arc.c revision 288583
1/* 2 * CDDL HEADER START 3 * 4 * The contents of this file are subject to the terms of the 5 * Common Development and Distribution License (the "License"). 6 * You may not use this file except in compliance with the License. 7 * 8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE 9 * or http://www.opensolaris.org/os/licensing. 10 * See the License for the specific language governing permissions 11 * and limitations under the License. 12 * 13 * When distributing Covered Code, include this CDDL HEADER in each 14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE. 15 * If applicable, add the following below this CDDL HEADER, with the 16 * fields enclosed by brackets "[]" replaced with your own identifying 17 * information: Portions Copyright [yyyy] [name of copyright owner] 18 * 19 * CDDL HEADER END 20 */ 21/* 22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved. 23 * Copyright (c) 2012, Joyent, Inc. All rights reserved. 24 * Copyright (c) 2011, 2015 by Delphix. All rights reserved. 25 * Copyright (c) 2014 by Saso Kiselkov. All rights reserved. 26 * Copyright 2015 Nexenta Systems, Inc. All rights reserved. 27 */ 28 29/* 30 * DVA-based Adjustable Replacement Cache 31 * 32 * While much of the theory of operation used here is 33 * based on the self-tuning, low overhead replacement cache 34 * presented by Megiddo and Modha at FAST 2003, there are some 35 * significant differences: 36 * 37 * 1. The Megiddo and Modha model assumes any page is evictable. 38 * Pages in its cache cannot be "locked" into memory. This makes 39 * the eviction algorithm simple: evict the last page in the list. 40 * This also make the performance characteristics easy to reason 41 * about. Our cache is not so simple. At any given moment, some 42 * subset of the blocks in the cache are un-evictable because we 43 * have handed out a reference to them. Blocks are only evictable 44 * when there are no external references active. This makes 45 * eviction far more problematic: we choose to evict the evictable 46 * blocks that are the "lowest" in the list. 47 * 48 * There are times when it is not possible to evict the requested 49 * space. In these circumstances we are unable to adjust the cache 50 * size. To prevent the cache growing unbounded at these times we 51 * implement a "cache throttle" that slows the flow of new data 52 * into the cache until we can make space available. 53 * 54 * 2. The Megiddo and Modha model assumes a fixed cache size. 55 * Pages are evicted when the cache is full and there is a cache 56 * miss. Our model has a variable sized cache. It grows with 57 * high use, but also tries to react to memory pressure from the 58 * operating system: decreasing its size when system memory is 59 * tight. 60 * 61 * 3. The Megiddo and Modha model assumes a fixed page size. All 62 * elements of the cache are therefore exactly the same size. So 63 * when adjusting the cache size following a cache miss, its simply 64 * a matter of choosing a single page to evict. In our model, we 65 * have variable sized cache blocks (rangeing from 512 bytes to 66 * 128K bytes). We therefore choose a set of blocks to evict to make 67 * space for a cache miss that approximates as closely as possible 68 * the space used by the new block. 69 * 70 * See also: "ARC: A Self-Tuning, Low Overhead Replacement Cache" 71 * by N. Megiddo & D. Modha, FAST 2003 72 */ 73 74/* 75 * The locking model: 76 * 77 * A new reference to a cache buffer can be obtained in two 78 * ways: 1) via a hash table lookup using the DVA as a key, 79 * or 2) via one of the ARC lists. The arc_read() interface 80 * uses method 1, while the internal arc algorithms for 81 * adjusting the cache use method 2. We therefore provide two 82 * types of locks: 1) the hash table lock array, and 2) the 83 * arc list locks. 84 * 85 * Buffers do not have their own mutexs, rather they rely on the 86 * hash table mutexs for the bulk of their protection (i.e. most 87 * fields in the arc_buf_hdr_t are protected by these mutexs). 88 * 89 * buf_hash_find() returns the appropriate mutex (held) when it 90 * locates the requested buffer in the hash table. It returns 91 * NULL for the mutex if the buffer was not in the table. 92 * 93 * buf_hash_remove() expects the appropriate hash mutex to be 94 * already held before it is invoked. 95 * 96 * Each arc state also has a mutex which is used to protect the 97 * buffer list associated with the state. When attempting to 98 * obtain a hash table lock while holding an arc list lock you 99 * must use: mutex_tryenter() to avoid deadlock. Also note that 100 * the active state mutex must be held before the ghost state mutex. 101 * 102 * Arc buffers may have an associated eviction callback function. 103 * This function will be invoked prior to removing the buffer (e.g. 104 * in arc_do_user_evicts()). Note however that the data associated 105 * with the buffer may be evicted prior to the callback. The callback 106 * must be made with *no locks held* (to prevent deadlock). Additionally, 107 * the users of callbacks must ensure that their private data is 108 * protected from simultaneous callbacks from arc_clear_callback() 109 * and arc_do_user_evicts(). 110 * 111 * Note that the majority of the performance stats are manipulated 112 * with atomic operations. 113 * 114 * The L2ARC uses the l2ad_mtx on each vdev for the following: 115 * 116 * - L2ARC buflist creation 117 * - L2ARC buflist eviction 118 * - L2ARC write completion, which walks L2ARC buflists 119 * - ARC header destruction, as it removes from L2ARC buflists 120 * - ARC header release, as it removes from L2ARC buflists 121 */ 122 123#include <sys/spa.h> 124#include <sys/zio.h> 125#include <sys/zio_compress.h> 126#include <sys/zfs_context.h> 127#include <sys/arc.h> 128#include <sys/refcount.h> 129#include <sys/vdev.h> 130#include <sys/vdev_impl.h> 131#include <sys/dsl_pool.h> 132#include <sys/multilist.h> 133#ifdef _KERNEL 134#include <sys/dnlc.h> 135#endif 136#include <sys/callb.h> 137#include <sys/kstat.h> 138#include <sys/trim_map.h> 139#include <zfs_fletcher.h> 140#include <sys/sdt.h> 141 142#include <vm/vm_pageout.h> 143#include <machine/vmparam.h> 144 145#ifdef illumos 146#ifndef _KERNEL 147/* set with ZFS_DEBUG=watch, to enable watchpoints on frozen buffers */ 148boolean_t arc_watch = B_FALSE; 149int arc_procfd; 150#endif 151#endif /* illumos */ 152 153static kmutex_t arc_reclaim_lock; 154static kcondvar_t arc_reclaim_thread_cv; 155static boolean_t arc_reclaim_thread_exit; 156static kcondvar_t arc_reclaim_waiters_cv; 157 158static kmutex_t arc_user_evicts_lock; 159static kcondvar_t arc_user_evicts_cv; 160static boolean_t arc_user_evicts_thread_exit; 161 162uint_t arc_reduce_dnlc_percent = 3; 163 164/* 165 * The number of headers to evict in arc_evict_state_impl() before 166 * dropping the sublist lock and evicting from another sublist. A lower 167 * value means we're more likely to evict the "correct" header (i.e. the 168 * oldest header in the arc state), but comes with higher overhead 169 * (i.e. more invocations of arc_evict_state_impl()). 170 */ 171int zfs_arc_evict_batch_limit = 10; 172 173/* 174 * The number of sublists used for each of the arc state lists. If this 175 * is not set to a suitable value by the user, it will be configured to 176 * the number of CPUs on the system in arc_init(). 177 */ 178int zfs_arc_num_sublists_per_state = 0; 179 180/* number of seconds before growing cache again */ 181static int arc_grow_retry = 60; 182 183/* shift of arc_c for calculating overflow limit in arc_get_data_buf */ 184int zfs_arc_overflow_shift = 8; 185 186/* shift of arc_c for calculating both min and max arc_p */ 187static int arc_p_min_shift = 4; 188 189/* log2(fraction of arc to reclaim) */ 190static int arc_shrink_shift = 7; 191 192/* 193 * log2(fraction of ARC which must be free to allow growing). 194 * I.e. If there is less than arc_c >> arc_no_grow_shift free memory, 195 * when reading a new block into the ARC, we will evict an equal-sized block 196 * from the ARC. 197 * 198 * This must be less than arc_shrink_shift, so that when we shrink the ARC, 199 * we will still not allow it to grow. 200 */ 201int arc_no_grow_shift = 5; 202 203 204/* 205 * minimum lifespan of a prefetch block in clock ticks 206 * (initialized in arc_init()) 207 */ 208static int arc_min_prefetch_lifespan; 209 210/* 211 * If this percent of memory is free, don't throttle. 212 */ 213int arc_lotsfree_percent = 10; 214 215static int arc_dead; 216extern int zfs_prefetch_disable; 217 218/* 219 * The arc has filled available memory and has now warmed up. 220 */ 221static boolean_t arc_warm; 222 223/* 224 * These tunables are for performance analysis. 225 */ 226uint64_t zfs_arc_max; 227uint64_t zfs_arc_min; 228uint64_t zfs_arc_meta_limit = 0; 229uint64_t zfs_arc_meta_min = 0; 230int zfs_arc_grow_retry = 0; 231int zfs_arc_shrink_shift = 0; 232int zfs_arc_p_min_shift = 0; 233int zfs_disable_dup_eviction = 0; 234uint64_t zfs_arc_average_blocksize = 8 * 1024; /* 8KB */ 235u_int zfs_arc_free_target = 0; 236 237static int sysctl_vfs_zfs_arc_free_target(SYSCTL_HANDLER_ARGS); 238static int sysctl_vfs_zfs_arc_meta_limit(SYSCTL_HANDLER_ARGS); 239 240#ifdef _KERNEL 241static void 242arc_free_target_init(void *unused __unused) 243{ 244 245 zfs_arc_free_target = vm_pageout_wakeup_thresh; 246} 247SYSINIT(arc_free_target_init, SI_SUB_KTHREAD_PAGE, SI_ORDER_ANY, 248 arc_free_target_init, NULL); 249 250TUNABLE_QUAD("vfs.zfs.arc_max", &zfs_arc_max); 251TUNABLE_QUAD("vfs.zfs.arc_min", &zfs_arc_min); 252TUNABLE_QUAD("vfs.zfs.arc_meta_limit", &zfs_arc_meta_limit); 253TUNABLE_QUAD("vfs.zfs.arc_meta_min", &zfs_arc_meta_min); 254TUNABLE_QUAD("vfs.zfs.arc_average_blocksize", &zfs_arc_average_blocksize); 255TUNABLE_INT("vfs.zfs.arc_shrink_shift", &zfs_arc_shrink_shift); 256SYSCTL_DECL(_vfs_zfs); 257SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_max, CTLFLAG_RDTUN, &zfs_arc_max, 0, 258 "Maximum ARC size"); 259SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_min, CTLFLAG_RDTUN, &zfs_arc_min, 0, 260 "Minimum ARC size"); 261SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_average_blocksize, CTLFLAG_RDTUN, 262 &zfs_arc_average_blocksize, 0, 263 "ARC average blocksize"); 264SYSCTL_INT(_vfs_zfs, OID_AUTO, arc_shrink_shift, CTLFLAG_RW, 265 &arc_shrink_shift, 0, 266 "log2(fraction of arc to reclaim)"); 267 268/* 269 * We don't have a tunable for arc_free_target due to the dependency on 270 * pagedaemon initialisation. 271 */ 272SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_free_target, 273 CTLTYPE_UINT | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, sizeof(u_int), 274 sysctl_vfs_zfs_arc_free_target, "IU", 275 "Desired number of free pages below which ARC triggers reclaim"); 276 277static int 278sysctl_vfs_zfs_arc_free_target(SYSCTL_HANDLER_ARGS) 279{ 280 u_int val; 281 int err; 282 283 val = zfs_arc_free_target; 284 err = sysctl_handle_int(oidp, &val, 0, req); 285 if (err != 0 || req->newptr == NULL) 286 return (err); 287 288 if (val < minfree) 289 return (EINVAL); 290 if (val > cnt.v_page_count) 291 return (EINVAL); 292 293 zfs_arc_free_target = val; 294 295 return (0); 296} 297 298/* 299 * Must be declared here, before the definition of corresponding kstat 300 * macro which uses the same names will confuse the compiler. 301 */ 302SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_meta_limit, 303 CTLTYPE_U64 | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, sizeof(uint64_t), 304 sysctl_vfs_zfs_arc_meta_limit, "QU", 305 "ARC metadata limit"); 306#endif 307 308/* 309 * Note that buffers can be in one of 6 states: 310 * ARC_anon - anonymous (discussed below) 311 * ARC_mru - recently used, currently cached 312 * ARC_mru_ghost - recentely used, no longer in cache 313 * ARC_mfu - frequently used, currently cached 314 * ARC_mfu_ghost - frequently used, no longer in cache 315 * ARC_l2c_only - exists in L2ARC but not other states 316 * When there are no active references to the buffer, they are 317 * are linked onto a list in one of these arc states. These are 318 * the only buffers that can be evicted or deleted. Within each 319 * state there are multiple lists, one for meta-data and one for 320 * non-meta-data. Meta-data (indirect blocks, blocks of dnodes, 321 * etc.) is tracked separately so that it can be managed more 322 * explicitly: favored over data, limited explicitly. 323 * 324 * Anonymous buffers are buffers that are not associated with 325 * a DVA. These are buffers that hold dirty block copies 326 * before they are written to stable storage. By definition, 327 * they are "ref'd" and are considered part of arc_mru 328 * that cannot be freed. Generally, they will aquire a DVA 329 * as they are written and migrate onto the arc_mru list. 330 * 331 * The ARC_l2c_only state is for buffers that are in the second 332 * level ARC but no longer in any of the ARC_m* lists. The second 333 * level ARC itself may also contain buffers that are in any of 334 * the ARC_m* states - meaning that a buffer can exist in two 335 * places. The reason for the ARC_l2c_only state is to keep the 336 * buffer header in the hash table, so that reads that hit the 337 * second level ARC benefit from these fast lookups. 338 */ 339 340typedef struct arc_state { 341 /* 342 * list of evictable buffers 343 */ 344 multilist_t arcs_list[ARC_BUFC_NUMTYPES]; 345 /* 346 * total amount of evictable data in this state 347 */ 348 uint64_t arcs_lsize[ARC_BUFC_NUMTYPES]; 349 /* 350 * total amount of data in this state; this includes: evictable, 351 * non-evictable, ARC_BUFC_DATA, and ARC_BUFC_METADATA. 352 */ 353 refcount_t arcs_size; 354} arc_state_t; 355 356/* The 6 states: */ 357static arc_state_t ARC_anon; 358static arc_state_t ARC_mru; 359static arc_state_t ARC_mru_ghost; 360static arc_state_t ARC_mfu; 361static arc_state_t ARC_mfu_ghost; 362static arc_state_t ARC_l2c_only; 363 364typedef struct arc_stats { 365 kstat_named_t arcstat_hits; 366 kstat_named_t arcstat_misses; 367 kstat_named_t arcstat_demand_data_hits; 368 kstat_named_t arcstat_demand_data_misses; 369 kstat_named_t arcstat_demand_metadata_hits; 370 kstat_named_t arcstat_demand_metadata_misses; 371 kstat_named_t arcstat_prefetch_data_hits; 372 kstat_named_t arcstat_prefetch_data_misses; 373 kstat_named_t arcstat_prefetch_metadata_hits; 374 kstat_named_t arcstat_prefetch_metadata_misses; 375 kstat_named_t arcstat_mru_hits; 376 kstat_named_t arcstat_mru_ghost_hits; 377 kstat_named_t arcstat_mfu_hits; 378 kstat_named_t arcstat_mfu_ghost_hits; 379 kstat_named_t arcstat_allocated; 380 kstat_named_t arcstat_deleted; 381 /* 382 * Number of buffers that could not be evicted because the hash lock 383 * was held by another thread. The lock may not necessarily be held 384 * by something using the same buffer, since hash locks are shared 385 * by multiple buffers. 386 */ 387 kstat_named_t arcstat_mutex_miss; 388 /* 389 * Number of buffers skipped because they have I/O in progress, are 390 * indrect prefetch buffers that have not lived long enough, or are 391 * not from the spa we're trying to evict from. 392 */ 393 kstat_named_t arcstat_evict_skip; 394 /* 395 * Number of times arc_evict_state() was unable to evict enough 396 * buffers to reach it's target amount. 397 */ 398 kstat_named_t arcstat_evict_not_enough; 399 kstat_named_t arcstat_evict_l2_cached; 400 kstat_named_t arcstat_evict_l2_eligible; 401 kstat_named_t arcstat_evict_l2_ineligible; 402 kstat_named_t arcstat_evict_l2_skip; 403 kstat_named_t arcstat_hash_elements; 404 kstat_named_t arcstat_hash_elements_max; 405 kstat_named_t arcstat_hash_collisions; 406 kstat_named_t arcstat_hash_chains; 407 kstat_named_t arcstat_hash_chain_max; 408 kstat_named_t arcstat_p; 409 kstat_named_t arcstat_c; 410 kstat_named_t arcstat_c_min; 411 kstat_named_t arcstat_c_max; 412 kstat_named_t arcstat_size; 413 /* 414 * Number of bytes consumed by internal ARC structures necessary 415 * for tracking purposes; these structures are not actually 416 * backed by ARC buffers. This includes arc_buf_hdr_t structures 417 * (allocated via arc_buf_hdr_t_full and arc_buf_hdr_t_l2only 418 * caches), and arc_buf_t structures (allocated via arc_buf_t 419 * cache). 420 */ 421 kstat_named_t arcstat_hdr_size; 422 /* 423 * Number of bytes consumed by ARC buffers of type equal to 424 * ARC_BUFC_DATA. This is generally consumed by buffers backing 425 * on disk user data (e.g. plain file contents). 426 */ 427 kstat_named_t arcstat_data_size; 428 /* 429 * Number of bytes consumed by ARC buffers of type equal to 430 * ARC_BUFC_METADATA. This is generally consumed by buffers 431 * backing on disk data that is used for internal ZFS 432 * structures (e.g. ZAP, dnode, indirect blocks, etc). 433 */ 434 kstat_named_t arcstat_metadata_size; 435 /* 436 * Number of bytes consumed by various buffers and structures 437 * not actually backed with ARC buffers. This includes bonus 438 * buffers (allocated directly via zio_buf_* functions), 439 * dmu_buf_impl_t structures (allocated via dmu_buf_impl_t 440 * cache), and dnode_t structures (allocated via dnode_t cache). 441 */ 442 kstat_named_t arcstat_other_size; 443 /* 444 * Total number of bytes consumed by ARC buffers residing in the 445 * arc_anon state. This includes *all* buffers in the arc_anon 446 * state; e.g. data, metadata, evictable, and unevictable buffers 447 * are all included in this value. 448 */ 449 kstat_named_t arcstat_anon_size; 450 /* 451 * Number of bytes consumed by ARC buffers that meet the 452 * following criteria: backing buffers of type ARC_BUFC_DATA, 453 * residing in the arc_anon state, and are eligible for eviction 454 * (e.g. have no outstanding holds on the buffer). 455 */ 456 kstat_named_t arcstat_anon_evictable_data; 457 /* 458 * Number of bytes consumed by ARC buffers that meet the 459 * following criteria: backing buffers of type ARC_BUFC_METADATA, 460 * residing in the arc_anon state, and are eligible for eviction 461 * (e.g. have no outstanding holds on the buffer). 462 */ 463 kstat_named_t arcstat_anon_evictable_metadata; 464 /* 465 * Total number of bytes consumed by ARC buffers residing in the 466 * arc_mru state. This includes *all* buffers in the arc_mru 467 * state; e.g. data, metadata, evictable, and unevictable buffers 468 * are all included in this value. 469 */ 470 kstat_named_t arcstat_mru_size; 471 /* 472 * Number of bytes consumed by ARC buffers that meet the 473 * following criteria: backing buffers of type ARC_BUFC_DATA, 474 * residing in the arc_mru state, and are eligible for eviction 475 * (e.g. have no outstanding holds on the buffer). 476 */ 477 kstat_named_t arcstat_mru_evictable_data; 478 /* 479 * Number of bytes consumed by ARC buffers that meet the 480 * following criteria: backing buffers of type ARC_BUFC_METADATA, 481 * residing in the arc_mru state, and are eligible for eviction 482 * (e.g. have no outstanding holds on the buffer). 483 */ 484 kstat_named_t arcstat_mru_evictable_metadata; 485 /* 486 * Total number of bytes that *would have been* consumed by ARC 487 * buffers in the arc_mru_ghost state. The key thing to note 488 * here, is the fact that this size doesn't actually indicate 489 * RAM consumption. The ghost lists only consist of headers and 490 * don't actually have ARC buffers linked off of these headers. 491 * Thus, *if* the headers had associated ARC buffers, these 492 * buffers *would have* consumed this number of bytes. 493 */ 494 kstat_named_t arcstat_mru_ghost_size; 495 /* 496 * Number of bytes that *would have been* consumed by ARC 497 * buffers that are eligible for eviction, of type 498 * ARC_BUFC_DATA, and linked off the arc_mru_ghost state. 499 */ 500 kstat_named_t arcstat_mru_ghost_evictable_data; 501 /* 502 * Number of bytes that *would have been* consumed by ARC 503 * buffers that are eligible for eviction, of type 504 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state. 505 */ 506 kstat_named_t arcstat_mru_ghost_evictable_metadata; 507 /* 508 * Total number of bytes consumed by ARC buffers residing in the 509 * arc_mfu state. This includes *all* buffers in the arc_mfu 510 * state; e.g. data, metadata, evictable, and unevictable buffers 511 * are all included in this value. 512 */ 513 kstat_named_t arcstat_mfu_size; 514 /* 515 * Number of bytes consumed by ARC buffers that are eligible for 516 * eviction, of type ARC_BUFC_DATA, and reside in the arc_mfu 517 * state. 518 */ 519 kstat_named_t arcstat_mfu_evictable_data; 520 /* 521 * Number of bytes consumed by ARC buffers that are eligible for 522 * eviction, of type ARC_BUFC_METADATA, and reside in the 523 * arc_mfu state. 524 */ 525 kstat_named_t arcstat_mfu_evictable_metadata; 526 /* 527 * Total number of bytes that *would have been* consumed by ARC 528 * buffers in the arc_mfu_ghost state. See the comment above 529 * arcstat_mru_ghost_size for more details. 530 */ 531 kstat_named_t arcstat_mfu_ghost_size; 532 /* 533 * Number of bytes that *would have been* consumed by ARC 534 * buffers that are eligible for eviction, of type 535 * ARC_BUFC_DATA, and linked off the arc_mfu_ghost state. 536 */ 537 kstat_named_t arcstat_mfu_ghost_evictable_data; 538 /* 539 * Number of bytes that *would have been* consumed by ARC 540 * buffers that are eligible for eviction, of type 541 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state. 542 */ 543 kstat_named_t arcstat_mfu_ghost_evictable_metadata; 544 kstat_named_t arcstat_l2_hits; 545 kstat_named_t arcstat_l2_misses; 546 kstat_named_t arcstat_l2_feeds; 547 kstat_named_t arcstat_l2_rw_clash; 548 kstat_named_t arcstat_l2_read_bytes; 549 kstat_named_t arcstat_l2_write_bytes; 550 kstat_named_t arcstat_l2_writes_sent; 551 kstat_named_t arcstat_l2_writes_done; 552 kstat_named_t arcstat_l2_writes_error; 553 kstat_named_t arcstat_l2_writes_lock_retry; 554 kstat_named_t arcstat_l2_evict_lock_retry; 555 kstat_named_t arcstat_l2_evict_reading; 556 kstat_named_t arcstat_l2_evict_l1cached; 557 kstat_named_t arcstat_l2_free_on_write; 558 kstat_named_t arcstat_l2_cdata_free_on_write; 559 kstat_named_t arcstat_l2_abort_lowmem; 560 kstat_named_t arcstat_l2_cksum_bad; 561 kstat_named_t arcstat_l2_io_error; 562 kstat_named_t arcstat_l2_size; 563 kstat_named_t arcstat_l2_asize; 564 kstat_named_t arcstat_l2_hdr_size; 565 kstat_named_t arcstat_l2_compress_successes; 566 kstat_named_t arcstat_l2_compress_zeros; 567 kstat_named_t arcstat_l2_compress_failures; 568 kstat_named_t arcstat_l2_write_trylock_fail; 569 kstat_named_t arcstat_l2_write_passed_headroom; 570 kstat_named_t arcstat_l2_write_spa_mismatch; 571 kstat_named_t arcstat_l2_write_in_l2; 572 kstat_named_t arcstat_l2_write_hdr_io_in_progress; 573 kstat_named_t arcstat_l2_write_not_cacheable; 574 kstat_named_t arcstat_l2_write_full; 575 kstat_named_t arcstat_l2_write_buffer_iter; 576 kstat_named_t arcstat_l2_write_pios; 577 kstat_named_t arcstat_l2_write_buffer_bytes_scanned; 578 kstat_named_t arcstat_l2_write_buffer_list_iter; 579 kstat_named_t arcstat_l2_write_buffer_list_null_iter; 580 kstat_named_t arcstat_memory_throttle_count; 581 kstat_named_t arcstat_duplicate_buffers; 582 kstat_named_t arcstat_duplicate_buffers_size; 583 kstat_named_t arcstat_duplicate_reads; 584 kstat_named_t arcstat_meta_used; 585 kstat_named_t arcstat_meta_limit; 586 kstat_named_t arcstat_meta_max; 587 kstat_named_t arcstat_meta_min; 588} arc_stats_t; 589 590static arc_stats_t arc_stats = { 591 { "hits", KSTAT_DATA_UINT64 }, 592 { "misses", KSTAT_DATA_UINT64 }, 593 { "demand_data_hits", KSTAT_DATA_UINT64 }, 594 { "demand_data_misses", KSTAT_DATA_UINT64 }, 595 { "demand_metadata_hits", KSTAT_DATA_UINT64 }, 596 { "demand_metadata_misses", KSTAT_DATA_UINT64 }, 597 { "prefetch_data_hits", KSTAT_DATA_UINT64 }, 598 { "prefetch_data_misses", KSTAT_DATA_UINT64 }, 599 { "prefetch_metadata_hits", KSTAT_DATA_UINT64 }, 600 { "prefetch_metadata_misses", KSTAT_DATA_UINT64 }, 601 { "mru_hits", KSTAT_DATA_UINT64 }, 602 { "mru_ghost_hits", KSTAT_DATA_UINT64 }, 603 { "mfu_hits", KSTAT_DATA_UINT64 }, 604 { "mfu_ghost_hits", KSTAT_DATA_UINT64 }, 605 { "allocated", KSTAT_DATA_UINT64 }, 606 { "deleted", KSTAT_DATA_UINT64 }, 607 { "mutex_miss", KSTAT_DATA_UINT64 }, 608 { "evict_skip", KSTAT_DATA_UINT64 }, 609 { "evict_not_enough", KSTAT_DATA_UINT64 }, 610 { "evict_l2_cached", KSTAT_DATA_UINT64 }, 611 { "evict_l2_eligible", KSTAT_DATA_UINT64 }, 612 { "evict_l2_ineligible", KSTAT_DATA_UINT64 }, 613 { "evict_l2_skip", KSTAT_DATA_UINT64 }, 614 { "hash_elements", KSTAT_DATA_UINT64 }, 615 { "hash_elements_max", KSTAT_DATA_UINT64 }, 616 { "hash_collisions", KSTAT_DATA_UINT64 }, 617 { "hash_chains", KSTAT_DATA_UINT64 }, 618 { "hash_chain_max", KSTAT_DATA_UINT64 }, 619 { "p", KSTAT_DATA_UINT64 }, 620 { "c", KSTAT_DATA_UINT64 }, 621 { "c_min", KSTAT_DATA_UINT64 }, 622 { "c_max", KSTAT_DATA_UINT64 }, 623 { "size", KSTAT_DATA_UINT64 }, 624 { "hdr_size", KSTAT_DATA_UINT64 }, 625 { "data_size", KSTAT_DATA_UINT64 }, 626 { "metadata_size", KSTAT_DATA_UINT64 }, 627 { "other_size", KSTAT_DATA_UINT64 }, 628 { "anon_size", KSTAT_DATA_UINT64 }, 629 { "anon_evictable_data", KSTAT_DATA_UINT64 }, 630 { "anon_evictable_metadata", KSTAT_DATA_UINT64 }, 631 { "mru_size", KSTAT_DATA_UINT64 }, 632 { "mru_evictable_data", KSTAT_DATA_UINT64 }, 633 { "mru_evictable_metadata", KSTAT_DATA_UINT64 }, 634 { "mru_ghost_size", KSTAT_DATA_UINT64 }, 635 { "mru_ghost_evictable_data", KSTAT_DATA_UINT64 }, 636 { "mru_ghost_evictable_metadata", KSTAT_DATA_UINT64 }, 637 { "mfu_size", KSTAT_DATA_UINT64 }, 638 { "mfu_evictable_data", KSTAT_DATA_UINT64 }, 639 { "mfu_evictable_metadata", KSTAT_DATA_UINT64 }, 640 { "mfu_ghost_size", KSTAT_DATA_UINT64 }, 641 { "mfu_ghost_evictable_data", KSTAT_DATA_UINT64 }, 642 { "mfu_ghost_evictable_metadata", KSTAT_DATA_UINT64 }, 643 { "l2_hits", KSTAT_DATA_UINT64 }, 644 { "l2_misses", KSTAT_DATA_UINT64 }, 645 { "l2_feeds", KSTAT_DATA_UINT64 }, 646 { "l2_rw_clash", KSTAT_DATA_UINT64 }, 647 { "l2_read_bytes", KSTAT_DATA_UINT64 }, 648 { "l2_write_bytes", KSTAT_DATA_UINT64 }, 649 { "l2_writes_sent", KSTAT_DATA_UINT64 }, 650 { "l2_writes_done", KSTAT_DATA_UINT64 }, 651 { "l2_writes_error", KSTAT_DATA_UINT64 }, 652 { "l2_writes_lock_retry", KSTAT_DATA_UINT64 }, 653 { "l2_evict_lock_retry", KSTAT_DATA_UINT64 }, 654 { "l2_evict_reading", KSTAT_DATA_UINT64 }, 655 { "l2_evict_l1cached", KSTAT_DATA_UINT64 }, 656 { "l2_free_on_write", KSTAT_DATA_UINT64 }, 657 { "l2_cdata_free_on_write", KSTAT_DATA_UINT64 }, 658 { "l2_abort_lowmem", KSTAT_DATA_UINT64 }, 659 { "l2_cksum_bad", KSTAT_DATA_UINT64 }, 660 { "l2_io_error", KSTAT_DATA_UINT64 }, 661 { "l2_size", KSTAT_DATA_UINT64 }, 662 { "l2_asize", KSTAT_DATA_UINT64 }, 663 { "l2_hdr_size", KSTAT_DATA_UINT64 }, 664 { "l2_compress_successes", KSTAT_DATA_UINT64 }, 665 { "l2_compress_zeros", KSTAT_DATA_UINT64 }, 666 { "l2_compress_failures", KSTAT_DATA_UINT64 }, 667 { "l2_write_trylock_fail", KSTAT_DATA_UINT64 }, 668 { "l2_write_passed_headroom", KSTAT_DATA_UINT64 }, 669 { "l2_write_spa_mismatch", KSTAT_DATA_UINT64 }, 670 { "l2_write_in_l2", KSTAT_DATA_UINT64 }, 671 { "l2_write_io_in_progress", KSTAT_DATA_UINT64 }, 672 { "l2_write_not_cacheable", KSTAT_DATA_UINT64 }, 673 { "l2_write_full", KSTAT_DATA_UINT64 }, 674 { "l2_write_buffer_iter", KSTAT_DATA_UINT64 }, 675 { "l2_write_pios", KSTAT_DATA_UINT64 }, 676 { "l2_write_buffer_bytes_scanned", KSTAT_DATA_UINT64 }, 677 { "l2_write_buffer_list_iter", KSTAT_DATA_UINT64 }, 678 { "l2_write_buffer_list_null_iter", KSTAT_DATA_UINT64 }, 679 { "memory_throttle_count", KSTAT_DATA_UINT64 }, 680 { "duplicate_buffers", KSTAT_DATA_UINT64 }, 681 { "duplicate_buffers_size", KSTAT_DATA_UINT64 }, 682 { "duplicate_reads", KSTAT_DATA_UINT64 }, 683 { "arc_meta_used", KSTAT_DATA_UINT64 }, 684 { "arc_meta_limit", KSTAT_DATA_UINT64 }, 685 { "arc_meta_max", KSTAT_DATA_UINT64 }, 686 { "arc_meta_min", KSTAT_DATA_UINT64 } 687}; 688 689#define ARCSTAT(stat) (arc_stats.stat.value.ui64) 690 691#define ARCSTAT_INCR(stat, val) \ 692 atomic_add_64(&arc_stats.stat.value.ui64, (val)) 693 694#define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1) 695#define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1) 696 697#define ARCSTAT_MAX(stat, val) { \ 698 uint64_t m; \ 699 while ((val) > (m = arc_stats.stat.value.ui64) && \ 700 (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \ 701 continue; \ 702} 703 704#define ARCSTAT_MAXSTAT(stat) \ 705 ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64) 706 707/* 708 * We define a macro to allow ARC hits/misses to be easily broken down by 709 * two separate conditions, giving a total of four different subtypes for 710 * each of hits and misses (so eight statistics total). 711 */ 712#define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \ 713 if (cond1) { \ 714 if (cond2) { \ 715 ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \ 716 } else { \ 717 ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \ 718 } \ 719 } else { \ 720 if (cond2) { \ 721 ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \ 722 } else { \ 723 ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\ 724 } \ 725 } 726 727kstat_t *arc_ksp; 728static arc_state_t *arc_anon; 729static arc_state_t *arc_mru; 730static arc_state_t *arc_mru_ghost; 731static arc_state_t *arc_mfu; 732static arc_state_t *arc_mfu_ghost; 733static arc_state_t *arc_l2c_only; 734 735/* 736 * There are several ARC variables that are critical to export as kstats -- 737 * but we don't want to have to grovel around in the kstat whenever we wish to 738 * manipulate them. For these variables, we therefore define them to be in 739 * terms of the statistic variable. This assures that we are not introducing 740 * the possibility of inconsistency by having shadow copies of the variables, 741 * while still allowing the code to be readable. 742 */ 743#define arc_size ARCSTAT(arcstat_size) /* actual total arc size */ 744#define arc_p ARCSTAT(arcstat_p) /* target size of MRU */ 745#define arc_c ARCSTAT(arcstat_c) /* target size of cache */ 746#define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */ 747#define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */ 748#define arc_meta_limit ARCSTAT(arcstat_meta_limit) /* max size for metadata */ 749#define arc_meta_min ARCSTAT(arcstat_meta_min) /* min size for metadata */ 750#define arc_meta_used ARCSTAT(arcstat_meta_used) /* size of metadata */ 751#define arc_meta_max ARCSTAT(arcstat_meta_max) /* max size of metadata */ 752 753#define L2ARC_IS_VALID_COMPRESS(_c_) \ 754 ((_c_) == ZIO_COMPRESS_LZ4 || (_c_) == ZIO_COMPRESS_EMPTY) 755 756static int arc_no_grow; /* Don't try to grow cache size */ 757static uint64_t arc_tempreserve; 758static uint64_t arc_loaned_bytes; 759 760typedef struct arc_callback arc_callback_t; 761 762struct arc_callback { 763 void *acb_private; 764 arc_done_func_t *acb_done; 765 arc_buf_t *acb_buf; 766 zio_t *acb_zio_dummy; 767 arc_callback_t *acb_next; 768}; 769 770typedef struct arc_write_callback arc_write_callback_t; 771 772struct arc_write_callback { 773 void *awcb_private; 774 arc_done_func_t *awcb_ready; 775 arc_done_func_t *awcb_physdone; 776 arc_done_func_t *awcb_done; 777 arc_buf_t *awcb_buf; 778}; 779 780/* 781 * ARC buffers are separated into multiple structs as a memory saving measure: 782 * - Common fields struct, always defined, and embedded within it: 783 * - L2-only fields, always allocated but undefined when not in L2ARC 784 * - L1-only fields, only allocated when in L1ARC 785 * 786 * Buffer in L1 Buffer only in L2 787 * +------------------------+ +------------------------+ 788 * | arc_buf_hdr_t | | arc_buf_hdr_t | 789 * | | | | 790 * | | | | 791 * | | | | 792 * +------------------------+ +------------------------+ 793 * | l2arc_buf_hdr_t | | l2arc_buf_hdr_t | 794 * | (undefined if L1-only) | | | 795 * +------------------------+ +------------------------+ 796 * | l1arc_buf_hdr_t | 797 * | | 798 * | | 799 * | | 800 * | | 801 * +------------------------+ 802 * 803 * Because it's possible for the L2ARC to become extremely large, we can wind 804 * up eating a lot of memory in L2ARC buffer headers, so the size of a header 805 * is minimized by only allocating the fields necessary for an L1-cached buffer 806 * when a header is actually in the L1 cache. The sub-headers (l1arc_buf_hdr and 807 * l2arc_buf_hdr) are embedded rather than allocated separately to save a couple 808 * words in pointers. arc_hdr_realloc() is used to switch a header between 809 * these two allocation states. 810 */ 811typedef struct l1arc_buf_hdr { 812 kmutex_t b_freeze_lock; 813#ifdef ZFS_DEBUG 814 /* 815 * used for debugging wtih kmem_flags - by allocating and freeing 816 * b_thawed when the buffer is thawed, we get a record of the stack 817 * trace that thawed it. 818 */ 819 void *b_thawed; 820#endif 821 822 arc_buf_t *b_buf; 823 uint32_t b_datacnt; 824 /* for waiting on writes to complete */ 825 kcondvar_t b_cv; 826 827 /* protected by arc state mutex */ 828 arc_state_t *b_state; 829 multilist_node_t b_arc_node; 830 831 /* updated atomically */ 832 clock_t b_arc_access; 833 834 /* self protecting */ 835 refcount_t b_refcnt; 836 837 arc_callback_t *b_acb; 838 /* temporary buffer holder for in-flight compressed data */ 839 void *b_tmp_cdata; 840} l1arc_buf_hdr_t; 841 842typedef struct l2arc_dev l2arc_dev_t; 843 844typedef struct l2arc_buf_hdr { 845 /* protected by arc_buf_hdr mutex */ 846 l2arc_dev_t *b_dev; /* L2ARC device */ 847 uint64_t b_daddr; /* disk address, offset byte */ 848 /* real alloc'd buffer size depending on b_compress applied */ 849 int32_t b_asize; 850 851 list_node_t b_l2node; 852} l2arc_buf_hdr_t; 853 854struct arc_buf_hdr { 855 /* protected by hash lock */ 856 dva_t b_dva; 857 uint64_t b_birth; 858 /* 859 * Even though this checksum is only set/verified when a buffer is in 860 * the L1 cache, it needs to be in the set of common fields because it 861 * must be preserved from the time before a buffer is written out to 862 * L2ARC until after it is read back in. 863 */ 864 zio_cksum_t *b_freeze_cksum; 865 866 arc_buf_hdr_t *b_hash_next; 867 arc_flags_t b_flags; 868 869 /* immutable */ 870 int32_t b_size; 871 uint64_t b_spa; 872 873 /* L2ARC fields. Undefined when not in L2ARC. */ 874 l2arc_buf_hdr_t b_l2hdr; 875 /* L1ARC fields. Undefined when in l2arc_only state */ 876 l1arc_buf_hdr_t b_l1hdr; 877}; 878 879#ifdef _KERNEL 880static int 881sysctl_vfs_zfs_arc_meta_limit(SYSCTL_HANDLER_ARGS) 882{ 883 uint64_t val; 884 int err; 885 886 val = arc_meta_limit; 887 err = sysctl_handle_64(oidp, &val, 0, req); 888 if (err != 0 || req->newptr == NULL) 889 return (err); 890 891 if (val <= 0 || val > arc_c_max) 892 return (EINVAL); 893 894 arc_meta_limit = val; 895 return (0); 896} 897#endif 898 899static arc_buf_t *arc_eviction_list; 900static arc_buf_hdr_t arc_eviction_hdr; 901 902#define GHOST_STATE(state) \ 903 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \ 904 (state) == arc_l2c_only) 905 906#define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_FLAG_IN_HASH_TABLE) 907#define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) 908#define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_FLAG_IO_ERROR) 909#define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_FLAG_PREFETCH) 910#define HDR_FREED_IN_READ(hdr) ((hdr)->b_flags & ARC_FLAG_FREED_IN_READ) 911#define HDR_BUF_AVAILABLE(hdr) ((hdr)->b_flags & ARC_FLAG_BUF_AVAILABLE) 912 913#define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_FLAG_L2CACHE) 914#define HDR_L2COMPRESS(hdr) ((hdr)->b_flags & ARC_FLAG_L2COMPRESS) 915#define HDR_L2_READING(hdr) \ 916 (((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) && \ 917 ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR)) 918#define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITING) 919#define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_FLAG_L2_EVICTED) 920#define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITE_HEAD) 921 922#define HDR_ISTYPE_METADATA(hdr) \ 923 ((hdr)->b_flags & ARC_FLAG_BUFC_METADATA) 924#define HDR_ISTYPE_DATA(hdr) (!HDR_ISTYPE_METADATA(hdr)) 925 926#define HDR_HAS_L1HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L1HDR) 927#define HDR_HAS_L2HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR) 928 929/* For storing compression mode in b_flags */ 930#define HDR_COMPRESS_OFFSET 24 931#define HDR_COMPRESS_NBITS 7 932 933#define HDR_GET_COMPRESS(hdr) ((enum zio_compress)BF32_GET(hdr->b_flags, \ 934 HDR_COMPRESS_OFFSET, HDR_COMPRESS_NBITS)) 935#define HDR_SET_COMPRESS(hdr, cmp) BF32_SET(hdr->b_flags, \ 936 HDR_COMPRESS_OFFSET, HDR_COMPRESS_NBITS, (cmp)) 937 938/* 939 * Other sizes 940 */ 941 942#define HDR_FULL_SIZE ((int64_t)sizeof (arc_buf_hdr_t)) 943#define HDR_L2ONLY_SIZE ((int64_t)offsetof(arc_buf_hdr_t, b_l1hdr)) 944 945/* 946 * Hash table routines 947 */ 948 949#define HT_LOCK_PAD CACHE_LINE_SIZE 950 951struct ht_lock { 952 kmutex_t ht_lock; 953#ifdef _KERNEL 954 unsigned char pad[(HT_LOCK_PAD - sizeof (kmutex_t))]; 955#endif 956}; 957 958#define BUF_LOCKS 256 959typedef struct buf_hash_table { 960 uint64_t ht_mask; 961 arc_buf_hdr_t **ht_table; 962 struct ht_lock ht_locks[BUF_LOCKS] __aligned(CACHE_LINE_SIZE); 963} buf_hash_table_t; 964 965static buf_hash_table_t buf_hash_table; 966 967#define BUF_HASH_INDEX(spa, dva, birth) \ 968 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask) 969#define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)]) 970#define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock)) 971#define HDR_LOCK(hdr) \ 972 (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth))) 973 974uint64_t zfs_crc64_table[256]; 975 976/* 977 * Level 2 ARC 978 */ 979 980#define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */ 981#define L2ARC_HEADROOM 2 /* num of writes */ 982/* 983 * If we discover during ARC scan any buffers to be compressed, we boost 984 * our headroom for the next scanning cycle by this percentage multiple. 985 */ 986#define L2ARC_HEADROOM_BOOST 200 987#define L2ARC_FEED_SECS 1 /* caching interval secs */ 988#define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */ 989 990/* 991 * Used to distinguish headers that are being process by 992 * l2arc_write_buffers(), but have yet to be assigned to a l2arc disk 993 * address. This can happen when the header is added to the l2arc's list 994 * of buffers to write in the first stage of l2arc_write_buffers(), but 995 * has not yet been written out which happens in the second stage of 996 * l2arc_write_buffers(). 997 */ 998#define L2ARC_ADDR_UNSET ((uint64_t)(-1)) 999 1000#define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent) 1001#define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done) 1002 1003/* L2ARC Performance Tunables */ 1004uint64_t l2arc_write_max = L2ARC_WRITE_SIZE; /* default max write size */ 1005uint64_t l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra write during warmup */ 1006uint64_t l2arc_headroom = L2ARC_HEADROOM; /* number of dev writes */ 1007uint64_t l2arc_headroom_boost = L2ARC_HEADROOM_BOOST; 1008uint64_t l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */ 1009uint64_t l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval milliseconds */ 1010boolean_t l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */ 1011boolean_t l2arc_feed_again = B_TRUE; /* turbo warmup */ 1012boolean_t l2arc_norw = B_TRUE; /* no reads during writes */ 1013 1014SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_max, CTLFLAG_RW, 1015 &l2arc_write_max, 0, "max write size"); 1016SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_boost, CTLFLAG_RW, 1017 &l2arc_write_boost, 0, "extra write during warmup"); 1018SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_headroom, CTLFLAG_RW, 1019 &l2arc_headroom, 0, "number of dev writes"); 1020SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_secs, CTLFLAG_RW, 1021 &l2arc_feed_secs, 0, "interval seconds"); 1022SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_min_ms, CTLFLAG_RW, 1023 &l2arc_feed_min_ms, 0, "min interval milliseconds"); 1024 1025SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_noprefetch, CTLFLAG_RW, 1026 &l2arc_noprefetch, 0, "don't cache prefetch bufs"); 1027SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_feed_again, CTLFLAG_RW, 1028 &l2arc_feed_again, 0, "turbo warmup"); 1029SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_norw, CTLFLAG_RW, 1030 &l2arc_norw, 0, "no reads during writes"); 1031 1032SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_size, CTLFLAG_RD, 1033 &ARC_anon.arcs_size, 0, "size of anonymous state"); 1034SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_metadata_lsize, CTLFLAG_RD, 1035 &ARC_anon.arcs_lsize[ARC_BUFC_METADATA], 0, "size of anonymous state"); 1036SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_data_lsize, CTLFLAG_RD, 1037 &ARC_anon.arcs_lsize[ARC_BUFC_DATA], 0, "size of anonymous state"); 1038 1039SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_size, CTLFLAG_RD, 1040 &ARC_mru.arcs_size, 0, "size of mru state"); 1041SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_metadata_lsize, CTLFLAG_RD, 1042 &ARC_mru.arcs_lsize[ARC_BUFC_METADATA], 0, "size of metadata in mru state"); 1043SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_data_lsize, CTLFLAG_RD, 1044 &ARC_mru.arcs_lsize[ARC_BUFC_DATA], 0, "size of data in mru state"); 1045 1046SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_size, CTLFLAG_RD, 1047 &ARC_mru_ghost.arcs_size, 0, "size of mru ghost state"); 1048SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_metadata_lsize, CTLFLAG_RD, 1049 &ARC_mru_ghost.arcs_lsize[ARC_BUFC_METADATA], 0, 1050 "size of metadata in mru ghost state"); 1051SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_data_lsize, CTLFLAG_RD, 1052 &ARC_mru_ghost.arcs_lsize[ARC_BUFC_DATA], 0, 1053 "size of data in mru ghost state"); 1054 1055SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_size, CTLFLAG_RD, 1056 &ARC_mfu.arcs_size, 0, "size of mfu state"); 1057SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_metadata_lsize, CTLFLAG_RD, 1058 &ARC_mfu.arcs_lsize[ARC_BUFC_METADATA], 0, "size of metadata in mfu state"); 1059SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_data_lsize, CTLFLAG_RD, 1060 &ARC_mfu.arcs_lsize[ARC_BUFC_DATA], 0, "size of data in mfu state"); 1061 1062SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_size, CTLFLAG_RD, 1063 &ARC_mfu_ghost.arcs_size, 0, "size of mfu ghost state"); 1064SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_metadata_lsize, CTLFLAG_RD, 1065 &ARC_mfu_ghost.arcs_lsize[ARC_BUFC_METADATA], 0, 1066 "size of metadata in mfu ghost state"); 1067SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_data_lsize, CTLFLAG_RD, 1068 &ARC_mfu_ghost.arcs_lsize[ARC_BUFC_DATA], 0, 1069 "size of data in mfu ghost state"); 1070 1071SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2c_only_size, CTLFLAG_RD, 1072 &ARC_l2c_only.arcs_size, 0, "size of mru state"); 1073 1074/* 1075 * L2ARC Internals 1076 */ 1077struct l2arc_dev { 1078 vdev_t *l2ad_vdev; /* vdev */ 1079 spa_t *l2ad_spa; /* spa */ 1080 uint64_t l2ad_hand; /* next write location */ 1081 uint64_t l2ad_start; /* first addr on device */ 1082 uint64_t l2ad_end; /* last addr on device */ 1083 boolean_t l2ad_first; /* first sweep through */ 1084 boolean_t l2ad_writing; /* currently writing */ 1085 kmutex_t l2ad_mtx; /* lock for buffer list */ 1086 list_t l2ad_buflist; /* buffer list */ 1087 list_node_t l2ad_node; /* device list node */ 1088 refcount_t l2ad_alloc; /* allocated bytes */ 1089}; 1090 1091static list_t L2ARC_dev_list; /* device list */ 1092static list_t *l2arc_dev_list; /* device list pointer */ 1093static kmutex_t l2arc_dev_mtx; /* device list mutex */ 1094static l2arc_dev_t *l2arc_dev_last; /* last device used */ 1095static list_t L2ARC_free_on_write; /* free after write buf list */ 1096static list_t *l2arc_free_on_write; /* free after write list ptr */ 1097static kmutex_t l2arc_free_on_write_mtx; /* mutex for list */ 1098static uint64_t l2arc_ndev; /* number of devices */ 1099 1100typedef struct l2arc_read_callback { 1101 arc_buf_t *l2rcb_buf; /* read buffer */ 1102 spa_t *l2rcb_spa; /* spa */ 1103 blkptr_t l2rcb_bp; /* original blkptr */ 1104 zbookmark_phys_t l2rcb_zb; /* original bookmark */ 1105 int l2rcb_flags; /* original flags */ 1106 enum zio_compress l2rcb_compress; /* applied compress */ 1107} l2arc_read_callback_t; 1108 1109typedef struct l2arc_write_callback { 1110 l2arc_dev_t *l2wcb_dev; /* device info */ 1111 arc_buf_hdr_t *l2wcb_head; /* head of write buflist */ 1112} l2arc_write_callback_t; 1113 1114typedef struct l2arc_data_free { 1115 /* protected by l2arc_free_on_write_mtx */ 1116 void *l2df_data; 1117 size_t l2df_size; 1118 void (*l2df_func)(void *, size_t); 1119 list_node_t l2df_list_node; 1120} l2arc_data_free_t; 1121 1122static kmutex_t l2arc_feed_thr_lock; 1123static kcondvar_t l2arc_feed_thr_cv; 1124static uint8_t l2arc_thread_exit; 1125 1126static void arc_get_data_buf(arc_buf_t *); 1127static void arc_access(arc_buf_hdr_t *, kmutex_t *); 1128static boolean_t arc_is_overflowing(); 1129static void arc_buf_watch(arc_buf_t *); 1130 1131static arc_buf_contents_t arc_buf_type(arc_buf_hdr_t *); 1132static uint32_t arc_bufc_to_flags(arc_buf_contents_t); 1133 1134static boolean_t l2arc_write_eligible(uint64_t, arc_buf_hdr_t *); 1135static void l2arc_read_done(zio_t *); 1136 1137static boolean_t l2arc_compress_buf(arc_buf_hdr_t *); 1138static void l2arc_decompress_zio(zio_t *, arc_buf_hdr_t *, enum zio_compress); 1139static void l2arc_release_cdata_buf(arc_buf_hdr_t *); 1140 1141static uint64_t 1142buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth) 1143{ 1144 uint8_t *vdva = (uint8_t *)dva; 1145 uint64_t crc = -1ULL; 1146 int i; 1147 1148 ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY); 1149 1150 for (i = 0; i < sizeof (dva_t); i++) 1151 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ vdva[i]) & 0xFF]; 1152 1153 crc ^= (spa>>8) ^ birth; 1154 1155 return (crc); 1156} 1157 1158#define BUF_EMPTY(buf) \ 1159 ((buf)->b_dva.dva_word[0] == 0 && \ 1160 (buf)->b_dva.dva_word[1] == 0) 1161 1162#define BUF_EQUAL(spa, dva, birth, buf) \ 1163 ((buf)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \ 1164 ((buf)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \ 1165 ((buf)->b_birth == birth) && ((buf)->b_spa == spa) 1166 1167static void 1168buf_discard_identity(arc_buf_hdr_t *hdr) 1169{ 1170 hdr->b_dva.dva_word[0] = 0; 1171 hdr->b_dva.dva_word[1] = 0; 1172 hdr->b_birth = 0; 1173} 1174 1175static arc_buf_hdr_t * 1176buf_hash_find(uint64_t spa, const blkptr_t *bp, kmutex_t **lockp) 1177{ 1178 const dva_t *dva = BP_IDENTITY(bp); 1179 uint64_t birth = BP_PHYSICAL_BIRTH(bp); 1180 uint64_t idx = BUF_HASH_INDEX(spa, dva, birth); 1181 kmutex_t *hash_lock = BUF_HASH_LOCK(idx); 1182 arc_buf_hdr_t *hdr; 1183 1184 mutex_enter(hash_lock); 1185 for (hdr = buf_hash_table.ht_table[idx]; hdr != NULL; 1186 hdr = hdr->b_hash_next) { 1187 if (BUF_EQUAL(spa, dva, birth, hdr)) { 1188 *lockp = hash_lock; 1189 return (hdr); 1190 } 1191 } 1192 mutex_exit(hash_lock); 1193 *lockp = NULL; 1194 return (NULL); 1195} 1196 1197/* 1198 * Insert an entry into the hash table. If there is already an element 1199 * equal to elem in the hash table, then the already existing element 1200 * will be returned and the new element will not be inserted. 1201 * Otherwise returns NULL. 1202 * If lockp == NULL, the caller is assumed to already hold the hash lock. 1203 */ 1204static arc_buf_hdr_t * 1205buf_hash_insert(arc_buf_hdr_t *hdr, kmutex_t **lockp) 1206{ 1207 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth); 1208 kmutex_t *hash_lock = BUF_HASH_LOCK(idx); 1209 arc_buf_hdr_t *fhdr; 1210 uint32_t i; 1211 1212 ASSERT(!DVA_IS_EMPTY(&hdr->b_dva)); 1213 ASSERT(hdr->b_birth != 0); 1214 ASSERT(!HDR_IN_HASH_TABLE(hdr)); 1215 1216 if (lockp != NULL) { 1217 *lockp = hash_lock; 1218 mutex_enter(hash_lock); 1219 } else { 1220 ASSERT(MUTEX_HELD(hash_lock)); 1221 } 1222 1223 for (fhdr = buf_hash_table.ht_table[idx], i = 0; fhdr != NULL; 1224 fhdr = fhdr->b_hash_next, i++) { 1225 if (BUF_EQUAL(hdr->b_spa, &hdr->b_dva, hdr->b_birth, fhdr)) 1226 return (fhdr); 1227 } 1228 1229 hdr->b_hash_next = buf_hash_table.ht_table[idx]; 1230 buf_hash_table.ht_table[idx] = hdr; 1231 hdr->b_flags |= ARC_FLAG_IN_HASH_TABLE; 1232 1233 /* collect some hash table performance data */ 1234 if (i > 0) { 1235 ARCSTAT_BUMP(arcstat_hash_collisions); 1236 if (i == 1) 1237 ARCSTAT_BUMP(arcstat_hash_chains); 1238 1239 ARCSTAT_MAX(arcstat_hash_chain_max, i); 1240 } 1241 1242 ARCSTAT_BUMP(arcstat_hash_elements); 1243 ARCSTAT_MAXSTAT(arcstat_hash_elements); 1244 1245 return (NULL); 1246} 1247 1248static void 1249buf_hash_remove(arc_buf_hdr_t *hdr) 1250{ 1251 arc_buf_hdr_t *fhdr, **hdrp; 1252 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth); 1253 1254 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx))); 1255 ASSERT(HDR_IN_HASH_TABLE(hdr)); 1256 1257 hdrp = &buf_hash_table.ht_table[idx]; 1258 while ((fhdr = *hdrp) != hdr) { 1259 ASSERT(fhdr != NULL); 1260 hdrp = &fhdr->b_hash_next; 1261 } 1262 *hdrp = hdr->b_hash_next; 1263 hdr->b_hash_next = NULL; 1264 hdr->b_flags &= ~ARC_FLAG_IN_HASH_TABLE; 1265 1266 /* collect some hash table performance data */ 1267 ARCSTAT_BUMPDOWN(arcstat_hash_elements); 1268 1269 if (buf_hash_table.ht_table[idx] && 1270 buf_hash_table.ht_table[idx]->b_hash_next == NULL) 1271 ARCSTAT_BUMPDOWN(arcstat_hash_chains); 1272} 1273 1274/* 1275 * Global data structures and functions for the buf kmem cache. 1276 */ 1277static kmem_cache_t *hdr_full_cache; 1278static kmem_cache_t *hdr_l2only_cache; 1279static kmem_cache_t *buf_cache; 1280 1281static void 1282buf_fini(void) 1283{ 1284 int i; 1285 1286 kmem_free(buf_hash_table.ht_table, 1287 (buf_hash_table.ht_mask + 1) * sizeof (void *)); 1288 for (i = 0; i < BUF_LOCKS; i++) 1289 mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock); 1290 kmem_cache_destroy(hdr_full_cache); 1291 kmem_cache_destroy(hdr_l2only_cache); 1292 kmem_cache_destroy(buf_cache); 1293} 1294 1295/* 1296 * Constructor callback - called when the cache is empty 1297 * and a new buf is requested. 1298 */ 1299/* ARGSUSED */ 1300static int 1301hdr_full_cons(void *vbuf, void *unused, int kmflag) 1302{ 1303 arc_buf_hdr_t *hdr = vbuf; 1304 1305 bzero(hdr, HDR_FULL_SIZE); 1306 cv_init(&hdr->b_l1hdr.b_cv, NULL, CV_DEFAULT, NULL); 1307 refcount_create(&hdr->b_l1hdr.b_refcnt); 1308 mutex_init(&hdr->b_l1hdr.b_freeze_lock, NULL, MUTEX_DEFAULT, NULL); 1309 multilist_link_init(&hdr->b_l1hdr.b_arc_node); 1310 arc_space_consume(HDR_FULL_SIZE, ARC_SPACE_HDRS); 1311 1312 return (0); 1313} 1314 1315/* ARGSUSED */ 1316static int 1317hdr_l2only_cons(void *vbuf, void *unused, int kmflag) 1318{ 1319 arc_buf_hdr_t *hdr = vbuf; 1320 1321 bzero(hdr, HDR_L2ONLY_SIZE); 1322 arc_space_consume(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS); 1323 1324 return (0); 1325} 1326 1327/* ARGSUSED */ 1328static int 1329buf_cons(void *vbuf, void *unused, int kmflag) 1330{ 1331 arc_buf_t *buf = vbuf; 1332 1333 bzero(buf, sizeof (arc_buf_t)); 1334 mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL); 1335 arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS); 1336 1337 return (0); 1338} 1339 1340/* 1341 * Destructor callback - called when a cached buf is 1342 * no longer required. 1343 */ 1344/* ARGSUSED */ 1345static void 1346hdr_full_dest(void *vbuf, void *unused) 1347{ 1348 arc_buf_hdr_t *hdr = vbuf; 1349 1350 ASSERT(BUF_EMPTY(hdr)); 1351 cv_destroy(&hdr->b_l1hdr.b_cv); 1352 refcount_destroy(&hdr->b_l1hdr.b_refcnt); 1353 mutex_destroy(&hdr->b_l1hdr.b_freeze_lock); 1354 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 1355 arc_space_return(HDR_FULL_SIZE, ARC_SPACE_HDRS); 1356} 1357 1358/* ARGSUSED */ 1359static void 1360hdr_l2only_dest(void *vbuf, void *unused) 1361{ 1362 arc_buf_hdr_t *hdr = vbuf; 1363 1364 ASSERT(BUF_EMPTY(hdr)); 1365 arc_space_return(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS); 1366} 1367 1368/* ARGSUSED */ 1369static void 1370buf_dest(void *vbuf, void *unused) 1371{ 1372 arc_buf_t *buf = vbuf; 1373 1374 mutex_destroy(&buf->b_evict_lock); 1375 arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS); 1376} 1377 1378/* 1379 * Reclaim callback -- invoked when memory is low. 1380 */ 1381/* ARGSUSED */ 1382static void 1383hdr_recl(void *unused) 1384{ 1385 dprintf("hdr_recl called\n"); 1386 /* 1387 * umem calls the reclaim func when we destroy the buf cache, 1388 * which is after we do arc_fini(). 1389 */ 1390 if (!arc_dead) 1391 cv_signal(&arc_reclaim_thread_cv); 1392} 1393 1394static void 1395buf_init(void) 1396{ 1397 uint64_t *ct; 1398 uint64_t hsize = 1ULL << 12; 1399 int i, j; 1400 1401 /* 1402 * The hash table is big enough to fill all of physical memory 1403 * with an average block size of zfs_arc_average_blocksize (default 8K). 1404 * By default, the table will take up 1405 * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers). 1406 */ 1407 while (hsize * zfs_arc_average_blocksize < (uint64_t)physmem * PAGESIZE) 1408 hsize <<= 1; 1409retry: 1410 buf_hash_table.ht_mask = hsize - 1; 1411 buf_hash_table.ht_table = 1412 kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP); 1413 if (buf_hash_table.ht_table == NULL) { 1414 ASSERT(hsize > (1ULL << 8)); 1415 hsize >>= 1; 1416 goto retry; 1417 } 1418 1419 hdr_full_cache = kmem_cache_create("arc_buf_hdr_t_full", HDR_FULL_SIZE, 1420 0, hdr_full_cons, hdr_full_dest, hdr_recl, NULL, NULL, 0); 1421 hdr_l2only_cache = kmem_cache_create("arc_buf_hdr_t_l2only", 1422 HDR_L2ONLY_SIZE, 0, hdr_l2only_cons, hdr_l2only_dest, hdr_recl, 1423 NULL, NULL, 0); 1424 buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t), 1425 0, buf_cons, buf_dest, NULL, NULL, NULL, 0); 1426 1427 for (i = 0; i < 256; i++) 1428 for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--) 1429 *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY); 1430 1431 for (i = 0; i < BUF_LOCKS; i++) { 1432 mutex_init(&buf_hash_table.ht_locks[i].ht_lock, 1433 NULL, MUTEX_DEFAULT, NULL); 1434 } 1435} 1436 1437/* 1438 * Transition between the two allocation states for the arc_buf_hdr struct. 1439 * The arc_buf_hdr struct can be allocated with (hdr_full_cache) or without 1440 * (hdr_l2only_cache) the fields necessary for the L1 cache - the smaller 1441 * version is used when a cache buffer is only in the L2ARC in order to reduce 1442 * memory usage. 1443 */ 1444static arc_buf_hdr_t * 1445arc_hdr_realloc(arc_buf_hdr_t *hdr, kmem_cache_t *old, kmem_cache_t *new) 1446{ 1447 ASSERT(HDR_HAS_L2HDR(hdr)); 1448 1449 arc_buf_hdr_t *nhdr; 1450 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev; 1451 1452 ASSERT((old == hdr_full_cache && new == hdr_l2only_cache) || 1453 (old == hdr_l2only_cache && new == hdr_full_cache)); 1454 1455 nhdr = kmem_cache_alloc(new, KM_PUSHPAGE); 1456 1457 ASSERT(MUTEX_HELD(HDR_LOCK(hdr))); 1458 buf_hash_remove(hdr); 1459 1460 bcopy(hdr, nhdr, HDR_L2ONLY_SIZE); 1461 1462 if (new == hdr_full_cache) { 1463 nhdr->b_flags |= ARC_FLAG_HAS_L1HDR; 1464 /* 1465 * arc_access and arc_change_state need to be aware that a 1466 * header has just come out of L2ARC, so we set its state to 1467 * l2c_only even though it's about to change. 1468 */ 1469 nhdr->b_l1hdr.b_state = arc_l2c_only; 1470 1471 /* Verify previous threads set to NULL before freeing */ 1472 ASSERT3P(nhdr->b_l1hdr.b_tmp_cdata, ==, NULL); 1473 } else { 1474 ASSERT(hdr->b_l1hdr.b_buf == NULL); 1475 ASSERT0(hdr->b_l1hdr.b_datacnt); 1476 1477 /* 1478 * If we've reached here, We must have been called from 1479 * arc_evict_hdr(), as such we should have already been 1480 * removed from any ghost list we were previously on 1481 * (which protects us from racing with arc_evict_state), 1482 * thus no locking is needed during this check. 1483 */ 1484 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 1485 1486 /* 1487 * A buffer must not be moved into the arc_l2c_only 1488 * state if it's not finished being written out to the 1489 * l2arc device. Otherwise, the b_l1hdr.b_tmp_cdata field 1490 * might try to be accessed, even though it was removed. 1491 */ 1492 VERIFY(!HDR_L2_WRITING(hdr)); 1493 VERIFY3P(hdr->b_l1hdr.b_tmp_cdata, ==, NULL); 1494 1495 nhdr->b_flags &= ~ARC_FLAG_HAS_L1HDR; 1496 } 1497 /* 1498 * The header has been reallocated so we need to re-insert it into any 1499 * lists it was on. 1500 */ 1501 (void) buf_hash_insert(nhdr, NULL); 1502 1503 ASSERT(list_link_active(&hdr->b_l2hdr.b_l2node)); 1504 1505 mutex_enter(&dev->l2ad_mtx); 1506 1507 /* 1508 * We must place the realloc'ed header back into the list at 1509 * the same spot. Otherwise, if it's placed earlier in the list, 1510 * l2arc_write_buffers() could find it during the function's 1511 * write phase, and try to write it out to the l2arc. 1512 */ 1513 list_insert_after(&dev->l2ad_buflist, hdr, nhdr); 1514 list_remove(&dev->l2ad_buflist, hdr); 1515 1516 mutex_exit(&dev->l2ad_mtx); 1517 1518 /* 1519 * Since we're using the pointer address as the tag when 1520 * incrementing and decrementing the l2ad_alloc refcount, we 1521 * must remove the old pointer (that we're about to destroy) and 1522 * add the new pointer to the refcount. Otherwise we'd remove 1523 * the wrong pointer address when calling arc_hdr_destroy() later. 1524 */ 1525 1526 (void) refcount_remove_many(&dev->l2ad_alloc, 1527 hdr->b_l2hdr.b_asize, hdr); 1528 1529 (void) refcount_add_many(&dev->l2ad_alloc, 1530 nhdr->b_l2hdr.b_asize, nhdr); 1531 1532 buf_discard_identity(hdr); 1533 hdr->b_freeze_cksum = NULL; 1534 kmem_cache_free(old, hdr); 1535 1536 return (nhdr); 1537} 1538 1539 1540#define ARC_MINTIME (hz>>4) /* 62 ms */ 1541 1542static void 1543arc_cksum_verify(arc_buf_t *buf) 1544{ 1545 zio_cksum_t zc; 1546 1547 if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 1548 return; 1549 1550 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1551 if (buf->b_hdr->b_freeze_cksum == NULL || HDR_IO_ERROR(buf->b_hdr)) { 1552 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1553 return; 1554 } 1555 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc); 1556 if (!ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc)) 1557 panic("buffer modified while frozen!"); 1558 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1559} 1560 1561static int 1562arc_cksum_equal(arc_buf_t *buf) 1563{ 1564 zio_cksum_t zc; 1565 int equal; 1566 1567 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1568 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc); 1569 equal = ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc); 1570 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1571 1572 return (equal); 1573} 1574 1575static void 1576arc_cksum_compute(arc_buf_t *buf, boolean_t force) 1577{ 1578 if (!force && !(zfs_flags & ZFS_DEBUG_MODIFY)) 1579 return; 1580 1581 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1582 if (buf->b_hdr->b_freeze_cksum != NULL) { 1583 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1584 return; 1585 } 1586 buf->b_hdr->b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t), KM_SLEEP); 1587 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, 1588 buf->b_hdr->b_freeze_cksum); 1589 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1590#ifdef illumos 1591 arc_buf_watch(buf); 1592#endif /* illumos */ 1593} 1594 1595#ifdef illumos 1596#ifndef _KERNEL 1597typedef struct procctl { 1598 long cmd; 1599 prwatch_t prwatch; 1600} procctl_t; 1601#endif 1602 1603/* ARGSUSED */ 1604static void 1605arc_buf_unwatch(arc_buf_t *buf) 1606{ 1607#ifndef _KERNEL 1608 if (arc_watch) { 1609 int result; 1610 procctl_t ctl; 1611 ctl.cmd = PCWATCH; 1612 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data; 1613 ctl.prwatch.pr_size = 0; 1614 ctl.prwatch.pr_wflags = 0; 1615 result = write(arc_procfd, &ctl, sizeof (ctl)); 1616 ASSERT3U(result, ==, sizeof (ctl)); 1617 } 1618#endif 1619} 1620 1621/* ARGSUSED */ 1622static void 1623arc_buf_watch(arc_buf_t *buf) 1624{ 1625#ifndef _KERNEL 1626 if (arc_watch) { 1627 int result; 1628 procctl_t ctl; 1629 ctl.cmd = PCWATCH; 1630 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data; 1631 ctl.prwatch.pr_size = buf->b_hdr->b_size; 1632 ctl.prwatch.pr_wflags = WA_WRITE; 1633 result = write(arc_procfd, &ctl, sizeof (ctl)); 1634 ASSERT3U(result, ==, sizeof (ctl)); 1635 } 1636#endif 1637} 1638#endif /* illumos */ 1639 1640static arc_buf_contents_t 1641arc_buf_type(arc_buf_hdr_t *hdr) 1642{ 1643 if (HDR_ISTYPE_METADATA(hdr)) { 1644 return (ARC_BUFC_METADATA); 1645 } else { 1646 return (ARC_BUFC_DATA); 1647 } 1648} 1649 1650static uint32_t 1651arc_bufc_to_flags(arc_buf_contents_t type) 1652{ 1653 switch (type) { 1654 case ARC_BUFC_DATA: 1655 /* metadata field is 0 if buffer contains normal data */ 1656 return (0); 1657 case ARC_BUFC_METADATA: 1658 return (ARC_FLAG_BUFC_METADATA); 1659 default: 1660 break; 1661 } 1662 panic("undefined ARC buffer type!"); 1663 return ((uint32_t)-1); 1664} 1665 1666void 1667arc_buf_thaw(arc_buf_t *buf) 1668{ 1669 if (zfs_flags & ZFS_DEBUG_MODIFY) { 1670 if (buf->b_hdr->b_l1hdr.b_state != arc_anon) 1671 panic("modifying non-anon buffer!"); 1672 if (HDR_IO_IN_PROGRESS(buf->b_hdr)) 1673 panic("modifying buffer while i/o in progress!"); 1674 arc_cksum_verify(buf); 1675 } 1676 1677 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1678 if (buf->b_hdr->b_freeze_cksum != NULL) { 1679 kmem_free(buf->b_hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 1680 buf->b_hdr->b_freeze_cksum = NULL; 1681 } 1682 1683#ifdef ZFS_DEBUG 1684 if (zfs_flags & ZFS_DEBUG_MODIFY) { 1685 if (buf->b_hdr->b_l1hdr.b_thawed != NULL) 1686 kmem_free(buf->b_hdr->b_l1hdr.b_thawed, 1); 1687 buf->b_hdr->b_l1hdr.b_thawed = kmem_alloc(1, KM_SLEEP); 1688 } 1689#endif 1690 1691 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1692 1693#ifdef illumos 1694 arc_buf_unwatch(buf); 1695#endif /* illumos */ 1696} 1697 1698void 1699arc_buf_freeze(arc_buf_t *buf) 1700{ 1701 kmutex_t *hash_lock; 1702 1703 if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 1704 return; 1705 1706 hash_lock = HDR_LOCK(buf->b_hdr); 1707 mutex_enter(hash_lock); 1708 1709 ASSERT(buf->b_hdr->b_freeze_cksum != NULL || 1710 buf->b_hdr->b_l1hdr.b_state == arc_anon); 1711 arc_cksum_compute(buf, B_FALSE); 1712 mutex_exit(hash_lock); 1713 1714} 1715 1716static void 1717add_reference(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, void *tag) 1718{ 1719 ASSERT(HDR_HAS_L1HDR(hdr)); 1720 ASSERT(MUTEX_HELD(hash_lock)); 1721 arc_state_t *state = hdr->b_l1hdr.b_state; 1722 1723 if ((refcount_add(&hdr->b_l1hdr.b_refcnt, tag) == 1) && 1724 (state != arc_anon)) { 1725 /* We don't use the L2-only state list. */ 1726 if (state != arc_l2c_only) { 1727 arc_buf_contents_t type = arc_buf_type(hdr); 1728 uint64_t delta = hdr->b_size * hdr->b_l1hdr.b_datacnt; 1729 multilist_t *list = &state->arcs_list[type]; 1730 uint64_t *size = &state->arcs_lsize[type]; 1731 1732 multilist_remove(list, hdr); 1733 1734 if (GHOST_STATE(state)) { 1735 ASSERT0(hdr->b_l1hdr.b_datacnt); 1736 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 1737 delta = hdr->b_size; 1738 } 1739 ASSERT(delta > 0); 1740 ASSERT3U(*size, >=, delta); 1741 atomic_add_64(size, -delta); 1742 } 1743 /* remove the prefetch flag if we get a reference */ 1744 hdr->b_flags &= ~ARC_FLAG_PREFETCH; 1745 } 1746} 1747 1748static int 1749remove_reference(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, void *tag) 1750{ 1751 int cnt; 1752 arc_state_t *state = hdr->b_l1hdr.b_state; 1753 1754 ASSERT(HDR_HAS_L1HDR(hdr)); 1755 ASSERT(state == arc_anon || MUTEX_HELD(hash_lock)); 1756 ASSERT(!GHOST_STATE(state)); 1757 1758 /* 1759 * arc_l2c_only counts as a ghost state so we don't need to explicitly 1760 * check to prevent usage of the arc_l2c_only list. 1761 */ 1762 if (((cnt = refcount_remove(&hdr->b_l1hdr.b_refcnt, tag)) == 0) && 1763 (state != arc_anon)) { 1764 arc_buf_contents_t type = arc_buf_type(hdr); 1765 multilist_t *list = &state->arcs_list[type]; 1766 uint64_t *size = &state->arcs_lsize[type]; 1767 1768 multilist_insert(list, hdr); 1769 1770 ASSERT(hdr->b_l1hdr.b_datacnt > 0); 1771 atomic_add_64(size, hdr->b_size * 1772 hdr->b_l1hdr.b_datacnt); 1773 } 1774 return (cnt); 1775} 1776 1777/* 1778 * Move the supplied buffer to the indicated state. The hash lock 1779 * for the buffer must be held by the caller. 1780 */ 1781static void 1782arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *hdr, 1783 kmutex_t *hash_lock) 1784{ 1785 arc_state_t *old_state; 1786 int64_t refcnt; 1787 uint32_t datacnt; 1788 uint64_t from_delta, to_delta; 1789 arc_buf_contents_t buftype = arc_buf_type(hdr); 1790 1791 /* 1792 * We almost always have an L1 hdr here, since we call arc_hdr_realloc() 1793 * in arc_read() when bringing a buffer out of the L2ARC. However, the 1794 * L1 hdr doesn't always exist when we change state to arc_anon before 1795 * destroying a header, in which case reallocating to add the L1 hdr is 1796 * pointless. 1797 */ 1798 if (HDR_HAS_L1HDR(hdr)) { 1799 old_state = hdr->b_l1hdr.b_state; 1800 refcnt = refcount_count(&hdr->b_l1hdr.b_refcnt); 1801 datacnt = hdr->b_l1hdr.b_datacnt; 1802 } else { 1803 old_state = arc_l2c_only; 1804 refcnt = 0; 1805 datacnt = 0; 1806 } 1807 1808 ASSERT(MUTEX_HELD(hash_lock)); 1809 ASSERT3P(new_state, !=, old_state); 1810 ASSERT(refcnt == 0 || datacnt > 0); 1811 ASSERT(!GHOST_STATE(new_state) || datacnt == 0); 1812 ASSERT(old_state != arc_anon || datacnt <= 1); 1813 1814 from_delta = to_delta = datacnt * hdr->b_size; 1815 1816 /* 1817 * If this buffer is evictable, transfer it from the 1818 * old state list to the new state list. 1819 */ 1820 if (refcnt == 0) { 1821 if (old_state != arc_anon && old_state != arc_l2c_only) { 1822 uint64_t *size = &old_state->arcs_lsize[buftype]; 1823 1824 ASSERT(HDR_HAS_L1HDR(hdr)); 1825 multilist_remove(&old_state->arcs_list[buftype], hdr); 1826 1827 /* 1828 * If prefetching out of the ghost cache, 1829 * we will have a non-zero datacnt. 1830 */ 1831 if (GHOST_STATE(old_state) && datacnt == 0) { 1832 /* ghost elements have a ghost size */ 1833 ASSERT(hdr->b_l1hdr.b_buf == NULL); 1834 from_delta = hdr->b_size; 1835 } 1836 ASSERT3U(*size, >=, from_delta); 1837 atomic_add_64(size, -from_delta); 1838 } 1839 if (new_state != arc_anon && new_state != arc_l2c_only) { 1840 uint64_t *size = &new_state->arcs_lsize[buftype]; 1841 1842 /* 1843 * An L1 header always exists here, since if we're 1844 * moving to some L1-cached state (i.e. not l2c_only or 1845 * anonymous), we realloc the header to add an L1hdr 1846 * beforehand. 1847 */ 1848 ASSERT(HDR_HAS_L1HDR(hdr)); 1849 multilist_insert(&new_state->arcs_list[buftype], hdr); 1850 1851 /* ghost elements have a ghost size */ 1852 if (GHOST_STATE(new_state)) { 1853 ASSERT0(datacnt); 1854 ASSERT(hdr->b_l1hdr.b_buf == NULL); 1855 to_delta = hdr->b_size; 1856 } 1857 atomic_add_64(size, to_delta); 1858 } 1859 } 1860 1861 ASSERT(!BUF_EMPTY(hdr)); 1862 if (new_state == arc_anon && HDR_IN_HASH_TABLE(hdr)) 1863 buf_hash_remove(hdr); 1864 1865 /* adjust state sizes (ignore arc_l2c_only) */ 1866 1867 if (to_delta && new_state != arc_l2c_only) { 1868 ASSERT(HDR_HAS_L1HDR(hdr)); 1869 if (GHOST_STATE(new_state)) { 1870 ASSERT0(datacnt); 1871 1872 /* 1873 * We moving a header to a ghost state, we first 1874 * remove all arc buffers. Thus, we'll have a 1875 * datacnt of zero, and no arc buffer to use for 1876 * the reference. As a result, we use the arc 1877 * header pointer for the reference. 1878 */ 1879 (void) refcount_add_many(&new_state->arcs_size, 1880 hdr->b_size, hdr); 1881 } else { 1882 ASSERT3U(datacnt, !=, 0); 1883 1884 /* 1885 * Each individual buffer holds a unique reference, 1886 * thus we must remove each of these references one 1887 * at a time. 1888 */ 1889 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL; 1890 buf = buf->b_next) { 1891 (void) refcount_add_many(&new_state->arcs_size, 1892 hdr->b_size, buf); 1893 } 1894 } 1895 } 1896 1897 if (from_delta && old_state != arc_l2c_only) { 1898 ASSERT(HDR_HAS_L1HDR(hdr)); 1899 if (GHOST_STATE(old_state)) { 1900 /* 1901 * When moving a header off of a ghost state, 1902 * there's the possibility for datacnt to be 1903 * non-zero. This is because we first add the 1904 * arc buffer to the header prior to changing 1905 * the header's state. Since we used the header 1906 * for the reference when putting the header on 1907 * the ghost state, we must balance that and use 1908 * the header when removing off the ghost state 1909 * (even though datacnt is non zero). 1910 */ 1911 1912 IMPLY(datacnt == 0, new_state == arc_anon || 1913 new_state == arc_l2c_only); 1914 1915 (void) refcount_remove_many(&old_state->arcs_size, 1916 hdr->b_size, hdr); 1917 } else { 1918 ASSERT3P(datacnt, !=, 0); 1919 1920 /* 1921 * Each individual buffer holds a unique reference, 1922 * thus we must remove each of these references one 1923 * at a time. 1924 */ 1925 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL; 1926 buf = buf->b_next) { 1927 (void) refcount_remove_many( 1928 &old_state->arcs_size, hdr->b_size, buf); 1929 } 1930 } 1931 } 1932 1933 if (HDR_HAS_L1HDR(hdr)) 1934 hdr->b_l1hdr.b_state = new_state; 1935 1936 /* 1937 * L2 headers should never be on the L2 state list since they don't 1938 * have L1 headers allocated. 1939 */ 1940 ASSERT(multilist_is_empty(&arc_l2c_only->arcs_list[ARC_BUFC_DATA]) && 1941 multilist_is_empty(&arc_l2c_only->arcs_list[ARC_BUFC_METADATA])); 1942} 1943 1944void 1945arc_space_consume(uint64_t space, arc_space_type_t type) 1946{ 1947 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES); 1948 1949 switch (type) { 1950 case ARC_SPACE_DATA: 1951 ARCSTAT_INCR(arcstat_data_size, space); 1952 break; 1953 case ARC_SPACE_META: 1954 ARCSTAT_INCR(arcstat_metadata_size, space); 1955 break; 1956 case ARC_SPACE_OTHER: 1957 ARCSTAT_INCR(arcstat_other_size, space); 1958 break; 1959 case ARC_SPACE_HDRS: 1960 ARCSTAT_INCR(arcstat_hdr_size, space); 1961 break; 1962 case ARC_SPACE_L2HDRS: 1963 ARCSTAT_INCR(arcstat_l2_hdr_size, space); 1964 break; 1965 } 1966 1967 if (type != ARC_SPACE_DATA) 1968 ARCSTAT_INCR(arcstat_meta_used, space); 1969 1970 atomic_add_64(&arc_size, space); 1971} 1972 1973void 1974arc_space_return(uint64_t space, arc_space_type_t type) 1975{ 1976 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES); 1977 1978 switch (type) { 1979 case ARC_SPACE_DATA: 1980 ARCSTAT_INCR(arcstat_data_size, -space); 1981 break; 1982 case ARC_SPACE_META: 1983 ARCSTAT_INCR(arcstat_metadata_size, -space); 1984 break; 1985 case ARC_SPACE_OTHER: 1986 ARCSTAT_INCR(arcstat_other_size, -space); 1987 break; 1988 case ARC_SPACE_HDRS: 1989 ARCSTAT_INCR(arcstat_hdr_size, -space); 1990 break; 1991 case ARC_SPACE_L2HDRS: 1992 ARCSTAT_INCR(arcstat_l2_hdr_size, -space); 1993 break; 1994 } 1995 1996 if (type != ARC_SPACE_DATA) { 1997 ASSERT(arc_meta_used >= space); 1998 if (arc_meta_max < arc_meta_used) 1999 arc_meta_max = arc_meta_used; 2000 ARCSTAT_INCR(arcstat_meta_used, -space); 2001 } 2002 2003 ASSERT(arc_size >= space); 2004 atomic_add_64(&arc_size, -space); 2005} 2006 2007arc_buf_t * 2008arc_buf_alloc(spa_t *spa, int32_t size, void *tag, arc_buf_contents_t type) 2009{ 2010 arc_buf_hdr_t *hdr; 2011 arc_buf_t *buf; 2012 2013 ASSERT3U(size, >, 0); 2014 hdr = kmem_cache_alloc(hdr_full_cache, KM_PUSHPAGE); 2015 ASSERT(BUF_EMPTY(hdr)); 2016 ASSERT3P(hdr->b_freeze_cksum, ==, NULL); 2017 hdr->b_size = size; 2018 hdr->b_spa = spa_load_guid(spa); 2019 2020 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 2021 buf->b_hdr = hdr; 2022 buf->b_data = NULL; 2023 buf->b_efunc = NULL; 2024 buf->b_private = NULL; 2025 buf->b_next = NULL; 2026 2027 hdr->b_flags = arc_bufc_to_flags(type); 2028 hdr->b_flags |= ARC_FLAG_HAS_L1HDR; 2029 2030 hdr->b_l1hdr.b_buf = buf; 2031 hdr->b_l1hdr.b_state = arc_anon; 2032 hdr->b_l1hdr.b_arc_access = 0; 2033 hdr->b_l1hdr.b_datacnt = 1; 2034 hdr->b_l1hdr.b_tmp_cdata = NULL; 2035 2036 arc_get_data_buf(buf); 2037 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 2038 (void) refcount_add(&hdr->b_l1hdr.b_refcnt, tag); 2039 2040 return (buf); 2041} 2042 2043static char *arc_onloan_tag = "onloan"; 2044 2045/* 2046 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in 2047 * flight data by arc_tempreserve_space() until they are "returned". Loaned 2048 * buffers must be returned to the arc before they can be used by the DMU or 2049 * freed. 2050 */ 2051arc_buf_t * 2052arc_loan_buf(spa_t *spa, int size) 2053{ 2054 arc_buf_t *buf; 2055 2056 buf = arc_buf_alloc(spa, size, arc_onloan_tag, ARC_BUFC_DATA); 2057 2058 atomic_add_64(&arc_loaned_bytes, size); 2059 return (buf); 2060} 2061 2062/* 2063 * Return a loaned arc buffer to the arc. 2064 */ 2065void 2066arc_return_buf(arc_buf_t *buf, void *tag) 2067{ 2068 arc_buf_hdr_t *hdr = buf->b_hdr; 2069 2070 ASSERT(buf->b_data != NULL); 2071 ASSERT(HDR_HAS_L1HDR(hdr)); 2072 (void) refcount_add(&hdr->b_l1hdr.b_refcnt, tag); 2073 (void) refcount_remove(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag); 2074 2075 atomic_add_64(&arc_loaned_bytes, -hdr->b_size); 2076} 2077 2078/* Detach an arc_buf from a dbuf (tag) */ 2079void 2080arc_loan_inuse_buf(arc_buf_t *buf, void *tag) 2081{ 2082 arc_buf_hdr_t *hdr = buf->b_hdr; 2083 2084 ASSERT(buf->b_data != NULL); 2085 ASSERT(HDR_HAS_L1HDR(hdr)); 2086 (void) refcount_add(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag); 2087 (void) refcount_remove(&hdr->b_l1hdr.b_refcnt, tag); 2088 buf->b_efunc = NULL; 2089 buf->b_private = NULL; 2090 2091 atomic_add_64(&arc_loaned_bytes, hdr->b_size); 2092} 2093 2094static arc_buf_t * 2095arc_buf_clone(arc_buf_t *from) 2096{ 2097 arc_buf_t *buf; 2098 arc_buf_hdr_t *hdr = from->b_hdr; 2099 uint64_t size = hdr->b_size; 2100 2101 ASSERT(HDR_HAS_L1HDR(hdr)); 2102 ASSERT(hdr->b_l1hdr.b_state != arc_anon); 2103 2104 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 2105 buf->b_hdr = hdr; 2106 buf->b_data = NULL; 2107 buf->b_efunc = NULL; 2108 buf->b_private = NULL; 2109 buf->b_next = hdr->b_l1hdr.b_buf; 2110 hdr->b_l1hdr.b_buf = buf; 2111 arc_get_data_buf(buf); 2112 bcopy(from->b_data, buf->b_data, size); 2113 2114 /* 2115 * This buffer already exists in the arc so create a duplicate 2116 * copy for the caller. If the buffer is associated with user data 2117 * then track the size and number of duplicates. These stats will be 2118 * updated as duplicate buffers are created and destroyed. 2119 */ 2120 if (HDR_ISTYPE_DATA(hdr)) { 2121 ARCSTAT_BUMP(arcstat_duplicate_buffers); 2122 ARCSTAT_INCR(arcstat_duplicate_buffers_size, size); 2123 } 2124 hdr->b_l1hdr.b_datacnt += 1; 2125 return (buf); 2126} 2127 2128void 2129arc_buf_add_ref(arc_buf_t *buf, void* tag) 2130{ 2131 arc_buf_hdr_t *hdr; 2132 kmutex_t *hash_lock; 2133 2134 /* 2135 * Check to see if this buffer is evicted. Callers 2136 * must verify b_data != NULL to know if the add_ref 2137 * was successful. 2138 */ 2139 mutex_enter(&buf->b_evict_lock); 2140 if (buf->b_data == NULL) { 2141 mutex_exit(&buf->b_evict_lock); 2142 return; 2143 } 2144 hash_lock = HDR_LOCK(buf->b_hdr); 2145 mutex_enter(hash_lock); 2146 hdr = buf->b_hdr; 2147 ASSERT(HDR_HAS_L1HDR(hdr)); 2148 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 2149 mutex_exit(&buf->b_evict_lock); 2150 2151 ASSERT(hdr->b_l1hdr.b_state == arc_mru || 2152 hdr->b_l1hdr.b_state == arc_mfu); 2153 2154 add_reference(hdr, hash_lock, tag); 2155 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr); 2156 arc_access(hdr, hash_lock); 2157 mutex_exit(hash_lock); 2158 ARCSTAT_BUMP(arcstat_hits); 2159 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr), 2160 demand, prefetch, !HDR_ISTYPE_METADATA(hdr), 2161 data, metadata, hits); 2162} 2163 2164static void 2165arc_buf_free_on_write(void *data, size_t size, 2166 void (*free_func)(void *, size_t)) 2167{ 2168 l2arc_data_free_t *df; 2169 2170 df = kmem_alloc(sizeof (*df), KM_SLEEP); 2171 df->l2df_data = data; 2172 df->l2df_size = size; 2173 df->l2df_func = free_func; 2174 mutex_enter(&l2arc_free_on_write_mtx); 2175 list_insert_head(l2arc_free_on_write, df); 2176 mutex_exit(&l2arc_free_on_write_mtx); 2177} 2178 2179/* 2180 * Free the arc data buffer. If it is an l2arc write in progress, 2181 * the buffer is placed on l2arc_free_on_write to be freed later. 2182 */ 2183static void 2184arc_buf_data_free(arc_buf_t *buf, void (*free_func)(void *, size_t)) 2185{ 2186 arc_buf_hdr_t *hdr = buf->b_hdr; 2187 2188 if (HDR_L2_WRITING(hdr)) { 2189 arc_buf_free_on_write(buf->b_data, hdr->b_size, free_func); 2190 ARCSTAT_BUMP(arcstat_l2_free_on_write); 2191 } else { 2192 free_func(buf->b_data, hdr->b_size); 2193 } 2194} 2195 2196/* 2197 * Free up buf->b_data and if 'remove' is set, then pull the 2198 * arc_buf_t off of the the arc_buf_hdr_t's list and free it. 2199 */ 2200static void 2201arc_buf_l2_cdata_free(arc_buf_hdr_t *hdr) 2202{ 2203 ASSERT(HDR_HAS_L2HDR(hdr)); 2204 ASSERT(MUTEX_HELD(&hdr->b_l2hdr.b_dev->l2ad_mtx)); 2205 2206 /* 2207 * The b_tmp_cdata field is linked off of the b_l1hdr, so if 2208 * that doesn't exist, the header is in the arc_l2c_only state, 2209 * and there isn't anything to free (it's already been freed). 2210 */ 2211 if (!HDR_HAS_L1HDR(hdr)) 2212 return; 2213 2214 /* 2215 * The header isn't being written to the l2arc device, thus it 2216 * shouldn't have a b_tmp_cdata to free. 2217 */ 2218 if (!HDR_L2_WRITING(hdr)) { 2219 ASSERT3P(hdr->b_l1hdr.b_tmp_cdata, ==, NULL); 2220 return; 2221 } 2222 2223 /* 2224 * The header does not have compression enabled. This can be due 2225 * to the buffer not being compressible, or because we're 2226 * freeing the buffer before the second phase of 2227 * l2arc_write_buffer() has started (which does the compression 2228 * step). In either case, b_tmp_cdata does not point to a 2229 * separately compressed buffer, so there's nothing to free (it 2230 * points to the same buffer as the arc_buf_t's b_data field). 2231 */ 2232 if (HDR_GET_COMPRESS(hdr) == ZIO_COMPRESS_OFF) { 2233 hdr->b_l1hdr.b_tmp_cdata = NULL; 2234 return; 2235 } 2236 2237 /* 2238 * There's nothing to free since the buffer was all zero's and 2239 * compressed to a zero length buffer. 2240 */ 2241 if (HDR_GET_COMPRESS(hdr) == ZIO_COMPRESS_EMPTY) { 2242 ASSERT3P(hdr->b_l1hdr.b_tmp_cdata, ==, NULL); 2243 return; 2244 } 2245 2246 ASSERT(L2ARC_IS_VALID_COMPRESS(HDR_GET_COMPRESS(hdr))); 2247 2248 arc_buf_free_on_write(hdr->b_l1hdr.b_tmp_cdata, 2249 hdr->b_size, zio_data_buf_free); 2250 2251 ARCSTAT_BUMP(arcstat_l2_cdata_free_on_write); 2252 hdr->b_l1hdr.b_tmp_cdata = NULL; 2253} 2254 2255static void 2256arc_buf_destroy(arc_buf_t *buf, boolean_t remove) 2257{ 2258 arc_buf_t **bufp; 2259 2260 /* free up data associated with the buf */ 2261 if (buf->b_data != NULL) { 2262 arc_state_t *state = buf->b_hdr->b_l1hdr.b_state; 2263 uint64_t size = buf->b_hdr->b_size; 2264 arc_buf_contents_t type = arc_buf_type(buf->b_hdr); 2265 2266 arc_cksum_verify(buf); 2267#ifdef illumos 2268 arc_buf_unwatch(buf); 2269#endif /* illumos */ 2270 2271 if (type == ARC_BUFC_METADATA) { 2272 arc_buf_data_free(buf, zio_buf_free); 2273 arc_space_return(size, ARC_SPACE_META); 2274 } else { 2275 ASSERT(type == ARC_BUFC_DATA); 2276 arc_buf_data_free(buf, zio_data_buf_free); 2277 arc_space_return(size, ARC_SPACE_DATA); 2278 } 2279 2280 /* protected by hash lock, if in the hash table */ 2281 if (multilist_link_active(&buf->b_hdr->b_l1hdr.b_arc_node)) { 2282 uint64_t *cnt = &state->arcs_lsize[type]; 2283 2284 ASSERT(refcount_is_zero( 2285 &buf->b_hdr->b_l1hdr.b_refcnt)); 2286 ASSERT(state != arc_anon && state != arc_l2c_only); 2287 2288 ASSERT3U(*cnt, >=, size); 2289 atomic_add_64(cnt, -size); 2290 } 2291 2292 (void) refcount_remove_many(&state->arcs_size, size, buf); 2293 buf->b_data = NULL; 2294 2295 /* 2296 * If we're destroying a duplicate buffer make sure 2297 * that the appropriate statistics are updated. 2298 */ 2299 if (buf->b_hdr->b_l1hdr.b_datacnt > 1 && 2300 HDR_ISTYPE_DATA(buf->b_hdr)) { 2301 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers); 2302 ARCSTAT_INCR(arcstat_duplicate_buffers_size, -size); 2303 } 2304 ASSERT(buf->b_hdr->b_l1hdr.b_datacnt > 0); 2305 buf->b_hdr->b_l1hdr.b_datacnt -= 1; 2306 } 2307 2308 /* only remove the buf if requested */ 2309 if (!remove) 2310 return; 2311 2312 /* remove the buf from the hdr list */ 2313 for (bufp = &buf->b_hdr->b_l1hdr.b_buf; *bufp != buf; 2314 bufp = &(*bufp)->b_next) 2315 continue; 2316 *bufp = buf->b_next; 2317 buf->b_next = NULL; 2318 2319 ASSERT(buf->b_efunc == NULL); 2320 2321 /* clean up the buf */ 2322 buf->b_hdr = NULL; 2323 kmem_cache_free(buf_cache, buf); 2324} 2325 2326static void 2327arc_hdr_l2hdr_destroy(arc_buf_hdr_t *hdr) 2328{ 2329 l2arc_buf_hdr_t *l2hdr = &hdr->b_l2hdr; 2330 l2arc_dev_t *dev = l2hdr->b_dev; 2331 2332 ASSERT(MUTEX_HELD(&dev->l2ad_mtx)); 2333 ASSERT(HDR_HAS_L2HDR(hdr)); 2334 2335 list_remove(&dev->l2ad_buflist, hdr); 2336 2337 /* 2338 * We don't want to leak the b_tmp_cdata buffer that was 2339 * allocated in l2arc_write_buffers() 2340 */ 2341 arc_buf_l2_cdata_free(hdr); 2342 2343 /* 2344 * If the l2hdr's b_daddr is equal to L2ARC_ADDR_UNSET, then 2345 * this header is being processed by l2arc_write_buffers() (i.e. 2346 * it's in the first stage of l2arc_write_buffers()). 2347 * Re-affirming that truth here, just to serve as a reminder. If 2348 * b_daddr does not equal L2ARC_ADDR_UNSET, then the header may or 2349 * may not have its HDR_L2_WRITING flag set. (the write may have 2350 * completed, in which case HDR_L2_WRITING will be false and the 2351 * b_daddr field will point to the address of the buffer on disk). 2352 */ 2353 IMPLY(l2hdr->b_daddr == L2ARC_ADDR_UNSET, HDR_L2_WRITING(hdr)); 2354 2355 /* 2356 * If b_daddr is equal to L2ARC_ADDR_UNSET, we're racing with 2357 * l2arc_write_buffers(). Since we've just removed this header 2358 * from the l2arc buffer list, this header will never reach the 2359 * second stage of l2arc_write_buffers(), which increments the 2360 * accounting stats for this header. Thus, we must be careful 2361 * not to decrement them for this header either. 2362 */ 2363 if (l2hdr->b_daddr != L2ARC_ADDR_UNSET) { 2364 ARCSTAT_INCR(arcstat_l2_asize, -l2hdr->b_asize); 2365 ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size); 2366 2367 vdev_space_update(dev->l2ad_vdev, 2368 -l2hdr->b_asize, 0, 0); 2369 2370 (void) refcount_remove_many(&dev->l2ad_alloc, 2371 l2hdr->b_asize, hdr); 2372 } 2373 2374 hdr->b_flags &= ~ARC_FLAG_HAS_L2HDR; 2375} 2376 2377static void 2378arc_hdr_destroy(arc_buf_hdr_t *hdr) 2379{ 2380 if (HDR_HAS_L1HDR(hdr)) { 2381 ASSERT(hdr->b_l1hdr.b_buf == NULL || 2382 hdr->b_l1hdr.b_datacnt > 0); 2383 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 2384 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon); 2385 } 2386 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 2387 ASSERT(!HDR_IN_HASH_TABLE(hdr)); 2388 2389 if (HDR_HAS_L2HDR(hdr)) { 2390 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev; 2391 boolean_t buflist_held = MUTEX_HELD(&dev->l2ad_mtx); 2392 2393 if (!buflist_held) 2394 mutex_enter(&dev->l2ad_mtx); 2395 2396 /* 2397 * Even though we checked this conditional above, we 2398 * need to check this again now that we have the 2399 * l2ad_mtx. This is because we could be racing with 2400 * another thread calling l2arc_evict() which might have 2401 * destroyed this header's L2 portion as we were waiting 2402 * to acquire the l2ad_mtx. If that happens, we don't 2403 * want to re-destroy the header's L2 portion. 2404 */ 2405 if (HDR_HAS_L2HDR(hdr)) { 2406 if (hdr->b_l2hdr.b_daddr != L2ARC_ADDR_UNSET) 2407 trim_map_free(dev->l2ad_vdev, 2408 hdr->b_l2hdr.b_daddr, 2409 hdr->b_l2hdr.b_asize, 0); 2410 arc_hdr_l2hdr_destroy(hdr); 2411 } 2412 2413 if (!buflist_held) 2414 mutex_exit(&dev->l2ad_mtx); 2415 } 2416 2417 if (!BUF_EMPTY(hdr)) 2418 buf_discard_identity(hdr); 2419 if (hdr->b_freeze_cksum != NULL) { 2420 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 2421 hdr->b_freeze_cksum = NULL; 2422 } 2423 2424 if (HDR_HAS_L1HDR(hdr)) { 2425 while (hdr->b_l1hdr.b_buf) { 2426 arc_buf_t *buf = hdr->b_l1hdr.b_buf; 2427 2428 if (buf->b_efunc != NULL) { 2429 mutex_enter(&arc_user_evicts_lock); 2430 mutex_enter(&buf->b_evict_lock); 2431 ASSERT(buf->b_hdr != NULL); 2432 arc_buf_destroy(hdr->b_l1hdr.b_buf, FALSE); 2433 hdr->b_l1hdr.b_buf = buf->b_next; 2434 buf->b_hdr = &arc_eviction_hdr; 2435 buf->b_next = arc_eviction_list; 2436 arc_eviction_list = buf; 2437 mutex_exit(&buf->b_evict_lock); 2438 cv_signal(&arc_user_evicts_cv); 2439 mutex_exit(&arc_user_evicts_lock); 2440 } else { 2441 arc_buf_destroy(hdr->b_l1hdr.b_buf, TRUE); 2442 } 2443 } 2444#ifdef ZFS_DEBUG 2445 if (hdr->b_l1hdr.b_thawed != NULL) { 2446 kmem_free(hdr->b_l1hdr.b_thawed, 1); 2447 hdr->b_l1hdr.b_thawed = NULL; 2448 } 2449#endif 2450 } 2451 2452 ASSERT3P(hdr->b_hash_next, ==, NULL); 2453 if (HDR_HAS_L1HDR(hdr)) { 2454 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 2455 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL); 2456 kmem_cache_free(hdr_full_cache, hdr); 2457 } else { 2458 kmem_cache_free(hdr_l2only_cache, hdr); 2459 } 2460} 2461 2462void 2463arc_buf_free(arc_buf_t *buf, void *tag) 2464{ 2465 arc_buf_hdr_t *hdr = buf->b_hdr; 2466 int hashed = hdr->b_l1hdr.b_state != arc_anon; 2467 2468 ASSERT(buf->b_efunc == NULL); 2469 ASSERT(buf->b_data != NULL); 2470 2471 if (hashed) { 2472 kmutex_t *hash_lock = HDR_LOCK(hdr); 2473 2474 mutex_enter(hash_lock); 2475 hdr = buf->b_hdr; 2476 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 2477 2478 (void) remove_reference(hdr, hash_lock, tag); 2479 if (hdr->b_l1hdr.b_datacnt > 1) { 2480 arc_buf_destroy(buf, TRUE); 2481 } else { 2482 ASSERT(buf == hdr->b_l1hdr.b_buf); 2483 ASSERT(buf->b_efunc == NULL); 2484 hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE; 2485 } 2486 mutex_exit(hash_lock); 2487 } else if (HDR_IO_IN_PROGRESS(hdr)) { 2488 int destroy_hdr; 2489 /* 2490 * We are in the middle of an async write. Don't destroy 2491 * this buffer unless the write completes before we finish 2492 * decrementing the reference count. 2493 */ 2494 mutex_enter(&arc_user_evicts_lock); 2495 (void) remove_reference(hdr, NULL, tag); 2496 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 2497 destroy_hdr = !HDR_IO_IN_PROGRESS(hdr); 2498 mutex_exit(&arc_user_evicts_lock); 2499 if (destroy_hdr) 2500 arc_hdr_destroy(hdr); 2501 } else { 2502 if (remove_reference(hdr, NULL, tag) > 0) 2503 arc_buf_destroy(buf, TRUE); 2504 else 2505 arc_hdr_destroy(hdr); 2506 } 2507} 2508 2509boolean_t 2510arc_buf_remove_ref(arc_buf_t *buf, void* tag) 2511{ 2512 arc_buf_hdr_t *hdr = buf->b_hdr; 2513 kmutex_t *hash_lock = HDR_LOCK(hdr); 2514 boolean_t no_callback = (buf->b_efunc == NULL); 2515 2516 if (hdr->b_l1hdr.b_state == arc_anon) { 2517 ASSERT(hdr->b_l1hdr.b_datacnt == 1); 2518 arc_buf_free(buf, tag); 2519 return (no_callback); 2520 } 2521 2522 mutex_enter(hash_lock); 2523 hdr = buf->b_hdr; 2524 ASSERT(hdr->b_l1hdr.b_datacnt > 0); 2525 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 2526 ASSERT(hdr->b_l1hdr.b_state != arc_anon); 2527 ASSERT(buf->b_data != NULL); 2528 2529 (void) remove_reference(hdr, hash_lock, tag); 2530 if (hdr->b_l1hdr.b_datacnt > 1) { 2531 if (no_callback) 2532 arc_buf_destroy(buf, TRUE); 2533 } else if (no_callback) { 2534 ASSERT(hdr->b_l1hdr.b_buf == buf && buf->b_next == NULL); 2535 ASSERT(buf->b_efunc == NULL); 2536 hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE; 2537 } 2538 ASSERT(no_callback || hdr->b_l1hdr.b_datacnt > 1 || 2539 refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 2540 mutex_exit(hash_lock); 2541 return (no_callback); 2542} 2543 2544int32_t 2545arc_buf_size(arc_buf_t *buf) 2546{ 2547 return (buf->b_hdr->b_size); 2548} 2549 2550/* 2551 * Called from the DMU to determine if the current buffer should be 2552 * evicted. In order to ensure proper locking, the eviction must be initiated 2553 * from the DMU. Return true if the buffer is associated with user data and 2554 * duplicate buffers still exist. 2555 */ 2556boolean_t 2557arc_buf_eviction_needed(arc_buf_t *buf) 2558{ 2559 arc_buf_hdr_t *hdr; 2560 boolean_t evict_needed = B_FALSE; 2561 2562 if (zfs_disable_dup_eviction) 2563 return (B_FALSE); 2564 2565 mutex_enter(&buf->b_evict_lock); 2566 hdr = buf->b_hdr; 2567 if (hdr == NULL) { 2568 /* 2569 * We are in arc_do_user_evicts(); let that function 2570 * perform the eviction. 2571 */ 2572 ASSERT(buf->b_data == NULL); 2573 mutex_exit(&buf->b_evict_lock); 2574 return (B_FALSE); 2575 } else if (buf->b_data == NULL) { 2576 /* 2577 * We have already been added to the arc eviction list; 2578 * recommend eviction. 2579 */ 2580 ASSERT3P(hdr, ==, &arc_eviction_hdr); 2581 mutex_exit(&buf->b_evict_lock); 2582 return (B_TRUE); 2583 } 2584 2585 if (hdr->b_l1hdr.b_datacnt > 1 && HDR_ISTYPE_DATA(hdr)) 2586 evict_needed = B_TRUE; 2587 2588 mutex_exit(&buf->b_evict_lock); 2589 return (evict_needed); 2590} 2591 2592/* 2593 * Evict the arc_buf_hdr that is provided as a parameter. The resultant 2594 * state of the header is dependent on it's state prior to entering this 2595 * function. The following transitions are possible: 2596 * 2597 * - arc_mru -> arc_mru_ghost 2598 * - arc_mfu -> arc_mfu_ghost 2599 * - arc_mru_ghost -> arc_l2c_only 2600 * - arc_mru_ghost -> deleted 2601 * - arc_mfu_ghost -> arc_l2c_only 2602 * - arc_mfu_ghost -> deleted 2603 */ 2604static int64_t 2605arc_evict_hdr(arc_buf_hdr_t *hdr, kmutex_t *hash_lock) 2606{ 2607 arc_state_t *evicted_state, *state; 2608 int64_t bytes_evicted = 0; 2609 2610 ASSERT(MUTEX_HELD(hash_lock)); 2611 ASSERT(HDR_HAS_L1HDR(hdr)); 2612 2613 state = hdr->b_l1hdr.b_state; 2614 if (GHOST_STATE(state)) { 2615 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 2616 ASSERT(hdr->b_l1hdr.b_buf == NULL); 2617 2618 /* 2619 * l2arc_write_buffers() relies on a header's L1 portion 2620 * (i.e. it's b_tmp_cdata field) during it's write phase. 2621 * Thus, we cannot push a header onto the arc_l2c_only 2622 * state (removing it's L1 piece) until the header is 2623 * done being written to the l2arc. 2624 */ 2625 if (HDR_HAS_L2HDR(hdr) && HDR_L2_WRITING(hdr)) { 2626 ARCSTAT_BUMP(arcstat_evict_l2_skip); 2627 return (bytes_evicted); 2628 } 2629 2630 ARCSTAT_BUMP(arcstat_deleted); 2631 bytes_evicted += hdr->b_size; 2632 2633 DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, hdr); 2634 2635 if (HDR_HAS_L2HDR(hdr)) { 2636 /* 2637 * This buffer is cached on the 2nd Level ARC; 2638 * don't destroy the header. 2639 */ 2640 arc_change_state(arc_l2c_only, hdr, hash_lock); 2641 /* 2642 * dropping from L1+L2 cached to L2-only, 2643 * realloc to remove the L1 header. 2644 */ 2645 hdr = arc_hdr_realloc(hdr, hdr_full_cache, 2646 hdr_l2only_cache); 2647 } else { 2648 arc_change_state(arc_anon, hdr, hash_lock); 2649 arc_hdr_destroy(hdr); 2650 } 2651 return (bytes_evicted); 2652 } 2653 2654 ASSERT(state == arc_mru || state == arc_mfu); 2655 evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost; 2656 2657 /* prefetch buffers have a minimum lifespan */ 2658 if (HDR_IO_IN_PROGRESS(hdr) || 2659 ((hdr->b_flags & (ARC_FLAG_PREFETCH | ARC_FLAG_INDIRECT)) && 2660 ddi_get_lbolt() - hdr->b_l1hdr.b_arc_access < 2661 arc_min_prefetch_lifespan)) { 2662 ARCSTAT_BUMP(arcstat_evict_skip); 2663 return (bytes_evicted); 2664 } 2665 2666 ASSERT0(refcount_count(&hdr->b_l1hdr.b_refcnt)); 2667 ASSERT3U(hdr->b_l1hdr.b_datacnt, >, 0); 2668 while (hdr->b_l1hdr.b_buf) { 2669 arc_buf_t *buf = hdr->b_l1hdr.b_buf; 2670 if (!mutex_tryenter(&buf->b_evict_lock)) { 2671 ARCSTAT_BUMP(arcstat_mutex_miss); 2672 break; 2673 } 2674 if (buf->b_data != NULL) 2675 bytes_evicted += hdr->b_size; 2676 if (buf->b_efunc != NULL) { 2677 mutex_enter(&arc_user_evicts_lock); 2678 arc_buf_destroy(buf, FALSE); 2679 hdr->b_l1hdr.b_buf = buf->b_next; 2680 buf->b_hdr = &arc_eviction_hdr; 2681 buf->b_next = arc_eviction_list; 2682 arc_eviction_list = buf; 2683 cv_signal(&arc_user_evicts_cv); 2684 mutex_exit(&arc_user_evicts_lock); 2685 mutex_exit(&buf->b_evict_lock); 2686 } else { 2687 mutex_exit(&buf->b_evict_lock); 2688 arc_buf_destroy(buf, TRUE); 2689 } 2690 } 2691 2692 if (HDR_HAS_L2HDR(hdr)) { 2693 ARCSTAT_INCR(arcstat_evict_l2_cached, hdr->b_size); 2694 } else { 2695 if (l2arc_write_eligible(hdr->b_spa, hdr)) 2696 ARCSTAT_INCR(arcstat_evict_l2_eligible, hdr->b_size); 2697 else 2698 ARCSTAT_INCR(arcstat_evict_l2_ineligible, hdr->b_size); 2699 } 2700 2701 if (hdr->b_l1hdr.b_datacnt == 0) { 2702 arc_change_state(evicted_state, hdr, hash_lock); 2703 ASSERT(HDR_IN_HASH_TABLE(hdr)); 2704 hdr->b_flags |= ARC_FLAG_IN_HASH_TABLE; 2705 hdr->b_flags &= ~ARC_FLAG_BUF_AVAILABLE; 2706 DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, hdr); 2707 } 2708 2709 return (bytes_evicted); 2710} 2711 2712static uint64_t 2713arc_evict_state_impl(multilist_t *ml, int idx, arc_buf_hdr_t *marker, 2714 uint64_t spa, int64_t bytes) 2715{ 2716 multilist_sublist_t *mls; 2717 uint64_t bytes_evicted = 0; 2718 arc_buf_hdr_t *hdr; 2719 kmutex_t *hash_lock; 2720 int evict_count = 0; 2721 2722 ASSERT3P(marker, !=, NULL); 2723 IMPLY(bytes < 0, bytes == ARC_EVICT_ALL); 2724 2725 mls = multilist_sublist_lock(ml, idx); 2726 2727 for (hdr = multilist_sublist_prev(mls, marker); hdr != NULL; 2728 hdr = multilist_sublist_prev(mls, marker)) { 2729 if ((bytes != ARC_EVICT_ALL && bytes_evicted >= bytes) || 2730 (evict_count >= zfs_arc_evict_batch_limit)) 2731 break; 2732 2733 /* 2734 * To keep our iteration location, move the marker 2735 * forward. Since we're not holding hdr's hash lock, we 2736 * must be very careful and not remove 'hdr' from the 2737 * sublist. Otherwise, other consumers might mistake the 2738 * 'hdr' as not being on a sublist when they call the 2739 * multilist_link_active() function (they all rely on 2740 * the hash lock protecting concurrent insertions and 2741 * removals). multilist_sublist_move_forward() was 2742 * specifically implemented to ensure this is the case 2743 * (only 'marker' will be removed and re-inserted). 2744 */ 2745 multilist_sublist_move_forward(mls, marker); 2746 2747 /* 2748 * The only case where the b_spa field should ever be 2749 * zero, is the marker headers inserted by 2750 * arc_evict_state(). It's possible for multiple threads 2751 * to be calling arc_evict_state() concurrently (e.g. 2752 * dsl_pool_close() and zio_inject_fault()), so we must 2753 * skip any markers we see from these other threads. 2754 */ 2755 if (hdr->b_spa == 0) 2756 continue; 2757 2758 /* we're only interested in evicting buffers of a certain spa */ 2759 if (spa != 0 && hdr->b_spa != spa) { 2760 ARCSTAT_BUMP(arcstat_evict_skip); 2761 continue; 2762 } 2763 2764 hash_lock = HDR_LOCK(hdr); 2765 2766 /* 2767 * We aren't calling this function from any code path 2768 * that would already be holding a hash lock, so we're 2769 * asserting on this assumption to be defensive in case 2770 * this ever changes. Without this check, it would be 2771 * possible to incorrectly increment arcstat_mutex_miss 2772 * below (e.g. if the code changed such that we called 2773 * this function with a hash lock held). 2774 */ 2775 ASSERT(!MUTEX_HELD(hash_lock)); 2776 2777 if (mutex_tryenter(hash_lock)) { 2778 uint64_t evicted = arc_evict_hdr(hdr, hash_lock); 2779 mutex_exit(hash_lock); 2780 2781 bytes_evicted += evicted; 2782 2783 /* 2784 * If evicted is zero, arc_evict_hdr() must have 2785 * decided to skip this header, don't increment 2786 * evict_count in this case. 2787 */ 2788 if (evicted != 0) 2789 evict_count++; 2790 2791 /* 2792 * If arc_size isn't overflowing, signal any 2793 * threads that might happen to be waiting. 2794 * 2795 * For each header evicted, we wake up a single 2796 * thread. If we used cv_broadcast, we could 2797 * wake up "too many" threads causing arc_size 2798 * to significantly overflow arc_c; since 2799 * arc_get_data_buf() doesn't check for overflow 2800 * when it's woken up (it doesn't because it's 2801 * possible for the ARC to be overflowing while 2802 * full of un-evictable buffers, and the 2803 * function should proceed in this case). 2804 * 2805 * If threads are left sleeping, due to not 2806 * using cv_broadcast, they will be woken up 2807 * just before arc_reclaim_thread() sleeps. 2808 */ 2809 mutex_enter(&arc_reclaim_lock); 2810 if (!arc_is_overflowing()) 2811 cv_signal(&arc_reclaim_waiters_cv); 2812 mutex_exit(&arc_reclaim_lock); 2813 } else { 2814 ARCSTAT_BUMP(arcstat_mutex_miss); 2815 } 2816 } 2817 2818 multilist_sublist_unlock(mls); 2819 2820 return (bytes_evicted); 2821} 2822 2823/* 2824 * Evict buffers from the given arc state, until we've removed the 2825 * specified number of bytes. Move the removed buffers to the 2826 * appropriate evict state. 2827 * 2828 * This function makes a "best effort". It skips over any buffers 2829 * it can't get a hash_lock on, and so, may not catch all candidates. 2830 * It may also return without evicting as much space as requested. 2831 * 2832 * If bytes is specified using the special value ARC_EVICT_ALL, this 2833 * will evict all available (i.e. unlocked and evictable) buffers from 2834 * the given arc state; which is used by arc_flush(). 2835 */ 2836static uint64_t 2837arc_evict_state(arc_state_t *state, uint64_t spa, int64_t bytes, 2838 arc_buf_contents_t type) 2839{ 2840 uint64_t total_evicted = 0; 2841 multilist_t *ml = &state->arcs_list[type]; 2842 int num_sublists; 2843 arc_buf_hdr_t **markers; 2844 2845 IMPLY(bytes < 0, bytes == ARC_EVICT_ALL); 2846 2847 num_sublists = multilist_get_num_sublists(ml); 2848 2849 /* 2850 * If we've tried to evict from each sublist, made some 2851 * progress, but still have not hit the target number of bytes 2852 * to evict, we want to keep trying. The markers allow us to 2853 * pick up where we left off for each individual sublist, rather 2854 * than starting from the tail each time. 2855 */ 2856 markers = kmem_zalloc(sizeof (*markers) * num_sublists, KM_SLEEP); 2857 for (int i = 0; i < num_sublists; i++) { 2858 markers[i] = kmem_cache_alloc(hdr_full_cache, KM_SLEEP); 2859 2860 /* 2861 * A b_spa of 0 is used to indicate that this header is 2862 * a marker. This fact is used in arc_adjust_type() and 2863 * arc_evict_state_impl(). 2864 */ 2865 markers[i]->b_spa = 0; 2866 2867 multilist_sublist_t *mls = multilist_sublist_lock(ml, i); 2868 multilist_sublist_insert_tail(mls, markers[i]); 2869 multilist_sublist_unlock(mls); 2870 } 2871 2872 /* 2873 * While we haven't hit our target number of bytes to evict, or 2874 * we're evicting all available buffers. 2875 */ 2876 while (total_evicted < bytes || bytes == ARC_EVICT_ALL) { 2877 /* 2878 * Start eviction using a randomly selected sublist, 2879 * this is to try and evenly balance eviction across all 2880 * sublists. Always starting at the same sublist 2881 * (e.g. index 0) would cause evictions to favor certain 2882 * sublists over others. 2883 */ 2884 int sublist_idx = multilist_get_random_index(ml); 2885 uint64_t scan_evicted = 0; 2886 2887 for (int i = 0; i < num_sublists; i++) { 2888 uint64_t bytes_remaining; 2889 uint64_t bytes_evicted; 2890 2891 if (bytes == ARC_EVICT_ALL) 2892 bytes_remaining = ARC_EVICT_ALL; 2893 else if (total_evicted < bytes) 2894 bytes_remaining = bytes - total_evicted; 2895 else 2896 break; 2897 2898 bytes_evicted = arc_evict_state_impl(ml, sublist_idx, 2899 markers[sublist_idx], spa, bytes_remaining); 2900 2901 scan_evicted += bytes_evicted; 2902 total_evicted += bytes_evicted; 2903 2904 /* we've reached the end, wrap to the beginning */ 2905 if (++sublist_idx >= num_sublists) 2906 sublist_idx = 0; 2907 } 2908 2909 /* 2910 * If we didn't evict anything during this scan, we have 2911 * no reason to believe we'll evict more during another 2912 * scan, so break the loop. 2913 */ 2914 if (scan_evicted == 0) { 2915 /* This isn't possible, let's make that obvious */ 2916 ASSERT3S(bytes, !=, 0); 2917 2918 /* 2919 * When bytes is ARC_EVICT_ALL, the only way to 2920 * break the loop is when scan_evicted is zero. 2921 * In that case, we actually have evicted enough, 2922 * so we don't want to increment the kstat. 2923 */ 2924 if (bytes != ARC_EVICT_ALL) { 2925 ASSERT3S(total_evicted, <, bytes); 2926 ARCSTAT_BUMP(arcstat_evict_not_enough); 2927 } 2928 2929 break; 2930 } 2931 } 2932 2933 for (int i = 0; i < num_sublists; i++) { 2934 multilist_sublist_t *mls = multilist_sublist_lock(ml, i); 2935 multilist_sublist_remove(mls, markers[i]); 2936 multilist_sublist_unlock(mls); 2937 2938 kmem_cache_free(hdr_full_cache, markers[i]); 2939 } 2940 kmem_free(markers, sizeof (*markers) * num_sublists); 2941 2942 return (total_evicted); 2943} 2944 2945/* 2946 * Flush all "evictable" data of the given type from the arc state 2947 * specified. This will not evict any "active" buffers (i.e. referenced). 2948 * 2949 * When 'retry' is set to FALSE, the function will make a single pass 2950 * over the state and evict any buffers that it can. Since it doesn't 2951 * continually retry the eviction, it might end up leaving some buffers 2952 * in the ARC due to lock misses. 2953 * 2954 * When 'retry' is set to TRUE, the function will continually retry the 2955 * eviction until *all* evictable buffers have been removed from the 2956 * state. As a result, if concurrent insertions into the state are 2957 * allowed (e.g. if the ARC isn't shutting down), this function might 2958 * wind up in an infinite loop, continually trying to evict buffers. 2959 */ 2960static uint64_t 2961arc_flush_state(arc_state_t *state, uint64_t spa, arc_buf_contents_t type, 2962 boolean_t retry) 2963{ 2964 uint64_t evicted = 0; 2965 2966 while (state->arcs_lsize[type] != 0) { 2967 evicted += arc_evict_state(state, spa, ARC_EVICT_ALL, type); 2968 2969 if (!retry) 2970 break; 2971 } 2972 2973 return (evicted); 2974} 2975 2976/* 2977 * Evict the specified number of bytes from the state specified, 2978 * restricting eviction to the spa and type given. This function 2979 * prevents us from trying to evict more from a state's list than 2980 * is "evictable", and to skip evicting altogether when passed a 2981 * negative value for "bytes". In contrast, arc_evict_state() will 2982 * evict everything it can, when passed a negative value for "bytes". 2983 */ 2984static uint64_t 2985arc_adjust_impl(arc_state_t *state, uint64_t spa, int64_t bytes, 2986 arc_buf_contents_t type) 2987{ 2988 int64_t delta; 2989 2990 if (bytes > 0 && state->arcs_lsize[type] > 0) { 2991 delta = MIN(state->arcs_lsize[type], bytes); 2992 return (arc_evict_state(state, spa, delta, type)); 2993 } 2994 2995 return (0); 2996} 2997 2998/* 2999 * Evict metadata buffers from the cache, such that arc_meta_used is 3000 * capped by the arc_meta_limit tunable. 3001 */ 3002static uint64_t 3003arc_adjust_meta(void) 3004{ 3005 uint64_t total_evicted = 0; 3006 int64_t target; 3007 3008 /* 3009 * If we're over the meta limit, we want to evict enough 3010 * metadata to get back under the meta limit. We don't want to 3011 * evict so much that we drop the MRU below arc_p, though. If 3012 * we're over the meta limit more than we're over arc_p, we 3013 * evict some from the MRU here, and some from the MFU below. 3014 */ 3015 target = MIN((int64_t)(arc_meta_used - arc_meta_limit), 3016 (int64_t)(refcount_count(&arc_anon->arcs_size) + 3017 refcount_count(&arc_mru->arcs_size) - arc_p)); 3018 3019 total_evicted += arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA); 3020 3021 /* 3022 * Similar to the above, we want to evict enough bytes to get us 3023 * below the meta limit, but not so much as to drop us below the 3024 * space alloted to the MFU (which is defined as arc_c - arc_p). 3025 */ 3026 target = MIN((int64_t)(arc_meta_used - arc_meta_limit), 3027 (int64_t)(refcount_count(&arc_mfu->arcs_size) - (arc_c - arc_p))); 3028 3029 total_evicted += arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA); 3030 3031 return (total_evicted); 3032} 3033 3034/* 3035 * Return the type of the oldest buffer in the given arc state 3036 * 3037 * This function will select a random sublist of type ARC_BUFC_DATA and 3038 * a random sublist of type ARC_BUFC_METADATA. The tail of each sublist 3039 * is compared, and the type which contains the "older" buffer will be 3040 * returned. 3041 */ 3042static arc_buf_contents_t 3043arc_adjust_type(arc_state_t *state) 3044{ 3045 multilist_t *data_ml = &state->arcs_list[ARC_BUFC_DATA]; 3046 multilist_t *meta_ml = &state->arcs_list[ARC_BUFC_METADATA]; 3047 int data_idx = multilist_get_random_index(data_ml); 3048 int meta_idx = multilist_get_random_index(meta_ml); 3049 multilist_sublist_t *data_mls; 3050 multilist_sublist_t *meta_mls; 3051 arc_buf_contents_t type; 3052 arc_buf_hdr_t *data_hdr; 3053 arc_buf_hdr_t *meta_hdr; 3054 3055 /* 3056 * We keep the sublist lock until we're finished, to prevent 3057 * the headers from being destroyed via arc_evict_state(). 3058 */ 3059 data_mls = multilist_sublist_lock(data_ml, data_idx); 3060 meta_mls = multilist_sublist_lock(meta_ml, meta_idx); 3061 3062 /* 3063 * These two loops are to ensure we skip any markers that 3064 * might be at the tail of the lists due to arc_evict_state(). 3065 */ 3066 3067 for (data_hdr = multilist_sublist_tail(data_mls); data_hdr != NULL; 3068 data_hdr = multilist_sublist_prev(data_mls, data_hdr)) { 3069 if (data_hdr->b_spa != 0) 3070 break; 3071 } 3072 3073 for (meta_hdr = multilist_sublist_tail(meta_mls); meta_hdr != NULL; 3074 meta_hdr = multilist_sublist_prev(meta_mls, meta_hdr)) { 3075 if (meta_hdr->b_spa != 0) 3076 break; 3077 } 3078 3079 if (data_hdr == NULL && meta_hdr == NULL) { 3080 type = ARC_BUFC_DATA; 3081 } else if (data_hdr == NULL) { 3082 ASSERT3P(meta_hdr, !=, NULL); 3083 type = ARC_BUFC_METADATA; 3084 } else if (meta_hdr == NULL) { 3085 ASSERT3P(data_hdr, !=, NULL); 3086 type = ARC_BUFC_DATA; 3087 } else { 3088 ASSERT3P(data_hdr, !=, NULL); 3089 ASSERT3P(meta_hdr, !=, NULL); 3090 3091 /* The headers can't be on the sublist without an L1 header */ 3092 ASSERT(HDR_HAS_L1HDR(data_hdr)); 3093 ASSERT(HDR_HAS_L1HDR(meta_hdr)); 3094 3095 if (data_hdr->b_l1hdr.b_arc_access < 3096 meta_hdr->b_l1hdr.b_arc_access) { 3097 type = ARC_BUFC_DATA; 3098 } else { 3099 type = ARC_BUFC_METADATA; 3100 } 3101 } 3102 3103 multilist_sublist_unlock(meta_mls); 3104 multilist_sublist_unlock(data_mls); 3105 3106 return (type); 3107} 3108 3109/* 3110 * Evict buffers from the cache, such that arc_size is capped by arc_c. 3111 */ 3112static uint64_t 3113arc_adjust(void) 3114{ 3115 uint64_t total_evicted = 0; 3116 uint64_t bytes; 3117 int64_t target; 3118 3119 /* 3120 * If we're over arc_meta_limit, we want to correct that before 3121 * potentially evicting data buffers below. 3122 */ 3123 total_evicted += arc_adjust_meta(); 3124 3125 /* 3126 * Adjust MRU size 3127 * 3128 * If we're over the target cache size, we want to evict enough 3129 * from the list to get back to our target size. We don't want 3130 * to evict too much from the MRU, such that it drops below 3131 * arc_p. So, if we're over our target cache size more than 3132 * the MRU is over arc_p, we'll evict enough to get back to 3133 * arc_p here, and then evict more from the MFU below. 3134 */ 3135 target = MIN((int64_t)(arc_size - arc_c), 3136 (int64_t)(refcount_count(&arc_anon->arcs_size) + 3137 refcount_count(&arc_mru->arcs_size) + arc_meta_used - arc_p)); 3138 3139 /* 3140 * If we're below arc_meta_min, always prefer to evict data. 3141 * Otherwise, try to satisfy the requested number of bytes to 3142 * evict from the type which contains older buffers; in an 3143 * effort to keep newer buffers in the cache regardless of their 3144 * type. If we cannot satisfy the number of bytes from this 3145 * type, spill over into the next type. 3146 */ 3147 if (arc_adjust_type(arc_mru) == ARC_BUFC_METADATA && 3148 arc_meta_used > arc_meta_min) { 3149 bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA); 3150 total_evicted += bytes; 3151 3152 /* 3153 * If we couldn't evict our target number of bytes from 3154 * metadata, we try to get the rest from data. 3155 */ 3156 target -= bytes; 3157 3158 total_evicted += 3159 arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA); 3160 } else { 3161 bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA); 3162 total_evicted += bytes; 3163 3164 /* 3165 * If we couldn't evict our target number of bytes from 3166 * data, we try to get the rest from metadata. 3167 */ 3168 target -= bytes; 3169 3170 total_evicted += 3171 arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA); 3172 } 3173 3174 /* 3175 * Adjust MFU size 3176 * 3177 * Now that we've tried to evict enough from the MRU to get its 3178 * size back to arc_p, if we're still above the target cache 3179 * size, we evict the rest from the MFU. 3180 */ 3181 target = arc_size - arc_c; 3182 3183 if (arc_adjust_type(arc_mfu) == ARC_BUFC_METADATA && 3184 arc_meta_used > arc_meta_min) { 3185 bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA); 3186 total_evicted += bytes; 3187 3188 /* 3189 * If we couldn't evict our target number of bytes from 3190 * metadata, we try to get the rest from data. 3191 */ 3192 target -= bytes; 3193 3194 total_evicted += 3195 arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA); 3196 } else { 3197 bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA); 3198 total_evicted += bytes; 3199 3200 /* 3201 * If we couldn't evict our target number of bytes from 3202 * data, we try to get the rest from data. 3203 */ 3204 target -= bytes; 3205 3206 total_evicted += 3207 arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA); 3208 } 3209 3210 /* 3211 * Adjust ghost lists 3212 * 3213 * In addition to the above, the ARC also defines target values 3214 * for the ghost lists. The sum of the mru list and mru ghost 3215 * list should never exceed the target size of the cache, and 3216 * the sum of the mru list, mfu list, mru ghost list, and mfu 3217 * ghost list should never exceed twice the target size of the 3218 * cache. The following logic enforces these limits on the ghost 3219 * caches, and evicts from them as needed. 3220 */ 3221 target = refcount_count(&arc_mru->arcs_size) + 3222 refcount_count(&arc_mru_ghost->arcs_size) - arc_c; 3223 3224 bytes = arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_DATA); 3225 total_evicted += bytes; 3226 3227 target -= bytes; 3228 3229 total_evicted += 3230 arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_METADATA); 3231 3232 /* 3233 * We assume the sum of the mru list and mfu list is less than 3234 * or equal to arc_c (we enforced this above), which means we 3235 * can use the simpler of the two equations below: 3236 * 3237 * mru + mfu + mru ghost + mfu ghost <= 2 * arc_c 3238 * mru ghost + mfu ghost <= arc_c 3239 */ 3240 target = refcount_count(&arc_mru_ghost->arcs_size) + 3241 refcount_count(&arc_mfu_ghost->arcs_size) - arc_c; 3242 3243 bytes = arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_DATA); 3244 total_evicted += bytes; 3245 3246 target -= bytes; 3247 3248 total_evicted += 3249 arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_METADATA); 3250 3251 return (total_evicted); 3252} 3253 3254static void 3255arc_do_user_evicts(void) 3256{ 3257 mutex_enter(&arc_user_evicts_lock); 3258 while (arc_eviction_list != NULL) { 3259 arc_buf_t *buf = arc_eviction_list; 3260 arc_eviction_list = buf->b_next; 3261 mutex_enter(&buf->b_evict_lock); 3262 buf->b_hdr = NULL; 3263 mutex_exit(&buf->b_evict_lock); 3264 mutex_exit(&arc_user_evicts_lock); 3265 3266 if (buf->b_efunc != NULL) 3267 VERIFY0(buf->b_efunc(buf->b_private)); 3268 3269 buf->b_efunc = NULL; 3270 buf->b_private = NULL; 3271 kmem_cache_free(buf_cache, buf); 3272 mutex_enter(&arc_user_evicts_lock); 3273 } 3274 mutex_exit(&arc_user_evicts_lock); 3275} 3276 3277void 3278arc_flush(spa_t *spa, boolean_t retry) 3279{ 3280 uint64_t guid = 0; 3281 3282 /* 3283 * If retry is TRUE, a spa must not be specified since we have 3284 * no good way to determine if all of a spa's buffers have been 3285 * evicted from an arc state. 3286 */ 3287 ASSERT(!retry || spa == 0); 3288 3289 if (spa != NULL) 3290 guid = spa_load_guid(spa); 3291 3292 (void) arc_flush_state(arc_mru, guid, ARC_BUFC_DATA, retry); 3293 (void) arc_flush_state(arc_mru, guid, ARC_BUFC_METADATA, retry); 3294 3295 (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_DATA, retry); 3296 (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_METADATA, retry); 3297 3298 (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_DATA, retry); 3299 (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_METADATA, retry); 3300 3301 (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_DATA, retry); 3302 (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_METADATA, retry); 3303 3304 arc_do_user_evicts(); 3305 ASSERT(spa || arc_eviction_list == NULL); 3306} 3307 3308void 3309arc_shrink(int64_t to_free) 3310{ 3311 if (arc_c > arc_c_min) { 3312 DTRACE_PROBE4(arc__shrink, uint64_t, arc_c, uint64_t, 3313 arc_c_min, uint64_t, arc_p, uint64_t, to_free); 3314 if (arc_c > arc_c_min + to_free) 3315 atomic_add_64(&arc_c, -to_free); 3316 else 3317 arc_c = arc_c_min; 3318 3319 atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift)); 3320 if (arc_c > arc_size) 3321 arc_c = MAX(arc_size, arc_c_min); 3322 if (arc_p > arc_c) 3323 arc_p = (arc_c >> 1); 3324 3325 DTRACE_PROBE2(arc__shrunk, uint64_t, arc_c, uint64_t, 3326 arc_p); 3327 3328 ASSERT(arc_c >= arc_c_min); 3329 ASSERT((int64_t)arc_p >= 0); 3330 } 3331 3332 if (arc_size > arc_c) { 3333 DTRACE_PROBE2(arc__shrink_adjust, uint64_t, arc_size, 3334 uint64_t, arc_c); 3335 (void) arc_adjust(); 3336 } 3337} 3338 3339static long needfree = 0; 3340 3341typedef enum free_memory_reason_t { 3342 FMR_UNKNOWN, 3343 FMR_NEEDFREE, 3344 FMR_LOTSFREE, 3345 FMR_SWAPFS_MINFREE, 3346 FMR_PAGES_PP_MAXIMUM, 3347 FMR_HEAP_ARENA, 3348 FMR_ZIO_ARENA, 3349 FMR_ZIO_FRAG, 3350} free_memory_reason_t; 3351 3352int64_t last_free_memory; 3353free_memory_reason_t last_free_reason; 3354 3355/* 3356 * Additional reserve of pages for pp_reserve. 3357 */ 3358int64_t arc_pages_pp_reserve = 64; 3359 3360/* 3361 * Additional reserve of pages for swapfs. 3362 */ 3363int64_t arc_swapfs_reserve = 64; 3364 3365/* 3366 * Return the amount of memory that can be consumed before reclaim will be 3367 * needed. Positive if there is sufficient free memory, negative indicates 3368 * the amount of memory that needs to be freed up. 3369 */ 3370static int64_t 3371arc_available_memory(void) 3372{ 3373 int64_t lowest = INT64_MAX; 3374 int64_t n; 3375 free_memory_reason_t r = FMR_UNKNOWN; 3376 3377#ifdef _KERNEL 3378 if (needfree > 0) { 3379 n = PAGESIZE * (-needfree); 3380 if (n < lowest) { 3381 lowest = n; 3382 r = FMR_NEEDFREE; 3383 } 3384 } 3385 3386 /* 3387 * Cooperate with pagedaemon when it's time for it to scan 3388 * and reclaim some pages. 3389 */ 3390 n = PAGESIZE * ((int64_t)freemem - zfs_arc_free_target); 3391 if (n < lowest) { 3392 lowest = n; 3393 r = FMR_LOTSFREE; 3394 } 3395 3396#ifdef sun 3397 /* 3398 * check that we're out of range of the pageout scanner. It starts to 3399 * schedule paging if freemem is less than lotsfree and needfree. 3400 * lotsfree is the high-water mark for pageout, and needfree is the 3401 * number of needed free pages. We add extra pages here to make sure 3402 * the scanner doesn't start up while we're freeing memory. 3403 */ 3404 n = PAGESIZE * (freemem - lotsfree - needfree - desfree); 3405 if (n < lowest) { 3406 lowest = n; 3407 r = FMR_LOTSFREE; 3408 } 3409 3410 /* 3411 * check to make sure that swapfs has enough space so that anon 3412 * reservations can still succeed. anon_resvmem() checks that the 3413 * availrmem is greater than swapfs_minfree, and the number of reserved 3414 * swap pages. We also add a bit of extra here just to prevent 3415 * circumstances from getting really dire. 3416 */ 3417 n = PAGESIZE * (availrmem - swapfs_minfree - swapfs_reserve - 3418 desfree - arc_swapfs_reserve); 3419 if (n < lowest) { 3420 lowest = n; 3421 r = FMR_SWAPFS_MINFREE; 3422 } 3423 3424 3425 /* 3426 * Check that we have enough availrmem that memory locking (e.g., via 3427 * mlock(3C) or memcntl(2)) can still succeed. (pages_pp_maximum 3428 * stores the number of pages that cannot be locked; when availrmem 3429 * drops below pages_pp_maximum, page locking mechanisms such as 3430 * page_pp_lock() will fail.) 3431 */ 3432 n = PAGESIZE * (availrmem - pages_pp_maximum - 3433 arc_pages_pp_reserve); 3434 if (n < lowest) { 3435 lowest = n; 3436 r = FMR_PAGES_PP_MAXIMUM; 3437 } 3438 3439#endif /* sun */ 3440#if defined(__i386) || !defined(UMA_MD_SMALL_ALLOC) 3441 /* 3442 * If we're on an i386 platform, it's possible that we'll exhaust the 3443 * kernel heap space before we ever run out of available physical 3444 * memory. Most checks of the size of the heap_area compare against 3445 * tune.t_minarmem, which is the minimum available real memory that we 3446 * can have in the system. However, this is generally fixed at 25 pages 3447 * which is so low that it's useless. In this comparison, we seek to 3448 * calculate the total heap-size, and reclaim if more than 3/4ths of the 3449 * heap is allocated. (Or, in the calculation, if less than 1/4th is 3450 * free) 3451 */ 3452 n = (int64_t)vmem_size(heap_arena, VMEM_FREE) - 3453 (vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC) >> 2); 3454 if (n < lowest) { 3455 lowest = n; 3456 r = FMR_HEAP_ARENA; 3457 } 3458#define zio_arena NULL 3459#else 3460#define zio_arena heap_arena 3461#endif 3462 3463 /* 3464 * If zio data pages are being allocated out of a separate heap segment, 3465 * then enforce that the size of available vmem for this arena remains 3466 * above about 1/16th free. 3467 * 3468 * Note: The 1/16th arena free requirement was put in place 3469 * to aggressively evict memory from the arc in order to avoid 3470 * memory fragmentation issues. 3471 */ 3472 if (zio_arena != NULL) { 3473 n = (int64_t)vmem_size(zio_arena, VMEM_FREE) - 3474 (vmem_size(zio_arena, VMEM_ALLOC) >> 4); 3475 if (n < lowest) { 3476 lowest = n; 3477 r = FMR_ZIO_ARENA; 3478 } 3479 } 3480 3481 /* 3482 * Above limits know nothing about real level of KVA fragmentation. 3483 * Start aggressive reclamation if too little sequential KVA left. 3484 */ 3485 if (lowest > 0) { 3486 n = (vmem_size(heap_arena, VMEM_MAXFREE) < zfs_max_recordsize) ? 3487 -((int64_t)vmem_size(heap_arena, VMEM_ALLOC) >> 4) : 3488 INT64_MAX; 3489 if (n < lowest) { 3490 lowest = n; 3491 r = FMR_ZIO_FRAG; 3492 } 3493 } 3494 3495#else /* _KERNEL */ 3496 /* Every 100 calls, free a small amount */ 3497 if (spa_get_random(100) == 0) 3498 lowest = -1024; 3499#endif /* _KERNEL */ 3500 3501 last_free_memory = lowest; 3502 last_free_reason = r; 3503 DTRACE_PROBE2(arc__available_memory, int64_t, lowest, int, r); 3504 return (lowest); 3505} 3506 3507 3508/* 3509 * Determine if the system is under memory pressure and is asking 3510 * to reclaim memory. A return value of TRUE indicates that the system 3511 * is under memory pressure and that the arc should adjust accordingly. 3512 */ 3513static boolean_t 3514arc_reclaim_needed(void) 3515{ 3516 return (arc_available_memory() < 0); 3517} 3518 3519extern kmem_cache_t *zio_buf_cache[]; 3520extern kmem_cache_t *zio_data_buf_cache[]; 3521extern kmem_cache_t *range_seg_cache; 3522 3523static __noinline void 3524arc_kmem_reap_now(void) 3525{ 3526 size_t i; 3527 kmem_cache_t *prev_cache = NULL; 3528 kmem_cache_t *prev_data_cache = NULL; 3529 3530 DTRACE_PROBE(arc__kmem_reap_start); 3531#ifdef _KERNEL 3532 if (arc_meta_used >= arc_meta_limit) { 3533 /* 3534 * We are exceeding our meta-data cache limit. 3535 * Purge some DNLC entries to release holds on meta-data. 3536 */ 3537 dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent); 3538 } 3539#if defined(__i386) 3540 /* 3541 * Reclaim unused memory from all kmem caches. 3542 */ 3543 kmem_reap(); 3544#endif 3545#endif 3546 3547 for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) { 3548 if (zio_buf_cache[i] != prev_cache) { 3549 prev_cache = zio_buf_cache[i]; 3550 kmem_cache_reap_now(zio_buf_cache[i]); 3551 } 3552 if (zio_data_buf_cache[i] != prev_data_cache) { 3553 prev_data_cache = zio_data_buf_cache[i]; 3554 kmem_cache_reap_now(zio_data_buf_cache[i]); 3555 } 3556 } 3557 kmem_cache_reap_now(buf_cache); 3558 kmem_cache_reap_now(hdr_full_cache); 3559 kmem_cache_reap_now(hdr_l2only_cache); 3560 kmem_cache_reap_now(range_seg_cache); 3561 3562#ifdef sun 3563 if (zio_arena != NULL) { 3564 /* 3565 * Ask the vmem arena to reclaim unused memory from its 3566 * quantum caches. 3567 */ 3568 vmem_qcache_reap(zio_arena); 3569 } 3570#endif 3571 DTRACE_PROBE(arc__kmem_reap_end); 3572} 3573 3574/* 3575 * Threads can block in arc_get_data_buf() waiting for this thread to evict 3576 * enough data and signal them to proceed. When this happens, the threads in 3577 * arc_get_data_buf() are sleeping while holding the hash lock for their 3578 * particular arc header. Thus, we must be careful to never sleep on a 3579 * hash lock in this thread. This is to prevent the following deadlock: 3580 * 3581 * - Thread A sleeps on CV in arc_get_data_buf() holding hash lock "L", 3582 * waiting for the reclaim thread to signal it. 3583 * 3584 * - arc_reclaim_thread() tries to acquire hash lock "L" using mutex_enter, 3585 * fails, and goes to sleep forever. 3586 * 3587 * This possible deadlock is avoided by always acquiring a hash lock 3588 * using mutex_tryenter() from arc_reclaim_thread(). 3589 */ 3590static void 3591arc_reclaim_thread(void *dummy __unused) 3592{ 3593 clock_t growtime = 0; 3594 callb_cpr_t cpr; 3595 3596 CALLB_CPR_INIT(&cpr, &arc_reclaim_lock, callb_generic_cpr, FTAG); 3597 3598 mutex_enter(&arc_reclaim_lock); 3599 while (!arc_reclaim_thread_exit) { 3600 int64_t free_memory = arc_available_memory(); 3601 uint64_t evicted = 0; 3602 3603 mutex_exit(&arc_reclaim_lock); 3604 3605 if (free_memory < 0) { 3606 3607 arc_no_grow = B_TRUE; 3608 arc_warm = B_TRUE; 3609 3610 /* 3611 * Wait at least zfs_grow_retry (default 60) seconds 3612 * before considering growing. 3613 */ 3614 growtime = ddi_get_lbolt() + (arc_grow_retry * hz); 3615 3616 arc_kmem_reap_now(); 3617 3618 /* 3619 * If we are still low on memory, shrink the ARC 3620 * so that we have arc_shrink_min free space. 3621 */ 3622 free_memory = arc_available_memory(); 3623 3624 int64_t to_free = 3625 (arc_c >> arc_shrink_shift) - free_memory; 3626 if (to_free > 0) { 3627#ifdef _KERNEL 3628 to_free = MAX(to_free, ptob(needfree)); 3629#endif 3630 arc_shrink(to_free); 3631 } 3632 } else if (free_memory < arc_c >> arc_no_grow_shift) { 3633 arc_no_grow = B_TRUE; 3634 } else if (ddi_get_lbolt() >= growtime) { 3635 arc_no_grow = B_FALSE; 3636 } 3637 3638 evicted = arc_adjust(); 3639 3640 mutex_enter(&arc_reclaim_lock); 3641 3642 /* 3643 * If evicted is zero, we couldn't evict anything via 3644 * arc_adjust(). This could be due to hash lock 3645 * collisions, but more likely due to the majority of 3646 * arc buffers being unevictable. Therefore, even if 3647 * arc_size is above arc_c, another pass is unlikely to 3648 * be helpful and could potentially cause us to enter an 3649 * infinite loop. 3650 */ 3651 if (arc_size <= arc_c || evicted == 0) { 3652#ifdef _KERNEL 3653 needfree = 0; 3654#endif 3655 /* 3656 * We're either no longer overflowing, or we 3657 * can't evict anything more, so we should wake 3658 * up any threads before we go to sleep. 3659 */ 3660 cv_broadcast(&arc_reclaim_waiters_cv); 3661 3662 /* 3663 * Block until signaled, or after one second (we 3664 * might need to perform arc_kmem_reap_now() 3665 * even if we aren't being signalled) 3666 */ 3667 CALLB_CPR_SAFE_BEGIN(&cpr); 3668 (void) cv_timedwait(&arc_reclaim_thread_cv, 3669 &arc_reclaim_lock, hz); 3670 CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_lock); 3671 } 3672 } 3673 3674 arc_reclaim_thread_exit = FALSE; 3675 cv_broadcast(&arc_reclaim_thread_cv); 3676 CALLB_CPR_EXIT(&cpr); /* drops arc_reclaim_lock */ 3677 thread_exit(); 3678} 3679 3680static void 3681arc_user_evicts_thread(void *dummy __unused) 3682{ 3683 callb_cpr_t cpr; 3684 3685 CALLB_CPR_INIT(&cpr, &arc_user_evicts_lock, callb_generic_cpr, FTAG); 3686 3687 mutex_enter(&arc_user_evicts_lock); 3688 while (!arc_user_evicts_thread_exit) { 3689 mutex_exit(&arc_user_evicts_lock); 3690 3691 arc_do_user_evicts(); 3692 3693 /* 3694 * This is necessary in order for the mdb ::arc dcmd to 3695 * show up to date information. Since the ::arc command 3696 * does not call the kstat's update function, without 3697 * this call, the command may show stale stats for the 3698 * anon, mru, mru_ghost, mfu, and mfu_ghost lists. Even 3699 * with this change, the data might be up to 1 second 3700 * out of date; but that should suffice. The arc_state_t 3701 * structures can be queried directly if more accurate 3702 * information is needed. 3703 */ 3704 if (arc_ksp != NULL) 3705 arc_ksp->ks_update(arc_ksp, KSTAT_READ); 3706 3707 mutex_enter(&arc_user_evicts_lock); 3708 3709 /* 3710 * Block until signaled, or after one second (we need to 3711 * call the arc's kstat update function regularly). 3712 */ 3713 CALLB_CPR_SAFE_BEGIN(&cpr); 3714 (void) cv_timedwait(&arc_user_evicts_cv, 3715 &arc_user_evicts_lock, hz); 3716 CALLB_CPR_SAFE_END(&cpr, &arc_user_evicts_lock); 3717 } 3718 3719 arc_user_evicts_thread_exit = FALSE; 3720 cv_broadcast(&arc_user_evicts_cv); 3721 CALLB_CPR_EXIT(&cpr); /* drops arc_user_evicts_lock */ 3722 thread_exit(); 3723} 3724 3725/* 3726 * Adapt arc info given the number of bytes we are trying to add and 3727 * the state that we are comming from. This function is only called 3728 * when we are adding new content to the cache. 3729 */ 3730static void 3731arc_adapt(int bytes, arc_state_t *state) 3732{ 3733 int mult; 3734 uint64_t arc_p_min = (arc_c >> arc_p_min_shift); 3735 int64_t mrug_size = refcount_count(&arc_mru_ghost->arcs_size); 3736 int64_t mfug_size = refcount_count(&arc_mfu_ghost->arcs_size); 3737 3738 if (state == arc_l2c_only) 3739 return; 3740 3741 ASSERT(bytes > 0); 3742 /* 3743 * Adapt the target size of the MRU list: 3744 * - if we just hit in the MRU ghost list, then increase 3745 * the target size of the MRU list. 3746 * - if we just hit in the MFU ghost list, then increase 3747 * the target size of the MFU list by decreasing the 3748 * target size of the MRU list. 3749 */ 3750 if (state == arc_mru_ghost) { 3751 mult = (mrug_size >= mfug_size) ? 1 : (mfug_size / mrug_size); 3752 mult = MIN(mult, 10); /* avoid wild arc_p adjustment */ 3753 3754 arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult); 3755 } else if (state == arc_mfu_ghost) { 3756 uint64_t delta; 3757 3758 mult = (mfug_size >= mrug_size) ? 1 : (mrug_size / mfug_size); 3759 mult = MIN(mult, 10); 3760 3761 delta = MIN(bytes * mult, arc_p); 3762 arc_p = MAX(arc_p_min, arc_p - delta); 3763 } 3764 ASSERT((int64_t)arc_p >= 0); 3765 3766 if (arc_reclaim_needed()) { 3767 cv_signal(&arc_reclaim_thread_cv); 3768 return; 3769 } 3770 3771 if (arc_no_grow) 3772 return; 3773 3774 if (arc_c >= arc_c_max) 3775 return; 3776 3777 /* 3778 * If we're within (2 * maxblocksize) bytes of the target 3779 * cache size, increment the target cache size 3780 */ 3781 if (arc_size > arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) { 3782 DTRACE_PROBE1(arc__inc_adapt, int, bytes); 3783 atomic_add_64(&arc_c, (int64_t)bytes); 3784 if (arc_c > arc_c_max) 3785 arc_c = arc_c_max; 3786 else if (state == arc_anon) 3787 atomic_add_64(&arc_p, (int64_t)bytes); 3788 if (arc_p > arc_c) 3789 arc_p = arc_c; 3790 } 3791 ASSERT((int64_t)arc_p >= 0); 3792} 3793 3794/* 3795 * Check if arc_size has grown past our upper threshold, determined by 3796 * zfs_arc_overflow_shift. 3797 */ 3798static boolean_t 3799arc_is_overflowing(void) 3800{ 3801 /* Always allow at least one block of overflow */ 3802 uint64_t overflow = MAX(SPA_MAXBLOCKSIZE, 3803 arc_c >> zfs_arc_overflow_shift); 3804 3805 return (arc_size >= arc_c + overflow); 3806} 3807 3808/* 3809 * The buffer, supplied as the first argument, needs a data block. If we 3810 * are hitting the hard limit for the cache size, we must sleep, waiting 3811 * for the eviction thread to catch up. If we're past the target size 3812 * but below the hard limit, we'll only signal the reclaim thread and 3813 * continue on. 3814 */ 3815static void 3816arc_get_data_buf(arc_buf_t *buf) 3817{ 3818 arc_state_t *state = buf->b_hdr->b_l1hdr.b_state; 3819 uint64_t size = buf->b_hdr->b_size; 3820 arc_buf_contents_t type = arc_buf_type(buf->b_hdr); 3821 3822 arc_adapt(size, state); 3823 3824 /* 3825 * If arc_size is currently overflowing, and has grown past our 3826 * upper limit, we must be adding data faster than the evict 3827 * thread can evict. Thus, to ensure we don't compound the 3828 * problem by adding more data and forcing arc_size to grow even 3829 * further past it's target size, we halt and wait for the 3830 * eviction thread to catch up. 3831 * 3832 * It's also possible that the reclaim thread is unable to evict 3833 * enough buffers to get arc_size below the overflow limit (e.g. 3834 * due to buffers being un-evictable, or hash lock collisions). 3835 * In this case, we want to proceed regardless if we're 3836 * overflowing; thus we don't use a while loop here. 3837 */ 3838 if (arc_is_overflowing()) { 3839 mutex_enter(&arc_reclaim_lock); 3840 3841 /* 3842 * Now that we've acquired the lock, we may no longer be 3843 * over the overflow limit, lets check. 3844 * 3845 * We're ignoring the case of spurious wake ups. If that 3846 * were to happen, it'd let this thread consume an ARC 3847 * buffer before it should have (i.e. before we're under 3848 * the overflow limit and were signalled by the reclaim 3849 * thread). As long as that is a rare occurrence, it 3850 * shouldn't cause any harm. 3851 */ 3852 if (arc_is_overflowing()) { 3853 cv_signal(&arc_reclaim_thread_cv); 3854 cv_wait(&arc_reclaim_waiters_cv, &arc_reclaim_lock); 3855 } 3856 3857 mutex_exit(&arc_reclaim_lock); 3858 } 3859 3860 if (type == ARC_BUFC_METADATA) { 3861 buf->b_data = zio_buf_alloc(size); 3862 arc_space_consume(size, ARC_SPACE_META); 3863 } else { 3864 ASSERT(type == ARC_BUFC_DATA); 3865 buf->b_data = zio_data_buf_alloc(size); 3866 arc_space_consume(size, ARC_SPACE_DATA); 3867 } 3868 3869 /* 3870 * Update the state size. Note that ghost states have a 3871 * "ghost size" and so don't need to be updated. 3872 */ 3873 if (!GHOST_STATE(buf->b_hdr->b_l1hdr.b_state)) { 3874 arc_buf_hdr_t *hdr = buf->b_hdr; 3875 arc_state_t *state = hdr->b_l1hdr.b_state; 3876 3877 (void) refcount_add_many(&state->arcs_size, size, buf); 3878 3879 /* 3880 * If this is reached via arc_read, the link is 3881 * protected by the hash lock. If reached via 3882 * arc_buf_alloc, the header should not be accessed by 3883 * any other thread. And, if reached via arc_read_done, 3884 * the hash lock will protect it if it's found in the 3885 * hash table; otherwise no other thread should be 3886 * trying to [add|remove]_reference it. 3887 */ 3888 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) { 3889 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 3890 atomic_add_64(&hdr->b_l1hdr.b_state->arcs_lsize[type], 3891 size); 3892 } 3893 /* 3894 * If we are growing the cache, and we are adding anonymous 3895 * data, and we have outgrown arc_p, update arc_p 3896 */ 3897 if (arc_size < arc_c && hdr->b_l1hdr.b_state == arc_anon && 3898 (refcount_count(&arc_anon->arcs_size) + 3899 refcount_count(&arc_mru->arcs_size) > arc_p)) 3900 arc_p = MIN(arc_c, arc_p + size); 3901 } 3902 ARCSTAT_BUMP(arcstat_allocated); 3903} 3904 3905/* 3906 * This routine is called whenever a buffer is accessed. 3907 * NOTE: the hash lock is dropped in this function. 3908 */ 3909static void 3910arc_access(arc_buf_hdr_t *hdr, kmutex_t *hash_lock) 3911{ 3912 clock_t now; 3913 3914 ASSERT(MUTEX_HELD(hash_lock)); 3915 ASSERT(HDR_HAS_L1HDR(hdr)); 3916 3917 if (hdr->b_l1hdr.b_state == arc_anon) { 3918 /* 3919 * This buffer is not in the cache, and does not 3920 * appear in our "ghost" list. Add the new buffer 3921 * to the MRU state. 3922 */ 3923 3924 ASSERT0(hdr->b_l1hdr.b_arc_access); 3925 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 3926 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr); 3927 arc_change_state(arc_mru, hdr, hash_lock); 3928 3929 } else if (hdr->b_l1hdr.b_state == arc_mru) { 3930 now = ddi_get_lbolt(); 3931 3932 /* 3933 * If this buffer is here because of a prefetch, then either: 3934 * - clear the flag if this is a "referencing" read 3935 * (any subsequent access will bump this into the MFU state). 3936 * or 3937 * - move the buffer to the head of the list if this is 3938 * another prefetch (to make it less likely to be evicted). 3939 */ 3940 if (HDR_PREFETCH(hdr)) { 3941 if (refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) { 3942 /* link protected by hash lock */ 3943 ASSERT(multilist_link_active( 3944 &hdr->b_l1hdr.b_arc_node)); 3945 } else { 3946 hdr->b_flags &= ~ARC_FLAG_PREFETCH; 3947 ARCSTAT_BUMP(arcstat_mru_hits); 3948 } 3949 hdr->b_l1hdr.b_arc_access = now; 3950 return; 3951 } 3952 3953 /* 3954 * This buffer has been "accessed" only once so far, 3955 * but it is still in the cache. Move it to the MFU 3956 * state. 3957 */ 3958 if (now > hdr->b_l1hdr.b_arc_access + ARC_MINTIME) { 3959 /* 3960 * More than 125ms have passed since we 3961 * instantiated this buffer. Move it to the 3962 * most frequently used state. 3963 */ 3964 hdr->b_l1hdr.b_arc_access = now; 3965 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 3966 arc_change_state(arc_mfu, hdr, hash_lock); 3967 } 3968 ARCSTAT_BUMP(arcstat_mru_hits); 3969 } else if (hdr->b_l1hdr.b_state == arc_mru_ghost) { 3970 arc_state_t *new_state; 3971 /* 3972 * This buffer has been "accessed" recently, but 3973 * was evicted from the cache. Move it to the 3974 * MFU state. 3975 */ 3976 3977 if (HDR_PREFETCH(hdr)) { 3978 new_state = arc_mru; 3979 if (refcount_count(&hdr->b_l1hdr.b_refcnt) > 0) 3980 hdr->b_flags &= ~ARC_FLAG_PREFETCH; 3981 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr); 3982 } else { 3983 new_state = arc_mfu; 3984 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 3985 } 3986 3987 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 3988 arc_change_state(new_state, hdr, hash_lock); 3989 3990 ARCSTAT_BUMP(arcstat_mru_ghost_hits); 3991 } else if (hdr->b_l1hdr.b_state == arc_mfu) { 3992 /* 3993 * This buffer has been accessed more than once and is 3994 * still in the cache. Keep it in the MFU state. 3995 * 3996 * NOTE: an add_reference() that occurred when we did 3997 * the arc_read() will have kicked this off the list. 3998 * If it was a prefetch, we will explicitly move it to 3999 * the head of the list now. 4000 */ 4001 if ((HDR_PREFETCH(hdr)) != 0) { 4002 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 4003 /* link protected by hash_lock */ 4004 ASSERT(multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 4005 } 4006 ARCSTAT_BUMP(arcstat_mfu_hits); 4007 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 4008 } else if (hdr->b_l1hdr.b_state == arc_mfu_ghost) { 4009 arc_state_t *new_state = arc_mfu; 4010 /* 4011 * This buffer has been accessed more than once but has 4012 * been evicted from the cache. Move it back to the 4013 * MFU state. 4014 */ 4015 4016 if (HDR_PREFETCH(hdr)) { 4017 /* 4018 * This is a prefetch access... 4019 * move this block back to the MRU state. 4020 */ 4021 ASSERT0(refcount_count(&hdr->b_l1hdr.b_refcnt)); 4022 new_state = arc_mru; 4023 } 4024 4025 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 4026 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 4027 arc_change_state(new_state, hdr, hash_lock); 4028 4029 ARCSTAT_BUMP(arcstat_mfu_ghost_hits); 4030 } else if (hdr->b_l1hdr.b_state == arc_l2c_only) { 4031 /* 4032 * This buffer is on the 2nd Level ARC. 4033 */ 4034 4035 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 4036 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 4037 arc_change_state(arc_mfu, hdr, hash_lock); 4038 } else { 4039 ASSERT(!"invalid arc state"); 4040 } 4041} 4042 4043/* a generic arc_done_func_t which you can use */ 4044/* ARGSUSED */ 4045void 4046arc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg) 4047{ 4048 if (zio == NULL || zio->io_error == 0) 4049 bcopy(buf->b_data, arg, buf->b_hdr->b_size); 4050 VERIFY(arc_buf_remove_ref(buf, arg)); 4051} 4052 4053/* a generic arc_done_func_t */ 4054void 4055arc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg) 4056{ 4057 arc_buf_t **bufp = arg; 4058 if (zio && zio->io_error) { 4059 VERIFY(arc_buf_remove_ref(buf, arg)); 4060 *bufp = NULL; 4061 } else { 4062 *bufp = buf; 4063 ASSERT(buf->b_data); 4064 } 4065} 4066 4067static void 4068arc_read_done(zio_t *zio) 4069{ 4070 arc_buf_hdr_t *hdr; 4071 arc_buf_t *buf; 4072 arc_buf_t *abuf; /* buffer we're assigning to callback */ 4073 kmutex_t *hash_lock = NULL; 4074 arc_callback_t *callback_list, *acb; 4075 int freeable = FALSE; 4076 4077 buf = zio->io_private; 4078 hdr = buf->b_hdr; 4079 4080 /* 4081 * The hdr was inserted into hash-table and removed from lists 4082 * prior to starting I/O. We should find this header, since 4083 * it's in the hash table, and it should be legit since it's 4084 * not possible to evict it during the I/O. The only possible 4085 * reason for it not to be found is if we were freed during the 4086 * read. 4087 */ 4088 if (HDR_IN_HASH_TABLE(hdr)) { 4089 ASSERT3U(hdr->b_birth, ==, BP_PHYSICAL_BIRTH(zio->io_bp)); 4090 ASSERT3U(hdr->b_dva.dva_word[0], ==, 4091 BP_IDENTITY(zio->io_bp)->dva_word[0]); 4092 ASSERT3U(hdr->b_dva.dva_word[1], ==, 4093 BP_IDENTITY(zio->io_bp)->dva_word[1]); 4094 4095 arc_buf_hdr_t *found = buf_hash_find(hdr->b_spa, zio->io_bp, 4096 &hash_lock); 4097 4098 ASSERT((found == NULL && HDR_FREED_IN_READ(hdr) && 4099 hash_lock == NULL) || 4100 (found == hdr && 4101 DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) || 4102 (found == hdr && HDR_L2_READING(hdr))); 4103 } 4104 4105 hdr->b_flags &= ~ARC_FLAG_L2_EVICTED; 4106 if (l2arc_noprefetch && HDR_PREFETCH(hdr)) 4107 hdr->b_flags &= ~ARC_FLAG_L2CACHE; 4108 4109 /* byteswap if necessary */ 4110 callback_list = hdr->b_l1hdr.b_acb; 4111 ASSERT(callback_list != NULL); 4112 if (BP_SHOULD_BYTESWAP(zio->io_bp) && zio->io_error == 0) { 4113 dmu_object_byteswap_t bswap = 4114 DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp)); 4115 arc_byteswap_func_t *func = BP_GET_LEVEL(zio->io_bp) > 0 ? 4116 byteswap_uint64_array : 4117 dmu_ot_byteswap[bswap].ob_func; 4118 func(buf->b_data, hdr->b_size); 4119 } 4120 4121 arc_cksum_compute(buf, B_FALSE); 4122#ifdef illumos 4123 arc_buf_watch(buf); 4124#endif /* illumos */ 4125 4126 if (hash_lock && zio->io_error == 0 && 4127 hdr->b_l1hdr.b_state == arc_anon) { 4128 /* 4129 * Only call arc_access on anonymous buffers. This is because 4130 * if we've issued an I/O for an evicted buffer, we've already 4131 * called arc_access (to prevent any simultaneous readers from 4132 * getting confused). 4133 */ 4134 arc_access(hdr, hash_lock); 4135 } 4136 4137 /* create copies of the data buffer for the callers */ 4138 abuf = buf; 4139 for (acb = callback_list; acb; acb = acb->acb_next) { 4140 if (acb->acb_done) { 4141 if (abuf == NULL) { 4142 ARCSTAT_BUMP(arcstat_duplicate_reads); 4143 abuf = arc_buf_clone(buf); 4144 } 4145 acb->acb_buf = abuf; 4146 abuf = NULL; 4147 } 4148 } 4149 hdr->b_l1hdr.b_acb = NULL; 4150 hdr->b_flags &= ~ARC_FLAG_IO_IN_PROGRESS; 4151 ASSERT(!HDR_BUF_AVAILABLE(hdr)); 4152 if (abuf == buf) { 4153 ASSERT(buf->b_efunc == NULL); 4154 ASSERT(hdr->b_l1hdr.b_datacnt == 1); 4155 hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE; 4156 } 4157 4158 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt) || 4159 callback_list != NULL); 4160 4161 if (zio->io_error != 0) { 4162 hdr->b_flags |= ARC_FLAG_IO_ERROR; 4163 if (hdr->b_l1hdr.b_state != arc_anon) 4164 arc_change_state(arc_anon, hdr, hash_lock); 4165 if (HDR_IN_HASH_TABLE(hdr)) 4166 buf_hash_remove(hdr); 4167 freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt); 4168 } 4169 4170 /* 4171 * Broadcast before we drop the hash_lock to avoid the possibility 4172 * that the hdr (and hence the cv) might be freed before we get to 4173 * the cv_broadcast(). 4174 */ 4175 cv_broadcast(&hdr->b_l1hdr.b_cv); 4176 4177 if (hash_lock != NULL) { 4178 mutex_exit(hash_lock); 4179 } else { 4180 /* 4181 * This block was freed while we waited for the read to 4182 * complete. It has been removed from the hash table and 4183 * moved to the anonymous state (so that it won't show up 4184 * in the cache). 4185 */ 4186 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon); 4187 freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt); 4188 } 4189 4190 /* execute each callback and free its structure */ 4191 while ((acb = callback_list) != NULL) { 4192 if (acb->acb_done) 4193 acb->acb_done(zio, acb->acb_buf, acb->acb_private); 4194 4195 if (acb->acb_zio_dummy != NULL) { 4196 acb->acb_zio_dummy->io_error = zio->io_error; 4197 zio_nowait(acb->acb_zio_dummy); 4198 } 4199 4200 callback_list = acb->acb_next; 4201 kmem_free(acb, sizeof (arc_callback_t)); 4202 } 4203 4204 if (freeable) 4205 arc_hdr_destroy(hdr); 4206} 4207 4208/* 4209 * "Read" the block at the specified DVA (in bp) via the 4210 * cache. If the block is found in the cache, invoke the provided 4211 * callback immediately and return. Note that the `zio' parameter 4212 * in the callback will be NULL in this case, since no IO was 4213 * required. If the block is not in the cache pass the read request 4214 * on to the spa with a substitute callback function, so that the 4215 * requested block will be added to the cache. 4216 * 4217 * If a read request arrives for a block that has a read in-progress, 4218 * either wait for the in-progress read to complete (and return the 4219 * results); or, if this is a read with a "done" func, add a record 4220 * to the read to invoke the "done" func when the read completes, 4221 * and return; or just return. 4222 * 4223 * arc_read_done() will invoke all the requested "done" functions 4224 * for readers of this block. 4225 */ 4226int 4227arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_done_func_t *done, 4228 void *private, zio_priority_t priority, int zio_flags, 4229 arc_flags_t *arc_flags, const zbookmark_phys_t *zb) 4230{ 4231 arc_buf_hdr_t *hdr = NULL; 4232 arc_buf_t *buf = NULL; 4233 kmutex_t *hash_lock = NULL; 4234 zio_t *rzio; 4235 uint64_t guid = spa_load_guid(spa); 4236 4237 ASSERT(!BP_IS_EMBEDDED(bp) || 4238 BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA); 4239 4240top: 4241 if (!BP_IS_EMBEDDED(bp)) { 4242 /* 4243 * Embedded BP's have no DVA and require no I/O to "read". 4244 * Create an anonymous arc buf to back it. 4245 */ 4246 hdr = buf_hash_find(guid, bp, &hash_lock); 4247 } 4248 4249 if (hdr != NULL && HDR_HAS_L1HDR(hdr) && hdr->b_l1hdr.b_datacnt > 0) { 4250 4251 *arc_flags |= ARC_FLAG_CACHED; 4252 4253 if (HDR_IO_IN_PROGRESS(hdr)) { 4254 4255 if (*arc_flags & ARC_FLAG_WAIT) { 4256 cv_wait(&hdr->b_l1hdr.b_cv, hash_lock); 4257 mutex_exit(hash_lock); 4258 goto top; 4259 } 4260 ASSERT(*arc_flags & ARC_FLAG_NOWAIT); 4261 4262 if (done) { 4263 arc_callback_t *acb = NULL; 4264 4265 acb = kmem_zalloc(sizeof (arc_callback_t), 4266 KM_SLEEP); 4267 acb->acb_done = done; 4268 acb->acb_private = private; 4269 if (pio != NULL) 4270 acb->acb_zio_dummy = zio_null(pio, 4271 spa, NULL, NULL, NULL, zio_flags); 4272 4273 ASSERT(acb->acb_done != NULL); 4274 acb->acb_next = hdr->b_l1hdr.b_acb; 4275 hdr->b_l1hdr.b_acb = acb; 4276 add_reference(hdr, hash_lock, private); 4277 mutex_exit(hash_lock); 4278 return (0); 4279 } 4280 mutex_exit(hash_lock); 4281 return (0); 4282 } 4283 4284 ASSERT(hdr->b_l1hdr.b_state == arc_mru || 4285 hdr->b_l1hdr.b_state == arc_mfu); 4286 4287 if (done) { 4288 add_reference(hdr, hash_lock, private); 4289 /* 4290 * If this block is already in use, create a new 4291 * copy of the data so that we will be guaranteed 4292 * that arc_release() will always succeed. 4293 */ 4294 buf = hdr->b_l1hdr.b_buf; 4295 ASSERT(buf); 4296 ASSERT(buf->b_data); 4297 if (HDR_BUF_AVAILABLE(hdr)) { 4298 ASSERT(buf->b_efunc == NULL); 4299 hdr->b_flags &= ~ARC_FLAG_BUF_AVAILABLE; 4300 } else { 4301 buf = arc_buf_clone(buf); 4302 } 4303 4304 } else if (*arc_flags & ARC_FLAG_PREFETCH && 4305 refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) { 4306 hdr->b_flags |= ARC_FLAG_PREFETCH; 4307 } 4308 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr); 4309 arc_access(hdr, hash_lock); 4310 if (*arc_flags & ARC_FLAG_L2CACHE) 4311 hdr->b_flags |= ARC_FLAG_L2CACHE; 4312 if (*arc_flags & ARC_FLAG_L2COMPRESS) 4313 hdr->b_flags |= ARC_FLAG_L2COMPRESS; 4314 mutex_exit(hash_lock); 4315 ARCSTAT_BUMP(arcstat_hits); 4316 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr), 4317 demand, prefetch, !HDR_ISTYPE_METADATA(hdr), 4318 data, metadata, hits); 4319 4320 if (done) 4321 done(NULL, buf, private); 4322 } else { 4323 uint64_t size = BP_GET_LSIZE(bp); 4324 arc_callback_t *acb; 4325 vdev_t *vd = NULL; 4326 uint64_t addr = 0; 4327 boolean_t devw = B_FALSE; 4328 enum zio_compress b_compress = ZIO_COMPRESS_OFF; 4329 int32_t b_asize = 0; 4330 4331 if (hdr == NULL) { 4332 /* this block is not in the cache */ 4333 arc_buf_hdr_t *exists = NULL; 4334 arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp); 4335 buf = arc_buf_alloc(spa, size, private, type); 4336 hdr = buf->b_hdr; 4337 if (!BP_IS_EMBEDDED(bp)) { 4338 hdr->b_dva = *BP_IDENTITY(bp); 4339 hdr->b_birth = BP_PHYSICAL_BIRTH(bp); 4340 exists = buf_hash_insert(hdr, &hash_lock); 4341 } 4342 if (exists != NULL) { 4343 /* somebody beat us to the hash insert */ 4344 mutex_exit(hash_lock); 4345 buf_discard_identity(hdr); 4346 (void) arc_buf_remove_ref(buf, private); 4347 goto top; /* restart the IO request */ 4348 } 4349 4350 /* if this is a prefetch, we don't have a reference */ 4351 if (*arc_flags & ARC_FLAG_PREFETCH) { 4352 (void) remove_reference(hdr, hash_lock, 4353 private); 4354 hdr->b_flags |= ARC_FLAG_PREFETCH; 4355 } 4356 if (*arc_flags & ARC_FLAG_L2CACHE) 4357 hdr->b_flags |= ARC_FLAG_L2CACHE; 4358 if (*arc_flags & ARC_FLAG_L2COMPRESS) 4359 hdr->b_flags |= ARC_FLAG_L2COMPRESS; 4360 if (BP_GET_LEVEL(bp) > 0) 4361 hdr->b_flags |= ARC_FLAG_INDIRECT; 4362 } else { 4363 /* 4364 * This block is in the ghost cache. If it was L2-only 4365 * (and thus didn't have an L1 hdr), we realloc the 4366 * header to add an L1 hdr. 4367 */ 4368 if (!HDR_HAS_L1HDR(hdr)) { 4369 hdr = arc_hdr_realloc(hdr, hdr_l2only_cache, 4370 hdr_full_cache); 4371 } 4372 4373 ASSERT(GHOST_STATE(hdr->b_l1hdr.b_state)); 4374 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 4375 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 4376 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 4377 4378 /* if this is a prefetch, we don't have a reference */ 4379 if (*arc_flags & ARC_FLAG_PREFETCH) 4380 hdr->b_flags |= ARC_FLAG_PREFETCH; 4381 else 4382 add_reference(hdr, hash_lock, private); 4383 if (*arc_flags & ARC_FLAG_L2CACHE) 4384 hdr->b_flags |= ARC_FLAG_L2CACHE; 4385 if (*arc_flags & ARC_FLAG_L2COMPRESS) 4386 hdr->b_flags |= ARC_FLAG_L2COMPRESS; 4387 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 4388 buf->b_hdr = hdr; 4389 buf->b_data = NULL; 4390 buf->b_efunc = NULL; 4391 buf->b_private = NULL; 4392 buf->b_next = NULL; 4393 hdr->b_l1hdr.b_buf = buf; 4394 ASSERT0(hdr->b_l1hdr.b_datacnt); 4395 hdr->b_l1hdr.b_datacnt = 1; 4396 arc_get_data_buf(buf); 4397 arc_access(hdr, hash_lock); 4398 } 4399 4400 ASSERT(!GHOST_STATE(hdr->b_l1hdr.b_state)); 4401 4402 acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP); 4403 acb->acb_done = done; 4404 acb->acb_private = private; 4405 4406 ASSERT(hdr->b_l1hdr.b_acb == NULL); 4407 hdr->b_l1hdr.b_acb = acb; 4408 hdr->b_flags |= ARC_FLAG_IO_IN_PROGRESS; 4409 4410 if (HDR_HAS_L2HDR(hdr) && 4411 (vd = hdr->b_l2hdr.b_dev->l2ad_vdev) != NULL) { 4412 devw = hdr->b_l2hdr.b_dev->l2ad_writing; 4413 addr = hdr->b_l2hdr.b_daddr; 4414 b_compress = HDR_GET_COMPRESS(hdr); 4415 b_asize = hdr->b_l2hdr.b_asize; 4416 /* 4417 * Lock out device removal. 4418 */ 4419 if (vdev_is_dead(vd) || 4420 !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER)) 4421 vd = NULL; 4422 } 4423 4424 if (hash_lock != NULL) 4425 mutex_exit(hash_lock); 4426 4427 /* 4428 * At this point, we have a level 1 cache miss. Try again in 4429 * L2ARC if possible. 4430 */ 4431 ASSERT3U(hdr->b_size, ==, size); 4432 DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp, 4433 uint64_t, size, zbookmark_phys_t *, zb); 4434 ARCSTAT_BUMP(arcstat_misses); 4435 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr), 4436 demand, prefetch, !HDR_ISTYPE_METADATA(hdr), 4437 data, metadata, misses); 4438#ifdef _KERNEL 4439 curthread->td_ru.ru_inblock++; 4440#endif 4441 4442 if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) { 4443 /* 4444 * Read from the L2ARC if the following are true: 4445 * 1. The L2ARC vdev was previously cached. 4446 * 2. This buffer still has L2ARC metadata. 4447 * 3. This buffer isn't currently writing to the L2ARC. 4448 * 4. The L2ARC entry wasn't evicted, which may 4449 * also have invalidated the vdev. 4450 * 5. This isn't prefetch and l2arc_noprefetch is set. 4451 */ 4452 if (HDR_HAS_L2HDR(hdr) && 4453 !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) && 4454 !(l2arc_noprefetch && HDR_PREFETCH(hdr))) { 4455 l2arc_read_callback_t *cb; 4456 4457 DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr); 4458 ARCSTAT_BUMP(arcstat_l2_hits); 4459 4460 cb = kmem_zalloc(sizeof (l2arc_read_callback_t), 4461 KM_SLEEP); 4462 cb->l2rcb_buf = buf; 4463 cb->l2rcb_spa = spa; 4464 cb->l2rcb_bp = *bp; 4465 cb->l2rcb_zb = *zb; 4466 cb->l2rcb_flags = zio_flags; 4467 cb->l2rcb_compress = b_compress; 4468 4469 ASSERT(addr >= VDEV_LABEL_START_SIZE && 4470 addr + size < vd->vdev_psize - 4471 VDEV_LABEL_END_SIZE); 4472 4473 /* 4474 * l2arc read. The SCL_L2ARC lock will be 4475 * released by l2arc_read_done(). 4476 * Issue a null zio if the underlying buffer 4477 * was squashed to zero size by compression. 4478 */ 4479 if (b_compress == ZIO_COMPRESS_EMPTY) { 4480 rzio = zio_null(pio, spa, vd, 4481 l2arc_read_done, cb, 4482 zio_flags | ZIO_FLAG_DONT_CACHE | 4483 ZIO_FLAG_CANFAIL | 4484 ZIO_FLAG_DONT_PROPAGATE | 4485 ZIO_FLAG_DONT_RETRY); 4486 } else { 4487 rzio = zio_read_phys(pio, vd, addr, 4488 b_asize, buf->b_data, 4489 ZIO_CHECKSUM_OFF, 4490 l2arc_read_done, cb, priority, 4491 zio_flags | ZIO_FLAG_DONT_CACHE | 4492 ZIO_FLAG_CANFAIL | 4493 ZIO_FLAG_DONT_PROPAGATE | 4494 ZIO_FLAG_DONT_RETRY, B_FALSE); 4495 } 4496 DTRACE_PROBE2(l2arc__read, vdev_t *, vd, 4497 zio_t *, rzio); 4498 ARCSTAT_INCR(arcstat_l2_read_bytes, b_asize); 4499 4500 if (*arc_flags & ARC_FLAG_NOWAIT) { 4501 zio_nowait(rzio); 4502 return (0); 4503 } 4504 4505 ASSERT(*arc_flags & ARC_FLAG_WAIT); 4506 if (zio_wait(rzio) == 0) 4507 return (0); 4508 4509 /* l2arc read error; goto zio_read() */ 4510 } else { 4511 DTRACE_PROBE1(l2arc__miss, 4512 arc_buf_hdr_t *, hdr); 4513 ARCSTAT_BUMP(arcstat_l2_misses); 4514 if (HDR_L2_WRITING(hdr)) 4515 ARCSTAT_BUMP(arcstat_l2_rw_clash); 4516 spa_config_exit(spa, SCL_L2ARC, vd); 4517 } 4518 } else { 4519 if (vd != NULL) 4520 spa_config_exit(spa, SCL_L2ARC, vd); 4521 if (l2arc_ndev != 0) { 4522 DTRACE_PROBE1(l2arc__miss, 4523 arc_buf_hdr_t *, hdr); 4524 ARCSTAT_BUMP(arcstat_l2_misses); 4525 } 4526 } 4527 4528 rzio = zio_read(pio, spa, bp, buf->b_data, size, 4529 arc_read_done, buf, priority, zio_flags, zb); 4530 4531 if (*arc_flags & ARC_FLAG_WAIT) 4532 return (zio_wait(rzio)); 4533 4534 ASSERT(*arc_flags & ARC_FLAG_NOWAIT); 4535 zio_nowait(rzio); 4536 } 4537 return (0); 4538} 4539 4540void 4541arc_set_callback(arc_buf_t *buf, arc_evict_func_t *func, void *private) 4542{ 4543 ASSERT(buf->b_hdr != NULL); 4544 ASSERT(buf->b_hdr->b_l1hdr.b_state != arc_anon); 4545 ASSERT(!refcount_is_zero(&buf->b_hdr->b_l1hdr.b_refcnt) || 4546 func == NULL); 4547 ASSERT(buf->b_efunc == NULL); 4548 ASSERT(!HDR_BUF_AVAILABLE(buf->b_hdr)); 4549 4550 buf->b_efunc = func; 4551 buf->b_private = private; 4552} 4553 4554/* 4555 * Notify the arc that a block was freed, and thus will never be used again. 4556 */ 4557void 4558arc_freed(spa_t *spa, const blkptr_t *bp) 4559{ 4560 arc_buf_hdr_t *hdr; 4561 kmutex_t *hash_lock; 4562 uint64_t guid = spa_load_guid(spa); 4563 4564 ASSERT(!BP_IS_EMBEDDED(bp)); 4565 4566 hdr = buf_hash_find(guid, bp, &hash_lock); 4567 if (hdr == NULL) 4568 return; 4569 if (HDR_BUF_AVAILABLE(hdr)) { 4570 arc_buf_t *buf = hdr->b_l1hdr.b_buf; 4571 add_reference(hdr, hash_lock, FTAG); 4572 hdr->b_flags &= ~ARC_FLAG_BUF_AVAILABLE; 4573 mutex_exit(hash_lock); 4574 4575 arc_release(buf, FTAG); 4576 (void) arc_buf_remove_ref(buf, FTAG); 4577 } else { 4578 mutex_exit(hash_lock); 4579 } 4580 4581} 4582 4583/* 4584 * Clear the user eviction callback set by arc_set_callback(), first calling 4585 * it if it exists. Because the presence of a callback keeps an arc_buf cached 4586 * clearing the callback may result in the arc_buf being destroyed. However, 4587 * it will not result in the *last* arc_buf being destroyed, hence the data 4588 * will remain cached in the ARC. We make a copy of the arc buffer here so 4589 * that we can process the callback without holding any locks. 4590 * 4591 * It's possible that the callback is already in the process of being cleared 4592 * by another thread. In this case we can not clear the callback. 4593 * 4594 * Returns B_TRUE if the callback was successfully called and cleared. 4595 */ 4596boolean_t 4597arc_clear_callback(arc_buf_t *buf) 4598{ 4599 arc_buf_hdr_t *hdr; 4600 kmutex_t *hash_lock; 4601 arc_evict_func_t *efunc = buf->b_efunc; 4602 void *private = buf->b_private; 4603 4604 mutex_enter(&buf->b_evict_lock); 4605 hdr = buf->b_hdr; 4606 if (hdr == NULL) { 4607 /* 4608 * We are in arc_do_user_evicts(). 4609 */ 4610 ASSERT(buf->b_data == NULL); 4611 mutex_exit(&buf->b_evict_lock); 4612 return (B_FALSE); 4613 } else if (buf->b_data == NULL) { 4614 /* 4615 * We are on the eviction list; process this buffer now 4616 * but let arc_do_user_evicts() do the reaping. 4617 */ 4618 buf->b_efunc = NULL; 4619 mutex_exit(&buf->b_evict_lock); 4620 VERIFY0(efunc(private)); 4621 return (B_TRUE); 4622 } 4623 hash_lock = HDR_LOCK(hdr); 4624 mutex_enter(hash_lock); 4625 hdr = buf->b_hdr; 4626 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 4627 4628 ASSERT3U(refcount_count(&hdr->b_l1hdr.b_refcnt), <, 4629 hdr->b_l1hdr.b_datacnt); 4630 ASSERT(hdr->b_l1hdr.b_state == arc_mru || 4631 hdr->b_l1hdr.b_state == arc_mfu); 4632 4633 buf->b_efunc = NULL; 4634 buf->b_private = NULL; 4635 4636 if (hdr->b_l1hdr.b_datacnt > 1) { 4637 mutex_exit(&buf->b_evict_lock); 4638 arc_buf_destroy(buf, TRUE); 4639 } else { 4640 ASSERT(buf == hdr->b_l1hdr.b_buf); 4641 hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE; 4642 mutex_exit(&buf->b_evict_lock); 4643 } 4644 4645 mutex_exit(hash_lock); 4646 VERIFY0(efunc(private)); 4647 return (B_TRUE); 4648} 4649 4650/* 4651 * Release this buffer from the cache, making it an anonymous buffer. This 4652 * must be done after a read and prior to modifying the buffer contents. 4653 * If the buffer has more than one reference, we must make 4654 * a new hdr for the buffer. 4655 */ 4656void 4657arc_release(arc_buf_t *buf, void *tag) 4658{ 4659 arc_buf_hdr_t *hdr = buf->b_hdr; 4660 4661 ASSERT(HDR_HAS_L1HDR(hdr)); 4662 4663 /* 4664 * It would be nice to assert that if it's DMU metadata (level > 4665 * 0 || it's the dnode file), then it must be syncing context. 4666 * But we don't know that information at this level. 4667 */ 4668 4669 mutex_enter(&buf->b_evict_lock); 4670 /* 4671 * We don't grab the hash lock prior to this check, because if 4672 * the buffer's header is in the arc_anon state, it won't be 4673 * linked into the hash table. 4674 */ 4675 if (hdr->b_l1hdr.b_state == arc_anon) { 4676 mutex_exit(&buf->b_evict_lock); 4677 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 4678 ASSERT(!HDR_IN_HASH_TABLE(hdr)); 4679 ASSERT(!HDR_HAS_L2HDR(hdr)); 4680 ASSERT(BUF_EMPTY(hdr)); 4681 ASSERT3U(hdr->b_l1hdr.b_datacnt, ==, 1); 4682 ASSERT3S(refcount_count(&hdr->b_l1hdr.b_refcnt), ==, 1); 4683 ASSERT(!list_link_active(&hdr->b_l1hdr.b_arc_node)); 4684 4685 ASSERT3P(buf->b_efunc, ==, NULL); 4686 ASSERT3P(buf->b_private, ==, NULL); 4687 4688 hdr->b_l1hdr.b_arc_access = 0; 4689 arc_buf_thaw(buf); 4690 4691 return; 4692 } 4693 4694 kmutex_t *hash_lock = HDR_LOCK(hdr); 4695 mutex_enter(hash_lock); 4696 4697 /* 4698 * This assignment is only valid as long as the hash_lock is 4699 * held, we must be careful not to reference state or the 4700 * b_state field after dropping the lock. 4701 */ 4702 arc_state_t *state = hdr->b_l1hdr.b_state; 4703 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 4704 ASSERT3P(state, !=, arc_anon); 4705 4706 /* this buffer is not on any list */ 4707 ASSERT(refcount_count(&hdr->b_l1hdr.b_refcnt) > 0); 4708 4709 if (HDR_HAS_L2HDR(hdr)) { 4710 mutex_enter(&hdr->b_l2hdr.b_dev->l2ad_mtx); 4711 4712 /* 4713 * We have to recheck this conditional again now that 4714 * we're holding the l2ad_mtx to prevent a race with 4715 * another thread which might be concurrently calling 4716 * l2arc_evict(). In that case, l2arc_evict() might have 4717 * destroyed the header's L2 portion as we were waiting 4718 * to acquire the l2ad_mtx. 4719 */ 4720 if (HDR_HAS_L2HDR(hdr)) { 4721 if (hdr->b_l2hdr.b_daddr != L2ARC_ADDR_UNSET) 4722 trim_map_free(hdr->b_l2hdr.b_dev->l2ad_vdev, 4723 hdr->b_l2hdr.b_daddr, 4724 hdr->b_l2hdr.b_asize, 0); 4725 arc_hdr_l2hdr_destroy(hdr); 4726 } 4727 4728 mutex_exit(&hdr->b_l2hdr.b_dev->l2ad_mtx); 4729 } 4730 4731 /* 4732 * Do we have more than one buf? 4733 */ 4734 if (hdr->b_l1hdr.b_datacnt > 1) { 4735 arc_buf_hdr_t *nhdr; 4736 arc_buf_t **bufp; 4737 uint64_t blksz = hdr->b_size; 4738 uint64_t spa = hdr->b_spa; 4739 arc_buf_contents_t type = arc_buf_type(hdr); 4740 uint32_t flags = hdr->b_flags; 4741 4742 ASSERT(hdr->b_l1hdr.b_buf != buf || buf->b_next != NULL); 4743 /* 4744 * Pull the data off of this hdr and attach it to 4745 * a new anonymous hdr. 4746 */ 4747 (void) remove_reference(hdr, hash_lock, tag); 4748 bufp = &hdr->b_l1hdr.b_buf; 4749 while (*bufp != buf) 4750 bufp = &(*bufp)->b_next; 4751 *bufp = buf->b_next; 4752 buf->b_next = NULL; 4753 4754 ASSERT3P(state, !=, arc_l2c_only); 4755 4756 (void) refcount_remove_many( 4757 &state->arcs_size, hdr->b_size, buf); 4758 4759 if (refcount_is_zero(&hdr->b_l1hdr.b_refcnt)) { 4760 ASSERT3P(state, !=, arc_l2c_only); 4761 uint64_t *size = &state->arcs_lsize[type]; 4762 ASSERT3U(*size, >=, hdr->b_size); 4763 atomic_add_64(size, -hdr->b_size); 4764 } 4765 4766 /* 4767 * We're releasing a duplicate user data buffer, update 4768 * our statistics accordingly. 4769 */ 4770 if (HDR_ISTYPE_DATA(hdr)) { 4771 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers); 4772 ARCSTAT_INCR(arcstat_duplicate_buffers_size, 4773 -hdr->b_size); 4774 } 4775 hdr->b_l1hdr.b_datacnt -= 1; 4776 arc_cksum_verify(buf); 4777#ifdef illumos 4778 arc_buf_unwatch(buf); 4779#endif /* illumos */ 4780 4781 mutex_exit(hash_lock); 4782 4783 nhdr = kmem_cache_alloc(hdr_full_cache, KM_PUSHPAGE); 4784 nhdr->b_size = blksz; 4785 nhdr->b_spa = spa; 4786 4787 nhdr->b_flags = flags & ARC_FLAG_L2_WRITING; 4788 nhdr->b_flags |= arc_bufc_to_flags(type); 4789 nhdr->b_flags |= ARC_FLAG_HAS_L1HDR; 4790 4791 nhdr->b_l1hdr.b_buf = buf; 4792 nhdr->b_l1hdr.b_datacnt = 1; 4793 nhdr->b_l1hdr.b_state = arc_anon; 4794 nhdr->b_l1hdr.b_arc_access = 0; 4795 nhdr->b_l1hdr.b_tmp_cdata = NULL; 4796 nhdr->b_freeze_cksum = NULL; 4797 4798 (void) refcount_add(&nhdr->b_l1hdr.b_refcnt, tag); 4799 buf->b_hdr = nhdr; 4800 mutex_exit(&buf->b_evict_lock); 4801 (void) refcount_add_many(&arc_anon->arcs_size, blksz, buf); 4802 } else { 4803 mutex_exit(&buf->b_evict_lock); 4804 ASSERT(refcount_count(&hdr->b_l1hdr.b_refcnt) == 1); 4805 /* protected by hash lock, or hdr is on arc_anon */ 4806 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 4807 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 4808 arc_change_state(arc_anon, hdr, hash_lock); 4809 hdr->b_l1hdr.b_arc_access = 0; 4810 mutex_exit(hash_lock); 4811 4812 buf_discard_identity(hdr); 4813 arc_buf_thaw(buf); 4814 } 4815 buf->b_efunc = NULL; 4816 buf->b_private = NULL; 4817} 4818 4819int 4820arc_released(arc_buf_t *buf) 4821{ 4822 int released; 4823 4824 mutex_enter(&buf->b_evict_lock); 4825 released = (buf->b_data != NULL && 4826 buf->b_hdr->b_l1hdr.b_state == arc_anon); 4827 mutex_exit(&buf->b_evict_lock); 4828 return (released); 4829} 4830 4831#ifdef ZFS_DEBUG 4832int 4833arc_referenced(arc_buf_t *buf) 4834{ 4835 int referenced; 4836 4837 mutex_enter(&buf->b_evict_lock); 4838 referenced = (refcount_count(&buf->b_hdr->b_l1hdr.b_refcnt)); 4839 mutex_exit(&buf->b_evict_lock); 4840 return (referenced); 4841} 4842#endif 4843 4844static void 4845arc_write_ready(zio_t *zio) 4846{ 4847 arc_write_callback_t *callback = zio->io_private; 4848 arc_buf_t *buf = callback->awcb_buf; 4849 arc_buf_hdr_t *hdr = buf->b_hdr; 4850 4851 ASSERT(HDR_HAS_L1HDR(hdr)); 4852 ASSERT(!refcount_is_zero(&buf->b_hdr->b_l1hdr.b_refcnt)); 4853 ASSERT(hdr->b_l1hdr.b_datacnt > 0); 4854 callback->awcb_ready(zio, buf, callback->awcb_private); 4855 4856 /* 4857 * If the IO is already in progress, then this is a re-write 4858 * attempt, so we need to thaw and re-compute the cksum. 4859 * It is the responsibility of the callback to handle the 4860 * accounting for any re-write attempt. 4861 */ 4862 if (HDR_IO_IN_PROGRESS(hdr)) { 4863 mutex_enter(&hdr->b_l1hdr.b_freeze_lock); 4864 if (hdr->b_freeze_cksum != NULL) { 4865 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 4866 hdr->b_freeze_cksum = NULL; 4867 } 4868 mutex_exit(&hdr->b_l1hdr.b_freeze_lock); 4869 } 4870 arc_cksum_compute(buf, B_FALSE); 4871 hdr->b_flags |= ARC_FLAG_IO_IN_PROGRESS; 4872} 4873 4874/* 4875 * The SPA calls this callback for each physical write that happens on behalf 4876 * of a logical write. See the comment in dbuf_write_physdone() for details. 4877 */ 4878static void 4879arc_write_physdone(zio_t *zio) 4880{ 4881 arc_write_callback_t *cb = zio->io_private; 4882 if (cb->awcb_physdone != NULL) 4883 cb->awcb_physdone(zio, cb->awcb_buf, cb->awcb_private); 4884} 4885 4886static void 4887arc_write_done(zio_t *zio) 4888{ 4889 arc_write_callback_t *callback = zio->io_private; 4890 arc_buf_t *buf = callback->awcb_buf; 4891 arc_buf_hdr_t *hdr = buf->b_hdr; 4892 4893 ASSERT(hdr->b_l1hdr.b_acb == NULL); 4894 4895 if (zio->io_error == 0) { 4896 if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) { 4897 buf_discard_identity(hdr); 4898 } else { 4899 hdr->b_dva = *BP_IDENTITY(zio->io_bp); 4900 hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp); 4901 } 4902 } else { 4903 ASSERT(BUF_EMPTY(hdr)); 4904 } 4905 4906 /* 4907 * If the block to be written was all-zero or compressed enough to be 4908 * embedded in the BP, no write was performed so there will be no 4909 * dva/birth/checksum. The buffer must therefore remain anonymous 4910 * (and uncached). 4911 */ 4912 if (!BUF_EMPTY(hdr)) { 4913 arc_buf_hdr_t *exists; 4914 kmutex_t *hash_lock; 4915 4916 ASSERT(zio->io_error == 0); 4917 4918 arc_cksum_verify(buf); 4919 4920 exists = buf_hash_insert(hdr, &hash_lock); 4921 if (exists != NULL) { 4922 /* 4923 * This can only happen if we overwrite for 4924 * sync-to-convergence, because we remove 4925 * buffers from the hash table when we arc_free(). 4926 */ 4927 if (zio->io_flags & ZIO_FLAG_IO_REWRITE) { 4928 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) 4929 panic("bad overwrite, hdr=%p exists=%p", 4930 (void *)hdr, (void *)exists); 4931 ASSERT(refcount_is_zero( 4932 &exists->b_l1hdr.b_refcnt)); 4933 arc_change_state(arc_anon, exists, hash_lock); 4934 mutex_exit(hash_lock); 4935 arc_hdr_destroy(exists); 4936 exists = buf_hash_insert(hdr, &hash_lock); 4937 ASSERT3P(exists, ==, NULL); 4938 } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) { 4939 /* nopwrite */ 4940 ASSERT(zio->io_prop.zp_nopwrite); 4941 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) 4942 panic("bad nopwrite, hdr=%p exists=%p", 4943 (void *)hdr, (void *)exists); 4944 } else { 4945 /* Dedup */ 4946 ASSERT(hdr->b_l1hdr.b_datacnt == 1); 4947 ASSERT(hdr->b_l1hdr.b_state == arc_anon); 4948 ASSERT(BP_GET_DEDUP(zio->io_bp)); 4949 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0); 4950 } 4951 } 4952 hdr->b_flags &= ~ARC_FLAG_IO_IN_PROGRESS; 4953 /* if it's not anon, we are doing a scrub */ 4954 if (exists == NULL && hdr->b_l1hdr.b_state == arc_anon) 4955 arc_access(hdr, hash_lock); 4956 mutex_exit(hash_lock); 4957 } else { 4958 hdr->b_flags &= ~ARC_FLAG_IO_IN_PROGRESS; 4959 } 4960 4961 ASSERT(!refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 4962 callback->awcb_done(zio, buf, callback->awcb_private); 4963 4964 kmem_free(callback, sizeof (arc_write_callback_t)); 4965} 4966 4967zio_t * 4968arc_write(zio_t *pio, spa_t *spa, uint64_t txg, 4969 blkptr_t *bp, arc_buf_t *buf, boolean_t l2arc, boolean_t l2arc_compress, 4970 const zio_prop_t *zp, arc_done_func_t *ready, arc_done_func_t *physdone, 4971 arc_done_func_t *done, void *private, zio_priority_t priority, 4972 int zio_flags, const zbookmark_phys_t *zb) 4973{ 4974 arc_buf_hdr_t *hdr = buf->b_hdr; 4975 arc_write_callback_t *callback; 4976 zio_t *zio; 4977 4978 ASSERT(ready != NULL); 4979 ASSERT(done != NULL); 4980 ASSERT(!HDR_IO_ERROR(hdr)); 4981 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 4982 ASSERT(hdr->b_l1hdr.b_acb == NULL); 4983 ASSERT(hdr->b_l1hdr.b_datacnt > 0); 4984 if (l2arc) 4985 hdr->b_flags |= ARC_FLAG_L2CACHE; 4986 if (l2arc_compress) 4987 hdr->b_flags |= ARC_FLAG_L2COMPRESS; 4988 callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP); 4989 callback->awcb_ready = ready; 4990 callback->awcb_physdone = physdone; 4991 callback->awcb_done = done; 4992 callback->awcb_private = private; 4993 callback->awcb_buf = buf; 4994 4995 zio = zio_write(pio, spa, txg, bp, buf->b_data, hdr->b_size, zp, 4996 arc_write_ready, arc_write_physdone, arc_write_done, callback, 4997 priority, zio_flags, zb); 4998 4999 return (zio); 5000} 5001 5002static int 5003arc_memory_throttle(uint64_t reserve, uint64_t txg) 5004{ 5005#ifdef _KERNEL 5006 uint64_t available_memory = ptob(freemem); 5007 static uint64_t page_load = 0; 5008 static uint64_t last_txg = 0; 5009 5010#if defined(__i386) || !defined(UMA_MD_SMALL_ALLOC) 5011 available_memory = 5012 MIN(available_memory, ptob(vmem_size(heap_arena, VMEM_FREE))); 5013#endif 5014 5015 if (freemem > (uint64_t)physmem * arc_lotsfree_percent / 100) 5016 return (0); 5017 5018 if (txg > last_txg) { 5019 last_txg = txg; 5020 page_load = 0; 5021 } 5022 /* 5023 * If we are in pageout, we know that memory is already tight, 5024 * the arc is already going to be evicting, so we just want to 5025 * continue to let page writes occur as quickly as possible. 5026 */ 5027 if (curproc == pageproc) { 5028 if (page_load > MAX(ptob(minfree), available_memory) / 4) 5029 return (SET_ERROR(ERESTART)); 5030 /* Note: reserve is inflated, so we deflate */ 5031 page_load += reserve / 8; 5032 return (0); 5033 } else if (page_load > 0 && arc_reclaim_needed()) { 5034 /* memory is low, delay before restarting */ 5035 ARCSTAT_INCR(arcstat_memory_throttle_count, 1); 5036 return (SET_ERROR(EAGAIN)); 5037 } 5038 page_load = 0; 5039#endif 5040 return (0); 5041} 5042 5043void 5044arc_tempreserve_clear(uint64_t reserve) 5045{ 5046 atomic_add_64(&arc_tempreserve, -reserve); 5047 ASSERT((int64_t)arc_tempreserve >= 0); 5048} 5049 5050int 5051arc_tempreserve_space(uint64_t reserve, uint64_t txg) 5052{ 5053 int error; 5054 uint64_t anon_size; 5055 5056 if (reserve > arc_c/4 && !arc_no_grow) { 5057 arc_c = MIN(arc_c_max, reserve * 4); 5058 DTRACE_PROBE1(arc__set_reserve, uint64_t, arc_c); 5059 } 5060 if (reserve > arc_c) 5061 return (SET_ERROR(ENOMEM)); 5062 5063 /* 5064 * Don't count loaned bufs as in flight dirty data to prevent long 5065 * network delays from blocking transactions that are ready to be 5066 * assigned to a txg. 5067 */ 5068 anon_size = MAX((int64_t)(refcount_count(&arc_anon->arcs_size) - 5069 arc_loaned_bytes), 0); 5070 5071 /* 5072 * Writes will, almost always, require additional memory allocations 5073 * in order to compress/encrypt/etc the data. We therefore need to 5074 * make sure that there is sufficient available memory for this. 5075 */ 5076 error = arc_memory_throttle(reserve, txg); 5077 if (error != 0) 5078 return (error); 5079 5080 /* 5081 * Throttle writes when the amount of dirty data in the cache 5082 * gets too large. We try to keep the cache less than half full 5083 * of dirty blocks so that our sync times don't grow too large. 5084 * Note: if two requests come in concurrently, we might let them 5085 * both succeed, when one of them should fail. Not a huge deal. 5086 */ 5087 5088 if (reserve + arc_tempreserve + anon_size > arc_c / 2 && 5089 anon_size > arc_c / 4) { 5090 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK " 5091 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n", 5092 arc_tempreserve>>10, 5093 arc_anon->arcs_lsize[ARC_BUFC_METADATA]>>10, 5094 arc_anon->arcs_lsize[ARC_BUFC_DATA]>>10, 5095 reserve>>10, arc_c>>10); 5096 return (SET_ERROR(ERESTART)); 5097 } 5098 atomic_add_64(&arc_tempreserve, reserve); 5099 return (0); 5100} 5101 5102static void 5103arc_kstat_update_state(arc_state_t *state, kstat_named_t *size, 5104 kstat_named_t *evict_data, kstat_named_t *evict_metadata) 5105{ 5106 size->value.ui64 = refcount_count(&state->arcs_size); 5107 evict_data->value.ui64 = state->arcs_lsize[ARC_BUFC_DATA]; 5108 evict_metadata->value.ui64 = state->arcs_lsize[ARC_BUFC_METADATA]; 5109} 5110 5111static int 5112arc_kstat_update(kstat_t *ksp, int rw) 5113{ 5114 arc_stats_t *as = ksp->ks_data; 5115 5116 if (rw == KSTAT_WRITE) { 5117 return (EACCES); 5118 } else { 5119 arc_kstat_update_state(arc_anon, 5120 &as->arcstat_anon_size, 5121 &as->arcstat_anon_evictable_data, 5122 &as->arcstat_anon_evictable_metadata); 5123 arc_kstat_update_state(arc_mru, 5124 &as->arcstat_mru_size, 5125 &as->arcstat_mru_evictable_data, 5126 &as->arcstat_mru_evictable_metadata); 5127 arc_kstat_update_state(arc_mru_ghost, 5128 &as->arcstat_mru_ghost_size, 5129 &as->arcstat_mru_ghost_evictable_data, 5130 &as->arcstat_mru_ghost_evictable_metadata); 5131 arc_kstat_update_state(arc_mfu, 5132 &as->arcstat_mfu_size, 5133 &as->arcstat_mfu_evictable_data, 5134 &as->arcstat_mfu_evictable_metadata); 5135 arc_kstat_update_state(arc_mfu_ghost, 5136 &as->arcstat_mfu_ghost_size, 5137 &as->arcstat_mfu_ghost_evictable_data, 5138 &as->arcstat_mfu_ghost_evictable_metadata); 5139 } 5140 5141 return (0); 5142} 5143 5144/* 5145 * This function *must* return indices evenly distributed between all 5146 * sublists of the multilist. This is needed due to how the ARC eviction 5147 * code is laid out; arc_evict_state() assumes ARC buffers are evenly 5148 * distributed between all sublists and uses this assumption when 5149 * deciding which sublist to evict from and how much to evict from it. 5150 */ 5151unsigned int 5152arc_state_multilist_index_func(multilist_t *ml, void *obj) 5153{ 5154 arc_buf_hdr_t *hdr = obj; 5155 5156 /* 5157 * We rely on b_dva to generate evenly distributed index 5158 * numbers using buf_hash below. So, as an added precaution, 5159 * let's make sure we never add empty buffers to the arc lists. 5160 */ 5161 ASSERT(!BUF_EMPTY(hdr)); 5162 5163 /* 5164 * The assumption here, is the hash value for a given 5165 * arc_buf_hdr_t will remain constant throughout it's lifetime 5166 * (i.e. it's b_spa, b_dva, and b_birth fields don't change). 5167 * Thus, we don't need to store the header's sublist index 5168 * on insertion, as this index can be recalculated on removal. 5169 * 5170 * Also, the low order bits of the hash value are thought to be 5171 * distributed evenly. Otherwise, in the case that the multilist 5172 * has a power of two number of sublists, each sublists' usage 5173 * would not be evenly distributed. 5174 */ 5175 return (buf_hash(hdr->b_spa, &hdr->b_dva, hdr->b_birth) % 5176 multilist_get_num_sublists(ml)); 5177} 5178 5179#ifdef _KERNEL 5180static eventhandler_tag arc_event_lowmem = NULL; 5181 5182static void 5183arc_lowmem(void *arg __unused, int howto __unused) 5184{ 5185 5186 mutex_enter(&arc_reclaim_lock); 5187 /* XXX: Memory deficit should be passed as argument. */ 5188 needfree = btoc(arc_c >> arc_shrink_shift); 5189 DTRACE_PROBE(arc__needfree); 5190 cv_signal(&arc_reclaim_thread_cv); 5191 5192 /* 5193 * It is unsafe to block here in arbitrary threads, because we can come 5194 * here from ARC itself and may hold ARC locks and thus risk a deadlock 5195 * with ARC reclaim thread. 5196 */ 5197 if (curproc == pageproc) 5198 (void) cv_wait(&arc_reclaim_waiters_cv, &arc_reclaim_lock); 5199 mutex_exit(&arc_reclaim_lock); 5200} 5201#endif 5202 5203void 5204arc_init(void) 5205{ 5206 int i, prefetch_tunable_set = 0; 5207 5208 mutex_init(&arc_reclaim_lock, NULL, MUTEX_DEFAULT, NULL); 5209 cv_init(&arc_reclaim_thread_cv, NULL, CV_DEFAULT, NULL); 5210 cv_init(&arc_reclaim_waiters_cv, NULL, CV_DEFAULT, NULL); 5211 5212 mutex_init(&arc_user_evicts_lock, NULL, MUTEX_DEFAULT, NULL); 5213 cv_init(&arc_user_evicts_cv, NULL, CV_DEFAULT, NULL); 5214 5215 /* Convert seconds to clock ticks */ 5216 arc_min_prefetch_lifespan = 1 * hz; 5217 5218 /* Start out with 1/8 of all memory */ 5219 arc_c = kmem_size() / 8; 5220 5221#ifdef sun 5222#ifdef _KERNEL 5223 /* 5224 * On architectures where the physical memory can be larger 5225 * than the addressable space (intel in 32-bit mode), we may 5226 * need to limit the cache to 1/8 of VM size. 5227 */ 5228 arc_c = MIN(arc_c, vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 8); 5229#endif 5230#endif /* sun */ 5231 /* set min cache to 1/32 of all memory, or 16MB, whichever is more */ 5232 arc_c_min = MAX(arc_c / 4, 16 << 20); 5233 /* set max to 1/2 of all memory, or all but 1GB, whichever is more */ 5234 if (arc_c * 8 >= 1 << 30) 5235 arc_c_max = (arc_c * 8) - (1 << 30); 5236 else 5237 arc_c_max = arc_c_min; 5238 arc_c_max = MAX(arc_c * 5, arc_c_max); 5239 5240#ifdef _KERNEL 5241 /* 5242 * Allow the tunables to override our calculations if they are 5243 * reasonable (ie. over 16MB) 5244 */ 5245 if (zfs_arc_max > 16 << 20 && zfs_arc_max < kmem_size()) 5246 arc_c_max = zfs_arc_max; 5247 if (zfs_arc_min > 16 << 20 && zfs_arc_min <= arc_c_max) 5248 arc_c_min = zfs_arc_min; 5249#endif 5250 5251 arc_c = arc_c_max; 5252 arc_p = (arc_c >> 1); 5253 5254 /* limit meta-data to 1/4 of the arc capacity */ 5255 arc_meta_limit = arc_c_max / 4; 5256 5257 /* Allow the tunable to override if it is reasonable */ 5258 if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max) 5259 arc_meta_limit = zfs_arc_meta_limit; 5260 5261 if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0) 5262 arc_c_min = arc_meta_limit / 2; 5263 5264 if (zfs_arc_meta_min > 0) { 5265 arc_meta_min = zfs_arc_meta_min; 5266 } else { 5267 arc_meta_min = arc_c_min / 2; 5268 } 5269 5270 if (zfs_arc_grow_retry > 0) 5271 arc_grow_retry = zfs_arc_grow_retry; 5272 5273 if (zfs_arc_shrink_shift > 0) 5274 arc_shrink_shift = zfs_arc_shrink_shift; 5275 5276 /* 5277 * Ensure that arc_no_grow_shift is less than arc_shrink_shift. 5278 */ 5279 if (arc_no_grow_shift >= arc_shrink_shift) 5280 arc_no_grow_shift = arc_shrink_shift - 1; 5281 5282 if (zfs_arc_p_min_shift > 0) 5283 arc_p_min_shift = zfs_arc_p_min_shift; 5284 5285 if (zfs_arc_num_sublists_per_state < 1) 5286 zfs_arc_num_sublists_per_state = MAX(max_ncpus, 1); 5287 5288 /* if kmem_flags are set, lets try to use less memory */ 5289 if (kmem_debugging()) 5290 arc_c = arc_c / 2; 5291 if (arc_c < arc_c_min) 5292 arc_c = arc_c_min; 5293 5294 zfs_arc_min = arc_c_min; 5295 zfs_arc_max = arc_c_max; 5296 5297 arc_anon = &ARC_anon; 5298 arc_mru = &ARC_mru; 5299 arc_mru_ghost = &ARC_mru_ghost; 5300 arc_mfu = &ARC_mfu; 5301 arc_mfu_ghost = &ARC_mfu_ghost; 5302 arc_l2c_only = &ARC_l2c_only; 5303 arc_size = 0; 5304 5305 multilist_create(&arc_mru->arcs_list[ARC_BUFC_METADATA], 5306 sizeof (arc_buf_hdr_t), 5307 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5308 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5309 multilist_create(&arc_mru->arcs_list[ARC_BUFC_DATA], 5310 sizeof (arc_buf_hdr_t), 5311 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5312 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5313 multilist_create(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA], 5314 sizeof (arc_buf_hdr_t), 5315 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5316 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5317 multilist_create(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA], 5318 sizeof (arc_buf_hdr_t), 5319 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5320 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5321 multilist_create(&arc_mfu->arcs_list[ARC_BUFC_METADATA], 5322 sizeof (arc_buf_hdr_t), 5323 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5324 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5325 multilist_create(&arc_mfu->arcs_list[ARC_BUFC_DATA], 5326 sizeof (arc_buf_hdr_t), 5327 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5328 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5329 multilist_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA], 5330 sizeof (arc_buf_hdr_t), 5331 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5332 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5333 multilist_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA], 5334 sizeof (arc_buf_hdr_t), 5335 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5336 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5337 multilist_create(&arc_l2c_only->arcs_list[ARC_BUFC_METADATA], 5338 sizeof (arc_buf_hdr_t), 5339 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5340 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5341 multilist_create(&arc_l2c_only->arcs_list[ARC_BUFC_DATA], 5342 sizeof (arc_buf_hdr_t), 5343 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5344 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5345 5346 refcount_create(&arc_anon->arcs_size); 5347 refcount_create(&arc_mru->arcs_size); 5348 refcount_create(&arc_mru_ghost->arcs_size); 5349 refcount_create(&arc_mfu->arcs_size); 5350 refcount_create(&arc_mfu_ghost->arcs_size); 5351 refcount_create(&arc_l2c_only->arcs_size); 5352 5353 buf_init(); 5354 5355 arc_reclaim_thread_exit = FALSE; 5356 arc_user_evicts_thread_exit = FALSE; 5357 arc_eviction_list = NULL; 5358 bzero(&arc_eviction_hdr, sizeof (arc_buf_hdr_t)); 5359 5360 arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED, 5361 sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL); 5362 5363 if (arc_ksp != NULL) { 5364 arc_ksp->ks_data = &arc_stats; 5365 arc_ksp->ks_update = arc_kstat_update; 5366 kstat_install(arc_ksp); 5367 } 5368 5369 (void) thread_create(NULL, 0, arc_reclaim_thread, NULL, 0, &p0, 5370 TS_RUN, minclsyspri); 5371 5372#ifdef _KERNEL 5373 arc_event_lowmem = EVENTHANDLER_REGISTER(vm_lowmem, arc_lowmem, NULL, 5374 EVENTHANDLER_PRI_FIRST); 5375#endif 5376 5377 (void) thread_create(NULL, 0, arc_user_evicts_thread, NULL, 0, &p0, 5378 TS_RUN, minclsyspri); 5379 5380 arc_dead = FALSE; 5381 arc_warm = B_FALSE; 5382 5383 /* 5384 * Calculate maximum amount of dirty data per pool. 5385 * 5386 * If it has been set by /etc/system, take that. 5387 * Otherwise, use a percentage of physical memory defined by 5388 * zfs_dirty_data_max_percent (default 10%) with a cap at 5389 * zfs_dirty_data_max_max (default 4GB). 5390 */ 5391 if (zfs_dirty_data_max == 0) { 5392 zfs_dirty_data_max = ptob(physmem) * 5393 zfs_dirty_data_max_percent / 100; 5394 zfs_dirty_data_max = MIN(zfs_dirty_data_max, 5395 zfs_dirty_data_max_max); 5396 } 5397 5398#ifdef _KERNEL 5399 if (TUNABLE_INT_FETCH("vfs.zfs.prefetch_disable", &zfs_prefetch_disable)) 5400 prefetch_tunable_set = 1; 5401 5402#ifdef __i386__ 5403 if (prefetch_tunable_set == 0) { 5404 printf("ZFS NOTICE: Prefetch is disabled by default on i386 " 5405 "-- to enable,\n"); 5406 printf(" add \"vfs.zfs.prefetch_disable=0\" " 5407 "to /boot/loader.conf.\n"); 5408 zfs_prefetch_disable = 1; 5409 } 5410#else 5411 if ((((uint64_t)physmem * PAGESIZE) < (1ULL << 32)) && 5412 prefetch_tunable_set == 0) { 5413 printf("ZFS NOTICE: Prefetch is disabled by default if less " 5414 "than 4GB of RAM is present;\n" 5415 " to enable, add \"vfs.zfs.prefetch_disable=0\" " 5416 "to /boot/loader.conf.\n"); 5417 zfs_prefetch_disable = 1; 5418 } 5419#endif 5420 /* Warn about ZFS memory and address space requirements. */ 5421 if (((uint64_t)physmem * PAGESIZE) < (256 + 128 + 64) * (1 << 20)) { 5422 printf("ZFS WARNING: Recommended minimum RAM size is 512MB; " 5423 "expect unstable behavior.\n"); 5424 } 5425 if (kmem_size() < 512 * (1 << 20)) { 5426 printf("ZFS WARNING: Recommended minimum kmem_size is 512MB; " 5427 "expect unstable behavior.\n"); 5428 printf(" Consider tuning vm.kmem_size and " 5429 "vm.kmem_size_max\n"); 5430 printf(" in /boot/loader.conf.\n"); 5431 } 5432#endif 5433} 5434 5435void 5436arc_fini(void) 5437{ 5438 mutex_enter(&arc_reclaim_lock); 5439 arc_reclaim_thread_exit = TRUE; 5440 /* 5441 * The reclaim thread will set arc_reclaim_thread_exit back to 5442 * FALSE when it is finished exiting; we're waiting for that. 5443 */ 5444 while (arc_reclaim_thread_exit) { 5445 cv_signal(&arc_reclaim_thread_cv); 5446 cv_wait(&arc_reclaim_thread_cv, &arc_reclaim_lock); 5447 } 5448 mutex_exit(&arc_reclaim_lock); 5449 5450 mutex_enter(&arc_user_evicts_lock); 5451 arc_user_evicts_thread_exit = TRUE; 5452 /* 5453 * The user evicts thread will set arc_user_evicts_thread_exit 5454 * to FALSE when it is finished exiting; we're waiting for that. 5455 */ 5456 while (arc_user_evicts_thread_exit) { 5457 cv_signal(&arc_user_evicts_cv); 5458 cv_wait(&arc_user_evicts_cv, &arc_user_evicts_lock); 5459 } 5460 mutex_exit(&arc_user_evicts_lock); 5461 5462 /* Use TRUE to ensure *all* buffers are evicted */ 5463 arc_flush(NULL, TRUE); 5464 5465 arc_dead = TRUE; 5466 5467 if (arc_ksp != NULL) { 5468 kstat_delete(arc_ksp); 5469 arc_ksp = NULL; 5470 } 5471 5472 mutex_destroy(&arc_reclaim_lock); 5473 cv_destroy(&arc_reclaim_thread_cv); 5474 cv_destroy(&arc_reclaim_waiters_cv); 5475 5476 mutex_destroy(&arc_user_evicts_lock); 5477 cv_destroy(&arc_user_evicts_cv); 5478 5479 refcount_destroy(&arc_anon->arcs_size); 5480 refcount_destroy(&arc_mru->arcs_size); 5481 refcount_destroy(&arc_mru_ghost->arcs_size); 5482 refcount_destroy(&arc_mfu->arcs_size); 5483 refcount_destroy(&arc_mfu_ghost->arcs_size); 5484 refcount_destroy(&arc_l2c_only->arcs_size); 5485 5486 multilist_destroy(&arc_mru->arcs_list[ARC_BUFC_METADATA]); 5487 multilist_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA]); 5488 multilist_destroy(&arc_mfu->arcs_list[ARC_BUFC_METADATA]); 5489 multilist_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA]); 5490 multilist_destroy(&arc_mru->arcs_list[ARC_BUFC_DATA]); 5491 multilist_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA]); 5492 multilist_destroy(&arc_mfu->arcs_list[ARC_BUFC_DATA]); 5493 multilist_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA]); 5494 5495 buf_fini(); 5496 5497 ASSERT0(arc_loaned_bytes); 5498 5499#ifdef _KERNEL 5500 if (arc_event_lowmem != NULL) 5501 EVENTHANDLER_DEREGISTER(vm_lowmem, arc_event_lowmem); 5502#endif 5503} 5504 5505/* 5506 * Level 2 ARC 5507 * 5508 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk. 5509 * It uses dedicated storage devices to hold cached data, which are populated 5510 * using large infrequent writes. The main role of this cache is to boost 5511 * the performance of random read workloads. The intended L2ARC devices 5512 * include short-stroked disks, solid state disks, and other media with 5513 * substantially faster read latency than disk. 5514 * 5515 * +-----------------------+ 5516 * | ARC | 5517 * +-----------------------+ 5518 * | ^ ^ 5519 * | | | 5520 * l2arc_feed_thread() arc_read() 5521 * | | | 5522 * | l2arc read | 5523 * V | | 5524 * +---------------+ | 5525 * | L2ARC | | 5526 * +---------------+ | 5527 * | ^ | 5528 * l2arc_write() | | 5529 * | | | 5530 * V | | 5531 * +-------+ +-------+ 5532 * | vdev | | vdev | 5533 * | cache | | cache | 5534 * +-------+ +-------+ 5535 * +=========+ .-----. 5536 * : L2ARC : |-_____-| 5537 * : devices : | Disks | 5538 * +=========+ `-_____-' 5539 * 5540 * Read requests are satisfied from the following sources, in order: 5541 * 5542 * 1) ARC 5543 * 2) vdev cache of L2ARC devices 5544 * 3) L2ARC devices 5545 * 4) vdev cache of disks 5546 * 5) disks 5547 * 5548 * Some L2ARC device types exhibit extremely slow write performance. 5549 * To accommodate for this there are some significant differences between 5550 * the L2ARC and traditional cache design: 5551 * 5552 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from 5553 * the ARC behave as usual, freeing buffers and placing headers on ghost 5554 * lists. The ARC does not send buffers to the L2ARC during eviction as 5555 * this would add inflated write latencies for all ARC memory pressure. 5556 * 5557 * 2. The L2ARC attempts to cache data from the ARC before it is evicted. 5558 * It does this by periodically scanning buffers from the eviction-end of 5559 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are 5560 * not already there. It scans until a headroom of buffers is satisfied, 5561 * which itself is a buffer for ARC eviction. If a compressible buffer is 5562 * found during scanning and selected for writing to an L2ARC device, we 5563 * temporarily boost scanning headroom during the next scan cycle to make 5564 * sure we adapt to compression effects (which might significantly reduce 5565 * the data volume we write to L2ARC). The thread that does this is 5566 * l2arc_feed_thread(), illustrated below; example sizes are included to 5567 * provide a better sense of ratio than this diagram: 5568 * 5569 * head --> tail 5570 * +---------------------+----------+ 5571 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC 5572 * +---------------------+----------+ | o L2ARC eligible 5573 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer 5574 * +---------------------+----------+ | 5575 * 15.9 Gbytes ^ 32 Mbytes | 5576 * headroom | 5577 * l2arc_feed_thread() 5578 * | 5579 * l2arc write hand <--[oooo]--' 5580 * | 8 Mbyte 5581 * | write max 5582 * V 5583 * +==============================+ 5584 * L2ARC dev |####|#|###|###| |####| ... | 5585 * +==============================+ 5586 * 32 Gbytes 5587 * 5588 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of 5589 * evicted, then the L2ARC has cached a buffer much sooner than it probably 5590 * needed to, potentially wasting L2ARC device bandwidth and storage. It is 5591 * safe to say that this is an uncommon case, since buffers at the end of 5592 * the ARC lists have moved there due to inactivity. 5593 * 5594 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom, 5595 * then the L2ARC simply misses copying some buffers. This serves as a 5596 * pressure valve to prevent heavy read workloads from both stalling the ARC 5597 * with waits and clogging the L2ARC with writes. This also helps prevent 5598 * the potential for the L2ARC to churn if it attempts to cache content too 5599 * quickly, such as during backups of the entire pool. 5600 * 5601 * 5. After system boot and before the ARC has filled main memory, there are 5602 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru 5603 * lists can remain mostly static. Instead of searching from tail of these 5604 * lists as pictured, the l2arc_feed_thread() will search from the list heads 5605 * for eligible buffers, greatly increasing its chance of finding them. 5606 * 5607 * The L2ARC device write speed is also boosted during this time so that 5608 * the L2ARC warms up faster. Since there have been no ARC evictions yet, 5609 * there are no L2ARC reads, and no fear of degrading read performance 5610 * through increased writes. 5611 * 5612 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that 5613 * the vdev queue can aggregate them into larger and fewer writes. Each 5614 * device is written to in a rotor fashion, sweeping writes through 5615 * available space then repeating. 5616 * 5617 * 7. The L2ARC does not store dirty content. It never needs to flush 5618 * write buffers back to disk based storage. 5619 * 5620 * 8. If an ARC buffer is written (and dirtied) which also exists in the 5621 * L2ARC, the now stale L2ARC buffer is immediately dropped. 5622 * 5623 * The performance of the L2ARC can be tweaked by a number of tunables, which 5624 * may be necessary for different workloads: 5625 * 5626 * l2arc_write_max max write bytes per interval 5627 * l2arc_write_boost extra write bytes during device warmup 5628 * l2arc_noprefetch skip caching prefetched buffers 5629 * l2arc_headroom number of max device writes to precache 5630 * l2arc_headroom_boost when we find compressed buffers during ARC 5631 * scanning, we multiply headroom by this 5632 * percentage factor for the next scan cycle, 5633 * since more compressed buffers are likely to 5634 * be present 5635 * l2arc_feed_secs seconds between L2ARC writing 5636 * 5637 * Tunables may be removed or added as future performance improvements are 5638 * integrated, and also may become zpool properties. 5639 * 5640 * There are three key functions that control how the L2ARC warms up: 5641 * 5642 * l2arc_write_eligible() check if a buffer is eligible to cache 5643 * l2arc_write_size() calculate how much to write 5644 * l2arc_write_interval() calculate sleep delay between writes 5645 * 5646 * These three functions determine what to write, how much, and how quickly 5647 * to send writes. 5648 */ 5649 5650static boolean_t 5651l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *hdr) 5652{ 5653 /* 5654 * A buffer is *not* eligible for the L2ARC if it: 5655 * 1. belongs to a different spa. 5656 * 2. is already cached on the L2ARC. 5657 * 3. has an I/O in progress (it may be an incomplete read). 5658 * 4. is flagged not eligible (zfs property). 5659 */ 5660 if (hdr->b_spa != spa_guid) { 5661 ARCSTAT_BUMP(arcstat_l2_write_spa_mismatch); 5662 return (B_FALSE); 5663 } 5664 if (HDR_HAS_L2HDR(hdr)) { 5665 ARCSTAT_BUMP(arcstat_l2_write_in_l2); 5666 return (B_FALSE); 5667 } 5668 if (HDR_IO_IN_PROGRESS(hdr)) { 5669 ARCSTAT_BUMP(arcstat_l2_write_hdr_io_in_progress); 5670 return (B_FALSE); 5671 } 5672 if (!HDR_L2CACHE(hdr)) { 5673 ARCSTAT_BUMP(arcstat_l2_write_not_cacheable); 5674 return (B_FALSE); 5675 } 5676 5677 return (B_TRUE); 5678} 5679 5680static uint64_t 5681l2arc_write_size(void) 5682{ 5683 uint64_t size; 5684 5685 /* 5686 * Make sure our globals have meaningful values in case the user 5687 * altered them. 5688 */ 5689 size = l2arc_write_max; 5690 if (size == 0) { 5691 cmn_err(CE_NOTE, "Bad value for l2arc_write_max, value must " 5692 "be greater than zero, resetting it to the default (%d)", 5693 L2ARC_WRITE_SIZE); 5694 size = l2arc_write_max = L2ARC_WRITE_SIZE; 5695 } 5696 5697 if (arc_warm == B_FALSE) 5698 size += l2arc_write_boost; 5699 5700 return (size); 5701 5702} 5703 5704static clock_t 5705l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote) 5706{ 5707 clock_t interval, next, now; 5708 5709 /* 5710 * If the ARC lists are busy, increase our write rate; if the 5711 * lists are stale, idle back. This is achieved by checking 5712 * how much we previously wrote - if it was more than half of 5713 * what we wanted, schedule the next write much sooner. 5714 */ 5715 if (l2arc_feed_again && wrote > (wanted / 2)) 5716 interval = (hz * l2arc_feed_min_ms) / 1000; 5717 else 5718 interval = hz * l2arc_feed_secs; 5719 5720 now = ddi_get_lbolt(); 5721 next = MAX(now, MIN(now + interval, began + interval)); 5722 5723 return (next); 5724} 5725 5726/* 5727 * Cycle through L2ARC devices. This is how L2ARC load balances. 5728 * If a device is returned, this also returns holding the spa config lock. 5729 */ 5730static l2arc_dev_t * 5731l2arc_dev_get_next(void) 5732{ 5733 l2arc_dev_t *first, *next = NULL; 5734 5735 /* 5736 * Lock out the removal of spas (spa_namespace_lock), then removal 5737 * of cache devices (l2arc_dev_mtx). Once a device has been selected, 5738 * both locks will be dropped and a spa config lock held instead. 5739 */ 5740 mutex_enter(&spa_namespace_lock); 5741 mutex_enter(&l2arc_dev_mtx); 5742 5743 /* if there are no vdevs, there is nothing to do */ 5744 if (l2arc_ndev == 0) 5745 goto out; 5746 5747 first = NULL; 5748 next = l2arc_dev_last; 5749 do { 5750 /* loop around the list looking for a non-faulted vdev */ 5751 if (next == NULL) { 5752 next = list_head(l2arc_dev_list); 5753 } else { 5754 next = list_next(l2arc_dev_list, next); 5755 if (next == NULL) 5756 next = list_head(l2arc_dev_list); 5757 } 5758 5759 /* if we have come back to the start, bail out */ 5760 if (first == NULL) 5761 first = next; 5762 else if (next == first) 5763 break; 5764 5765 } while (vdev_is_dead(next->l2ad_vdev)); 5766 5767 /* if we were unable to find any usable vdevs, return NULL */ 5768 if (vdev_is_dead(next->l2ad_vdev)) 5769 next = NULL; 5770 5771 l2arc_dev_last = next; 5772 5773out: 5774 mutex_exit(&l2arc_dev_mtx); 5775 5776 /* 5777 * Grab the config lock to prevent the 'next' device from being 5778 * removed while we are writing to it. 5779 */ 5780 if (next != NULL) 5781 spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER); 5782 mutex_exit(&spa_namespace_lock); 5783 5784 return (next); 5785} 5786 5787/* 5788 * Free buffers that were tagged for destruction. 5789 */ 5790static void 5791l2arc_do_free_on_write() 5792{ 5793 list_t *buflist; 5794 l2arc_data_free_t *df, *df_prev; 5795 5796 mutex_enter(&l2arc_free_on_write_mtx); 5797 buflist = l2arc_free_on_write; 5798 5799 for (df = list_tail(buflist); df; df = df_prev) { 5800 df_prev = list_prev(buflist, df); 5801 ASSERT(df->l2df_data != NULL); 5802 ASSERT(df->l2df_func != NULL); 5803 df->l2df_func(df->l2df_data, df->l2df_size); 5804 list_remove(buflist, df); 5805 kmem_free(df, sizeof (l2arc_data_free_t)); 5806 } 5807 5808 mutex_exit(&l2arc_free_on_write_mtx); 5809} 5810 5811/* 5812 * A write to a cache device has completed. Update all headers to allow 5813 * reads from these buffers to begin. 5814 */ 5815static void 5816l2arc_write_done(zio_t *zio) 5817{ 5818 l2arc_write_callback_t *cb; 5819 l2arc_dev_t *dev; 5820 list_t *buflist; 5821 arc_buf_hdr_t *head, *hdr, *hdr_prev; 5822 kmutex_t *hash_lock; 5823 int64_t bytes_dropped = 0; 5824 5825 cb = zio->io_private; 5826 ASSERT(cb != NULL); 5827 dev = cb->l2wcb_dev; 5828 ASSERT(dev != NULL); 5829 head = cb->l2wcb_head; 5830 ASSERT(head != NULL); 5831 buflist = &dev->l2ad_buflist; 5832 ASSERT(buflist != NULL); 5833 DTRACE_PROBE2(l2arc__iodone, zio_t *, zio, 5834 l2arc_write_callback_t *, cb); 5835 5836 if (zio->io_error != 0) 5837 ARCSTAT_BUMP(arcstat_l2_writes_error); 5838 5839 /* 5840 * All writes completed, or an error was hit. 5841 */ 5842top: 5843 mutex_enter(&dev->l2ad_mtx); 5844 for (hdr = list_prev(buflist, head); hdr; hdr = hdr_prev) { 5845 hdr_prev = list_prev(buflist, hdr); 5846 5847 hash_lock = HDR_LOCK(hdr); 5848 5849 /* 5850 * We cannot use mutex_enter or else we can deadlock 5851 * with l2arc_write_buffers (due to swapping the order 5852 * the hash lock and l2ad_mtx are taken). 5853 */ 5854 if (!mutex_tryenter(hash_lock)) { 5855 /* 5856 * Missed the hash lock. We must retry so we 5857 * don't leave the ARC_FLAG_L2_WRITING bit set. 5858 */ 5859 ARCSTAT_BUMP(arcstat_l2_writes_lock_retry); 5860 5861 /* 5862 * We don't want to rescan the headers we've 5863 * already marked as having been written out, so 5864 * we reinsert the head node so we can pick up 5865 * where we left off. 5866 */ 5867 list_remove(buflist, head); 5868 list_insert_after(buflist, hdr, head); 5869 5870 mutex_exit(&dev->l2ad_mtx); 5871 5872 /* 5873 * We wait for the hash lock to become available 5874 * to try and prevent busy waiting, and increase 5875 * the chance we'll be able to acquire the lock 5876 * the next time around. 5877 */ 5878 mutex_enter(hash_lock); 5879 mutex_exit(hash_lock); 5880 goto top; 5881 } 5882 5883 /* 5884 * We could not have been moved into the arc_l2c_only 5885 * state while in-flight due to our ARC_FLAG_L2_WRITING 5886 * bit being set. Let's just ensure that's being enforced. 5887 */ 5888 ASSERT(HDR_HAS_L1HDR(hdr)); 5889 5890 /* 5891 * We may have allocated a buffer for L2ARC compression, 5892 * we must release it to avoid leaking this data. 5893 */ 5894 l2arc_release_cdata_buf(hdr); 5895 5896 if (zio->io_error != 0) { 5897 /* 5898 * Error - drop L2ARC entry. 5899 */ 5900 trim_map_free(hdr->b_l2hdr.b_dev->l2ad_vdev, 5901 hdr->b_l2hdr.b_daddr, hdr->b_l2hdr.b_asize, 0); 5902 hdr->b_flags &= ~ARC_FLAG_HAS_L2HDR; 5903 5904 ARCSTAT_INCR(arcstat_l2_asize, -hdr->b_l2hdr.b_asize); 5905 ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size); 5906 5907 bytes_dropped += hdr->b_l2hdr.b_asize; 5908 (void) refcount_remove_many(&dev->l2ad_alloc, 5909 hdr->b_l2hdr.b_asize, hdr); 5910 } 5911 5912 /* 5913 * Allow ARC to begin reads and ghost list evictions to 5914 * this L2ARC entry. 5915 */ 5916 hdr->b_flags &= ~ARC_FLAG_L2_WRITING; 5917 5918 mutex_exit(hash_lock); 5919 } 5920 5921 atomic_inc_64(&l2arc_writes_done); 5922 list_remove(buflist, head); 5923 ASSERT(!HDR_HAS_L1HDR(head)); 5924 kmem_cache_free(hdr_l2only_cache, head); 5925 mutex_exit(&dev->l2ad_mtx); 5926 5927 vdev_space_update(dev->l2ad_vdev, -bytes_dropped, 0, 0); 5928 5929 l2arc_do_free_on_write(); 5930 5931 kmem_free(cb, sizeof (l2arc_write_callback_t)); 5932} 5933 5934/* 5935 * A read to a cache device completed. Validate buffer contents before 5936 * handing over to the regular ARC routines. 5937 */ 5938static void 5939l2arc_read_done(zio_t *zio) 5940{ 5941 l2arc_read_callback_t *cb; 5942 arc_buf_hdr_t *hdr; 5943 arc_buf_t *buf; 5944 kmutex_t *hash_lock; 5945 int equal; 5946 5947 ASSERT(zio->io_vd != NULL); 5948 ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE); 5949 5950 spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd); 5951 5952 cb = zio->io_private; 5953 ASSERT(cb != NULL); 5954 buf = cb->l2rcb_buf; 5955 ASSERT(buf != NULL); 5956 5957 hash_lock = HDR_LOCK(buf->b_hdr); 5958 mutex_enter(hash_lock); 5959 hdr = buf->b_hdr; 5960 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 5961 5962 /* 5963 * If the buffer was compressed, decompress it first. 5964 */ 5965 if (cb->l2rcb_compress != ZIO_COMPRESS_OFF) 5966 l2arc_decompress_zio(zio, hdr, cb->l2rcb_compress); 5967 ASSERT(zio->io_data != NULL); 5968 5969 /* 5970 * Check this survived the L2ARC journey. 5971 */ 5972 equal = arc_cksum_equal(buf); 5973 if (equal && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) { 5974 mutex_exit(hash_lock); 5975 zio->io_private = buf; 5976 zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */ 5977 zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */ 5978 arc_read_done(zio); 5979 } else { 5980 mutex_exit(hash_lock); 5981 /* 5982 * Buffer didn't survive caching. Increment stats and 5983 * reissue to the original storage device. 5984 */ 5985 if (zio->io_error != 0) { 5986 ARCSTAT_BUMP(arcstat_l2_io_error); 5987 } else { 5988 zio->io_error = SET_ERROR(EIO); 5989 } 5990 if (!equal) 5991 ARCSTAT_BUMP(arcstat_l2_cksum_bad); 5992 5993 /* 5994 * If there's no waiter, issue an async i/o to the primary 5995 * storage now. If there *is* a waiter, the caller must 5996 * issue the i/o in a context where it's OK to block. 5997 */ 5998 if (zio->io_waiter == NULL) { 5999 zio_t *pio = zio_unique_parent(zio); 6000 6001 ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL); 6002 6003 zio_nowait(zio_read(pio, cb->l2rcb_spa, &cb->l2rcb_bp, 6004 buf->b_data, zio->io_size, arc_read_done, buf, 6005 zio->io_priority, cb->l2rcb_flags, &cb->l2rcb_zb)); 6006 } 6007 } 6008 6009 kmem_free(cb, sizeof (l2arc_read_callback_t)); 6010} 6011 6012/* 6013 * This is the list priority from which the L2ARC will search for pages to 6014 * cache. This is used within loops (0..3) to cycle through lists in the 6015 * desired order. This order can have a significant effect on cache 6016 * performance. 6017 * 6018 * Currently the metadata lists are hit first, MFU then MRU, followed by 6019 * the data lists. This function returns a locked list, and also returns 6020 * the lock pointer. 6021 */ 6022static multilist_sublist_t * 6023l2arc_sublist_lock(int list_num) 6024{ 6025 multilist_t *ml = NULL; 6026 unsigned int idx; 6027 6028 ASSERT(list_num >= 0 && list_num <= 3); 6029 6030 switch (list_num) { 6031 case 0: 6032 ml = &arc_mfu->arcs_list[ARC_BUFC_METADATA]; 6033 break; 6034 case 1: 6035 ml = &arc_mru->arcs_list[ARC_BUFC_METADATA]; 6036 break; 6037 case 2: 6038 ml = &arc_mfu->arcs_list[ARC_BUFC_DATA]; 6039 break; 6040 case 3: 6041 ml = &arc_mru->arcs_list[ARC_BUFC_DATA]; 6042 break; 6043 } 6044 6045 /* 6046 * Return a randomly-selected sublist. This is acceptable 6047 * because the caller feeds only a little bit of data for each 6048 * call (8MB). Subsequent calls will result in different 6049 * sublists being selected. 6050 */ 6051 idx = multilist_get_random_index(ml); 6052 return (multilist_sublist_lock(ml, idx)); 6053} 6054 6055/* 6056 * Evict buffers from the device write hand to the distance specified in 6057 * bytes. This distance may span populated buffers, it may span nothing. 6058 * This is clearing a region on the L2ARC device ready for writing. 6059 * If the 'all' boolean is set, every buffer is evicted. 6060 */ 6061static void 6062l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all) 6063{ 6064 list_t *buflist; 6065 arc_buf_hdr_t *hdr, *hdr_prev; 6066 kmutex_t *hash_lock; 6067 uint64_t taddr; 6068 6069 buflist = &dev->l2ad_buflist; 6070 6071 if (!all && dev->l2ad_first) { 6072 /* 6073 * This is the first sweep through the device. There is 6074 * nothing to evict. 6075 */ 6076 return; 6077 } 6078 6079 if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) { 6080 /* 6081 * When nearing the end of the device, evict to the end 6082 * before the device write hand jumps to the start. 6083 */ 6084 taddr = dev->l2ad_end; 6085 } else { 6086 taddr = dev->l2ad_hand + distance; 6087 } 6088 DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist, 6089 uint64_t, taddr, boolean_t, all); 6090 6091top: 6092 mutex_enter(&dev->l2ad_mtx); 6093 for (hdr = list_tail(buflist); hdr; hdr = hdr_prev) { 6094 hdr_prev = list_prev(buflist, hdr); 6095 6096 hash_lock = HDR_LOCK(hdr); 6097 6098 /* 6099 * We cannot use mutex_enter or else we can deadlock 6100 * with l2arc_write_buffers (due to swapping the order 6101 * the hash lock and l2ad_mtx are taken). 6102 */ 6103 if (!mutex_tryenter(hash_lock)) { 6104 /* 6105 * Missed the hash lock. Retry. 6106 */ 6107 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry); 6108 mutex_exit(&dev->l2ad_mtx); 6109 mutex_enter(hash_lock); 6110 mutex_exit(hash_lock); 6111 goto top; 6112 } 6113 6114 if (HDR_L2_WRITE_HEAD(hdr)) { 6115 /* 6116 * We hit a write head node. Leave it for 6117 * l2arc_write_done(). 6118 */ 6119 list_remove(buflist, hdr); 6120 mutex_exit(hash_lock); 6121 continue; 6122 } 6123 6124 if (!all && HDR_HAS_L2HDR(hdr) && 6125 (hdr->b_l2hdr.b_daddr > taddr || 6126 hdr->b_l2hdr.b_daddr < dev->l2ad_hand)) { 6127 /* 6128 * We've evicted to the target address, 6129 * or the end of the device. 6130 */ 6131 mutex_exit(hash_lock); 6132 break; 6133 } 6134 6135 ASSERT(HDR_HAS_L2HDR(hdr)); 6136 if (!HDR_HAS_L1HDR(hdr)) { 6137 ASSERT(!HDR_L2_READING(hdr)); 6138 /* 6139 * This doesn't exist in the ARC. Destroy. 6140 * arc_hdr_destroy() will call list_remove() 6141 * and decrement arcstat_l2_size. 6142 */ 6143 arc_change_state(arc_anon, hdr, hash_lock); 6144 arc_hdr_destroy(hdr); 6145 } else { 6146 ASSERT(hdr->b_l1hdr.b_state != arc_l2c_only); 6147 ARCSTAT_BUMP(arcstat_l2_evict_l1cached); 6148 /* 6149 * Invalidate issued or about to be issued 6150 * reads, since we may be about to write 6151 * over this location. 6152 */ 6153 if (HDR_L2_READING(hdr)) { 6154 ARCSTAT_BUMP(arcstat_l2_evict_reading); 6155 hdr->b_flags |= ARC_FLAG_L2_EVICTED; 6156 } 6157 6158 /* Ensure this header has finished being written */ 6159 ASSERT(!HDR_L2_WRITING(hdr)); 6160 ASSERT3P(hdr->b_l1hdr.b_tmp_cdata, ==, NULL); 6161 6162 arc_hdr_l2hdr_destroy(hdr); 6163 } 6164 mutex_exit(hash_lock); 6165 } 6166 mutex_exit(&dev->l2ad_mtx); 6167} 6168 6169/* 6170 * Find and write ARC buffers to the L2ARC device. 6171 * 6172 * An ARC_FLAG_L2_WRITING flag is set so that the L2ARC buffers are not valid 6173 * for reading until they have completed writing. 6174 * The headroom_boost is an in-out parameter used to maintain headroom boost 6175 * state between calls to this function. 6176 * 6177 * Returns the number of bytes actually written (which may be smaller than 6178 * the delta by which the device hand has changed due to alignment). 6179 */ 6180static uint64_t 6181l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz, 6182 boolean_t *headroom_boost) 6183{ 6184 arc_buf_hdr_t *hdr, *hdr_prev, *head; 6185 uint64_t write_asize, write_sz, headroom, buf_compress_minsz; 6186 void *buf_data; 6187 boolean_t full; 6188 l2arc_write_callback_t *cb; 6189 zio_t *pio, *wzio; 6190 uint64_t guid = spa_load_guid(spa); 6191 const boolean_t do_headroom_boost = *headroom_boost; 6192 int try; 6193 6194 ASSERT(dev->l2ad_vdev != NULL); 6195 6196 /* Lower the flag now, we might want to raise it again later. */ 6197 *headroom_boost = B_FALSE; 6198 6199 pio = NULL; 6200 write_sz = write_asize = 0; 6201 full = B_FALSE; 6202 head = kmem_cache_alloc(hdr_l2only_cache, KM_PUSHPAGE); 6203 head->b_flags |= ARC_FLAG_L2_WRITE_HEAD; 6204 head->b_flags |= ARC_FLAG_HAS_L2HDR; 6205 6206 ARCSTAT_BUMP(arcstat_l2_write_buffer_iter); 6207 /* 6208 * We will want to try to compress buffers that are at least 2x the 6209 * device sector size. 6210 */ 6211 buf_compress_minsz = 2 << dev->l2ad_vdev->vdev_ashift; 6212 6213 /* 6214 * Copy buffers for L2ARC writing. 6215 */ 6216 for (try = 0; try <= 3; try++) { 6217 multilist_sublist_t *mls = l2arc_sublist_lock(try); 6218 uint64_t passed_sz = 0; 6219 6220 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_iter); 6221 6222 /* 6223 * L2ARC fast warmup. 6224 * 6225 * Until the ARC is warm and starts to evict, read from the 6226 * head of the ARC lists rather than the tail. 6227 */ 6228 if (arc_warm == B_FALSE) 6229 hdr = multilist_sublist_head(mls); 6230 else 6231 hdr = multilist_sublist_tail(mls); 6232 if (hdr == NULL) 6233 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_null_iter); 6234 6235 headroom = target_sz * l2arc_headroom; 6236 if (do_headroom_boost) 6237 headroom = (headroom * l2arc_headroom_boost) / 100; 6238 6239 for (; hdr; hdr = hdr_prev) { 6240 kmutex_t *hash_lock; 6241 uint64_t buf_sz; 6242 uint64_t buf_a_sz; 6243 6244 if (arc_warm == B_FALSE) 6245 hdr_prev = multilist_sublist_next(mls, hdr); 6246 else 6247 hdr_prev = multilist_sublist_prev(mls, hdr); 6248 ARCSTAT_INCR(arcstat_l2_write_buffer_bytes_scanned, hdr->b_size); 6249 6250 hash_lock = HDR_LOCK(hdr); 6251 if (!mutex_tryenter(hash_lock)) { 6252 ARCSTAT_BUMP(arcstat_l2_write_trylock_fail); 6253 /* 6254 * Skip this buffer rather than waiting. 6255 */ 6256 continue; 6257 } 6258 6259 passed_sz += hdr->b_size; 6260 if (passed_sz > headroom) { 6261 /* 6262 * Searched too far. 6263 */ 6264 mutex_exit(hash_lock); 6265 ARCSTAT_BUMP(arcstat_l2_write_passed_headroom); 6266 break; 6267 } 6268 6269 if (!l2arc_write_eligible(guid, hdr)) { 6270 mutex_exit(hash_lock); 6271 continue; 6272 } 6273 6274 /* 6275 * Assume that the buffer is not going to be compressed 6276 * and could take more space on disk because of a larger 6277 * disk block size. 6278 */ 6279 buf_sz = hdr->b_size; 6280 buf_a_sz = vdev_psize_to_asize(dev->l2ad_vdev, buf_sz); 6281 6282 if ((write_asize + buf_a_sz) > target_sz) { 6283 full = B_TRUE; 6284 mutex_exit(hash_lock); 6285 ARCSTAT_BUMP(arcstat_l2_write_full); 6286 break; 6287 } 6288 6289 if (pio == NULL) { 6290 /* 6291 * Insert a dummy header on the buflist so 6292 * l2arc_write_done() can find where the 6293 * write buffers begin without searching. 6294 */ 6295 mutex_enter(&dev->l2ad_mtx); 6296 list_insert_head(&dev->l2ad_buflist, head); 6297 mutex_exit(&dev->l2ad_mtx); 6298 6299 cb = kmem_alloc( 6300 sizeof (l2arc_write_callback_t), KM_SLEEP); 6301 cb->l2wcb_dev = dev; 6302 cb->l2wcb_head = head; 6303 pio = zio_root(spa, l2arc_write_done, cb, 6304 ZIO_FLAG_CANFAIL); 6305 ARCSTAT_BUMP(arcstat_l2_write_pios); 6306 } 6307 6308 /* 6309 * Create and add a new L2ARC header. 6310 */ 6311 hdr->b_l2hdr.b_dev = dev; 6312 hdr->b_flags |= ARC_FLAG_L2_WRITING; 6313 /* 6314 * Temporarily stash the data buffer in b_tmp_cdata. 6315 * The subsequent write step will pick it up from 6316 * there. This is because can't access b_l1hdr.b_buf 6317 * without holding the hash_lock, which we in turn 6318 * can't access without holding the ARC list locks 6319 * (which we want to avoid during compression/writing). 6320 */ 6321 HDR_SET_COMPRESS(hdr, ZIO_COMPRESS_OFF); 6322 hdr->b_l2hdr.b_asize = hdr->b_size; 6323 hdr->b_l1hdr.b_tmp_cdata = hdr->b_l1hdr.b_buf->b_data; 6324 6325 /* 6326 * Explicitly set the b_daddr field to a known 6327 * value which means "invalid address". This 6328 * enables us to differentiate which stage of 6329 * l2arc_write_buffers() the particular header 6330 * is in (e.g. this loop, or the one below). 6331 * ARC_FLAG_L2_WRITING is not enough to make 6332 * this distinction, and we need to know in 6333 * order to do proper l2arc vdev accounting in 6334 * arc_release() and arc_hdr_destroy(). 6335 * 6336 * Note, we can't use a new flag to distinguish 6337 * the two stages because we don't hold the 6338 * header's hash_lock below, in the second stage 6339 * of this function. Thus, we can't simply 6340 * change the b_flags field to denote that the 6341 * IO has been sent. We can change the b_daddr 6342 * field of the L2 portion, though, since we'll 6343 * be holding the l2ad_mtx; which is why we're 6344 * using it to denote the header's state change. 6345 */ 6346 hdr->b_l2hdr.b_daddr = L2ARC_ADDR_UNSET; 6347 hdr->b_flags |= ARC_FLAG_HAS_L2HDR; 6348 6349 mutex_enter(&dev->l2ad_mtx); 6350 list_insert_head(&dev->l2ad_buflist, hdr); 6351 mutex_exit(&dev->l2ad_mtx); 6352 6353 /* 6354 * Compute and store the buffer cksum before 6355 * writing. On debug the cksum is verified first. 6356 */ 6357 arc_cksum_verify(hdr->b_l1hdr.b_buf); 6358 arc_cksum_compute(hdr->b_l1hdr.b_buf, B_TRUE); 6359 6360 mutex_exit(hash_lock); 6361 6362 write_sz += buf_sz; 6363 write_asize += buf_a_sz; 6364 } 6365 6366 multilist_sublist_unlock(mls); 6367 6368 if (full == B_TRUE) 6369 break; 6370 } 6371 6372 /* No buffers selected for writing? */ 6373 if (pio == NULL) { 6374 ASSERT0(write_sz); 6375 ASSERT(!HDR_HAS_L1HDR(head)); 6376 kmem_cache_free(hdr_l2only_cache, head); 6377 return (0); 6378 } 6379 6380 mutex_enter(&dev->l2ad_mtx); 6381 6382 /* 6383 * Note that elsewhere in this file arcstat_l2_asize 6384 * and the used space on l2ad_vdev are updated using b_asize, 6385 * which is not necessarily rounded up to the device block size. 6386 * Too keep accounting consistent we do the same here as well: 6387 * stats_size accumulates the sum of b_asize of the written buffers, 6388 * while write_asize accumulates the sum of b_asize rounded up 6389 * to the device block size. 6390 * The latter sum is used only to validate the corectness of the code. 6391 */ 6392 uint64_t stats_size = 0; 6393 write_asize = 0; 6394 6395 /* 6396 * Now start writing the buffers. We're starting at the write head 6397 * and work backwards, retracing the course of the buffer selector 6398 * loop above. 6399 */ 6400 for (hdr = list_prev(&dev->l2ad_buflist, head); hdr; 6401 hdr = list_prev(&dev->l2ad_buflist, hdr)) { 6402 uint64_t buf_sz; 6403 6404 /* 6405 * We rely on the L1 portion of the header below, so 6406 * it's invalid for this header to have been evicted out 6407 * of the ghost cache, prior to being written out. The 6408 * ARC_FLAG_L2_WRITING bit ensures this won't happen. 6409 */ 6410 ASSERT(HDR_HAS_L1HDR(hdr)); 6411 6412 /* 6413 * We shouldn't need to lock the buffer here, since we flagged 6414 * it as ARC_FLAG_L2_WRITING in the previous step, but we must 6415 * take care to only access its L2 cache parameters. In 6416 * particular, hdr->l1hdr.b_buf may be invalid by now due to 6417 * ARC eviction. 6418 */ 6419 hdr->b_l2hdr.b_daddr = dev->l2ad_hand; 6420 6421 if ((HDR_L2COMPRESS(hdr)) && 6422 hdr->b_l2hdr.b_asize >= buf_compress_minsz) { 6423 if (l2arc_compress_buf(hdr)) { 6424 /* 6425 * If compression succeeded, enable headroom 6426 * boost on the next scan cycle. 6427 */ 6428 *headroom_boost = B_TRUE; 6429 } 6430 } 6431 6432 /* 6433 * Pick up the buffer data we had previously stashed away 6434 * (and now potentially also compressed). 6435 */ 6436 buf_data = hdr->b_l1hdr.b_tmp_cdata; 6437 buf_sz = hdr->b_l2hdr.b_asize; 6438 6439 /* 6440 * If the data has not been compressed, then clear b_tmp_cdata 6441 * to make sure that it points only to a temporary compression 6442 * buffer. 6443 */ 6444 if (!L2ARC_IS_VALID_COMPRESS(HDR_GET_COMPRESS(hdr))) 6445 hdr->b_l1hdr.b_tmp_cdata = NULL; 6446 6447 /* 6448 * We need to do this regardless if buf_sz is zero or 6449 * not, otherwise, when this l2hdr is evicted we'll 6450 * remove a reference that was never added. 6451 */ 6452 (void) refcount_add_many(&dev->l2ad_alloc, buf_sz, hdr); 6453 6454 /* Compression may have squashed the buffer to zero length. */ 6455 if (buf_sz != 0) { 6456 uint64_t buf_a_sz; 6457 6458 wzio = zio_write_phys(pio, dev->l2ad_vdev, 6459 dev->l2ad_hand, buf_sz, buf_data, ZIO_CHECKSUM_OFF, 6460 NULL, NULL, ZIO_PRIORITY_ASYNC_WRITE, 6461 ZIO_FLAG_CANFAIL, B_FALSE); 6462 6463 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev, 6464 zio_t *, wzio); 6465 (void) zio_nowait(wzio); 6466 6467 stats_size += buf_sz; 6468 6469 /* 6470 * Keep the clock hand suitably device-aligned. 6471 */ 6472 buf_a_sz = vdev_psize_to_asize(dev->l2ad_vdev, buf_sz); 6473 write_asize += buf_a_sz; 6474 dev->l2ad_hand += buf_a_sz; 6475 } 6476 } 6477 6478 mutex_exit(&dev->l2ad_mtx); 6479 6480 ASSERT3U(write_asize, <=, target_sz); 6481 ARCSTAT_BUMP(arcstat_l2_writes_sent); 6482 ARCSTAT_INCR(arcstat_l2_write_bytes, write_asize); 6483 ARCSTAT_INCR(arcstat_l2_size, write_sz); 6484 ARCSTAT_INCR(arcstat_l2_asize, stats_size); 6485 vdev_space_update(dev->l2ad_vdev, stats_size, 0, 0); 6486 6487 /* 6488 * Bump device hand to the device start if it is approaching the end. 6489 * l2arc_evict() will already have evicted ahead for this case. 6490 */ 6491 if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) { 6492 dev->l2ad_hand = dev->l2ad_start; 6493 dev->l2ad_first = B_FALSE; 6494 } 6495 6496 dev->l2ad_writing = B_TRUE; 6497 (void) zio_wait(pio); 6498 dev->l2ad_writing = B_FALSE; 6499 6500 return (write_asize); 6501} 6502 6503/* 6504 * Compresses an L2ARC buffer. 6505 * The data to be compressed must be prefilled in l1hdr.b_tmp_cdata and its 6506 * size in l2hdr->b_asize. This routine tries to compress the data and 6507 * depending on the compression result there are three possible outcomes: 6508 * *) The buffer was incompressible. The original l2hdr contents were left 6509 * untouched and are ready for writing to an L2 device. 6510 * *) The buffer was all-zeros, so there is no need to write it to an L2 6511 * device. To indicate this situation b_tmp_cdata is NULL'ed, b_asize is 6512 * set to zero and b_compress is set to ZIO_COMPRESS_EMPTY. 6513 * *) Compression succeeded and b_tmp_cdata was replaced with a temporary 6514 * data buffer which holds the compressed data to be written, and b_asize 6515 * tells us how much data there is. b_compress is set to the appropriate 6516 * compression algorithm. Once writing is done, invoke 6517 * l2arc_release_cdata_buf on this l2hdr to free this temporary buffer. 6518 * 6519 * Returns B_TRUE if compression succeeded, or B_FALSE if it didn't (the 6520 * buffer was incompressible). 6521 */ 6522static boolean_t 6523l2arc_compress_buf(arc_buf_hdr_t *hdr) 6524{ 6525 void *cdata; 6526 size_t csize, len, rounded; 6527 ASSERT(HDR_HAS_L2HDR(hdr)); 6528 l2arc_buf_hdr_t *l2hdr = &hdr->b_l2hdr; 6529 6530 ASSERT(HDR_HAS_L1HDR(hdr)); 6531 ASSERT(HDR_GET_COMPRESS(hdr) == ZIO_COMPRESS_OFF); 6532 ASSERT(hdr->b_l1hdr.b_tmp_cdata != NULL); 6533 6534 len = l2hdr->b_asize; 6535 cdata = zio_data_buf_alloc(len); 6536 ASSERT3P(cdata, !=, NULL); 6537 csize = zio_compress_data(ZIO_COMPRESS_LZ4, hdr->b_l1hdr.b_tmp_cdata, 6538 cdata, l2hdr->b_asize); 6539 6540 if (csize == 0) { 6541 /* zero block, indicate that there's nothing to write */ 6542 zio_data_buf_free(cdata, len); 6543 HDR_SET_COMPRESS(hdr, ZIO_COMPRESS_EMPTY); 6544 l2hdr->b_asize = 0; 6545 hdr->b_l1hdr.b_tmp_cdata = NULL; 6546 ARCSTAT_BUMP(arcstat_l2_compress_zeros); 6547 return (B_TRUE); 6548 } 6549 6550 rounded = P2ROUNDUP(csize, 6551 (size_t)1 << l2hdr->b_dev->l2ad_vdev->vdev_ashift); 6552 if (rounded < len) { 6553 /* 6554 * Compression succeeded, we'll keep the cdata around for 6555 * writing and release it afterwards. 6556 */ 6557 if (rounded > csize) { 6558 bzero((char *)cdata + csize, rounded - csize); 6559 csize = rounded; 6560 } 6561 HDR_SET_COMPRESS(hdr, ZIO_COMPRESS_LZ4); 6562 l2hdr->b_asize = csize; 6563 hdr->b_l1hdr.b_tmp_cdata = cdata; 6564 ARCSTAT_BUMP(arcstat_l2_compress_successes); 6565 return (B_TRUE); 6566 } else { 6567 /* 6568 * Compression failed, release the compressed buffer. 6569 * l2hdr will be left unmodified. 6570 */ 6571 zio_data_buf_free(cdata, len); 6572 ARCSTAT_BUMP(arcstat_l2_compress_failures); 6573 return (B_FALSE); 6574 } 6575} 6576 6577/* 6578 * Decompresses a zio read back from an l2arc device. On success, the 6579 * underlying zio's io_data buffer is overwritten by the uncompressed 6580 * version. On decompression error (corrupt compressed stream), the 6581 * zio->io_error value is set to signal an I/O error. 6582 * 6583 * Please note that the compressed data stream is not checksummed, so 6584 * if the underlying device is experiencing data corruption, we may feed 6585 * corrupt data to the decompressor, so the decompressor needs to be 6586 * able to handle this situation (LZ4 does). 6587 */ 6588static void 6589l2arc_decompress_zio(zio_t *zio, arc_buf_hdr_t *hdr, enum zio_compress c) 6590{ 6591 ASSERT(L2ARC_IS_VALID_COMPRESS(c)); 6592 6593 if (zio->io_error != 0) { 6594 /* 6595 * An io error has occured, just restore the original io 6596 * size in preparation for a main pool read. 6597 */ 6598 zio->io_orig_size = zio->io_size = hdr->b_size; 6599 return; 6600 } 6601 6602 if (c == ZIO_COMPRESS_EMPTY) { 6603 /* 6604 * An empty buffer results in a null zio, which means we 6605 * need to fill its io_data after we're done restoring the 6606 * buffer's contents. 6607 */ 6608 ASSERT(hdr->b_l1hdr.b_buf != NULL); 6609 bzero(hdr->b_l1hdr.b_buf->b_data, hdr->b_size); 6610 zio->io_data = zio->io_orig_data = hdr->b_l1hdr.b_buf->b_data; 6611 } else { 6612 ASSERT(zio->io_data != NULL); 6613 /* 6614 * We copy the compressed data from the start of the arc buffer 6615 * (the zio_read will have pulled in only what we need, the 6616 * rest is garbage which we will overwrite at decompression) 6617 * and then decompress back to the ARC data buffer. This way we 6618 * can minimize copying by simply decompressing back over the 6619 * original compressed data (rather than decompressing to an 6620 * aux buffer and then copying back the uncompressed buffer, 6621 * which is likely to be much larger). 6622 */ 6623 uint64_t csize; 6624 void *cdata; 6625 6626 csize = zio->io_size; 6627 cdata = zio_data_buf_alloc(csize); 6628 bcopy(zio->io_data, cdata, csize); 6629 if (zio_decompress_data(c, cdata, zio->io_data, csize, 6630 hdr->b_size) != 0) 6631 zio->io_error = EIO; 6632 zio_data_buf_free(cdata, csize); 6633 } 6634 6635 /* Restore the expected uncompressed IO size. */ 6636 zio->io_orig_size = zio->io_size = hdr->b_size; 6637} 6638 6639/* 6640 * Releases the temporary b_tmp_cdata buffer in an l2arc header structure. 6641 * This buffer serves as a temporary holder of compressed data while 6642 * the buffer entry is being written to an l2arc device. Once that is 6643 * done, we can dispose of it. 6644 */ 6645static void 6646l2arc_release_cdata_buf(arc_buf_hdr_t *hdr) 6647{ 6648 enum zio_compress comp = HDR_GET_COMPRESS(hdr); 6649 6650 ASSERT(HDR_HAS_L1HDR(hdr)); 6651 ASSERT(comp == ZIO_COMPRESS_OFF || L2ARC_IS_VALID_COMPRESS(comp)); 6652 6653 if (comp == ZIO_COMPRESS_OFF) { 6654 /* 6655 * In this case, b_tmp_cdata points to the same buffer 6656 * as the arc_buf_t's b_data field. We don't want to 6657 * free it, since the arc_buf_t will handle that. 6658 */ 6659 hdr->b_l1hdr.b_tmp_cdata = NULL; 6660 } else if (comp == ZIO_COMPRESS_EMPTY) { 6661 /* 6662 * In this case, b_tmp_cdata was compressed to an empty 6663 * buffer, thus there's nothing to free and b_tmp_cdata 6664 * should have been set to NULL in l2arc_write_buffers(). 6665 */ 6666 ASSERT3P(hdr->b_l1hdr.b_tmp_cdata, ==, NULL); 6667 } else { 6668 /* 6669 * If the data was compressed, then we've allocated a 6670 * temporary buffer for it, so now we need to release it. 6671 */ 6672 ASSERT(hdr->b_l1hdr.b_tmp_cdata != NULL); 6673 zio_data_buf_free(hdr->b_l1hdr.b_tmp_cdata, 6674 hdr->b_size); 6675 hdr->b_l1hdr.b_tmp_cdata = NULL; 6676 } 6677} 6678 6679/* 6680 * This thread feeds the L2ARC at regular intervals. This is the beating 6681 * heart of the L2ARC. 6682 */ 6683static void 6684l2arc_feed_thread(void *dummy __unused) 6685{ 6686 callb_cpr_t cpr; 6687 l2arc_dev_t *dev; 6688 spa_t *spa; 6689 uint64_t size, wrote; 6690 clock_t begin, next = ddi_get_lbolt(); 6691 boolean_t headroom_boost = B_FALSE; 6692 6693 CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG); 6694 6695 mutex_enter(&l2arc_feed_thr_lock); 6696 6697 while (l2arc_thread_exit == 0) { 6698 CALLB_CPR_SAFE_BEGIN(&cpr); 6699 (void) cv_timedwait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock, 6700 next - ddi_get_lbolt()); 6701 CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock); 6702 next = ddi_get_lbolt() + hz; 6703 6704 /* 6705 * Quick check for L2ARC devices. 6706 */ 6707 mutex_enter(&l2arc_dev_mtx); 6708 if (l2arc_ndev == 0) { 6709 mutex_exit(&l2arc_dev_mtx); 6710 continue; 6711 } 6712 mutex_exit(&l2arc_dev_mtx); 6713 begin = ddi_get_lbolt(); 6714 6715 /* 6716 * This selects the next l2arc device to write to, and in 6717 * doing so the next spa to feed from: dev->l2ad_spa. This 6718 * will return NULL if there are now no l2arc devices or if 6719 * they are all faulted. 6720 * 6721 * If a device is returned, its spa's config lock is also 6722 * held to prevent device removal. l2arc_dev_get_next() 6723 * will grab and release l2arc_dev_mtx. 6724 */ 6725 if ((dev = l2arc_dev_get_next()) == NULL) 6726 continue; 6727 6728 spa = dev->l2ad_spa; 6729 ASSERT(spa != NULL); 6730 6731 /* 6732 * If the pool is read-only then force the feed thread to 6733 * sleep a little longer. 6734 */ 6735 if (!spa_writeable(spa)) { 6736 next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz; 6737 spa_config_exit(spa, SCL_L2ARC, dev); 6738 continue; 6739 } 6740 6741 /* 6742 * Avoid contributing to memory pressure. 6743 */ 6744 if (arc_reclaim_needed()) { 6745 ARCSTAT_BUMP(arcstat_l2_abort_lowmem); 6746 spa_config_exit(spa, SCL_L2ARC, dev); 6747 continue; 6748 } 6749 6750 ARCSTAT_BUMP(arcstat_l2_feeds); 6751 6752 size = l2arc_write_size(); 6753 6754 /* 6755 * Evict L2ARC buffers that will be overwritten. 6756 */ 6757 l2arc_evict(dev, size, B_FALSE); 6758 6759 /* 6760 * Write ARC buffers. 6761 */ 6762 wrote = l2arc_write_buffers(spa, dev, size, &headroom_boost); 6763 6764 /* 6765 * Calculate interval between writes. 6766 */ 6767 next = l2arc_write_interval(begin, size, wrote); 6768 spa_config_exit(spa, SCL_L2ARC, dev); 6769 } 6770 6771 l2arc_thread_exit = 0; 6772 cv_broadcast(&l2arc_feed_thr_cv); 6773 CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */ 6774 thread_exit(); 6775} 6776 6777boolean_t 6778l2arc_vdev_present(vdev_t *vd) 6779{ 6780 l2arc_dev_t *dev; 6781 6782 mutex_enter(&l2arc_dev_mtx); 6783 for (dev = list_head(l2arc_dev_list); dev != NULL; 6784 dev = list_next(l2arc_dev_list, dev)) { 6785 if (dev->l2ad_vdev == vd) 6786 break; 6787 } 6788 mutex_exit(&l2arc_dev_mtx); 6789 6790 return (dev != NULL); 6791} 6792 6793/* 6794 * Add a vdev for use by the L2ARC. By this point the spa has already 6795 * validated the vdev and opened it. 6796 */ 6797void 6798l2arc_add_vdev(spa_t *spa, vdev_t *vd) 6799{ 6800 l2arc_dev_t *adddev; 6801 6802 ASSERT(!l2arc_vdev_present(vd)); 6803 6804 vdev_ashift_optimize(vd); 6805 6806 /* 6807 * Create a new l2arc device entry. 6808 */ 6809 adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP); 6810 adddev->l2ad_spa = spa; 6811 adddev->l2ad_vdev = vd; 6812 adddev->l2ad_start = VDEV_LABEL_START_SIZE; 6813 adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd); 6814 adddev->l2ad_hand = adddev->l2ad_start; 6815 adddev->l2ad_first = B_TRUE; 6816 adddev->l2ad_writing = B_FALSE; 6817 6818 mutex_init(&adddev->l2ad_mtx, NULL, MUTEX_DEFAULT, NULL); 6819 /* 6820 * This is a list of all ARC buffers that are still valid on the 6821 * device. 6822 */ 6823 list_create(&adddev->l2ad_buflist, sizeof (arc_buf_hdr_t), 6824 offsetof(arc_buf_hdr_t, b_l2hdr.b_l2node)); 6825 6826 vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand); 6827 refcount_create(&adddev->l2ad_alloc); 6828 6829 /* 6830 * Add device to global list 6831 */ 6832 mutex_enter(&l2arc_dev_mtx); 6833 list_insert_head(l2arc_dev_list, adddev); 6834 atomic_inc_64(&l2arc_ndev); 6835 mutex_exit(&l2arc_dev_mtx); 6836} 6837 6838/* 6839 * Remove a vdev from the L2ARC. 6840 */ 6841void 6842l2arc_remove_vdev(vdev_t *vd) 6843{ 6844 l2arc_dev_t *dev, *nextdev, *remdev = NULL; 6845 6846 /* 6847 * Find the device by vdev 6848 */ 6849 mutex_enter(&l2arc_dev_mtx); 6850 for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) { 6851 nextdev = list_next(l2arc_dev_list, dev); 6852 if (vd == dev->l2ad_vdev) { 6853 remdev = dev; 6854 break; 6855 } 6856 } 6857 ASSERT(remdev != NULL); 6858 6859 /* 6860 * Remove device from global list 6861 */ 6862 list_remove(l2arc_dev_list, remdev); 6863 l2arc_dev_last = NULL; /* may have been invalidated */ 6864 atomic_dec_64(&l2arc_ndev); 6865 mutex_exit(&l2arc_dev_mtx); 6866 6867 /* 6868 * Clear all buflists and ARC references. L2ARC device flush. 6869 */ 6870 l2arc_evict(remdev, 0, B_TRUE); 6871 list_destroy(&remdev->l2ad_buflist); 6872 mutex_destroy(&remdev->l2ad_mtx); 6873 refcount_destroy(&remdev->l2ad_alloc); 6874 kmem_free(remdev, sizeof (l2arc_dev_t)); 6875} 6876 6877void 6878l2arc_init(void) 6879{ 6880 l2arc_thread_exit = 0; 6881 l2arc_ndev = 0; 6882 l2arc_writes_sent = 0; 6883 l2arc_writes_done = 0; 6884 6885 mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL); 6886 cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL); 6887 mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL); 6888 mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL); 6889 6890 l2arc_dev_list = &L2ARC_dev_list; 6891 l2arc_free_on_write = &L2ARC_free_on_write; 6892 list_create(l2arc_dev_list, sizeof (l2arc_dev_t), 6893 offsetof(l2arc_dev_t, l2ad_node)); 6894 list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t), 6895 offsetof(l2arc_data_free_t, l2df_list_node)); 6896} 6897 6898void 6899l2arc_fini(void) 6900{ 6901 /* 6902 * This is called from dmu_fini(), which is called from spa_fini(); 6903 * Because of this, we can assume that all l2arc devices have 6904 * already been removed when the pools themselves were removed. 6905 */ 6906 6907 l2arc_do_free_on_write(); 6908 6909 mutex_destroy(&l2arc_feed_thr_lock); 6910 cv_destroy(&l2arc_feed_thr_cv); 6911 mutex_destroy(&l2arc_dev_mtx); 6912 mutex_destroy(&l2arc_free_on_write_mtx); 6913 6914 list_destroy(l2arc_dev_list); 6915 list_destroy(l2arc_free_on_write); 6916} 6917 6918void 6919l2arc_start(void) 6920{ 6921 if (!(spa_mode_global & FWRITE)) 6922 return; 6923 6924 (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0, 6925 TS_RUN, minclsyspri); 6926} 6927 6928void 6929l2arc_stop(void) 6930{ 6931 if (!(spa_mode_global & FWRITE)) 6932 return; 6933 6934 mutex_enter(&l2arc_feed_thr_lock); 6935 cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */ 6936 l2arc_thread_exit = 1; 6937 while (l2arc_thread_exit != 0) 6938 cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock); 6939 mutex_exit(&l2arc_feed_thr_lock); 6940} 6941