva[i]) & 0xFF]; 890 891 crc ^= (spa>>8) ^ birth; 892 893 return (crc); 894} 895 896#define BUF_EMPTY(buf) \ 897 ((buf)->b_dva.dva_word[0] == 0 && \ 898 (buf)->b_dva.dva_word[1] == 0 && \ 899 (buf)->b_cksum0 == 0) 900 901#define BUF_EQUAL(spa, dva, birth, buf) \ 902 ((buf)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \ 903 ((buf)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \ 904 ((buf)->b_birth == birth) && ((buf)->b_spa == spa) 905 906static void 907buf_discard_identity(arc_buf_hdr_t *hdr) 908{ 909 hdr->b_dva.dva_word[0] = 0; 910 hdr->b_dva.dva_word[1] = 0; 911 hdr->b_birth = 0; 912 hdr->b_cksum0 = 0; 913} 914 915static arc_buf_hdr_t * 916buf_hash_find(uint64_t spa, const blkptr_t *bp, kmutex_t **lockp) 917{ 918 const dva_t *dva = BP_IDENTITY(bp); 919 uint64_t birth = BP_PHYSICAL_BIRTH(bp); 920 uint64_t idx = BUF_HASH_INDEX(spa, dva, birth); 921 kmutex_t *hash_lock = BUF_HASH_LOCK(idx); 922 arc_buf_hdr_t *buf; 923 924 mutex_enter(hash_lock); 925 for (buf = buf_hash_table.ht_table[idx]; buf != NULL; 926 buf = buf->b_hash_next) { 927 if (BUF_EQUAL(spa, dva, birth, buf)) { 928 *lockp = hash_lock; 929 return (buf); 930 } 931 } 932 mutex_exit(hash_lock); 933 *lockp = NULL; 934 return (NULL); 935} 936 937/* 938 * Insert an entry into the hash table. If there is already an element 939 * equal to elem in the hash table, then the already existing element 940 * will be returned and the new element will not be inserted. 941 * Otherwise returns NULL. 942 */ 943static arc_buf_hdr_t * 944buf_hash_insert(arc_buf_hdr_t *buf, kmutex_t **lockp) 945{ 946 uint64_t idx = BUF_HASH_INDEX(buf->b_spa, &buf->b_dva, buf->b_birth); 947 kmutex_t *hash_lock = BUF_HASH_LOCK(idx); 948 arc_buf_hdr_t *fbuf; 949 uint32_t i; 950 951 ASSERT(!DVA_IS_EMPTY(&buf->b_dva)); 952 ASSERT(buf->b_birth != 0); 953 ASSERT(!HDR_IN_HASH_TABLE(buf)); 954 *lockp = hash_lock; 955 mutex_enter(hash_lock); 956 for (fbuf = buf_hash_table.ht_table[idx], i = 0; fbuf != NULL; 957 fbuf = fbuf->b_hash_next, i++) { 958 if (BUF_EQUAL(buf->b_spa, &buf->b_dva, buf->b_birth, fbuf)) 959 return (fbuf); 960 } 961 962 buf->b_hash_next = buf_hash_table.ht_table[idx]; 963 buf_hash_table.ht_table[idx] = buf; 964 buf->b_flags |= ARC_IN_HASH_TABLE; 965 966 /* collect some hash table performance data */ 967 if (i > 0) { 968 ARCSTAT_BUMP(arcstat_hash_collisions); 969 if (i == 1) 970 ARCSTAT_BUMP(arcstat_hash_chains); 971 972 ARCSTAT_MAX(arcstat_hash_chain_max, i); 973 } 974 975 ARCSTAT_BUMP(arcstat_hash_elements); 976 ARCSTAT_MAXSTAT(arcstat_hash_elements); 977 978 return (NULL); 979} 980 981static void 982buf_hash_remove(arc_buf_hdr_t *buf) 983{ 984 arc_buf_hdr_t *fbuf, **bufp; 985 uint64_t idx = BUF_HASH_INDEX(buf->b_spa, &buf->b_dva, buf->b_birth); 986 987 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx))); 988 ASSERT(HDR_IN_HASH_TABLE(buf)); 989 990 bufp = &buf_hash_table.ht_table[idx]; 991 while ((fbuf = *bufp) != buf) { 992 ASSERT(fbuf != NULL); 993 bufp = &fbuf->b_hash_next; 994 } 995 *bufp = buf->b_hash_next; 996 buf->b_hash_next = NULL; 997 buf->b_flags &= ~ARC_IN_HASH_TABLE; 998 999 /* collect some hash table performance data */ 1000 ARCSTAT_BUMPDOWN(arcstat_hash_elements); 1001 1002 if (buf_hash_table.ht_table[idx] && 1003 buf_hash_table.ht_table[idx]->b_hash_next == NULL) 1004 ARCSTAT_BUMPDOWN(arcstat_hash_chains); 1005} 1006 1007/* 1008 * Global data structures and functions for the buf kmem cache. 1009 */ 1010static kmem_cache_t *hdr_cache; 1011static kmem_cache_t *buf_cache; 1012 1013static void 1014buf_fini(void) 1015{ 1016 int i; 1017 1018 kmem_free(buf_hash_table.ht_table, 1019 (buf_hash_table.ht_mask + 1) * sizeof (void *)); 1020 for (i = 0; i < BUF_LOCKS; i++) 1021 mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock); 1022 kmem_cache_destroy(hdr_cache); 1023 kmem_cache_destroy(buf_cache); 1024} 1025 1026/* 1027 * Constructor callback - called when the cache is empty 1028 * and a new buf is requested. 1029 */ 1030/* ARGSUSED */ 1031static int 1032hdr_cons(void *vbuf, void *unused, int kmflag) 1033{ 1034 arc_buf_hdr_t *buf = vbuf; 1035 1036 bzero(buf, sizeof (arc_buf_hdr_t)); 1037 refcount_create(&buf->b_refcnt); 1038 cv_init(&buf->b_cv, NULL, CV_DEFAULT, NULL); 1039 mutex_init(&buf->b_freeze_lock, NULL, MUTEX_DEFAULT, NULL); 1040 arc_space_consume(sizeof (arc_buf_hdr_t), ARC_SPACE_HDRS); 1041 1042 return (0); 1043} 1044 1045/* ARGSUSED */ 1046static int 1047buf_cons(void *vbuf, void *unused, int kmflag) 1048{ 1049 arc_buf_t *buf = vbuf; 1050 1051 bzero(buf, sizeof (arc_buf_t)); 1052 mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL); 1053 arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS); 1054 1055 return (0); 1056} 1057 1058/* 1059 * Destructor callback - called when a cached buf is 1060 * no longer required. 1061 */ 1062/* ARGSUSED */ 1063static void 1064hdr_dest(void *vbuf, void *unused) 1065{ 1066 arc_buf_hdr_t *buf = vbuf; 1067 1068 ASSERT(BUF_EMPTY(buf)); 1069 refcount_destroy(&buf->b_refcnt); 1070 cv_destroy(&buf->b_cv); 1071 mutex_destroy(&buf->b_freeze_lock); 1072 arc_space_return(sizeof (arc_buf_hdr_t), ARC_SPACE_HDRS); 1073} 1074 1075/* ARGSUSED */ 1076static void 1077buf_dest(void *vbuf, void *unused) 1078{ 1079 arc_buf_t *buf = vbuf; 1080 1081 mutex_destroy(&buf->b_evict_lock); 1082 arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS); 1083} 1084 1085/* 1086 * Reclaim callback -- invoked when memory is low. 1087 */ 1088/* ARGSUSED */ 1089static void 1090hdr_recl(void *unused) 1091{ 1092 dprintf("hdr_recl called\n"); 1093 /* 1094 * umem calls the reclaim func when we destroy the buf cache, 1095 * which is after we do arc_fini(). 1096 */ 1097 if (!arc_dead) 1098 cv_signal(&arc_reclaim_thr_cv); 1099} 1100 1101static void 1102buf_init(void) 1103{ 1104 uint64_t *ct; 1105 uint64_t hsize = 1ULL << 12; 1106 int i, j; 1107 1108 /* 1109 * The hash table is big enough to fill all of physical memory 1110 * with an average block size of zfs_arc_average_blocksize (default 8K). 1111 * By default, the table will take up 1112 * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers). 1113 */ 1114 while (hsize * zfs_arc_average_blocksize < (uint64_t)physmem * PAGESIZE) 1115 hsize <<= 1; 1116retry: 1117 buf_hash_table.ht_mask = hsize - 1; 1118 buf_hash_table.ht_table = 1119 kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP); 1120 if (buf_hash_table.ht_table == NULL) { 1121 ASSERT(hsize > (1ULL << 8)); 1122 hsize >>= 1; 1123 goto retry; 1124 } 1125 1126 hdr_cache = kmem_cache_create("arc_buf_hdr_t", sizeof (arc_buf_hdr_t), 1127 0, hdr_cons, hdr_dest, hdr_recl, NULL, NULL, 0); 1128 buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t), 1129 0, buf_cons, buf_dest, NULL, NULL, NULL, 0); 1130 1131 for (i = 0; i < 256; i++) 1132 for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--) 1133 *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY); 1134 1135 for (i = 0; i < BUF_LOCKS; i++) { 1136 mutex_init(&buf_hash_table.ht_locks[i].ht_lock, 1137 NULL, MUTEX_DEFAULT, NULL); 1138 } 1139} 1140 1141#define ARC_MINTIME (hz>>4) /* 62 ms */ 1142 1143static void 1144arc_cksum_verify(arc_buf_t *buf) 1145{ 1146 zio_cksum_t zc; 1147 1148 if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 1149 return; 1150 1151 mutex_enter(&buf->b_hdr->b_freeze_lock); 1152 if (buf->b_hdr->b_freeze_cksum == NULL || 1153 (buf->b_hdr->b_flags & ARC_IO_ERROR)) { 1154 mutex_exit(&buf->b_hdr->b_freeze_lock); 1155 return; 1156 } 1157 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc); 1158 if (!ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc)) 1159 panic("buffer modified while frozen!"); 1160 mutex_exit(&buf->b_hdr->b_freeze_lock); 1161} 1162 1163static int 1164arc_cksum_equal(arc_buf_t *buf) 1165{ 1166 zio_cksum_t zc; 1167 int equal; 1168 1169 mutex_enter(&buf->b_hdr->b_freeze_lock); 1170 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc); 1171 equal = ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc); 1172 mutex_exit(&buf->b_hdr->b_freeze_lock); 1173 1174 return (equal); 1175} 1176 1177static void 1178arc_cksum_compute(arc_buf_t *buf, boolean_t force) 1179{ 1180 if (!force && !(zfs_flags & ZFS_DEBUG_MODIFY)) 1181 return; 1182 1183 mutex_enter(&buf->b_hdr->b_freeze_lock); 1184 if (buf->b_hdr->b_freeze_cksum != NULL) { 1185 mutex_exit(&buf->b_hdr->b_freeze_lock); 1186 return; 1187 } 1188 buf->b_hdr->b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t), KM_SLEEP); 1189 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, 1190 buf->b_hdr->b_freeze_cksum); 1191 mutex_exit(&buf->b_hdr->b_freeze_lock); 1192#ifdef illumos 1193 arc_buf_watch(buf); 1194#endif /* illumos */ 1195} 1196 1197#ifdef illumos 1198#ifndef _KERNEL 1199typedef struct procctl { 1200 long cmd; 1201 prwatch_t prwatch; 1202} procctl_t; 1203#endif 1204 1205/* ARGSUSED */ 1206static void 1207arc_buf_unwatch(arc_buf_t *buf) 1208{ 1209#ifndef _KERNEL 1210 if (arc_watch) { 1211 int result; 1212 procctl_t ctl; 1213 ctl.cmd = PCWATCH; 1214 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data; 1215 ctl.prwatch.pr_size = 0; 1216 ctl.prwatch.pr_wflags = 0; 1217 result = write(arc_procfd, &ctl, sizeof (ctl)); 1218 ASSERT3U(result, ==, sizeof (ctl)); 1219 } 1220#endif 1221} 1222 1223/* ARGSUSED */ 1224static void 1225arc_buf_watch(arc_buf_t *buf) 1226{ 1227#ifndef _KERNEL 1228 if (arc_watch) { 1229 int result; 1230 procctl_t ctl; 1231 ctl.cmd = PCWATCH; 1232 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data; 1233 ctl.prwatch.pr_size = buf->b_hdr->b_size; 1234 ctl.prwatch.pr_wflags = WA_WRITE; 1235 result = write(arc_procfd, &ctl, sizeof (ctl)); 1236 ASSERT3U(result, ==, sizeof (ctl)); 1237 } 1238#endif 1239} 1240#endif /* illumos */ 1241 1242void 1243arc_buf_thaw(arc_buf_t *buf) 1244{ 1245 if (zfs_flags & ZFS_DEBUG_MODIFY) { 1246 if (buf->b_hdr->b_state != arc_anon) 1247 panic("modifying non-anon buffer!"); 1248 if (buf->b_hdr->b_flags & ARC_IO_IN_PROGRESS) 1249 panic("modifying buffer while i/o in progress!"); 1250 arc_cksum_verify(buf); 1251 } 1252 1253 mutex_enter(&buf->b_hdr->b_freeze_lock); 1254 if (buf->b_hdr->b_freeze_cksum != NULL) { 1255 kmem_free(buf->b_hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 1256 buf->b_hdr->b_freeze_cksum = NULL; 1257 } 1258 1259 if (zfs_flags & ZFS_DEBUG_MODIFY) { 1260 if (buf->b_hdr->b_thawed) 1261 kmem_free(buf->b_hdr->b_thawed, 1); 1262 buf->b_hdr->b_thawed = kmem_alloc(1, KM_SLEEP); 1263 } 1264 1265 mutex_exit(&buf->b_hdr->b_freeze_lock); 1266 1267#ifdef illumos 1268 arc_buf_unwatch(buf); 1269#endif /* illumos */ 1270} 1271 1272void 1273arc_buf_freeze(arc_buf_t *buf) 1274{ 1275 kmutex_t *hash_lock; 1276 1277 if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 1278 return; 1279 1280 hash_lock = HDR_LOCK(buf->b_hdr); 1281 mutex_enter(hash_lock); 1282 1283 ASSERT(buf->b_hdr->b_freeze_cksum != NULL || 1284 buf->b_hdr->b_state == arc_anon); 1285 arc_cksum_compute(buf, B_FALSE); 1286 mutex_exit(hash_lock); 1287 1288} 1289 1290static void 1291get_buf_info(arc_buf_hdr_t *ab, arc_state_t *state, list_t **list, kmutex_t **lock) 1292{ 1293 uint64_t buf_hashid = buf_hash(ab->b_spa, &ab->b_dva, ab->b_birth); 1294 1295 if (ab->b_type == ARC_BUFC_METADATA) 1296 buf_hashid &= (ARC_BUFC_NUMMETADATALISTS - 1); 1297 else { 1298 buf_hashid &= (ARC_BUFC_NUMDATALISTS - 1); 1299 buf_hashid += ARC_BUFC_NUMMETADATALISTS; 1300 } 1301 1302 *list = &state->arcs_lists[buf_hashid]; 1303 *lock = ARCS_LOCK(state, buf_hashid); 1304} 1305 1306 1307static void 1308add_reference(arc_buf_hdr_t *ab, kmutex_t *hash_lock, void *tag) 1309{ 1310 ASSERT(MUTEX_HELD(hash_lock)); 1311 1312 if ((refcount_add(&ab->b_refcnt, tag) == 1) && 1313 (ab->b_state != arc_anon)) { 1314 uint64_t delta = ab->b_size * ab->b_datacnt; 1315 uint64_t *size = &ab->b_state->arcs_lsize[ab->b_type]; 1316 list_t *list; 1317 kmutex_t *lock; 1318 1319 get_buf_info(ab, ab->b_state, &list, &lock); 1320 ASSERT(!MUTEX_HELD(lock)); 1321 mutex_enter(lock); 1322 ASSERT(list_link_active(&ab->b_arc_node)); 1323 list_remove(list, ab); 1324 if (GHOST_STATE(ab->b_state)) { 1325 ASSERT0(ab->b_datacnt); 1326 ASSERT3P(ab->b_buf, ==, NULL); 1327 delta = ab->b_size; 1328 } 1329 ASSERT(delta > 0); 1330 ASSERT3U(*size, >=, delta); 1331 atomic_add_64(size, -delta); 1332 mutex_exit(lock); 1333 /* remove the prefetch flag if we get a reference */ 1334 if (ab->b_flags & ARC_PREFETCH) 1335 ab->b_flags &= ~ARC_PREFETCH; 1336 } 1337} 1338 1339static int 1340remove_reference(arc_buf_hdr_t *ab, kmutex_t *hash_lock, void *tag) 1341{ 1342 int cnt; 1343 arc_state_t *state = ab->b_state; 1344 1345 ASSERT(state == arc_anon || MUTEX_HELD(hash_lock)); 1346 ASSERT(!GHOST_STATE(state)); 1347 1348 if (((cnt = refcount_remove(&ab->b_refcnt, tag)) == 0) && 1349 (state != arc_anon)) { 1350 uint64_t *size = &state->arcs_lsize[ab->b_type]; 1351 list_t *list; 1352 kmutex_t *lock; 1353 1354 get_buf_info(ab, state, &list, &lock); 1355 ASSERT(!MUTEX_HELD(lock)); 1356 mutex_enter(lock); 1357 ASSERT(!list_link_active(&ab->b_arc_node)); 1358 list_insert_head(list, ab); 1359 ASSERT(ab->b_datacnt > 0); 1360 atomic_add_64(size, ab->b_size * ab->b_datacnt); 1361 mutex_exit(lock); 1362 } 1363 return (cnt); 1364} 1365 1366/* 1367 * Move the supplied buffer to the indicated state. The mutex 1368 * for the buffer must be held by the caller. 1369 */ 1370static void 1371arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *ab, kmutex_t *hash_lock) 1372{ 1373 arc_state_t *old_state = ab->b_state; 1374 int64_t refcnt = refcount_count(&ab->b_refcnt); 1375 uint64_t from_delta, to_delta; 1376 list_t *list; 1377 kmutex_t *lock; 1378 1379 ASSERT(MUTEX_HELD(hash_lock)); 1380 ASSERT3P(new_state, !=, old_state); 1381 ASSERT(refcnt == 0 || ab->b_datacnt > 0); 1382 ASSERT(ab->b_datacnt == 0 || !GHOST_STATE(new_state)); 1383 ASSERT(ab->b_datacnt <= 1 || old_state != arc_anon); 1384 1385 from_delta = to_delta = ab->b_datacnt * ab->b_size; 1386 1387 /* 1388 * If this buffer is evictable, transfer it from the 1389 * old state list to the new state list. 1390 */ 1391 if (refcnt == 0) { 1392 if (old_state != arc_anon) { 1393 int use_mutex; 1394 uint64_t *size = &old_state->arcs_lsize[ab->b_type]; 1395 1396 get_buf_info(ab, old_state, &list, &lock); 1397 use_mutex = !MUTEX_HELD(lock); 1398 if (use_mutex) 1399 mutex_enter(lock); 1400 1401 ASSERT(list_link_active(&ab->b_arc_node)); 1402 list_remove(list, ab); 1403 1404 /* 1405 * If prefetching out of the ghost cache, 1406 * we will have a non-zero datacnt. 1407 */ 1408 if (GHOST_STATE(old_state) && ab->b_datacnt == 0) { 1409 /* ghost elements have a ghost size */ 1410 ASSERT(ab->b_buf == NULL); 1411 from_delta = ab->b_size; 1412 } 1413 ASSERT3U(*size, >=, from_delta); 1414 atomic_add_64(size, -from_delta); 1415 1416 if (use_mutex) 1417 mutex_exit(lock); 1418 } 1419 if (new_state != arc_anon) { 1420 int use_mutex; 1421 uint64_t *size = &new_state->arcs_lsize[ab->b_type]; 1422 1423 get_buf_info(ab, new_state, &list, &lock); 1424 use_mutex = !MUTEX_HELD(lock); 1425 if (use_mutex) 1426 mutex_enter(lock); 1427 1428 list_insert_head(list, ab); 1429 1430 /* ghost elements have a ghost size */ 1431 if (GHOST_STATE(new_state)) { 1432 ASSERT(ab->b_datacnt == 0); 1433 ASSERT(ab->b_buf == NULL); 1434 to_delta = ab->b_size; 1435 } 1436 atomic_add_64(size, to_delta); 1437 1438 if (use_mutex) 1439 mutex_exit(lock); 1440 } 1441 } 1442 1443 ASSERT(!BUF_EMPTY(ab)); 1444 if (new_state == arc_anon && HDR_IN_HASH_TABLE(ab)) 1445 buf_hash_remove(ab); 1446 1447 /* adjust state sizes */ 1448 if (to_delta) 1449 atomic_add_64(&new_state->arcs_size, to_delta); 1450 if (from_delta) { 1451 ASSERT3U(old_state->arcs_size, >=, from_delta); 1452 atomic_add_64(&old_state->arcs_size, -from_delta); 1453 } 1454 ab->b_state = new_state; 1455 1456 /* adjust l2arc hdr stats */ 1457 if (new_state == arc_l2c_only) 1458 l2arc_hdr_stat_add(); 1459 else if (old_state == arc_l2c_only) 1460 l2arc_hdr_stat_remove(); 1461} 1462 1463void 1464arc_space_consume(uint64_t space, arc_space_type_t type) 1465{ 1466 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES); 1467 1468 switch (type) { 1469 case ARC_SPACE_DATA: 1470 ARCSTAT_INCR(arcstat_data_size, space); 1471 break; 1472 case ARC_SPACE_OTHER: 1473 ARCSTAT_INCR(arcstat_other_size, space); 1474 break; 1475 case ARC_SPACE_HDRS: 1476 ARCSTAT_INCR(arcstat_hdr_size, space); 1477 break; 1478 case ARC_SPACE_L2HDRS: 1479 ARCSTAT_INCR(arcstat_l2_hdr_size, space); 1480 break; 1481 } 1482 1483 atomic_add_64(&arc_meta_used, space); 1484 atomic_add_64(&arc_size, space); 1485} 1486 1487void 1488arc_space_return(uint64_t space, arc_space_type_t type) 1489{ 1490 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES); 1491 1492 switch (type) { 1493 case ARC_SPACE_DATA: 1494 ARCSTAT_INCR(arcstat_data_size, -space); 1495 break; 1496 case ARC_SPACE_OTHER: 1497 ARCSTAT_INCR(arcstat_other_size, -space); 1498 break; 1499 case ARC_SPACE_HDRS: 1500 ARCSTAT_INCR(arcstat_hdr_size, -space); 1501 break; 1502 case ARC_SPACE_L2HDRS: 1503 ARCSTAT_INCR(arcstat_l2_hdr_size, -space); 1504 break; 1505 } 1506 1507 ASSERT(arc_meta_used >= space); 1508 if (arc_meta_max < arc_meta_used) 1509 arc_meta_max = arc_meta_used; 1510 atomic_add_64(&arc_meta_used, -space); 1511 ASSERT(arc_size >= space); 1512 atomic_add_64(&arc_size, -space); 1513} 1514 1515arc_buf_t * 1516arc_buf_alloc(spa_t *spa, int size, void *tag, arc_buf_contents_t type) 1517{ 1518 arc_buf_hdr_t *hdr; 1519 arc_buf_t *buf; 1520 1521 ASSERT3U(size, >, 0); 1522 hdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE); 1523 ASSERT(BUF_EMPTY(hdr)); 1524 hdr->b_size = size; 1525 hdr->b_type = type; 1526 hdr->b_spa = spa_load_guid(spa); 1527 hdr->b_state = arc_anon; 1528 hdr->b_arc_access = 0; 1529 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 1530 buf->b_hdr = hdr; 1531 buf->b_data = NULL; 1532 buf->b_efunc = NULL; 1533 buf->b_private = NULL; 1534 buf->b_next = NULL; 1535 hdr->b_buf = buf; 1536 arc_get_data_buf(buf); 1537 hdr->b_datacnt = 1; 1538 hdr->b_flags = 0; 1539 ASSERT(refcount_is_zero(&hdr->b_refcnt)); 1540 (void) refcount_add(&hdr->b_refcnt, tag); 1541 1542 return (buf); 1543} 1544 1545static char *arc_onloan_tag = "onloan"; 1546 1547/* 1548 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in 1549 * flight data by arc_tempreserve_space() until they are "returned". Loaned 1550 * buffers must be returned to the arc before they can be used by the DMU or 1551 * freed. 1552 */ 1553arc_buf_t * 1554arc_loan_buf(spa_t *spa, int size) 1555{ 1556 arc_buf_t *buf; 1557 1558 buf = arc_buf_alloc(spa, size, arc_onloan_tag, ARC_BUFC_DATA); 1559 1560 atomic_add_64(&arc_loaned_bytes, size); 1561 return (buf); 1562} 1563 1564/* 1565 * Return a loaned arc buffer to the arc. 1566 */ 1567void 1568arc_return_buf(arc_buf_t *buf, void *tag) 1569{ 1570 arc_buf_hdr_t *hdr = buf->b_hdr; 1571 1572 ASSERT(buf->b_data != NULL); 1573 (void) refcount_add(&hdr->b_refcnt, tag); 1574 (void) refcount_remove(&hdr->b_refcnt, arc_onloan_tag); 1575 1576 atomic_add_64(&arc_loaned_bytes, -hdr->b_size); 1577} 1578 1579/* Detach an arc_buf from a dbuf (tag) */ 1580void 1581arc_loan_inuse_buf(arc_buf_t *buf, void *tag) 1582{ 1583 arc_buf_hdr_t *hdr; 1584 1585 ASSERT(buf->b_data != NULL); 1586 hdr = buf->b_hdr; 1587 (void) refcount_add(&hdr->b_refcnt, arc_onloan_tag); 1588 (void) refcount_remove(&hdr->b_refcnt, tag); 1589 buf->b_efunc = NULL; 1590 buf->b_private = NULL; 1591 1592 atomic_add_64(&arc_loaned_bytes, hdr->b_size); 1593} 1594 1595static arc_buf_t * 1596arc_buf_clone(arc_buf_t *from) 1597{ 1598 arc_buf_t *buf; 1599 arc_buf_hdr_t *hdr = from->b_hdr; 1600 uint64_t size = hdr->b_size; 1601 1602 ASSERT(hdr->b_state != arc_anon); 1603 1604 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 1605 buf->b_hdr = hdr; 1606 buf->b_data = NULL; 1607 buf->b_efunc = NULL; 1608 buf->b_private = NULL; 1609 buf->b_next = hdr->b_buf; 1610 hdr->b_buf = buf; 1611 arc_get_data_buf(buf); 1612 bcopy(from->b_data, buf->b_data, size); 1613 1614 /* 1615 * This buffer already exists in the arc so create a duplicate 1616 * copy for the caller. If the buffer is associated with user data 1617 * then track the size and number of duplicates. These stats will be 1618 * updated as duplicate buffers are created and destroyed. 1619 */ 1620 if (hdr->b_type == ARC_BUFC_DATA) { 1621 ARCSTAT_BUMP(arcstat_duplicate_buffers); 1622 ARCSTAT_INCR(arcstat_duplicate_buffers_size, size); 1623 } 1624 hdr->b_datacnt += 1; 1625 return (buf); 1626} 1627 1628void 1629arc_buf_add_ref(arc_buf_t *buf, void* tag) 1630{ 1631 arc_buf_hdr_t *hdr; 1632 kmutex_t *hash_lock; 1633 1634 /* 1635 * Check to see if this buffer is evicted. Callers 1636 * must verify b_data != NULL to know if the add_ref 1637 * was successful. 1638 */ 1639 mutex_enter(&buf->b_evict_lock); 1640 if (buf->b_data == NULL) { 1641 mutex_exit(&buf->b_evict_lock); 1642 return; 1643 } 1644 hash_lock = HDR_LOCK(buf->b_hdr); 1645 mutex_enter(hash_lock); 1646 hdr = buf->b_hdr; 1647 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 1648 mutex_exit(&buf->b_evict_lock); 1649 1650 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu); 1651 add_reference(hdr, hash_lock, tag); 1652 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr); 1653 arc_access(hdr, hash_lock); 1654 mutex_exit(hash_lock); 1655 ARCSTAT_BUMP(arcstat_hits); 1656 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH), 1657 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA, 1658 data, metadata, hits); 1659} 1660 1661/* 1662 * Free the arc data buffer. If it is an l2arc write in progress, 1663 * the buffer is placed on l2arc_free_on_write to be freed later. 1664 */ 1665static void 1666arc_buf_data_free(arc_buf_t *buf, void (*free_func)(void *, size_t)) 1667{ 1668 arc_buf_hdr_t *hdr = buf->b_hdr; 1669 1670 if (HDR_L2_WRITING(hdr)) { 1671 l2arc_data_free_t *df; 1672 df = kmem_alloc(sizeof (l2arc_data_free_t), KM_SLEEP); 1673 df->l2df_data = buf->b_data; 1674 df->l2df_size = hdr->b_size; 1675 df->l2df_func = free_func; 1676 mutex_enter(&l2arc_free_on_write_mtx); 1677 list_insert_head(l2arc_free_on_write, df); 1678 mutex_exit(&l2arc_free_on_write_mtx); 1679 ARCSTAT_BUMP(arcstat_l2_free_on_write); 1680 } else { 1681 free_func(buf->b_data, hdr->b_size); 1682 } 1683} 1684 1685/* 1686 * Free up buf->b_data and if 'remove' is set, then pull the 1687 * arc_buf_t off of the the arc_buf_hdr_t's list and free it. 1688 */ 1689static void 1690arc_buf_destroy(arc_buf_t *buf, boolean_t recycle, boolean_t remove) 1691{ 1692 arc_buf_t **bufp; 1693 1694 /* free up data associated with the buf */ 1695 if (buf->b_data) { 1696 arc_state_t *state = buf->b_hdr->b_state; 1697 uint64_t size = buf->b_hdr->b_size; 1698 arc_buf_contents_t type = buf->b_hdr->b_type; 1699 1700 arc_cksum_verify(buf); 1701#ifdef illumos 1702 arc_buf_unwatch(buf); 1703#endif /* illumos */ 1704 1705 if (!recycle) { 1706 if (type == ARC_BUFC_METADATA) { 1707 arc_buf_data_free(buf, zio_buf_free); 1708 arc_space_return(size, ARC_SPACE_DATA); 1709 } else { 1710 ASSERT(type == ARC_BUFC_DATA); 1711 arc_buf_data_free(buf, zio_data_buf_free); 1712 ARCSTAT_INCR(arcstat_data_size, -size); 1713 atomic_add_64(&arc_size, -size); 1714 } 1715 } 1716 if (list_link_active(&buf->b_hdr->b_arc_node)) { 1717 uint64_t *cnt = &state->arcs_lsize[type]; 1718 1719 ASSERT(refcount_is_zero(&buf->b_hdr->b_refcnt)); 1720 ASSERT(state != arc_anon); 1721 1722 ASSERT3U(*cnt, >=, size); 1723 atomic_add_64(cnt, -size); 1724 } 1725 ASSERT3U(state->arcs_size, >=, size); 1726 atomic_add_64(&state->arcs_size, -size); 1727 buf->b_data = NULL; 1728 1729 /* 1730 * If we're destroying a duplicate buffer make sure 1731 * that the appropriate statistics are updated. 1732 */ 1733 if (buf->b_hdr->b_datacnt > 1 && 1734 buf->b_hdr->b_type == ARC_BUFC_DATA) { 1735 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers); 1736 ARCSTAT_INCR(arcstat_duplicate_buffers_size, -size); 1737 } 1738 ASSERT(buf->b_hdr->b_datacnt > 0); 1739 buf->b_hdr->b_datacnt -= 1; 1740 } 1741 1742 /* only remove the buf if requested */ 1743 if (!remove) 1744 return; 1745 1746 /* remove the buf from the hdr list */ 1747 for (bufp = &buf->b_hdr->b_buf; *bufp != buf; bufp = &(*bufp)->b_next) 1748 continue; 1749 *bufp = buf->b_next; 1750 buf->b_next = NULL; 1751 1752 ASSERT(buf->b_efunc == NULL); 1753 1754 /* clean up the buf */ 1755 buf->b_hdr = NULL; 1756 kmem_cache_free(buf_cache, buf); 1757} 1758 1759static void 1760arc_hdr_destroy(arc_buf_hdr_t *hdr) 1761{ 1762 ASSERT(refcount_is_zero(&hdr->b_refcnt)); 1763 ASSERT3P(hdr->b_state, ==, arc_anon); 1764 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 1765 l2arc_buf_hdr_t *l2hdr = hdr->b_l2hdr; 1766 1767 if (l2hdr != NULL) { 1768 boolean_t buflist_held = MUTEX_HELD(&l2arc_buflist_mtx); 1769 /* 1770 * To prevent arc_free() and l2arc_evict() from 1771 * attempting to free the same buffer at the same time, 1772 * a FREE_IN_PROGRESS flag is given to arc_free() to 1773 * give it priority. l2arc_evict() can't destroy this 1774 * header while we are waiting on l2arc_buflist_mtx. 1775 * 1776 * The hdr may be removed from l2ad_buflist before we 1777 * grab l2arc_buflist_mtx, so b_l2hdr is rechecked. 1778 */ 1779 if (!buflist_held) { 1780 mutex_enter(&l2arc_buflist_mtx); 1781 l2hdr = hdr->b_l2hdr; 1782 } 1783 1784 if (l2hdr != NULL) { 1785 trim_map_free(l2hdr->b_dev->l2ad_vdev, l2hdr->b_daddr, 1786 hdr->b_size, 0); 1787 list_remove(l2hdr->b_dev->l2ad_buflist, hdr); 1788 ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size); 1789 ARCSTAT_INCR(arcstat_l2_asize, -l2hdr->b_asize); 1790 vdev_space_update(l2hdr->b_dev->l2ad_vdev, 1791 -l2hdr->b_asize, 0, 0); 1792 kmem_free(l2hdr, sizeof (l2arc_buf_hdr_t)); 1793 if (hdr->b_state == arc_l2c_only) 1794 l2arc_hdr_stat_remove(); 1795 hdr->b_l2hdr = NULL; 1796 } 1797 1798 if (!buflist_held) 1799 mutex_exit(&l2arc_buflist_mtx); 1800 } 1801 1802 if (!BUF_EMPTY(hdr)) { 1803 ASSERT(!HDR_IN_HASH_TABLE(hdr)); 1804 buf_discard_identity(hdr); 1805 } 1806 while (hdr->b_buf) { 1807 arc_buf_t *buf = hdr->b_buf; 1808 1809 if (buf->b_efunc) { 1810 mutex_enter(&arc_eviction_mtx); 1811 mutex_enter(&buf->b_evict_lock); 1812 ASSERT(buf->b_hdr != NULL); 1813 arc_buf_destroy(hdr->b_buf, FALSE, FALSE); 1814 hdr->b_buf = buf->b_next; 1815 buf->b_hdr = &arc_eviction_hdr; 1816 buf->b_next = arc_eviction_list; 1817 arc_eviction_list = buf; 1818 mutex_exit(&buf->b_evict_lock); 1819 mutex_exit(&arc_eviction_mtx); 1820 } else { 1821 arc_buf_destroy(hdr->b_buf, FALSE, TRUE); 1822 } 1823 } 1824 if (hdr->b_freeze_cksum != NULL) { 1825 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 1826 hdr->b_freeze_cksum = NULL; 1827 } 1828 if (hdr->b_thawed) { 1829 kmem_free(hdr->b_thawed, 1); 1830 hdr->b_thawed = NULL; 1831 } 1832 1833 ASSERT(!list_link_active(&hdr->b_arc_node)); 1834 ASSERT3P(hdr->b_hash_next, ==, NULL); 1835 ASSERT3P(hdr->b_acb, ==, NULL); 1836 kmem_cache_free(hdr_cache, hdr); 1837} 1838 1839void 1840arc_buf_free(arc_buf_t *buf, void *tag) 1841{ 1842 arc_buf_hdr_t *hdr = buf->b_hdr; 1843 int hashed = hdr->b_state != arc_anon; 1844 1845 ASSERT(buf->b_efunc == NULL); 1846 ASSERT(buf->b_data != NULL); 1847 1848 if (hashed) { 1849 kmutex_t *hash_lock = HDR_LOCK(hdr); 1850 1851 mutex_enter(hash_lock); 1852 hdr = buf->b_hdr; 1853 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 1854 1855 (void) remove_reference(hdr, hash_lock, tag); 1856 if (hdr->b_datacnt > 1) { 1857 arc_buf_destroy(buf, FALSE, TRUE); 1858 } else { 1859 ASSERT(buf == hdr->b_buf); 1860 ASSERT(buf->b_efunc == NULL); 1861 hdr->b_flags |= ARC_BUF_AVAILABLE; 1862 } 1863 mutex_exit(hash_lock); 1864 } else if (HDR_IO_IN_PROGRESS(hdr)) { 1865 int destroy_hdr; 1866 /* 1867 * We are in the middle of an async write. Don't destroy 1868 * this buffer unless the write completes before we finish 1869 * decrementing the reference count. 1870 */ 1871 mutex_enter(&arc_eviction_mtx); 1872 (void) remove_reference(hdr, NULL, tag); 1873 ASSERT(refcount_is_zero(&hdr->b_refcnt)); 1874 destroy_hdr = !HDR_IO_IN_PROGRESS(hdr); 1875 mutex_exit(&arc_eviction_mtx); 1876 if (destroy_hdr) 1877 arc_hdr_destroy(hdr); 1878 } else { 1879 if (remove_reference(hdr, NULL, tag) > 0) 1880 arc_buf_destroy(buf, FALSE, TRUE); 1881 else 1882 arc_hdr_destroy(hdr); 1883 } 1884} 1885 1886boolean_t 1887arc_buf_remove_ref(arc_buf_t *buf, void* tag) 1888{ 1889 arc_buf_hdr_t *hdr = buf->b_hdr; 1890 kmutex_t *hash_lock = HDR_LOCK(hdr); 1891 boolean_t no_callback = (buf->b_efunc == NULL); 1892 1893 if (hdr->b_state == arc_anon) { 1894 ASSERT(hdr->b_datacnt == 1); 1895 arc_buf_free(buf, tag); 1896 return (no_callback); 1897 } 1898 1899 mutex_enter(hash_lock); 1900 hdr = buf->b_hdr; 1901 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 1902 ASSERT(hdr->b_state != arc_anon); 1903 ASSERT(buf->b_data != NULL); 1904 1905 (void) remove_reference(hdr, hash_lock, tag); 1906 if (hdr->b_datacnt > 1) { 1907 if (no_callback) 1908 arc_buf_destroy(buf, FALSE, TRUE); 1909 } else if (no_callback) { 1910 ASSERT(hdr->b_buf == buf && buf->b_next == NULL); 1911 ASSERT(buf->b_efunc == NULL); 1912 hdr->b_flags |= ARC_BUF_AVAILABLE; 1913 } 1914 ASSERT(no_callback || hdr->b_datacnt > 1 || 1915 refcount_is_zero(&hdr->b_refcnt)); 1916 mutex_exit(hash_lock); 1917 return (no_callback); 1918} 1919 1920int 1921arc_buf_size(arc_buf_t *buf) 1922{ 1923 return (buf->b_hdr->b_size); 1924} 1925 1926/* 1927 * Called from the DMU to determine if the current buffer should be 1928 * evicted. In order to ensure proper locking, the eviction must be initiated 1929 * from the DMU. Return true if the buffer is associated with user data and 1930 * duplicate buffers still exist. 1931 */ 1932boolean_t 1933arc_buf_eviction_needed(arc_buf_t *buf) 1934{ 1935 arc_buf_hdr_t *hdr; 1936 boolean_t evict_needed = B_FALSE; 1937 1938 if (zfs_disable_dup_eviction) 1939 return (B_FALSE); 1940 1941 mutex_enter(&buf->b_evict_lock); 1942 hdr = buf->b_hdr; 1943 if (hdr == NULL) { 1944 /* 1945 * We are in arc_do_user_evicts(); let that function 1946 * perform the eviction. 1947 */ 1948 ASSERT(buf->b_data == NULL); 1949 mutex_exit(&buf->b_evict_lock); 1950 return (B_FALSE); 1951 } else if (buf->b_data == NULL) { 1952 /* 1953 * We have already been added to the arc eviction list; 1954 * recommend eviction. 1955 */ 1956 ASSERT3P(hdr, ==, &arc_eviction_hdr); 1957 mutex_exit(&buf->b_evict_lock); 1958 return (B_TRUE); 1959 } 1960 1961 if (hdr->b_datacnt > 1 && hdr->b_type == ARC_BUFC_DATA) 1962 evict_needed = B_TRUE; 1963 1964 mutex_exit(&buf->b_evict_lock); 1965 return (evict_needed); 1966} 1967 1968/* 1969 * Evict buffers from list until we've removed the specified number of 1970 * bytes. Move the removed buffers to the appropriate evict state. 1971 * If the recycle flag is set, then attempt to "recycle" a buffer: 1972 * - look for a buffer to evict that is `bytes' long. 1973 * - return the data block from this buffer rather than freeing it. 1974 * This flag is used by callers that are trying to make space for a 1975 * new buffer in a full arc cache. 1976 * 1977 * This function makes a "best effort". It skips over any buffers 1978 * it can't get a hash_lock on, and so may not catch all candidates. 1979 * It may also return without evicting as much space as requested. 1980 */ 1981static void * 1982arc_evict(arc_state_t *state, uint64_t spa, int64_t bytes, boolean_t recycle, 1983 arc_buf_contents_t type) 1984{ 1985 arc_state_t *evicted_state; 1986 uint64_t bytes_evicted = 0, skipped = 0, missed = 0; 1987 int64_t bytes_remaining; 1988 arc_buf_hdr_t *ab, *ab_prev = NULL; 1989 list_t *evicted_list, *list, *evicted_list_start, *list_start; 1990 kmutex_t *lock, *evicted_lock; 1991 kmutex_t *hash_lock; 1992 boolean_t have_lock; 1993 void *stolen = NULL; 1994 arc_buf_hdr_t marker = { 0 }; 1995 int count = 0; 1996 static int evict_metadata_offset, evict_data_offset; 1997 int i, idx, offset, list_count, lists; 1998 1999 ASSERT(state == arc_mru || state == arc_mfu); 2000 2001 evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost; 2002 2003 if (type == ARC_BUFC_METADATA) { 2004 offset = 0; 2005 list_count = ARC_BUFC_NUMMETADATALISTS; 2006 list_start = &state->arcs_lists[0]; 2007 evicted_list_start = &evicted_state->arcs_lists[0]; 2008 idx = evict_metadata_offset; 2009 } else { 2010 offset = ARC_BUFC_NUMMETADATALISTS; 2011 list_start = &state->arcs_lists[offset]; 2012 evicted_list_start = &evicted_state->arcs_lists[offset]; 2013 list_count = ARC_BUFC_NUMDATALISTS; 2014 idx = evict_data_offset; 2015 } 2016 bytes_remaining = evicted_state->arcs_lsize[type]; 2017 lists = 0; 2018 2019evict_start: 2020 list = &list_start[idx]; 2021 evicted_list = &evicted_list_start[idx]; 2022 lock = ARCS_LOCK(state, (offset + idx)); 2023 evicted_lock = ARCS_LOCK(evicted_state, (offset + idx)); 2024 2025 mutex_enter(lock); 2026 mutex_enter(evicted_lock); 2027 2028 for (ab = list_tail(list); ab; ab = ab_prev) { 2029 ab_prev = list_prev(list, ab); 2030 bytes_remaining -= (ab->b_size * ab->b_datacnt); 2031 /* prefetch buffers have a minimum lifespan */ 2032 if (HDR_IO_IN_PROGRESS(ab) || 2033 (spa && ab->b_spa != spa) || 2034 (ab->b_flags & (ARC_PREFETCH|ARC_INDIRECT) && 2035 ddi_get_lbolt() - ab->b_arc_access < 2036 arc_min_prefetch_lifespan)) { 2037 skipped++; 2038 continue; 2039 } 2040 /* "lookahead" for better eviction candidate */ 2041 if (recycle && ab->b_size != bytes && 2042 ab_prev && ab_prev->b_size == bytes) 2043 continue; 2044 2045 /* ignore markers */ 2046 if (ab->b_spa == 0) 2047 continue; 2048 2049 /* 2050 * It may take a long time to evict all the bufs requested. 2051 * To avoid blocking all arc activity, periodically drop 2052 * the arcs_mtx and give other threads a chance to run 2053 * before reacquiring the lock. 2054 * 2055 * If we are looking for a buffer to recycle, we are in 2056 * the hot code path, so don't sleep. 2057 */ 2058 if (!recycle && count++ > arc_evict_iterations) { 2059 list_insert_after(list, ab, &marker); 2060 mutex_exit(evicted_lock); 2061 mutex_exit(lock); 2062 kpreempt(KPREEMPT_SYNC); 2063 mutex_enter(lock); 2064 mutex_enter(evicted_lock); 2065 ab_prev = list_prev(list, &marker); 2066 list_remove(list, &marker); 2067 count = 0; 2068 continue; 2069 } 2070 2071 hash_lock = HDR_LOCK(ab); 2072 have_lock = MUTEX_HELD(hash_lock); 2073 if (have_lock || mutex_tryenter(hash_lock)) { 2074 ASSERT0(refcount_count(&ab->b_refcnt)); 2075 ASSERT(ab->b_datacnt > 0); 2076 while (ab->b_buf) { 2077 arc_buf_t *buf = ab->b_buf; 2078 if (!mutex_tryenter(&buf->b_evict_lock)) { 2079 missed += 1; 2080 break; 2081 } 2082 if (buf->b_data) { 2083 bytes_evicted += ab->b_size; 2084 if (recycle && ab->b_type == type && 2085 ab->b_size == bytes && 2086 !HDR_L2_WRITING(ab)) { 2087 stolen = buf->b_data; 2088 recycle = FALSE; 2089 } 2090 } 2091 if (buf->b_efunc) { 2092 mutex_enter(&arc_eviction_mtx); 2093 arc_buf_destroy(buf, 2094 buf->b_data == stolen, FALSE); 2095 ab->b_buf = buf->b_next; 2096 buf->b_hdr = &arc_eviction_hdr; 2097 buf->b_next = arc_eviction_list; 2098 arc_eviction_list = buf; 2099 mutex_exit(&arc_eviction_mtx); 2100 mutex_exit(&buf->b_evict_lock); 2101 } else { 2102 mutex_exit(&buf->b_evict_lock); 2103 arc_buf_destroy(buf, 2104 buf->b_data == stolen, TRUE); 2105 } 2106 } 2107 2108 if (ab->b_l2hdr) { 2109 ARCSTAT_INCR(arcstat_evict_l2_cached, 2110 ab->b_size); 2111 } else { 2112 if (l2arc_write_eligible(ab->b_spa, ab)) { 2113 ARCSTAT_INCR(arcstat_evict_l2_eligible, 2114 ab->b_size); 2115 } else { 2116 ARCSTAT_INCR( 2117 arcstat_evict_l2_ineligible, 2118 ab->b_size); 2119 } 2120 } 2121 2122 if (ab->b_datacnt == 0) { 2123 arc_change_state(evicted_state, ab, hash_lock); 2124 ASSERT(HDR_IN_HASH_TABLE(ab)); 2125 ab->b_flags |= ARC_IN_HASH_TABLE; 2126 ab->b_flags &= ~ARC_BUF_AVAILABLE; 2127 DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, ab); 2128 } 2129 if (!have_lock) 2130 mutex_exit(hash_lock); 2131 if (bytes >= 0 && bytes_evicted >= bytes) 2132 break; 2133 if (bytes_remaining > 0) { 2134 mutex_exit(evicted_lock); 2135 mutex_exit(lock); 2136 idx = ((idx + 1) & (list_count - 1)); 2137 lists++; 2138 goto evict_start; 2139 } 2140 } else { 2141 missed += 1; 2142 } 2143 } 2144 2145 mutex_exit(evicted_lock); 2146 mutex_exit(lock); 2147 2148 idx = ((idx + 1) & (list_count - 1)); 2149 lists++; 2150 2151 if (bytes_evicted < bytes) { 2152 if (lists < list_count) 2153 goto evict_start; 2154 else 2155 dprintf("only evicted %lld bytes from %x", 2156 (longlong_t)bytes_evicted, state); 2157 } 2158 if (type == ARC_BUFC_METADATA) 2159 evict_metadata_offset = idx; 2160 else 2161 evict_data_offset = idx; 2162 2163 if (skipped) 2164 ARCSTAT_INCR(arcstat_evict_skip, skipped); 2165 2166 if (missed) 2167 ARCSTAT_INCR(arcstat_mutex_miss, missed); 2168 2169 /* 2170 * Note: we have just evicted some data into the ghost state, 2171 * potentially putting the ghost size over the desired size. Rather 2172 * that evicting from the ghost list in this hot code path, leave 2173 * this chore to the arc_reclaim_thread(). 2174 */ 2175 2176 if (stolen) 2177 ARCSTAT_BUMP(arcstat_stolen); 2178 return (stolen); 2179} 2180 2181/* 2182 * Remove buffers from list until we've removed the specified number of 2183 * bytes. Destroy the buffers that are removed. 2184 */ 2185static void 2186arc_evict_ghost(arc_state_t *state, uint64_t spa, int64_t bytes) 2187{ 2188 arc_buf_hdr_t *ab, *ab_prev; 2189 arc_buf_hdr_t marker = { 0 }; 2190 list_t *list, *list_start; 2191 kmutex_t *hash_lock, *lock; 2192 uint64_t bytes_deleted = 0; 2193 uint64_t bufs_skipped = 0; 2194 int count = 0; 2195 static int evict_offset; 2196 int list_count, idx = evict_offset; 2197 int offset, lists = 0; 2198 2199 ASSERT(GHOST_STATE(state)); 2200 2201 /* 2202 * data lists come after metadata lists 2203 */ 2204 list_start = &state->arcs_lists[ARC_BUFC_NUMMETADATALISTS]; 2205 list_count = ARC_BUFC_NUMDATALISTS; 2206 offset = ARC_BUFC_NUMMETADATALISTS; 2207 2208evict_start: 2209 list = &list_start[idx]; 2210 lock = ARCS_LOCK(state, idx + offset); 2211 2212 mutex_enter(lock); 2213 for (ab = list_tail(list); ab; ab = ab_prev) { 2214 ab_prev = list_prev(list, ab); 2215 if (ab->b_type > ARC_BUFC_NUMTYPES) 2216 panic("invalid ab=%p", (void *)ab); 2217 if (spa && ab->b_spa != spa) 2218 continue; 2219 2220 /* ignore markers */ 2221 if (ab->b_spa == 0) 2222 continue; 2223 2224 hash_lock = HDR_LOCK(ab); 2225 /* caller may be trying to modify this buffer, skip it */ 2226 if (MUTEX_HELD(hash_lock)) 2227 continue; 2228 2229 /* 2230 * It may take a long time to evict all the bufs requested. 2231 * To avoid blocking all arc activity, periodically drop 2232 * the arcs_mtx and give other threads a chance to run 2233 * before reacquiring the lock. 2234 */ 2235 if (count++ > arc_evict_iterations) { 2236 list_insert_after(list, ab, &marker); 2237 mutex_exit(lock); 2238 kpreempt(KPREEMPT_SYNC); 2239 mutex_enter(lock); 2240 ab_prev = list_prev(list, &marker); 2241 list_remove(list, &marker); 2242 count = 0; 2243 continue; 2244 } 2245 if (mutex_tryenter(hash_lock)) { 2246 ASSERT(!HDR_IO_IN_PROGRESS(ab)); 2247 ASSERT(ab->b_buf == NULL); 2248 ARCSTAT_BUMP(arcstat_deleted); 2249 bytes_deleted += ab->b_size; 2250 2251 if (ab->b_l2hdr != NULL) { 2252 /* 2253 * This buffer is cached on the 2nd Level ARC; 2254 * don't destroy the header. 2255 */ 2256 arc_change_state(arc_l2c_only, ab, hash_lock); 2257 mutex_exit(hash_lock); 2258 } else { 2259 arc_change_state(arc_anon, ab, hash_lock); 2260 mutex_exit(hash_lock); 2261 arc_hdr_destroy(ab); 2262 } 2263 2264 DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, ab); 2265 if (bytes >= 0 && bytes_deleted >= bytes) 2266 break; 2267 } else if (bytes < 0) { 2268 /* 2269 * Insert a list marker and then wait for the 2270 * hash lock to become available. Once its 2271 * available, restart from where we left off. 2272 */ 2273 list_insert_after(list, ab, &marker); 2274 mutex_exit(lock); 2275 mutex_enter(hash_lock); 2276 mutex_exit(hash_lock); 2277 mutex_enter(lock); 2278 ab_prev = list_prev(list, &marker); 2279 list_remove(list, &marker); 2280 } else { 2281 bufs_skipped += 1; 2282 } 2283 2284 } 2285 mutex_exit(lock); 2286 idx = ((idx + 1) & (ARC_BUFC_NUMDATALISTS - 1)); 2287 lists++; 2288 2289 if (lists < list_count) 2290 goto evict_start; 2291 2292 evict_offset = idx; 2293 if ((uintptr_t)list > (uintptr_t)&state->arcs_lists[ARC_BUFC_NUMMETADATALISTS] && 2294 (bytes < 0 || bytes_deleted < bytes)) { 2295 list_start = &state->arcs_lists[0]; 2296 list_count = ARC_BUFC_NUMMETADATALISTS; 2297 offset = lists = 0; 2298 goto evict_start; 2299 } 2300 2301 if (bufs_skipped) { 2302 ARCSTAT_INCR(arcstat_mutex_miss, bufs_skipped); 2303 ASSERT(bytes >= 0); 2304 } 2305 2306 if (bytes_deleted < bytes) 2307 dprintf("only deleted %lld bytes from %p", 2308 (longlong_t)bytes_deleted, state); 2309} 2310 2311static void 2312arc_adjust(void) 2313{ 2314 int64_t adjustment, delta; 2315 2316 /* 2317 * Adjust MRU size 2318 */ 2319 2320 adjustment = MIN((int64_t)(arc_size - arc_c), 2321 (int64_t)(arc_anon->arcs_size + arc_mru->arcs_size + arc_meta_used - 2322 arc_p)); 2323 2324 if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_DATA] > 0) { 2325 delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_DATA], adjustment); 2326 (void) arc_evict(arc_mru, 0, delta, FALSE, ARC_BUFC_DATA); 2327 adjustment -= delta; 2328 } 2329 2330 if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_METADATA] > 0) { 2331 delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_METADATA], adjustment); 2332 (void) arc_evict(arc_mru, 0, delta, FALSE, 2333 ARC_BUFC_METADATA); 2334 } 2335 2336 /* 2337 * Adjust MFU size 2338 */ 2339 2340 adjustment = arc_size - arc_c; 2341 2342 if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_DATA] > 0) { 2343 delta = MIN(adjustment, arc_mfu->arcs_lsize[ARC_BUFC_DATA]); 2344 (void) arc_evict(arc_mfu, 0, delta, FALSE, ARC_BUFC_DATA); 2345 adjustment -= delta; 2346 } 2347 2348 if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_METADATA] > 0) { 2349 int64_t delta = MIN(adjustment, 2350 arc_mfu->arcs_lsize[ARC_BUFC_METADATA]); 2351 (void) arc_evict(arc_mfu, 0, delta, FALSE, 2352 ARC_BUFC_METADATA); 2353 } 2354 2355 /* 2356 * Adjust ghost lists 2357 */ 2358 2359 adjustment = arc_mru->arcs_size + arc_mru_ghost->arcs_size - arc_c; 2360 2361 if (adjustment > 0 && arc_mru_ghost->arcs_size > 0) { 2362 delta = MIN(arc_mru_ghost->arcs_size, adjustment); 2363 arc_evict_ghost(arc_mru_ghost, 0, delta); 2364 } 2365 2366 adjustment = 2367 arc_mru_ghost->arcs_size + arc_mfu_ghost->arcs_size - arc_c; 2368 2369 if (adjustment > 0 && arc_mfu_ghost->arcs_size > 0) { 2370 delta = MIN(arc_mfu_ghost->arcs_size, adjustment); 2371 arc_evict_ghost(arc_mfu_ghost, 0, delta); 2372 } 2373} 2374 2375static void 2376arc_do_user_evicts(void) 2377{ 2378 static arc_buf_t *tmp_arc_eviction_list; 2379 2380 /* 2381 * Move list over to avoid LOR 2382 */ 2383restart: 2384 mutex_enter(&arc_eviction_mtx); 2385 tmp_arc_eviction_list = arc_eviction_list; 2386 arc_eviction_list = NULL; 2387 mutex_exit(&arc_eviction_mtx); 2388 2389 while (tmp_arc_eviction_list != NULL) { 2390 arc_buf_t *buf = tmp_arc_eviction_list; 2391 tmp_arc_eviction_list = buf->b_next; 2392 mutex_enter(&buf->b_evict_lock); 2393 buf->b_hdr = NULL; 2394 mutex_exit(&buf->b_evict_lock); 2395 2396 if (buf->b_efunc != NULL) 2397 VERIFY0(buf->b_efunc(buf->b_private)); 2398 2399 buf->b_efunc = NULL; 2400 buf->b_private = NULL; 2401 kmem_cache_free(buf_cache, buf); 2402 } 2403 2404 if (arc_eviction_list != NULL) 2405 goto restart; 2406} 2407 2408/* 2409 * Flush all *evictable* data from the cache for the given spa. 2410 * NOTE: this will not touch "active" (i.e. referenced) data. 2411 */ 2412void 2413arc_flush(spa_t *spa) 2414{ 2415 uint64_t guid = 0; 2416 2417 if (spa) 2418 guid = spa_load_guid(spa); 2419 2420 while (arc_mru->arcs_lsize[ARC_BUFC_DATA]) { 2421 (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_DATA); 2422 if (spa) 2423 break; 2424 } 2425 while (arc_mru->arcs_lsize[ARC_BUFC_METADATA]) { 2426 (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_METADATA); 2427 if (spa) 2428 break; 2429 } 2430 while (arc_mfu->arcs_lsize[ARC_BUFC_DATA]) { 2431 (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_DATA); 2432 if (spa) 2433 break; 2434 } 2435 while (arc_mfu->arcs_lsize[ARC_BUFC_METADATA]) { 2436 (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_METADATA); 2437 if (spa) 2438 break; 2439 } 2440 2441 arc_evict_ghost(arc_mru_ghost, guid, -1); 2442 arc_evict_ghost(arc_mfu_ghost, guid, -1); 2443 2444 mutex_enter(&arc_reclaim_thr_lock); 2445 arc_do_user_evicts(); 2446 mutex_exit(&arc_reclaim_thr_lock); 2447 ASSERT(spa || arc_eviction_list == NULL); 2448} 2449 2450void 2451arc_shrink(void) 2452{ 2453 2454 if (arc_c > arc_c_min) { 2455 uint64_t to_free; 2456 2457 DTRACE_PROBE4(arc__shrink, uint64_t, arc_c, uint64_t, 2458 arc_c_min, uint64_t, arc_p, uint64_t, to_free); 2459#ifdef _KERNEL 2460 to_free = arc_c >> arc_shrink_shift; 2461#else 2462 to_free = arc_c >> arc_shrink_shift; 2463#endif 2464 if (arc_c > arc_c_min + to_free) 2465 atomic_add_64(&arc_c, -to_free); 2466 else 2467 arc_c = arc_c_min; 2468 2469 atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift)); 2470 if (arc_c > arc_size) 2471 arc_c = MAX(arc_size, arc_c_min); 2472 if (arc_p > arc_c) 2473 arc_p = (arc_c >> 1); 2474 2475 DTRACE_PROBE2(arc__shrunk, uint64_t, arc_c, uint64_t, 2476 arc_p); 2477 2478 ASSERT(arc_c >= arc_c_min); 2479 ASSERT((int64_t)arc_p >= 0); 2480 } 2481 2482 if (arc_size > arc_c) { 2483 DTRACE_PROBE2(arc__shrink_adjust, uint64_t, arc_size, 2484 uint64_t, arc_c); 2485 arc_adjust(); 2486 } 2487} 2488 2489static int needfree = 0; 2490 2491static int 2492arc_reclaim_needed(void) 2493{ 2494 2495#ifdef _KERNEL 2496 2497 if (needfree) { 2498 DTRACE_PROBE(arc__reclaim_needfree); 2499 return (1); 2500 } 2501 2502 /* 2503 * Cooperate with pagedaemon when it's time for it to scan 2504 * and reclaim some pages. 2505 */ 2506 if (freemem < zfs_arc_free_target) { 2507 DTRACE_PROBE2(arc__reclaim_freemem, uint64_t, 2508 freemem, uint64_t, zfs_arc_free_target); 2509 return (1); 2510 } 2511 2512#ifdef sun 2513 /* 2514 * take 'desfree' extra pages, so we reclaim sooner, rather than later 2515 */ 2516 extra = desfree; 2517 2518 /* 2519 * check that we're out of range of the pageout scanner. It starts to 2520 * schedule paging if freemem is less than lotsfree and needfree. 2521 * lotsfree is the high-water mark for pageout, and needfree is the 2522 * number of needed free pages. We add extra pages here to make sure 2523 * the scanner doesn't start up while we're freeing memory. 2524 */ 2525 if (freemem < lotsfree + needfree + extra) 2526 return (1); 2527 2528 /* 2529 * check to make sure that swapfs has enough space so that anon 2530 * reservations can still succeed. anon_resvmem() checks that the 2531 * availrmem is greater than swapfs_minfree, and the number of reserved 2532 * swap pages. We also add a bit of extra here just to prevent 2533 * circumstances from getting really dire. 2534 */ 2535 if (availrmem < swapfs_minfree + swapfs_reserve + extra) 2536 return (1); 2537 2538 /* 2539 * Check that we have enough availrmem that memory locking (e.g., via 2540 * mlock(3C) or memcntl(2)) can still succeed. (pages_pp_maximum 2541 * stores the number of pages that cannot be locked; when availrmem 2542 * drops below pages_pp_maximum, page locking mechanisms such as 2543 * page_pp_lock() will fail.) 2544 */ 2545 if (availrmem <= pages_pp_maximum) 2546 return (1); 2547 2548#endif /* sun */ 2549#if defined(__i386) || !defined(UMA_MD_SMALL_ALLOC) 2550 /* 2551 * If we're on an i386 platform, it's possible that we'll exhaust the 2552 * kernel heap space before we ever run out of available physical 2553 * memory. Most checks of the size of the heap_area compare against 2554 * tune.t_minarmem, which is the minimum available real memory that we 2555 * can have in the system. However, this is generally fixed at 25 pages 2556 * which is so low that it's useless. In this comparison, we seek to 2557 * calculate the total heap-size, and reclaim if more than 3/4ths of the 2558 * heap is allocated. (Or, in the calculation, if less than 1/4th is 2559 * free) 2560 */ 2561 if (vmem_size(heap_arena, VMEM_FREE) < 2562 (vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC) >> 2)) { 2563 DTRACE_PROBE2(arc__reclaim_used, uint64_t, 2564 vmem_size(heap_arena, VMEM_FREE), uint64_t, 2565 (vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC)) >> 2); 2566 return (1); 2567 } 2568#endif 2569#ifdef sun 2570 /* 2571 * If zio data pages are being allocated out of a separate heap segment, 2572 * then enforce that the size of available vmem for this arena remains 2573 * above about 1/16th free. 2574 * 2575 * Note: The 1/16th arena free requirement was put in place 2576 * to aggressively evict memory from the arc in order to avoid 2577 * memory fragmentation issues. 2578 */ 2579 if (zio_arena != NULL && 2580 vmem_size(zio_arena, VMEM_FREE) < 2581 (vmem_size(zio_arena, VMEM_ALLOC) >> 4)) 2582 return (1); 2583#endif /* sun */ 2584#else /* _KERNEL */ 2585 if (spa_get_random(100) == 0) 2586 return (1); 2587#endif /* _KERNEL */ 2588 DTRACE_PROBE(arc__reclaim_no); 2589 2590 return (0); 2591} 2592 2593extern kmem_cache_t *zio_buf_cache[]; 2594extern kmem_cache_t *zio_data_buf_cache[]; 2595extern kmem_cache_t *range_seg_cache; 2596 2597static void __noinline 2598arc_kmem_reap_now(arc_reclaim_strategy_t strat) 2599{ 2600 size_t i; 2601 kmem_cache_t *prev_cache = NULL; 2602 kmem_cache_t *prev_data_cache = NULL; 2603 2604 DTRACE_PROBE(arc__kmem_reap_start); 2605#ifdef _KERNEL 2606 if (arc_meta_used >= arc_meta_limit) { 2607 /* 2608 * We are exceeding our meta-data cache limit. 2609 * Purge some DNLC entries to release holds on meta-data. 2610 */ 2611 dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent); 2612 } 2613#if defined(__i386) 2614 /* 2615 * Reclaim unused memory from all kmem caches. 2616 */ 2617 kmem_reap(); 2618#endif 2619#endif 2620 2621 /* 2622 * An aggressive reclamation will shrink the cache size as well as 2623 * reap free buffers from the arc kmem caches. 2624 */ 2625 if (strat == ARC_RECLAIM_AGGR) 2626 arc_shrink(); 2627 2628 for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) { 2629 if (zio_buf_cache[i] != prev_cache) { 2630 prev_cache = zio_buf_cache[i]; 2631 kmem_cache_reap_now(zio_buf_cache[i]); 2632 } 2633 if (zio_data_buf_cache[i] != prev_data_cache) { 2634 prev_data_cache = zio_data_buf_cache[i]; 2635 kmem_cache_reap_now(zio_data_buf_cache[i]); 2636 } 2637 } 2638 kmem_cache_reap_now(buf_cache); 2639 kmem_cache_reap_now(hdr_cache); 2640 kmem_cache_reap_now(range_seg_cache); 2641 2642#ifdef sun 2643 /* 2644 * Ask the vmem arena to reclaim unused memory from its 2645 * quantum caches. 2646 */ 2647 if (zio_arena != NULL && strat == ARC_RECLAIM_AGGR) 2648 vmem_qcache_reap(zio_arena); 2649#endif 2650 DTRACE_PROBE(arc__kmem_reap_end); 2651} 2652 2653static void 2654arc_reclaim_thread(void *dummy __unused) 2655{ 2656 clock_t growtime = 0; 2657 arc_reclaim_strategy_t last_reclaim = ARC_RECLAIM_CONS; 2658 callb_cpr_t cpr; 2659 2660 CALLB_CPR_INIT(&cpr, &arc_reclaim_thr_lock, callb_generic_cpr, FTAG); 2661 2662 mutex_enter(&arc_reclaim_thr_lock); 2663 while (arc_thread_exit == 0) { 2664 if (arc_reclaim_needed()) { 2665 2666 if (arc_no_grow) { 2667 if (last_reclaim == ARC_RECLAIM_CONS) { 2668 DTRACE_PROBE(arc__reclaim_aggr_no_grow); 2669 last_reclaim = ARC_RECLAIM_AGGR; 2670 } else { 2671 last_reclaim = ARC_RECLAIM_CONS; 2672 } 2673 } else { 2674 arc_no_grow = TRUE; 2675 last_reclaim = ARC_RECLAIM_AGGR; 2676 DTRACE_PROBE(arc__reclaim_aggr); 2677 membar_producer(); 2678 } 2679 2680 /* reset the growth delay for every reclaim */ 2681 growtime = ddi_get_lbolt() + (arc_grow_retry * hz); 2682 2683 if (needfree && last_reclaim == ARC_RECLAIM_CONS) { 2684 /* 2685 * If needfree is TRUE our vm_lowmem hook 2686 * was called and in that case we must free some 2687 * memory, so switch to aggressive mode. 2688 */ 2689 arc_no_grow = TRUE; 2690 last_reclaim = ARC_RECLAIM_AGGR; 2691 } 2692 arc_kmem_reap_now(last_reclaim); 2693 arc_warm = B_TRUE; 2694 2695 } else if (arc_no_grow && ddi_get_lbolt() >= growtime) { 2696 arc_no_grow = FALSE; 2697 } 2698 2699 arc_adjust(); 2700 2701 if (arc_eviction_list != NULL) 2702 arc_do_user_evicts(); 2703 2704#ifdef _KERNEL 2705 if (needfree) { 2706 needfree = 0; 2707 wakeup(&needfree); 2708 } 2709#endif 2710 2711 /* block until needed, or one second, whichever is shorter */ 2712 CALLB_CPR_SAFE_BEGIN(&cpr); 2713 (void) cv_timedwait(&arc_reclaim_thr_cv, 2714 &arc_reclaim_thr_lock, hz); 2715 CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_thr_lock); 2716 } 2717 2718 arc_thread_exit = 0; 2719 cv_broadcast(&arc_reclaim_thr_cv); 2720 CALLB_CPR_EXIT(&cpr); /* drops arc_reclaim_thr_lock */ 2721 thread_exit(); 2722} 2723 2724/* 2725 * Adapt arc info given the number of bytes we are trying to add and 2726 * the state that we are comming from. This function is only called 2727 * when we are adding new content to the cache. 2728 */ 2729static void 2730arc_adapt(int bytes, arc_state_t *state) 2731{ 2732 int mult; 2733 uint64_t arc_p_min = (arc_c >> arc_p_min_shift); 2734 2735 if (state == arc_l2c_only) 2736 return; 2737 2738 ASSERT(bytes > 0); 2739 /* 2740 * Adapt the target size of the MRU list: 2741 * - if we just hit in the MRU ghost list, then increase 2742 * the target size of the MRU list. 2743 * - if we just hit in the MFU ghost list, then increase 2744 * the target size of the MFU list by decreasing the 2745 * target size of the MRU list. 2746 */ 2747 if (state == arc_mru_ghost) { 2748 mult = ((arc_mru_ghost->arcs_size >= arc_mfu_ghost->arcs_size) ? 2749 1 : (arc_mfu_ghost->arcs_size/arc_mru_ghost->arcs_size)); 2750 mult = MIN(mult, 10); /* avoid wild arc_p adjustment */ 2751 2752 arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult); 2753 } else if (state == arc_mfu_ghost) { 2754 uint64_t delta; 2755 2756 mult = ((arc_mfu_ghost->arcs_size >= arc_mru_ghost->arcs_size) ? 2757 1 : (arc_mru_ghost->arcs_size/arc_mfu_ghost->arcs_size)); 2758 mult = MIN(mult, 10); 2759 2760 delta = MIN(bytes * mult, arc_p); 2761 arc_p = MAX(arc_p_min, arc_p - delta); 2762 } 2763 ASSERT((int64_t)arc_p >= 0); 2764 2765 if (arc_reclaim_needed()) { 2766 cv_signal(&arc_reclaim_thr_cv); 2767 return; 2768 } 2769 2770 if (arc_no_grow) 2771 return; 2772 2773 if (arc_c >= arc_c_max) 2774 return; 2775 2776 /* 2777 * If we're within (2 * maxblocksize) bytes of the target 2778 * cache size, increment the target cache size 2779 */ 2780 if (arc_size > arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) { 2781 DTRACE_PROBE1(arc__inc_adapt, int, bytes); 2782 atomic_add_64(&arc_c, (int64_t)bytes); 2783 if (arc_c > arc_c_max) 2784 arc_c = arc_c_max; 2785 else if (state == arc_anon) 2786 atomic_add_64(&arc_p, (int64_t)bytes); 2787 if (arc_p > arc_c) 2788 arc_p = arc_c; 2789 } 2790 ASSERT((int64_t)arc_p >= 0); 2791} 2792 2793/* 2794 * Check if the cache has reached its limits and eviction is required 2795 * prior to insert. 2796 */ 2797static int 2798arc_evict_needed(arc_buf_contents_t type) 2799{ 2800 if (type == ARC_BUFC_METADATA && arc_meta_used >= arc_meta_limit) 2801 return (1); 2802 2803 if (arc_reclaim_needed()) 2804 return (1); 2805 2806 return (arc_size > arc_c); 2807} 2808 2809/* 2810 * The buffer, supplied as the first argument, needs a data block. 2811 * So, if we are at cache max, determine which cache should be victimized. 2812 * We have the following cases: 2813 * 2814 * 1. Insert for MRU, p > sizeof(arc_anon + arc_mru) -> 2815 * In this situation if we're out of space, but the resident size of the MFU is 2816 * under the limit, victimize the MFU cache to satisfy this insertion request. 2817 * 2818 * 2. Insert for MRU, p <= sizeof(arc_anon + arc_mru) -> 2819 * Here, we've used up all of the available space for the MRU, so we need to 2820 * evict from our own cache instead. Evict from the set of resident MRU 2821 * entries. 2822 * 2823 * 3. Insert for MFU (c - p) > sizeof(arc_mfu) -> 2824 * c minus p represents the MFU space in the cache, since p is the size of the 2825 * cache that is dedicated to the MRU. In this situation there's still space on 2826 * the MFU side, so the MRU side needs to be victimized. 2827 * 2828 * 4. Insert for MFU (c - p) < sizeof(arc_mfu) -> 2829 * MFU's resident set is consuming more space than it has been allotted. In 2830 * this situation, we must victimize our own cache, the MFU, for this insertion. 2831 */ 2832static void 2833arc_get_data_buf(arc_buf_t *buf) 2834{ 2835 arc_state_t *state = buf->b_hdr->b_state; 2836 uint64_t size = buf->b_hdr->b_size; 2837 arc_buf_contents_t type = buf->b_hdr->b_type; 2838 2839 arc_adapt(size, state); 2840 2841 /* 2842 * We have not yet reached cache maximum size, 2843 * just allocate a new buffer. 2844 */ 2845 if (!arc_evict_needed(type)) { 2846 if (type == ARC_BUFC_METADATA) { 2847 buf->b_data = zio_buf_alloc(size); 2848 arc_space_consume(size, ARC_SPACE_DATA); 2849 } else { 2850 ASSERT(type == ARC_BUFC_DATA); 2851 buf->b_data = zio_data_buf_alloc(size); 2852 ARCSTAT_INCR(arcstat_data_size, size); 2853 atomic_add_64(&arc_size, size); 2854 } 2855 goto out; 2856 } 2857 2858 /* 2859 * If we are prefetching from the mfu ghost list, this buffer 2860 * will end up on the mru list; so steal space from there. 2861 */ 2862 if (state == arc_mfu_ghost) 2863 state = buf->b_hdr->b_flags & ARC_PREFETCH ? arc_mru : arc_mfu; 2864 else if (state == arc_mru_ghost) 2865 state = arc_mru; 2866 2867 if (state == arc_mru || state == arc_anon) { 2868 uint64_t mru_used = arc_anon->arcs_size + arc_mru->arcs_size; 2869 state = (arc_mfu->arcs_lsize[type] >= size && 2870 arc_p > mru_used) ? arc_mfu : arc_mru; 2871 } else { 2872 /* MFU cases */ 2873 uint64_t mfu_space = arc_c - arc_p; 2874 state = (arc_mru->arcs_lsize[type] >= size && 2875 mfu_space > arc_mfu->arcs_size) ? arc_mru : arc_mfu; 2876 } 2877 if ((buf->b_data = arc_evict(state, 0, size, TRUE, type)) == NULL) { 2878 if (type == ARC_BUFC_METADATA) { 2879 buf->b_data = zio_buf_alloc(size); 2880 arc_space_consume(size, ARC_SPACE_DATA); 2881 } else { 2882 ASSERT(type == ARC_BUFC_DATA); 2883 buf->b_data = zio_data_buf_alloc(size); 2884 ARCSTAT_INCR(arcstat_data_size, size); 2885 atomic_add_64(&arc_size, size); 2886 } 2887 ARCSTAT_BUMP(arcstat_recycle_miss); 2888 } 2889 ASSERT(buf->b_data != NULL); 2890out: 2891 /* 2892 * Update the state size. Note that ghost states have a 2893 * "ghost size" and so don't need to be updated. 2894 */ 2895 if (!GHOST_STATE(buf->b_hdr->b_state)) { 2896 arc_buf_hdr_t *hdr = buf->b_hdr; 2897 2898 atomic_add_64(&hdr->b_state->arcs_size, size); 2899 if (list_link_active(&hdr->b_arc_node)) { 2900 ASSERT(refcount_is_zero(&hdr->b_refcnt)); 2901 atomic_add_64(&hdr->b_state->arcs_lsize[type], size); 2902 } 2903 /* 2904 * If we are growing the cache, and we are adding anonymous 2905 * data, and we have outgrown arc_p, update arc_p 2906 */ 2907 if (arc_size < arc_c && hdr->b_state == arc_anon && 2908 arc_anon->arcs_size + arc_mru->arcs_size > arc_p) 2909 arc_p = MIN(arc_c, arc_p + size); 2910 } 2911 ARCSTAT_BUMP(arcstat_allocated); 2912} 2913 2914/* 2915 * This routine is called whenever a buffer is accessed. 2916 * NOTE: the hash lock is dropped in this function. 2917 */ 2918static void 2919arc_access(arc_buf_hdr_t *buf, kmutex_t *hash_lock) 2920{ 2921 clock_t now; 2922 2923 ASSERT(MUTEX_HELD(hash_lock)); 2924 2925 if (buf->b_state == arc_anon) { 2926 /* 2927 * This buffer is not in the cache, and does not 2928 * appear in our "ghost" list. Add the new buffer 2929 * to the MRU state. 2930 */ 2931 2932 ASSERT(buf->b_arc_access == 0); 2933 buf->b_arc_access = ddi_get_lbolt(); 2934 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, buf); 2935 arc_change_state(arc_mru, buf, hash_lock); 2936 2937 } else if (buf->b_state == arc_mru) { 2938 now = ddi_get_lbolt(); 2939 2940 /* 2941 * If this buffer is here because of a prefetch, then either: 2942 * - clear the flag if this is a "referencing" read 2943 * (any subsequent access will bump this into the MFU state). 2944 * or 2945 * - move the buffer to the head of the list if this is 2946 * another prefetch (to make it less likely to be evicted). 2947 */ 2948 if ((buf->b_flags & ARC_PREFETCH) != 0) { 2949 if (refcount_count(&buf->b_refcnt) == 0) { 2950 ASSERT(list_link_active(&buf->b_arc_node)); 2951 } else { 2952 buf->b_flags &= ~ARC_PREFETCH; 2953 ARCSTAT_BUMP(arcstat_mru_hits); 2954 } 2955 buf->b_arc_access = now; 2956 return; 2957 } 2958 2959 /* 2960 * This buffer has been "accessed" only once so far, 2961 * but it is still in the cache. Move it to the MFU 2962 * state. 2963 */ 2964 if (now > buf->b_arc_access + ARC_MINTIME) { 2965 /* 2966 * More than 125ms have passed since we 2967 * instantiated this buffer. Move it to the 2968 * most frequently used state. 2969 */ 2970 buf->b_arc_access = now; 2971 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf); 2972 arc_change_state(arc_mfu, buf, hash_lock); 2973 } 2974 ARCSTAT_BUMP(arcstat_mru_hits); 2975 } else if (buf->b_state == arc_mru_ghost) { 2976 arc_state_t *new_state; 2977 /* 2978 * This buffer has been "accessed" recently, but 2979 * was evicted from the cache. Move it to the 2980 * MFU state. 2981 */ 2982 2983 if (buf->b_flags & ARC_PREFETCH) { 2984 new_state = arc_mru; 2985 if (refcount_count(&buf->b_refcnt) > 0) 2986 buf->b_flags &= ~ARC_PREFETCH; 2987 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, buf); 2988 } else { 2989 new_state = arc_mfu; 2990 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf); 2991 } 2992 2993 buf->b_arc_access = ddi_get_lbolt(); 2994 arc_change_state(new_state, buf, hash_lock); 2995 2996 ARCSTAT_BUMP(arcstat_mru_ghost_hits); 2997 } else if (buf->b_state == arc_mfu) { 2998 /* 2999 * This buffer has been accessed more than once and is 3000 * still in the cache. Keep it in the MFU state. 3001 * 3002 * NOTE: an add_reference() that occurred when we did 3003 * the arc_read() will have kicked this off the list. 3004 * If it was a prefetch, we will explicitly move it to 3005 * the head of the list now. 3006 */ 3007 if ((buf->b_flags & ARC_PREFETCH) != 0) { 3008 ASSERT(refcount_count(&buf->b_refcnt) == 0); 3009 ASSERT(list_link_active(&buf->b_arc_node)); 3010 } 3011 ARCSTAT_BUMP(arcstat_mfu_hits); 3012 buf->b_arc_access = ddi_get_lbolt(); 3013 } else if (buf->b_state == arc_mfu_ghost) { 3014 arc_state_t *new_state = arc_mfu; 3015 /* 3016 * This buffer has been accessed more than once but has 3017 * been evicted from the cache. Move it back to the 3018 * MFU state. 3019 */ 3020 3021 if (buf->b_flags & ARC_PREFETCH) { 3022 /* 3023 * This is a prefetch access... 3024 * move this block back to the MRU state. 3025 */ 3026 ASSERT0(refcount_count(&buf->b_refcnt)); 3027 new_state = arc_mru; 3028 } 3029 3030 buf->b_arc_access = ddi_get_lbolt(); 3031 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf); 3032 arc_change_state(new_state, buf, hash_lock); 3033 3034 ARCSTAT_BUMP(arcstat_mfu_ghost_hits); 3035 } else if (buf->b_state == arc_l2c_only) { 3036 /* 3037 * This buffer is on the 2nd Level ARC. 3038 */ 3039 3040 buf->b_arc_access = ddi_get_lbolt(); 3041 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, buf); 3042 arc_change_state(arc_mfu, buf, hash_lock); 3043 } else { 3044 ASSERT(!"invalid arc state"); 3045 } 3046} 3047 3048/* a generic arc_done_func_t which you can use */ 3049/* ARGSUSED */ 3050void 3051arc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg) 3052{ 3053 if (zio == NULL || zio->io_error == 0) 3054 bcopy(buf->b_data, arg, buf->b_hdr->b_size); 3055 VERIFY(arc_buf_remove_ref(buf, arg)); 3056} 3057 3058/* a generic arc_done_func_t */ 3059void 3060arc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg) 3061{ 3062 arc_buf_t **bufp = arg; 3063 if (zio && zio->io_error) { 3064 VERIFY(arc_buf_remove_ref(buf, arg)); 3065 *bufp = NULL; 3066 } else { 3067 *bufp = buf; 3068 ASSERT(buf->b_data); 3069 } 3070} 3071 3072static void 3073arc_read_done(zio_t *zio) 3074{ 3075 arc_buf_hdr_t *hdr; 3076 arc_buf_t *buf; 3077 arc_buf_t *abuf; /* buffer we're assigning to callback */ 3078 kmutex_t *hash_lock = NULL; 3079 arc_callback_t *callback_list, *acb; 3080 int freeable = FALSE; 3081 3082 buf = zio->io_private; 3083 hdr = buf->b_hdr; 3084 3085 /* 3086 * The hdr was inserted into hash-table and removed from lists 3087 * prior to starting I/O. We should find this header, since 3088 * it's in the hash table, and it should be legit since it's 3089 * not possible to evict it during the I/O. The only possible 3090 * reason for it not to be found is if we were freed during the 3091 * read. 3092 */ 3093 if (HDR_IN_HASH_TABLE(hdr)) { 3094 ASSERT3U(hdr->b_birth, ==, BP_PHYSICAL_BIRTH(zio->io_bp)); 3095 ASSERT3U(hdr->b_dva.dva_word[0], ==, 3096 BP_IDENTITY(zio->io_bp)->dva_word[0]); 3097 ASSERT3U(hdr->b_dva.dva_word[1], ==, 3098 BP_IDENTITY(zio->io_bp)->dva_word[1]); 3099 3100 arc_buf_hdr_t *found = buf_hash_find(hdr->b_spa, zio->io_bp, 3101 &hash_lock); 3102 3103 ASSERT((found == NULL && HDR_FREED_IN_READ(hdr) && 3104 hash_lock == NULL) || 3105 (found == hdr && 3106 DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) || 3107 (found == hdr && HDR_L2_READING(hdr))); 3108 } 3109 3110 hdr->b_flags &= ~ARC_L2_EVICTED; 3111 if (l2arc_noprefetch && (hdr->b_flags & ARC_PREFETCH)) 3112 hdr->b_flags &= ~ARC_L2CACHE; 3113 3114 /* byteswap if necessary */ 3115 callback_list = hdr->b_acb; 3116 ASSERT(callback_list != NULL); 3117 if (BP_SHOULD_BYTESWAP(zio->io_bp) && zio->io_error == 0) { 3118 dmu_object_byteswap_t bswap = 3119 DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp)); 3120 arc_byteswap_func_t *func = BP_GET_LEVEL(zio->io_bp) > 0 ? 3121 byteswap_uint64_array : 3122 dmu_ot_byteswap[bswap].ob_func; 3123 func(buf->b_data, hdr->b_size); 3124 } 3125 3126 arc_cksum_compute(buf, B_FALSE); 3127#ifdef illumos 3128 arc_buf_watch(buf); 3129#endif /* illumos */ 3130 3131 if (hash_lock && zio->io_error == 0 && hdr->b_state == arc_anon) { 3132 /* 3133 * Only call arc_access on anonymous buffers. This is because 3134 * if we've issued an I/O for an evicted buffer, we've already 3135 * called arc_access (to prevent any simultaneous readers from 3136 * getting confused). 3137 */ 3138 arc_access(hdr, hash_lock); 3139 } 3140 3141 /* create copies of the data buffer for the callers */ 3142 abuf = buf; 3143 for (acb = callback_list; acb; acb = acb->acb_next) { 3144 if (acb->acb_done) { 3145 if (abuf == NULL) { 3146 ARCSTAT_BUMP(arcstat_duplicate_reads); 3147 abuf = arc_buf_clone(buf); 3148 } 3149 acb->acb_buf = abuf; 3150 abuf = NULL; 3151 } 3152 } 3153 hdr->b_acb = NULL; 3154 hdr->b_flags &= ~ARC_IO_IN_PROGRESS; 3155 ASSERT(!HDR_BUF_AVAILABLE(hdr)); 3156 if (abuf == buf) { 3157 ASSERT(buf->b_efunc == NULL); 3158 ASSERT(hdr->b_datacnt == 1); 3159 hdr->b_flags |= ARC_BUF_AVAILABLE; 3160 } 3161 3162 ASSERT(refcount_is_zero(&hdr->b_refcnt) || callback_list != NULL); 3163 3164 if (zio->io_error != 0) { 3165 hdr->b_flags |= ARC_IO_ERROR; 3166 if (hdr->b_state != arc_anon) 3167 arc_change_state(arc_anon, hdr, hash_lock); 3168 if (HDR_IN_HASH_TABLE(hdr)) 3169 buf_hash_remove(hdr); 3170 freeable = refcount_is_zero(&hdr->b_refcnt); 3171 } 3172 3173 /* 3174 * Broadcast before we drop the hash_lock to avoid the possibility 3175 * that the hdr (and hence the cv) might be freed before we get to 3176 * the cv_broadcast(). 3177 */ 3178 cv_broadcast(&hdr->b_cv); 3179 3180 if (hash_lock) { 3181 mutex_exit(hash_lock); 3182 } else { 3183 /* 3184 * This block was freed while we waited for the read to 3185 * complete. It has been removed from the hash table and 3186 * moved to the anonymous state (so that it won't show up 3187 * in the cache). 3188 */ 3189 ASSERT3P(hdr->b_state, ==, arc_anon); 3190 freeable = refcount_is_zero(&hdr->b_refcnt); 3191 } 3192 3193 /* execute each callback and free its structure */ 3194 while ((acb = callback_list) != NULL) { 3195 if (acb->acb_done) 3196 acb->acb_done(zio, acb->acb_buf, acb->acb_private); 3197 3198 if (acb->acb_zio_dummy != NULL) { 3199 acb->acb_zio_dummy->io_error = zio->io_error; 3200 zio_nowait(acb->acb_zio_dummy); 3201 } 3202 3203 callback_list = acb->acb_next; 3204 kmem_free(acb, sizeof (arc_callback_t)); 3205 } 3206 3207 if (freeable) 3208 arc_hdr_destroy(hdr); 3209} 3210 3211/* 3212 * "Read" the block block at the specified DVA (in bp) via the 3213 * cache. If the block is found in the cache, invoke the provided 3214 * callback immediately and return. Note that the `zio' parameter 3215 * in the callback will be NULL in this case, since no IO was 3216 * required. If the block is not in the cache pass the read request 3217 * on to the spa with a substitute callback function, so that the 3218 * requested block will be added to the cache. 3219 * 3220 * If a read request arrives for a block that has a read in-progress, 3221 * either wait for the in-progress read to complete (and return the 3222 * results); or, if this is a read with a "done" func, add a record 3223 * to the read to invoke the "done" func when the read completes, 3224 * and return; or just return. 3225 * 3226 * arc_read_done() will invoke all the requested "done" functions 3227 * for readers of this block. 3228 */ 3229int 3230arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_done_func_t *done, 3231 void *private, zio_priority_t priority, int zio_flags, uint32_t *arc_flags, 3232 const zbookmark_phys_t *zb) 3233{ 3234 arc_buf_hdr_t *hdr = NULL; 3235 arc_buf_t *buf = NULL; 3236 kmutex_t *hash_lock = NULL; 3237 zio_t *rzio; 3238 uint64_t guid = spa_load_guid(spa); 3239 3240 ASSERT(!BP_IS_EMBEDDED(bp) || 3241 BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA); 3242 3243top: 3244 if (!BP_IS_EMBEDDED(bp)) { 3245 /* 3246 * Embedded BP's have no DVA and require no I/O to "read". 3247 * Create an anonymous arc buf to back it. 3248 */ 3249 hdr = buf_hash_find(guid, bp, &hash_lock); 3250 } 3251 3252 if (hdr != NULL && hdr->b_datacnt > 0) { 3253 3254 *arc_flags |= ARC_CACHED; 3255 3256 if (HDR_IO_IN_PROGRESS(hdr)) { 3257 3258 if (*arc_flags & ARC_WAIT) { 3259 cv_wait(&hdr->b_cv, hash_lock); 3260 mutex_exit(hash_lock); 3261 goto top; 3262 } 3263 ASSERT(*arc_flags & ARC_NOWAIT); 3264 3265 if (done) { 3266 arc_callback_t *acb = NULL; 3267 3268 acb = kmem_zalloc(sizeof (arc_callback_t), 3269 KM_SLEEP); 3270 acb->acb_done = done; 3271 acb->acb_private = private; 3272 if (pio != NULL) 3273 acb->acb_zio_dummy = zio_null(pio, 3274 spa, NULL, NULL, NULL, zio_flags); 3275 3276 ASSERT(acb->acb_done != NULL); 3277 acb->acb_next = hdr->b_acb; 3278 hdr->b_acb = acb; 3279 add_reference(hdr, hash_lock, private); 3280 mutex_exit(hash_lock); 3281 return (0); 3282 } 3283 mutex_exit(hash_lock); 3284 return (0); 3285 } 3286 3287 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu); 3288 3289 if (done) { 3290 add_reference(hdr, hash_lock, private); 3291 /* 3292 * If this block is already in use, create a new 3293 * copy of the data so that we will be guaranteed 3294 * that arc_release() will always succeed. 3295 */ 3296 buf = hdr->b_buf; 3297 ASSERT(buf); 3298 ASSERT(buf->b_data); 3299 if (HDR_BUF_AVAILABLE(hdr)) { 3300 ASSERT(buf->b_efunc == NULL); 3301 hdr->b_flags &= ~ARC_BUF_AVAILABLE; 3302 } else { 3303 buf = arc_buf_clone(buf); 3304 } 3305 3306 } else if (*arc_flags & ARC_PREFETCH && 3307 refcount_count(&hdr->b_refcnt) == 0) { 3308 hdr->b_flags |= ARC_PREFETCH; 3309 } 3310 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr); 3311 arc_access(hdr, hash_lock); 3312 if (*arc_flags & ARC_L2CACHE) 3313 hdr->b_flags |= ARC_L2CACHE; 3314 if (*arc_flags & ARC_L2COMPRESS) 3315 hdr->b_flags |= ARC_L2COMPRESS; 3316 mutex_exit(hash_lock); 3317 ARCSTAT_BUMP(arcstat_hits); 3318 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH), 3319 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA, 3320 data, metadata, hits); 3321 3322 if (done) 3323 done(NULL, buf, private); 3324 } else { 3325 uint64_t size = BP_GET_LSIZE(bp); 3326 arc_callback_t *acb; 3327 vdev_t *vd = NULL; 3328 uint64_t addr = 0; 3329 boolean_t devw = B_FALSE; 3330 enum zio_compress b_compress = ZIO_COMPRESS_OFF; 3331 uint64_t b_asize = 0; 3332 3333 if (hdr == NULL) { 3334 /* this block is not in the cache */ 3335 arc_buf_hdr_t *exists = NULL; 3336 arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp); 3337 buf = arc_buf_alloc(spa, size, private, type); 3338 hdr = buf->b_hdr; 3339 if (!BP_IS_EMBEDDED(bp)) { 3340 hdr->b_dva = *BP_IDENTITY(bp); 3341 hdr->b_birth = BP_PHYSICAL_BIRTH(bp); 3342 hdr->b_cksum0 = bp->blk_cksum.zc_word[0]; 3343 exists = buf_hash_insert(hdr, &hash_lock); 3344 } 3345 if (exists != NULL) { 3346 /* somebody beat us to the hash insert */ 3347 mutex_exit(hash_lock); 3348 buf_discard_identity(hdr); 3349 (void) arc_buf_remove_ref(buf, private); 3350 goto top; /* restart the IO request */ 3351 } 3352 /* if this is a prefetch, we don't have a reference */ 3353 if (*arc_flags & ARC_PREFETCH) { 3354 (void) remove_reference(hdr, hash_lock, 3355 private); 3356 hdr->b_flags |= ARC_PREFETCH; 3357 } 3358 if (*arc_flags & ARC_L2CACHE) 3359 hdr->b_flags |= ARC_L2CACHE; 3360 if (*arc_flags & ARC_L2COMPRESS) 3361 hdr->b_flags |= ARC_L2COMPRESS; 3362 if (BP_GET_LEVEL(bp) > 0) 3363 hdr->b_flags |= ARC_INDIRECT; 3364 } else { 3365 /* this block is in the ghost cache */ 3366 ASSERT(GHOST_STATE(hdr->b_state)); 3367 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 3368 ASSERT0(refcount_count(&hdr->b_refcnt)); 3369 ASSERT(hdr->b_buf == NULL); 3370 3371 /* if this is a prefetch, we don't have a reference */ 3372 if (*arc_flags & ARC_PREFETCH) 3373 hdr->b_flags |= ARC_PREFETCH; 3374 else 3375 add_reference(hdr, hash_lock, private); 3376 if (*arc_flags & ARC_L2CACHE) 3377 hdr->b_flags |= ARC_L2CACHE; 3378 if (*arc_flags & ARC_L2COMPRESS) 3379 hdr->b_flags |= ARC_L2COMPRESS; 3380 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 3381 buf->b_hdr = hdr; 3382 buf->b_data = NULL; 3383 buf->b_efunc = NULL; 3384 buf->b_private = NULL; 3385 buf->b_next = NULL; 3386 hdr->b_buf = buf; 3387 ASSERT(hdr->b_datacnt == 0); 3388 hdr->b_datacnt = 1; 3389 arc_get_data_buf(buf); 3390 arc_access(hdr, hash_lock); 3391 } 3392 3393 ASSERT(!GHOST_STATE(hdr->b_state)); 3394 3395 acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP); 3396 acb->acb_done = done; 3397 acb->acb_private = private; 3398 3399 ASSERT(hdr->b_acb == NULL); 3400 hdr->b_acb = acb; 3401 hdr->b_flags |= ARC_IO_IN_PROGRESS; 3402 3403 if (hdr->b_l2hdr != NULL && 3404 (vd = hdr->b_l2hdr->b_dev->l2ad_vdev) != NULL) { 3405 devw = hdr->b_l2hdr->b_dev->l2ad_writing; 3406 addr = hdr->b_l2hdr->b_daddr; 3407 b_compress = hdr->b_l2hdr->b_compress; 3408 b_asize = hdr->b_l2hdr->b_asize; 3409 /* 3410 * Lock out device removal. 3411 */ 3412 if (vdev_is_dead(vd) || 3413 !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER)) 3414 vd = NULL; 3415 } 3416 3417 if (hash_lock != NULL) 3418 mutex_exit(hash_lock); 3419 3420 /* 3421 * At this point, we have a level 1 cache miss. Try again in 3422 * L2ARC if possible. 3423 */ 3424 ASSERT3U(hdr->b_size, ==, size); 3425 DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp, 3426 uint64_t, size, zbookmark_phys_t *, zb); 3427 ARCSTAT_BUMP(arcstat_misses); 3428 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_PREFETCH), 3429 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA, 3430 data, metadata, misses); 3431#ifdef _KERNEL 3432 curthread->td_ru.ru_inblock++; 3433#endif 3434 3435 if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) { 3436 /* 3437 * Read from the L2ARC if the following are true: 3438 * 1. The L2ARC vdev was previously cached. 3439 * 2. This buffer still has L2ARC metadata. 3440 * 3. This buffer isn't currently writing to the L2ARC. 3441 * 4. The L2ARC entry wasn't evicted, which may 3442 * also have invalidated the vdev. 3443 * 5. This isn't prefetch and l2arc_noprefetch is set. 3444 */ 3445 if (hdr->b_l2hdr != NULL && 3446 !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) && 3447 !(l2arc_noprefetch && HDR_PREFETCH(hdr))) { 3448 l2arc_read_callback_t *cb; 3449 3450 DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr); 3451 ARCSTAT_BUMP(arcstat_l2_hits); 3452 3453 cb = kmem_zalloc(sizeof (l2arc_read_callback_t), 3454 KM_SLEEP); 3455 cb->l2rcb_buf = buf; 3456 cb->l2rcb_spa = spa; 3457 cb->l2rcb_bp = *bp; 3458 cb->l2rcb_zb = *zb; 3459 cb->l2rcb_flags = zio_flags; 3460 cb->l2rcb_compress = b_compress; 3461 3462 ASSERT(addr >= VDEV_LABEL_START_SIZE && 3463 addr + size < vd->vdev_psize - 3464 VDEV_LABEL_END_SIZE); 3465 3466 /* 3467 * l2arc read. The SCL_L2ARC lock will be 3468 * released by l2arc_read_done(). 3469 * Issue a null zio if the underlying buffer 3470 * was squashed to zero size by compression. 3471 */ 3472 if (b_compress == ZIO_COMPRESS_EMPTY) { 3473 rzio = zio_null(pio, spa, vd, 3474 l2arc_read_done, cb, 3475 zio_flags | ZIO_FLAG_DONT_CACHE | 3476 ZIO_FLAG_CANFAIL | 3477 ZIO_FLAG_DONT_PROPAGATE | 3478 ZIO_FLAG_DONT_RETRY); 3479 } else { 3480 rzio = zio_read_phys(pio, vd, addr, 3481 b_asize, buf->b_data, 3482 ZIO_CHECKSUM_OFF, 3483 l2arc_read_done, cb, priority, 3484 zio_flags | ZIO_FLAG_DONT_CACHE | 3485 ZIO_FLAG_CANFAIL | 3486 ZIO_FLAG_DONT_PROPAGATE | 3487 ZIO_FLAG_DONT_RETRY, B_FALSE); 3488 } 3489 DTRACE_PROBE2(l2arc__read, vdev_t *, vd, 3490 zio_t *, rzio); 3491 ARCSTAT_INCR(arcstat_l2_read_bytes, b_asize); 3492 3493 if (*arc_flags & ARC_NOWAIT) { 3494 zio_nowait(rzio); 3495 return (0); 3496 } 3497 3498 ASSERT(*arc_flags & ARC_WAIT); 3499 if (zio_wait(rzio) == 0) 3500 return (0); 3501 3502 /* l2arc read error; goto zio_read() */ 3503 } else { 3504 DTRACE_PROBE1(l2arc__miss, 3505 arc_buf_hdr_t *, hdr); 3506 ARCSTAT_BUMP(arcstat_l2_misses); 3507 if (HDR_L2_WRITING(hdr)) 3508 ARCSTAT_BUMP(arcstat_l2_rw_clash); 3509 spa_config_exit(spa, SCL_L2ARC, vd); 3510 } 3511 } else { 3512 if (vd != NULL) 3513 spa_config_exit(spa, SCL_L2ARC, vd); 3514 if (l2arc_ndev != 0) { 3515 DTRACE_PROBE1(l2arc__miss, 3516 arc_buf_hdr_t *, hdr); 3517 ARCSTAT_BUMP(arcstat_l2_misses); 3518 } 3519 } 3520 3521 rzio = zio_read(pio, spa, bp, buf->b_data, size, 3522 arc_read_done, buf, priority, zio_flags, zb); 3523 3524 if (*arc_flags & ARC_WAIT) 3525 return (zio_wait(rzio)); 3526 3527 ASSERT(*arc_flags & ARC_NOWAIT); 3528 zio_nowait(rzio); 3529 } 3530 return (0); 3531} 3532 3533void 3534arc_set_callback(arc_buf_t *buf, arc_evict_func_t *func, void *private) 3535{ 3536 ASSERT(buf->b_hdr != NULL); 3537 ASSERT(buf->b_hdr->b_state != arc_anon); 3538 ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt) || func == NULL); 3539 ASSERT(buf->b_efunc == NULL); 3540 ASSERT(!HDR_BUF_AVAILABLE(buf->b_hdr)); 3541 3542 buf->b_efunc = func; 3543 buf->b_private = private; 3544} 3545 3546/* 3547 * Notify the arc that a block was freed, and thus will never be used again. 3548 */ 3549void 3550arc_freed(spa_t *spa, const blkptr_t *bp) 3551{ 3552 arc_buf_hdr_t *hdr; 3553 kmutex_t *hash_lock; 3554 uint64_t guid = spa_load_guid(spa); 3555 3556 ASSERT(!BP_IS_EMBEDDED(bp)); 3557 3558 hdr = buf_hash_find(guid, bp, &hash_lock); 3559 if (hdr == NULL) 3560 return; 3561 if (HDR_BUF_AVAILABLE(hdr)) { 3562 arc_buf_t *buf = hdr->b_buf; 3563 add_reference(hdr, hash_lock, FTAG); 3564 hdr->b_flags &= ~ARC_BUF_AVAILABLE; 3565 mutex_exit(hash_lock); 3566 3567 arc_release(buf, FTAG); 3568 (void) arc_buf_remove_ref(buf, FTAG); 3569 } else { 3570 mutex_exit(hash_lock); 3571 } 3572 3573} 3574 3575/* 3576 * Clear the user eviction callback set by arc_set_callback(), first calling 3577 * it if it exists. Because the presence of a callback keeps an arc_buf cached 3578 * clearing the callback may result in the arc_buf being destroyed. However, 3579 * it will not result in the *last* arc_buf being destroyed, hence the data 3580 * will remain cached in the ARC. We make a copy of the arc buffer here so 3581 * that we can process the callback without holding any locks. 3582 * 3583 * It's possible that the callback is already in the process of being cleared 3584 * by another thread. In this case we can not clear the callback. 3585 * 3586 * Returns B_TRUE if the callback was successfully called and cleared. 3587 */ 3588boolean_t 3589arc_clear_callback(arc_buf_t *buf) 3590{ 3591 arc_buf_hdr_t *hdr; 3592 kmutex_t *hash_lock; 3593 arc_evict_func_t *efunc = buf->b_efunc; 3594 void *private = buf->b_private; 3595 list_t *list, *evicted_list; 3596 kmutex_t *lock, *evicted_lock; 3597 3598 mutex_enter(&buf->b_evict_lock); 3599 hdr = buf->b_hdr; 3600 if (hdr == NULL) { 3601 /* 3602 * We are in arc_do_user_evicts(). 3603 */ 3604 ASSERT(buf->b_data == NULL); 3605 mutex_exit(&buf->b_evict_lock); 3606 return (B_FALSE); 3607 } else if (buf->b_data == NULL) { 3608 /* 3609 * We are on the eviction list; process this buffer now 3610 * but let arc_do_user_evicts() do the reaping. 3611 */ 3612 buf->b_efunc = NULL; 3613 mutex_exit(&buf->b_evict_lock); 3614 VERIFY0(efunc(private)); 3615 return (B_TRUE); 3616 } 3617 hash_lock = HDR_LOCK(hdr); 3618 mutex_enter(hash_lock); 3619 hdr = buf->b_hdr; 3620 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 3621 3622 ASSERT3U(refcount_count(&hdr->b_refcnt), <, hdr->b_datacnt); 3623 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu); 3624 3625 buf->b_efunc = NULL; 3626 buf->b_private = NULL; 3627 3628 if (hdr->b_datacnt > 1) { 3629 mutex_exit(&buf->b_evict_lock); 3630 arc_buf_destroy(buf, FALSE, TRUE); 3631 } else { 3632 ASSERT(buf == hdr->b_buf); 3633 hdr->b_flags |= ARC_BUF_AVAILABLE; 3634 mutex_exit(&buf->b_evict_lock); 3635 } 3636 3637 mutex_exit(hash_lock); 3638 VERIFY0(efunc(private)); 3639 return (B_TRUE); 3640} 3641 3642/* 3643 * Release this buffer from the cache, making it an anonymous buffer. This 3644 * must be done after a read and prior to modifying the buffer contents. 3645 * If the buffer has more than one reference, we must make 3646 * a new hdr for the buffer. 3647 */ 3648void 3649arc_release(arc_buf_t *buf, void *tag) 3650{ 3651 arc_buf_hdr_t *hdr; 3652 kmutex_t *hash_lock = NULL; 3653 l2arc_buf_hdr_t *l2hdr; 3654 uint64_t buf_size; 3655 3656 /* 3657 * It would be nice to assert that if it's DMU metadata (level > 3658 * 0 || it's the dnode file), then it must be syncing context. 3659 * But we don't know that information at this level. 3660 */ 3661 3662 mutex_enter(&buf->b_evict_lock); 3663 hdr = buf->b_hdr; 3664 3665 /* this buffer is not on any list */ 3666 ASSERT(refcount_count(&hdr->b_refcnt) > 0); 3667 3668 if (hdr->b_state == arc_anon) { 3669 /* this buffer is already released */ 3670 ASSERT(buf->b_efunc == NULL); 3671 } else { 3672 hash_lock = HDR_LOCK(hdr); 3673 mutex_enter(hash_lock); 3674 hdr = buf->b_hdr; 3675 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 3676 } 3677 3678 l2hdr = hdr->b_l2hdr; 3679 if (l2hdr) { 3680 mutex_enter(&l2arc_buflist_mtx); 3681 hdr->b_l2hdr = NULL; 3682 list_remove(l2hdr->b_dev->l2ad_buflist, hdr); 3683 } 3684 buf_size = hdr->b_size; 3685 3686 /* 3687 * Do we have more than one buf? 3688 */ 3689 if (hdr->b_datacnt > 1) { 3690 arc_buf_hdr_t *nhdr; 3691 arc_buf_t **bufp; 3692 uint64_t blksz = hdr->b_size; 3693 uint64_t spa = hdr->b_spa; 3694 arc_buf_contents_t type = hdr->b_type; 3695 uint32_t flags = hdr->b_flags; 3696 3697 ASSERT(hdr->b_buf != buf || buf->b_next != NULL); 3698 /* 3699 * Pull the data off of this hdr and attach it to 3700 * a new anonymous hdr. 3701 */ 3702 (void) remove_reference(hdr, hash_lock, tag); 3703 bufp = &hdr->b_buf; 3704 while (*bufp != buf) 3705 bufp = &(*bufp)->b_next; 3706 *bufp = buf->b_next; 3707 buf->b_next = NULL; 3708 3709 ASSERT3U(hdr->b_state->arcs_size, >=, hdr->b_size); 3710 atomic_add_64(&hdr->b_state->arcs_size, -hdr->b_size); 3711 if (refcount_is_zero(&hdr->b_refcnt)) { 3712 uint64_t *size = &hdr->b_state->arcs_lsize[hdr->b_type]; 3713 ASSERT3U(*size, >=, hdr->b_size); 3714 atomic_add_64(size, -hdr->b_size); 3715 } 3716 3717 /* 3718 * We're releasing a duplicate user data buffer, update 3719 * our statistics accordingly. 3720 */ 3721 if (hdr->b_type == ARC_BUFC_DATA) { 3722 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers); 3723 ARCSTAT_INCR(arcstat_duplicate_buffers_size, 3724 -hdr->b_size); 3725 } 3726 hdr->b_datacnt -= 1; 3727 arc_cksum_verify(buf); 3728#ifdef illumos 3729 arc_buf_unwatch(buf); 3730#endif /* illumos */ 3731 3732 mutex_exit(hash_lock); 3733 3734 nhdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE); 3735 nhdr->b_size = blksz; 3736 nhdr->b_spa = spa; 3737 nhdr->b_type = type; 3738 nhdr->b_buf = buf; 3739 nhdr->b_state = arc_anon; 3740 nhdr->b_arc_access = 0; 3741 nhdr->b_flags = flags & ARC_L2_WRITING; 3742 nhdr->b_l2hdr = NULL; 3743 nhdr->b_datacnt = 1; 3744 nhdr->b_freeze_cksum = NULL; 3745 (void) refcount_add(&nhdr->b_refcnt, tag); 3746 buf->b_hdr = nhdr; 3747 mutex_exit(&buf->b_evict_lock); 3748 atomic_add_64(&arc_anon->arcs_size, blksz); 3749 } else { 3750 mutex_exit(&buf->b_evict_lock); 3751 ASSERT(refcount_count(&hdr->b_refcnt) == 1); 3752 ASSERT(!list_link_active(&hdr->b_arc_node)); 3753 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 3754 if (hdr->b_state != arc_anon) 3755 arc_change_state(arc_anon, hdr, hash_lock); 3756 hdr->b_arc_access = 0; 3757 if (hash_lock) 3758 mutex_exit(hash_lock); 3759 3760 buf_discard_identity(hdr); 3761 arc_buf_thaw(buf); 3762 } 3763 buf->b_efunc = NULL; 3764 buf->b_private = NULL; 3765 3766 if (l2hdr) { 3767 ARCSTAT_INCR(arcstat_l2_asize, -l2hdr->b_asize); 3768 vdev_space_update(l2hdr->b_dev->l2ad_vdev, 3769 -l2hdr->b_asize, 0, 0); 3770 trim_map_free(l2hdr->b_dev->l2ad_vdev, l2hdr->b_daddr, 3771 hdr->b_size, 0); 3772 kmem_free(l2hdr, sizeof (l2arc_buf_hdr_t)); 3773 ARCSTAT_INCR(arcstat_l2_size, -buf_size); 3774 mutex_exit(&l2arc_buflist_mtx); 3775 } 3776} 3777 3778int 3779arc_released(arc_buf_t *buf) 3780{ 3781 int released; 3782 3783 mutex_enter(&buf->b_evict_lock); 3784 released = (buf->b_data != NULL && buf->b_hdr->b_state == arc_anon); 3785 mutex_exit(&buf->b_evict_lock); 3786 return (released); 3787} 3788 3789#ifdef ZFS_DEBUG 3790int 3791arc_referenced(arc_buf_t *buf) 3792{ 3793 int referenced; 3794 3795 mutex_enter(&buf->b_evict_lock); 3796 referenced = (refcount_count(&buf->b_hdr->b_refcnt)); 3797 mutex_exit(&buf->b_evict_lock); 3798 return (referenced); 3799} 3800#endif 3801 3802static void 3803arc_write_ready(zio_t *zio) 3804{ 3805 arc_write_callback_t *callback = zio->io_private; 3806 arc_buf_t *buf = callback->awcb_buf; 3807 arc_buf_hdr_t *hdr = buf->b_hdr; 3808 3809 ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt)); 3810 callback->awcb_ready(zio, buf, callback->awcb_private); 3811 3812 /* 3813 * If the IO is already in progress, then this is a re-write 3814 * attempt, so we need to thaw and re-compute the cksum. 3815 * It is the responsibility of the callback to handle the 3816 * accounting for any re-write attempt. 3817 */ 3818 if (HDR_IO_IN_PROGRESS(hdr)) { 3819 mutex_enter(&hdr->b_freeze_lock); 3820 if (hdr->b_freeze_cksum != NULL) { 3821 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 3822 hdr->b_freeze_cksum = NULL; 3823 } 3824 mutex_exit(&hdr->b_freeze_lock); 3825 } 3826 arc_cksum_compute(buf, B_FALSE); 3827 hdr->b_flags |= ARC_IO_IN_PROGRESS; 3828} 3829 3830/* 3831 * The SPA calls this callback for each physical write that happens on behalf 3832 * of a logical write. See the comment in dbuf_write_physdone() for details. 3833 */ 3834static void 3835arc_write_physdone(zio_t *zio) 3836{ 3837 arc_write_callback_t *cb = zio->io_private; 3838 if (cb->awcb_physdone != NULL) 3839 cb->awcb_physdone(zio, cb->awcb_buf, cb->awcb_private); 3840} 3841 3842static void 3843arc_write_done(zio_t *zio) 3844{ 3845 arc_write_callback_t *callback = zio->io_private; 3846 arc_buf_t *buf = callback->awcb_buf; 3847 arc_buf_hdr_t *hdr = buf->b_hdr; 3848 3849 ASSERT(hdr->b_acb == NULL); 3850 3851 if (zio->io_error == 0) { 3852 if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) { 3853 buf_discard_identity(hdr); 3854 } else { 3855 hdr->b_dva = *BP_IDENTITY(zio->io_bp); 3856 hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp); 3857 hdr->b_cksum0 = zio->io_bp->blk_cksum.zc_word[0]; 3858 } 3859 } else { 3860 ASSERT(BUF_EMPTY(hdr)); 3861 } 3862 3863 /* 3864 * If the block to be written was all-zero or compressed enough to be 3865 * embedded in the BP, no write was performed so there will be no 3866 * dva/birth/checksum. The buffer must therefore remain anonymous 3867 * (and uncached). 3868 */ 3869 if (!BUF_EMPTY(hdr)) { 3870 arc_buf_hdr_t *exists; 3871 kmutex_t *hash_lock; 3872 3873 ASSERT(zio->io_error == 0); 3874 3875 arc_cksum_verify(buf); 3876 3877 exists = buf_hash_insert(hdr, &hash_lock); 3878 if (exists) { 3879 /* 3880 * This can only happen if we overwrite for 3881 * sync-to-convergence, because we remove 3882 * buffers from the hash table when we arc_free(). 3883 */ 3884 if (zio->io_flags & ZIO_FLAG_IO_REWRITE) { 3885 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) 3886 panic("bad overwrite, hdr=%p exists=%p", 3887 (void *)hdr, (void *)exists); 3888 ASSERT(refcount_is_zero(&exists->b_refcnt)); 3889 arc_change_state(arc_anon, exists, hash_lock); 3890 mutex_exit(hash_lock); 3891 arc_hdr_destroy(exists); 3892 exists = buf_hash_insert(hdr, &hash_lock); 3893 ASSERT3P(exists, ==, NULL); 3894 } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) { 3895 /* nopwrite */ 3896 ASSERT(zio->io_prop.zp_nopwrite); 3897 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) 3898 panic("bad nopwrite, hdr=%p exists=%p", 3899 (void *)hdr, (void *)exists); 3900 } else { 3901 /* Dedup */ 3902 ASSERT(hdr->b_datacnt == 1); 3903 ASSERT(hdr->b_state == arc_anon); 3904 ASSERT(BP_GET_DEDUP(zio->io_bp)); 3905 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0); 3906 } 3907 } 3908 hdr->b_flags &= ~ARC_IO_IN_PROGRESS; 3909 /* if it's not anon, we are doing a scrub */ 3910 if (!exists && hdr->b_state == arc_anon) 3911 arc_access(hdr, hash_lock); 3912 mutex_exit(hash_lock); 3913 } else { 3914 hdr->b_flags &= ~ARC_IO_IN_PROGRESS; 3915 } 3916 3917 ASSERT(!refcount_is_zero(&hdr->b_refcnt)); 3918 callback->awcb_done(zio, buf, callback->awcb_private); 3919 3920 kmem_free(callback, sizeof (arc_write_callback_t)); 3921} 3922 3923zio_t * 3924arc_write(zio_t *pio, spa_t *spa, uint64_t txg, 3925 blkptr_t *bp, arc_buf_t *buf, boolean_t l2arc, boolean_t l2arc_compress, 3926 const zio_prop_t *zp, arc_done_func_t *ready, arc_done_func_t *physdone, 3927 arc_done_func_t *done, void *private, zio_priority_t priority, 3928 int zio_flags, const zbookmark_phys_t *zb) 3929{ 3930 arc_buf_hdr_t *hdr = buf->b_hdr; 3931 arc_write_callback_t *callback; 3932 zio_t *zio; 3933 3934 ASSERT(ready != NULL); 3935 ASSERT(done != NULL); 3936 ASSERT(!HDR_IO_ERROR(hdr)); 3937 ASSERT((hdr->b_flags & ARC_IO_IN_PROGRESS) == 0); 3938 ASSERT(hdr->b_acb == NULL); 3939 if (l2arc) 3940 hdr->b_flags |= ARC_L2CACHE; 3941 if (l2arc_compress) 3942 hdr->b_flags |= ARC_L2COMPRESS; 3943 callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP); 3944 callback->awcb_ready = ready; 3945 callback->awcb_physdone = physdone; 3946 callback->awcb_done = done; 3947 callback->awcb_private = private; 3948 callback->awcb_buf = buf; 3949 3950 zio = zio_write(pio, spa, txg, bp, buf->b_data, hdr->b_size, zp, 3951 arc_write_ready, arc_write_physdone, arc_write_done, callback, 3952 priority, zio_flags, zb); 3953 3954 return (zio); 3955} 3956 3957static int 3958arc_memory_throttle(uint64_t reserve, uint64_t txg) 3959{ 3960#ifdef _KERNEL 3961 uint64_t available_memory = ptob(freemem); 3962 static uint64_t page_load = 0; 3963 static uint64_t last_txg = 0; 3964 3965#if defined(__i386) || !defined(UMA_MD_SMALL_ALLOC) 3966 available_memory = 3967 MIN(available_memory, ptob(vmem_size(heap_arena, VMEM_FREE))); 3968#endif 3969 3970 if (freemem > (uint64_t)physmem * arc_lotsfree_percent / 100) 3971 return (0); 3972 3973 if (txg > last_txg) { 3974 last_txg = txg; 3975 page_load = 0; 3976 } 3977 /* 3978 * If we are in pageout, we know that memory is already tight, 3979 * the arc is already going to be evicting, so we just want to 3980 * continue to let page writes occur as quickly as possible. 3981 */ 3982 if (curproc == pageproc) { 3983 if (page_load > MAX(ptob(minfree), available_memory) / 4) 3984 return (SET_ERROR(ERESTART)); 3985 /* Note: reserve is inflated, so we deflate */ 3986 page_load += reserve / 8; 3987 return (0); 3988 } else if (page_load > 0 && arc_reclaim_needed()) { 3989 /* memory is low, delay before restarting */ 3990 ARCSTAT_INCR(arcstat_memory_throttle_count, 1); 3991 return (SET_ERROR(EAGAIN)); 3992 } 3993 page_load = 0; 3994#endif 3995 return (0); 3996} 3997 3998void 3999arc_tempreserve_clear(uint64_t reserve) 4000{ 4001 atomic_add_64(&arc_tempreserve, -reserve); 4002 ASSERT((int64_t)arc_tempreserve >= 0); 4003} 4004 4005int 4006arc_tempreserve_space(uint64_t reserve, uint64_t txg) 4007{ 4008 int error; 4009 uint64_t anon_size; 4010 4011 if (reserve > arc_c/4 && !arc_no_grow) { 4012 arc_c = MIN(arc_c_max, reserve * 4); 4013 DTRACE_PROBE1(arc__set_reserve, uint64_t, arc_c); 4014 } 4015 if (reserve > arc_c) 4016 return (SET_ERROR(ENOMEM)); 4017 4018 /* 4019 * Don't count loaned bufs as in flight dirty data to prevent long 4020 * network delays from blocking transactions that are ready to be 4021 * assigned to a txg. 4022 */ 4023 anon_size = MAX((int64_t)(arc_anon->arcs_size - arc_loaned_bytes), 0); 4024 4025 /* 4026 * Writes will, almost always, require additional memory allocations 4027 * in order to compress/encrypt/etc the data. We therefore need to 4028 * make sure that there is sufficient available memory for this. 4029 */ 4030 error = arc_memory_throttle(reserve, txg); 4031 if (error != 0) 4032 return (error); 4033 4034 /* 4035 * Throttle writes when the amount of dirty data in the cache 4036 * gets too large. We try to keep the cache less than half full 4037 * of dirty blocks so that our sync times don't grow too large. 4038 * Note: if two requests come in concurrently, we might let them 4039 * both succeed, when one of them should fail. Not a huge deal. 4040 */ 4041 4042 if (reserve + arc_tempreserve + anon_size > arc_c / 2 && 4043 anon_size > arc_c / 4) { 4044 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK " 4045 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n", 4046 arc_tempreserve>>10, 4047 arc_anon->arcs_lsize[ARC_BUFC_METADATA]>>10, 4048 arc_anon->arcs_lsize[ARC_BUFC_DATA]>>10, 4049 reserve>>10, arc_c>>10); 4050 return (SET_ERROR(ERESTART)); 4051 } 4052 atomic_add_64(&arc_tempreserve, reserve); 4053 return (0); 4054} 4055 4056static kmutex_t arc_lowmem_lock; 4057#ifdef _KERNEL 4058static eventhandler_tag arc_event_lowmem = NULL; 4059 4060static void 4061arc_lowmem(void *arg __unused, int howto __unused) 4062{ 4063 4064 /* Serialize access via arc_lowmem_lock. */ 4065 mutex_enter(&arc_lowmem_lock); 4066 mutex_enter(&arc_reclaim_thr_lock); 4067 needfree = 1; 4068 DTRACE_PROBE(arc__needfree); 4069 cv_signal(&arc_reclaim_thr_cv); 4070 4071 /* 4072 * It is unsafe to block here in arbitrary threads, because we can come 4073 * here from ARC itself and may hold ARC locks and thus risk a deadlock 4074 * with ARC reclaim thread. 4075 */ 4076 if (curproc == pageproc) { 4077 while (needfree) 4078 msleep(&needfree, &arc_reclaim_thr_lock, 0, "zfs:lowmem", 0); 4079 } 4080 mutex_exit(&arc_reclaim_thr_lock); 4081 mutex_exit(&arc_lowmem_lock); 4082} 4083#endif 4084 4085void 4086arc_init(void) 4087{ 4088 int i, prefetch_tunable_set = 0; 4089 4090 mutex_init(&arc_reclaim_thr_lock, NULL, MUTEX_DEFAULT, NULL); 4091 cv_init(&arc_reclaim_thr_cv, NULL, CV_DEFAULT, NULL); 4092 mutex_init(&arc_lowmem_lock, NULL, MUTEX_DEFAULT, NULL); 4093 4094 /* Convert seconds to clock ticks */ 4095 arc_min_prefetch_lifespan = 1 * hz; 4096 4097 /* Start out with 1/8 of all memory */ 4098 arc_c = kmem_size() / 8; 4099 4100#ifdef sun 4101#ifdef _KERNEL 4102 /* 4103 * On architectures where the physical memory can be larger 4104 * than the addressable space (intel in 32-bit mode), we may 4105 * need to limit the cache to 1/8 of VM size. 4106 */ 4107 arc_c = MIN(arc_c, vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 8); 4108#endif 4109#endif /* sun */ 4110 /* set min cache to 1/32 of all memory, or 16MB, whichever is more */ 4111 arc_c_min = MAX(arc_c / 4, 64<<18); 4112 /* set max to 1/2 of all memory, or all but 1GB, whichever is more */ 4113 if (arc_c * 8 >= 1<<30) 4114 arc_c_max = (arc_c * 8) - (1<<30); 4115 else 4116 arc_c_max = arc_c_min; 4117 arc_c_max = MAX(arc_c * 5, arc_c_max); 4118 4119#ifdef _KERNEL 4120 /* 4121 * Allow the tunables to override our calculations if they are 4122 * reasonable (ie. over 16MB) 4123 */ 4124 if (zfs_arc_max > 64<<18 && zfs_arc_max < kmem_size()) 4125 arc_c_max = zfs_arc_max; 4126 if (zfs_arc_min > 64<<18 && zfs_arc_min <= arc_c_max) 4127 arc_c_min = zfs_arc_min; 4128#endif 4129 4130 arc_c = arc_c_max; 4131 arc_p = (arc_c >> 1); 4132 4133 /* limit meta-data to 1/4 of the arc capacity */ 4134 arc_meta_limit = arc_c_max / 4; 4135 4136 /* Allow the tunable to override if it is reasonable */ 4137 if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max) 4138 arc_meta_limit = zfs_arc_meta_limit; 4139 4140 if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0) 4141 arc_c_min = arc_meta_limit / 2; 4142 4143 if (zfs_arc_grow_retry > 0) 4144 arc_grow_retry = zfs_arc_grow_retry; 4145 4146 if (zfs_arc_shrink_shift > 0) 4147 arc_shrink_shift = zfs_arc_shrink_shift; 4148 4149 if (zfs_arc_p_min_shift > 0) 4150 arc_p_min_shift = zfs_arc_p_min_shift; 4151 4152 /* if kmem_flags are set, lets try to use less memory */ 4153 if (kmem_debugging()) 4154 arc_c = arc_c / 2; 4155 if (arc_c < arc_c_min) 4156 arc_c = arc_c_min; 4157 4158 zfs_arc_min = arc_c_min; 4159 zfs_arc_max = arc_c_max; 4160 4161 arc_anon = &ARC_anon; 4162 arc_mru = &ARC_mru; 4163 arc_mru_ghost = &ARC_mru_ghost; 4164 arc_mfu = &ARC_mfu; 4165 arc_mfu_ghost = &ARC_mfu_ghost; 4166 arc_l2c_only = &ARC_l2c_only; 4167 arc_size = 0; 4168 4169 for (i = 0; i < ARC_BUFC_NUMLISTS; i++) { 4170 mutex_init(&arc_anon->arcs_locks[i].arcs_lock, 4171 NULL, MUTEX_DEFAULT, NULL); 4172 mutex_init(&arc_mru->arcs_locks[i].arcs_lock, 4173 NULL, MUTEX_DEFAULT, NULL); 4174 mutex_init(&arc_mru_ghost->arcs_locks[i].arcs_lock, 4175 NULL, MUTEX_DEFAULT, NULL); 4176 mutex_init(&arc_mfu->arcs_locks[i].arcs_lock, 4177 NULL, MUTEX_DEFAULT, NULL); 4178 mutex_init(&arc_mfu_ghost->arcs_locks[i].arcs_lock, 4179 NULL, MUTEX_DEFAULT, NULL); 4180 mutex_init(&arc_l2c_only->arcs_locks[i].arcs_lock, 4181 NULL, MUTEX_DEFAULT, NULL); 4182 4183 list_create(&arc_mru->arcs_lists[i], 4184 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4185 list_create(&arc_mru_ghost->arcs_lists[i], 4186 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4187 list_create(&arc_mfu->arcs_lists[i], 4188 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4189 list_create(&arc_mfu_ghost->arcs_lists[i], 4190 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4191 list_create(&arc_mfu_ghost->arcs_lists[i], 4192 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4193 list_create(&arc_l2c_only->arcs_lists[i], 4194 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4195 } 4196 4197 buf_init(); 4198 4199 arc_thread_exit = 0; 4200 arc_eviction_list = NULL; 4201 mutex_init(&arc_eviction_mtx, NULL, MUTEX_DEFAULT, NULL); 4202 bzero(&arc_eviction_hdr, sizeof (arc_buf_hdr_t)); 4203 4204 arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED, 4205 sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL); 4206 4207 if (arc_ksp != NULL) { 4208 arc_ksp->ks_data = &arc_stats; 4209 kstat_install(arc_ksp); 4210 } 4211 4212 (void) thread_create(NULL, 0, arc_reclaim_thread, NULL, 0, &p0, 4213 TS_RUN, minclsyspri); 4214 4215#ifdef _KERNEL 4216 arc_event_lowmem = EVENTHANDLER_REGISTER(vm_lowmem, arc_lowmem, NULL, 4217 EVENTHANDLER_PRI_FIRST); 4218#endif 4219 4220 arc_dead = FALSE; 4221 arc_warm = B_FALSE; 4222 4223 /* 4224 * Calculate maximum amount of dirty data per pool. 4225 * 4226 * If it has been set by /etc/system, take that. 4227 * Otherwise, use a percentage of physical memory defined by 4228 * zfs_dirty_data_max_percent (default 10%) with a cap at 4229 * zfs_dirty_data_max_max (default 4GB). 4230 */ 4231 if (zfs_dirty_data_max == 0) { 4232 zfs_dirty_data_max = ptob(physmem) * 4233 zfs_dirty_data_max_percent / 100; 4234 zfs_dirty_data_max = MIN(zfs_dirty_data_max, 4235 zfs_dirty_data_max_max); 4236 } 4237 4238#ifdef _KERNEL 4239 if (TUNABLE_INT_FETCH("vfs.zfs.prefetch_disable", &zfs_prefetch_disable)) 4240 prefetch_tunable_set = 1; 4241 4242#ifdef __i386__ 4243 if (prefetch_tunable_set == 0) { 4244 printf("ZFS NOTICE: Prefetch is disabled by default on i386 " 4245 "-- to enable,\n"); 4246 printf(" add \"vfs.zfs.prefetch_disable=0\" " 4247 "to /boot/loader.conf.\n"); 4248 zfs_prefetch_disable = 1; 4249 } 4250#else 4251 if ((((uint64_t)physmem * PAGESIZE) < (1ULL << 32)) && 4252 prefetch_tunable_set == 0) { 4253 printf("ZFS NOTICE: Prefetch is disabled by default if less " 4254 "than 4GB of RAM is present;\n" 4255 " to enable, add \"vfs.zfs.prefetch_disable=0\" " 4256 "to /boot/loader.conf.\n"); 4257 zfs_prefetch_disable = 1; 4258 } 4259#endif 4260 /* Warn about ZFS memory and address space requirements. */ 4261 if (((uint64_t)physmem * PAGESIZE) < (256 + 128 + 64) * (1 << 20)) { 4262 printf("ZFS WARNING: Recommended minimum RAM size is 512MB; " 4263 "expect unstable behavior.\n"); 4264 } 4265 if (kmem_size() < 512 * (1 << 20)) { 4266 printf("ZFS WARNING: Recommended minimum kmem_size is 512MB; " 4267 "expect unstable behavior.\n"); 4268 printf(" Consider tuning vm.kmem_size and " 4269 "vm.kmem_size_max\n"); 4270 printf(" in /boot/loader.conf.\n"); 4271 } 4272#endif 4273} 4274 4275void 4276arc_fini(void) 4277{ 4278 int i; 4279 4280 mutex_enter(&arc_reclaim_thr_lock); 4281 arc_thread_exit = 1; 4282 cv_signal(&arc_reclaim_thr_cv); 4283 while (arc_thread_exit != 0) 4284 cv_wait(&arc_reclaim_thr_cv, &arc_reclaim_thr_lock); 4285 mutex_exit(&arc_reclaim_thr_lock); 4286 4287 arc_flush(NULL); 4288 4289 arc_dead = TRUE; 4290 4291 if (arc_ksp != NULL) { 4292 kstat_delete(arc_ksp); 4293 arc_ksp = NULL; 4294 } 4295 4296 mutex_destroy(&arc_eviction_mtx); 4297 mutex_destroy(&arc_reclaim_thr_lock); 4298 cv_destroy(&arc_reclaim_thr_cv); 4299 4300 for (i = 0; i < ARC_BUFC_NUMLISTS; i++) { 4301 list_destroy(&arc_mru->arcs_lists[i]); 4302 list_destroy(&arc_mru_ghost->arcs_lists[i]); 4303 list_destroy(&arc_mfu->arcs_lists[i]); 4304 list_destroy(&arc_mfu_ghost->arcs_lists[i]); 4305 list_destroy(&arc_l2c_only->arcs_lists[i]); 4306 4307 mutex_destroy(&arc_anon->arcs_locks[i].arcs_lock); 4308 mutex_destroy(&arc_mru->arcs_locks[i].arcs_lock); 4309 mutex_destroy(&arc_mru_ghost->arcs_locks[i].arcs_lock); 4310 mutex_destroy(&arc_mfu->arcs_locks[i].arcs_lock); 4311 mutex_destroy(&arc_mfu_ghost->arcs_locks[i].arcs_lock); 4312 mutex_destroy(&arc_l2c_only->arcs_locks[i].arcs_lock); 4313 } 4314 4315 buf_fini(); 4316 4317 ASSERT(arc_loaned_bytes == 0); 4318 4319 mutex_destroy(&arc_lowmem_lock); 4320#ifdef _KERNEL 4321 if (arc_event_lowmem != NULL) 4322 EVENTHANDLER_DEREGISTER(vm_lowmem, arc_event_lowmem); 4323#endif 4324} 4325 4326/* 4327 * Level 2 ARC 4328 * 4329 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk. 4330 * It uses dedicated storage devices to hold cached data, which are populated 4331 * using large infrequent writes. The main role of this cache is to boost 4332 * the performance of random read workloads. The intended L2ARC devices 4333 * include short-stroked disks, solid state disks, and other media with 4334 * substantially faster read latency than disk. 4335 * 4336 * +-----------------------+ 4337 * | ARC | 4338 * +-----------------------+ 4339 * | ^ ^ 4340 * | | | 4341 * l2arc_feed_thread() arc_read() 4342 * | | | 4343 * | l2arc read | 4344 * V | | 4345 * +---------------+ | 4346 * | L2ARC | | 4347 * +---------------+ | 4348 * | ^ | 4349 * l2arc_write() | | 4350 * | | | 4351 * V | | 4352 * +-------+ +-------+ 4353 * | vdev | | vdev | 4354 * | cache | | cache | 4355 * +-------+ +-------+ 4356 * +=========+ .-----. 4357 * : L2ARC : |-_____-| 4358 * : devices : | Disks | 4359 * +=========+ `-_____-' 4360 * 4361 * Read requests are satisfied from the following sources, in order: 4362 * 4363 * 1) ARC 4364 * 2) vdev cache of L2ARC devices 4365 * 3) L2ARC devices 4366 * 4) vdev cache of disks 4367 * 5) disks 4368 * 4369 * Some L2ARC device types exhibit extremely slow write performance. 4370 * To accommodate for this there are some significant differences between 4371 * the L2ARC and traditional cache design: 4372 * 4373 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from 4374 * the ARC behave as usual, freeing buffers and placing headers on ghost 4375 * lists. The ARC does not send buffers to the L2ARC during eviction as 4376 * this would add inflated write latencies for all ARC memory pressure. 4377 * 4378 * 2. The L2ARC attempts to cache data from the ARC before it is evicted. 4379 * It does this by periodically scanning buffers from the eviction-end of 4380 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are 4381 * not already there. It scans until a headroom of buffers is satisfied, 4382 * which itself is a buffer for ARC eviction. If a compressible buffer is 4383 * found during scanning and selected for writing to an L2ARC device, we 4384 * temporarily boost scanning headroom during the next scan cycle to make 4385 * sure we adapt to compression effects (which might significantly reduce 4386 * the data volume we write to L2ARC). The thread that does this is 4387 * l2arc_feed_thread(), illustrated below; example sizes are included to 4388 * provide a better sense of ratio than this diagram: 4389 * 4390 * head --> tail 4391 * +---------------------+----------+ 4392 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC 4393 * +---------------------+----------+ | o L2ARC eligible 4394 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer 4395 * +---------------------+----------+ | 4396 * 15.9 Gbytes ^ 32 Mbytes | 4397 * headroom | 4398 * l2arc_feed_thread() 4399 * | 4400 * l2arc write hand <--[oooo]--' 4401 * | 8 Mbyte 4402 * | write max 4403 * V 4404 * +==============================+ 4405 * L2ARC dev |####|#|###|###| |####| ... | 4406 * +==============================+ 4407 * 32 Gbytes 4408 * 4409 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of 4410 * evicted, then the L2ARC has cached a buffer much sooner than it probably 4411 * needed to, potentially wasting L2ARC device bandwidth and storage. It is 4412 * safe to say that this is an uncommon case, since buffers at the end of 4413 * the ARC lists have moved there due to inactivity. 4414 * 4415 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom, 4416 * then the L2ARC simply misses copying some buffers. This serves as a 4417 * pressure valve to prevent heavy read workloads from both stalling the ARC 4418 * with waits and clogging the L2ARC with writes. This also helps prevent 4419 * the potential for the L2ARC to churn if it attempts to cache content too 4420 * quickly, such as during backups of the entire pool. 4421 * 4422 * 5. After system boot and before the ARC has filled main memory, there are 4423 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru 4424 * lists can remain mostly static. Instead of searching from tail of these 4425 * lists as pictured, the l2arc_feed_thread() will search from the list heads 4426 * for eligible buffers, greatly increasing its chance of finding them. 4427 * 4428 * The L2ARC device write speed is also boosted during this time so that 4429 * the L2ARC warms up faster. Since there have been no ARC evictions yet, 4430 * there are no L2ARC reads, and no fear of degrading read performance 4431 * through increased writes. 4432 * 4433 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that 4434 * the vdev queue can aggregate them into larger and fewer writes. Each 4435 * device is written to in a rotor fashion, sweeping writes through 4436 * available space then repeating. 4437 * 4438 * 7. The L2ARC does not store dirty content. It never needs to flush 4439 * write buffers back to disk based storage. 4440 * 4441 * 8. If an ARC buffer is written (and dirtied) which also exists in the 4442 * L2ARC, the now stale L2ARC buffer is immediately dropped. 4443 * 4444 * The performance of the L2ARC can be tweaked by a number of tunables, which 4445 * may be necessary for different workloads: 4446 * 4447 * l2arc_write_max max write bytes per interval 4448 * l2arc_write_boost extra write bytes during device warmup 4449 * l2arc_noprefetch skip caching prefetched buffers 4450 * l2arc_headroom number of max device writes to precache 4451 * l2arc_headroom_boost when we find compressed buffers during ARC 4452 * scanning, we multiply headroom by this 4453 * percentage factor for the next scan cycle, 4454 * since more compressed buffers are likely to 4455 * be present 4456 * l2arc_feed_secs seconds between L2ARC writing 4457 * 4458 * Tunables may be removed or added as future performance improvements are 4459 * integrated, and also may become zpool properties. 4460 * 4461 * There are three key functions that control how the L2ARC warms up: 4462 * 4463 * l2arc_write_eligible() check if a buffer is eligible to cache 4464 * l2arc_write_size() calculate how much to write 4465 * l2arc_write_interval() calculate sleep delay between writes 4466 * 4467 * These three functions determine what to write, how much, and how quickly 4468 * to send writes. 4469 */ 4470 4471static boolean_t 4472l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *ab) 4473{ 4474 /* 4475 * A buffer is *not* eligible for the L2ARC if it: 4476 * 1. belongs to a different spa. 4477 * 2. is already cached on the L2ARC. 4478 * 3. has an I/O in progress (it may be an incomplete read). 4479 * 4. is flagged not eligible (zfs property). 4480 */ 4481 if (ab->b_spa != spa_guid) { 4482 ARCSTAT_BUMP(arcstat_l2_write_spa_mismatch); 4483 return (B_FALSE); 4484 } 4485 if (ab->b_l2hdr != NULL) { 4486 ARCSTAT_BUMP(arcstat_l2_write_in_l2); 4487 return (B_FALSE); 4488 } 4489 if (HDR_IO_IN_PROGRESS(ab)) { 4490 ARCSTAT_BUMP(arcstat_l2_write_hdr_io_in_progress); 4491 return (B_FALSE); 4492 } 4493 if (!HDR_L2CACHE(ab)) { 4494 ARCSTAT_BUMP(arcstat_l2_write_not_cacheable); 4495 return (B_FALSE); 4496 } 4497 4498 return (B_TRUE); 4499} 4500 4501static uint64_t 4502l2arc_write_size(void) 4503{ 4504 uint64_t size; 4505 4506 /* 4507 * Make sure our globals have meaningful values in case the user 4508 * altered them. 4509 */ 4510 size = l2arc_write_max; 4511 if (size == 0) { 4512 cmn_err(CE_NOTE, "Bad value for l2arc_write_max, value must " 4513 "be greater than zero, resetting it to the default (%d)", 4514 L2ARC_WRITE_SIZE); 4515 size = l2arc_write_max = L2ARC_WRITE_SIZE; 4516 } 4517 4518 if (arc_warm == B_FALSE) 4519 size += l2arc_write_boost; 4520 4521 return (size); 4522 4523} 4524 4525static clock_t 4526l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote) 4527{ 4528 clock_t interval, next, now; 4529 4530 /* 4531 * If the ARC lists are busy, increase our write rate; if the 4532 * lists are stale, idle back. This is achieved by checking 4533 * how much we previously wrote - if it was more than half of 4534 * what we wanted, schedule the next write much sooner. 4535 */ 4536 if (l2arc_feed_again && wrote > (wanted / 2)) 4537 interval = (hz * l2arc_feed_min_ms) / 1000; 4538 else 4539 interval = hz * l2arc_feed_secs; 4540 4541 now = ddi_get_lbolt(); 4542 next = MAX(now, MIN(now + interval, began + interval)); 4543 4544 return (next); 4545} 4546 4547static void 4548l2arc_hdr_stat_add(void) 4549{ 4550 ARCSTAT_INCR(arcstat_l2_hdr_size, HDR_SIZE + L2HDR_SIZE); 4551 ARCSTAT_INCR(arcstat_hdr_size, -HDR_SIZE); 4552} 4553 4554static void 4555l2arc_hdr_stat_remove(void) 4556{ 4557 ARCSTAT_INCR(arcstat_l2_hdr_size, -(HDR_SIZE + L2HDR_SIZE)); 4558 ARCSTAT_INCR(arcstat_hdr_size, HDR_SIZE); 4559} 4560 4561/* 4562 * Cycle through L2ARC devices. This is how L2ARC load balances. 4563 * If a device is returned, this also returns holding the spa config lock. 4564 */ 4565static l2arc_dev_t * 4566l2arc_dev_get_next(void) 4567{ 4568 l2arc_dev_t *first, *next = NULL; 4569 4570 /* 4571 * Lock out the removal of spas (spa_namespace_lock), then removal 4572 * of cache devices (l2arc_dev_mtx). Once a device has been selected, 4573 * both locks will be dropped and a spa config lock held instead. 4574 */ 4575 mutex_enter(&spa_namespace_lock); 4576 mutex_enter(&l2arc_dev_mtx); 4577 4578 /* if there are no vdevs, there is nothing to do */ 4579 if (l2arc_ndev == 0) 4580 goto out; 4581 4582 first = NULL; 4583 next = l2arc_dev_last; 4584 do { 4585 /* loop around the list looking for a non-faulted vdev */ 4586 if (next == NULL) { 4587 next = list_head(l2arc_dev_list); 4588 } else { 4589 next = list_next(l2arc_dev_list, next); 4590 if (next == NULL) 4591 next = list_head(l2arc_dev_list); 4592 } 4593 4594 /* if we have come back to the start, bail out */ 4595 if (first == NULL) 4596 first = next; 4597 else if (next == first) 4598 break; 4599 4600 } while (vdev_is_dead(next->l2ad_vdev)); 4601 4602 /* if we were unable to find any usable vdevs, return NULL */ 4603 if (vdev_is_dead(next->l2ad_vdev)) 4604 next = NULL; 4605 4606 l2arc_dev_last = next; 4607 4608out: 4609 mutex_exit(&l2arc_dev_mtx); 4610 4611 /* 4612 * Grab the config lock to prevent the 'next' device from being 4613 * removed while we are writing to it. 4614 */ 4615 if (next != NULL) 4616 spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER); 4617 mutex_exit(&spa_namespace_lock); 4618 4619 return (next); 4620} 4621 4622/* 4623 * Free buffers that were tagged for destruction. 4624 */ 4625static void 4626l2arc_do_free_on_write() 4627{ 4628 list_t *buflist; 4629 l2arc_data_free_t *df, *df_prev; 4630 4631 mutex_enter(&l2arc_free_on_write_mtx); 4632 buflist = l2arc_free_on_write; 4633 4634 for (df = list_tail(buflist); df; df = df_prev) { 4635 df_prev = list_prev(buflist, df); 4636 ASSERT(df->l2df_data != NULL); 4637 ASSERT(df->l2df_func != NULL); 4638 df->l2df_func(df->l2df_data, df->l2df_size); 4639 list_remove(buflist, df); 4640 kmem_free(df, sizeof (l2arc_data_free_t)); 4641 } 4642 4643 mutex_exit(&l2arc_free_on_write_mtx); 4644} 4645 4646/* 4647 * A write to a cache device has completed. Update all headers to allow 4648 * reads from these buffers to begin. 4649 */ 4650static void 4651l2arc_write_done(zio_t *zio) 4652{ 4653 l2arc_write_callback_t *cb; 4654 l2arc_dev_t *dev; 4655 list_t *buflist; 4656 arc_buf_hdr_t *head, *ab, *ab_prev; 4657 l2arc_buf_hdr_t *abl2; 4658 kmutex_t *hash_lock; 4659 int64_t bytes_dropped = 0; 4660 4661 cb = zio->io_private; 4662 ASSERT(cb != NULL); 4663 dev = cb->l2wcb_dev; 4664 ASSERT(dev != NULL); 4665 head = cb->l2wcb_head; 4666 ASSERT(head != NULL); 4667 buflist = dev->l2ad_buflist; 4668 ASSERT(buflist != NULL); 4669 DTRACE_PROBE2(l2arc__iodone, zio_t *, zio, 4670 l2arc_write_callback_t *, cb); 4671 4672 if (zio->io_error != 0) 4673 ARCSTAT_BUMP(arcstat_l2_writes_error); 4674 4675 mutex_enter(&l2arc_buflist_mtx); 4676 4677 /* 4678 * All writes completed, or an error was hit. 4679 */ 4680 for (ab = list_prev(buflist, head); ab; ab = ab_prev) { 4681 ab_prev = list_prev(buflist, ab); 4682 abl2 = ab->b_l2hdr; 4683 4684 /* 4685 * Release the temporary compressed buffer as soon as possible. 4686 */ 4687 if (abl2->b_compress != ZIO_COMPRESS_OFF) 4688 l2arc_release_cdata_buf(ab); 4689 4690 hash_lock = HDR_LOCK(ab); 4691 if (!mutex_tryenter(hash_lock)) { 4692 /* 4693 * This buffer misses out. It may be in a stage 4694 * of eviction. Its ARC_L2_WRITING flag will be 4695 * left set, denying reads to this buffer. 4696 */ 4697 ARCSTAT_BUMP(arcstat_l2_writes_hdr_miss); 4698 continue; 4699 } 4700 4701 if (zio->io_error != 0) { 4702 /* 4703 * Error - drop L2ARC entry. 4704 */ 4705 list_remove(buflist, ab); 4706 ARCSTAT_INCR(arcstat_l2_asize, -abl2->b_asize); 4707 bytes_dropped += abl2->b_asize; 4708 ab->b_l2hdr = NULL; 4709 trim_map_free(abl2->b_dev->l2ad_vdev, abl2->b_daddr, 4710 ab->b_size, 0); 4711 kmem_free(abl2, sizeof (l2arc_buf_hdr_t)); 4712 ARCSTAT_INCR(arcstat_l2_size, -ab->b_size); 4713 } 4714 4715 /* 4716 * Allow ARC to begin reads to this L2ARC entry. 4717 */ 4718 ab->b_flags &= ~ARC_L2_WRITING; 4719 4720 mutex_exit(hash_lock); 4721 } 4722 4723 atomic_inc_64(&l2arc_writes_done); 4724 list_remove(buflist, head); 4725 kmem_cache_free(hdr_cache, head); 4726 mutex_exit(&l2arc_buflist_mtx); 4727 4728 vdev_space_update(dev->l2ad_vdev, -bytes_dropped, 0, 0); 4729 4730 l2arc_do_free_on_write(); 4731 4732 kmem_free(cb, sizeof (l2arc_write_callback_t)); 4733} 4734 4735/* 4736 * A read to a cache device completed. Validate buffer contents before 4737 * handing over to the regular ARC routines. 4738 */ 4739static void 4740l2arc_read_done(zio_t *zio) 4741{ 4742 l2arc_read_callback_t *cb; 4743 arc_buf_hdr_t *hdr; 4744 arc_buf_t *buf; 4745 kmutex_t *hash_lock; 4746 int equal; 4747 4748 ASSERT(zio->io_vd != NULL); 4749 ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE); 4750 4751 spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd); 4752 4753 cb = zio->io_private; 4754 ASSERT(cb != NULL); 4755 buf = cb->l2rcb_buf; 4756 ASSERT(buf != NULL); 4757 4758 hash_lock = HDR_LOCK(buf->b_hdr); 4759 mutex_enter(hash_lock); 4760 hdr = buf->b_hdr; 4761 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 4762 4763 /* 4764 * If the buffer was compressed, decompress it first. 4765 */ 4766 if (cb->l2rcb_compress != ZIO_COMPRESS_OFF) 4767 l2arc_decompress_zio(zio, hdr, cb->l2rcb_compress); 4768 ASSERT(zio->io_data != NULL); 4769 4770 /* 4771 * Check this survived the L2ARC journey. 4772 */ 4773 equal = arc_cksum_equal(buf); 4774 if (equal && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) { 4775 mutex_exit(hash_lock); 4776 zio->io_private = buf; 4777 zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */ 4778 zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */ 4779 arc_read_done(zio); 4780 } else { 4781 mutex_exit(hash_lock); 4782 /* 4783 * Buffer didn't survive caching. Increment stats and 4784 * reissue to the original storage device. 4785 */ 4786 if (zio->io_error != 0) { 4787 ARCSTAT_BUMP(arcstat_l2_io_error); 4788 } else { 4789 zio->io_error = SET_ERROR(EIO); 4790 } 4791 if (!equal) 4792 ARCSTAT_BUMP(arcstat_l2_cksum_bad); 4793 4794 /* 4795 * If there's no waiter, issue an async i/o to the primary 4796 * storage now. If there *is* a waiter, the caller must 4797 * issue the i/o in a context where it's OK to block. 4798 */ 4799 if (zio->io_waiter == NULL) { 4800 zio_t *pio = zio_unique_parent(zio); 4801 4802 ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL); 4803 4804 zio_nowait(zio_read(pio, cb->l2rcb_spa, &cb->l2rcb_bp, 4805 buf->b_data, zio->io_size, arc_read_done, buf, 4806 zio->io_priority, cb->l2rcb_flags, &cb->l2rcb_zb)); 4807 } 4808 } 4809 4810 kmem_free(cb, sizeof (l2arc_read_callback_t)); 4811} 4812 4813/* 4814 * This is the list priority from which the L2ARC will search for pages to 4815 * cache. This is used within loops (0..3) to cycle through lists in the 4816 * desired order. This order can have a significant effect on cache 4817 * performance. 4818 * 4819 * Currently the metadata lists are hit first, MFU then MRU, followed by 4820 * the data lists. This function returns a locked list, and also returns 4821 * the lock pointer. 4822 */ 4823static list_t * 4824l2arc_list_locked(int list_num, kmutex_t **lock) 4825{ 4826 list_t *list = NULL; 4827 int idx; 4828 4829 ASSERT(list_num >= 0 && list_num < 2 * ARC_BUFC_NUMLISTS); 4830 4831 if (list_num < ARC_BUFC_NUMMETADATALISTS) { 4832 idx = list_num; 4833 list = &arc_mfu->arcs_lists[idx]; 4834 *lock = ARCS_LOCK(arc_mfu, idx); 4835 } else if (list_num < ARC_BUFC_NUMMETADATALISTS * 2) { 4836 idx = list_num - ARC_BUFC_NUMMETADATALISTS; 4837 list = &arc_mru->arcs_lists[idx]; 4838 *lock = ARCS_LOCK(arc_mru, idx); 4839 } else if (list_num < (ARC_BUFC_NUMMETADATALISTS * 2 + 4840 ARC_BUFC_NUMDATALISTS)) { 4841 idx = list_num - ARC_BUFC_NUMMETADATALISTS; 4842 list = &arc_mfu->arcs_lists[idx]; 4843 *lock = ARCS_LOCK(arc_mfu, idx); 4844 } else { 4845 idx = list_num - ARC_BUFC_NUMLISTS; 4846 list = &arc_mru->arcs_lists[idx]; 4847 *lock = ARCS_LOCK(arc_mru, idx); 4848 } 4849 4850 ASSERT(!(MUTEX_HELD(*lock))); 4851 mutex_enter(*lock); 4852 return (list); 4853} 4854 4855/* 4856 * Evict buffers from the device write hand to the distance specified in 4857 * bytes. This distance may span populated buffers, it may span nothing. 4858 * This is clearing a region on the L2ARC device ready for writing. 4859 * If the 'all' boolean is set, every buffer is evicted. 4860 */ 4861static void 4862l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all) 4863{ 4864 list_t *buflist; 4865 l2arc_buf_hdr_t *abl2; 4866 arc_buf_hdr_t *ab, *ab_prev; 4867 kmutex_t *hash_lock; 4868 uint64_t taddr; 4869 int64_t bytes_evicted = 0; 4870 4871 buflist = dev->l2ad_buflist; 4872 4873 if (buflist == NULL) 4874 return; 4875 4876 if (!all && dev->l2ad_first) { 4877 /* 4878 * This is the first sweep through the device. There is 4879 * nothing to evict. 4880 */ 4881 return; 4882 } 4883 4884 if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) { 4885 /* 4886 * When nearing the end of the device, evict to the end 4887 * before the device write hand jumps to the start. 4888 */ 4889 taddr = dev->l2ad_end; 4890 } else { 4891 taddr = dev->l2ad_hand + distance; 4892 } 4893 DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist, 4894 uint64_t, taddr, boolean_t, all); 4895 4896top: 4897 mutex_enter(&l2arc_buflist_mtx); 4898 for (ab = list_tail(buflist); ab; ab = ab_prev) { 4899 ab_prev = list_prev(buflist, ab); 4900 4901 hash_lock = HDR_LOCK(ab); 4902 if (!mutex_tryenter(hash_lock)) { 4903 /* 4904 * Missed the hash lock. Retry. 4905 */ 4906 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry); 4907 mutex_exit(&l2arc_buflist_mtx); 4908 mutex_enter(hash_lock); 4909 mutex_exit(hash_lock); 4910 goto top; 4911 } 4912 4913 if (HDR_L2_WRITE_HEAD(ab)) { 4914 /* 4915 * We hit a write head node. Leave it for 4916 * l2arc_write_done(). 4917 */ 4918 list_remove(buflist, ab); 4919 mutex_exit(hash_lock); 4920 continue; 4921 } 4922 4923 if (!all && ab->b_l2hdr != NULL && 4924 (ab->b_l2hdr->b_daddr > taddr || 4925 ab->b_l2hdr->b_daddr < dev->l2ad_hand)) { 4926 /* 4927 * We've evicted to the target address, 4928 * or the end of the device. 4929 */ 4930 mutex_exit(hash_lock); 4931 break; 4932 } 4933 4934 if (HDR_FREE_IN_PROGRESS(ab)) { 4935 /* 4936 * Already on the path to destruction. 4937 */ 4938 mutex_exit(hash_lock); 4939 continue; 4940 } 4941 4942 if (ab->b_state == arc_l2c_only) { 4943 ASSERT(!HDR_L2_READING(ab)); 4944 /* 4945 * This doesn't exist in the ARC. Destroy. 4946 * arc_hdr_destroy() will call list_remove() 4947 * and decrement arcstat_l2_size. 4948 */ 4949 arc_change_state(arc_anon, ab, hash_lock); 4950 arc_hdr_destroy(ab); 4951 } else { 4952 /* 4953 * Invalidate issued or about to be issued 4954 * reads, since we may be about to write 4955 * over this location. 4956 */ 4957 if (HDR_L2_READING(ab)) { 4958 ARCSTAT_BUMP(arcstat_l2_evict_reading); 4959 ab->b_flags |= ARC_L2_EVICTED; 4960 } 4961 4962 /* 4963 * Tell ARC this no longer exists in L2ARC. 4964 */ 4965 if (ab->b_l2hdr != NULL) { 4966 abl2 = ab->b_l2hdr; 4967 ARCSTAT_INCR(arcstat_l2_asize, -abl2->b_asize); 4968 bytes_evicted += abl2->b_asize; 4969 ab->b_l2hdr = NULL; 4970 kmem_free(abl2, sizeof (l2arc_buf_hdr_t)); 4971 ARCSTAT_INCR(arcstat_l2_size, -ab->b_size); 4972 } 4973 list_remove(buflist, ab); 4974 4975 /* 4976 * This may have been leftover after a 4977 * failed write. 4978 */ 4979 ab->b_flags &= ~ARC_L2_WRITING; 4980 } 4981 mutex_exit(hash_lock); 4982 } 4983 mutex_exit(&l2arc_buflist_mtx); 4984 4985 vdev_space_update(dev->l2ad_vdev, -bytes_evicted, 0, 0); 4986 dev->l2ad_evict = taddr; 4987} 4988 4989/* 4990 * Find and write ARC buffers to the L2ARC device. 4991 * 4992 * An ARC_L2_WRITING flag is set so that the L2ARC buffers are not valid 4993 * for reading until they have completed writing. 4994 * The headroom_boost is an in-out parameter used to maintain headroom boost 4995 * state between calls to this function. 4996 * 4997 * Returns the number of bytes actually written (which may be smaller than 4998 * the delta by which the device hand has changed due to alignment). 4999 */ 5000static uint64_t 5001l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz, 5002 boolean_t *headroom_boost) 5003{ 5004 arc_buf_hdr_t *ab, *ab_prev, *head; 5005 list_t *list; 5006 uint64_t write_asize, write_psize, write_sz, headroom, 5007 buf_compress_minsz; 5008 void *buf_data; 5009 kmutex_t *list_lock; 5010 boolean_t full; 5011 l2arc_write_callback_t *cb; 5012 zio_t *pio, *wzio; 5013 uint64_t guid = spa_load_guid(spa); 5014 const boolean_t do_headroom_boost = *headroom_boost; 5015 int try; 5016 5017 ASSERT(dev->l2ad_vdev != NULL); 5018 5019 /* Lower the flag now, we might want to raise it again later. */ 5020 *headroom_boost = B_FALSE; 5021 5022 pio = NULL; 5023 write_sz = write_asize = write_psize = 0; 5024 full = B_FALSE; 5025 head = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE); 5026 head->b_flags |= ARC_L2_WRITE_HEAD; 5027 5028 ARCSTAT_BUMP(arcstat_l2_write_buffer_iter); 5029 /* 5030 * We will want to try to compress buffers that are at least 2x the 5031 * device sector size. 5032 */ 5033 buf_compress_minsz = 2 << dev->l2ad_vdev->vdev_ashift; 5034 5035 /* 5036 * Copy buffers for L2ARC writing. 5037 */ 5038 mutex_enter(&l2arc_buflist_mtx); 5039 for (try = 0; try < 2 * ARC_BUFC_NUMLISTS; try++) { 5040 uint64_t passed_sz = 0; 5041 5042 list = l2arc_list_locked(try, &list_lock); 5043 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_iter); 5044 5045 /* 5046 * L2ARC fast warmup. 5047 * 5048 * Until the ARC is warm and starts to evict, read from the 5049 * head of the ARC lists rather than the tail. 5050 */ 5051 if (arc_warm == B_FALSE) 5052 ab = list_head(list); 5053 else 5054 ab = list_tail(list); 5055 if (ab == NULL) 5056 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_null_iter); 5057 5058 headroom = target_sz * l2arc_headroom * 2 / ARC_BUFC_NUMLISTS; 5059 if (do_headroom_boost) 5060 headroom = (headroom * l2arc_headroom_boost) / 100; 5061 5062 for (; ab; ab = ab_prev) { 5063 l2arc_buf_hdr_t *l2hdr; 5064 kmutex_t *hash_lock; 5065 uint64_t buf_sz; 5066 5067 if (arc_warm == B_FALSE) 5068 ab_prev = list_next(list, ab); 5069 else 5070 ab_prev = list_prev(list, ab); 5071 ARCSTAT_INCR(arcstat_l2_write_buffer_bytes_scanned, ab->b_size); 5072 5073 hash_lock = HDR_LOCK(ab); 5074 if (!mutex_tryenter(hash_lock)) { 5075 ARCSTAT_BUMP(arcstat_l2_write_trylock_fail); 5076 /* 5077 * Skip this buffer rather than waiting. 5078 */ 5079 continue; 5080 } 5081 5082 passed_sz += ab->b_size; 5083 if (passed_sz > headroom) { 5084 /* 5085 * Searched too far. 5086 */ 5087 mutex_exit(hash_lock); 5088 ARCSTAT_BUMP(arcstat_l2_write_passed_headroom); 5089 break; 5090 } 5091 5092 if (!l2arc_write_eligible(guid, ab)) { 5093 mutex_exit(hash_lock); 5094 continue; 5095 } 5096 5097 if ((write_sz + ab->b_size) > target_sz) { 5098 full = B_TRUE; 5099 mutex_exit(hash_lock); 5100 ARCSTAT_BUMP(arcstat_l2_write_full); 5101 break; 5102 } 5103 5104 if (pio == NULL) { 5105 /* 5106 * Insert a dummy header on the buflist so 5107 * l2arc_write_done() can find where the 5108 * write buffers begin without searching. 5109 */ 5110 list_insert_head(dev->l2ad_buflist, head); 5111 5112 cb = kmem_alloc( 5113 sizeof (l2arc_write_callback_t), KM_SLEEP); 5114 cb->l2wcb_dev = dev; 5115 cb->l2wcb_head = head; 5116 pio = zio_root(spa, l2arc_write_done, cb, 5117 ZIO_FLAG_CANFAIL); 5118 ARCSTAT_BUMP(arcstat_l2_write_pios); 5119 } 5120 5121 /* 5122 * Create and add a new L2ARC header. 5123 */ 5124 l2hdr = kmem_zalloc(sizeof (l2arc_buf_hdr_t), KM_SLEEP); 5125 l2hdr->b_dev = dev; 5126 ab->b_flags |= ARC_L2_WRITING; 5127 5128 /* 5129 * Temporarily stash the data buffer in b_tmp_cdata. 5130 * The subsequent write step will pick it up from 5131 * there. This is because can't access ab->b_buf 5132 * without holding the hash_lock, which we in turn 5133 * can't access without holding the ARC list locks 5134 * (which we want to avoid during compression/writing). 5135 */ 5136 l2hdr->b_compress = ZIO_COMPRESS_OFF; 5137 l2hdr->b_asize = ab->b_size; 5138 l2hdr->b_tmp_cdata = ab->b_buf->b_data; 5139 5140 buf_sz = ab->b_size; 5141 ab->b_l2hdr = l2hdr; 5142 5143 list_insert_head(dev->l2ad_buflist, ab); 5144 5145 /* 5146 * Compute and store the buffer cksum before 5147 * writing. On debug the cksum is verified first. 5148 */ 5149 arc_cksum_verify(ab->b_buf); 5150 arc_cksum_compute(ab->b_buf, B_TRUE); 5151 5152 mutex_exit(hash_lock); 5153 5154 write_sz += buf_sz; 5155 } 5156 5157 mutex_exit(list_lock); 5158 5159 if (full == B_TRUE) 5160 break; 5161 } 5162 5163 /* No buffers selected for writing? */ 5164 if (pio == NULL) { 5165 ASSERT0(write_sz); 5166 mutex_exit(&l2arc_buflist_mtx); 5167 kmem_cache_free(hdr_cache, head); 5168 return (0); 5169 } 5170 5171 /* 5172 * Now start writing the buffers. We're starting at the write head 5173 * and work backwards, retracing the course of the buffer selector 5174 * loop above. 5175 */ 5176 for (ab = list_prev(dev->l2ad_buflist, head); ab; 5177 ab = list_prev(dev->l2ad_buflist, ab)) { 5178 l2arc_buf_hdr_t *l2hdr; 5179 uint64_t buf_sz; 5180 5181 /* 5182 * We shouldn't need to lock the buffer here, since we flagged 5183 * it as ARC_L2_WRITING in the previous step, but we must take 5184 * care to only access its L2 cache parameters. In particular, 5185 * ab->b_buf may be invalid by now due to ARC eviction. 5186 */ 5187 l2hdr = ab->b_l2hdr; 5188 l2hdr->b_daddr = dev->l2ad_hand; 5189 5190 if ((ab->b_flags & ARC_L2COMPRESS) && 5191 l2hdr->b_asize >= buf_compress_minsz) { 5192 if (l2arc_compress_buf(l2hdr)) { 5193 /* 5194 * If compression succeeded, enable headroom 5195 * boost on the next scan cycle. 5196 */ 5197 *headroom_boost = B_TRUE; 5198 } 5199 } 5200 5201 /* 5202 * Pick up the buffer data we had previously stashed away 5203 * (and now potentially also compressed). 5204 */ 5205 buf_data = l2hdr->b_tmp_cdata; 5206 buf_sz = l2hdr->b_asize; 5207 5208 /* Compression may have squashed the buffer to zero length. */ 5209 if (buf_sz != 0) { 5210 uint64_t buf_p_sz; 5211 5212 wzio = zio_write_phys(pio, dev->l2ad_vdev, 5213 dev->l2ad_hand, buf_sz, buf_data, ZIO_CHECKSUM_OFF, 5214 NULL, NULL, ZIO_PRIORITY_ASYNC_WRITE, 5215 ZIO_FLAG_CANFAIL, B_FALSE); 5216 5217 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev, 5218 zio_t *, wzio); 5219 (void) zio_nowait(wzio); 5220 5221 write_asize += buf_sz; 5222 /* 5223 * Keep the clock hand suitably device-aligned. 5224 */ 5225 buf_p_sz = vdev_psize_to_asize(dev->l2ad_vdev, buf_sz); 5226 write_psize += buf_p_sz; 5227 dev->l2ad_hand += buf_p_sz; 5228 } 5229 } 5230 5231 mutex_exit(&l2arc_buflist_mtx); 5232 5233 ASSERT3U(write_asize, <=, target_sz); 5234 ARCSTAT_BUMP(arcstat_l2_writes_sent); 5235 ARCSTAT_INCR(arcstat_l2_write_bytes, write_asize); 5236 ARCSTAT_INCR(arcstat_l2_size, write_sz); 5237 ARCSTAT_INCR(arcstat_l2_asize, write_asize); 5238 vdev_space_update(dev->l2ad_vdev, write_psize, 0, 0); 5239 5240 /* 5241 * Bump device hand to the device start if it is approaching the end. 5242 * l2arc_evict() will already have evicted ahead for this case. 5243 */ 5244 if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) { 5245 dev->l2ad_hand = dev->l2ad_start; 5246 dev->l2ad_evict = dev->l2ad_start; 5247 dev->l2ad_first = B_FALSE; 5248 } 5249 5250 dev->l2ad_writing = B_TRUE; 5251 (void) zio_wait(pio); 5252 dev->l2ad_writing = B_FALSE; 5253 5254 return (write_asize); 5255} 5256 5257/* 5258 * Compresses an L2ARC buffer. 5259 * The data to be compressed must be prefilled in l2hdr->b_tmp_cdata and its 5260 * size in l2hdr->b_asize. This routine tries to compress the data and 5261 * depending on the compression result there are three possible outcomes: 5262 * *) The buffer was incompressible. The original l2hdr contents were left 5263 * untouched and are ready for writing to an L2 device. 5264 * *) The buffer was all-zeros, so there is no need to write it to an L2 5265 * device. To indicate this situation b_tmp_cdata is NULL'ed, b_asize is 5266 * set to zero and b_compress is set to ZIO_COMPRESS_EMPTY. 5267 * *) Compression succeeded and b_tmp_cdata was replaced with a temporary 5268 * data buffer which holds the compressed data to be written, and b_asize 5269 * tells us how much data there is. b_compress is set to the appropriate 5270 * compression algorithm. Once writing is done, invoke 5271 * l2arc_release_cdata_buf on this l2hdr to free this temporary buffer. 5272 * 5273 * Returns B_TRUE if compression succeeded, or B_FALSE if it didn't (the 5274 * buffer was incompressible). 5275 */ 5276static boolean_t 5277l2arc_compress_buf(l2arc_buf_hdr_t *l2hdr) 5278{ 5279 void *cdata; 5280 size_t csize, len, rounded; 5281 5282 ASSERT(l2hdr->b_compress == ZIO_COMPRESS_OFF); 5283 ASSERT(l2hdr->b_tmp_cdata != NULL); 5284 5285 len = l2hdr->b_asize; 5286 cdata = zio_data_buf_alloc(len); 5287 csize = zio_compress_data(ZIO_COMPRESS_LZ4, l2hdr->b_tmp_cdata, 5288 cdata, l2hdr->b_asize); 5289 5290 rounded = P2ROUNDUP(csize, (size_t)SPA_MINBLOCKSIZE); 5291 if (rounded > csize) { 5292 bzero((char *)cdata + csize, rounded - csize); 5293 csize = rounded; 5294 } 5295 5296 if (csize == 0) { 5297 /* zero block, indicate that there's nothing to write */ 5298 zio_data_buf_free(cdata, len); 5299 l2hdr->b_compress = ZIO_COMPRESS_EMPTY; 5300 l2hdr->b_asize = 0; 5301 l2hdr->b_tmp_cdata = NULL; 5302 ARCSTAT_BUMP(arcstat_l2_compress_zeros); 5303 return (B_TRUE); 5304 } else if (csize > 0 && csize < len) { 5305 /* 5306 * Compression succeeded, we'll keep the cdata around for 5307 * writing and release it afterwards. 5308 */ 5309 l2hdr->b_compress = ZIO_COMPRESS_LZ4; 5310 l2hdr->b_asize = csize; 5311 l2hdr->b_tmp_cdata = cdata; 5312 ARCSTAT_BUMP(arcstat_l2_compress_successes); 5313 return (B_TRUE); 5314 } else { 5315 /* 5316 * Compression failed, release the compressed buffer. 5317 * l2hdr will be left unmodified. 5318 */ 5319 zio_data_buf_free(cdata, len); 5320 ARCSTAT_BUMP(arcstat_l2_compress_failures); 5321 return (B_FALSE); 5322 } 5323} 5324 5325/* 5326 * Decompresses a zio read back from an l2arc device. On success, the 5327 * underlying zio's io_data buffer is overwritten by the uncompressed 5328 * version. On decompression error (corrupt compressed stream), the 5329 * zio->io_error value is set to signal an I/O error. 5330 * 5331 * Please note that the compressed data stream is not checksummed, so 5332 * if the underlying device is experiencing data corruption, we may feed 5333 * corrupt data to the decompressor, so the decompressor needs to be 5334 * able to handle this situation (LZ4 does). 5335 */ 5336static void 5337l2arc_decompress_zio(zio_t *zio, arc_buf_hdr_t *hdr, enum zio_compress c) 5338{ 5339 ASSERT(L2ARC_IS_VALID_COMPRESS(c)); 5340 5341 if (zio->io_error != 0) { 5342 /* 5343 * An io error has occured, just restore the original io 5344 * size in preparation for a main pool read. 5345 */ 5346 zio->io_orig_size = zio->io_size = hdr->b_size; 5347 return; 5348 } 5349 5350 if (c == ZIO_COMPRESS_EMPTY) { 5351 /* 5352 * An empty buffer results in a null zio, which means we 5353 * need to fill its io_data after we're done restoring the 5354 * buffer's contents. 5355 */ 5356 ASSERT(hdr->b_buf != NULL); 5357 bzero(hdr->b_buf->b_data, hdr->b_size); 5358 zio->io_data = zio->io_orig_data = hdr->b_buf->b_data; 5359 } else { 5360 ASSERT(zio->io_data != NULL); 5361 /* 5362 * We copy the compressed data from the start of the arc buffer 5363 * (the zio_read will have pulled in only what we need, the 5364 * rest is garbage which we will overwrite at decompression) 5365 * and then decompress back to the ARC data buffer. This way we 5366 * can minimize copying by simply decompressing back over the 5367 * original compressed data (rather than decompressing to an 5368 * aux buffer and then copying back the uncompressed buffer, 5369 * which is likely to be much larger). 5370 */ 5371 uint64_t csize; 5372 void *cdata; 5373 5374 csize = zio->io_size; 5375 cdata = zio_data_buf_alloc(csize); 5376 bcopy(zio->io_data, cdata, csize); 5377 if (zio_decompress_data(c, cdata, zio->io_data, csize, 5378 hdr->b_size) != 0) 5379 zio->io_error = EIO; 5380 zio_data_buf_free(cdata, csize); 5381 } 5382 5383 /* Restore the expected uncompressed IO size. */ 5384 zio->io_orig_size = zio->io_size = hdr->b_size; 5385} 5386 5387/* 5388 * Releases the temporary b_tmp_cdata buffer in an l2arc header structure. 5389 * This buffer serves as a temporary holder of compressed data while 5390 * the buffer entry is being written to an l2arc device. Once that is 5391 * done, we can dispose of it. 5392 */ 5393static void 5394l2arc_release_cdata_buf(arc_buf_hdr_t *ab) 5395{ 5396 l2arc_buf_hdr_t *l2hdr = ab->b_l2hdr; 5397 5398 if (l2hdr->b_compress == ZIO_COMPRESS_LZ4) { 5399 /* 5400 * If the data was compressed, then we've allocated a 5401 * temporary buffer for it, so now we need to release it. 5402 */ 5403 ASSERT(l2hdr->b_tmp_cdata != NULL); 5404 zio_data_buf_free(l2hdr->b_tmp_cdata, ab->b_size); 5405 } 5406 l2hdr->b_tmp_cdata = NULL; 5407} 5408 5409/* 5410 * This thread feeds the L2ARC at regular intervals. This is the beating 5411 * heart of the L2ARC. 5412 */ 5413static void 5414l2arc_feed_thread(void *dummy __unused) 5415{ 5416 callb_cpr_t cpr; 5417 l2arc_dev_t *dev; 5418 spa_t *spa; 5419 uint64_t size, wrote; 5420 clock_t begin, next = ddi_get_lbolt(); 5421 boolean_t headroom_boost = B_FALSE; 5422 5423 CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG); 5424 5425 mutex_enter(&l2arc_feed_thr_lock); 5426 5427 while (l2arc_thread_exit == 0) { 5428 CALLB_CPR_SAFE_BEGIN(&cpr); 5429 (void) cv_timedwait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock, 5430 next - ddi_get_lbolt()); 5431 CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock); 5432 next = ddi_get_lbolt() + hz; 5433 5434 /* 5435 * Quick check for L2ARC devices. 5436 */ 5437 mutex_enter(&l2arc_dev_mtx); 5438 if (l2arc_ndev == 0) { 5439 mutex_exit(&l2arc_dev_mtx); 5440 continue; 5441 } 5442 mutex_exit(&l2arc_dev_mtx); 5443 begin = ddi_get_lbolt(); 5444 5445 /* 5446 * This selects the next l2arc device to write to, and in 5447 * doing so the next spa to feed from: dev->l2ad_spa. This 5448 * will return NULL if there are now no l2arc devices or if 5449 * they are all faulted. 5450 * 5451 * If a device is returned, its spa's config lock is also 5452 * held to prevent device removal. l2arc_dev_get_next() 5453 * will grab and release l2arc_dev_mtx. 5454 */ 5455 if ((dev = l2arc_dev_get_next()) == NULL) 5456 continue; 5457 5458 spa = dev->l2ad_spa; 5459 ASSERT(spa != NULL); 5460 5461 /* 5462 * If the pool is read-only then force the feed thread to 5463 * sleep a little longer. 5464 */ 5465 if (!spa_writeable(spa)) { 5466 next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz; 5467 spa_config_exit(spa, SCL_L2ARC, dev); 5468 continue; 5469 } 5470 5471 /* 5472 * Avoid contributing to memory pressure. 5473 */ 5474 if (arc_reclaim_needed()) { 5475 ARCSTAT_BUMP(arcstat_l2_abort_lowmem); 5476 spa_config_exit(spa, SCL_L2ARC, dev); 5477 continue; 5478 } 5479 5480 ARCSTAT_BUMP(arcstat_l2_feeds); 5481 5482 size = l2arc_write_size(); 5483 5484 /* 5485 * Evict L2ARC buffers that will be overwritten. 5486 */ 5487 l2arc_evict(dev, size, B_FALSE); 5488 5489 /* 5490 * Write ARC buffers. 5491 */ 5492 wrote = l2arc_write_buffers(spa, dev, size, &headroom_boost); 5493 5494 /* 5495 * Calculate interval between writes. 5496 */ 5497 next = l2arc_write_interval(begin, size, wrote); 5498 spa_config_exit(spa, SCL_L2ARC, dev); 5499 } 5500 5501 l2arc_thread_exit = 0; 5502 cv_broadcast(&l2arc_feed_thr_cv); 5503 CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */ 5504 thread_exit(); 5505} 5506 5507boolean_t 5508l2arc_vdev_present(vdev_t *vd) 5509{ 5510 l2arc_dev_t *dev; 5511 5512 mutex_enter(&l2arc_dev_mtx); 5513 for (dev = list_head(l2arc_dev_list); dev != NULL; 5514 dev = list_next(l2arc_dev_list, dev)) { 5515 if (dev->l2ad_vdev == vd) 5516 break; 5517 } 5518 mutex_exit(&l2arc_dev_mtx); 5519 5520 return (dev != NULL); 5521} 5522 5523/* 5524 * Add a vdev for use by the L2ARC. By this point the spa has already 5525 * validated the vdev and opened it. 5526 */ 5527void 5528l2arc_add_vdev(spa_t *spa, vdev_t *vd) 5529{ 5530 l2arc_dev_t *adddev; 5531 5532 ASSERT(!l2arc_vdev_present(vd)); 5533 5534 vdev_ashift_optimize(vd); 5535 5536 /* 5537 * Create a new l2arc device entry. 5538 */ 5539 adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP); 5540 adddev->l2ad_spa = spa; 5541 adddev->l2ad_vdev = vd; 5542 adddev->l2ad_start = VDEV_LABEL_START_SIZE; 5543 adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd); 5544 adddev->l2ad_hand = adddev->l2ad_start; 5545 adddev->l2ad_evict = adddev->l2ad_start; 5546 adddev->l2ad_first = B_TRUE; 5547 adddev->l2ad_writing = B_FALSE; 5548 5549 /* 5550 * This is a list of all ARC buffers that are still valid on the 5551 * device. 5552 */ 5553 adddev->l2ad_buflist = kmem_zalloc(sizeof (list_t), KM_SLEEP); 5554 list_create(adddev->l2ad_buflist, sizeof (arc_buf_hdr_t), 5555 offsetof(arc_buf_hdr_t, b_l2node)); 5556 5557 vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand); 5558 5559 /* 5560 * Add device to global list 5561 */ 5562 mutex_enter(&l2arc_dev_mtx); 5563 list_insert_head(l2arc_dev_list, adddev); 5564 atomic_inc_64(&l2arc_ndev); 5565 mutex_exit(&l2arc_dev_mtx); 5566} 5567 5568/* 5569 * Remove a vdev from the L2ARC. 5570 */ 5571void 5572l2arc_remove_vdev(vdev_t *vd) 5573{ 5574 l2arc_dev_t *dev, *nextdev, *remdev = NULL; 5575 5576 /* 5577 * Find the device by vdev 5578 */ 5579 mutex_enter(&l2arc_dev_mtx); 5580 for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) { 5581 nextdev = list_next(l2arc_dev_list, dev); 5582 if (vd == dev->l2ad_vdev) { 5583 remdev = dev; 5584 break; 5585 } 5586 } 5587 ASSERT(remdev != NULL); 5588 5589 /* 5590 * Remove device from global list 5591 */ 5592 list_remove(l2arc_dev_list, remdev); 5593 l2arc_dev_last = NULL; /* may have been invalidated */ 5594 atomic_dec_64(&l2arc_ndev); 5595 mutex_exit(&l2arc_dev_mtx); 5596 5597 /* 5598 * Clear all buflists and ARC references. L2ARC device flush. 5599 */ 5600 l2arc_evict(remdev, 0, B_TRUE); 5601 list_destroy(remdev->l2ad_buflist); 5602 kmem_free(remdev->l2ad_buflist, sizeof (list_t)); 5603 kmem_free(remdev, sizeof (l2arc_dev_t)); 5604} 5605 5606void 5607l2arc_init(void) 5608{ 5609 l2arc_thread_exit = 0; 5610 l2arc_ndev = 0; 5611 l2arc_writes_sent = 0; 5612 l2arc_writes_done = 0; 5613 5614 mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL); 5615 cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL); 5616 mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL); 5617 mutex_init(&l2arc_buflist_mtx, NULL, MUTEX_DEFAULT, NULL); 5618 mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL); 5619 5620 l2arc_dev_list = &L2ARC_dev_list; 5621 l2arc_free_on_write = &L2ARC_free_on_write; 5622 list_create(l2arc_dev_list, sizeof (l2arc_dev_t), 5623 offsetof(l2arc_dev_t, l2ad_node)); 5624 list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t), 5625 offsetof(l2arc_data_free_t, l2df_list_node)); 5626} 5627 5628void 5629l2arc_fini(void) 5630{ 5631 /* 5632 * This is called from dmu_fini(), which is called from spa_fini(); 5633 * Because of this, we can assume that all l2arc devices have 5634 * already been removed when the pools themselves were removed. 5635 */ 5636 5637 l2arc_do_free_on_write(); 5638 5639 mutex_destroy(&l2arc_feed_thr_lock); 5640 cv_destroy(&l2arc_feed_thr_cv); 5641 mutex_destroy(&l2arc_dev_mtx); 5642 mutex_destroy(&l2arc_buflist_mtx); 5643 mutex_destroy(&l2arc_free_on_write_mtx); 5644 5645 list_destroy(l2arc_dev_list); 5646 list_destroy(l2arc_free_on_write); 5647} 5648 5649void 5650l2arc_start(void) 5651{ 5652 if (!(spa_mode_global & FWRITE)) 5653 return; 5654 5655 (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0, 5656 TS_RUN, minclsyspri); 5657} 5658 5659void 5660l2arc_stop(void) 5661{ 5662 if (!(spa_mode_global & FWRITE)) 5663 return; 5664 5665 mutex_enter(&l2arc_feed_thr_lock); 5666 cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */ 5667 l2arc_thread_exit = 1; 5668 while (l2arc_thread_exit != 0) 5669 cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock); 5670 mutex_exit(&l2arc_feed_thr_lock); 5671} 5672