arc.c revision 277826
198937Sdes/* 298937Sdes * CDDL HEADER START 398937Sdes * 498937Sdes * The contents of this file are subject to the terms of the 598937Sdes * Common Development and Distribution License (the "License"). 698937Sdes * You may not use this file except in compliance with the License. 798937Sdes * 898937Sdes * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE 998937Sdes * or http://www.opensolaris.org/os/licensing. 1098937Sdes * See the License for the specific language governing permissions 1198937Sdes * and limitations under the License. 1298937Sdes * 1398937Sdes * When distributing Covered Code, include this CDDL HEADER in each 1498937Sdes * file and include the License file at usr/src/OPENSOLARIS.LICENSE. 1598937Sdes * If applicable, add the following below this CDDL HEADER, with the 1698937Sdes * fields enclosed by brackets "[]" replaced with your own identifying 1798937Sdes * information: Portions Copyright [yyyy] [name of copyright owner] 1898937Sdes * 1998937Sdes * CDDL HEADER END 2098937Sdes */ 2198937Sdes/* 2298937Sdes * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved. 2398937Sdes * Copyright (c) 2012, Joyent, Inc. All rights reserved. 2498937Sdes * Copyright (c) 2011, 2014 by Delphix. All rights reserved. 2598937Sdes * Copyright (c) 2014 by Saso Kiselkov. All rights reserved. 2698937Sdes * Copyright 2014 Nexenta Systems, Inc. All rights reserved. 2798937Sdes */ 2898937Sdes 29124208Sdes/* 3098937Sdes * DVA-based Adjustable Replacement Cache 3198937Sdes * 32 * While much of the theory of operation used here is 33 * based on the self-tuning, low overhead replacement cache 34 * presented by Megiddo and Modha at FAST 2003, there are some 35 * significant differences: 36 * 37 * 1. The Megiddo and Modha model assumes any page is evictable. 38 * Pages in its cache cannot be "locked" into memory. This makes 39 * the eviction algorithm simple: evict the last page in the list. 40 * This also make the performance characteristics easy to reason 41 * about. Our cache is not so simple. At any given moment, some 42 * subset of the blocks in the cache are un-evictable because we 43 * have handed out a reference to them. Blocks are only evictable 44 * when there are no external references active. This makes 45 * eviction far more problematic: we choose to evict the evictable 46 * blocks that are the "lowest" in the list. 47 * 48 * There are times when it is not possible to evict the requested 49 * space. In these circumstances we are unable to adjust the cache 50 * size. To prevent the cache growing unbounded at these times we 51 * implement a "cache throttle" that slows the flow of new data 52 * into the cache until we can make space available. 53 * 54 * 2. The Megiddo and Modha model assumes a fixed cache size. 55 * Pages are evicted when the cache is full and there is a cache 56 * miss. Our model has a variable sized cache. It grows with 57 * high use, but also tries to react to memory pressure from the 58 * operating system: decreasing its size when system memory is 59 * tight. 60 * 61 * 3. The Megiddo and Modha model assumes a fixed page size. All 62 * elements of the cache are therefore exactly the same size. So 63 * when adjusting the cache size following a cache miss, its simply 64 * a matter of choosing a single page to evict. In our model, we 65 * have variable sized cache blocks (rangeing from 512 bytes to 66 * 128K bytes). We therefore choose a set of blocks to evict to make 67 * space for a cache miss that approximates as closely as possible 68 * the space used by the new block. 69 * 70 * See also: "ARC: A Self-Tuning, Low Overhead Replacement Cache" 71 * by N. Megiddo & D. Modha, FAST 2003 72 */ 73 74/* 75 * The locking model: 76 * 77 * A new reference to a cache buffer can be obtained in two 78 * ways: 1) via a hash table lookup using the DVA as a key, 79 * or 2) via one of the ARC lists. The arc_read() interface 80 * uses method 1, while the internal arc algorithms for 81 * adjusting the cache use method 2. We therefore provide two 82 * types of locks: 1) the hash table lock array, and 2) the 83 * arc list locks. 84 * 85 * Buffers do not have their own mutexs, rather they rely on the 86 * hash table mutexs for the bulk of their protection (i.e. most 87 * fields in the arc_buf_hdr_t are protected by these mutexs). 88 * 89 * buf_hash_find() returns the appropriate mutex (held) when it 90 * locates the requested buffer in the hash table. It returns 91 * NULL for the mutex if the buffer was not in the table. 92 * 93 * buf_hash_remove() expects the appropriate hash mutex to be 94 * already held before it is invoked. 95 * 96 * Each arc state also has a mutex which is used to protect the 97 * buffer list associated with the state. When attempting to 98 * obtain a hash table lock while holding an arc list lock you 99 * must use: mutex_tryenter() to avoid deadlock. Also note that 100 * the active state mutex must be held before the ghost state mutex. 101 * 102 * Arc buffers may have an associated eviction callback function. 103 * This function will be invoked prior to removing the buffer (e.g. 104 * in arc_do_user_evicts()). Note however that the data associated 105 * with the buffer may be evicted prior to the callback. The callback 106 * must be made with *no locks held* (to prevent deadlock). Additionally, 107 * the users of callbacks must ensure that their private data is 108 * protected from simultaneous callbacks from arc_clear_callback() 109 * and arc_do_user_evicts(). 110 * 111 * Note that the majority of the performance stats are manipulated 112 * with atomic operations. 113 * 114 * The L2ARC uses the l2arc_buflist_mtx global mutex for the following: 115 * 116 * - L2ARC buflist creation 117 * - L2ARC buflist eviction 118 * - L2ARC write completion, which walks L2ARC buflists 119 * - ARC header destruction, as it removes from L2ARC buflists 120 * - ARC header release, as it removes from L2ARC buflists 121 */ 122 123#include <sys/spa.h> 124#include <sys/zio.h> 125#include <sys/zio_compress.h> 126#include <sys/zfs_context.h> 127#include <sys/arc.h> 128#include <sys/refcount.h> 129#include <sys/vdev.h> 130#include <sys/vdev_impl.h> 131#include <sys/dsl_pool.h> 132#ifdef _KERNEL 133#include <sys/dnlc.h> 134#endif 135#include <sys/callb.h> 136#include <sys/kstat.h> 137#include <sys/trim_map.h> 138#include <zfs_fletcher.h> 139#include <sys/sdt.h> 140 141#include <vm/vm_pageout.h> 142#include <machine/vmparam.h> 143 144#ifdef illumos 145#ifndef _KERNEL 146/* set with ZFS_DEBUG=watch, to enable watchpoints on frozen buffers */ 147boolean_t arc_watch = B_FALSE; 148int arc_procfd; 149#endif 150#endif /* illumos */ 151 152static kmutex_t arc_reclaim_thr_lock; 153static kcondvar_t arc_reclaim_thr_cv; /* used to signal reclaim thr */ 154static uint8_t arc_thread_exit; 155 156#define ARC_REDUCE_DNLC_PERCENT 3 157uint_t arc_reduce_dnlc_percent = ARC_REDUCE_DNLC_PERCENT; 158 159typedef enum arc_reclaim_strategy { 160 ARC_RECLAIM_AGGR, /* Aggressive reclaim strategy */ 161 ARC_RECLAIM_CONS /* Conservative reclaim strategy */ 162} arc_reclaim_strategy_t; 163 164/* 165 * The number of iterations through arc_evict_*() before we 166 * drop & reacquire the lock. 167 */ 168int arc_evict_iterations = 100; 169 170/* number of seconds before growing cache again */ 171static int arc_grow_retry = 60; 172 173/* shift of arc_c for calculating both min and max arc_p */ 174static int arc_p_min_shift = 4; 175 176/* log2(fraction of arc to reclaim) */ 177static int arc_shrink_shift = 5; 178 179/* 180 * minimum lifespan of a prefetch block in clock ticks 181 * (initialized in arc_init()) 182 */ 183static int arc_min_prefetch_lifespan; 184 185/* 186 * If this percent of memory is free, don't throttle. 187 */ 188int arc_lotsfree_percent = 10; 189 190static int arc_dead; 191extern int zfs_prefetch_disable; 192 193/* 194 * The arc has filled available memory and has now warmed up. 195 */ 196static boolean_t arc_warm; 197 198uint64_t zfs_arc_max; 199uint64_t zfs_arc_min; 200uint64_t zfs_arc_meta_limit = 0; 201uint64_t zfs_arc_meta_min = 0; 202int zfs_arc_grow_retry = 0; 203int zfs_arc_shrink_shift = 0; 204int zfs_arc_p_min_shift = 0; 205int zfs_disable_dup_eviction = 0; 206uint64_t zfs_arc_average_blocksize = 8 * 1024; /* 8KB */ 207u_int zfs_arc_free_target = 0; 208 209static int sysctl_vfs_zfs_arc_free_target(SYSCTL_HANDLER_ARGS); 210static int sysctl_vfs_zfs_arc_meta_limit(SYSCTL_HANDLER_ARGS); 211 212#ifdef _KERNEL 213static void 214arc_free_target_init(void *unused __unused) 215{ 216 217 zfs_arc_free_target = vm_pageout_wakeup_thresh; 218} 219SYSINIT(arc_free_target_init, SI_SUB_KTHREAD_PAGE, SI_ORDER_ANY, 220 arc_free_target_init, NULL); 221 222TUNABLE_QUAD("vfs.zfs.arc_meta_limit", &zfs_arc_meta_limit); 223TUNABLE_QUAD("vfs.zfs.arc_meta_min", &zfs_arc_meta_min); 224TUNABLE_INT("vfs.zfs.arc_shrink_shift", &zfs_arc_shrink_shift); 225SYSCTL_DECL(_vfs_zfs); 226SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_max, CTLFLAG_RDTUN, &zfs_arc_max, 0, 227 "Maximum ARC size"); 228SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_min, CTLFLAG_RDTUN, &zfs_arc_min, 0, 229 "Minimum ARC size"); 230SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_average_blocksize, CTLFLAG_RDTUN, 231 &zfs_arc_average_blocksize, 0, 232 "ARC average blocksize"); 233SYSCTL_INT(_vfs_zfs, OID_AUTO, arc_shrink_shift, CTLFLAG_RW, 234 &arc_shrink_shift, 0, 235 "log2(fraction of arc to reclaim)"); 236 237/* 238 * We don't have a tunable for arc_free_target due to the dependency on 239 * pagedaemon initialisation. 240 */ 241SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_free_target, 242 CTLTYPE_UINT | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, sizeof(u_int), 243 sysctl_vfs_zfs_arc_free_target, "IU", 244 "Desired number of free pages below which ARC triggers reclaim"); 245 246static int 247sysctl_vfs_zfs_arc_free_target(SYSCTL_HANDLER_ARGS) 248{ 249 u_int val; 250 int err; 251 252 val = zfs_arc_free_target; 253 err = sysctl_handle_int(oidp, &val, 0, req); 254 if (err != 0 || req->newptr == NULL) 255 return (err); 256 257 if (val < minfree) 258 return (EINVAL); 259 if (val > vm_cnt.v_page_count) 260 return (EINVAL); 261 262 zfs_arc_free_target = val; 263 264 return (0); 265} 266 267/* 268 * Must be declared here, before the definition of corresponding kstat 269 * macro which uses the same names will confuse the compiler. 270 */ 271SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_meta_limit, 272 CTLTYPE_U64 | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, sizeof(uint64_t), 273 sysctl_vfs_zfs_arc_meta_limit, "QU", 274 "ARC metadata limit"); 275#endif 276 277/* 278 * Note that buffers can be in one of 6 states: 279 * ARC_anon - anonymous (discussed below) 280 * ARC_mru - recently used, currently cached 281 * ARC_mru_ghost - recentely used, no longer in cache 282 * ARC_mfu - frequently used, currently cached 283 * ARC_mfu_ghost - frequently used, no longer in cache 284 * ARC_l2c_only - exists in L2ARC but not other states 285 * When there are no active references to the buffer, they are 286 * are linked onto a list in one of these arc states. These are 287 * the only buffers that can be evicted or deleted. Within each 288 * state there are multiple lists, one for meta-data and one for 289 * non-meta-data. Meta-data (indirect blocks, blocks of dnodes, 290 * etc.) is tracked separately so that it can be managed more 291 * explicitly: favored over data, limited explicitly. 292 * 293 * Anonymous buffers are buffers that are not associated with 294 * a DVA. These are buffers that hold dirty block copies 295 * before they are written to stable storage. By definition, 296 * they are "ref'd" and are considered part of arc_mru 297 * that cannot be freed. Generally, they will aquire a DVA 298 * as they are written and migrate onto the arc_mru list. 299 * 300 * The ARC_l2c_only state is for buffers that are in the second 301 * level ARC but no longer in any of the ARC_m* lists. The second 302 * level ARC itself may also contain buffers that are in any of 303 * the ARC_m* states - meaning that a buffer can exist in two 304 * places. The reason for the ARC_l2c_only state is to keep the 305 * buffer header in the hash table, so that reads that hit the 306 * second level ARC benefit from these fast lookups. 307 */ 308 309#define ARCS_LOCK_PAD CACHE_LINE_SIZE 310struct arcs_lock { 311 kmutex_t arcs_lock; 312#ifdef _KERNEL 313 unsigned char pad[(ARCS_LOCK_PAD - sizeof (kmutex_t))]; 314#endif 315}; 316 317/* 318 * must be power of two for mask use to work 319 * 320 */ 321#define ARC_BUFC_NUMDATALISTS 16 322#define ARC_BUFC_NUMMETADATALISTS 16 323#define ARC_BUFC_NUMLISTS (ARC_BUFC_NUMMETADATALISTS + ARC_BUFC_NUMDATALISTS) 324 325typedef struct arc_state { 326 uint64_t arcs_lsize[ARC_BUFC_NUMTYPES]; /* amount of evictable data */ 327 uint64_t arcs_size; /* total amount of data in this state */ 328 list_t arcs_lists[ARC_BUFC_NUMLISTS]; /* list of evictable buffers */ 329 struct arcs_lock arcs_locks[ARC_BUFC_NUMLISTS] __aligned(CACHE_LINE_SIZE); 330} arc_state_t; 331 332#define ARCS_LOCK(s, i) (&((s)->arcs_locks[(i)].arcs_lock)) 333 334/* The 6 states: */ 335static arc_state_t ARC_anon; 336static arc_state_t ARC_mru; 337static arc_state_t ARC_mru_ghost; 338static arc_state_t ARC_mfu; 339static arc_state_t ARC_mfu_ghost; 340static arc_state_t ARC_l2c_only; 341 342typedef struct arc_stats { 343 kstat_named_t arcstat_hits; 344 kstat_named_t arcstat_misses; 345 kstat_named_t arcstat_demand_data_hits; 346 kstat_named_t arcstat_demand_data_misses; 347 kstat_named_t arcstat_demand_metadata_hits; 348 kstat_named_t arcstat_demand_metadata_misses; 349 kstat_named_t arcstat_prefetch_data_hits; 350 kstat_named_t arcstat_prefetch_data_misses; 351 kstat_named_t arcstat_prefetch_metadata_hits; 352 kstat_named_t arcstat_prefetch_metadata_misses; 353 kstat_named_t arcstat_mru_hits; 354 kstat_named_t arcstat_mru_ghost_hits; 355 kstat_named_t arcstat_mfu_hits; 356 kstat_named_t arcstat_mfu_ghost_hits; 357 kstat_named_t arcstat_allocated; 358 kstat_named_t arcstat_deleted; 359 kstat_named_t arcstat_stolen; 360 kstat_named_t arcstat_recycle_miss; 361 /* 362 * Number of buffers that could not be evicted because the hash lock 363 * was held by another thread. The lock may not necessarily be held 364 * by something using the same buffer, since hash locks are shared 365 * by multiple buffers. 366 */ 367 kstat_named_t arcstat_mutex_miss; 368 /* 369 * Number of buffers skipped because they have I/O in progress, are 370 * indrect prefetch buffers that have not lived long enough, or are 371 * not from the spa we're trying to evict from. 372 */ 373 kstat_named_t arcstat_evict_skip; 374 kstat_named_t arcstat_evict_l2_cached; 375 kstat_named_t arcstat_evict_l2_eligible; 376 kstat_named_t arcstat_evict_l2_ineligible; 377 kstat_named_t arcstat_hash_elements; 378 kstat_named_t arcstat_hash_elements_max; 379 kstat_named_t arcstat_hash_collisions; 380 kstat_named_t arcstat_hash_chains; 381 kstat_named_t arcstat_hash_chain_max; 382 kstat_named_t arcstat_p; 383 kstat_named_t arcstat_c; 384 kstat_named_t arcstat_c_min; 385 kstat_named_t arcstat_c_max; 386 kstat_named_t arcstat_size; 387 kstat_named_t arcstat_hdr_size; 388 kstat_named_t arcstat_data_size; 389 kstat_named_t arcstat_other_size; 390 kstat_named_t arcstat_l2_hits; 391 kstat_named_t arcstat_l2_misses; 392 kstat_named_t arcstat_l2_feeds; 393 kstat_named_t arcstat_l2_rw_clash; 394 kstat_named_t arcstat_l2_read_bytes; 395 kstat_named_t arcstat_l2_write_bytes; 396 kstat_named_t arcstat_l2_writes_sent; 397 kstat_named_t arcstat_l2_writes_done; 398 kstat_named_t arcstat_l2_writes_error; 399 kstat_named_t arcstat_l2_writes_hdr_miss; 400 kstat_named_t arcstat_l2_evict_lock_retry; 401 kstat_named_t arcstat_l2_evict_reading; 402 kstat_named_t arcstat_l2_free_on_write; 403 kstat_named_t arcstat_l2_cdata_free_on_write; 404 kstat_named_t arcstat_l2_abort_lowmem; 405 kstat_named_t arcstat_l2_cksum_bad; 406 kstat_named_t arcstat_l2_io_error; 407 kstat_named_t arcstat_l2_size; 408 kstat_named_t arcstat_l2_asize; 409 kstat_named_t arcstat_l2_hdr_size; 410 kstat_named_t arcstat_l2_compress_successes; 411 kstat_named_t arcstat_l2_compress_zeros; 412 kstat_named_t arcstat_l2_compress_failures; 413 kstat_named_t arcstat_l2_write_trylock_fail; 414 kstat_named_t arcstat_l2_write_passed_headroom; 415 kstat_named_t arcstat_l2_write_spa_mismatch; 416 kstat_named_t arcstat_l2_write_in_l2; 417 kstat_named_t arcstat_l2_write_hdr_io_in_progress; 418 kstat_named_t arcstat_l2_write_not_cacheable; 419 kstat_named_t arcstat_l2_write_full; 420 kstat_named_t arcstat_l2_write_buffer_iter; 421 kstat_named_t arcstat_l2_write_pios; 422 kstat_named_t arcstat_l2_write_buffer_bytes_scanned; 423 kstat_named_t arcstat_l2_write_buffer_list_iter; 424 kstat_named_t arcstat_l2_write_buffer_list_null_iter; 425 kstat_named_t arcstat_memory_throttle_count; 426 kstat_named_t arcstat_duplicate_buffers; 427 kstat_named_t arcstat_duplicate_buffers_size; 428 kstat_named_t arcstat_duplicate_reads; 429 kstat_named_t arcstat_meta_used; 430 kstat_named_t arcstat_meta_limit; 431 kstat_named_t arcstat_meta_max; 432 kstat_named_t arcstat_meta_min; 433} arc_stats_t; 434 435static arc_stats_t arc_stats = { 436 { "hits", KSTAT_DATA_UINT64 }, 437 { "misses", KSTAT_DATA_UINT64 }, 438 { "demand_data_hits", KSTAT_DATA_UINT64 }, 439 { "demand_data_misses", KSTAT_DATA_UINT64 }, 440 { "demand_metadata_hits", KSTAT_DATA_UINT64 }, 441 { "demand_metadata_misses", KSTAT_DATA_UINT64 }, 442 { "prefetch_data_hits", KSTAT_DATA_UINT64 }, 443 { "prefetch_data_misses", KSTAT_DATA_UINT64 }, 444 { "prefetch_metadata_hits", KSTAT_DATA_UINT64 }, 445 { "prefetch_metadata_misses", KSTAT_DATA_UINT64 }, 446 { "mru_hits", KSTAT_DATA_UINT64 }, 447 { "mru_ghost_hits", KSTAT_DATA_UINT64 }, 448 { "mfu_hits", KSTAT_DATA_UINT64 }, 449 { "mfu_ghost_hits", KSTAT_DATA_UINT64 }, 450 { "allocated", KSTAT_DATA_UINT64 }, 451 { "deleted", KSTAT_DATA_UINT64 }, 452 { "stolen", KSTAT_DATA_UINT64 }, 453 { "recycle_miss", KSTAT_DATA_UINT64 }, 454 { "mutex_miss", KSTAT_DATA_UINT64 }, 455 { "evict_skip", KSTAT_DATA_UINT64 }, 456 { "evict_l2_cached", KSTAT_DATA_UINT64 }, 457 { "evict_l2_eligible", KSTAT_DATA_UINT64 }, 458 { "evict_l2_ineligible", KSTAT_DATA_UINT64 }, 459 { "hash_elements", KSTAT_DATA_UINT64 }, 460 { "hash_elements_max", KSTAT_DATA_UINT64 }, 461 { "hash_collisions", KSTAT_DATA_UINT64 }, 462 { "hash_chains", KSTAT_DATA_UINT64 }, 463 { "hash_chain_max", KSTAT_DATA_UINT64 }, 464 { "p", KSTAT_DATA_UINT64 }, 465 { "c", KSTAT_DATA_UINT64 }, 466 { "c_min", KSTAT_DATA_UINT64 }, 467 { "c_max", KSTAT_DATA_UINT64 }, 468 { "size", KSTAT_DATA_UINT64 }, 469 { "hdr_size", KSTAT_DATA_UINT64 }, 470 { "data_size", KSTAT_DATA_UINT64 }, 471 { "other_size", KSTAT_DATA_UINT64 }, 472 { "l2_hits", KSTAT_DATA_UINT64 }, 473 { "l2_misses", KSTAT_DATA_UINT64 }, 474 { "l2_feeds", KSTAT_DATA_UINT64 }, 475 { "l2_rw_clash", KSTAT_DATA_UINT64 }, 476 { "l2_read_bytes", KSTAT_DATA_UINT64 }, 477 { "l2_write_bytes", KSTAT_DATA_UINT64 }, 478 { "l2_writes_sent", KSTAT_DATA_UINT64 }, 479 { "l2_writes_done", KSTAT_DATA_UINT64 }, 480 { "l2_writes_error", KSTAT_DATA_UINT64 }, 481 { "l2_writes_hdr_miss", KSTAT_DATA_UINT64 }, 482 { "l2_evict_lock_retry", KSTAT_DATA_UINT64 }, 483 { "l2_evict_reading", KSTAT_DATA_UINT64 }, 484 { "l2_free_on_write", KSTAT_DATA_UINT64 }, 485 { "l2_cdata_free_on_write", KSTAT_DATA_UINT64 }, 486 { "l2_abort_lowmem", KSTAT_DATA_UINT64 }, 487 { "l2_cksum_bad", KSTAT_DATA_UINT64 }, 488 { "l2_io_error", KSTAT_DATA_UINT64 }, 489 { "l2_size", KSTAT_DATA_UINT64 }, 490 { "l2_asize", KSTAT_DATA_UINT64 }, 491 { "l2_hdr_size", KSTAT_DATA_UINT64 }, 492 { "l2_compress_successes", KSTAT_DATA_UINT64 }, 493 { "l2_compress_zeros", KSTAT_DATA_UINT64 }, 494 { "l2_compress_failures", KSTAT_DATA_UINT64 }, 495 { "l2_write_trylock_fail", KSTAT_DATA_UINT64 }, 496 { "l2_write_passed_headroom", KSTAT_DATA_UINT64 }, 497 { "l2_write_spa_mismatch", KSTAT_DATA_UINT64 }, 498 { "l2_write_in_l2", KSTAT_DATA_UINT64 }, 499 { "l2_write_io_in_progress", KSTAT_DATA_UINT64 }, 500 { "l2_write_not_cacheable", KSTAT_DATA_UINT64 }, 501 { "l2_write_full", KSTAT_DATA_UINT64 }, 502 { "l2_write_buffer_iter", KSTAT_DATA_UINT64 }, 503 { "l2_write_pios", KSTAT_DATA_UINT64 }, 504 { "l2_write_buffer_bytes_scanned", KSTAT_DATA_UINT64 }, 505 { "l2_write_buffer_list_iter", KSTAT_DATA_UINT64 }, 506 { "l2_write_buffer_list_null_iter", KSTAT_DATA_UINT64 }, 507 { "memory_throttle_count", KSTAT_DATA_UINT64 }, 508 { "duplicate_buffers", KSTAT_DATA_UINT64 }, 509 { "duplicate_buffers_size", KSTAT_DATA_UINT64 }, 510 { "duplicate_reads", KSTAT_DATA_UINT64 }, 511 { "arc_meta_used", KSTAT_DATA_UINT64 }, 512 { "arc_meta_limit", KSTAT_DATA_UINT64 }, 513 { "arc_meta_max", KSTAT_DATA_UINT64 }, 514 { "arc_meta_min", KSTAT_DATA_UINT64 } 515}; 516 517#define ARCSTAT(stat) (arc_stats.stat.value.ui64) 518 519#define ARCSTAT_INCR(stat, val) \ 520 atomic_add_64(&arc_stats.stat.value.ui64, (val)) 521 522#define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1) 523#define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1) 524 525#define ARCSTAT_MAX(stat, val) { \ 526 uint64_t m; \ 527 while ((val) > (m = arc_stats.stat.value.ui64) && \ 528 (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \ 529 continue; \ 530} 531 532#define ARCSTAT_MAXSTAT(stat) \ 533 ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64) 534 535/* 536 * We define a macro to allow ARC hits/misses to be easily broken down by 537 * two separate conditions, giving a total of four different subtypes for 538 * each of hits and misses (so eight statistics total). 539 */ 540#define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \ 541 if (cond1) { \ 542 if (cond2) { \ 543 ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \ 544 } else { \ 545 ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \ 546 } \ 547 } else { \ 548 if (cond2) { \ 549 ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \ 550 } else { \ 551 ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\ 552 } \ 553 } 554 555kstat_t *arc_ksp; 556static arc_state_t *arc_anon; 557static arc_state_t *arc_mru; 558static arc_state_t *arc_mru_ghost; 559static arc_state_t *arc_mfu; 560static arc_state_t *arc_mfu_ghost; 561static arc_state_t *arc_l2c_only; 562 563/* 564 * There are several ARC variables that are critical to export as kstats -- 565 * but we don't want to have to grovel around in the kstat whenever we wish to 566 * manipulate them. For these variables, we therefore define them to be in 567 * terms of the statistic variable. This assures that we are not introducing 568 * the possibility of inconsistency by having shadow copies of the variables, 569 * while still allowing the code to be readable. 570 */ 571#define arc_size ARCSTAT(arcstat_size) /* actual total arc size */ 572#define arc_p ARCSTAT(arcstat_p) /* target size of MRU */ 573#define arc_c ARCSTAT(arcstat_c) /* target size of cache */ 574#define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */ 575#define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */ 576#define arc_meta_limit ARCSTAT(arcstat_meta_limit) /* max size for metadata */ 577#define arc_meta_min ARCSTAT(arcstat_meta_min) /* min size for metadata */ 578#define arc_meta_used ARCSTAT(arcstat_meta_used) /* size of metadata */ 579#define arc_meta_max ARCSTAT(arcstat_meta_max) /* max size of metadata */ 580 581#define L2ARC_IS_VALID_COMPRESS(_c_) \ 582 ((_c_) == ZIO_COMPRESS_LZ4 || (_c_) == ZIO_COMPRESS_EMPTY) 583 584static int arc_no_grow; /* Don't try to grow cache size */ 585static uint64_t arc_tempreserve; 586static uint64_t arc_loaned_bytes; 587 588typedef struct l2arc_buf_hdr l2arc_buf_hdr_t; 589 590typedef struct arc_callback arc_callback_t; 591 592struct arc_callback { 593 void *acb_private; 594 arc_done_func_t *acb_done; 595 arc_buf_t *acb_buf; 596 zio_t *acb_zio_dummy; 597 arc_callback_t *acb_next; 598}; 599 600typedef struct arc_write_callback arc_write_callback_t; 601 602struct arc_write_callback { 603 void *awcb_private; 604 arc_done_func_t *awcb_ready; 605 arc_done_func_t *awcb_physdone; 606 arc_done_func_t *awcb_done; 607 arc_buf_t *awcb_buf; 608}; 609 610struct arc_buf_hdr { 611 /* protected by hash lock */ 612 dva_t b_dva; 613 uint64_t b_birth; 614 uint64_t b_cksum0; 615 616 kmutex_t b_freeze_lock; 617 zio_cksum_t *b_freeze_cksum; 618 void *b_thawed; 619 620 arc_buf_hdr_t *b_hash_next; 621 arc_buf_t *b_buf; 622 arc_flags_t b_flags; 623 uint32_t b_datacnt; 624 625 arc_callback_t *b_acb; 626 kcondvar_t b_cv; 627 628 /* immutable */ 629 arc_buf_contents_t b_type; 630 uint64_t b_size; 631 uint64_t b_spa; 632 633 /* protected by arc state mutex */ 634 arc_state_t *b_state; 635 list_node_t b_arc_node; 636 637 /* updated atomically */ 638 clock_t b_arc_access; 639 640 /* self protecting */ 641 refcount_t b_refcnt; 642 643 l2arc_buf_hdr_t *b_l2hdr; 644 list_node_t b_l2node; 645}; 646 647#ifdef _KERNEL 648static int 649sysctl_vfs_zfs_arc_meta_limit(SYSCTL_HANDLER_ARGS) 650{ 651 uint64_t val; 652 int err; 653 654 val = arc_meta_limit; 655 err = sysctl_handle_64(oidp, &val, 0, req); 656 if (err != 0 || req->newptr == NULL) 657 return (err); 658 659 if (val <= 0 || val > arc_c_max) 660 return (EINVAL); 661 662 arc_meta_limit = val; 663 return (0); 664} 665#endif 666 667static arc_buf_t *arc_eviction_list; 668static kmutex_t arc_eviction_mtx; 669static arc_buf_hdr_t arc_eviction_hdr; 670 671#define GHOST_STATE(state) \ 672 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \ 673 (state) == arc_l2c_only) 674 675#define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_FLAG_IN_HASH_TABLE) 676#define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) 677#define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_FLAG_IO_ERROR) 678#define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_FLAG_PREFETCH) 679#define HDR_FREED_IN_READ(hdr) ((hdr)->b_flags & ARC_FLAG_FREED_IN_READ) 680#define HDR_BUF_AVAILABLE(hdr) ((hdr)->b_flags & ARC_FLAG_BUF_AVAILABLE) 681#define HDR_FREE_IN_PROGRESS(hdr) \ 682 ((hdr)->b_flags & ARC_FLAG_FREE_IN_PROGRESS) 683#define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_FLAG_L2CACHE) 684#define HDR_L2_READING(hdr) \ 685 ((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS && \ 686 (hdr)->b_l2hdr != NULL) 687#define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITING) 688#define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_FLAG_L2_EVICTED) 689#define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITE_HEAD) 690 691/* 692 * Other sizes 693 */ 694 695#define HDR_SIZE ((int64_t)sizeof (arc_buf_hdr_t)) 696#define L2HDR_SIZE ((int64_t)sizeof (l2arc_buf_hdr_t)) 697 698/* 699 * Hash table routines 700 */ 701 702#define HT_LOCK_PAD CACHE_LINE_SIZE 703 704struct ht_lock { 705 kmutex_t ht_lock; 706#ifdef _KERNEL 707 unsigned char pad[(HT_LOCK_PAD - sizeof (kmutex_t))]; 708#endif 709}; 710 711#define BUF_LOCKS 256 712typedef struct buf_hash_table { 713 uint64_t ht_mask; 714 arc_buf_hdr_t **ht_table; 715 struct ht_lock ht_locks[BUF_LOCKS] __aligned(CACHE_LINE_SIZE); 716} buf_hash_table_t; 717 718static buf_hash_table_t buf_hash_table; 719 720#define BUF_HASH_INDEX(spa, dva, birth) \ 721 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask) 722#define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)]) 723#define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock)) 724#define HDR_LOCK(hdr) \ 725 (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth))) 726 727uint64_t zfs_crc64_table[256]; 728 729/* 730 * Level 2 ARC 731 */ 732 733#define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */ 734#define L2ARC_HEADROOM 2 /* num of writes */ 735/* 736 * If we discover during ARC scan any buffers to be compressed, we boost 737 * our headroom for the next scanning cycle by this percentage multiple. 738 */ 739#define L2ARC_HEADROOM_BOOST 200 740#define L2ARC_FEED_SECS 1 /* caching interval secs */ 741#define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */ 742 743#define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent) 744#define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done) 745 746/* L2ARC Performance Tunables */ 747uint64_t l2arc_write_max = L2ARC_WRITE_SIZE; /* default max write size */ 748uint64_t l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra write during warmup */ 749uint64_t l2arc_headroom = L2ARC_HEADROOM; /* number of dev writes */ 750uint64_t l2arc_headroom_boost = L2ARC_HEADROOM_BOOST; 751uint64_t l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */ 752uint64_t l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval milliseconds */ 753boolean_t l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */ 754boolean_t l2arc_feed_again = B_TRUE; /* turbo warmup */ 755boolean_t l2arc_norw = B_TRUE; /* no reads during writes */ 756 757SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_max, CTLFLAG_RW, 758 &l2arc_write_max, 0, "max write size"); 759SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_boost, CTLFLAG_RW, 760 &l2arc_write_boost, 0, "extra write during warmup"); 761SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_headroom, CTLFLAG_RW, 762 &l2arc_headroom, 0, "number of dev writes"); 763SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_secs, CTLFLAG_RW, 764 &l2arc_feed_secs, 0, "interval seconds"); 765SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_min_ms, CTLFLAG_RW, 766 &l2arc_feed_min_ms, 0, "min interval milliseconds"); 767 768SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_noprefetch, CTLFLAG_RW, 769 &l2arc_noprefetch, 0, "don't cache prefetch bufs"); 770SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_feed_again, CTLFLAG_RW, 771 &l2arc_feed_again, 0, "turbo warmup"); 772SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_norw, CTLFLAG_RW, 773 &l2arc_norw, 0, "no reads during writes"); 774 775SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_size, CTLFLAG_RD, 776 &ARC_anon.arcs_size, 0, "size of anonymous state"); 777SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_metadata_lsize, CTLFLAG_RD, 778 &ARC_anon.arcs_lsize[ARC_BUFC_METADATA], 0, "size of anonymous state"); 779SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_data_lsize, CTLFLAG_RD, 780 &ARC_anon.arcs_lsize[ARC_BUFC_DATA], 0, "size of anonymous state"); 781 782SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_size, CTLFLAG_RD, 783 &ARC_mru.arcs_size, 0, "size of mru state"); 784SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_metadata_lsize, CTLFLAG_RD, 785 &ARC_mru.arcs_lsize[ARC_BUFC_METADATA], 0, "size of metadata in mru state"); 786SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_data_lsize, CTLFLAG_RD, 787 &ARC_mru.arcs_lsize[ARC_BUFC_DATA], 0, "size of data in mru state"); 788 789SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_size, CTLFLAG_RD, 790 &ARC_mru_ghost.arcs_size, 0, "size of mru ghost state"); 791SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_metadata_lsize, CTLFLAG_RD, 792 &ARC_mru_ghost.arcs_lsize[ARC_BUFC_METADATA], 0, 793 "size of metadata in mru ghost state"); 794SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_data_lsize, CTLFLAG_RD, 795 &ARC_mru_ghost.arcs_lsize[ARC_BUFC_DATA], 0, 796 "size of data in mru ghost state"); 797 798SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_size, CTLFLAG_RD, 799 &ARC_mfu.arcs_size, 0, "size of mfu state"); 800SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_metadata_lsize, CTLFLAG_RD, 801 &ARC_mfu.arcs_lsize[ARC_BUFC_METADATA], 0, "size of metadata in mfu state"); 802SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_data_lsize, CTLFLAG_RD, 803 &ARC_mfu.arcs_lsize[ARC_BUFC_DATA], 0, "size of data in mfu state"); 804 805SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_size, CTLFLAG_RD, 806 &ARC_mfu_ghost.arcs_size, 0, "size of mfu ghost state"); 807SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_metadata_lsize, CTLFLAG_RD, 808 &ARC_mfu_ghost.arcs_lsize[ARC_BUFC_METADATA], 0, 809 "size of metadata in mfu ghost state"); 810SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_data_lsize, CTLFLAG_RD, 811 &ARC_mfu_ghost.arcs_lsize[ARC_BUFC_DATA], 0, 812 "size of data in mfu ghost state"); 813 814SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2c_only_size, CTLFLAG_RD, 815 &ARC_l2c_only.arcs_size, 0, "size of mru state"); 816 817/* 818 * L2ARC Internals 819 */ 820typedef struct l2arc_dev { 821 vdev_t *l2ad_vdev; /* vdev */ 822 spa_t *l2ad_spa; /* spa */ 823 uint64_t l2ad_hand; /* next write location */ 824 uint64_t l2ad_start; /* first addr on device */ 825 uint64_t l2ad_end; /* last addr on device */ 826 uint64_t l2ad_evict; /* last addr eviction reached */ 827 boolean_t l2ad_first; /* first sweep through */ 828 boolean_t l2ad_writing; /* currently writing */ 829 list_t *l2ad_buflist; /* buffer list */ 830 list_node_t l2ad_node; /* device list node */ 831} l2arc_dev_t; 832 833static list_t L2ARC_dev_list; /* device list */ 834static list_t *l2arc_dev_list; /* device list pointer */ 835static kmutex_t l2arc_dev_mtx; /* device list mutex */ 836static l2arc_dev_t *l2arc_dev_last; /* last device used */ 837static kmutex_t l2arc_buflist_mtx; /* mutex for all buflists */ 838static list_t L2ARC_free_on_write; /* free after write buf list */ 839static list_t *l2arc_free_on_write; /* free after write list ptr */ 840static kmutex_t l2arc_free_on_write_mtx; /* mutex for list */ 841static uint64_t l2arc_ndev; /* number of devices */ 842 843typedef struct l2arc_read_callback { 844 arc_buf_t *l2rcb_buf; /* read buffer */ 845 spa_t *l2rcb_spa; /* spa */ 846 blkptr_t l2rcb_bp; /* original blkptr */ 847 zbookmark_phys_t l2rcb_zb; /* original bookmark */ 848 int l2rcb_flags; /* original flags */ 849 enum zio_compress l2rcb_compress; /* applied compress */ 850} l2arc_read_callback_t; 851 852typedef struct l2arc_write_callback { 853 l2arc_dev_t *l2wcb_dev; /* device info */ 854 arc_buf_hdr_t *l2wcb_head; /* head of write buflist */ 855} l2arc_write_callback_t; 856 857struct l2arc_buf_hdr { 858 /* protected by arc_buf_hdr mutex */ 859 l2arc_dev_t *b_dev; /* L2ARC device */ 860 uint64_t b_daddr; /* disk address, offset byte */ 861 /* compression applied to buffer data */ 862 enum zio_compress b_compress; 863 /* real alloc'd buffer size depending on b_compress applied */ 864 int b_asize; 865 /* temporary buffer holder for in-flight compressed data */ 866 void *b_tmp_cdata; 867}; 868 869typedef struct l2arc_data_free { 870 /* protected by l2arc_free_on_write_mtx */ 871 void *l2df_data; 872 size_t l2df_size; 873 void (*l2df_func)(void *, size_t); 874 list_node_t l2df_list_node; 875} l2arc_data_free_t; 876 877static kmutex_t l2arc_feed_thr_lock; 878static kcondvar_t l2arc_feed_thr_cv; 879static uint8_t l2arc_thread_exit; 880 881static void arc_get_data_buf(arc_buf_t *); 882static void arc_access(arc_buf_hdr_t *, kmutex_t *); 883static int arc_evict_needed(arc_buf_contents_t); 884static void arc_evict_ghost(arc_state_t *, uint64_t, int64_t); 885static void arc_buf_watch(arc_buf_t *); 886 887static boolean_t l2arc_write_eligible(uint64_t, arc_buf_hdr_t *); 888static void l2arc_read_done(zio_t *); 889static void l2arc_hdr_stat_add(void); 890static void l2arc_hdr_stat_remove(void); 891 892static boolean_t l2arc_compress_buf(l2arc_buf_hdr_t *); 893static void l2arc_decompress_zio(zio_t *, arc_buf_hdr_t *, enum zio_compress); 894static void l2arc_release_cdata_buf(arc_buf_hdr_t *); 895 896static uint64_t 897buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth) 898{ 899 uint8_t *vdva = (uint8_t *)dva; 900 uint64_t crc = -1ULL; 901 int i; 902 903 ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY); 904 905 for (i = 0; i < sizeof (dva_t); i++) 906 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ vdva[i]) & 0xFF]; 907 908 crc ^= (spa>>8) ^ birth; 909 910 return (crc); 911} 912 913#define BUF_EMPTY(buf) \ 914 ((buf)->b_dva.dva_word[0] == 0 && \ 915 (buf)->b_dva.dva_word[1] == 0 && \ 916 (buf)->b_cksum0 == 0) 917 918#define BUF_EQUAL(spa, dva, birth, buf) \ 919 ((buf)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \ 920 ((buf)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \ 921 ((buf)->b_birth == birth) && ((buf)->b_spa == spa) 922 923static void 924buf_discard_identity(arc_buf_hdr_t *hdr) 925{ 926 hdr->b_dva.dva_word[0] = 0; 927 hdr->b_dva.dva_word[1] = 0; 928 hdr->b_birth = 0; 929 hdr->b_cksum0 = 0; 930} 931 932static arc_buf_hdr_t * 933buf_hash_find(uint64_t spa, const blkptr_t *bp, kmutex_t **lockp) 934{ 935 const dva_t *dva = BP_IDENTITY(bp); 936 uint64_t birth = BP_PHYSICAL_BIRTH(bp); 937 uint64_t idx = BUF_HASH_INDEX(spa, dva, birth); 938 kmutex_t *hash_lock = BUF_HASH_LOCK(idx); 939 arc_buf_hdr_t *hdr; 940 941 mutex_enter(hash_lock); 942 for (hdr = buf_hash_table.ht_table[idx]; hdr != NULL; 943 hdr = hdr->b_hash_next) { 944 if (BUF_EQUAL(spa, dva, birth, hdr)) { 945 *lockp = hash_lock; 946 return (hdr); 947 } 948 } 949 mutex_exit(hash_lock); 950 *lockp = NULL; 951 return (NULL); 952} 953 954/* 955 * Insert an entry into the hash table. If there is already an element 956 * equal to elem in the hash table, then the already existing element 957 * will be returned and the new element will not be inserted. 958 * Otherwise returns NULL. 959 */ 960static arc_buf_hdr_t * 961buf_hash_insert(arc_buf_hdr_t *hdr, kmutex_t **lockp) 962{ 963 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth); 964 kmutex_t *hash_lock = BUF_HASH_LOCK(idx); 965 arc_buf_hdr_t *fhdr; 966 uint32_t i; 967 968 ASSERT(!DVA_IS_EMPTY(&hdr->b_dva)); 969 ASSERT(hdr->b_birth != 0); 970 ASSERT(!HDR_IN_HASH_TABLE(hdr)); 971 *lockp = hash_lock; 972 mutex_enter(hash_lock); 973 for (fhdr = buf_hash_table.ht_table[idx], i = 0; fhdr != NULL; 974 fhdr = fhdr->b_hash_next, i++) { 975 if (BUF_EQUAL(hdr->b_spa, &hdr->b_dva, hdr->b_birth, fhdr)) 976 return (fhdr); 977 } 978 979 hdr->b_hash_next = buf_hash_table.ht_table[idx]; 980 buf_hash_table.ht_table[idx] = hdr; 981 hdr->b_flags |= ARC_FLAG_IN_HASH_TABLE; 982 983 /* collect some hash table performance data */ 984 if (i > 0) { 985 ARCSTAT_BUMP(arcstat_hash_collisions); 986 if (i == 1) 987 ARCSTAT_BUMP(arcstat_hash_chains); 988 989 ARCSTAT_MAX(arcstat_hash_chain_max, i); 990 } 991 992 ARCSTAT_BUMP(arcstat_hash_elements); 993 ARCSTAT_MAXSTAT(arcstat_hash_elements); 994 995 return (NULL); 996} 997 998static void 999buf_hash_remove(arc_buf_hdr_t *hdr) 1000{ 1001 arc_buf_hdr_t *fhdr, **hdrp; 1002 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth); 1003 1004 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx))); 1005 ASSERT(HDR_IN_HASH_TABLE(hdr)); 1006 1007 hdrp = &buf_hash_table.ht_table[idx]; 1008 while ((fhdr = *hdrp) != hdr) { 1009 ASSERT(fhdr != NULL); 1010 hdrp = &fhdr->b_hash_next; 1011 } 1012 *hdrp = hdr->b_hash_next; 1013 hdr->b_hash_next = NULL; 1014 hdr->b_flags &= ~ARC_FLAG_IN_HASH_TABLE; 1015 1016 /* collect some hash table performance data */ 1017 ARCSTAT_BUMPDOWN(arcstat_hash_elements); 1018 1019 if (buf_hash_table.ht_table[idx] && 1020 buf_hash_table.ht_table[idx]->b_hash_next == NULL) 1021 ARCSTAT_BUMPDOWN(arcstat_hash_chains); 1022} 1023 1024/* 1025 * Global data structures and functions for the buf kmem cache. 1026 */ 1027static kmem_cache_t *hdr_cache; 1028static kmem_cache_t *buf_cache; 1029 1030static void 1031buf_fini(void) 1032{ 1033 int i; 1034 1035 kmem_free(buf_hash_table.ht_table, 1036 (buf_hash_table.ht_mask + 1) * sizeof (void *)); 1037 for (i = 0; i < BUF_LOCKS; i++) 1038 mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock); 1039 kmem_cache_destroy(hdr_cache); 1040 kmem_cache_destroy(buf_cache); 1041} 1042 1043/* 1044 * Constructor callback - called when the cache is empty 1045 * and a new buf is requested. 1046 */ 1047/* ARGSUSED */ 1048static int 1049hdr_cons(void *vbuf, void *unused, int kmflag) 1050{ 1051 arc_buf_hdr_t *hdr = vbuf; 1052 1053 bzero(hdr, sizeof (arc_buf_hdr_t)); 1054 refcount_create(&hdr->b_refcnt); 1055 cv_init(&hdr->b_cv, NULL, CV_DEFAULT, NULL); 1056 mutex_init(&hdr->b_freeze_lock, NULL, MUTEX_DEFAULT, NULL); 1057 arc_space_consume(sizeof (arc_buf_hdr_t), ARC_SPACE_HDRS); 1058 1059 return (0); 1060} 1061 1062/* ARGSUSED */ 1063static int 1064buf_cons(void *vbuf, void *unused, int kmflag) 1065{ 1066 arc_buf_t *buf = vbuf; 1067 1068 bzero(buf, sizeof (arc_buf_t)); 1069 mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL); 1070 arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS); 1071 1072 return (0); 1073} 1074 1075/* 1076 * Destructor callback - called when a cached buf is 1077 * no longer required. 1078 */ 1079/* ARGSUSED */ 1080static void 1081hdr_dest(void *vbuf, void *unused) 1082{ 1083 arc_buf_hdr_t *hdr = vbuf; 1084 1085 ASSERT(BUF_EMPTY(hdr)); 1086 refcount_destroy(&hdr->b_refcnt); 1087 cv_destroy(&hdr->b_cv); 1088 mutex_destroy(&hdr->b_freeze_lock); 1089 arc_space_return(sizeof (arc_buf_hdr_t), ARC_SPACE_HDRS); 1090} 1091 1092/* ARGSUSED */ 1093static void 1094buf_dest(void *vbuf, void *unused) 1095{ 1096 arc_buf_t *buf = vbuf; 1097 1098 mutex_destroy(&buf->b_evict_lock); 1099 arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS); 1100} 1101 1102/* 1103 * Reclaim callback -- invoked when memory is low. 1104 */ 1105/* ARGSUSED */ 1106static void 1107hdr_recl(void *unused) 1108{ 1109 dprintf("hdr_recl called\n"); 1110 /* 1111 * umem calls the reclaim func when we destroy the buf cache, 1112 * which is after we do arc_fini(). 1113 */ 1114 if (!arc_dead) 1115 cv_signal(&arc_reclaim_thr_cv); 1116} 1117 1118static void 1119buf_init(void) 1120{ 1121 uint64_t *ct; 1122 uint64_t hsize = 1ULL << 12; 1123 int i, j; 1124 1125 /* 1126 * The hash table is big enough to fill all of physical memory 1127 * with an average block size of zfs_arc_average_blocksize (default 8K). 1128 * By default, the table will take up 1129 * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers). 1130 */ 1131 while (hsize * zfs_arc_average_blocksize < (uint64_t)physmem * PAGESIZE) 1132 hsize <<= 1; 1133retry: 1134 buf_hash_table.ht_mask = hsize - 1; 1135 buf_hash_table.ht_table = 1136 kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP); 1137 if (buf_hash_table.ht_table == NULL) { 1138 ASSERT(hsize > (1ULL << 8)); 1139 hsize >>= 1; 1140 goto retry; 1141 } 1142 1143 hdr_cache = kmem_cache_create("arc_buf_hdr_t", sizeof (arc_buf_hdr_t), 1144 0, hdr_cons, hdr_dest, hdr_recl, NULL, NULL, 0); 1145 buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t), 1146 0, buf_cons, buf_dest, NULL, NULL, NULL, 0); 1147 1148 for (i = 0; i < 256; i++) 1149 for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--) 1150 *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY); 1151 1152 for (i = 0; i < BUF_LOCKS; i++) { 1153 mutex_init(&buf_hash_table.ht_locks[i].ht_lock, 1154 NULL, MUTEX_DEFAULT, NULL); 1155 } 1156} 1157 1158#define ARC_MINTIME (hz>>4) /* 62 ms */ 1159 1160static void 1161arc_cksum_verify(arc_buf_t *buf) 1162{ 1163 zio_cksum_t zc; 1164 1165 if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 1166 return; 1167 1168 mutex_enter(&buf->b_hdr->b_freeze_lock); 1169 if (buf->b_hdr->b_freeze_cksum == NULL || 1170 (buf->b_hdr->b_flags & ARC_FLAG_IO_ERROR)) { 1171 mutex_exit(&buf->b_hdr->b_freeze_lock); 1172 return; 1173 } 1174 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc); 1175 if (!ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc)) 1176 panic("buffer modified while frozen!"); 1177 mutex_exit(&buf->b_hdr->b_freeze_lock); 1178} 1179 1180static int 1181arc_cksum_equal(arc_buf_t *buf) 1182{ 1183 zio_cksum_t zc; 1184 int equal; 1185 1186 mutex_enter(&buf->b_hdr->b_freeze_lock); 1187 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc); 1188 equal = ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc); 1189 mutex_exit(&buf->b_hdr->b_freeze_lock); 1190 1191 return (equal); 1192} 1193 1194static void 1195arc_cksum_compute(arc_buf_t *buf, boolean_t force) 1196{ 1197 if (!force && !(zfs_flags & ZFS_DEBUG_MODIFY)) 1198 return; 1199 1200 mutex_enter(&buf->b_hdr->b_freeze_lock); 1201 if (buf->b_hdr->b_freeze_cksum != NULL) { 1202 mutex_exit(&buf->b_hdr->b_freeze_lock); 1203 return; 1204 } 1205 buf->b_hdr->b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t), KM_SLEEP); 1206 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, 1207 buf->b_hdr->b_freeze_cksum); 1208 mutex_exit(&buf->b_hdr->b_freeze_lock); 1209#ifdef illumos 1210 arc_buf_watch(buf); 1211#endif 1212} 1213 1214#ifdef illumos 1215#ifndef _KERNEL 1216typedef struct procctl { 1217 long cmd; 1218 prwatch_t prwatch; 1219} procctl_t; 1220#endif 1221 1222/* ARGSUSED */ 1223static void 1224arc_buf_unwatch(arc_buf_t *buf) 1225{ 1226#ifndef _KERNEL 1227 if (arc_watch) { 1228 int result; 1229 procctl_t ctl; 1230 ctl.cmd = PCWATCH; 1231 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data; 1232 ctl.prwatch.pr_size = 0; 1233 ctl.prwatch.pr_wflags = 0; 1234 result = write(arc_procfd, &ctl, sizeof (ctl)); 1235 ASSERT3U(result, ==, sizeof (ctl)); 1236 } 1237#endif 1238} 1239 1240/* ARGSUSED */ 1241static void 1242arc_buf_watch(arc_buf_t *buf) 1243{ 1244#ifndef _KERNEL 1245 if (arc_watch) { 1246 int result; 1247 procctl_t ctl; 1248 ctl.cmd = PCWATCH; 1249 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data; 1250 ctl.prwatch.pr_size = buf->b_hdr->b_size; 1251 ctl.prwatch.pr_wflags = WA_WRITE; 1252 result = write(arc_procfd, &ctl, sizeof (ctl)); 1253 ASSERT3U(result, ==, sizeof (ctl)); 1254 } 1255#endif 1256} 1257#endif /* illumos */ 1258 1259void 1260arc_buf_thaw(arc_buf_t *buf) 1261{ 1262 if (zfs_flags & ZFS_DEBUG_MODIFY) { 1263 if (buf->b_hdr->b_state != arc_anon) 1264 panic("modifying non-anon buffer!"); 1265 if (buf->b_hdr->b_flags & ARC_FLAG_IO_IN_PROGRESS) 1266 panic("modifying buffer while i/o in progress!"); 1267 arc_cksum_verify(buf); 1268 } 1269 1270 mutex_enter(&buf->b_hdr->b_freeze_lock); 1271 if (buf->b_hdr->b_freeze_cksum != NULL) { 1272 kmem_free(buf->b_hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 1273 buf->b_hdr->b_freeze_cksum = NULL; 1274 } 1275 1276 if (zfs_flags & ZFS_DEBUG_MODIFY) { 1277 if (buf->b_hdr->b_thawed) 1278 kmem_free(buf->b_hdr->b_thawed, 1); 1279 buf->b_hdr->b_thawed = kmem_alloc(1, KM_SLEEP); 1280 } 1281 1282 mutex_exit(&buf->b_hdr->b_freeze_lock); 1283 1284#ifdef illumos 1285 arc_buf_unwatch(buf); 1286#endif 1287} 1288 1289void 1290arc_buf_freeze(arc_buf_t *buf) 1291{ 1292 kmutex_t *hash_lock; 1293 1294 if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 1295 return; 1296 1297 hash_lock = HDR_LOCK(buf->b_hdr); 1298 mutex_enter(hash_lock); 1299 1300 ASSERT(buf->b_hdr->b_freeze_cksum != NULL || 1301 buf->b_hdr->b_state == arc_anon); 1302 arc_cksum_compute(buf, B_FALSE); 1303 mutex_exit(hash_lock); 1304 1305} 1306 1307static void 1308get_buf_info(arc_buf_hdr_t *hdr, arc_state_t *state, list_t **list, kmutex_t **lock) 1309{ 1310 uint64_t buf_hashid = buf_hash(hdr->b_spa, &hdr->b_dva, hdr->b_birth); 1311 1312 if (hdr->b_type == ARC_BUFC_METADATA) 1313 buf_hashid &= (ARC_BUFC_NUMMETADATALISTS - 1); 1314 else { 1315 buf_hashid &= (ARC_BUFC_NUMDATALISTS - 1); 1316 buf_hashid += ARC_BUFC_NUMMETADATALISTS; 1317 } 1318 1319 *list = &state->arcs_lists[buf_hashid]; 1320 *lock = ARCS_LOCK(state, buf_hashid); 1321} 1322 1323 1324static void 1325add_reference(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, void *tag) 1326{ 1327 ASSERT(MUTEX_HELD(hash_lock)); 1328 1329 if ((refcount_add(&hdr->b_refcnt, tag) == 1) && 1330 (hdr->b_state != arc_anon)) { 1331 uint64_t delta = hdr->b_size * hdr->b_datacnt; 1332 uint64_t *size = &hdr->b_state->arcs_lsize[hdr->b_type]; 1333 list_t *list; 1334 kmutex_t *lock; 1335 1336 get_buf_info(hdr, hdr->b_state, &list, &lock); 1337 ASSERT(!MUTEX_HELD(lock)); 1338 mutex_enter(lock); 1339 ASSERT(list_link_active(&hdr->b_arc_node)); 1340 list_remove(list, hdr); 1341 if (GHOST_STATE(hdr->b_state)) { 1342 ASSERT0(hdr->b_datacnt); 1343 ASSERT3P(hdr->b_buf, ==, NULL); 1344 delta = hdr->b_size; 1345 } 1346 ASSERT(delta > 0); 1347 ASSERT3U(*size, >=, delta); 1348 atomic_add_64(size, -delta); 1349 mutex_exit(lock); 1350 /* remove the prefetch flag if we get a reference */ 1351 if (hdr->b_flags & ARC_FLAG_PREFETCH) 1352 hdr->b_flags &= ~ARC_FLAG_PREFETCH; 1353 } 1354} 1355 1356static int 1357remove_reference(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, void *tag) 1358{ 1359 int cnt; 1360 arc_state_t *state = hdr->b_state; 1361 1362 ASSERT(state == arc_anon || MUTEX_HELD(hash_lock)); 1363 ASSERT(!GHOST_STATE(state)); 1364 1365 if (((cnt = refcount_remove(&hdr->b_refcnt, tag)) == 0) && 1366 (state != arc_anon)) { 1367 uint64_t *size = &state->arcs_lsize[hdr->b_type]; 1368 list_t *list; 1369 kmutex_t *lock; 1370 1371 get_buf_info(hdr, state, &list, &lock); 1372 ASSERT(!MUTEX_HELD(lock)); 1373 mutex_enter(lock); 1374 ASSERT(!list_link_active(&hdr->b_arc_node)); 1375 list_insert_head(list, hdr); 1376 ASSERT(hdr->b_datacnt > 0); 1377 atomic_add_64(size, hdr->b_size * hdr->b_datacnt); 1378 mutex_exit(lock); 1379 } 1380 return (cnt); 1381} 1382 1383/* 1384 * Move the supplied buffer to the indicated state. The mutex 1385 * for the buffer must be held by the caller. 1386 */ 1387static void 1388arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *hdr, 1389 kmutex_t *hash_lock) 1390{ 1391 arc_state_t *old_state = hdr->b_state; 1392 int64_t refcnt = refcount_count(&hdr->b_refcnt); 1393 uint64_t from_delta, to_delta; 1394 list_t *list; 1395 kmutex_t *lock; 1396 1397 ASSERT(MUTEX_HELD(hash_lock)); 1398 ASSERT3P(new_state, !=, old_state); 1399 ASSERT(refcnt == 0 || hdr->b_datacnt > 0); 1400 ASSERT(hdr->b_datacnt == 0 || !GHOST_STATE(new_state)); 1401 ASSERT(hdr->b_datacnt <= 1 || old_state != arc_anon); 1402 1403 from_delta = to_delta = hdr->b_datacnt * hdr->b_size; 1404 1405 /* 1406 * If this buffer is evictable, transfer it from the 1407 * old state list to the new state list. 1408 */ 1409 if (refcnt == 0) { 1410 if (old_state != arc_anon) { 1411 int use_mutex; 1412 uint64_t *size = &old_state->arcs_lsize[hdr->b_type]; 1413 1414 get_buf_info(hdr, old_state, &list, &lock); 1415 use_mutex = !MUTEX_HELD(lock); 1416 if (use_mutex) 1417 mutex_enter(lock); 1418 1419 ASSERT(list_link_active(&hdr->b_arc_node)); 1420 list_remove(list, hdr); 1421 1422 /* 1423 * If prefetching out of the ghost cache, 1424 * we will have a non-zero datacnt. 1425 */ 1426 if (GHOST_STATE(old_state) && hdr->b_datacnt == 0) { 1427 /* ghost elements have a ghost size */ 1428 ASSERT(hdr->b_buf == NULL); 1429 from_delta = hdr->b_size; 1430 } 1431 ASSERT3U(*size, >=, from_delta); 1432 atomic_add_64(size, -from_delta); 1433 1434 if (use_mutex) 1435 mutex_exit(lock); 1436 } 1437 if (new_state != arc_anon) { 1438 int use_mutex; 1439 uint64_t *size = &new_state->arcs_lsize[hdr->b_type]; 1440 1441 get_buf_info(hdr, new_state, &list, &lock); 1442 use_mutex = !MUTEX_HELD(lock); 1443 if (use_mutex) 1444 mutex_enter(lock); 1445 1446 list_insert_head(list, hdr); 1447 1448 /* ghost elements have a ghost size */ 1449 if (GHOST_STATE(new_state)) { 1450 ASSERT(hdr->b_datacnt == 0); 1451 ASSERT(hdr->b_buf == NULL); 1452 to_delta = hdr->b_size; 1453 } 1454 atomic_add_64(size, to_delta); 1455 1456 if (use_mutex) 1457 mutex_exit(lock); 1458 } 1459 } 1460 1461 ASSERT(!BUF_EMPTY(hdr)); 1462 if (new_state == arc_anon && HDR_IN_HASH_TABLE(hdr)) 1463 buf_hash_remove(hdr); 1464 1465 /* adjust state sizes */ 1466 if (to_delta) 1467 atomic_add_64(&new_state->arcs_size, to_delta); 1468 if (from_delta) { 1469 ASSERT3U(old_state->arcs_size, >=, from_delta); 1470 atomic_add_64(&old_state->arcs_size, -from_delta); 1471 } 1472 hdr->b_state = new_state; 1473 1474 /* adjust l2arc hdr stats */ 1475 if (new_state == arc_l2c_only) 1476 l2arc_hdr_stat_add(); 1477 else if (old_state == arc_l2c_only) 1478 l2arc_hdr_stat_remove(); 1479} 1480 1481void 1482arc_space_consume(uint64_t space, arc_space_type_t type) 1483{ 1484 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES); 1485 1486 switch (type) { 1487 case ARC_SPACE_DATA: 1488 ARCSTAT_INCR(arcstat_data_size, space); 1489 break; 1490 case ARC_SPACE_OTHER: 1491 ARCSTAT_INCR(arcstat_other_size, space); 1492 break; 1493 case ARC_SPACE_HDRS: 1494 ARCSTAT_INCR(arcstat_hdr_size, space); 1495 break; 1496 case ARC_SPACE_L2HDRS: 1497 ARCSTAT_INCR(arcstat_l2_hdr_size, space); 1498 break; 1499 } 1500 1501 ARCSTAT_INCR(arcstat_meta_used, space); 1502 atomic_add_64(&arc_size, space); 1503} 1504 1505void 1506arc_space_return(uint64_t space, arc_space_type_t type) 1507{ 1508 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES); 1509 1510 switch (type) { 1511 case ARC_SPACE_DATA: 1512 ARCSTAT_INCR(arcstat_data_size, -space); 1513 break; 1514 case ARC_SPACE_OTHER: 1515 ARCSTAT_INCR(arcstat_other_size, -space); 1516 break; 1517 case ARC_SPACE_HDRS: 1518 ARCSTAT_INCR(arcstat_hdr_size, -space); 1519 break; 1520 case ARC_SPACE_L2HDRS: 1521 ARCSTAT_INCR(arcstat_l2_hdr_size, -space); 1522 break; 1523 } 1524 1525 ASSERT(arc_meta_used >= space); 1526 if (arc_meta_max < arc_meta_used) 1527 arc_meta_max = arc_meta_used; 1528 ARCSTAT_INCR(arcstat_meta_used, -space); 1529 ASSERT(arc_size >= space); 1530 atomic_add_64(&arc_size, -space); 1531} 1532 1533arc_buf_t * 1534arc_buf_alloc(spa_t *spa, int size, void *tag, arc_buf_contents_t type) 1535{ 1536 arc_buf_hdr_t *hdr; 1537 arc_buf_t *buf; 1538 1539 ASSERT3U(size, >, 0); 1540 hdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE); 1541 ASSERT(BUF_EMPTY(hdr)); 1542 hdr->b_size = size; 1543 hdr->b_type = type; 1544 hdr->b_spa = spa_load_guid(spa); 1545 hdr->b_state = arc_anon; 1546 hdr->b_arc_access = 0; 1547 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 1548 buf->b_hdr = hdr; 1549 buf->b_data = NULL; 1550 buf->b_efunc = NULL; 1551 buf->b_private = NULL; 1552 buf->b_next = NULL; 1553 hdr->b_buf = buf; 1554 arc_get_data_buf(buf); 1555 hdr->b_datacnt = 1; 1556 hdr->b_flags = 0; 1557 ASSERT(refcount_is_zero(&hdr->b_refcnt)); 1558 (void) refcount_add(&hdr->b_refcnt, tag); 1559 1560 return (buf); 1561} 1562 1563static char *arc_onloan_tag = "onloan"; 1564 1565/* 1566 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in 1567 * flight data by arc_tempreserve_space() until they are "returned". Loaned 1568 * buffers must be returned to the arc before they can be used by the DMU or 1569 * freed. 1570 */ 1571arc_buf_t * 1572arc_loan_buf(spa_t *spa, int size) 1573{ 1574 arc_buf_t *buf; 1575 1576 buf = arc_buf_alloc(spa, size, arc_onloan_tag, ARC_BUFC_DATA); 1577 1578 atomic_add_64(&arc_loaned_bytes, size); 1579 return (buf); 1580} 1581 1582/* 1583 * Return a loaned arc buffer to the arc. 1584 */ 1585void 1586arc_return_buf(arc_buf_t *buf, void *tag) 1587{ 1588 arc_buf_hdr_t *hdr = buf->b_hdr; 1589 1590 ASSERT(buf->b_data != NULL); 1591 (void) refcount_add(&hdr->b_refcnt, tag); 1592 (void) refcount_remove(&hdr->b_refcnt, arc_onloan_tag); 1593 1594 atomic_add_64(&arc_loaned_bytes, -hdr->b_size); 1595} 1596 1597/* Detach an arc_buf from a dbuf (tag) */ 1598void 1599arc_loan_inuse_buf(arc_buf_t *buf, void *tag) 1600{ 1601 arc_buf_hdr_t *hdr; 1602 1603 ASSERT(buf->b_data != NULL); 1604 hdr = buf->b_hdr; 1605 (void) refcount_add(&hdr->b_refcnt, arc_onloan_tag); 1606 (void) refcount_remove(&hdr->b_refcnt, tag); 1607 buf->b_efunc = NULL; 1608 buf->b_private = NULL; 1609 1610 atomic_add_64(&arc_loaned_bytes, hdr->b_size); 1611} 1612 1613static arc_buf_t * 1614arc_buf_clone(arc_buf_t *from) 1615{ 1616 arc_buf_t *buf; 1617 arc_buf_hdr_t *hdr = from->b_hdr; 1618 uint64_t size = hdr->b_size; 1619 1620 ASSERT(hdr->b_state != arc_anon); 1621 1622 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 1623 buf->b_hdr = hdr; 1624 buf->b_data = NULL; 1625 buf->b_efunc = NULL; 1626 buf->b_private = NULL; 1627 buf->b_next = hdr->b_buf; 1628 hdr->b_buf = buf; 1629 arc_get_data_buf(buf); 1630 bcopy(from->b_data, buf->b_data, size); 1631 1632 /* 1633 * This buffer already exists in the arc so create a duplicate 1634 * copy for the caller. If the buffer is associated with user data 1635 * then track the size and number of duplicates. These stats will be 1636 * updated as duplicate buffers are created and destroyed. 1637 */ 1638 if (hdr->b_type == ARC_BUFC_DATA) { 1639 ARCSTAT_BUMP(arcstat_duplicate_buffers); 1640 ARCSTAT_INCR(arcstat_duplicate_buffers_size, size); 1641 } 1642 hdr->b_datacnt += 1; 1643 return (buf); 1644} 1645 1646void 1647arc_buf_add_ref(arc_buf_t *buf, void* tag) 1648{ 1649 arc_buf_hdr_t *hdr; 1650 kmutex_t *hash_lock; 1651 1652 /* 1653 * Check to see if this buffer is evicted. Callers 1654 * must verify b_data != NULL to know if the add_ref 1655 * was successful. 1656 */ 1657 mutex_enter(&buf->b_evict_lock); 1658 if (buf->b_data == NULL) { 1659 mutex_exit(&buf->b_evict_lock); 1660 return; 1661 } 1662 hash_lock = HDR_LOCK(buf->b_hdr); 1663 mutex_enter(hash_lock); 1664 hdr = buf->b_hdr; 1665 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 1666 mutex_exit(&buf->b_evict_lock); 1667 1668 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu); 1669 add_reference(hdr, hash_lock, tag); 1670 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr); 1671 arc_access(hdr, hash_lock); 1672 mutex_exit(hash_lock); 1673 ARCSTAT_BUMP(arcstat_hits); 1674 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_FLAG_PREFETCH), 1675 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA, 1676 data, metadata, hits); 1677} 1678 1679static void 1680arc_buf_free_on_write(void *data, size_t size, 1681 void (*free_func)(void *, size_t)) 1682{ 1683 l2arc_data_free_t *df; 1684 1685 df = kmem_alloc(sizeof (l2arc_data_free_t), KM_SLEEP); 1686 df->l2df_data = data; 1687 df->l2df_size = size; 1688 df->l2df_func = free_func; 1689 mutex_enter(&l2arc_free_on_write_mtx); 1690 list_insert_head(l2arc_free_on_write, df); 1691 mutex_exit(&l2arc_free_on_write_mtx); 1692} 1693 1694/* 1695 * Free the arc data buffer. If it is an l2arc write in progress, 1696 * the buffer is placed on l2arc_free_on_write to be freed later. 1697 */ 1698static void 1699arc_buf_data_free(arc_buf_t *buf, void (*free_func)(void *, size_t)) 1700{ 1701 arc_buf_hdr_t *hdr = buf->b_hdr; 1702 1703 if (HDR_L2_WRITING(hdr)) { 1704 arc_buf_free_on_write(buf->b_data, hdr->b_size, free_func); 1705 ARCSTAT_BUMP(arcstat_l2_free_on_write); 1706 } else { 1707 free_func(buf->b_data, hdr->b_size); 1708 } 1709} 1710 1711/* 1712 * Free up buf->b_data and if 'remove' is set, then pull the 1713 * arc_buf_t off of the the arc_buf_hdr_t's list and free it. 1714 */ 1715static void 1716arc_buf_l2_cdata_free(arc_buf_hdr_t *hdr) 1717{ 1718 l2arc_buf_hdr_t *l2hdr = hdr->b_l2hdr; 1719 1720 ASSERT(MUTEX_HELD(&l2arc_buflist_mtx)); 1721 1722 if (l2hdr->b_tmp_cdata == NULL) 1723 return; 1724 1725 ASSERT(HDR_L2_WRITING(hdr)); 1726 arc_buf_free_on_write(l2hdr->b_tmp_cdata, hdr->b_size, 1727 zio_data_buf_free); 1728 ARCSTAT_BUMP(arcstat_l2_cdata_free_on_write); 1729 l2hdr->b_tmp_cdata = NULL; 1730} 1731 1732static void 1733arc_buf_destroy(arc_buf_t *buf, boolean_t recycle, boolean_t remove) 1734{ 1735 arc_buf_t **bufp; 1736 1737 /* free up data associated with the buf */ 1738 if (buf->b_data) { 1739 arc_state_t *state = buf->b_hdr->b_state; 1740 uint64_t size = buf->b_hdr->b_size; 1741 arc_buf_contents_t type = buf->b_hdr->b_type; 1742 1743 arc_cksum_verify(buf); 1744#ifdef illumos 1745 arc_buf_unwatch(buf); 1746#endif 1747 1748 if (!recycle) { 1749 if (type == ARC_BUFC_METADATA) { 1750 arc_buf_data_free(buf, zio_buf_free); 1751 arc_space_return(size, ARC_SPACE_DATA); 1752 } else { 1753 ASSERT(type == ARC_BUFC_DATA); 1754 arc_buf_data_free(buf, zio_data_buf_free); 1755 ARCSTAT_INCR(arcstat_data_size, -size); 1756 atomic_add_64(&arc_size, -size); 1757 } 1758 } 1759 if (list_link_active(&buf->b_hdr->b_arc_node)) { 1760 uint64_t *cnt = &state->arcs_lsize[type]; 1761 1762 ASSERT(refcount_is_zero(&buf->b_hdr->b_refcnt)); 1763 ASSERT(state != arc_anon); 1764 1765 ASSERT3U(*cnt, >=, size); 1766 atomic_add_64(cnt, -size); 1767 } 1768 ASSERT3U(state->arcs_size, >=, size); 1769 atomic_add_64(&state->arcs_size, -size); 1770 buf->b_data = NULL; 1771 1772 /* 1773 * If we're destroying a duplicate buffer make sure 1774 * that the appropriate statistics are updated. 1775 */ 1776 if (buf->b_hdr->b_datacnt > 1 && 1777 buf->b_hdr->b_type == ARC_BUFC_DATA) { 1778 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers); 1779 ARCSTAT_INCR(arcstat_duplicate_buffers_size, -size); 1780 } 1781 ASSERT(buf->b_hdr->b_datacnt > 0); 1782 buf->b_hdr->b_datacnt -= 1; 1783 } 1784 1785 /* only remove the buf if requested */ 1786 if (!remove) 1787 return; 1788 1789 /* remove the buf from the hdr list */ 1790 for (bufp = &buf->b_hdr->b_buf; *bufp != buf; bufp = &(*bufp)->b_next) 1791 continue; 1792 *bufp = buf->b_next; 1793 buf->b_next = NULL; 1794 1795 ASSERT(buf->b_efunc == NULL); 1796 1797 /* clean up the buf */ 1798 buf->b_hdr = NULL; 1799 kmem_cache_free(buf_cache, buf); 1800} 1801 1802static void 1803arc_hdr_destroy(arc_buf_hdr_t *hdr) 1804{ 1805 ASSERT(refcount_is_zero(&hdr->b_refcnt)); 1806 ASSERT3P(hdr->b_state, ==, arc_anon); 1807 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 1808 l2arc_buf_hdr_t *l2hdr = hdr->b_l2hdr; 1809 1810 if (l2hdr != NULL) { 1811 boolean_t buflist_held = MUTEX_HELD(&l2arc_buflist_mtx); 1812 /* 1813 * To prevent arc_free() and l2arc_evict() from 1814 * attempting to free the same buffer at the same time, 1815 * a FREE_IN_PROGRESS flag is given to arc_free() to 1816 * give it priority. l2arc_evict() can't destroy this 1817 * header while we are waiting on l2arc_buflist_mtx. 1818 * 1819 * The hdr may be removed from l2ad_buflist before we 1820 * grab l2arc_buflist_mtx, so b_l2hdr is rechecked. 1821 */ 1822 if (!buflist_held) { 1823 mutex_enter(&l2arc_buflist_mtx); 1824 l2hdr = hdr->b_l2hdr; 1825 } 1826 1827 if (l2hdr != NULL) { 1828 trim_map_free(l2hdr->b_dev->l2ad_vdev, l2hdr->b_daddr, 1829 hdr->b_size, 0); 1830 list_remove(l2hdr->b_dev->l2ad_buflist, hdr); 1831 arc_buf_l2_cdata_free(hdr); 1832 ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size); 1833 ARCSTAT_INCR(arcstat_l2_asize, -l2hdr->b_asize); 1834 vdev_space_update(l2hdr->b_dev->l2ad_vdev, 1835 -l2hdr->b_asize, 0, 0); 1836 kmem_free(l2hdr, sizeof (l2arc_buf_hdr_t)); 1837 if (hdr->b_state == arc_l2c_only) 1838 l2arc_hdr_stat_remove(); 1839 hdr->b_l2hdr = NULL; 1840 } 1841 1842 if (!buflist_held) 1843 mutex_exit(&l2arc_buflist_mtx); 1844 } 1845 1846 if (!BUF_EMPTY(hdr)) { 1847 ASSERT(!HDR_IN_HASH_TABLE(hdr)); 1848 buf_discard_identity(hdr); 1849 } 1850 while (hdr->b_buf) { 1851 arc_buf_t *buf = hdr->b_buf; 1852 1853 if (buf->b_efunc) { 1854 mutex_enter(&arc_eviction_mtx); 1855 mutex_enter(&buf->b_evict_lock); 1856 ASSERT(buf->b_hdr != NULL); 1857 arc_buf_destroy(hdr->b_buf, FALSE, FALSE); 1858 hdr->b_buf = buf->b_next; 1859 buf->b_hdr = &arc_eviction_hdr; 1860 buf->b_next = arc_eviction_list; 1861 arc_eviction_list = buf; 1862 mutex_exit(&buf->b_evict_lock); 1863 mutex_exit(&arc_eviction_mtx); 1864 } else { 1865 arc_buf_destroy(hdr->b_buf, FALSE, TRUE); 1866 } 1867 } 1868 if (hdr->b_freeze_cksum != NULL) { 1869 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 1870 hdr->b_freeze_cksum = NULL; 1871 } 1872 if (hdr->b_thawed) { 1873 kmem_free(hdr->b_thawed, 1); 1874 hdr->b_thawed = NULL; 1875 } 1876 1877 ASSERT(!list_link_active(&hdr->b_arc_node)); 1878 ASSERT3P(hdr->b_hash_next, ==, NULL); 1879 ASSERT3P(hdr->b_acb, ==, NULL); 1880 kmem_cache_free(hdr_cache, hdr); 1881} 1882 1883void 1884arc_buf_free(arc_buf_t *buf, void *tag) 1885{ 1886 arc_buf_hdr_t *hdr = buf->b_hdr; 1887 int hashed = hdr->b_state != arc_anon; 1888 1889 ASSERT(buf->b_efunc == NULL); 1890 ASSERT(buf->b_data != NULL); 1891 1892 if (hashed) { 1893 kmutex_t *hash_lock = HDR_LOCK(hdr); 1894 1895 mutex_enter(hash_lock); 1896 hdr = buf->b_hdr; 1897 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 1898 1899 (void) remove_reference(hdr, hash_lock, tag); 1900 if (hdr->b_datacnt > 1) { 1901 arc_buf_destroy(buf, FALSE, TRUE); 1902 } else { 1903 ASSERT(buf == hdr->b_buf); 1904 ASSERT(buf->b_efunc == NULL); 1905 hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE; 1906 } 1907 mutex_exit(hash_lock); 1908 } else if (HDR_IO_IN_PROGRESS(hdr)) { 1909 int destroy_hdr; 1910 /* 1911 * We are in the middle of an async write. Don't destroy 1912 * this buffer unless the write completes before we finish 1913 * decrementing the reference count. 1914 */ 1915 mutex_enter(&arc_eviction_mtx); 1916 (void) remove_reference(hdr, NULL, tag); 1917 ASSERT(refcount_is_zero(&hdr->b_refcnt)); 1918 destroy_hdr = !HDR_IO_IN_PROGRESS(hdr); 1919 mutex_exit(&arc_eviction_mtx); 1920 if (destroy_hdr) 1921 arc_hdr_destroy(hdr); 1922 } else { 1923 if (remove_reference(hdr, NULL, tag) > 0) 1924 arc_buf_destroy(buf, FALSE, TRUE); 1925 else 1926 arc_hdr_destroy(hdr); 1927 } 1928} 1929 1930boolean_t 1931arc_buf_remove_ref(arc_buf_t *buf, void* tag) 1932{ 1933 arc_buf_hdr_t *hdr = buf->b_hdr; 1934 kmutex_t *hash_lock = HDR_LOCK(hdr); 1935 boolean_t no_callback = (buf->b_efunc == NULL); 1936 1937 if (hdr->b_state == arc_anon) { 1938 ASSERT(hdr->b_datacnt == 1); 1939 arc_buf_free(buf, tag); 1940 return (no_callback); 1941 } 1942 1943 mutex_enter(hash_lock); 1944 hdr = buf->b_hdr; 1945 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 1946 ASSERT(hdr->b_state != arc_anon); 1947 ASSERT(buf->b_data != NULL); 1948 1949 (void) remove_reference(hdr, hash_lock, tag); 1950 if (hdr->b_datacnt > 1) { 1951 if (no_callback) 1952 arc_buf_destroy(buf, FALSE, TRUE); 1953 } else if (no_callback) { 1954 ASSERT(hdr->b_buf == buf && buf->b_next == NULL); 1955 ASSERT(buf->b_efunc == NULL); 1956 hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE; 1957 } 1958 ASSERT(no_callback || hdr->b_datacnt > 1 || 1959 refcount_is_zero(&hdr->b_refcnt)); 1960 mutex_exit(hash_lock); 1961 return (no_callback); 1962} 1963 1964int 1965arc_buf_size(arc_buf_t *buf) 1966{ 1967 return (buf->b_hdr->b_size); 1968} 1969 1970/* 1971 * Called from the DMU to determine if the current buffer should be 1972 * evicted. In order to ensure proper locking, the eviction must be initiated 1973 * from the DMU. Return true if the buffer is associated with user data and 1974 * duplicate buffers still exist. 1975 */ 1976boolean_t 1977arc_buf_eviction_needed(arc_buf_t *buf) 1978{ 1979 arc_buf_hdr_t *hdr; 1980 boolean_t evict_needed = B_FALSE; 1981 1982 if (zfs_disable_dup_eviction) 1983 return (B_FALSE); 1984 1985 mutex_enter(&buf->b_evict_lock); 1986 hdr = buf->b_hdr; 1987 if (hdr == NULL) { 1988 /* 1989 * We are in arc_do_user_evicts(); let that function 1990 * perform the eviction. 1991 */ 1992 ASSERT(buf->b_data == NULL); 1993 mutex_exit(&buf->b_evict_lock); 1994 return (B_FALSE); 1995 } else if (buf->b_data == NULL) { 1996 /* 1997 * We have already been added to the arc eviction list; 1998 * recommend eviction. 1999 */ 2000 ASSERT3P(hdr, ==, &arc_eviction_hdr); 2001 mutex_exit(&buf->b_evict_lock); 2002 return (B_TRUE); 2003 } 2004 2005 if (hdr->b_datacnt > 1 && hdr->b_type == ARC_BUFC_DATA) 2006 evict_needed = B_TRUE; 2007 2008 mutex_exit(&buf->b_evict_lock); 2009 return (evict_needed); 2010} 2011 2012/* 2013 * Evict buffers from list until we've removed the specified number of 2014 * bytes. Move the removed buffers to the appropriate evict state. 2015 * If the recycle flag is set, then attempt to "recycle" a buffer: 2016 * - look for a buffer to evict that is `bytes' long. 2017 * - return the data block from this buffer rather than freeing it. 2018 * This flag is used by callers that are trying to make space for a 2019 * new buffer in a full arc cache. 2020 * 2021 * This function makes a "best effort". It skips over any buffers 2022 * it can't get a hash_lock on, and so may not catch all candidates. 2023 * It may also return without evicting as much space as requested. 2024 */ 2025static void * 2026arc_evict(arc_state_t *state, uint64_t spa, int64_t bytes, boolean_t recycle, 2027 arc_buf_contents_t type) 2028{ 2029 arc_state_t *evicted_state; 2030 uint64_t bytes_evicted = 0, skipped = 0, missed = 0; 2031 int64_t bytes_remaining; 2032 arc_buf_hdr_t *hdr, *hdr_prev = NULL; 2033 list_t *evicted_list, *list, *evicted_list_start, *list_start; 2034 kmutex_t *lock, *evicted_lock; 2035 kmutex_t *hash_lock; 2036 boolean_t have_lock; 2037 void *stolen = NULL; 2038 arc_buf_hdr_t marker = { 0 }; 2039 int count = 0; 2040 static int evict_metadata_offset, evict_data_offset; 2041 int i, idx, offset, list_count, lists; 2042 2043 ASSERT(state == arc_mru || state == arc_mfu); 2044 2045 evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost; 2046 2047 /* 2048 * Decide which "type" (data vs metadata) to recycle from. 2049 * 2050 * If we are over the metadata limit, recycle from metadata. 2051 * If we are under the metadata minimum, recycle from data. 2052 * Otherwise, recycle from whichever type has the oldest (least 2053 * recently accessed) header. This is not yet implemented. 2054 */ 2055 if (recycle) { 2056 arc_buf_contents_t realtype; 2057 if (state->arcs_lsize[ARC_BUFC_DATA] == 0) { 2058 realtype = ARC_BUFC_METADATA; 2059 } else if (state->arcs_lsize[ARC_BUFC_METADATA] == 0) { 2060 realtype = ARC_BUFC_DATA; 2061 } else if (arc_meta_used >= arc_meta_limit) { 2062 realtype = ARC_BUFC_METADATA; 2063 } else if (arc_meta_used <= arc_meta_min) { 2064 realtype = ARC_BUFC_DATA; 2065 } else { 2066#ifdef illumos 2067 if (data_hdr->b_arc_access < 2068 metadata_hdr->b_arc_access) { 2069 realtype = ARC_BUFC_DATA; 2070 } else { 2071 realtype = ARC_BUFC_METADATA; 2072 } 2073#else 2074 /* TODO */ 2075 realtype = type; 2076#endif 2077 } 2078 if (realtype != type) { 2079 /* 2080 * If we want to evict from a different list, 2081 * we can not recycle, because DATA vs METADATA 2082 * buffers are segregated into different kmem 2083 * caches (and vmem arenas). 2084 */ 2085 type = realtype; 2086 recycle = B_FALSE; 2087 } 2088 } 2089 2090 if (type == ARC_BUFC_METADATA) { 2091 offset = 0; 2092 list_count = ARC_BUFC_NUMMETADATALISTS; 2093 list_start = &state->arcs_lists[0]; 2094 evicted_list_start = &evicted_state->arcs_lists[0]; 2095 idx = evict_metadata_offset; 2096 } else { 2097 offset = ARC_BUFC_NUMMETADATALISTS; 2098 list_start = &state->arcs_lists[offset]; 2099 evicted_list_start = &evicted_state->arcs_lists[offset]; 2100 list_count = ARC_BUFC_NUMDATALISTS; 2101 idx = evict_data_offset; 2102 } 2103 bytes_remaining = evicted_state->arcs_lsize[type]; 2104 lists = 0; 2105 2106evict_start: 2107 list = &list_start[idx]; 2108 evicted_list = &evicted_list_start[idx]; 2109 lock = ARCS_LOCK(state, (offset + idx)); 2110 evicted_lock = ARCS_LOCK(evicted_state, (offset + idx)); 2111 2112 mutex_enter(lock); 2113 mutex_enter(evicted_lock); 2114 2115 for (hdr = list_tail(list); hdr; hdr = hdr_prev) { 2116 hdr_prev = list_prev(list, hdr); 2117 bytes_remaining -= (hdr->b_size * hdr->b_datacnt); 2118 /* prefetch buffers have a minimum lifespan */ 2119 if (HDR_IO_IN_PROGRESS(hdr) || 2120 (spa && hdr->b_spa != spa) || 2121 (hdr->b_flags & (ARC_FLAG_PREFETCH | ARC_FLAG_INDIRECT) && 2122 ddi_get_lbolt() - hdr->b_arc_access < 2123 arc_min_prefetch_lifespan)) { 2124 skipped++; 2125 continue; 2126 } 2127 /* "lookahead" for better eviction candidate */ 2128 if (recycle && hdr->b_size != bytes && 2129 hdr_prev && hdr_prev->b_size == bytes) 2130 continue; 2131 2132 /* ignore markers */ 2133 if (hdr->b_spa == 0) 2134 continue; 2135 2136 /* 2137 * It may take a long time to evict all the bufs requested. 2138 * To avoid blocking all arc activity, periodically drop 2139 * the arcs_mtx and give other threads a chance to run 2140 * before reacquiring the lock. 2141 * 2142 * If we are looking for a buffer to recycle, we are in 2143 * the hot code path, so don't sleep. 2144 */ 2145 if (!recycle && count++ > arc_evict_iterations) { 2146 list_insert_after(list, hdr, &marker); 2147 mutex_exit(evicted_lock); 2148 mutex_exit(lock); 2149 kpreempt(KPREEMPT_SYNC); 2150 mutex_enter(lock); 2151 mutex_enter(evicted_lock); 2152 hdr_prev = list_prev(list, &marker); 2153 list_remove(list, &marker); 2154 count = 0; 2155 continue; 2156 } 2157 2158 hash_lock = HDR_LOCK(hdr); 2159 have_lock = MUTEX_HELD(hash_lock); 2160 if (have_lock || mutex_tryenter(hash_lock)) { 2161 ASSERT0(refcount_count(&hdr->b_refcnt)); 2162 ASSERT(hdr->b_datacnt > 0); 2163 while (hdr->b_buf) { 2164 arc_buf_t *buf = hdr->b_buf; 2165 if (!mutex_tryenter(&buf->b_evict_lock)) { 2166 missed += 1; 2167 break; 2168 } 2169 if (buf->b_data) { 2170 bytes_evicted += hdr->b_size; 2171 if (recycle && hdr->b_type == type && 2172 hdr->b_size == bytes && 2173 !HDR_L2_WRITING(hdr)) { 2174 stolen = buf->b_data; 2175 recycle = FALSE; 2176 } 2177 } 2178 if (buf->b_efunc) { 2179 mutex_enter(&arc_eviction_mtx); 2180 arc_buf_destroy(buf, 2181 buf->b_data == stolen, FALSE); 2182 hdr->b_buf = buf->b_next; 2183 buf->b_hdr = &arc_eviction_hdr; 2184 buf->b_next = arc_eviction_list; 2185 arc_eviction_list = buf; 2186 mutex_exit(&arc_eviction_mtx); 2187 mutex_exit(&buf->b_evict_lock); 2188 } else { 2189 mutex_exit(&buf->b_evict_lock); 2190 arc_buf_destroy(buf, 2191 buf->b_data == stolen, TRUE); 2192 } 2193 } 2194 2195 if (hdr->b_l2hdr) { 2196 ARCSTAT_INCR(arcstat_evict_l2_cached, 2197 hdr->b_size); 2198 } else { 2199 if (l2arc_write_eligible(hdr->b_spa, hdr)) { 2200 ARCSTAT_INCR(arcstat_evict_l2_eligible, 2201 hdr->b_size); 2202 } else { 2203 ARCSTAT_INCR( 2204 arcstat_evict_l2_ineligible, 2205 hdr->b_size); 2206 } 2207 } 2208 2209 if (hdr->b_datacnt == 0) { 2210 arc_change_state(evicted_state, hdr, hash_lock); 2211 ASSERT(HDR_IN_HASH_TABLE(hdr)); 2212 hdr->b_flags |= ARC_FLAG_IN_HASH_TABLE; 2213 hdr->b_flags &= ~ARC_FLAG_BUF_AVAILABLE; 2214 DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, hdr); 2215 } 2216 if (!have_lock) 2217 mutex_exit(hash_lock); 2218 if (bytes >= 0 && bytes_evicted >= bytes) 2219 break; 2220 if (bytes_remaining > 0) { 2221 mutex_exit(evicted_lock); 2222 mutex_exit(lock); 2223 idx = ((idx + 1) & (list_count - 1)); 2224 lists++; 2225 goto evict_start; 2226 } 2227 } else { 2228 missed += 1; 2229 } 2230 } 2231 2232 mutex_exit(evicted_lock); 2233 mutex_exit(lock); 2234 2235 idx = ((idx + 1) & (list_count - 1)); 2236 lists++; 2237 2238 if (bytes_evicted < bytes) { 2239 if (lists < list_count) 2240 goto evict_start; 2241 else 2242 dprintf("only evicted %lld bytes from %x", 2243 (longlong_t)bytes_evicted, state); 2244 } 2245 if (type == ARC_BUFC_METADATA) 2246 evict_metadata_offset = idx; 2247 else 2248 evict_data_offset = idx; 2249 2250 if (skipped) 2251 ARCSTAT_INCR(arcstat_evict_skip, skipped); 2252 2253 if (missed) 2254 ARCSTAT_INCR(arcstat_mutex_miss, missed); 2255 2256 /* 2257 * Note: we have just evicted some data into the ghost state, 2258 * potentially putting the ghost size over the desired size. Rather 2259 * that evicting from the ghost list in this hot code path, leave 2260 * this chore to the arc_reclaim_thread(). 2261 */ 2262 2263 if (stolen) 2264 ARCSTAT_BUMP(arcstat_stolen); 2265 return (stolen); 2266} 2267 2268/* 2269 * Remove buffers from list until we've removed the specified number of 2270 * bytes. Destroy the buffers that are removed. 2271 */ 2272static void 2273arc_evict_ghost(arc_state_t *state, uint64_t spa, int64_t bytes) 2274{ 2275 arc_buf_hdr_t *hdr, *hdr_prev; 2276 arc_buf_hdr_t marker = { 0 }; 2277 list_t *list, *list_start; 2278 kmutex_t *hash_lock, *lock; 2279 uint64_t bytes_deleted = 0; 2280 uint64_t bufs_skipped = 0; 2281 int count = 0; 2282 static int evict_offset; 2283 int list_count, idx = evict_offset; 2284 int offset, lists = 0; 2285 2286 ASSERT(GHOST_STATE(state)); 2287 2288 /* 2289 * data lists come after metadata lists 2290 */ 2291 list_start = &state->arcs_lists[ARC_BUFC_NUMMETADATALISTS]; 2292 list_count = ARC_BUFC_NUMDATALISTS; 2293 offset = ARC_BUFC_NUMMETADATALISTS; 2294 2295evict_start: 2296 list = &list_start[idx]; 2297 lock = ARCS_LOCK(state, idx + offset); 2298 2299 mutex_enter(lock); 2300 for (hdr = list_tail(list); hdr; hdr = hdr_prev) { 2301 hdr_prev = list_prev(list, hdr); 2302 if (hdr->b_type > ARC_BUFC_NUMTYPES) 2303 panic("invalid hdr=%p", (void *)hdr); 2304 if (spa && hdr->b_spa != spa) 2305 continue; 2306 2307 /* ignore markers */ 2308 if (hdr->b_spa == 0) 2309 continue; 2310 2311 hash_lock = HDR_LOCK(hdr); 2312 /* caller may be trying to modify this buffer, skip it */ 2313 if (MUTEX_HELD(hash_lock)) 2314 continue; 2315 2316 /* 2317 * It may take a long time to evict all the bufs requested. 2318 * To avoid blocking all arc activity, periodically drop 2319 * the arcs_mtx and give other threads a chance to run 2320 * before reacquiring the lock. 2321 */ 2322 if (count++ > arc_evict_iterations) { 2323 list_insert_after(list, hdr, &marker); 2324 mutex_exit(lock); 2325 kpreempt(KPREEMPT_SYNC); 2326 mutex_enter(lock); 2327 hdr_prev = list_prev(list, &marker); 2328 list_remove(list, &marker); 2329 count = 0; 2330 continue; 2331 } 2332 if (mutex_tryenter(hash_lock)) { 2333 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 2334 ASSERT(hdr->b_buf == NULL); 2335 ARCSTAT_BUMP(arcstat_deleted); 2336 bytes_deleted += hdr->b_size; 2337 2338 if (hdr->b_l2hdr != NULL) { 2339 /* 2340 * This buffer is cached on the 2nd Level ARC; 2341 * don't destroy the header. 2342 */ 2343 arc_change_state(arc_l2c_only, hdr, hash_lock); 2344 mutex_exit(hash_lock); 2345 } else { 2346 arc_change_state(arc_anon, hdr, hash_lock); 2347 mutex_exit(hash_lock); 2348 arc_hdr_destroy(hdr); 2349 } 2350 2351 DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, hdr); 2352 if (bytes >= 0 && bytes_deleted >= bytes) 2353 break; 2354 } else if (bytes < 0) { 2355 /* 2356 * Insert a list marker and then wait for the 2357 * hash lock to become available. Once its 2358 * available, restart from where we left off. 2359 */ 2360 list_insert_after(list, hdr, &marker); 2361 mutex_exit(lock); 2362 mutex_enter(hash_lock); 2363 mutex_exit(hash_lock); 2364 mutex_enter(lock); 2365 hdr_prev = list_prev(list, &marker); 2366 list_remove(list, &marker); 2367 } else { 2368 bufs_skipped += 1; 2369 } 2370 2371 } 2372 mutex_exit(lock); 2373 idx = ((idx + 1) & (ARC_BUFC_NUMDATALISTS - 1)); 2374 lists++; 2375 2376 if (lists < list_count) 2377 goto evict_start; 2378 2379 evict_offset = idx; 2380 if ((uintptr_t)list > (uintptr_t)&state->arcs_lists[ARC_BUFC_NUMMETADATALISTS] && 2381 (bytes < 0 || bytes_deleted < bytes)) { 2382 list_start = &state->arcs_lists[0]; 2383 list_count = ARC_BUFC_NUMMETADATALISTS; 2384 offset = lists = 0; 2385 goto evict_start; 2386 } 2387 2388 if (bufs_skipped) { 2389 ARCSTAT_INCR(arcstat_mutex_miss, bufs_skipped); 2390 ASSERT(bytes >= 0); 2391 } 2392 2393 if (bytes_deleted < bytes) 2394 dprintf("only deleted %lld bytes from %p", 2395 (longlong_t)bytes_deleted, state); 2396} 2397 2398static void 2399arc_adjust(void) 2400{ 2401 int64_t adjustment, delta; 2402 2403 /* 2404 * Adjust MRU size 2405 */ 2406 2407 adjustment = MIN((int64_t)(arc_size - arc_c), 2408 (int64_t)(arc_anon->arcs_size + arc_mru->arcs_size + arc_meta_used - 2409 arc_p)); 2410 2411 if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_DATA] > 0) { 2412 delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_DATA], adjustment); 2413 (void) arc_evict(arc_mru, 0, delta, FALSE, ARC_BUFC_DATA); 2414 adjustment -= delta; 2415 } 2416 2417 if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_METADATA] > 0) { 2418 delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_METADATA], adjustment); 2419 (void) arc_evict(arc_mru, 0, delta, FALSE, 2420 ARC_BUFC_METADATA); 2421 } 2422 2423 /* 2424 * Adjust MFU size 2425 */ 2426 2427 adjustment = arc_size - arc_c; 2428 2429 if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_DATA] > 0) { 2430 delta = MIN(adjustment, arc_mfu->arcs_lsize[ARC_BUFC_DATA]); 2431 (void) arc_evict(arc_mfu, 0, delta, FALSE, ARC_BUFC_DATA); 2432 adjustment -= delta; 2433 } 2434 2435 if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_METADATA] > 0) { 2436 int64_t delta = MIN(adjustment, 2437 arc_mfu->arcs_lsize[ARC_BUFC_METADATA]); 2438 (void) arc_evict(arc_mfu, 0, delta, FALSE, 2439 ARC_BUFC_METADATA); 2440 } 2441 2442 /* 2443 * Adjust ghost lists 2444 */ 2445 2446 adjustment = arc_mru->arcs_size + arc_mru_ghost->arcs_size - arc_c; 2447 2448 if (adjustment > 0 && arc_mru_ghost->arcs_size > 0) { 2449 delta = MIN(arc_mru_ghost->arcs_size, adjustment); 2450 arc_evict_ghost(arc_mru_ghost, 0, delta); 2451 } 2452 2453 adjustment = 2454 arc_mru_ghost->arcs_size + arc_mfu_ghost->arcs_size - arc_c; 2455 2456 if (adjustment > 0 && arc_mfu_ghost->arcs_size > 0) { 2457 delta = MIN(arc_mfu_ghost->arcs_size, adjustment); 2458 arc_evict_ghost(arc_mfu_ghost, 0, delta); 2459 } 2460} 2461 2462static void 2463arc_do_user_evicts(void) 2464{ 2465 static arc_buf_t *tmp_arc_eviction_list; 2466 2467 /* 2468 * Move list over to avoid LOR 2469 */ 2470restart: 2471 mutex_enter(&arc_eviction_mtx); 2472 tmp_arc_eviction_list = arc_eviction_list; 2473 arc_eviction_list = NULL; 2474 mutex_exit(&arc_eviction_mtx); 2475 2476 while (tmp_arc_eviction_list != NULL) { 2477 arc_buf_t *buf = tmp_arc_eviction_list; 2478 tmp_arc_eviction_list = buf->b_next; 2479 mutex_enter(&buf->b_evict_lock); 2480 buf->b_hdr = NULL; 2481 mutex_exit(&buf->b_evict_lock); 2482 2483 if (buf->b_efunc != NULL) 2484 VERIFY0(buf->b_efunc(buf->b_private)); 2485 2486 buf->b_efunc = NULL; 2487 buf->b_private = NULL; 2488 kmem_cache_free(buf_cache, buf); 2489 } 2490 2491 if (arc_eviction_list != NULL) 2492 goto restart; 2493} 2494 2495/* 2496 * Flush all *evictable* data from the cache for the given spa. 2497 * NOTE: this will not touch "active" (i.e. referenced) data. 2498 */ 2499void 2500arc_flush(spa_t *spa) 2501{ 2502 uint64_t guid = 0; 2503 2504 if (spa) 2505 guid = spa_load_guid(spa); 2506 2507 while (arc_mru->arcs_lsize[ARC_BUFC_DATA]) { 2508 (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_DATA); 2509 if (spa) 2510 break; 2511 } 2512 while (arc_mru->arcs_lsize[ARC_BUFC_METADATA]) { 2513 (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_METADATA); 2514 if (spa) 2515 break; 2516 } 2517 while (arc_mfu->arcs_lsize[ARC_BUFC_DATA]) { 2518 (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_DATA); 2519 if (spa) 2520 break; 2521 } 2522 while (arc_mfu->arcs_lsize[ARC_BUFC_METADATA]) { 2523 (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_METADATA); 2524 if (spa) 2525 break; 2526 } 2527 2528 arc_evict_ghost(arc_mru_ghost, guid, -1); 2529 arc_evict_ghost(arc_mfu_ghost, guid, -1); 2530 2531 mutex_enter(&arc_reclaim_thr_lock); 2532 arc_do_user_evicts(); 2533 mutex_exit(&arc_reclaim_thr_lock); 2534 ASSERT(spa || arc_eviction_list == NULL); 2535} 2536 2537void 2538arc_shrink(void) 2539{ 2540 2541 if (arc_c > arc_c_min) { 2542 uint64_t to_free; 2543 2544 to_free = arc_c >> arc_shrink_shift; 2545 DTRACE_PROBE4(arc__shrink, uint64_t, arc_c, uint64_t, 2546 arc_c_min, uint64_t, arc_p, uint64_t, to_free); 2547 if (arc_c > arc_c_min + to_free) 2548 atomic_add_64(&arc_c, -to_free); 2549 else 2550 arc_c = arc_c_min; 2551 2552 atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift)); 2553 if (arc_c > arc_size) 2554 arc_c = MAX(arc_size, arc_c_min); 2555 if (arc_p > arc_c) 2556 arc_p = (arc_c >> 1); 2557 2558 DTRACE_PROBE2(arc__shrunk, uint64_t, arc_c, uint64_t, 2559 arc_p); 2560 2561 ASSERT(arc_c >= arc_c_min); 2562 ASSERT((int64_t)arc_p >= 0); 2563 } 2564 2565 if (arc_size > arc_c) { 2566 DTRACE_PROBE2(arc__shrink_adjust, uint64_t, arc_size, 2567 uint64_t, arc_c); 2568 arc_adjust(); 2569 } 2570} 2571 2572static int needfree = 0; 2573 2574static int 2575arc_reclaim_needed(void) 2576{ 2577 2578#ifdef _KERNEL 2579 2580 if (needfree) { 2581 DTRACE_PROBE(arc__reclaim_needfree); 2582 return (1); 2583 } 2584 2585 /* 2586 * Cooperate with pagedaemon when it's time for it to scan 2587 * and reclaim some pages. 2588 */ 2589 if (freemem < zfs_arc_free_target) { 2590 DTRACE_PROBE2(arc__reclaim_freemem, uint64_t, 2591 freemem, uint64_t, zfs_arc_free_target); 2592 return (1); 2593 } 2594 2595#ifdef illumos 2596 /* 2597 * take 'desfree' extra pages, so we reclaim sooner, rather than later 2598 */ 2599 extra = desfree; 2600 2601 /* 2602 * check that we're out of range of the pageout scanner. It starts to 2603 * schedule paging if freemem is less than lotsfree and needfree. 2604 * lotsfree is the high-water mark for pageout, and needfree is the 2605 * number of needed free pages. We add extra pages here to make sure 2606 * the scanner doesn't start up while we're freeing memory. 2607 */ 2608 if (freemem < lotsfree + needfree + extra) 2609 return (1); 2610 2611 /* 2612 * check to make sure that swapfs has enough space so that anon 2613 * reservations can still succeed. anon_resvmem() checks that the 2614 * availrmem is greater than swapfs_minfree, and the number of reserved 2615 * swap pages. We also add a bit of extra here just to prevent 2616 * circumstances from getting really dire. 2617 */ 2618 if (availrmem < swapfs_minfree + swapfs_reserve + extra) 2619 return (1); 2620 2621 /* 2622 * Check that we have enough availrmem that memory locking (e.g., via 2623 * mlock(3C) or memcntl(2)) can still succeed. (pages_pp_maximum 2624 * stores the number of pages that cannot be locked; when availrmem 2625 * drops below pages_pp_maximum, page locking mechanisms such as 2626 * page_pp_lock() will fail.) 2627 */ 2628 if (availrmem <= pages_pp_maximum) 2629 return (1); 2630 2631#endif /* illumos */ 2632#if defined(__i386) || !defined(UMA_MD_SMALL_ALLOC) 2633 /* 2634 * If we're on an i386 platform, it's possible that we'll exhaust the 2635 * kernel heap space before we ever run out of available physical 2636 * memory. Most checks of the size of the heap_area compare against 2637 * tune.t_minarmem, which is the minimum available real memory that we 2638 * can have in the system. However, this is generally fixed at 25 pages 2639 * which is so low that it's useless. In this comparison, we seek to 2640 * calculate the total heap-size, and reclaim if more than 3/4ths of the 2641 * heap is allocated. (Or, in the calculation, if less than 1/4th is 2642 * free) 2643 */ 2644 if (vmem_size(heap_arena, VMEM_FREE) < 2645 (vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC) >> 2)) { 2646 DTRACE_PROBE2(arc__reclaim_used, uint64_t, 2647 vmem_size(heap_arena, VMEM_FREE), uint64_t, 2648 (vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC)) >> 2); 2649 return (1); 2650 } 2651#endif 2652#ifdef illumos 2653 /* 2654 * If zio data pages are being allocated out of a separate heap segment, 2655 * then enforce that the size of available vmem for this arena remains 2656 * above about 1/16th free. 2657 * 2658 * Note: The 1/16th arena free requirement was put in place 2659 * to aggressively evict memory from the arc in order to avoid 2660 * memory fragmentation issues. 2661 */ 2662 if (zio_arena != NULL && 2663 vmem_size(zio_arena, VMEM_FREE) < 2664 (vmem_size(zio_arena, VMEM_ALLOC) >> 4)) 2665 return (1); 2666#endif /* illumos */ 2667#else /* _KERNEL */ 2668 if (spa_get_random(100) == 0) 2669 return (1); 2670#endif /* _KERNEL */ 2671 DTRACE_PROBE(arc__reclaim_no); 2672 2673 return (0); 2674} 2675 2676extern kmem_cache_t *zio_buf_cache[]; 2677extern kmem_cache_t *zio_data_buf_cache[]; 2678extern kmem_cache_t *range_seg_cache; 2679 2680static void __noinline 2681arc_kmem_reap_now(arc_reclaim_strategy_t strat) 2682{ 2683 size_t i; 2684 kmem_cache_t *prev_cache = NULL; 2685 kmem_cache_t *prev_data_cache = NULL; 2686 2687 DTRACE_PROBE(arc__kmem_reap_start); 2688#ifdef _KERNEL 2689 if (arc_meta_used >= arc_meta_limit) { 2690 /* 2691 * We are exceeding our meta-data cache limit. 2692 * Purge some DNLC entries to release holds on meta-data. 2693 */ 2694 dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent); 2695 } 2696#if defined(__i386) 2697 /* 2698 * Reclaim unused memory from all kmem caches. 2699 */ 2700 kmem_reap(); 2701#endif 2702#endif 2703 2704 /* 2705 * An aggressive reclamation will shrink the cache size as well as 2706 * reap free buffers from the arc kmem caches. 2707 */ 2708 if (strat == ARC_RECLAIM_AGGR) 2709 arc_shrink(); 2710 2711 for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) { 2712 if (zio_buf_cache[i] != prev_cache) { 2713 prev_cache = zio_buf_cache[i]; 2714 kmem_cache_reap_now(zio_buf_cache[i]); 2715 } 2716 if (zio_data_buf_cache[i] != prev_data_cache) { 2717 prev_data_cache = zio_data_buf_cache[i]; 2718 kmem_cache_reap_now(zio_data_buf_cache[i]); 2719 } 2720 } 2721 kmem_cache_reap_now(buf_cache); 2722 kmem_cache_reap_now(hdr_cache); 2723 kmem_cache_reap_now(range_seg_cache); 2724 2725#ifdef illumos 2726 /* 2727 * Ask the vmem arena to reclaim unused memory from its 2728 * quantum caches. 2729 */ 2730 if (zio_arena != NULL && strat == ARC_RECLAIM_AGGR) 2731 vmem_qcache_reap(zio_arena); 2732#endif 2733 DTRACE_PROBE(arc__kmem_reap_end); 2734} 2735 2736static void 2737arc_reclaim_thread(void *dummy __unused) 2738{ 2739 clock_t growtime = 0; 2740 arc_reclaim_strategy_t last_reclaim = ARC_RECLAIM_CONS; 2741 callb_cpr_t cpr; 2742 2743 CALLB_CPR_INIT(&cpr, &arc_reclaim_thr_lock, callb_generic_cpr, FTAG); 2744 2745 mutex_enter(&arc_reclaim_thr_lock); 2746 while (arc_thread_exit == 0) { 2747 if (arc_reclaim_needed()) { 2748 2749 if (arc_no_grow) { 2750 if (last_reclaim == ARC_RECLAIM_CONS) { 2751 DTRACE_PROBE(arc__reclaim_aggr_no_grow); 2752 last_reclaim = ARC_RECLAIM_AGGR; 2753 } else { 2754 last_reclaim = ARC_RECLAIM_CONS; 2755 } 2756 } else { 2757 arc_no_grow = TRUE; 2758 last_reclaim = ARC_RECLAIM_AGGR; 2759 DTRACE_PROBE(arc__reclaim_aggr); 2760 membar_producer(); 2761 } 2762 2763 /* reset the growth delay for every reclaim */ 2764 growtime = ddi_get_lbolt() + (arc_grow_retry * hz); 2765 2766 if (needfree && last_reclaim == ARC_RECLAIM_CONS) { 2767 /* 2768 * If needfree is TRUE our vm_lowmem hook 2769 * was called and in that case we must free some 2770 * memory, so switch to aggressive mode. 2771 */ 2772 arc_no_grow = TRUE; 2773 last_reclaim = ARC_RECLAIM_AGGR; 2774 } 2775 arc_kmem_reap_now(last_reclaim); 2776 arc_warm = B_TRUE; 2777 2778 } else if (arc_no_grow && ddi_get_lbolt() >= growtime) { 2779 arc_no_grow = FALSE; 2780 } 2781 2782 arc_adjust(); 2783 2784 if (arc_eviction_list != NULL) 2785 arc_do_user_evicts(); 2786 2787#ifdef _KERNEL 2788 if (needfree) { 2789 needfree = 0; 2790 wakeup(&needfree); 2791 } 2792#endif 2793 2794 /* block until needed, or one second, whichever is shorter */ 2795 CALLB_CPR_SAFE_BEGIN(&cpr); 2796 (void) cv_timedwait(&arc_reclaim_thr_cv, 2797 &arc_reclaim_thr_lock, hz); 2798 CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_thr_lock); 2799 } 2800 2801 arc_thread_exit = 0; 2802 cv_broadcast(&arc_reclaim_thr_cv); 2803 CALLB_CPR_EXIT(&cpr); /* drops arc_reclaim_thr_lock */ 2804 thread_exit(); 2805} 2806 2807/* 2808 * Adapt arc info given the number of bytes we are trying to add and 2809 * the state that we are comming from. This function is only called 2810 * when we are adding new content to the cache. 2811 */ 2812static void 2813arc_adapt(int bytes, arc_state_t *state) 2814{ 2815 int mult; 2816 uint64_t arc_p_min = (arc_c >> arc_p_min_shift); 2817 2818 if (state == arc_l2c_only) 2819 return; 2820 2821 ASSERT(bytes > 0); 2822 /* 2823 * Adapt the target size of the MRU list: 2824 * - if we just hit in the MRU ghost list, then increase 2825 * the target size of the MRU list. 2826 * - if we just hit in the MFU ghost list, then increase 2827 * the target size of the MFU list by decreasing the 2828 * target size of the MRU list. 2829 */ 2830 if (state == arc_mru_ghost) { 2831 mult = ((arc_mru_ghost->arcs_size >= arc_mfu_ghost->arcs_size) ? 2832 1 : (arc_mfu_ghost->arcs_size/arc_mru_ghost->arcs_size)); 2833 mult = MIN(mult, 10); /* avoid wild arc_p adjustment */ 2834 2835 arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult); 2836 } else if (state == arc_mfu_ghost) { 2837 uint64_t delta; 2838 2839 mult = ((arc_mfu_ghost->arcs_size >= arc_mru_ghost->arcs_size) ? 2840 1 : (arc_mru_ghost->arcs_size/arc_mfu_ghost->arcs_size)); 2841 mult = MIN(mult, 10); 2842 2843 delta = MIN(bytes * mult, arc_p); 2844 arc_p = MAX(arc_p_min, arc_p - delta); 2845 } 2846 ASSERT((int64_t)arc_p >= 0); 2847 2848 if (arc_reclaim_needed()) { 2849 cv_signal(&arc_reclaim_thr_cv); 2850 return; 2851 } 2852 2853 if (arc_no_grow) 2854 return; 2855 2856 if (arc_c >= arc_c_max) 2857 return; 2858 2859 /* 2860 * If we're within (2 * maxblocksize) bytes of the target 2861 * cache size, increment the target cache size 2862 */ 2863 if (arc_size > arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) { 2864 DTRACE_PROBE1(arc__inc_adapt, int, bytes); 2865 atomic_add_64(&arc_c, (int64_t)bytes); 2866 if (arc_c > arc_c_max) 2867 arc_c = arc_c_max; 2868 else if (state == arc_anon) 2869 atomic_add_64(&arc_p, (int64_t)bytes); 2870 if (arc_p > arc_c) 2871 arc_p = arc_c; 2872 } 2873 ASSERT((int64_t)arc_p >= 0); 2874} 2875 2876/* 2877 * Check if the cache has reached its limits and eviction is required 2878 * prior to insert. 2879 */ 2880static int 2881arc_evict_needed(arc_buf_contents_t type) 2882{ 2883 if (type == ARC_BUFC_METADATA && arc_meta_used >= arc_meta_limit) 2884 return (1); 2885 2886 if (arc_reclaim_needed()) 2887 return (1); 2888 2889 return (arc_size > arc_c); 2890} 2891 2892/* 2893 * The buffer, supplied as the first argument, needs a data block. 2894 * So, if we are at cache max, determine which cache should be victimized. 2895 * We have the following cases: 2896 * 2897 * 1. Insert for MRU, p > sizeof(arc_anon + arc_mru) -> 2898 * In this situation if we're out of space, but the resident size of the MFU is 2899 * under the limit, victimize the MFU cache to satisfy this insertion request. 2900 * 2901 * 2. Insert for MRU, p <= sizeof(arc_anon + arc_mru) -> 2902 * Here, we've used up all of the available space for the MRU, so we need to 2903 * evict from our own cache instead. Evict from the set of resident MRU 2904 * entries. 2905 * 2906 * 3. Insert for MFU (c - p) > sizeof(arc_mfu) -> 2907 * c minus p represents the MFU space in the cache, since p is the size of the 2908 * cache that is dedicated to the MRU. In this situation there's still space on 2909 * the MFU side, so the MRU side needs to be victimized. 2910 * 2911 * 4. Insert for MFU (c - p) < sizeof(arc_mfu) -> 2912 * MFU's resident set is consuming more space than it has been allotted. In 2913 * this situation, we must victimize our own cache, the MFU, for this insertion. 2914 */ 2915static void 2916arc_get_data_buf(arc_buf_t *buf) 2917{ 2918 arc_state_t *state = buf->b_hdr->b_state; 2919 uint64_t size = buf->b_hdr->b_size; 2920 arc_buf_contents_t type = buf->b_hdr->b_type; 2921 2922 arc_adapt(size, state); 2923 2924 /* 2925 * We have not yet reached cache maximum size, 2926 * just allocate a new buffer. 2927 */ 2928 if (!arc_evict_needed(type)) { 2929 if (type == ARC_BUFC_METADATA) { 2930 buf->b_data = zio_buf_alloc(size); 2931 arc_space_consume(size, ARC_SPACE_DATA); 2932 } else { 2933 ASSERT(type == ARC_BUFC_DATA); 2934 buf->b_data = zio_data_buf_alloc(size); 2935 ARCSTAT_INCR(arcstat_data_size, size); 2936 atomic_add_64(&arc_size, size); 2937 } 2938 goto out; 2939 } 2940 2941 /* 2942 * If we are prefetching from the mfu ghost list, this buffer 2943 * will end up on the mru list; so steal space from there. 2944 */ 2945 if (state == arc_mfu_ghost) 2946 state = buf->b_hdr->b_flags & ARC_FLAG_PREFETCH ? 2947 arc_mru : arc_mfu; 2948 else if (state == arc_mru_ghost) 2949 state = arc_mru; 2950 2951 if (state == arc_mru || state == arc_anon) { 2952 uint64_t mru_used = arc_anon->arcs_size + arc_mru->arcs_size; 2953 state = (arc_mfu->arcs_lsize[type] >= size && 2954 arc_p > mru_used) ? arc_mfu : arc_mru; 2955 } else { 2956 /* MFU cases */ 2957 uint64_t mfu_space = arc_c - arc_p; 2958 state = (arc_mru->arcs_lsize[type] >= size && 2959 mfu_space > arc_mfu->arcs_size) ? arc_mru : arc_mfu; 2960 } 2961 if ((buf->b_data = arc_evict(state, 0, size, TRUE, type)) == NULL) { 2962 if (type == ARC_BUFC_METADATA) { 2963 buf->b_data = zio_buf_alloc(size); 2964 arc_space_consume(size, ARC_SPACE_DATA); 2965 } else { 2966 ASSERT(type == ARC_BUFC_DATA); 2967 buf->b_data = zio_data_buf_alloc(size); 2968 ARCSTAT_INCR(arcstat_data_size, size); 2969 atomic_add_64(&arc_size, size); 2970 } 2971 ARCSTAT_BUMP(arcstat_recycle_miss); 2972 } 2973 ASSERT(buf->b_data != NULL); 2974out: 2975 /* 2976 * Update the state size. Note that ghost states have a 2977 * "ghost size" and so don't need to be updated. 2978 */ 2979 if (!GHOST_STATE(buf->b_hdr->b_state)) { 2980 arc_buf_hdr_t *hdr = buf->b_hdr; 2981 2982 atomic_add_64(&hdr->b_state->arcs_size, size); 2983 if (list_link_active(&hdr->b_arc_node)) { 2984 ASSERT(refcount_is_zero(&hdr->b_refcnt)); 2985 atomic_add_64(&hdr->b_state->arcs_lsize[type], size); 2986 } 2987 /* 2988 * If we are growing the cache, and we are adding anonymous 2989 * data, and we have outgrown arc_p, update arc_p 2990 */ 2991 if (arc_size < arc_c && hdr->b_state == arc_anon && 2992 arc_anon->arcs_size + arc_mru->arcs_size > arc_p) 2993 arc_p = MIN(arc_c, arc_p + size); 2994 } 2995 ARCSTAT_BUMP(arcstat_allocated); 2996} 2997 2998/* 2999 * This routine is called whenever a buffer is accessed. 3000 * NOTE: the hash lock is dropped in this function. 3001 */ 3002static void 3003arc_access(arc_buf_hdr_t *hdr, kmutex_t *hash_lock) 3004{ 3005 clock_t now; 3006 3007 ASSERT(MUTEX_HELD(hash_lock)); 3008 3009 if (hdr->b_state == arc_anon) { 3010 /* 3011 * This buffer is not in the cache, and does not 3012 * appear in our "ghost" list. Add the new buffer 3013 * to the MRU state. 3014 */ 3015 3016 ASSERT(hdr->b_arc_access == 0); 3017 hdr->b_arc_access = ddi_get_lbolt(); 3018 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr); 3019 arc_change_state(arc_mru, hdr, hash_lock); 3020 3021 } else if (hdr->b_state == arc_mru) { 3022 now = ddi_get_lbolt(); 3023 3024 /* 3025 * If this buffer is here because of a prefetch, then either: 3026 * - clear the flag if this is a "referencing" read 3027 * (any subsequent access will bump this into the MFU state). 3028 * or 3029 * - move the buffer to the head of the list if this is 3030 * another prefetch (to make it less likely to be evicted). 3031 */ 3032 if ((hdr->b_flags & ARC_FLAG_PREFETCH) != 0) { 3033 if (refcount_count(&hdr->b_refcnt) == 0) { 3034 ASSERT(list_link_active(&hdr->b_arc_node)); 3035 } else { 3036 hdr->b_flags &= ~ARC_FLAG_PREFETCH; 3037 ARCSTAT_BUMP(arcstat_mru_hits); 3038 } 3039 hdr->b_arc_access = now; 3040 return; 3041 } 3042 3043 /* 3044 * This buffer has been "accessed" only once so far, 3045 * but it is still in the cache. Move it to the MFU 3046 * state. 3047 */ 3048 if (now > hdr->b_arc_access + ARC_MINTIME) { 3049 /* 3050 * More than 125ms have passed since we 3051 * instantiated this buffer. Move it to the 3052 * most frequently used state. 3053 */ 3054 hdr->b_arc_access = now; 3055 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 3056 arc_change_state(arc_mfu, hdr, hash_lock); 3057 } 3058 ARCSTAT_BUMP(arcstat_mru_hits); 3059 } else if (hdr->b_state == arc_mru_ghost) { 3060 arc_state_t *new_state; 3061 /* 3062 * This buffer has been "accessed" recently, but 3063 * was evicted from the cache. Move it to the 3064 * MFU state. 3065 */ 3066 3067 if (hdr->b_flags & ARC_FLAG_PREFETCH) { 3068 new_state = arc_mru; 3069 if (refcount_count(&hdr->b_refcnt) > 0) 3070 hdr->b_flags &= ~ARC_FLAG_PREFETCH; 3071 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr); 3072 } else { 3073 new_state = arc_mfu; 3074 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 3075 } 3076 3077 hdr->b_arc_access = ddi_get_lbolt(); 3078 arc_change_state(new_state, hdr, hash_lock); 3079 3080 ARCSTAT_BUMP(arcstat_mru_ghost_hits); 3081 } else if (hdr->b_state == arc_mfu) { 3082 /* 3083 * This buffer has been accessed more than once and is 3084 * still in the cache. Keep it in the MFU state. 3085 * 3086 * NOTE: an add_reference() that occurred when we did 3087 * the arc_read() will have kicked this off the list. 3088 * If it was a prefetch, we will explicitly move it to 3089 * the head of the list now. 3090 */ 3091 if ((hdr->b_flags & ARC_FLAG_PREFETCH) != 0) { 3092 ASSERT(refcount_count(&hdr->b_refcnt) == 0); 3093 ASSERT(list_link_active(&hdr->b_arc_node)); 3094 } 3095 ARCSTAT_BUMP(arcstat_mfu_hits); 3096 hdr->b_arc_access = ddi_get_lbolt(); 3097 } else if (hdr->b_state == arc_mfu_ghost) { 3098 arc_state_t *new_state = arc_mfu; 3099 /* 3100 * This buffer has been accessed more than once but has 3101 * been evicted from the cache. Move it back to the 3102 * MFU state. 3103 */ 3104 3105 if (hdr->b_flags & ARC_FLAG_PREFETCH) { 3106 /* 3107 * This is a prefetch access... 3108 * move this block back to the MRU state. 3109 */ 3110 ASSERT0(refcount_count(&hdr->b_refcnt)); 3111 new_state = arc_mru; 3112 } 3113 3114 hdr->b_arc_access = ddi_get_lbolt(); 3115 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 3116 arc_change_state(new_state, hdr, hash_lock); 3117 3118 ARCSTAT_BUMP(arcstat_mfu_ghost_hits); 3119 } else if (hdr->b_state == arc_l2c_only) { 3120 /* 3121 * This buffer is on the 2nd Level ARC. 3122 */ 3123 3124 hdr->b_arc_access = ddi_get_lbolt(); 3125 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 3126 arc_change_state(arc_mfu, hdr, hash_lock); 3127 } else { 3128 ASSERT(!"invalid arc state"); 3129 } 3130} 3131 3132/* a generic arc_done_func_t which you can use */ 3133/* ARGSUSED */ 3134void 3135arc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg) 3136{ 3137 if (zio == NULL || zio->io_error == 0) 3138 bcopy(buf->b_data, arg, buf->b_hdr->b_size); 3139 VERIFY(arc_buf_remove_ref(buf, arg)); 3140} 3141 3142/* a generic arc_done_func_t */ 3143void 3144arc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg) 3145{ 3146 arc_buf_t **bufp = arg; 3147 if (zio && zio->io_error) { 3148 VERIFY(arc_buf_remove_ref(buf, arg)); 3149 *bufp = NULL; 3150 } else { 3151 *bufp = buf; 3152 ASSERT(buf->b_data); 3153 } 3154} 3155 3156static void 3157arc_read_done(zio_t *zio) 3158{ 3159 arc_buf_hdr_t *hdr; 3160 arc_buf_t *buf; 3161 arc_buf_t *abuf; /* buffer we're assigning to callback */ 3162 kmutex_t *hash_lock = NULL; 3163 arc_callback_t *callback_list, *acb; 3164 int freeable = FALSE; 3165 3166 buf = zio->io_private; 3167 hdr = buf->b_hdr; 3168 3169 /* 3170 * The hdr was inserted into hash-table and removed from lists 3171 * prior to starting I/O. We should find this header, since 3172 * it's in the hash table, and it should be legit since it's 3173 * not possible to evict it during the I/O. The only possible 3174 * reason for it not to be found is if we were freed during the 3175 * read. 3176 */ 3177 if (HDR_IN_HASH_TABLE(hdr)) { 3178 ASSERT3U(hdr->b_birth, ==, BP_PHYSICAL_BIRTH(zio->io_bp)); 3179 ASSERT3U(hdr->b_dva.dva_word[0], ==, 3180 BP_IDENTITY(zio->io_bp)->dva_word[0]); 3181 ASSERT3U(hdr->b_dva.dva_word[1], ==, 3182 BP_IDENTITY(zio->io_bp)->dva_word[1]); 3183 3184 arc_buf_hdr_t *found = buf_hash_find(hdr->b_spa, zio->io_bp, 3185 &hash_lock); 3186 3187 ASSERT((found == NULL && HDR_FREED_IN_READ(hdr) && 3188 hash_lock == NULL) || 3189 (found == hdr && 3190 DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) || 3191 (found == hdr && HDR_L2_READING(hdr))); 3192 } 3193 3194 hdr->b_flags &= ~ARC_FLAG_L2_EVICTED; 3195 if (l2arc_noprefetch && (hdr->b_flags & ARC_FLAG_PREFETCH)) 3196 hdr->b_flags &= ~ARC_FLAG_L2CACHE; 3197 3198 /* byteswap if necessary */ 3199 callback_list = hdr->b_acb; 3200 ASSERT(callback_list != NULL); 3201 if (BP_SHOULD_BYTESWAP(zio->io_bp) && zio->io_error == 0) { 3202 dmu_object_byteswap_t bswap = 3203 DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp)); 3204 arc_byteswap_func_t *func = BP_GET_LEVEL(zio->io_bp) > 0 ? 3205 byteswap_uint64_array : 3206 dmu_ot_byteswap[bswap].ob_func; 3207 func(buf->b_data, hdr->b_size); 3208 } 3209 3210 arc_cksum_compute(buf, B_FALSE); 3211#ifdef illumos 3212 arc_buf_watch(buf); 3213#endif 3214 3215 if (hash_lock && zio->io_error == 0 && hdr->b_state == arc_anon) { 3216 /* 3217 * Only call arc_access on anonymous buffers. This is because 3218 * if we've issued an I/O for an evicted buffer, we've already 3219 * called arc_access (to prevent any simultaneous readers from 3220 * getting confused). 3221 */ 3222 arc_access(hdr, hash_lock); 3223 } 3224 3225 /* create copies of the data buffer for the callers */ 3226 abuf = buf; 3227 for (acb = callback_list; acb; acb = acb->acb_next) { 3228 if (acb->acb_done) { 3229 if (abuf == NULL) { 3230 ARCSTAT_BUMP(arcstat_duplicate_reads); 3231 abuf = arc_buf_clone(buf); 3232 } 3233 acb->acb_buf = abuf; 3234 abuf = NULL; 3235 } 3236 } 3237 hdr->b_acb = NULL; 3238 hdr->b_flags &= ~ARC_FLAG_IO_IN_PROGRESS; 3239 ASSERT(!HDR_BUF_AVAILABLE(hdr)); 3240 if (abuf == buf) { 3241 ASSERT(buf->b_efunc == NULL); 3242 ASSERT(hdr->b_datacnt == 1); 3243 hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE; 3244 } 3245 3246 ASSERT(refcount_is_zero(&hdr->b_refcnt) || callback_list != NULL); 3247 3248 if (zio->io_error != 0) { 3249 hdr->b_flags |= ARC_FLAG_IO_ERROR; 3250 if (hdr->b_state != arc_anon) 3251 arc_change_state(arc_anon, hdr, hash_lock); 3252 if (HDR_IN_HASH_TABLE(hdr)) 3253 buf_hash_remove(hdr); 3254 freeable = refcount_is_zero(&hdr->b_refcnt); 3255 } 3256 3257 /* 3258 * Broadcast before we drop the hash_lock to avoid the possibility 3259 * that the hdr (and hence the cv) might be freed before we get to 3260 * the cv_broadcast(). 3261 */ 3262 cv_broadcast(&hdr->b_cv); 3263 3264 if (hash_lock) { 3265 mutex_exit(hash_lock); 3266 } else { 3267 /* 3268 * This block was freed while we waited for the read to 3269 * complete. It has been removed from the hash table and 3270 * moved to the anonymous state (so that it won't show up 3271 * in the cache). 3272 */ 3273 ASSERT3P(hdr->b_state, ==, arc_anon); 3274 freeable = refcount_is_zero(&hdr->b_refcnt); 3275 } 3276 3277 /* execute each callback and free its structure */ 3278 while ((acb = callback_list) != NULL) { 3279 if (acb->acb_done) 3280 acb->acb_done(zio, acb->acb_buf, acb->acb_private); 3281 3282 if (acb->acb_zio_dummy != NULL) { 3283 acb->acb_zio_dummy->io_error = zio->io_error; 3284 zio_nowait(acb->acb_zio_dummy); 3285 } 3286 3287 callback_list = acb->acb_next; 3288 kmem_free(acb, sizeof (arc_callback_t)); 3289 } 3290 3291 if (freeable) 3292 arc_hdr_destroy(hdr); 3293} 3294 3295/* 3296 * "Read" the block block at the specified DVA (in bp) via the 3297 * cache. If the block is found in the cache, invoke the provided 3298 * callback immediately and return. Note that the `zio' parameter 3299 * in the callback will be NULL in this case, since no IO was 3300 * required. If the block is not in the cache pass the read request 3301 * on to the spa with a substitute callback function, so that the 3302 * requested block will be added to the cache. 3303 * 3304 * If a read request arrives for a block that has a read in-progress, 3305 * either wait for the in-progress read to complete (and return the 3306 * results); or, if this is a read with a "done" func, add a record 3307 * to the read to invoke the "done" func when the read completes, 3308 * and return; or just return. 3309 * 3310 * arc_read_done() will invoke all the requested "done" functions 3311 * for readers of this block. 3312 */ 3313int 3314arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_done_func_t *done, 3315 void *private, zio_priority_t priority, int zio_flags, 3316 arc_flags_t *arc_flags, const zbookmark_phys_t *zb) 3317{ 3318 arc_buf_hdr_t *hdr = NULL; 3319 arc_buf_t *buf = NULL; 3320 kmutex_t *hash_lock = NULL; 3321 zio_t *rzio; 3322 uint64_t guid = spa_load_guid(spa); 3323 3324 ASSERT(!BP_IS_EMBEDDED(bp) || 3325 BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA); 3326 3327top: 3328 if (!BP_IS_EMBEDDED(bp)) { 3329 /* 3330 * Embedded BP's have no DVA and require no I/O to "read". 3331 * Create an anonymous arc buf to back it. 3332 */ 3333 hdr = buf_hash_find(guid, bp, &hash_lock); 3334 } 3335 3336 if (hdr != NULL && hdr->b_datacnt > 0) { 3337 3338 *arc_flags |= ARC_FLAG_CACHED; 3339 3340 if (HDR_IO_IN_PROGRESS(hdr)) { 3341 3342 if (*arc_flags & ARC_FLAG_WAIT) { 3343 cv_wait(&hdr->b_cv, hash_lock); 3344 mutex_exit(hash_lock); 3345 goto top; 3346 } 3347 ASSERT(*arc_flags & ARC_FLAG_NOWAIT); 3348 3349 if (done) { 3350 arc_callback_t *acb = NULL; 3351 3352 acb = kmem_zalloc(sizeof (arc_callback_t), 3353 KM_SLEEP); 3354 acb->acb_done = done; 3355 acb->acb_private = private; 3356 if (pio != NULL) 3357 acb->acb_zio_dummy = zio_null(pio, 3358 spa, NULL, NULL, NULL, zio_flags); 3359 3360 ASSERT(acb->acb_done != NULL); 3361 acb->acb_next = hdr->b_acb; 3362 hdr->b_acb = acb; 3363 add_reference(hdr, hash_lock, private); 3364 mutex_exit(hash_lock); 3365 return (0); 3366 } 3367 mutex_exit(hash_lock); 3368 return (0); 3369 } 3370 3371 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu); 3372 3373 if (done) { 3374 add_reference(hdr, hash_lock, private); 3375 /* 3376 * If this block is already in use, create a new 3377 * copy of the data so that we will be guaranteed 3378 * that arc_release() will always succeed. 3379 */ 3380 buf = hdr->b_buf; 3381 ASSERT(buf); 3382 ASSERT(buf->b_data); 3383 if (HDR_BUF_AVAILABLE(hdr)) { 3384 ASSERT(buf->b_efunc == NULL); 3385 hdr->b_flags &= ~ARC_FLAG_BUF_AVAILABLE; 3386 } else { 3387 buf = arc_buf_clone(buf); 3388 } 3389 3390 } else if (*arc_flags & ARC_FLAG_PREFETCH && 3391 refcount_count(&hdr->b_refcnt) == 0) { 3392 hdr->b_flags |= ARC_FLAG_PREFETCH; 3393 } 3394 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr); 3395 arc_access(hdr, hash_lock); 3396 if (*arc_flags & ARC_FLAG_L2CACHE) 3397 hdr->b_flags |= ARC_FLAG_L2CACHE; 3398 if (*arc_flags & ARC_FLAG_L2COMPRESS) 3399 hdr->b_flags |= ARC_FLAG_L2COMPRESS; 3400 mutex_exit(hash_lock); 3401 ARCSTAT_BUMP(arcstat_hits); 3402 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_FLAG_PREFETCH), 3403 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA, 3404 data, metadata, hits); 3405 3406 if (done) 3407 done(NULL, buf, private); 3408 } else { 3409 uint64_t size = BP_GET_LSIZE(bp); 3410 arc_callback_t *acb; 3411 vdev_t *vd = NULL; 3412 uint64_t addr = 0; 3413 boolean_t devw = B_FALSE; 3414 enum zio_compress b_compress = ZIO_COMPRESS_OFF; 3415 uint64_t b_asize = 0; 3416 3417 if (hdr == NULL) { 3418 /* this block is not in the cache */ 3419 arc_buf_hdr_t *exists = NULL; 3420 arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp); 3421 buf = arc_buf_alloc(spa, size, private, type); 3422 hdr = buf->b_hdr; 3423 if (!BP_IS_EMBEDDED(bp)) { 3424 hdr->b_dva = *BP_IDENTITY(bp); 3425 hdr->b_birth = BP_PHYSICAL_BIRTH(bp); 3426 hdr->b_cksum0 = bp->blk_cksum.zc_word[0]; 3427 exists = buf_hash_insert(hdr, &hash_lock); 3428 } 3429 if (exists != NULL) { 3430 /* somebody beat us to the hash insert */ 3431 mutex_exit(hash_lock); 3432 buf_discard_identity(hdr); 3433 (void) arc_buf_remove_ref(buf, private); 3434 goto top; /* restart the IO request */ 3435 } 3436 3437 /* if this is a prefetch, we don't have a reference */ 3438 if (*arc_flags & ARC_FLAG_PREFETCH) { 3439 (void) remove_reference(hdr, hash_lock, 3440 private); 3441 hdr->b_flags |= ARC_FLAG_PREFETCH; 3442 } 3443 if (*arc_flags & ARC_FLAG_L2CACHE) 3444 hdr->b_flags |= ARC_FLAG_L2CACHE; 3445 if (*arc_flags & ARC_FLAG_L2COMPRESS) 3446 hdr->b_flags |= ARC_FLAG_L2COMPRESS; 3447 if (BP_GET_LEVEL(bp) > 0) 3448 hdr->b_flags |= ARC_FLAG_INDIRECT; 3449 } else { 3450 /* this block is in the ghost cache */ 3451 ASSERT(GHOST_STATE(hdr->b_state)); 3452 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 3453 ASSERT0(refcount_count(&hdr->b_refcnt)); 3454 ASSERT(hdr->b_buf == NULL); 3455 3456 /* if this is a prefetch, we don't have a reference */ 3457 if (*arc_flags & ARC_FLAG_PREFETCH) 3458 hdr->b_flags |= ARC_FLAG_PREFETCH; 3459 else 3460 add_reference(hdr, hash_lock, private); 3461 if (*arc_flags & ARC_FLAG_L2CACHE) 3462 hdr->b_flags |= ARC_FLAG_L2CACHE; 3463 if (*arc_flags & ARC_FLAG_L2COMPRESS) 3464 hdr->b_flags |= ARC_FLAG_L2COMPRESS; 3465 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 3466 buf->b_hdr = hdr; 3467 buf->b_data = NULL; 3468 buf->b_efunc = NULL; 3469 buf->b_private = NULL; 3470 buf->b_next = NULL; 3471 hdr->b_buf = buf; 3472 ASSERT(hdr->b_datacnt == 0); 3473 hdr->b_datacnt = 1; 3474 arc_get_data_buf(buf); 3475 arc_access(hdr, hash_lock); 3476 } 3477 3478 ASSERT(!GHOST_STATE(hdr->b_state)); 3479 3480 acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP); 3481 acb->acb_done = done; 3482 acb->acb_private = private; 3483 3484 ASSERT(hdr->b_acb == NULL); 3485 hdr->b_acb = acb; 3486 hdr->b_flags |= ARC_FLAG_IO_IN_PROGRESS; 3487 3488 if (hdr->b_l2hdr != NULL && 3489 (vd = hdr->b_l2hdr->b_dev->l2ad_vdev) != NULL) { 3490 devw = hdr->b_l2hdr->b_dev->l2ad_writing; 3491 addr = hdr->b_l2hdr->b_daddr; 3492 b_compress = hdr->b_l2hdr->b_compress; 3493 b_asize = hdr->b_l2hdr->b_asize; 3494 /* 3495 * Lock out device removal. 3496 */ 3497 if (vdev_is_dead(vd) || 3498 !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER)) 3499 vd = NULL; 3500 } 3501 3502 if (hash_lock != NULL) 3503 mutex_exit(hash_lock); 3504 3505 /* 3506 * At this point, we have a level 1 cache miss. Try again in 3507 * L2ARC if possible. 3508 */ 3509 ASSERT3U(hdr->b_size, ==, size); 3510 DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp, 3511 uint64_t, size, zbookmark_phys_t *, zb); 3512 ARCSTAT_BUMP(arcstat_misses); 3513 ARCSTAT_CONDSTAT(!(hdr->b_flags & ARC_FLAG_PREFETCH), 3514 demand, prefetch, hdr->b_type != ARC_BUFC_METADATA, 3515 data, metadata, misses); 3516#ifdef _KERNEL 3517 curthread->td_ru.ru_inblock++; 3518#endif 3519 3520 if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) { 3521 /* 3522 * Read from the L2ARC if the following are true: 3523 * 1. The L2ARC vdev was previously cached. 3524 * 2. This buffer still has L2ARC metadata. 3525 * 3. This buffer isn't currently writing to the L2ARC. 3526 * 4. The L2ARC entry wasn't evicted, which may 3527 * also have invalidated the vdev. 3528 * 5. This isn't prefetch and l2arc_noprefetch is set. 3529 */ 3530 if (hdr->b_l2hdr != NULL && 3531 !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) && 3532 !(l2arc_noprefetch && HDR_PREFETCH(hdr))) { 3533 l2arc_read_callback_t *cb; 3534 3535 DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr); 3536 ARCSTAT_BUMP(arcstat_l2_hits); 3537 3538 cb = kmem_zalloc(sizeof (l2arc_read_callback_t), 3539 KM_SLEEP); 3540 cb->l2rcb_buf = buf; 3541 cb->l2rcb_spa = spa; 3542 cb->l2rcb_bp = *bp; 3543 cb->l2rcb_zb = *zb; 3544 cb->l2rcb_flags = zio_flags; 3545 cb->l2rcb_compress = b_compress; 3546 3547 ASSERT(addr >= VDEV_LABEL_START_SIZE && 3548 addr + size < vd->vdev_psize - 3549 VDEV_LABEL_END_SIZE); 3550 3551 /* 3552 * l2arc read. The SCL_L2ARC lock will be 3553 * released by l2arc_read_done(). 3554 * Issue a null zio if the underlying buffer 3555 * was squashed to zero size by compression. 3556 */ 3557 if (b_compress == ZIO_COMPRESS_EMPTY) { 3558 rzio = zio_null(pio, spa, vd, 3559 l2arc_read_done, cb, 3560 zio_flags | ZIO_FLAG_DONT_CACHE | 3561 ZIO_FLAG_CANFAIL | 3562 ZIO_FLAG_DONT_PROPAGATE | 3563 ZIO_FLAG_DONT_RETRY); 3564 } else { 3565 rzio = zio_read_phys(pio, vd, addr, 3566 b_asize, buf->b_data, 3567 ZIO_CHECKSUM_OFF, 3568 l2arc_read_done, cb, priority, 3569 zio_flags | ZIO_FLAG_DONT_CACHE | 3570 ZIO_FLAG_CANFAIL | 3571 ZIO_FLAG_DONT_PROPAGATE | 3572 ZIO_FLAG_DONT_RETRY, B_FALSE); 3573 } 3574 DTRACE_PROBE2(l2arc__read, vdev_t *, vd, 3575 zio_t *, rzio); 3576 ARCSTAT_INCR(arcstat_l2_read_bytes, b_asize); 3577 3578 if (*arc_flags & ARC_FLAG_NOWAIT) { 3579 zio_nowait(rzio); 3580 return (0); 3581 } 3582 3583 ASSERT(*arc_flags & ARC_FLAG_WAIT); 3584 if (zio_wait(rzio) == 0) 3585 return (0); 3586 3587 /* l2arc read error; goto zio_read() */ 3588 } else { 3589 DTRACE_PROBE1(l2arc__miss, 3590 arc_buf_hdr_t *, hdr); 3591 ARCSTAT_BUMP(arcstat_l2_misses); 3592 if (HDR_L2_WRITING(hdr)) 3593 ARCSTAT_BUMP(arcstat_l2_rw_clash); 3594 spa_config_exit(spa, SCL_L2ARC, vd); 3595 } 3596 } else { 3597 if (vd != NULL) 3598 spa_config_exit(spa, SCL_L2ARC, vd); 3599 if (l2arc_ndev != 0) { 3600 DTRACE_PROBE1(l2arc__miss, 3601 arc_buf_hdr_t *, hdr); 3602 ARCSTAT_BUMP(arcstat_l2_misses); 3603 } 3604 } 3605 3606 rzio = zio_read(pio, spa, bp, buf->b_data, size, 3607 arc_read_done, buf, priority, zio_flags, zb); 3608 3609 if (*arc_flags & ARC_FLAG_WAIT) 3610 return (zio_wait(rzio)); 3611 3612 ASSERT(*arc_flags & ARC_FLAG_NOWAIT); 3613 zio_nowait(rzio); 3614 } 3615 return (0); 3616} 3617 3618void 3619arc_set_callback(arc_buf_t *buf, arc_evict_func_t *func, void *private) 3620{ 3621 ASSERT(buf->b_hdr != NULL); 3622 ASSERT(buf->b_hdr->b_state != arc_anon); 3623 ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt) || func == NULL); 3624 ASSERT(buf->b_efunc == NULL); 3625 ASSERT(!HDR_BUF_AVAILABLE(buf->b_hdr)); 3626 3627 buf->b_efunc = func; 3628 buf->b_private = private; 3629} 3630 3631/* 3632 * Notify the arc that a block was freed, and thus will never be used again. 3633 */ 3634void 3635arc_freed(spa_t *spa, const blkptr_t *bp) 3636{ 3637 arc_buf_hdr_t *hdr; 3638 kmutex_t *hash_lock; 3639 uint64_t guid = spa_load_guid(spa); 3640 3641 ASSERT(!BP_IS_EMBEDDED(bp)); 3642 3643 hdr = buf_hash_find(guid, bp, &hash_lock); 3644 if (hdr == NULL) 3645 return; 3646 if (HDR_BUF_AVAILABLE(hdr)) { 3647 arc_buf_t *buf = hdr->b_buf; 3648 add_reference(hdr, hash_lock, FTAG); 3649 hdr->b_flags &= ~ARC_FLAG_BUF_AVAILABLE; 3650 mutex_exit(hash_lock); 3651 3652 arc_release(buf, FTAG); 3653 (void) arc_buf_remove_ref(buf, FTAG); 3654 } else { 3655 mutex_exit(hash_lock); 3656 } 3657 3658} 3659 3660/* 3661 * Clear the user eviction callback set by arc_set_callback(), first calling 3662 * it if it exists. Because the presence of a callback keeps an arc_buf cached 3663 * clearing the callback may result in the arc_buf being destroyed. However, 3664 * it will not result in the *last* arc_buf being destroyed, hence the data 3665 * will remain cached in the ARC. We make a copy of the arc buffer here so 3666 * that we can process the callback without holding any locks. 3667 * 3668 * It's possible that the callback is already in the process of being cleared 3669 * by another thread. In this case we can not clear the callback. 3670 * 3671 * Returns B_TRUE if the callback was successfully called and cleared. 3672 */ 3673boolean_t 3674arc_clear_callback(arc_buf_t *buf) 3675{ 3676 arc_buf_hdr_t *hdr; 3677 kmutex_t *hash_lock; 3678 arc_evict_func_t *efunc = buf->b_efunc; 3679 void *private = buf->b_private; 3680 list_t *list, *evicted_list; 3681 kmutex_t *lock, *evicted_lock; 3682 3683 mutex_enter(&buf->b_evict_lock); 3684 hdr = buf->b_hdr; 3685 if (hdr == NULL) { 3686 /* 3687 * We are in arc_do_user_evicts(). 3688 */ 3689 ASSERT(buf->b_data == NULL); 3690 mutex_exit(&buf->b_evict_lock); 3691 return (B_FALSE); 3692 } else if (buf->b_data == NULL) { 3693 /* 3694 * We are on the eviction list; process this buffer now 3695 * but let arc_do_user_evicts() do the reaping. 3696 */ 3697 buf->b_efunc = NULL; 3698 mutex_exit(&buf->b_evict_lock); 3699 VERIFY0(efunc(private)); 3700 return (B_TRUE); 3701 } 3702 hash_lock = HDR_LOCK(hdr); 3703 mutex_enter(hash_lock); 3704 hdr = buf->b_hdr; 3705 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 3706 3707 ASSERT3U(refcount_count(&hdr->b_refcnt), <, hdr->b_datacnt); 3708 ASSERT(hdr->b_state == arc_mru || hdr->b_state == arc_mfu); 3709 3710 buf->b_efunc = NULL; 3711 buf->b_private = NULL; 3712 3713 if (hdr->b_datacnt > 1) { 3714 mutex_exit(&buf->b_evict_lock); 3715 arc_buf_destroy(buf, FALSE, TRUE); 3716 } else { 3717 ASSERT(buf == hdr->b_buf); 3718 hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE; 3719 mutex_exit(&buf->b_evict_lock); 3720 } 3721 3722 mutex_exit(hash_lock); 3723 VERIFY0(efunc(private)); 3724 return (B_TRUE); 3725} 3726 3727/* 3728 * Release this buffer from the cache, making it an anonymous buffer. This 3729 * must be done after a read and prior to modifying the buffer contents. 3730 * If the buffer has more than one reference, we must make 3731 * a new hdr for the buffer. 3732 */ 3733void 3734arc_release(arc_buf_t *buf, void *tag) 3735{ 3736 arc_buf_hdr_t *hdr; 3737 kmutex_t *hash_lock = NULL; 3738 l2arc_buf_hdr_t *l2hdr; 3739 uint64_t buf_size; 3740 3741 /* 3742 * It would be nice to assert that if it's DMU metadata (level > 3743 * 0 || it's the dnode file), then it must be syncing context. 3744 * But we don't know that information at this level. 3745 */ 3746 3747 mutex_enter(&buf->b_evict_lock); 3748 hdr = buf->b_hdr; 3749 3750 /* this buffer is not on any list */ 3751 ASSERT(refcount_count(&hdr->b_refcnt) > 0); 3752 3753 if (hdr->b_state == arc_anon) { 3754 /* this buffer is already released */ 3755 ASSERT(buf->b_efunc == NULL); 3756 } else { 3757 hash_lock = HDR_LOCK(hdr); 3758 mutex_enter(hash_lock); 3759 hdr = buf->b_hdr; 3760 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 3761 } 3762 3763 l2hdr = hdr->b_l2hdr; 3764 if (l2hdr) { 3765 mutex_enter(&l2arc_buflist_mtx); 3766 arc_buf_l2_cdata_free(hdr); 3767 hdr->b_l2hdr = NULL; 3768 list_remove(l2hdr->b_dev->l2ad_buflist, hdr); 3769 } 3770 buf_size = hdr->b_size; 3771 3772 /* 3773 * Do we have more than one buf? 3774 */ 3775 if (hdr->b_datacnt > 1) { 3776 arc_buf_hdr_t *nhdr; 3777 arc_buf_t **bufp; 3778 uint64_t blksz = hdr->b_size; 3779 uint64_t spa = hdr->b_spa; 3780 arc_buf_contents_t type = hdr->b_type; 3781 uint32_t flags = hdr->b_flags; 3782 3783 ASSERT(hdr->b_buf != buf || buf->b_next != NULL); 3784 /* 3785 * Pull the data off of this hdr and attach it to 3786 * a new anonymous hdr. 3787 */ 3788 (void) remove_reference(hdr, hash_lock, tag); 3789 bufp = &hdr->b_buf; 3790 while (*bufp != buf) 3791 bufp = &(*bufp)->b_next; 3792 *bufp = buf->b_next; 3793 buf->b_next = NULL; 3794 3795 ASSERT3U(hdr->b_state->arcs_size, >=, hdr->b_size); 3796 atomic_add_64(&hdr->b_state->arcs_size, -hdr->b_size); 3797 if (refcount_is_zero(&hdr->b_refcnt)) { 3798 uint64_t *size = &hdr->b_state->arcs_lsize[hdr->b_type]; 3799 ASSERT3U(*size, >=, hdr->b_size); 3800 atomic_add_64(size, -hdr->b_size); 3801 } 3802 3803 /* 3804 * We're releasing a duplicate user data buffer, update 3805 * our statistics accordingly. 3806 */ 3807 if (hdr->b_type == ARC_BUFC_DATA) { 3808 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers); 3809 ARCSTAT_INCR(arcstat_duplicate_buffers_size, 3810 -hdr->b_size); 3811 } 3812 hdr->b_datacnt -= 1; 3813 arc_cksum_verify(buf); 3814#ifdef illumos 3815 arc_buf_unwatch(buf); 3816#endif 3817 3818 mutex_exit(hash_lock); 3819 3820 nhdr = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE); 3821 nhdr->b_size = blksz; 3822 nhdr->b_spa = spa; 3823 nhdr->b_type = type; 3824 nhdr->b_buf = buf; 3825 nhdr->b_state = arc_anon; 3826 nhdr->b_arc_access = 0; 3827 nhdr->b_flags = flags & ARC_FLAG_L2_WRITING; 3828 nhdr->b_l2hdr = NULL; 3829 nhdr->b_datacnt = 1; 3830 nhdr->b_freeze_cksum = NULL; 3831 (void) refcount_add(&nhdr->b_refcnt, tag); 3832 buf->b_hdr = nhdr; 3833 mutex_exit(&buf->b_evict_lock); 3834 atomic_add_64(&arc_anon->arcs_size, blksz); 3835 } else { 3836 mutex_exit(&buf->b_evict_lock); 3837 ASSERT(refcount_count(&hdr->b_refcnt) == 1); 3838 ASSERT(!list_link_active(&hdr->b_arc_node)); 3839 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 3840 if (hdr->b_state != arc_anon) 3841 arc_change_state(arc_anon, hdr, hash_lock); 3842 hdr->b_arc_access = 0; 3843 if (hash_lock) 3844 mutex_exit(hash_lock); 3845 3846 buf_discard_identity(hdr); 3847 arc_buf_thaw(buf); 3848 } 3849 buf->b_efunc = NULL; 3850 buf->b_private = NULL; 3851 3852 if (l2hdr) { 3853 ARCSTAT_INCR(arcstat_l2_asize, -l2hdr->b_asize); 3854 vdev_space_update(l2hdr->b_dev->l2ad_vdev, 3855 -l2hdr->b_asize, 0, 0); 3856 trim_map_free(l2hdr->b_dev->l2ad_vdev, l2hdr->b_daddr, 3857 hdr->b_size, 0); 3858 kmem_free(l2hdr, sizeof (l2arc_buf_hdr_t)); 3859 ARCSTAT_INCR(arcstat_l2_size, -buf_size); 3860 mutex_exit(&l2arc_buflist_mtx); 3861 } 3862} 3863 3864int 3865arc_released(arc_buf_t *buf) 3866{ 3867 int released; 3868 3869 mutex_enter(&buf->b_evict_lock); 3870 released = (buf->b_data != NULL && buf->b_hdr->b_state == arc_anon); 3871 mutex_exit(&buf->b_evict_lock); 3872 return (released); 3873} 3874 3875#ifdef ZFS_DEBUG 3876int 3877arc_referenced(arc_buf_t *buf) 3878{ 3879 int referenced; 3880 3881 mutex_enter(&buf->b_evict_lock); 3882 referenced = (refcount_count(&buf->b_hdr->b_refcnt)); 3883 mutex_exit(&buf->b_evict_lock); 3884 return (referenced); 3885} 3886#endif 3887 3888static void 3889arc_write_ready(zio_t *zio) 3890{ 3891 arc_write_callback_t *callback = zio->io_private; 3892 arc_buf_t *buf = callback->awcb_buf; 3893 arc_buf_hdr_t *hdr = buf->b_hdr; 3894 3895 ASSERT(!refcount_is_zero(&buf->b_hdr->b_refcnt)); 3896 callback->awcb_ready(zio, buf, callback->awcb_private); 3897 3898 /* 3899 * If the IO is already in progress, then this is a re-write 3900 * attempt, so we need to thaw and re-compute the cksum. 3901 * It is the responsibility of the callback to handle the 3902 * accounting for any re-write attempt. 3903 */ 3904 if (HDR_IO_IN_PROGRESS(hdr)) { 3905 mutex_enter(&hdr->b_freeze_lock); 3906 if (hdr->b_freeze_cksum != NULL) { 3907 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 3908 hdr->b_freeze_cksum = NULL; 3909 } 3910 mutex_exit(&hdr->b_freeze_lock); 3911 } 3912 arc_cksum_compute(buf, B_FALSE); 3913 hdr->b_flags |= ARC_FLAG_IO_IN_PROGRESS; 3914} 3915 3916/* 3917 * The SPA calls this callback for each physical write that happens on behalf 3918 * of a logical write. See the comment in dbuf_write_physdone() for details. 3919 */ 3920static void 3921arc_write_physdone(zio_t *zio) 3922{ 3923 arc_write_callback_t *cb = zio->io_private; 3924 if (cb->awcb_physdone != NULL) 3925 cb->awcb_physdone(zio, cb->awcb_buf, cb->awcb_private); 3926} 3927 3928static void 3929arc_write_done(zio_t *zio) 3930{ 3931 arc_write_callback_t *callback = zio->io_private; 3932 arc_buf_t *buf = callback->awcb_buf; 3933 arc_buf_hdr_t *hdr = buf->b_hdr; 3934 3935 ASSERT(hdr->b_acb == NULL); 3936 3937 if (zio->io_error == 0) { 3938 if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) { 3939 buf_discard_identity(hdr); 3940 } else { 3941 hdr->b_dva = *BP_IDENTITY(zio->io_bp); 3942 hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp); 3943 hdr->b_cksum0 = zio->io_bp->blk_cksum.zc_word[0]; 3944 } 3945 } else { 3946 ASSERT(BUF_EMPTY(hdr)); 3947 } 3948 3949 /* 3950 * If the block to be written was all-zero or compressed enough to be 3951 * embedded in the BP, no write was performed so there will be no 3952 * dva/birth/checksum. The buffer must therefore remain anonymous 3953 * (and uncached). 3954 */ 3955 if (!BUF_EMPTY(hdr)) { 3956 arc_buf_hdr_t *exists; 3957 kmutex_t *hash_lock; 3958 3959 ASSERT(zio->io_error == 0); 3960 3961 arc_cksum_verify(buf); 3962 3963 exists = buf_hash_insert(hdr, &hash_lock); 3964 if (exists) { 3965 /* 3966 * This can only happen if we overwrite for 3967 * sync-to-convergence, because we remove 3968 * buffers from the hash table when we arc_free(). 3969 */ 3970 if (zio->io_flags & ZIO_FLAG_IO_REWRITE) { 3971 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) 3972 panic("bad overwrite, hdr=%p exists=%p", 3973 (void *)hdr, (void *)exists); 3974 ASSERT(refcount_is_zero(&exists->b_refcnt)); 3975 arc_change_state(arc_anon, exists, hash_lock); 3976 mutex_exit(hash_lock); 3977 arc_hdr_destroy(exists); 3978 exists = buf_hash_insert(hdr, &hash_lock); 3979 ASSERT3P(exists, ==, NULL); 3980 } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) { 3981 /* nopwrite */ 3982 ASSERT(zio->io_prop.zp_nopwrite); 3983 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) 3984 panic("bad nopwrite, hdr=%p exists=%p", 3985 (void *)hdr, (void *)exists); 3986 } else { 3987 /* Dedup */ 3988 ASSERT(hdr->b_datacnt == 1); 3989 ASSERT(hdr->b_state == arc_anon); 3990 ASSERT(BP_GET_DEDUP(zio->io_bp)); 3991 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0); 3992 } 3993 } 3994 hdr->b_flags &= ~ARC_FLAG_IO_IN_PROGRESS; 3995 /* if it's not anon, we are doing a scrub */ 3996 if (!exists && hdr->b_state == arc_anon) 3997 arc_access(hdr, hash_lock); 3998 mutex_exit(hash_lock); 3999 } else { 4000 hdr->b_flags &= ~ARC_FLAG_IO_IN_PROGRESS; 4001 } 4002 4003 ASSERT(!refcount_is_zero(&hdr->b_refcnt)); 4004 callback->awcb_done(zio, buf, callback->awcb_private); 4005 4006 kmem_free(callback, sizeof (arc_write_callback_t)); 4007} 4008 4009zio_t * 4010arc_write(zio_t *pio, spa_t *spa, uint64_t txg, 4011 blkptr_t *bp, arc_buf_t *buf, boolean_t l2arc, boolean_t l2arc_compress, 4012 const zio_prop_t *zp, arc_done_func_t *ready, arc_done_func_t *physdone, 4013 arc_done_func_t *done, void *private, zio_priority_t priority, 4014 int zio_flags, const zbookmark_phys_t *zb) 4015{ 4016 arc_buf_hdr_t *hdr = buf->b_hdr; 4017 arc_write_callback_t *callback; 4018 zio_t *zio; 4019 4020 ASSERT(ready != NULL); 4021 ASSERT(done != NULL); 4022 ASSERT(!HDR_IO_ERROR(hdr)); 4023 ASSERT((hdr->b_flags & ARC_FLAG_IO_IN_PROGRESS) == 0); 4024 ASSERT(hdr->b_acb == NULL); 4025 if (l2arc) 4026 hdr->b_flags |= ARC_FLAG_L2CACHE; 4027 if (l2arc_compress) 4028 hdr->b_flags |= ARC_FLAG_L2COMPRESS; 4029 callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP); 4030 callback->awcb_ready = ready; 4031 callback->awcb_physdone = physdone; 4032 callback->awcb_done = done; 4033 callback->awcb_private = private; 4034 callback->awcb_buf = buf; 4035 4036 zio = zio_write(pio, spa, txg, bp, buf->b_data, hdr->b_size, zp, 4037 arc_write_ready, arc_write_physdone, arc_write_done, callback, 4038 priority, zio_flags, zb); 4039 4040 return (zio); 4041} 4042 4043static int 4044arc_memory_throttle(uint64_t reserve, uint64_t txg) 4045{ 4046#ifdef _KERNEL 4047 uint64_t available_memory = ptob(freemem); 4048 static uint64_t page_load = 0; 4049 static uint64_t last_txg = 0; 4050 4051#if defined(__i386) || !defined(UMA_MD_SMALL_ALLOC) 4052 available_memory = 4053 MIN(available_memory, ptob(vmem_size(heap_arena, VMEM_FREE))); 4054#endif 4055 4056 if (freemem > (uint64_t)physmem * arc_lotsfree_percent / 100) 4057 return (0); 4058 4059 if (txg > last_txg) { 4060 last_txg = txg; 4061 page_load = 0; 4062 } 4063 /* 4064 * If we are in pageout, we know that memory is already tight, 4065 * the arc is already going to be evicting, so we just want to 4066 * continue to let page writes occur as quickly as possible. 4067 */ 4068 if (curproc == pageproc) { 4069 if (page_load > MAX(ptob(minfree), available_memory) / 4) 4070 return (SET_ERROR(ERESTART)); 4071 /* Note: reserve is inflated, so we deflate */ 4072 page_load += reserve / 8; 4073 return (0); 4074 } else if (page_load > 0 && arc_reclaim_needed()) { 4075 /* memory is low, delay before restarting */ 4076 ARCSTAT_INCR(arcstat_memory_throttle_count, 1); 4077 return (SET_ERROR(EAGAIN)); 4078 } 4079 page_load = 0; 4080#endif 4081 return (0); 4082} 4083 4084void 4085arc_tempreserve_clear(uint64_t reserve) 4086{ 4087 atomic_add_64(&arc_tempreserve, -reserve); 4088 ASSERT((int64_t)arc_tempreserve >= 0); 4089} 4090 4091int 4092arc_tempreserve_space(uint64_t reserve, uint64_t txg) 4093{ 4094 int error; 4095 uint64_t anon_size; 4096 4097 if (reserve > arc_c/4 && !arc_no_grow) { 4098 arc_c = MIN(arc_c_max, reserve * 4); 4099 DTRACE_PROBE1(arc__set_reserve, uint64_t, arc_c); 4100 } 4101 if (reserve > arc_c) 4102 return (SET_ERROR(ENOMEM)); 4103 4104 /* 4105 * Don't count loaned bufs as in flight dirty data to prevent long 4106 * network delays from blocking transactions that are ready to be 4107 * assigned to a txg. 4108 */ 4109 anon_size = MAX((int64_t)(arc_anon->arcs_size - arc_loaned_bytes), 0); 4110 4111 /* 4112 * Writes will, almost always, require additional memory allocations 4113 * in order to compress/encrypt/etc the data. We therefore need to 4114 * make sure that there is sufficient available memory for this. 4115 */ 4116 error = arc_memory_throttle(reserve, txg); 4117 if (error != 0) 4118 return (error); 4119 4120 /* 4121 * Throttle writes when the amount of dirty data in the cache 4122 * gets too large. We try to keep the cache less than half full 4123 * of dirty blocks so that our sync times don't grow too large. 4124 * Note: if two requests come in concurrently, we might let them 4125 * both succeed, when one of them should fail. Not a huge deal. 4126 */ 4127 4128 if (reserve + arc_tempreserve + anon_size > arc_c / 2 && 4129 anon_size > arc_c / 4) { 4130 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK " 4131 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n", 4132 arc_tempreserve>>10, 4133 arc_anon->arcs_lsize[ARC_BUFC_METADATA]>>10, 4134 arc_anon->arcs_lsize[ARC_BUFC_DATA]>>10, 4135 reserve>>10, arc_c>>10); 4136 return (SET_ERROR(ERESTART)); 4137 } 4138 atomic_add_64(&arc_tempreserve, reserve); 4139 return (0); 4140} 4141 4142static kmutex_t arc_lowmem_lock; 4143#ifdef _KERNEL 4144static eventhandler_tag arc_event_lowmem = NULL; 4145 4146static void 4147arc_lowmem(void *arg __unused, int howto __unused) 4148{ 4149 4150 /* Serialize access via arc_lowmem_lock. */ 4151 mutex_enter(&arc_lowmem_lock); 4152 mutex_enter(&arc_reclaim_thr_lock); 4153 needfree = 1; 4154 DTRACE_PROBE(arc__needfree); 4155 cv_signal(&arc_reclaim_thr_cv); 4156 4157 /* 4158 * It is unsafe to block here in arbitrary threads, because we can come 4159 * here from ARC itself and may hold ARC locks and thus risk a deadlock 4160 * with ARC reclaim thread. 4161 */ 4162 if (curproc == pageproc) { 4163 while (needfree) 4164 msleep(&needfree, &arc_reclaim_thr_lock, 0, "zfs:lowmem", 0); 4165 } 4166 mutex_exit(&arc_reclaim_thr_lock); 4167 mutex_exit(&arc_lowmem_lock); 4168} 4169#endif 4170 4171void 4172arc_init(void) 4173{ 4174 int i, prefetch_tunable_set = 0; 4175 4176 mutex_init(&arc_reclaim_thr_lock, NULL, MUTEX_DEFAULT, NULL); 4177 cv_init(&arc_reclaim_thr_cv, NULL, CV_DEFAULT, NULL); 4178 mutex_init(&arc_lowmem_lock, NULL, MUTEX_DEFAULT, NULL); 4179 4180 /* Convert seconds to clock ticks */ 4181 arc_min_prefetch_lifespan = 1 * hz; 4182 4183 /* Start out with 1/8 of all memory */ 4184 arc_c = kmem_size() / 8; 4185 4186#ifdef illumos 4187#ifdef _KERNEL 4188 /* 4189 * On architectures where the physical memory can be larger 4190 * than the addressable space (intel in 32-bit mode), we may 4191 * need to limit the cache to 1/8 of VM size. 4192 */ 4193 arc_c = MIN(arc_c, vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 8); 4194#endif 4195#endif /* illumos */ 4196 /* set min cache to 1/32 of all memory, or 16MB, whichever is more */ 4197 arc_c_min = MAX(arc_c / 4, 64<<18); 4198 /* set max to 1/2 of all memory, or all but 1GB, whichever is more */ 4199 if (arc_c * 8 >= 1<<30) 4200 arc_c_max = (arc_c * 8) - (1<<30); 4201 else 4202 arc_c_max = arc_c_min; 4203 arc_c_max = MAX(arc_c * 5, arc_c_max); 4204 4205#ifdef _KERNEL 4206 /* 4207 * Allow the tunables to override our calculations if they are 4208 * reasonable (ie. over 16MB) 4209 */ 4210 if (zfs_arc_max > 64<<18 && zfs_arc_max < kmem_size()) 4211 arc_c_max = zfs_arc_max; 4212 if (zfs_arc_min > 64<<18 && zfs_arc_min <= arc_c_max) 4213 arc_c_min = zfs_arc_min; 4214#endif 4215 4216 arc_c = arc_c_max; 4217 arc_p = (arc_c >> 1); 4218 4219 /* limit meta-data to 1/4 of the arc capacity */ 4220 arc_meta_limit = arc_c_max / 4; 4221 4222 /* Allow the tunable to override if it is reasonable */ 4223 if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max) 4224 arc_meta_limit = zfs_arc_meta_limit; 4225 4226 if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0) 4227 arc_c_min = arc_meta_limit / 2; 4228 4229 if (zfs_arc_meta_min > 0) { 4230 arc_meta_min = zfs_arc_meta_min; 4231 } else { 4232 arc_meta_min = arc_c_min / 2; 4233 } 4234 4235 if (zfs_arc_grow_retry > 0) 4236 arc_grow_retry = zfs_arc_grow_retry; 4237 4238 if (zfs_arc_shrink_shift > 0) 4239 arc_shrink_shift = zfs_arc_shrink_shift; 4240 4241 if (zfs_arc_p_min_shift > 0) 4242 arc_p_min_shift = zfs_arc_p_min_shift; 4243 4244 /* if kmem_flags are set, lets try to use less memory */ 4245 if (kmem_debugging()) 4246 arc_c = arc_c / 2; 4247 if (arc_c < arc_c_min) 4248 arc_c = arc_c_min; 4249 4250 zfs_arc_min = arc_c_min; 4251 zfs_arc_max = arc_c_max; 4252 4253 arc_anon = &ARC_anon; 4254 arc_mru = &ARC_mru; 4255 arc_mru_ghost = &ARC_mru_ghost; 4256 arc_mfu = &ARC_mfu; 4257 arc_mfu_ghost = &ARC_mfu_ghost; 4258 arc_l2c_only = &ARC_l2c_only; 4259 arc_size = 0; 4260 4261 for (i = 0; i < ARC_BUFC_NUMLISTS; i++) { 4262 mutex_init(&arc_anon->arcs_locks[i].arcs_lock, 4263 NULL, MUTEX_DEFAULT, NULL); 4264 mutex_init(&arc_mru->arcs_locks[i].arcs_lock, 4265 NULL, MUTEX_DEFAULT, NULL); 4266 mutex_init(&arc_mru_ghost->arcs_locks[i].arcs_lock, 4267 NULL, MUTEX_DEFAULT, NULL); 4268 mutex_init(&arc_mfu->arcs_locks[i].arcs_lock, 4269 NULL, MUTEX_DEFAULT, NULL); 4270 mutex_init(&arc_mfu_ghost->arcs_locks[i].arcs_lock, 4271 NULL, MUTEX_DEFAULT, NULL); 4272 mutex_init(&arc_l2c_only->arcs_locks[i].arcs_lock, 4273 NULL, MUTEX_DEFAULT, NULL); 4274 4275 list_create(&arc_mru->arcs_lists[i], 4276 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4277 list_create(&arc_mru_ghost->arcs_lists[i], 4278 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4279 list_create(&arc_mfu->arcs_lists[i], 4280 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4281 list_create(&arc_mfu_ghost->arcs_lists[i], 4282 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4283 list_create(&arc_mfu_ghost->arcs_lists[i], 4284 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4285 list_create(&arc_l2c_only->arcs_lists[i], 4286 sizeof (arc_buf_hdr_t), offsetof(arc_buf_hdr_t, b_arc_node)); 4287 } 4288 4289 buf_init(); 4290 4291 arc_thread_exit = 0; 4292 arc_eviction_list = NULL; 4293 mutex_init(&arc_eviction_mtx, NULL, MUTEX_DEFAULT, NULL); 4294 bzero(&arc_eviction_hdr, sizeof (arc_buf_hdr_t)); 4295 4296 arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED, 4297 sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL); 4298 4299 if (arc_ksp != NULL) { 4300 arc_ksp->ks_data = &arc_stats; 4301 kstat_install(arc_ksp); 4302 } 4303 4304 (void) thread_create(NULL, 0, arc_reclaim_thread, NULL, 0, &p0, 4305 TS_RUN, minclsyspri); 4306 4307#ifdef _KERNEL 4308 arc_event_lowmem = EVENTHANDLER_REGISTER(vm_lowmem, arc_lowmem, NULL, 4309 EVENTHANDLER_PRI_FIRST); 4310#endif 4311 4312 arc_dead = FALSE; 4313 arc_warm = B_FALSE; 4314 4315 /* 4316 * Calculate maximum amount of dirty data per pool. 4317 * 4318 * If it has been set by /etc/system, take that. 4319 * Otherwise, use a percentage of physical memory defined by 4320 * zfs_dirty_data_max_percent (default 10%) with a cap at 4321 * zfs_dirty_data_max_max (default 4GB). 4322 */ 4323 if (zfs_dirty_data_max == 0) { 4324 zfs_dirty_data_max = ptob(physmem) * 4325 zfs_dirty_data_max_percent / 100; 4326 zfs_dirty_data_max = MIN(zfs_dirty_data_max, 4327 zfs_dirty_data_max_max); 4328 } 4329 4330#ifdef _KERNEL 4331 if (TUNABLE_INT_FETCH("vfs.zfs.prefetch_disable", &zfs_prefetch_disable)) 4332 prefetch_tunable_set = 1; 4333 4334#ifdef __i386__ 4335 if (prefetch_tunable_set == 0) { 4336 printf("ZFS NOTICE: Prefetch is disabled by default on i386 " 4337 "-- to enable,\n"); 4338 printf(" add \"vfs.zfs.prefetch_disable=0\" " 4339 "to /boot/loader.conf.\n"); 4340 zfs_prefetch_disable = 1; 4341 } 4342#else 4343 if ((((uint64_t)physmem * PAGESIZE) < (1ULL << 32)) && 4344 prefetch_tunable_set == 0) { 4345 printf("ZFS NOTICE: Prefetch is disabled by default if less " 4346 "than 4GB of RAM is present;\n" 4347 " to enable, add \"vfs.zfs.prefetch_disable=0\" " 4348 "to /boot/loader.conf.\n"); 4349 zfs_prefetch_disable = 1; 4350 } 4351#endif 4352 /* Warn about ZFS memory and address space requirements. */ 4353 if (((uint64_t)physmem * PAGESIZE) < (256 + 128 + 64) * (1 << 20)) { 4354 printf("ZFS WARNING: Recommended minimum RAM size is 512MB; " 4355 "expect unstable behavior.\n"); 4356 } 4357 if (kmem_size() < 512 * (1 << 20)) { 4358 printf("ZFS WARNING: Recommended minimum kmem_size is 512MB; " 4359 "expect unstable behavior.\n"); 4360 printf(" Consider tuning vm.kmem_size and " 4361 "vm.kmem_size_max\n"); 4362 printf(" in /boot/loader.conf.\n"); 4363 } 4364#endif 4365} 4366 4367void 4368arc_fini(void) 4369{ 4370 int i; 4371 4372 mutex_enter(&arc_reclaim_thr_lock); 4373 arc_thread_exit = 1; 4374 cv_signal(&arc_reclaim_thr_cv); 4375 while (arc_thread_exit != 0) 4376 cv_wait(&arc_reclaim_thr_cv, &arc_reclaim_thr_lock); 4377 mutex_exit(&arc_reclaim_thr_lock); 4378 4379 arc_flush(NULL); 4380 4381 arc_dead = TRUE; 4382 4383 if (arc_ksp != NULL) { 4384 kstat_delete(arc_ksp); 4385 arc_ksp = NULL; 4386 } 4387 4388 mutex_destroy(&arc_eviction_mtx); 4389 mutex_destroy(&arc_reclaim_thr_lock); 4390 cv_destroy(&arc_reclaim_thr_cv); 4391 4392 for (i = 0; i < ARC_BUFC_NUMLISTS; i++) { 4393 list_destroy(&arc_mru->arcs_lists[i]); 4394 list_destroy(&arc_mru_ghost->arcs_lists[i]); 4395 list_destroy(&arc_mfu->arcs_lists[i]); 4396 list_destroy(&arc_mfu_ghost->arcs_lists[i]); 4397 list_destroy(&arc_l2c_only->arcs_lists[i]); 4398 4399 mutex_destroy(&arc_anon->arcs_locks[i].arcs_lock); 4400 mutex_destroy(&arc_mru->arcs_locks[i].arcs_lock); 4401 mutex_destroy(&arc_mru_ghost->arcs_locks[i].arcs_lock); 4402 mutex_destroy(&arc_mfu->arcs_locks[i].arcs_lock); 4403 mutex_destroy(&arc_mfu_ghost->arcs_locks[i].arcs_lock); 4404 mutex_destroy(&arc_l2c_only->arcs_locks[i].arcs_lock); 4405 } 4406 4407 buf_fini(); 4408 4409 ASSERT(arc_loaned_bytes == 0); 4410 4411 mutex_destroy(&arc_lowmem_lock); 4412#ifdef _KERNEL 4413 if (arc_event_lowmem != NULL) 4414 EVENTHANDLER_DEREGISTER(vm_lowmem, arc_event_lowmem); 4415#endif 4416} 4417 4418/* 4419 * Level 2 ARC 4420 * 4421 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk. 4422 * It uses dedicated storage devices to hold cached data, which are populated 4423 * using large infrequent writes. The main role of this cache is to boost 4424 * the performance of random read workloads. The intended L2ARC devices 4425 * include short-stroked disks, solid state disks, and other media with 4426 * substantially faster read latency than disk. 4427 * 4428 * +-----------------------+ 4429 * | ARC | 4430 * +-----------------------+ 4431 * | ^ ^ 4432 * | | | 4433 * l2arc_feed_thread() arc_read() 4434 * | | | 4435 * | l2arc read | 4436 * V | | 4437 * +---------------+ | 4438 * | L2ARC | | 4439 * +---------------+ | 4440 * | ^ | 4441 * l2arc_write() | | 4442 * | | | 4443 * V | | 4444 * +-------+ +-------+ 4445 * | vdev | | vdev | 4446 * | cache | | cache | 4447 * +-------+ +-------+ 4448 * +=========+ .-----. 4449 * : L2ARC : |-_____-| 4450 * : devices : | Disks | 4451 * +=========+ `-_____-' 4452 * 4453 * Read requests are satisfied from the following sources, in order: 4454 * 4455 * 1) ARC 4456 * 2) vdev cache of L2ARC devices 4457 * 3) L2ARC devices 4458 * 4) vdev cache of disks 4459 * 5) disks 4460 * 4461 * Some L2ARC device types exhibit extremely slow write performance. 4462 * To accommodate for this there are some significant differences between 4463 * the L2ARC and traditional cache design: 4464 * 4465 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from 4466 * the ARC behave as usual, freeing buffers and placing headers on ghost 4467 * lists. The ARC does not send buffers to the L2ARC during eviction as 4468 * this would add inflated write latencies for all ARC memory pressure. 4469 * 4470 * 2. The L2ARC attempts to cache data from the ARC before it is evicted. 4471 * It does this by periodically scanning buffers from the eviction-end of 4472 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are 4473 * not already there. It scans until a headroom of buffers is satisfied, 4474 * which itself is a buffer for ARC eviction. If a compressible buffer is 4475 * found during scanning and selected for writing to an L2ARC device, we 4476 * temporarily boost scanning headroom during the next scan cycle to make 4477 * sure we adapt to compression effects (which might significantly reduce 4478 * the data volume we write to L2ARC). The thread that does this is 4479 * l2arc_feed_thread(), illustrated below; example sizes are included to 4480 * provide a better sense of ratio than this diagram: 4481 * 4482 * head --> tail 4483 * +---------------------+----------+ 4484 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC 4485 * +---------------------+----------+ | o L2ARC eligible 4486 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer 4487 * +---------------------+----------+ | 4488 * 15.9 Gbytes ^ 32 Mbytes | 4489 * headroom | 4490 * l2arc_feed_thread() 4491 * | 4492 * l2arc write hand <--[oooo]--' 4493 * | 8 Mbyte 4494 * | write max 4495 * V 4496 * +==============================+ 4497 * L2ARC dev |####|#|###|###| |####| ... | 4498 * +==============================+ 4499 * 32 Gbytes 4500 * 4501 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of 4502 * evicted, then the L2ARC has cached a buffer much sooner than it probably 4503 * needed to, potentially wasting L2ARC device bandwidth and storage. It is 4504 * safe to say that this is an uncommon case, since buffers at the end of 4505 * the ARC lists have moved there due to inactivity. 4506 * 4507 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom, 4508 * then the L2ARC simply misses copying some buffers. This serves as a 4509 * pressure valve to prevent heavy read workloads from both stalling the ARC 4510 * with waits and clogging the L2ARC with writes. This also helps prevent 4511 * the potential for the L2ARC to churn if it attempts to cache content too 4512 * quickly, such as during backups of the entire pool. 4513 * 4514 * 5. After system boot and before the ARC has filled main memory, there are 4515 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru 4516 * lists can remain mostly static. Instead of searching from tail of these 4517 * lists as pictured, the l2arc_feed_thread() will search from the list heads 4518 * for eligible buffers, greatly increasing its chance of finding them. 4519 * 4520 * The L2ARC device write speed is also boosted during this time so that 4521 * the L2ARC warms up faster. Since there have been no ARC evictions yet, 4522 * there are no L2ARC reads, and no fear of degrading read performance 4523 * through increased writes. 4524 * 4525 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that 4526 * the vdev queue can aggregate them into larger and fewer writes. Each 4527 * device is written to in a rotor fashion, sweeping writes through 4528 * available space then repeating. 4529 * 4530 * 7. The L2ARC does not store dirty content. It never needs to flush 4531 * write buffers back to disk based storage. 4532 * 4533 * 8. If an ARC buffer is written (and dirtied) which also exists in the 4534 * L2ARC, the now stale L2ARC buffer is immediately dropped. 4535 * 4536 * The performance of the L2ARC can be tweaked by a number of tunables, which 4537 * may be necessary for different workloads: 4538 * 4539 * l2arc_write_max max write bytes per interval 4540 * l2arc_write_boost extra write bytes during device warmup 4541 * l2arc_noprefetch skip caching prefetched buffers 4542 * l2arc_headroom number of max device writes to precache 4543 * l2arc_headroom_boost when we find compressed buffers during ARC 4544 * scanning, we multiply headroom by this 4545 * percentage factor for the next scan cycle, 4546 * since more compressed buffers are likely to 4547 * be present 4548 * l2arc_feed_secs seconds between L2ARC writing 4549 * 4550 * Tunables may be removed or added as future performance improvements are 4551 * integrated, and also may become zpool properties. 4552 * 4553 * There are three key functions that control how the L2ARC warms up: 4554 * 4555 * l2arc_write_eligible() check if a buffer is eligible to cache 4556 * l2arc_write_size() calculate how much to write 4557 * l2arc_write_interval() calculate sleep delay between writes 4558 * 4559 * These three functions determine what to write, how much, and how quickly 4560 * to send writes. 4561 */ 4562 4563static boolean_t 4564l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *hdr) 4565{ 4566 /* 4567 * A buffer is *not* eligible for the L2ARC if it: 4568 * 1. belongs to a different spa. 4569 * 2. is already cached on the L2ARC. 4570 * 3. has an I/O in progress (it may be an incomplete read). 4571 * 4. is flagged not eligible (zfs property). 4572 */ 4573 if (hdr->b_spa != spa_guid) { 4574 ARCSTAT_BUMP(arcstat_l2_write_spa_mismatch); 4575 return (B_FALSE); 4576 } 4577 if (hdr->b_l2hdr != NULL) { 4578 ARCSTAT_BUMP(arcstat_l2_write_in_l2); 4579 return (B_FALSE); 4580 } 4581 if (HDR_IO_IN_PROGRESS(hdr)) { 4582 ARCSTAT_BUMP(arcstat_l2_write_hdr_io_in_progress); 4583 return (B_FALSE); 4584 } 4585 if (!HDR_L2CACHE(hdr)) { 4586 ARCSTAT_BUMP(arcstat_l2_write_not_cacheable); 4587 return (B_FALSE); 4588 } 4589 4590 return (B_TRUE); 4591} 4592 4593static uint64_t 4594l2arc_write_size(void) 4595{ 4596 uint64_t size; 4597 4598 /* 4599 * Make sure our globals have meaningful values in case the user 4600 * altered them. 4601 */ 4602 size = l2arc_write_max; 4603 if (size == 0) { 4604 cmn_err(CE_NOTE, "Bad value for l2arc_write_max, value must " 4605 "be greater than zero, resetting it to the default (%d)", 4606 L2ARC_WRITE_SIZE); 4607 size = l2arc_write_max = L2ARC_WRITE_SIZE; 4608 } 4609 4610 if (arc_warm == B_FALSE) 4611 size += l2arc_write_boost; 4612 4613 return (size); 4614 4615} 4616 4617static clock_t 4618l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote) 4619{ 4620 clock_t interval, next, now; 4621 4622 /* 4623 * If the ARC lists are busy, increase our write rate; if the 4624 * lists are stale, idle back. This is achieved by checking 4625 * how much we previously wrote - if it was more than half of 4626 * what we wanted, schedule the next write much sooner. 4627 */ 4628 if (l2arc_feed_again && wrote > (wanted / 2)) 4629 interval = (hz * l2arc_feed_min_ms) / 1000; 4630 else 4631 interval = hz * l2arc_feed_secs; 4632 4633 now = ddi_get_lbolt(); 4634 next = MAX(now, MIN(now + interval, began + interval)); 4635 4636 return (next); 4637} 4638 4639static void 4640l2arc_hdr_stat_add(void) 4641{ 4642 ARCSTAT_INCR(arcstat_l2_hdr_size, HDR_SIZE + L2HDR_SIZE); 4643 ARCSTAT_INCR(arcstat_hdr_size, -HDR_SIZE); 4644} 4645 4646static void 4647l2arc_hdr_stat_remove(void) 4648{ 4649 ARCSTAT_INCR(arcstat_l2_hdr_size, -(HDR_SIZE + L2HDR_SIZE)); 4650 ARCSTAT_INCR(arcstat_hdr_size, HDR_SIZE); 4651} 4652 4653/* 4654 * Cycle through L2ARC devices. This is how L2ARC load balances. 4655 * If a device is returned, this also returns holding the spa config lock. 4656 */ 4657static l2arc_dev_t * 4658l2arc_dev_get_next(void) 4659{ 4660 l2arc_dev_t *first, *next = NULL; 4661 4662 /* 4663 * Lock out the removal of spas (spa_namespace_lock), then removal 4664 * of cache devices (l2arc_dev_mtx). Once a device has been selected, 4665 * both locks will be dropped and a spa config lock held instead. 4666 */ 4667 mutex_enter(&spa_namespace_lock); 4668 mutex_enter(&l2arc_dev_mtx); 4669 4670 /* if there are no vdevs, there is nothing to do */ 4671 if (l2arc_ndev == 0) 4672 goto out; 4673 4674 first = NULL; 4675 next = l2arc_dev_last; 4676 do { 4677 /* loop around the list looking for a non-faulted vdev */ 4678 if (next == NULL) { 4679 next = list_head(l2arc_dev_list); 4680 } else { 4681 next = list_next(l2arc_dev_list, next); 4682 if (next == NULL) 4683 next = list_head(l2arc_dev_list); 4684 } 4685 4686 /* if we have come back to the start, bail out */ 4687 if (first == NULL) 4688 first = next; 4689 else if (next == first) 4690 break; 4691 4692 } while (vdev_is_dead(next->l2ad_vdev)); 4693 4694 /* if we were unable to find any usable vdevs, return NULL */ 4695 if (vdev_is_dead(next->l2ad_vdev)) 4696 next = NULL; 4697 4698 l2arc_dev_last = next; 4699 4700out: 4701 mutex_exit(&l2arc_dev_mtx); 4702 4703 /* 4704 * Grab the config lock to prevent the 'next' device from being 4705 * removed while we are writing to it. 4706 */ 4707 if (next != NULL) 4708 spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER); 4709 mutex_exit(&spa_namespace_lock); 4710 4711 return (next); 4712} 4713 4714/* 4715 * Free buffers that were tagged for destruction. 4716 */ 4717static void 4718l2arc_do_free_on_write() 4719{ 4720 list_t *buflist; 4721 l2arc_data_free_t *df, *df_prev; 4722 4723 mutex_enter(&l2arc_free_on_write_mtx); 4724 buflist = l2arc_free_on_write; 4725 4726 for (df = list_tail(buflist); df; df = df_prev) { 4727 df_prev = list_prev(buflist, df); 4728 ASSERT(df->l2df_data != NULL); 4729 ASSERT(df->l2df_func != NULL); 4730 df->l2df_func(df->l2df_data, df->l2df_size); 4731 list_remove(buflist, df); 4732 kmem_free(df, sizeof (l2arc_data_free_t)); 4733 } 4734 4735 mutex_exit(&l2arc_free_on_write_mtx); 4736} 4737 4738/* 4739 * A write to a cache device has completed. Update all headers to allow 4740 * reads from these buffers to begin. 4741 */ 4742static void 4743l2arc_write_done(zio_t *zio) 4744{ 4745 l2arc_write_callback_t *cb; 4746 l2arc_dev_t *dev; 4747 list_t *buflist; 4748 arc_buf_hdr_t *head, *hdr, *hdr_prev; 4749 l2arc_buf_hdr_t *abl2; 4750 kmutex_t *hash_lock; 4751 int64_t bytes_dropped = 0; 4752 4753 cb = zio->io_private; 4754 ASSERT(cb != NULL); 4755 dev = cb->l2wcb_dev; 4756 ASSERT(dev != NULL); 4757 head = cb->l2wcb_head; 4758 ASSERT(head != NULL); 4759 buflist = dev->l2ad_buflist; 4760 ASSERT(buflist != NULL); 4761 DTRACE_PROBE2(l2arc__iodone, zio_t *, zio, 4762 l2arc_write_callback_t *, cb); 4763 4764 if (zio->io_error != 0) 4765 ARCSTAT_BUMP(arcstat_l2_writes_error); 4766 4767 mutex_enter(&l2arc_buflist_mtx); 4768 4769 /* 4770 * All writes completed, or an error was hit. 4771 */ 4772 for (hdr = list_prev(buflist, head); hdr; hdr = hdr_prev) { 4773 hdr_prev = list_prev(buflist, hdr); 4774 abl2 = hdr->b_l2hdr; 4775 4776 /* 4777 * Release the temporary compressed buffer as soon as possible. 4778 */ 4779 if (abl2->b_compress != ZIO_COMPRESS_OFF) 4780 l2arc_release_cdata_buf(hdr); 4781 4782 hash_lock = HDR_LOCK(hdr); 4783 if (!mutex_tryenter(hash_lock)) { 4784 /* 4785 * This buffer misses out. It may be in a stage 4786 * of eviction. Its ARC_L2_WRITING flag will be 4787 * left set, denying reads to this buffer. 4788 */ 4789 ARCSTAT_BUMP(arcstat_l2_writes_hdr_miss); 4790 continue; 4791 } 4792 4793 if (zio->io_error != 0) { 4794 /* 4795 * Error - drop L2ARC entry. 4796 */ 4797 list_remove(buflist, hdr); 4798 ARCSTAT_INCR(arcstat_l2_asize, -abl2->b_asize); 4799 bytes_dropped += abl2->b_asize; 4800 hdr->b_l2hdr = NULL; 4801 trim_map_free(abl2->b_dev->l2ad_vdev, abl2->b_daddr, 4802 hdr->b_size, 0); 4803 kmem_free(abl2, sizeof (l2arc_buf_hdr_t)); 4804 ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size); 4805 } 4806 4807 /* 4808 * Allow ARC to begin reads to this L2ARC entry. 4809 */ 4810 hdr->b_flags &= ~ARC_FLAG_L2_WRITING; 4811 4812 mutex_exit(hash_lock); 4813 } 4814 4815 atomic_inc_64(&l2arc_writes_done); 4816 list_remove(buflist, head); 4817 kmem_cache_free(hdr_cache, head); 4818 mutex_exit(&l2arc_buflist_mtx); 4819 4820 vdev_space_update(dev->l2ad_vdev, -bytes_dropped, 0, 0); 4821 4822 l2arc_do_free_on_write(); 4823 4824 kmem_free(cb, sizeof (l2arc_write_callback_t)); 4825} 4826 4827/* 4828 * A read to a cache device completed. Validate buffer contents before 4829 * handing over to the regular ARC routines. 4830 */ 4831static void 4832l2arc_read_done(zio_t *zio) 4833{ 4834 l2arc_read_callback_t *cb; 4835 arc_buf_hdr_t *hdr; 4836 arc_buf_t *buf; 4837 kmutex_t *hash_lock; 4838 int equal; 4839 4840 ASSERT(zio->io_vd != NULL); 4841 ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE); 4842 4843 spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd); 4844 4845 cb = zio->io_private; 4846 ASSERT(cb != NULL); 4847 buf = cb->l2rcb_buf; 4848 ASSERT(buf != NULL); 4849 4850 hash_lock = HDR_LOCK(buf->b_hdr); 4851 mutex_enter(hash_lock); 4852 hdr = buf->b_hdr; 4853 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 4854 4855 /* 4856 * If the buffer was compressed, decompress it first. 4857 */ 4858 if (cb->l2rcb_compress != ZIO_COMPRESS_OFF) 4859 l2arc_decompress_zio(zio, hdr, cb->l2rcb_compress); 4860 ASSERT(zio->io_data != NULL); 4861 4862 /* 4863 * Check this survived the L2ARC journey. 4864 */ 4865 equal = arc_cksum_equal(buf); 4866 if (equal && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) { 4867 mutex_exit(hash_lock); 4868 zio->io_private = buf; 4869 zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */ 4870 zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */ 4871 arc_read_done(zio); 4872 } else { 4873 mutex_exit(hash_lock); 4874 /* 4875 * Buffer didn't survive caching. Increment stats and 4876 * reissue to the original storage device. 4877 */ 4878 if (zio->io_error != 0) { 4879 ARCSTAT_BUMP(arcstat_l2_io_error); 4880 } else { 4881 zio->io_error = SET_ERROR(EIO); 4882 } 4883 if (!equal) 4884 ARCSTAT_BUMP(arcstat_l2_cksum_bad); 4885 4886 /* 4887 * If there's no waiter, issue an async i/o to the primary 4888 * storage now. If there *is* a waiter, the caller must 4889 * issue the i/o in a context where it's OK to block. 4890 */ 4891 if (zio->io_waiter == NULL) { 4892 zio_t *pio = zio_unique_parent(zio); 4893 4894 ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL); 4895 4896 zio_nowait(zio_read(pio, cb->l2rcb_spa, &cb->l2rcb_bp, 4897 buf->b_data, zio->io_size, arc_read_done, buf, 4898 zio->io_priority, cb->l2rcb_flags, &cb->l2rcb_zb)); 4899 } 4900 } 4901 4902 kmem_free(cb, sizeof (l2arc_read_callback_t)); 4903} 4904 4905/* 4906 * This is the list priority from which the L2ARC will search for pages to 4907 * cache. This is used within loops (0..3) to cycle through lists in the 4908 * desired order. This order can have a significant effect on cache 4909 * performance. 4910 * 4911 * Currently the metadata lists are hit first, MFU then MRU, followed by 4912 * the data lists. This function returns a locked list, and also returns 4913 * the lock pointer. 4914 */ 4915static list_t * 4916l2arc_list_locked(int list_num, kmutex_t **lock) 4917{ 4918 list_t *list = NULL; 4919 int idx; 4920 4921 ASSERT(list_num >= 0 && list_num < 2 * ARC_BUFC_NUMLISTS); 4922 4923 if (list_num < ARC_BUFC_NUMMETADATALISTS) { 4924 idx = list_num; 4925 list = &arc_mfu->arcs_lists[idx]; 4926 *lock = ARCS_LOCK(arc_mfu, idx); 4927 } else if (list_num < ARC_BUFC_NUMMETADATALISTS * 2) { 4928 idx = list_num - ARC_BUFC_NUMMETADATALISTS; 4929 list = &arc_mru->arcs_lists[idx]; 4930 *lock = ARCS_LOCK(arc_mru, idx); 4931 } else if (list_num < (ARC_BUFC_NUMMETADATALISTS * 2 + 4932 ARC_BUFC_NUMDATALISTS)) { 4933 idx = list_num - ARC_BUFC_NUMMETADATALISTS; 4934 list = &arc_mfu->arcs_lists[idx]; 4935 *lock = ARCS_LOCK(arc_mfu, idx); 4936 } else { 4937 idx = list_num - ARC_BUFC_NUMLISTS; 4938 list = &arc_mru->arcs_lists[idx]; 4939 *lock = ARCS_LOCK(arc_mru, idx); 4940 } 4941 4942 ASSERT(!(MUTEX_HELD(*lock))); 4943 mutex_enter(*lock); 4944 return (list); 4945} 4946 4947/* 4948 * Evict buffers from the device write hand to the distance specified in 4949 * bytes. This distance may span populated buffers, it may span nothing. 4950 * This is clearing a region on the L2ARC device ready for writing. 4951 * If the 'all' boolean is set, every buffer is evicted. 4952 */ 4953static void 4954l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all) 4955{ 4956 list_t *buflist; 4957 l2arc_buf_hdr_t *abl2; 4958 arc_buf_hdr_t *hdr, *hdr_prev; 4959 kmutex_t *hash_lock; 4960 uint64_t taddr; 4961 int64_t bytes_evicted = 0; 4962 4963 buflist = dev->l2ad_buflist; 4964 4965 if (buflist == NULL) 4966 return; 4967 4968 if (!all && dev->l2ad_first) { 4969 /* 4970 * This is the first sweep through the device. There is 4971 * nothing to evict. 4972 */ 4973 return; 4974 } 4975 4976 if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) { 4977 /* 4978 * When nearing the end of the device, evict to the end 4979 * before the device write hand jumps to the start. 4980 */ 4981 taddr = dev->l2ad_end; 4982 } else { 4983 taddr = dev->l2ad_hand + distance; 4984 } 4985 DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist, 4986 uint64_t, taddr, boolean_t, all); 4987 4988top: 4989 mutex_enter(&l2arc_buflist_mtx); 4990 for (hdr = list_tail(buflist); hdr; hdr = hdr_prev) { 4991 hdr_prev = list_prev(buflist, hdr); 4992 4993 hash_lock = HDR_LOCK(hdr); 4994 if (!mutex_tryenter(hash_lock)) { 4995 /* 4996 * Missed the hash lock. Retry. 4997 */ 4998 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry); 4999 mutex_exit(&l2arc_buflist_mtx); 5000 mutex_enter(hash_lock); 5001 mutex_exit(hash_lock); 5002 goto top; 5003 } 5004 5005 if (HDR_L2_WRITE_HEAD(hdr)) { 5006 /* 5007 * We hit a write head node. Leave it for 5008 * l2arc_write_done(). 5009 */ 5010 list_remove(buflist, hdr); 5011 mutex_exit(hash_lock); 5012 continue; 5013 } 5014 5015 if (!all && hdr->b_l2hdr != NULL && 5016 (hdr->b_l2hdr->b_daddr > taddr || 5017 hdr->b_l2hdr->b_daddr < dev->l2ad_hand)) { 5018 /* 5019 * We've evicted to the target address, 5020 * or the end of the device. 5021 */ 5022 mutex_exit(hash_lock); 5023 break; 5024 } 5025 5026 if (HDR_FREE_IN_PROGRESS(hdr)) { 5027 /* 5028 * Already on the path to destruction. 5029 */ 5030 mutex_exit(hash_lock); 5031 continue; 5032 } 5033 5034 if (hdr->b_state == arc_l2c_only) { 5035 ASSERT(!HDR_L2_READING(hdr)); 5036 /* 5037 * This doesn't exist in the ARC. Destroy. 5038 * arc_hdr_destroy() will call list_remove() 5039 * and decrement arcstat_l2_size. 5040 */ 5041 arc_change_state(arc_anon, hdr, hash_lock); 5042 arc_hdr_destroy(hdr); 5043 } else { 5044 /* 5045 * Invalidate issued or about to be issued 5046 * reads, since we may be about to write 5047 * over this location. 5048 */ 5049 if (HDR_L2_READING(hdr)) { 5050 ARCSTAT_BUMP(arcstat_l2_evict_reading); 5051 hdr->b_flags |= ARC_FLAG_L2_EVICTED; 5052 } 5053 5054 /* 5055 * Tell ARC this no longer exists in L2ARC. 5056 */ 5057 if (hdr->b_l2hdr != NULL) { 5058 abl2 = hdr->b_l2hdr; 5059 ARCSTAT_INCR(arcstat_l2_asize, -abl2->b_asize); 5060 bytes_evicted += abl2->b_asize; 5061 hdr->b_l2hdr = NULL; 5062 /* 5063 * We are destroying l2hdr, so ensure that 5064 * its compressed buffer, if any, is not leaked. 5065 */ 5066 ASSERT(abl2->b_tmp_cdata == NULL); 5067 kmem_free(abl2, sizeof (l2arc_buf_hdr_t)); 5068 ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size); 5069 } 5070 list_remove(buflist, hdr); 5071 5072 /* 5073 * This may have been leftover after a 5074 * failed write. 5075 */ 5076 hdr->b_flags &= ~ARC_FLAG_L2_WRITING; 5077 } 5078 mutex_exit(hash_lock); 5079 } 5080 mutex_exit(&l2arc_buflist_mtx); 5081 5082 vdev_space_update(dev->l2ad_vdev, -bytes_evicted, 0, 0); 5083 dev->l2ad_evict = taddr; 5084} 5085 5086/* 5087 * Find and write ARC buffers to the L2ARC device. 5088 * 5089 * An ARC_FLAG_L2_WRITING flag is set so that the L2ARC buffers are not valid 5090 * for reading until they have completed writing. 5091 * The headroom_boost is an in-out parameter used to maintain headroom boost 5092 * state between calls to this function. 5093 * 5094 * Returns the number of bytes actually written (which may be smaller than 5095 * the delta by which the device hand has changed due to alignment). 5096 */ 5097static uint64_t 5098l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz, 5099 boolean_t *headroom_boost) 5100{ 5101 arc_buf_hdr_t *hdr, *hdr_prev, *head; 5102 list_t *list; 5103 uint64_t write_asize, write_psize, write_sz, headroom, 5104 buf_compress_minsz; 5105 void *buf_data; 5106 kmutex_t *list_lock; 5107 boolean_t full; 5108 l2arc_write_callback_t *cb; 5109 zio_t *pio, *wzio; 5110 uint64_t guid = spa_load_guid(spa); 5111 const boolean_t do_headroom_boost = *headroom_boost; 5112 int try; 5113 5114 ASSERT(dev->l2ad_vdev != NULL); 5115 5116 /* Lower the flag now, we might want to raise it again later. */ 5117 *headroom_boost = B_FALSE; 5118 5119 pio = NULL; 5120 write_sz = write_asize = write_psize = 0; 5121 full = B_FALSE; 5122 head = kmem_cache_alloc(hdr_cache, KM_PUSHPAGE); 5123 head->b_flags |= ARC_FLAG_L2_WRITE_HEAD; 5124 5125 ARCSTAT_BUMP(arcstat_l2_write_buffer_iter); 5126 /* 5127 * We will want to try to compress buffers that are at least 2x the 5128 * device sector size. 5129 */ 5130 buf_compress_minsz = 2 << dev->l2ad_vdev->vdev_ashift; 5131 5132 /* 5133 * Copy buffers for L2ARC writing. 5134 */ 5135 mutex_enter(&l2arc_buflist_mtx); 5136 for (try = 0; try < 2 * ARC_BUFC_NUMLISTS; try++) { 5137 uint64_t passed_sz = 0; 5138 5139 list = l2arc_list_locked(try, &list_lock); 5140 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_iter); 5141 5142 /* 5143 * L2ARC fast warmup. 5144 * 5145 * Until the ARC is warm and starts to evict, read from the 5146 * head of the ARC lists rather than the tail. 5147 */ 5148 if (arc_warm == B_FALSE) 5149 hdr = list_head(list); 5150 else 5151 hdr = list_tail(list); 5152 if (hdr == NULL) 5153 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_null_iter); 5154 5155 headroom = target_sz * l2arc_headroom * 2 / ARC_BUFC_NUMLISTS; 5156 if (do_headroom_boost) 5157 headroom = (headroom * l2arc_headroom_boost) / 100; 5158 5159 for (; hdr; hdr = hdr_prev) { 5160 l2arc_buf_hdr_t *l2hdr; 5161 kmutex_t *hash_lock; 5162 uint64_t buf_sz; 5163 5164 if (arc_warm == B_FALSE) 5165 hdr_prev = list_next(list, hdr); 5166 else 5167 hdr_prev = list_prev(list, hdr); 5168 ARCSTAT_INCR(arcstat_l2_write_buffer_bytes_scanned, hdr->b_size); 5169 5170 hash_lock = HDR_LOCK(hdr); 5171 if (!mutex_tryenter(hash_lock)) { 5172 ARCSTAT_BUMP(arcstat_l2_write_trylock_fail); 5173 /* 5174 * Skip this buffer rather than waiting. 5175 */ 5176 continue; 5177 } 5178 5179 passed_sz += hdr->b_size; 5180 if (passed_sz > headroom) { 5181 /* 5182 * Searched too far. 5183 */ 5184 mutex_exit(hash_lock); 5185 ARCSTAT_BUMP(arcstat_l2_write_passed_headroom); 5186 break; 5187 } 5188 5189 if (!l2arc_write_eligible(guid, hdr)) { 5190 mutex_exit(hash_lock); 5191 continue; 5192 } 5193 5194 if ((write_sz + hdr->b_size) > target_sz) { 5195 full = B_TRUE; 5196 mutex_exit(hash_lock); 5197 ARCSTAT_BUMP(arcstat_l2_write_full); 5198 break; 5199 } 5200 5201 if (pio == NULL) { 5202 /* 5203 * Insert a dummy header on the buflist so 5204 * l2arc_write_done() can find where the 5205 * write buffers begin without searching. 5206 */ 5207 list_insert_head(dev->l2ad_buflist, head); 5208 5209 cb = kmem_alloc( 5210 sizeof (l2arc_write_callback_t), KM_SLEEP); 5211 cb->l2wcb_dev = dev; 5212 cb->l2wcb_head = head; 5213 pio = zio_root(spa, l2arc_write_done, cb, 5214 ZIO_FLAG_CANFAIL); 5215 ARCSTAT_BUMP(arcstat_l2_write_pios); 5216 } 5217 5218 /* 5219 * Create and add a new L2ARC header. 5220 */ 5221 l2hdr = kmem_zalloc(sizeof (l2arc_buf_hdr_t), KM_SLEEP); 5222 l2hdr->b_dev = dev; 5223 hdr->b_flags |= ARC_FLAG_L2_WRITING; 5224 5225 /* 5226 * Temporarily stash the data buffer in b_tmp_cdata. 5227 * The subsequent write step will pick it up from 5228 * there. This is because can't access hdr->b_buf 5229 * without holding the hash_lock, which we in turn 5230 * can't access without holding the ARC list locks 5231 * (which we want to avoid during compression/writing). 5232 */ 5233 l2hdr->b_compress = ZIO_COMPRESS_OFF; 5234 l2hdr->b_asize = hdr->b_size; 5235 l2hdr->b_tmp_cdata = hdr->b_buf->b_data; 5236 5237 buf_sz = hdr->b_size; 5238 hdr->b_l2hdr = l2hdr; 5239 5240 list_insert_head(dev->l2ad_buflist, hdr); 5241 5242 /* 5243 * Compute and store the buffer cksum before 5244 * writing. On debug the cksum is verified first. 5245 */ 5246 arc_cksum_verify(hdr->b_buf); 5247 arc_cksum_compute(hdr->b_buf, B_TRUE); 5248 5249 mutex_exit(hash_lock); 5250 5251 write_sz += buf_sz; 5252 } 5253 5254 mutex_exit(list_lock); 5255 5256 if (full == B_TRUE) 5257 break; 5258 } 5259 5260 /* No buffers selected for writing? */ 5261 if (pio == NULL) { 5262 ASSERT0(write_sz); 5263 mutex_exit(&l2arc_buflist_mtx); 5264 kmem_cache_free(hdr_cache, head); 5265 return (0); 5266 } 5267 5268 /* 5269 * Now start writing the buffers. We're starting at the write head 5270 * and work backwards, retracing the course of the buffer selector 5271 * loop above. 5272 */ 5273 for (hdr = list_prev(dev->l2ad_buflist, head); hdr; 5274 hdr = list_prev(dev->l2ad_buflist, hdr)) { 5275 l2arc_buf_hdr_t *l2hdr; 5276 uint64_t buf_sz; 5277 5278 /* 5279 * We shouldn't need to lock the buffer here, since we flagged 5280 * it as ARC_FLAG_L2_WRITING in the previous step, but we must 5281 * take care to only access its L2 cache parameters. In 5282 * particular, hdr->b_buf may be invalid by now due to 5283 * ARC eviction. 5284 */ 5285 l2hdr = hdr->b_l2hdr; 5286 l2hdr->b_daddr = dev->l2ad_hand; 5287 5288 if ((hdr->b_flags & ARC_FLAG_L2COMPRESS) && 5289 l2hdr->b_asize >= buf_compress_minsz) { 5290 if (l2arc_compress_buf(l2hdr)) { 5291 /* 5292 * If compression succeeded, enable headroom 5293 * boost on the next scan cycle. 5294 */ 5295 *headroom_boost = B_TRUE; 5296 } 5297 } 5298 5299 /* 5300 * Pick up the buffer data we had previously stashed away 5301 * (and now potentially also compressed). 5302 */ 5303 buf_data = l2hdr->b_tmp_cdata; 5304 buf_sz = l2hdr->b_asize; 5305 5306 /* 5307 * If the data has not been compressed, then clear b_tmp_cdata 5308 * to make sure that it points only to a temporary compression 5309 * buffer. 5310 */ 5311 if (!L2ARC_IS_VALID_COMPRESS(l2hdr->b_compress)) 5312 l2hdr->b_tmp_cdata = NULL; 5313 5314 /* Compression may have squashed the buffer to zero length. */ 5315 if (buf_sz != 0) { 5316 uint64_t buf_p_sz; 5317 5318 wzio = zio_write_phys(pio, dev->l2ad_vdev, 5319 dev->l2ad_hand, buf_sz, buf_data, ZIO_CHECKSUM_OFF, 5320 NULL, NULL, ZIO_PRIORITY_ASYNC_WRITE, 5321 ZIO_FLAG_CANFAIL, B_FALSE); 5322 5323 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev, 5324 zio_t *, wzio); 5325 (void) zio_nowait(wzio); 5326 5327 write_asize += buf_sz; 5328 /* 5329 * Keep the clock hand suitably device-aligned. 5330 */ 5331 buf_p_sz = vdev_psize_to_asize(dev->l2ad_vdev, buf_sz); 5332 write_psize += buf_p_sz; 5333 dev->l2ad_hand += buf_p_sz; 5334 } 5335 } 5336 5337 mutex_exit(&l2arc_buflist_mtx); 5338 5339 ASSERT3U(write_asize, <=, target_sz); 5340 ARCSTAT_BUMP(arcstat_l2_writes_sent); 5341 ARCSTAT_INCR(arcstat_l2_write_bytes, write_asize); 5342 ARCSTAT_INCR(arcstat_l2_size, write_sz); 5343 ARCSTAT_INCR(arcstat_l2_asize, write_asize); 5344 vdev_space_update(dev->l2ad_vdev, write_asize, 0, 0); 5345 5346 /* 5347 * Bump device hand to the device start if it is approaching the end. 5348 * l2arc_evict() will already have evicted ahead for this case. 5349 */ 5350 if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) { 5351 dev->l2ad_hand = dev->l2ad_start; 5352 dev->l2ad_evict = dev->l2ad_start; 5353 dev->l2ad_first = B_FALSE; 5354 } 5355 5356 dev->l2ad_writing = B_TRUE; 5357 (void) zio_wait(pio); 5358 dev->l2ad_writing = B_FALSE; 5359 5360 return (write_asize); 5361} 5362 5363/* 5364 * Compresses an L2ARC buffer. 5365 * The data to be compressed must be prefilled in l2hdr->b_tmp_cdata and its 5366 * size in l2hdr->b_asize. This routine tries to compress the data and 5367 * depending on the compression result there are three possible outcomes: 5368 * *) The buffer was incompressible. The original l2hdr contents were left 5369 * untouched and are ready for writing to an L2 device. 5370 * *) The buffer was all-zeros, so there is no need to write it to an L2 5371 * device. To indicate this situation b_tmp_cdata is NULL'ed, b_asize is 5372 * set to zero and b_compress is set to ZIO_COMPRESS_EMPTY. 5373 * *) Compression succeeded and b_tmp_cdata was replaced with a temporary 5374 * data buffer which holds the compressed data to be written, and b_asize 5375 * tells us how much data there is. b_compress is set to the appropriate 5376 * compression algorithm. Once writing is done, invoke 5377 * l2arc_release_cdata_buf on this l2hdr to free this temporary buffer. 5378 * 5379 * Returns B_TRUE if compression succeeded, or B_FALSE if it didn't (the 5380 * buffer was incompressible). 5381 */ 5382static boolean_t 5383l2arc_compress_buf(l2arc_buf_hdr_t *l2hdr) 5384{ 5385 void *cdata; 5386 size_t csize, len, rounded; 5387 5388 ASSERT(l2hdr->b_compress == ZIO_COMPRESS_OFF); 5389 ASSERT(l2hdr->b_tmp_cdata != NULL); 5390 5391 len = l2hdr->b_asize; 5392 cdata = zio_data_buf_alloc(len); 5393 csize = zio_compress_data(ZIO_COMPRESS_LZ4, l2hdr->b_tmp_cdata, 5394 cdata, l2hdr->b_asize); 5395 5396 if (csize == 0) { 5397 /* zero block, indicate that there's nothing to write */ 5398 zio_data_buf_free(cdata, len); 5399 l2hdr->b_compress = ZIO_COMPRESS_EMPTY; 5400 l2hdr->b_asize = 0; 5401 l2hdr->b_tmp_cdata = NULL; 5402 ARCSTAT_BUMP(arcstat_l2_compress_zeros); 5403 return (B_TRUE); 5404 } 5405 5406 rounded = P2ROUNDUP(csize, 5407 (size_t)1 << l2hdr->b_dev->l2ad_vdev->vdev_ashift); 5408 if (rounded < len) { 5409 /* 5410 * Compression succeeded, we'll keep the cdata around for 5411 * writing and release it afterwards. 5412 */ 5413 if (rounded > csize) { 5414 bzero((char *)cdata + csize, rounded - csize); 5415 csize = rounded; 5416 } 5417 l2hdr->b_compress = ZIO_COMPRESS_LZ4; 5418 l2hdr->b_asize = csize; 5419 l2hdr->b_tmp_cdata = cdata; 5420 ARCSTAT_BUMP(arcstat_l2_compress_successes); 5421 return (B_TRUE); 5422 } else { 5423 /* 5424 * Compression failed, release the compressed buffer. 5425 * l2hdr will be left unmodified. 5426 */ 5427 zio_data_buf_free(cdata, len); 5428 ARCSTAT_BUMP(arcstat_l2_compress_failures); 5429 return (B_FALSE); 5430 } 5431} 5432 5433/* 5434 * Decompresses a zio read back from an l2arc device. On success, the 5435 * underlying zio's io_data buffer is overwritten by the uncompressed 5436 * version. On decompression error (corrupt compressed stream), the 5437 * zio->io_error value is set to signal an I/O error. 5438 * 5439 * Please note that the compressed data stream is not checksummed, so 5440 * if the underlying device is experiencing data corruption, we may feed 5441 * corrupt data to the decompressor, so the decompressor needs to be 5442 * able to handle this situation (LZ4 does). 5443 */ 5444static void 5445l2arc_decompress_zio(zio_t *zio, arc_buf_hdr_t *hdr, enum zio_compress c) 5446{ 5447 ASSERT(L2ARC_IS_VALID_COMPRESS(c)); 5448 5449 if (zio->io_error != 0) { 5450 /* 5451 * An io error has occured, just restore the original io 5452 * size in preparation for a main pool read. 5453 */ 5454 zio->io_orig_size = zio->io_size = hdr->b_size; 5455 return; 5456 } 5457 5458 if (c == ZIO_COMPRESS_EMPTY) { 5459 /* 5460 * An empty buffer results in a null zio, which means we 5461 * need to fill its io_data after we're done restoring the 5462 * buffer's contents. 5463 */ 5464 ASSERT(hdr->b_buf != NULL); 5465 bzero(hdr->b_buf->b_data, hdr->b_size); 5466 zio->io_data = zio->io_orig_data = hdr->b_buf->b_data; 5467 } else { 5468 ASSERT(zio->io_data != NULL); 5469 /* 5470 * We copy the compressed data from the start of the arc buffer 5471 * (the zio_read will have pulled in only what we need, the 5472 * rest is garbage which we will overwrite at decompression) 5473 * and then decompress back to the ARC data buffer. This way we 5474 * can minimize copying by simply decompressing back over the 5475 * original compressed data (rather than decompressing to an 5476 * aux buffer and then copying back the uncompressed buffer, 5477 * which is likely to be much larger). 5478 */ 5479 uint64_t csize; 5480 void *cdata; 5481 5482 csize = zio->io_size; 5483 cdata = zio_data_buf_alloc(csize); 5484 bcopy(zio->io_data, cdata, csize); 5485 if (zio_decompress_data(c, cdata, zio->io_data, csize, 5486 hdr->b_size) != 0) 5487 zio->io_error = EIO; 5488 zio_data_buf_free(cdata, csize); 5489 } 5490 5491 /* Restore the expected uncompressed IO size. */ 5492 zio->io_orig_size = zio->io_size = hdr->b_size; 5493} 5494 5495/* 5496 * Releases the temporary b_tmp_cdata buffer in an l2arc header structure. 5497 * This buffer serves as a temporary holder of compressed data while 5498 * the buffer entry is being written to an l2arc device. Once that is 5499 * done, we can dispose of it. 5500 */ 5501static void 5502l2arc_release_cdata_buf(arc_buf_hdr_t *hdr) 5503{ 5504 l2arc_buf_hdr_t *l2hdr = hdr->b_l2hdr; 5505 5506 ASSERT(L2ARC_IS_VALID_COMPRESS(l2hdr->b_compress)); 5507 if (l2hdr->b_compress != ZIO_COMPRESS_EMPTY) { 5508 /* 5509 * If the data was compressed, then we've allocated a 5510 * temporary buffer for it, so now we need to release it. 5511 */ 5512 ASSERT(l2hdr->b_tmp_cdata != NULL); 5513 zio_data_buf_free(l2hdr->b_tmp_cdata, hdr->b_size); 5514 l2hdr->b_tmp_cdata = NULL; 5515 } else { 5516 ASSERT(l2hdr->b_tmp_cdata == NULL); 5517 } 5518} 5519 5520/* 5521 * This thread feeds the L2ARC at regular intervals. This is the beating 5522 * heart of the L2ARC. 5523 */ 5524static void 5525l2arc_feed_thread(void *dummy __unused) 5526{ 5527 callb_cpr_t cpr; 5528 l2arc_dev_t *dev; 5529 spa_t *spa; 5530 uint64_t size, wrote; 5531 clock_t begin, next = ddi_get_lbolt(); 5532 boolean_t headroom_boost = B_FALSE; 5533 5534 CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG); 5535 5536 mutex_enter(&l2arc_feed_thr_lock); 5537 5538 while (l2arc_thread_exit == 0) { 5539 CALLB_CPR_SAFE_BEGIN(&cpr); 5540 (void) cv_timedwait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock, 5541 next - ddi_get_lbolt()); 5542 CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock); 5543 next = ddi_get_lbolt() + hz; 5544 5545 /* 5546 * Quick check for L2ARC devices. 5547 */ 5548 mutex_enter(&l2arc_dev_mtx); 5549 if (l2arc_ndev == 0) { 5550 mutex_exit(&l2arc_dev_mtx); 5551 continue; 5552 } 5553 mutex_exit(&l2arc_dev_mtx); 5554 begin = ddi_get_lbolt(); 5555 5556 /* 5557 * This selects the next l2arc device to write to, and in 5558 * doing so the next spa to feed from: dev->l2ad_spa. This 5559 * will return NULL if there are now no l2arc devices or if 5560 * they are all faulted. 5561 * 5562 * If a device is returned, its spa's config lock is also 5563 * held to prevent device removal. l2arc_dev_get_next() 5564 * will grab and release l2arc_dev_mtx. 5565 */ 5566 if ((dev = l2arc_dev_get_next()) == NULL) 5567 continue; 5568 5569 spa = dev->l2ad_spa; 5570 ASSERT(spa != NULL); 5571 5572 /* 5573 * If the pool is read-only then force the feed thread to 5574 * sleep a little longer. 5575 */ 5576 if (!spa_writeable(spa)) { 5577 next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz; 5578 spa_config_exit(spa, SCL_L2ARC, dev); 5579 continue; 5580 } 5581 5582 /* 5583 * Avoid contributing to memory pressure. 5584 */ 5585 if (arc_reclaim_needed()) { 5586 ARCSTAT_BUMP(arcstat_l2_abort_lowmem); 5587 spa_config_exit(spa, SCL_L2ARC, dev); 5588 continue; 5589 } 5590 5591 ARCSTAT_BUMP(arcstat_l2_feeds); 5592 5593 size = l2arc_write_size(); 5594 5595 /* 5596 * Evict L2ARC buffers that will be overwritten. 5597 */ 5598 l2arc_evict(dev, size, B_FALSE); 5599 5600 /* 5601 * Write ARC buffers. 5602 */ 5603 wrote = l2arc_write_buffers(spa, dev, size, &headroom_boost); 5604 5605 /* 5606 * Calculate interval between writes. 5607 */ 5608 next = l2arc_write_interval(begin, size, wrote); 5609 spa_config_exit(spa, SCL_L2ARC, dev); 5610 } 5611 5612 l2arc_thread_exit = 0; 5613 cv_broadcast(&l2arc_feed_thr_cv); 5614 CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */ 5615 thread_exit(); 5616} 5617 5618boolean_t 5619l2arc_vdev_present(vdev_t *vd) 5620{ 5621 l2arc_dev_t *dev; 5622 5623 mutex_enter(&l2arc_dev_mtx); 5624 for (dev = list_head(l2arc_dev_list); dev != NULL; 5625 dev = list_next(l2arc_dev_list, dev)) { 5626 if (dev->l2ad_vdev == vd) 5627 break; 5628 } 5629 mutex_exit(&l2arc_dev_mtx); 5630 5631 return (dev != NULL); 5632} 5633 5634/* 5635 * Add a vdev for use by the L2ARC. By this point the spa has already 5636 * validated the vdev and opened it. 5637 */ 5638void 5639l2arc_add_vdev(spa_t *spa, vdev_t *vd) 5640{ 5641 l2arc_dev_t *adddev; 5642 5643 ASSERT(!l2arc_vdev_present(vd)); 5644 5645 vdev_ashift_optimize(vd); 5646 5647 /* 5648 * Create a new l2arc device entry. 5649 */ 5650 adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP); 5651 adddev->l2ad_spa = spa; 5652 adddev->l2ad_vdev = vd; 5653 adddev->l2ad_start = VDEV_LABEL_START_SIZE; 5654 adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd); 5655 adddev->l2ad_hand = adddev->l2ad_start; 5656 adddev->l2ad_evict = adddev->l2ad_start; 5657 adddev->l2ad_first = B_TRUE; 5658 adddev->l2ad_writing = B_FALSE; 5659 5660 /* 5661 * This is a list of all ARC buffers that are still valid on the 5662 * device. 5663 */ 5664 adddev->l2ad_buflist = kmem_zalloc(sizeof (list_t), KM_SLEEP); 5665 list_create(adddev->l2ad_buflist, sizeof (arc_buf_hdr_t), 5666 offsetof(arc_buf_hdr_t, b_l2node)); 5667 5668 vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand); 5669 5670 /* 5671 * Add device to global list 5672 */ 5673 mutex_enter(&l2arc_dev_mtx); 5674 list_insert_head(l2arc_dev_list, adddev); 5675 atomic_inc_64(&l2arc_ndev); 5676 mutex_exit(&l2arc_dev_mtx); 5677} 5678 5679/* 5680 * Remove a vdev from the L2ARC. 5681 */ 5682void 5683l2arc_remove_vdev(vdev_t *vd) 5684{ 5685 l2arc_dev_t *dev, *nextdev, *remdev = NULL; 5686 5687 /* 5688 * Find the device by vdev 5689 */ 5690 mutex_enter(&l2arc_dev_mtx); 5691 for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) { 5692 nextdev = list_next(l2arc_dev_list, dev); 5693 if (vd == dev->l2ad_vdev) { 5694 remdev = dev; 5695 break; 5696 } 5697 } 5698 ASSERT(remdev != NULL); 5699 5700 /* 5701 * Remove device from global list 5702 */ 5703 list_remove(l2arc_dev_list, remdev); 5704 l2arc_dev_last = NULL; /* may have been invalidated */ 5705 atomic_dec_64(&l2arc_ndev); 5706 mutex_exit(&l2arc_dev_mtx); 5707 5708 /* 5709 * Clear all buflists and ARC references. L2ARC device flush. 5710 */ 5711 l2arc_evict(remdev, 0, B_TRUE); 5712 list_destroy(remdev->l2ad_buflist); 5713 kmem_free(remdev->l2ad_buflist, sizeof (list_t)); 5714 kmem_free(remdev, sizeof (l2arc_dev_t)); 5715} 5716 5717void 5718l2arc_init(void) 5719{ 5720 l2arc_thread_exit = 0; 5721 l2arc_ndev = 0; 5722 l2arc_writes_sent = 0; 5723 l2arc_writes_done = 0; 5724 5725 mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL); 5726 cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL); 5727 mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL); 5728 mutex_init(&l2arc_buflist_mtx, NULL, MUTEX_DEFAULT, NULL); 5729 mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL); 5730 5731 l2arc_dev_list = &L2ARC_dev_list; 5732 l2arc_free_on_write = &L2ARC_free_on_write; 5733 list_create(l2arc_dev_list, sizeof (l2arc_dev_t), 5734 offsetof(l2arc_dev_t, l2ad_node)); 5735 list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t), 5736 offsetof(l2arc_data_free_t, l2df_list_node)); 5737} 5738 5739void 5740l2arc_fini(void) 5741{ 5742 /* 5743 * This is called from dmu_fini(), which is called from spa_fini(); 5744 * Because of this, we can assume that all l2arc devices have 5745 * already been removed when the pools themselves were removed. 5746 */ 5747 5748 l2arc_do_free_on_write(); 5749 5750 mutex_destroy(&l2arc_feed_thr_lock); 5751 cv_destroy(&l2arc_feed_thr_cv); 5752 mutex_destroy(&l2arc_dev_mtx); 5753 mutex_destroy(&l2arc_buflist_mtx); 5754 mutex_destroy(&l2arc_free_on_write_mtx); 5755 5756 list_destroy(l2arc_dev_list); 5757 list_destroy(l2arc_free_on_write); 5758} 5759 5760void 5761l2arc_start(void) 5762{ 5763 if (!(spa_mode_global & FWRITE)) 5764 return; 5765 5766 (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0, 5767 TS_RUN, minclsyspri); 5768} 5769 5770void 5771l2arc_stop(void) 5772{ 5773 if (!(spa_mode_global & FWRITE)) 5774 return; 5775 5776 mutex_enter(&l2arc_feed_thr_lock); 5777 cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */ 5778 l2arc_thread_exit = 1; 5779 while (l2arc_thread_exit != 0) 5780 cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock); 5781 mutex_exit(&l2arc_feed_thr_lock); 5782} 5783