arc.c revision 297077
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 mutexes, rather they rely on the 86 * hash table mutexes for the bulk of their protection (i.e. most 87 * fields in the arc_buf_hdr_t are protected by these mutexes). 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 boolean_t 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 kstat_named_t arcstat_sync_wait_for_async; 589 kstat_named_t arcstat_demand_hit_predictive_prefetch; 590} arc_stats_t; 591 592static arc_stats_t arc_stats = { 593 { "hits", KSTAT_DATA_UINT64 }, 594 { "misses", KSTAT_DATA_UINT64 }, 595 { "demand_data_hits", KSTAT_DATA_UINT64 }, 596 { "demand_data_misses", KSTAT_DATA_UINT64 }, 597 { "demand_metadata_hits", KSTAT_DATA_UINT64 }, 598 { "demand_metadata_misses", KSTAT_DATA_UINT64 }, 599 { "prefetch_data_hits", KSTAT_DATA_UINT64 }, 600 { "prefetch_data_misses", KSTAT_DATA_UINT64 }, 601 { "prefetch_metadata_hits", KSTAT_DATA_UINT64 }, 602 { "prefetch_metadata_misses", KSTAT_DATA_UINT64 }, 603 { "mru_hits", KSTAT_DATA_UINT64 }, 604 { "mru_ghost_hits", KSTAT_DATA_UINT64 }, 605 { "mfu_hits", KSTAT_DATA_UINT64 }, 606 { "mfu_ghost_hits", KSTAT_DATA_UINT64 }, 607 { "allocated", KSTAT_DATA_UINT64 }, 608 { "deleted", KSTAT_DATA_UINT64 }, 609 { "mutex_miss", KSTAT_DATA_UINT64 }, 610 { "evict_skip", KSTAT_DATA_UINT64 }, 611 { "evict_not_enough", KSTAT_DATA_UINT64 }, 612 { "evict_l2_cached", KSTAT_DATA_UINT64 }, 613 { "evict_l2_eligible", KSTAT_DATA_UINT64 }, 614 { "evict_l2_ineligible", KSTAT_DATA_UINT64 }, 615 { "evict_l2_skip", KSTAT_DATA_UINT64 }, 616 { "hash_elements", KSTAT_DATA_UINT64 }, 617 { "hash_elements_max", KSTAT_DATA_UINT64 }, 618 { "hash_collisions", KSTAT_DATA_UINT64 }, 619 { "hash_chains", KSTAT_DATA_UINT64 }, 620 { "hash_chain_max", KSTAT_DATA_UINT64 }, 621 { "p", KSTAT_DATA_UINT64 }, 622 { "c", KSTAT_DATA_UINT64 }, 623 { "c_min", KSTAT_DATA_UINT64 }, 624 { "c_max", KSTAT_DATA_UINT64 }, 625 { "size", KSTAT_DATA_UINT64 }, 626 { "hdr_size", KSTAT_DATA_UINT64 }, 627 { "data_size", KSTAT_DATA_UINT64 }, 628 { "metadata_size", KSTAT_DATA_UINT64 }, 629 { "other_size", KSTAT_DATA_UINT64 }, 630 { "anon_size", KSTAT_DATA_UINT64 }, 631 { "anon_evictable_data", KSTAT_DATA_UINT64 }, 632 { "anon_evictable_metadata", KSTAT_DATA_UINT64 }, 633 { "mru_size", KSTAT_DATA_UINT64 }, 634 { "mru_evictable_data", KSTAT_DATA_UINT64 }, 635 { "mru_evictable_metadata", KSTAT_DATA_UINT64 }, 636 { "mru_ghost_size", KSTAT_DATA_UINT64 }, 637 { "mru_ghost_evictable_data", KSTAT_DATA_UINT64 }, 638 { "mru_ghost_evictable_metadata", KSTAT_DATA_UINT64 }, 639 { "mfu_size", KSTAT_DATA_UINT64 }, 640 { "mfu_evictable_data", KSTAT_DATA_UINT64 }, 641 { "mfu_evictable_metadata", KSTAT_DATA_UINT64 }, 642 { "mfu_ghost_size", KSTAT_DATA_UINT64 }, 643 { "mfu_ghost_evictable_data", KSTAT_DATA_UINT64 }, 644 { "mfu_ghost_evictable_metadata", KSTAT_DATA_UINT64 }, 645 { "l2_hits", KSTAT_DATA_UINT64 }, 646 { "l2_misses", KSTAT_DATA_UINT64 }, 647 { "l2_feeds", KSTAT_DATA_UINT64 }, 648 { "l2_rw_clash", KSTAT_DATA_UINT64 }, 649 { "l2_read_bytes", KSTAT_DATA_UINT64 }, 650 { "l2_write_bytes", KSTAT_DATA_UINT64 }, 651 { "l2_writes_sent", KSTAT_DATA_UINT64 }, 652 { "l2_writes_done", KSTAT_DATA_UINT64 }, 653 { "l2_writes_error", KSTAT_DATA_UINT64 }, 654 { "l2_writes_lock_retry", KSTAT_DATA_UINT64 }, 655 { "l2_evict_lock_retry", KSTAT_DATA_UINT64 }, 656 { "l2_evict_reading", KSTAT_DATA_UINT64 }, 657 { "l2_evict_l1cached", KSTAT_DATA_UINT64 }, 658 { "l2_free_on_write", KSTAT_DATA_UINT64 }, 659 { "l2_cdata_free_on_write", KSTAT_DATA_UINT64 }, 660 { "l2_abort_lowmem", KSTAT_DATA_UINT64 }, 661 { "l2_cksum_bad", KSTAT_DATA_UINT64 }, 662 { "l2_io_error", KSTAT_DATA_UINT64 }, 663 { "l2_size", KSTAT_DATA_UINT64 }, 664 { "l2_asize", KSTAT_DATA_UINT64 }, 665 { "l2_hdr_size", KSTAT_DATA_UINT64 }, 666 { "l2_compress_successes", KSTAT_DATA_UINT64 }, 667 { "l2_compress_zeros", KSTAT_DATA_UINT64 }, 668 { "l2_compress_failures", KSTAT_DATA_UINT64 }, 669 { "l2_write_trylock_fail", KSTAT_DATA_UINT64 }, 670 { "l2_write_passed_headroom", KSTAT_DATA_UINT64 }, 671 { "l2_write_spa_mismatch", KSTAT_DATA_UINT64 }, 672 { "l2_write_in_l2", KSTAT_DATA_UINT64 }, 673 { "l2_write_io_in_progress", KSTAT_DATA_UINT64 }, 674 { "l2_write_not_cacheable", KSTAT_DATA_UINT64 }, 675 { "l2_write_full", KSTAT_DATA_UINT64 }, 676 { "l2_write_buffer_iter", KSTAT_DATA_UINT64 }, 677 { "l2_write_pios", KSTAT_DATA_UINT64 }, 678 { "l2_write_buffer_bytes_scanned", KSTAT_DATA_UINT64 }, 679 { "l2_write_buffer_list_iter", KSTAT_DATA_UINT64 }, 680 { "l2_write_buffer_list_null_iter", KSTAT_DATA_UINT64 }, 681 { "memory_throttle_count", KSTAT_DATA_UINT64 }, 682 { "duplicate_buffers", KSTAT_DATA_UINT64 }, 683 { "duplicate_buffers_size", KSTAT_DATA_UINT64 }, 684 { "duplicate_reads", KSTAT_DATA_UINT64 }, 685 { "arc_meta_used", KSTAT_DATA_UINT64 }, 686 { "arc_meta_limit", KSTAT_DATA_UINT64 }, 687 { "arc_meta_max", KSTAT_DATA_UINT64 }, 688 { "arc_meta_min", KSTAT_DATA_UINT64 }, 689 { "sync_wait_for_async", KSTAT_DATA_UINT64 }, 690 { "demand_hit_predictive_prefetch", KSTAT_DATA_UINT64 }, 691}; 692 693#define ARCSTAT(stat) (arc_stats.stat.value.ui64) 694 695#define ARCSTAT_INCR(stat, val) \ 696 atomic_add_64(&arc_stats.stat.value.ui64, (val)) 697 698#define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1) 699#define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1) 700 701#define ARCSTAT_MAX(stat, val) { \ 702 uint64_t m; \ 703 while ((val) > (m = arc_stats.stat.value.ui64) && \ 704 (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \ 705 continue; \ 706} 707 708#define ARCSTAT_MAXSTAT(stat) \ 709 ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64) 710 711/* 712 * We define a macro to allow ARC hits/misses to be easily broken down by 713 * two separate conditions, giving a total of four different subtypes for 714 * each of hits and misses (so eight statistics total). 715 */ 716#define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \ 717 if (cond1) { \ 718 if (cond2) { \ 719 ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \ 720 } else { \ 721 ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \ 722 } \ 723 } else { \ 724 if (cond2) { \ 725 ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \ 726 } else { \ 727 ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\ 728 } \ 729 } 730 731kstat_t *arc_ksp; 732static arc_state_t *arc_anon; 733static arc_state_t *arc_mru; 734static arc_state_t *arc_mru_ghost; 735static arc_state_t *arc_mfu; 736static arc_state_t *arc_mfu_ghost; 737static arc_state_t *arc_l2c_only; 738 739/* 740 * There are several ARC variables that are critical to export as kstats -- 741 * but we don't want to have to grovel around in the kstat whenever we wish to 742 * manipulate them. For these variables, we therefore define them to be in 743 * terms of the statistic variable. This assures that we are not introducing 744 * the possibility of inconsistency by having shadow copies of the variables, 745 * while still allowing the code to be readable. 746 */ 747#define arc_size ARCSTAT(arcstat_size) /* actual total arc size */ 748#define arc_p ARCSTAT(arcstat_p) /* target size of MRU */ 749#define arc_c ARCSTAT(arcstat_c) /* target size of cache */ 750#define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */ 751#define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */ 752#define arc_meta_limit ARCSTAT(arcstat_meta_limit) /* max size for metadata */ 753#define arc_meta_min ARCSTAT(arcstat_meta_min) /* min size for metadata */ 754#define arc_meta_used ARCSTAT(arcstat_meta_used) /* size of metadata */ 755#define arc_meta_max ARCSTAT(arcstat_meta_max) /* max size of metadata */ 756 757#define L2ARC_IS_VALID_COMPRESS(_c_) \ 758 ((_c_) == ZIO_COMPRESS_LZ4 || (_c_) == ZIO_COMPRESS_EMPTY) 759 760static int arc_no_grow; /* Don't try to grow cache size */ 761static uint64_t arc_tempreserve; 762static uint64_t arc_loaned_bytes; 763 764typedef struct arc_callback arc_callback_t; 765 766struct arc_callback { 767 void *acb_private; 768 arc_done_func_t *acb_done; 769 arc_buf_t *acb_buf; 770 zio_t *acb_zio_dummy; 771 arc_callback_t *acb_next; 772}; 773 774typedef struct arc_write_callback arc_write_callback_t; 775 776struct arc_write_callback { 777 void *awcb_private; 778 arc_done_func_t *awcb_ready; 779 arc_done_func_t *awcb_physdone; 780 arc_done_func_t *awcb_done; 781 arc_buf_t *awcb_buf; 782}; 783 784/* 785 * ARC buffers are separated into multiple structs as a memory saving measure: 786 * - Common fields struct, always defined, and embedded within it: 787 * - L2-only fields, always allocated but undefined when not in L2ARC 788 * - L1-only fields, only allocated when in L1ARC 789 * 790 * Buffer in L1 Buffer only in L2 791 * +------------------------+ +------------------------+ 792 * | arc_buf_hdr_t | | arc_buf_hdr_t | 793 * | | | | 794 * | | | | 795 * | | | | 796 * +------------------------+ +------------------------+ 797 * | l2arc_buf_hdr_t | | l2arc_buf_hdr_t | 798 * | (undefined if L1-only) | | | 799 * +------------------------+ +------------------------+ 800 * | l1arc_buf_hdr_t | 801 * | | 802 * | | 803 * | | 804 * | | 805 * +------------------------+ 806 * 807 * Because it's possible for the L2ARC to become extremely large, we can wind 808 * up eating a lot of memory in L2ARC buffer headers, so the size of a header 809 * is minimized by only allocating the fields necessary for an L1-cached buffer 810 * when a header is actually in the L1 cache. The sub-headers (l1arc_buf_hdr and 811 * l2arc_buf_hdr) are embedded rather than allocated separately to save a couple 812 * words in pointers. arc_hdr_realloc() is used to switch a header between 813 * these two allocation states. 814 */ 815typedef struct l1arc_buf_hdr { 816 kmutex_t b_freeze_lock; 817#ifdef ZFS_DEBUG 818 /* 819 * used for debugging wtih kmem_flags - by allocating and freeing 820 * b_thawed when the buffer is thawed, we get a record of the stack 821 * trace that thawed it. 822 */ 823 void *b_thawed; 824#endif 825 826 arc_buf_t *b_buf; 827 uint32_t b_datacnt; 828 /* for waiting on writes to complete */ 829 kcondvar_t b_cv; 830 831 /* protected by arc state mutex */ 832 arc_state_t *b_state; 833 multilist_node_t b_arc_node; 834 835 /* updated atomically */ 836 clock_t b_arc_access; 837 838 /* self protecting */ 839 refcount_t b_refcnt; 840 841 arc_callback_t *b_acb; 842 /* temporary buffer holder for in-flight compressed data */ 843 void *b_tmp_cdata; 844} l1arc_buf_hdr_t; 845 846typedef struct l2arc_dev l2arc_dev_t; 847 848typedef struct l2arc_buf_hdr { 849 /* protected by arc_buf_hdr mutex */ 850 l2arc_dev_t *b_dev; /* L2ARC device */ 851 uint64_t b_daddr; /* disk address, offset byte */ 852 /* real alloc'd buffer size depending on b_compress applied */ 853 int32_t b_asize; 854 uint8_t b_compress; 855 856 list_node_t b_l2node; 857} l2arc_buf_hdr_t; 858 859struct arc_buf_hdr { 860 /* protected by hash lock */ 861 dva_t b_dva; 862 uint64_t b_birth; 863 /* 864 * Even though this checksum is only set/verified when a buffer is in 865 * the L1 cache, it needs to be in the set of common fields because it 866 * must be preserved from the time before a buffer is written out to 867 * L2ARC until after it is read back in. 868 */ 869 zio_cksum_t *b_freeze_cksum; 870 871 arc_buf_hdr_t *b_hash_next; 872 arc_flags_t b_flags; 873 874 /* immutable */ 875 int32_t b_size; 876 uint64_t b_spa; 877 878 /* L2ARC fields. Undefined when not in L2ARC. */ 879 l2arc_buf_hdr_t b_l2hdr; 880 /* L1ARC fields. Undefined when in l2arc_only state */ 881 l1arc_buf_hdr_t b_l1hdr; 882}; 883 884#ifdef _KERNEL 885static int 886sysctl_vfs_zfs_arc_meta_limit(SYSCTL_HANDLER_ARGS) 887{ 888 uint64_t val; 889 int err; 890 891 val = arc_meta_limit; 892 err = sysctl_handle_64(oidp, &val, 0, req); 893 if (err != 0 || req->newptr == NULL) 894 return (err); 895 896 if (val <= 0 || val > arc_c_max) 897 return (EINVAL); 898 899 arc_meta_limit = val; 900 return (0); 901} 902#endif 903 904static arc_buf_t *arc_eviction_list; 905static arc_buf_hdr_t arc_eviction_hdr; 906 907#define GHOST_STATE(state) \ 908 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \ 909 (state) == arc_l2c_only) 910 911#define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_FLAG_IN_HASH_TABLE) 912#define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) 913#define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_FLAG_IO_ERROR) 914#define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_FLAG_PREFETCH) 915#define HDR_FREED_IN_READ(hdr) ((hdr)->b_flags & ARC_FLAG_FREED_IN_READ) 916#define HDR_BUF_AVAILABLE(hdr) ((hdr)->b_flags & ARC_FLAG_BUF_AVAILABLE) 917 918#define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_FLAG_L2CACHE) 919#define HDR_L2COMPRESS(hdr) ((hdr)->b_flags & ARC_FLAG_L2COMPRESS) 920#define HDR_L2_READING(hdr) \ 921 (((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) && \ 922 ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR)) 923#define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITING) 924#define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_FLAG_L2_EVICTED) 925#define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITE_HEAD) 926 927#define HDR_ISTYPE_METADATA(hdr) \ 928 ((hdr)->b_flags & ARC_FLAG_BUFC_METADATA) 929#define HDR_ISTYPE_DATA(hdr) (!HDR_ISTYPE_METADATA(hdr)) 930 931#define HDR_HAS_L1HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L1HDR) 932#define HDR_HAS_L2HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR) 933 934/* 935 * Other sizes 936 */ 937 938#define HDR_FULL_SIZE ((int64_t)sizeof (arc_buf_hdr_t)) 939#define HDR_L2ONLY_SIZE ((int64_t)offsetof(arc_buf_hdr_t, b_l1hdr)) 940 941/* 942 * Hash table routines 943 */ 944 945#define HT_LOCK_PAD CACHE_LINE_SIZE 946 947struct ht_lock { 948 kmutex_t ht_lock; 949#ifdef _KERNEL 950 unsigned char pad[(HT_LOCK_PAD - sizeof (kmutex_t))]; 951#endif 952}; 953 954#define BUF_LOCKS 256 955typedef struct buf_hash_table { 956 uint64_t ht_mask; 957 arc_buf_hdr_t **ht_table; 958 struct ht_lock ht_locks[BUF_LOCKS] __aligned(CACHE_LINE_SIZE); 959} buf_hash_table_t; 960 961static buf_hash_table_t buf_hash_table; 962 963#define BUF_HASH_INDEX(spa, dva, birth) \ 964 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask) 965#define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)]) 966#define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock)) 967#define HDR_LOCK(hdr) \ 968 (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth))) 969 970uint64_t zfs_crc64_table[256]; 971 972/* 973 * Level 2 ARC 974 */ 975 976#define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */ 977#define L2ARC_HEADROOM 2 /* num of writes */ 978/* 979 * If we discover during ARC scan any buffers to be compressed, we boost 980 * our headroom for the next scanning cycle by this percentage multiple. 981 */ 982#define L2ARC_HEADROOM_BOOST 200 983#define L2ARC_FEED_SECS 1 /* caching interval secs */ 984#define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */ 985 986/* 987 * Used to distinguish headers that are being process by 988 * l2arc_write_buffers(), but have yet to be assigned to a l2arc disk 989 * address. This can happen when the header is added to the l2arc's list 990 * of buffers to write in the first stage of l2arc_write_buffers(), but 991 * has not yet been written out which happens in the second stage of 992 * l2arc_write_buffers(). 993 */ 994#define L2ARC_ADDR_UNSET ((uint64_t)(-1)) 995 996#define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent) 997#define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done) 998 999/* L2ARC Performance Tunables */ 1000uint64_t l2arc_write_max = L2ARC_WRITE_SIZE; /* default max write size */ 1001uint64_t l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra write during warmup */ 1002uint64_t l2arc_headroom = L2ARC_HEADROOM; /* number of dev writes */ 1003uint64_t l2arc_headroom_boost = L2ARC_HEADROOM_BOOST; 1004uint64_t l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */ 1005uint64_t l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval milliseconds */ 1006boolean_t l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */ 1007boolean_t l2arc_feed_again = B_TRUE; /* turbo warmup */ 1008boolean_t l2arc_norw = B_TRUE; /* no reads during writes */ 1009 1010SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_max, CTLFLAG_RW, 1011 &l2arc_write_max, 0, "max write size"); 1012SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_boost, CTLFLAG_RW, 1013 &l2arc_write_boost, 0, "extra write during warmup"); 1014SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_headroom, CTLFLAG_RW, 1015 &l2arc_headroom, 0, "number of dev writes"); 1016SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_secs, CTLFLAG_RW, 1017 &l2arc_feed_secs, 0, "interval seconds"); 1018SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_min_ms, CTLFLAG_RW, 1019 &l2arc_feed_min_ms, 0, "min interval milliseconds"); 1020 1021SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_noprefetch, CTLFLAG_RW, 1022 &l2arc_noprefetch, 0, "don't cache prefetch bufs"); 1023SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_feed_again, CTLFLAG_RW, 1024 &l2arc_feed_again, 0, "turbo warmup"); 1025SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_norw, CTLFLAG_RW, 1026 &l2arc_norw, 0, "no reads during writes"); 1027 1028SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_size, CTLFLAG_RD, 1029 &ARC_anon.arcs_size.rc_count, 0, "size of anonymous state"); 1030SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_metadata_lsize, CTLFLAG_RD, 1031 &ARC_anon.arcs_lsize[ARC_BUFC_METADATA], 0, "size of anonymous state"); 1032SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_data_lsize, CTLFLAG_RD, 1033 &ARC_anon.arcs_lsize[ARC_BUFC_DATA], 0, "size of anonymous state"); 1034 1035SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_size, CTLFLAG_RD, 1036 &ARC_mru.arcs_size.rc_count, 0, "size of mru state"); 1037SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_metadata_lsize, CTLFLAG_RD, 1038 &ARC_mru.arcs_lsize[ARC_BUFC_METADATA], 0, "size of metadata in mru state"); 1039SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_data_lsize, CTLFLAG_RD, 1040 &ARC_mru.arcs_lsize[ARC_BUFC_DATA], 0, "size of data in mru state"); 1041 1042SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_size, CTLFLAG_RD, 1043 &ARC_mru_ghost.arcs_size.rc_count, 0, "size of mru ghost state"); 1044SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_metadata_lsize, CTLFLAG_RD, 1045 &ARC_mru_ghost.arcs_lsize[ARC_BUFC_METADATA], 0, 1046 "size of metadata in mru ghost state"); 1047SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_data_lsize, CTLFLAG_RD, 1048 &ARC_mru_ghost.arcs_lsize[ARC_BUFC_DATA], 0, 1049 "size of data in mru ghost state"); 1050 1051SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_size, CTLFLAG_RD, 1052 &ARC_mfu.arcs_size.rc_count, 0, "size of mfu state"); 1053SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_metadata_lsize, CTLFLAG_RD, 1054 &ARC_mfu.arcs_lsize[ARC_BUFC_METADATA], 0, "size of metadata in mfu state"); 1055SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_data_lsize, CTLFLAG_RD, 1056 &ARC_mfu.arcs_lsize[ARC_BUFC_DATA], 0, "size of data in mfu state"); 1057 1058SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_size, CTLFLAG_RD, 1059 &ARC_mfu_ghost.arcs_size.rc_count, 0, "size of mfu ghost state"); 1060SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_metadata_lsize, CTLFLAG_RD, 1061 &ARC_mfu_ghost.arcs_lsize[ARC_BUFC_METADATA], 0, 1062 "size of metadata in mfu ghost state"); 1063SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_data_lsize, CTLFLAG_RD, 1064 &ARC_mfu_ghost.arcs_lsize[ARC_BUFC_DATA], 0, 1065 "size of data in mfu ghost state"); 1066 1067SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2c_only_size, CTLFLAG_RD, 1068 &ARC_l2c_only.arcs_size.rc_count, 0, "size of mru state"); 1069 1070/* 1071 * L2ARC Internals 1072 */ 1073struct l2arc_dev { 1074 vdev_t *l2ad_vdev; /* vdev */ 1075 spa_t *l2ad_spa; /* spa */ 1076 uint64_t l2ad_hand; /* next write location */ 1077 uint64_t l2ad_start; /* first addr on device */ 1078 uint64_t l2ad_end; /* last addr on device */ 1079 boolean_t l2ad_first; /* first sweep through */ 1080 boolean_t l2ad_writing; /* currently writing */ 1081 kmutex_t l2ad_mtx; /* lock for buffer list */ 1082 list_t l2ad_buflist; /* buffer list */ 1083 list_node_t l2ad_node; /* device list node */ 1084 refcount_t l2ad_alloc; /* allocated bytes */ 1085}; 1086 1087static list_t L2ARC_dev_list; /* device list */ 1088static list_t *l2arc_dev_list; /* device list pointer */ 1089static kmutex_t l2arc_dev_mtx; /* device list mutex */ 1090static l2arc_dev_t *l2arc_dev_last; /* last device used */ 1091static list_t L2ARC_free_on_write; /* free after write buf list */ 1092static list_t *l2arc_free_on_write; /* free after write list ptr */ 1093static kmutex_t l2arc_free_on_write_mtx; /* mutex for list */ 1094static uint64_t l2arc_ndev; /* number of devices */ 1095 1096typedef struct l2arc_read_callback { 1097 arc_buf_t *l2rcb_buf; /* read buffer */ 1098 spa_t *l2rcb_spa; /* spa */ 1099 blkptr_t l2rcb_bp; /* original blkptr */ 1100 zbookmark_phys_t l2rcb_zb; /* original bookmark */ 1101 int l2rcb_flags; /* original flags */ 1102 enum zio_compress l2rcb_compress; /* applied compress */ 1103} l2arc_read_callback_t; 1104 1105typedef struct l2arc_write_callback { 1106 l2arc_dev_t *l2wcb_dev; /* device info */ 1107 arc_buf_hdr_t *l2wcb_head; /* head of write buflist */ 1108} l2arc_write_callback_t; 1109 1110typedef struct l2arc_data_free { 1111 /* protected by l2arc_free_on_write_mtx */ 1112 void *l2df_data; 1113 size_t l2df_size; 1114 void (*l2df_func)(void *, size_t); 1115 list_node_t l2df_list_node; 1116} l2arc_data_free_t; 1117 1118static kmutex_t l2arc_feed_thr_lock; 1119static kcondvar_t l2arc_feed_thr_cv; 1120static uint8_t l2arc_thread_exit; 1121 1122static void arc_get_data_buf(arc_buf_t *); 1123static void arc_access(arc_buf_hdr_t *, kmutex_t *); 1124static boolean_t arc_is_overflowing(); 1125static void arc_buf_watch(arc_buf_t *); 1126 1127static arc_buf_contents_t arc_buf_type(arc_buf_hdr_t *); 1128static uint32_t arc_bufc_to_flags(arc_buf_contents_t); 1129 1130static boolean_t l2arc_write_eligible(uint64_t, arc_buf_hdr_t *); 1131static void l2arc_read_done(zio_t *); 1132 1133static boolean_t l2arc_compress_buf(arc_buf_hdr_t *); 1134static void l2arc_decompress_zio(zio_t *, arc_buf_hdr_t *, enum zio_compress); 1135static void l2arc_release_cdata_buf(arc_buf_hdr_t *); 1136 1137static void 1138l2arc_trim(const arc_buf_hdr_t *hdr) 1139{ 1140 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev; 1141 1142 ASSERT(HDR_HAS_L2HDR(hdr)); 1143 ASSERT(MUTEX_HELD(&dev->l2ad_mtx)); 1144 1145 if (hdr->b_l2hdr.b_daddr == L2ARC_ADDR_UNSET) 1146 return; 1147 if (hdr->b_l2hdr.b_asize != 0) { 1148 trim_map_free(dev->l2ad_vdev, hdr->b_l2hdr.b_daddr, 1149 hdr->b_l2hdr.b_asize, 0); 1150 } else { 1151 ASSERT3U(hdr->b_l2hdr.b_compress, ==, ZIO_COMPRESS_EMPTY); 1152 } 1153} 1154 1155static uint64_t 1156buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth) 1157{ 1158 uint8_t *vdva = (uint8_t *)dva; 1159 uint64_t crc = -1ULL; 1160 int i; 1161 1162 ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY); 1163 1164 for (i = 0; i < sizeof (dva_t); i++) 1165 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ vdva[i]) & 0xFF]; 1166 1167 crc ^= (spa>>8) ^ birth; 1168 1169 return (crc); 1170} 1171 1172#define BUF_EMPTY(buf) \ 1173 ((buf)->b_dva.dva_word[0] == 0 && \ 1174 (buf)->b_dva.dva_word[1] == 0) 1175 1176#define BUF_EQUAL(spa, dva, birth, buf) \ 1177 ((buf)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \ 1178 ((buf)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \ 1179 ((buf)->b_birth == birth) && ((buf)->b_spa == spa) 1180 1181static void 1182buf_discard_identity(arc_buf_hdr_t *hdr) 1183{ 1184 hdr->b_dva.dva_word[0] = 0; 1185 hdr->b_dva.dva_word[1] = 0; 1186 hdr->b_birth = 0; 1187} 1188 1189static arc_buf_hdr_t * 1190buf_hash_find(uint64_t spa, const blkptr_t *bp, kmutex_t **lockp) 1191{ 1192 const dva_t *dva = BP_IDENTITY(bp); 1193 uint64_t birth = BP_PHYSICAL_BIRTH(bp); 1194 uint64_t idx = BUF_HASH_INDEX(spa, dva, birth); 1195 kmutex_t *hash_lock = BUF_HASH_LOCK(idx); 1196 arc_buf_hdr_t *hdr; 1197 1198 mutex_enter(hash_lock); 1199 for (hdr = buf_hash_table.ht_table[idx]; hdr != NULL; 1200 hdr = hdr->b_hash_next) { 1201 if (BUF_EQUAL(spa, dva, birth, hdr)) { 1202 *lockp = hash_lock; 1203 return (hdr); 1204 } 1205 } 1206 mutex_exit(hash_lock); 1207 *lockp = NULL; 1208 return (NULL); 1209} 1210 1211/* 1212 * Insert an entry into the hash table. If there is already an element 1213 * equal to elem in the hash table, then the already existing element 1214 * will be returned and the new element will not be inserted. 1215 * Otherwise returns NULL. 1216 * If lockp == NULL, the caller is assumed to already hold the hash lock. 1217 */ 1218static arc_buf_hdr_t * 1219buf_hash_insert(arc_buf_hdr_t *hdr, kmutex_t **lockp) 1220{ 1221 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth); 1222 kmutex_t *hash_lock = BUF_HASH_LOCK(idx); 1223 arc_buf_hdr_t *fhdr; 1224 uint32_t i; 1225 1226 ASSERT(!DVA_IS_EMPTY(&hdr->b_dva)); 1227 ASSERT(hdr->b_birth != 0); 1228 ASSERT(!HDR_IN_HASH_TABLE(hdr)); 1229 1230 if (lockp != NULL) { 1231 *lockp = hash_lock; 1232 mutex_enter(hash_lock); 1233 } else { 1234 ASSERT(MUTEX_HELD(hash_lock)); 1235 } 1236 1237 for (fhdr = buf_hash_table.ht_table[idx], i = 0; fhdr != NULL; 1238 fhdr = fhdr->b_hash_next, i++) { 1239 if (BUF_EQUAL(hdr->b_spa, &hdr->b_dva, hdr->b_birth, fhdr)) 1240 return (fhdr); 1241 } 1242 1243 hdr->b_hash_next = buf_hash_table.ht_table[idx]; 1244 buf_hash_table.ht_table[idx] = hdr; 1245 hdr->b_flags |= ARC_FLAG_IN_HASH_TABLE; 1246 1247 /* collect some hash table performance data */ 1248 if (i > 0) { 1249 ARCSTAT_BUMP(arcstat_hash_collisions); 1250 if (i == 1) 1251 ARCSTAT_BUMP(arcstat_hash_chains); 1252 1253 ARCSTAT_MAX(arcstat_hash_chain_max, i); 1254 } 1255 1256 ARCSTAT_BUMP(arcstat_hash_elements); 1257 ARCSTAT_MAXSTAT(arcstat_hash_elements); 1258 1259 return (NULL); 1260} 1261 1262static void 1263buf_hash_remove(arc_buf_hdr_t *hdr) 1264{ 1265 arc_buf_hdr_t *fhdr, **hdrp; 1266 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth); 1267 1268 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx))); 1269 ASSERT(HDR_IN_HASH_TABLE(hdr)); 1270 1271 hdrp = &buf_hash_table.ht_table[idx]; 1272 while ((fhdr = *hdrp) != hdr) { 1273 ASSERT(fhdr != NULL); 1274 hdrp = &fhdr->b_hash_next; 1275 } 1276 *hdrp = hdr->b_hash_next; 1277 hdr->b_hash_next = NULL; 1278 hdr->b_flags &= ~ARC_FLAG_IN_HASH_TABLE; 1279 1280 /* collect some hash table performance data */ 1281 ARCSTAT_BUMPDOWN(arcstat_hash_elements); 1282 1283 if (buf_hash_table.ht_table[idx] && 1284 buf_hash_table.ht_table[idx]->b_hash_next == NULL) 1285 ARCSTAT_BUMPDOWN(arcstat_hash_chains); 1286} 1287 1288/* 1289 * Global data structures and functions for the buf kmem cache. 1290 */ 1291static kmem_cache_t *hdr_full_cache; 1292static kmem_cache_t *hdr_l2only_cache; 1293static kmem_cache_t *buf_cache; 1294 1295static void 1296buf_fini(void) 1297{ 1298 int i; 1299 1300 kmem_free(buf_hash_table.ht_table, 1301 (buf_hash_table.ht_mask + 1) * sizeof (void *)); 1302 for (i = 0; i < BUF_LOCKS; i++) 1303 mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock); 1304 kmem_cache_destroy(hdr_full_cache); 1305 kmem_cache_destroy(hdr_l2only_cache); 1306 kmem_cache_destroy(buf_cache); 1307} 1308 1309/* 1310 * Constructor callback - called when the cache is empty 1311 * and a new buf is requested. 1312 */ 1313/* ARGSUSED */ 1314static int 1315hdr_full_cons(void *vbuf, void *unused, int kmflag) 1316{ 1317 arc_buf_hdr_t *hdr = vbuf; 1318 1319 bzero(hdr, HDR_FULL_SIZE); 1320 cv_init(&hdr->b_l1hdr.b_cv, NULL, CV_DEFAULT, NULL); 1321 refcount_create(&hdr->b_l1hdr.b_refcnt); 1322 mutex_init(&hdr->b_l1hdr.b_freeze_lock, NULL, MUTEX_DEFAULT, NULL); 1323 multilist_link_init(&hdr->b_l1hdr.b_arc_node); 1324 arc_space_consume(HDR_FULL_SIZE, ARC_SPACE_HDRS); 1325 1326 return (0); 1327} 1328 1329/* ARGSUSED */ 1330static int 1331hdr_l2only_cons(void *vbuf, void *unused, int kmflag) 1332{ 1333 arc_buf_hdr_t *hdr = vbuf; 1334 1335 bzero(hdr, HDR_L2ONLY_SIZE); 1336 arc_space_consume(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS); 1337 1338 return (0); 1339} 1340 1341/* ARGSUSED */ 1342static int 1343buf_cons(void *vbuf, void *unused, int kmflag) 1344{ 1345 arc_buf_t *buf = vbuf; 1346 1347 bzero(buf, sizeof (arc_buf_t)); 1348 mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL); 1349 arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS); 1350 1351 return (0); 1352} 1353 1354/* 1355 * Destructor callback - called when a cached buf is 1356 * no longer required. 1357 */ 1358/* ARGSUSED */ 1359static void 1360hdr_full_dest(void *vbuf, void *unused) 1361{ 1362 arc_buf_hdr_t *hdr = vbuf; 1363 1364 ASSERT(BUF_EMPTY(hdr)); 1365 cv_destroy(&hdr->b_l1hdr.b_cv); 1366 refcount_destroy(&hdr->b_l1hdr.b_refcnt); 1367 mutex_destroy(&hdr->b_l1hdr.b_freeze_lock); 1368 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 1369 arc_space_return(HDR_FULL_SIZE, ARC_SPACE_HDRS); 1370} 1371 1372/* ARGSUSED */ 1373static void 1374hdr_l2only_dest(void *vbuf, void *unused) 1375{ 1376 arc_buf_hdr_t *hdr = vbuf; 1377 1378 ASSERT(BUF_EMPTY(hdr)); 1379 arc_space_return(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS); 1380} 1381 1382/* ARGSUSED */ 1383static void 1384buf_dest(void *vbuf, void *unused) 1385{ 1386 arc_buf_t *buf = vbuf; 1387 1388 mutex_destroy(&buf->b_evict_lock); 1389 arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS); 1390} 1391 1392/* 1393 * Reclaim callback -- invoked when memory is low. 1394 */ 1395/* ARGSUSED */ 1396static void 1397hdr_recl(void *unused) 1398{ 1399 dprintf("hdr_recl called\n"); 1400 /* 1401 * umem calls the reclaim func when we destroy the buf cache, 1402 * which is after we do arc_fini(). 1403 */ 1404 if (!arc_dead) 1405 cv_signal(&arc_reclaim_thread_cv); 1406} 1407 1408static void 1409buf_init(void) 1410{ 1411 uint64_t *ct; 1412 uint64_t hsize = 1ULL << 12; 1413 int i, j; 1414 1415 /* 1416 * The hash table is big enough to fill all of physical memory 1417 * with an average block size of zfs_arc_average_blocksize (default 8K). 1418 * By default, the table will take up 1419 * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers). 1420 */ 1421 while (hsize * zfs_arc_average_blocksize < (uint64_t)physmem * PAGESIZE) 1422 hsize <<= 1; 1423retry: 1424 buf_hash_table.ht_mask = hsize - 1; 1425 buf_hash_table.ht_table = 1426 kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP); 1427 if (buf_hash_table.ht_table == NULL) { 1428 ASSERT(hsize > (1ULL << 8)); 1429 hsize >>= 1; 1430 goto retry; 1431 } 1432 1433 hdr_full_cache = kmem_cache_create("arc_buf_hdr_t_full", HDR_FULL_SIZE, 1434 0, hdr_full_cons, hdr_full_dest, hdr_recl, NULL, NULL, 0); 1435 hdr_l2only_cache = kmem_cache_create("arc_buf_hdr_t_l2only", 1436 HDR_L2ONLY_SIZE, 0, hdr_l2only_cons, hdr_l2only_dest, hdr_recl, 1437 NULL, NULL, 0); 1438 buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t), 1439 0, buf_cons, buf_dest, NULL, NULL, NULL, 0); 1440 1441 for (i = 0; i < 256; i++) 1442 for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--) 1443 *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY); 1444 1445 for (i = 0; i < BUF_LOCKS; i++) { 1446 mutex_init(&buf_hash_table.ht_locks[i].ht_lock, 1447 NULL, MUTEX_DEFAULT, NULL); 1448 } 1449} 1450 1451/* 1452 * Transition between the two allocation states for the arc_buf_hdr struct. 1453 * The arc_buf_hdr struct can be allocated with (hdr_full_cache) or without 1454 * (hdr_l2only_cache) the fields necessary for the L1 cache - the smaller 1455 * version is used when a cache buffer is only in the L2ARC in order to reduce 1456 * memory usage. 1457 */ 1458static arc_buf_hdr_t * 1459arc_hdr_realloc(arc_buf_hdr_t *hdr, kmem_cache_t *old, kmem_cache_t *new) 1460{ 1461 ASSERT(HDR_HAS_L2HDR(hdr)); 1462 1463 arc_buf_hdr_t *nhdr; 1464 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev; 1465 1466 ASSERT((old == hdr_full_cache && new == hdr_l2only_cache) || 1467 (old == hdr_l2only_cache && new == hdr_full_cache)); 1468 1469 nhdr = kmem_cache_alloc(new, KM_PUSHPAGE); 1470 1471 ASSERT(MUTEX_HELD(HDR_LOCK(hdr))); 1472 buf_hash_remove(hdr); 1473 1474 bcopy(hdr, nhdr, HDR_L2ONLY_SIZE); 1475 1476 if (new == hdr_full_cache) { 1477 nhdr->b_flags |= ARC_FLAG_HAS_L1HDR; 1478 /* 1479 * arc_access and arc_change_state need to be aware that a 1480 * header has just come out of L2ARC, so we set its state to 1481 * l2c_only even though it's about to change. 1482 */ 1483 nhdr->b_l1hdr.b_state = arc_l2c_only; 1484 1485 /* Verify previous threads set to NULL before freeing */ 1486 ASSERT3P(nhdr->b_l1hdr.b_tmp_cdata, ==, NULL); 1487 } else { 1488 ASSERT(hdr->b_l1hdr.b_buf == NULL); 1489 ASSERT0(hdr->b_l1hdr.b_datacnt); 1490 1491 /* 1492 * If we've reached here, We must have been called from 1493 * arc_evict_hdr(), as such we should have already been 1494 * removed from any ghost list we were previously on 1495 * (which protects us from racing with arc_evict_state), 1496 * thus no locking is needed during this check. 1497 */ 1498 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 1499 1500 /* 1501 * A buffer must not be moved into the arc_l2c_only 1502 * state if it's not finished being written out to the 1503 * l2arc device. Otherwise, the b_l1hdr.b_tmp_cdata field 1504 * might try to be accessed, even though it was removed. 1505 */ 1506 VERIFY(!HDR_L2_WRITING(hdr)); 1507 VERIFY3P(hdr->b_l1hdr.b_tmp_cdata, ==, NULL); 1508 1509#ifdef ZFS_DEBUG 1510 if (hdr->b_l1hdr.b_thawed != NULL) { 1511 kmem_free(hdr->b_l1hdr.b_thawed, 1); 1512 hdr->b_l1hdr.b_thawed = NULL; 1513 } 1514#endif 1515 1516 nhdr->b_flags &= ~ARC_FLAG_HAS_L1HDR; 1517 } 1518 /* 1519 * The header has been reallocated so we need to re-insert it into any 1520 * lists it was on. 1521 */ 1522 (void) buf_hash_insert(nhdr, NULL); 1523 1524 ASSERT(list_link_active(&hdr->b_l2hdr.b_l2node)); 1525 1526 mutex_enter(&dev->l2ad_mtx); 1527 1528 /* 1529 * We must place the realloc'ed header back into the list at 1530 * the same spot. Otherwise, if it's placed earlier in the list, 1531 * l2arc_write_buffers() could find it during the function's 1532 * write phase, and try to write it out to the l2arc. 1533 */ 1534 list_insert_after(&dev->l2ad_buflist, hdr, nhdr); 1535 list_remove(&dev->l2ad_buflist, hdr); 1536 1537 mutex_exit(&dev->l2ad_mtx); 1538 1539 /* 1540 * Since we're using the pointer address as the tag when 1541 * incrementing and decrementing the l2ad_alloc refcount, we 1542 * must remove the old pointer (that we're about to destroy) and 1543 * add the new pointer to the refcount. Otherwise we'd remove 1544 * the wrong pointer address when calling arc_hdr_destroy() later. 1545 */ 1546 1547 (void) refcount_remove_many(&dev->l2ad_alloc, 1548 hdr->b_l2hdr.b_asize, hdr); 1549 1550 (void) refcount_add_many(&dev->l2ad_alloc, 1551 nhdr->b_l2hdr.b_asize, nhdr); 1552 1553 buf_discard_identity(hdr); 1554 hdr->b_freeze_cksum = NULL; 1555 kmem_cache_free(old, hdr); 1556 1557 return (nhdr); 1558} 1559 1560 1561#define ARC_MINTIME (hz>>4) /* 62 ms */ 1562 1563static void 1564arc_cksum_verify(arc_buf_t *buf) 1565{ 1566 zio_cksum_t zc; 1567 1568 if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 1569 return; 1570 1571 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1572 if (buf->b_hdr->b_freeze_cksum == NULL || HDR_IO_ERROR(buf->b_hdr)) { 1573 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1574 return; 1575 } 1576 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, NULL, &zc); 1577 if (!ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc)) 1578 panic("buffer modified while frozen!"); 1579 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1580} 1581 1582static int 1583arc_cksum_equal(arc_buf_t *buf) 1584{ 1585 zio_cksum_t zc; 1586 int equal; 1587 1588 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1589 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, NULL, &zc); 1590 equal = ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc); 1591 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1592 1593 return (equal); 1594} 1595 1596static void 1597arc_cksum_compute(arc_buf_t *buf, boolean_t force) 1598{ 1599 if (!force && !(zfs_flags & ZFS_DEBUG_MODIFY)) 1600 return; 1601 1602 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1603 if (buf->b_hdr->b_freeze_cksum != NULL) { 1604 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1605 return; 1606 } 1607 buf->b_hdr->b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t), KM_SLEEP); 1608 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, 1609 NULL, buf->b_hdr->b_freeze_cksum); 1610 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1611#ifdef illumos 1612 arc_buf_watch(buf); 1613#endif 1614} 1615 1616#ifdef illumos 1617#ifndef _KERNEL 1618typedef struct procctl { 1619 long cmd; 1620 prwatch_t prwatch; 1621} procctl_t; 1622#endif 1623 1624/* ARGSUSED */ 1625static void 1626arc_buf_unwatch(arc_buf_t *buf) 1627{ 1628#ifndef _KERNEL 1629 if (arc_watch) { 1630 int result; 1631 procctl_t ctl; 1632 ctl.cmd = PCWATCH; 1633 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data; 1634 ctl.prwatch.pr_size = 0; 1635 ctl.prwatch.pr_wflags = 0; 1636 result = write(arc_procfd, &ctl, sizeof (ctl)); 1637 ASSERT3U(result, ==, sizeof (ctl)); 1638 } 1639#endif 1640} 1641 1642/* ARGSUSED */ 1643static void 1644arc_buf_watch(arc_buf_t *buf) 1645{ 1646#ifndef _KERNEL 1647 if (arc_watch) { 1648 int result; 1649 procctl_t ctl; 1650 ctl.cmd = PCWATCH; 1651 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data; 1652 ctl.prwatch.pr_size = buf->b_hdr->b_size; 1653 ctl.prwatch.pr_wflags = WA_WRITE; 1654 result = write(arc_procfd, &ctl, sizeof (ctl)); 1655 ASSERT3U(result, ==, sizeof (ctl)); 1656 } 1657#endif 1658} 1659#endif /* illumos */ 1660 1661static arc_buf_contents_t 1662arc_buf_type(arc_buf_hdr_t *hdr) 1663{ 1664 if (HDR_ISTYPE_METADATA(hdr)) { 1665 return (ARC_BUFC_METADATA); 1666 } else { 1667 return (ARC_BUFC_DATA); 1668 } 1669} 1670 1671static uint32_t 1672arc_bufc_to_flags(arc_buf_contents_t type) 1673{ 1674 switch (type) { 1675 case ARC_BUFC_DATA: 1676 /* metadata field is 0 if buffer contains normal data */ 1677 return (0); 1678 case ARC_BUFC_METADATA: 1679 return (ARC_FLAG_BUFC_METADATA); 1680 default: 1681 break; 1682 } 1683 panic("undefined ARC buffer type!"); 1684 return ((uint32_t)-1); 1685} 1686 1687void 1688arc_buf_thaw(arc_buf_t *buf) 1689{ 1690 if (zfs_flags & ZFS_DEBUG_MODIFY) { 1691 if (buf->b_hdr->b_l1hdr.b_state != arc_anon) 1692 panic("modifying non-anon buffer!"); 1693 if (HDR_IO_IN_PROGRESS(buf->b_hdr)) 1694 panic("modifying buffer while i/o in progress!"); 1695 arc_cksum_verify(buf); 1696 } 1697 1698 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1699 if (buf->b_hdr->b_freeze_cksum != NULL) { 1700 kmem_free(buf->b_hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 1701 buf->b_hdr->b_freeze_cksum = NULL; 1702 } 1703 1704#ifdef ZFS_DEBUG 1705 if (zfs_flags & ZFS_DEBUG_MODIFY) { 1706 if (buf->b_hdr->b_l1hdr.b_thawed != NULL) 1707 kmem_free(buf->b_hdr->b_l1hdr.b_thawed, 1); 1708 buf->b_hdr->b_l1hdr.b_thawed = kmem_alloc(1, KM_SLEEP); 1709 } 1710#endif 1711 1712 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1713 1714#ifdef illumos 1715 arc_buf_unwatch(buf); 1716#endif 1717} 1718 1719void 1720arc_buf_freeze(arc_buf_t *buf) 1721{ 1722 kmutex_t *hash_lock; 1723 1724 if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 1725 return; 1726 1727 hash_lock = HDR_LOCK(buf->b_hdr); 1728 mutex_enter(hash_lock); 1729 1730 ASSERT(buf->b_hdr->b_freeze_cksum != NULL || 1731 buf->b_hdr->b_l1hdr.b_state == arc_anon); 1732 arc_cksum_compute(buf, B_FALSE); 1733 mutex_exit(hash_lock); 1734 1735} 1736 1737static void 1738add_reference(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, void *tag) 1739{ 1740 ASSERT(HDR_HAS_L1HDR(hdr)); 1741 ASSERT(MUTEX_HELD(hash_lock)); 1742 arc_state_t *state = hdr->b_l1hdr.b_state; 1743 1744 if ((refcount_add(&hdr->b_l1hdr.b_refcnt, tag) == 1) && 1745 (state != arc_anon)) { 1746 /* We don't use the L2-only state list. */ 1747 if (state != arc_l2c_only) { 1748 arc_buf_contents_t type = arc_buf_type(hdr); 1749 uint64_t delta = hdr->b_size * hdr->b_l1hdr.b_datacnt; 1750 multilist_t *list = &state->arcs_list[type]; 1751 uint64_t *size = &state->arcs_lsize[type]; 1752 1753 multilist_remove(list, hdr); 1754 1755 if (GHOST_STATE(state)) { 1756 ASSERT0(hdr->b_l1hdr.b_datacnt); 1757 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 1758 delta = hdr->b_size; 1759 } 1760 ASSERT(delta > 0); 1761 ASSERT3U(*size, >=, delta); 1762 atomic_add_64(size, -delta); 1763 } 1764 /* remove the prefetch flag if we get a reference */ 1765 hdr->b_flags &= ~ARC_FLAG_PREFETCH; 1766 } 1767} 1768 1769static int 1770remove_reference(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, void *tag) 1771{ 1772 int cnt; 1773 arc_state_t *state = hdr->b_l1hdr.b_state; 1774 1775 ASSERT(HDR_HAS_L1HDR(hdr)); 1776 ASSERT(state == arc_anon || MUTEX_HELD(hash_lock)); 1777 ASSERT(!GHOST_STATE(state)); 1778 1779 /* 1780 * arc_l2c_only counts as a ghost state so we don't need to explicitly 1781 * check to prevent usage of the arc_l2c_only list. 1782 */ 1783 if (((cnt = refcount_remove(&hdr->b_l1hdr.b_refcnt, tag)) == 0) && 1784 (state != arc_anon)) { 1785 arc_buf_contents_t type = arc_buf_type(hdr); 1786 multilist_t *list = &state->arcs_list[type]; 1787 uint64_t *size = &state->arcs_lsize[type]; 1788 1789 multilist_insert(list, hdr); 1790 1791 ASSERT(hdr->b_l1hdr.b_datacnt > 0); 1792 atomic_add_64(size, hdr->b_size * 1793 hdr->b_l1hdr.b_datacnt); 1794 } 1795 return (cnt); 1796} 1797 1798/* 1799 * Move the supplied buffer to the indicated state. The hash lock 1800 * for the buffer must be held by the caller. 1801 */ 1802static void 1803arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *hdr, 1804 kmutex_t *hash_lock) 1805{ 1806 arc_state_t *old_state; 1807 int64_t refcnt; 1808 uint32_t datacnt; 1809 uint64_t from_delta, to_delta; 1810 arc_buf_contents_t buftype = arc_buf_type(hdr); 1811 1812 /* 1813 * We almost always have an L1 hdr here, since we call arc_hdr_realloc() 1814 * in arc_read() when bringing a buffer out of the L2ARC. However, the 1815 * L1 hdr doesn't always exist when we change state to arc_anon before 1816 * destroying a header, in which case reallocating to add the L1 hdr is 1817 * pointless. 1818 */ 1819 if (HDR_HAS_L1HDR(hdr)) { 1820 old_state = hdr->b_l1hdr.b_state; 1821 refcnt = refcount_count(&hdr->b_l1hdr.b_refcnt); 1822 datacnt = hdr->b_l1hdr.b_datacnt; 1823 } else { 1824 old_state = arc_l2c_only; 1825 refcnt = 0; 1826 datacnt = 0; 1827 } 1828 1829 ASSERT(MUTEX_HELD(hash_lock)); 1830 ASSERT3P(new_state, !=, old_state); 1831 ASSERT(refcnt == 0 || datacnt > 0); 1832 ASSERT(!GHOST_STATE(new_state) || datacnt == 0); 1833 ASSERT(old_state != arc_anon || datacnt <= 1); 1834 1835 from_delta = to_delta = datacnt * hdr->b_size; 1836 1837 /* 1838 * If this buffer is evictable, transfer it from the 1839 * old state list to the new state list. 1840 */ 1841 if (refcnt == 0) { 1842 if (old_state != arc_anon && old_state != arc_l2c_only) { 1843 uint64_t *size = &old_state->arcs_lsize[buftype]; 1844 1845 ASSERT(HDR_HAS_L1HDR(hdr)); 1846 multilist_remove(&old_state->arcs_list[buftype], hdr); 1847 1848 /* 1849 * If prefetching out of the ghost cache, 1850 * we will have a non-zero datacnt. 1851 */ 1852 if (GHOST_STATE(old_state) && datacnt == 0) { 1853 /* ghost elements have a ghost size */ 1854 ASSERT(hdr->b_l1hdr.b_buf == NULL); 1855 from_delta = hdr->b_size; 1856 } 1857 ASSERT3U(*size, >=, from_delta); 1858 atomic_add_64(size, -from_delta); 1859 } 1860 if (new_state != arc_anon && new_state != arc_l2c_only) { 1861 uint64_t *size = &new_state->arcs_lsize[buftype]; 1862 1863 /* 1864 * An L1 header always exists here, since if we're 1865 * moving to some L1-cached state (i.e. not l2c_only or 1866 * anonymous), we realloc the header to add an L1hdr 1867 * beforehand. 1868 */ 1869 ASSERT(HDR_HAS_L1HDR(hdr)); 1870 multilist_insert(&new_state->arcs_list[buftype], hdr); 1871 1872 /* ghost elements have a ghost size */ 1873 if (GHOST_STATE(new_state)) { 1874 ASSERT0(datacnt); 1875 ASSERT(hdr->b_l1hdr.b_buf == NULL); 1876 to_delta = hdr->b_size; 1877 } 1878 atomic_add_64(size, to_delta); 1879 } 1880 } 1881 1882 ASSERT(!BUF_EMPTY(hdr)); 1883 if (new_state == arc_anon && HDR_IN_HASH_TABLE(hdr)) 1884 buf_hash_remove(hdr); 1885 1886 /* adjust state sizes (ignore arc_l2c_only) */ 1887 1888 if (to_delta && new_state != arc_l2c_only) { 1889 ASSERT(HDR_HAS_L1HDR(hdr)); 1890 if (GHOST_STATE(new_state)) { 1891 ASSERT0(datacnt); 1892 1893 /* 1894 * We moving a header to a ghost state, we first 1895 * remove all arc buffers. Thus, we'll have a 1896 * datacnt of zero, and no arc buffer to use for 1897 * the reference. As a result, we use the arc 1898 * header pointer for the reference. 1899 */ 1900 (void) refcount_add_many(&new_state->arcs_size, 1901 hdr->b_size, hdr); 1902 } else { 1903 ASSERT3U(datacnt, !=, 0); 1904 1905 /* 1906 * Each individual buffer holds a unique reference, 1907 * thus we must remove each of these references one 1908 * at a time. 1909 */ 1910 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL; 1911 buf = buf->b_next) { 1912 (void) refcount_add_many(&new_state->arcs_size, 1913 hdr->b_size, buf); 1914 } 1915 } 1916 } 1917 1918 if (from_delta && old_state != arc_l2c_only) { 1919 ASSERT(HDR_HAS_L1HDR(hdr)); 1920 if (GHOST_STATE(old_state)) { 1921 /* 1922 * When moving a header off of a ghost state, 1923 * there's the possibility for datacnt to be 1924 * non-zero. This is because we first add the 1925 * arc buffer to the header prior to changing 1926 * the header's state. Since we used the header 1927 * for the reference when putting the header on 1928 * the ghost state, we must balance that and use 1929 * the header when removing off the ghost state 1930 * (even though datacnt is non zero). 1931 */ 1932 1933 IMPLY(datacnt == 0, new_state == arc_anon || 1934 new_state == arc_l2c_only); 1935 1936 (void) refcount_remove_many(&old_state->arcs_size, 1937 hdr->b_size, hdr); 1938 } else { 1939 ASSERT3P(datacnt, !=, 0); 1940 1941 /* 1942 * Each individual buffer holds a unique reference, 1943 * thus we must remove each of these references one 1944 * at a time. 1945 */ 1946 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL; 1947 buf = buf->b_next) { 1948 (void) refcount_remove_many( 1949 &old_state->arcs_size, hdr->b_size, buf); 1950 } 1951 } 1952 } 1953 1954 if (HDR_HAS_L1HDR(hdr)) 1955 hdr->b_l1hdr.b_state = new_state; 1956 1957 /* 1958 * L2 headers should never be on the L2 state list since they don't 1959 * have L1 headers allocated. 1960 */ 1961 ASSERT(multilist_is_empty(&arc_l2c_only->arcs_list[ARC_BUFC_DATA]) && 1962 multilist_is_empty(&arc_l2c_only->arcs_list[ARC_BUFC_METADATA])); 1963} 1964 1965void 1966arc_space_consume(uint64_t space, arc_space_type_t type) 1967{ 1968 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES); 1969 1970 switch (type) { 1971 case ARC_SPACE_DATA: 1972 ARCSTAT_INCR(arcstat_data_size, space); 1973 break; 1974 case ARC_SPACE_META: 1975 ARCSTAT_INCR(arcstat_metadata_size, space); 1976 break; 1977 case ARC_SPACE_OTHER: 1978 ARCSTAT_INCR(arcstat_other_size, space); 1979 break; 1980 case ARC_SPACE_HDRS: 1981 ARCSTAT_INCR(arcstat_hdr_size, space); 1982 break; 1983 case ARC_SPACE_L2HDRS: 1984 ARCSTAT_INCR(arcstat_l2_hdr_size, space); 1985 break; 1986 } 1987 1988 if (type != ARC_SPACE_DATA) 1989 ARCSTAT_INCR(arcstat_meta_used, space); 1990 1991 atomic_add_64(&arc_size, space); 1992} 1993 1994void 1995arc_space_return(uint64_t space, arc_space_type_t type) 1996{ 1997 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES); 1998 1999 switch (type) { 2000 case ARC_SPACE_DATA: 2001 ARCSTAT_INCR(arcstat_data_size, -space); 2002 break; 2003 case ARC_SPACE_META: 2004 ARCSTAT_INCR(arcstat_metadata_size, -space); 2005 break; 2006 case ARC_SPACE_OTHER: 2007 ARCSTAT_INCR(arcstat_other_size, -space); 2008 break; 2009 case ARC_SPACE_HDRS: 2010 ARCSTAT_INCR(arcstat_hdr_size, -space); 2011 break; 2012 case ARC_SPACE_L2HDRS: 2013 ARCSTAT_INCR(arcstat_l2_hdr_size, -space); 2014 break; 2015 } 2016 2017 if (type != ARC_SPACE_DATA) { 2018 ASSERT(arc_meta_used >= space); 2019 if (arc_meta_max < arc_meta_used) 2020 arc_meta_max = arc_meta_used; 2021 ARCSTAT_INCR(arcstat_meta_used, -space); 2022 } 2023 2024 ASSERT(arc_size >= space); 2025 atomic_add_64(&arc_size, -space); 2026} 2027 2028arc_buf_t * 2029arc_buf_alloc(spa_t *spa, int32_t size, void *tag, arc_buf_contents_t type) 2030{ 2031 arc_buf_hdr_t *hdr; 2032 arc_buf_t *buf; 2033 2034 ASSERT3U(size, >, 0); 2035 hdr = kmem_cache_alloc(hdr_full_cache, KM_PUSHPAGE); 2036 ASSERT(BUF_EMPTY(hdr)); 2037 ASSERT3P(hdr->b_freeze_cksum, ==, NULL); 2038 hdr->b_size = size; 2039 hdr->b_spa = spa_load_guid(spa); 2040 2041 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 2042 buf->b_hdr = hdr; 2043 buf->b_data = NULL; 2044 buf->b_efunc = NULL; 2045 buf->b_private = NULL; 2046 buf->b_next = NULL; 2047 2048 hdr->b_flags = arc_bufc_to_flags(type); 2049 hdr->b_flags |= ARC_FLAG_HAS_L1HDR; 2050 2051 hdr->b_l1hdr.b_buf = buf; 2052 hdr->b_l1hdr.b_state = arc_anon; 2053 hdr->b_l1hdr.b_arc_access = 0; 2054 hdr->b_l1hdr.b_datacnt = 1; 2055 hdr->b_l1hdr.b_tmp_cdata = NULL; 2056 2057 arc_get_data_buf(buf); 2058 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 2059 (void) refcount_add(&hdr->b_l1hdr.b_refcnt, tag); 2060 2061 return (buf); 2062} 2063 2064static char *arc_onloan_tag = "onloan"; 2065 2066/* 2067 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in 2068 * flight data by arc_tempreserve_space() until they are "returned". Loaned 2069 * buffers must be returned to the arc before they can be used by the DMU or 2070 * freed. 2071 */ 2072arc_buf_t * 2073arc_loan_buf(spa_t *spa, int size) 2074{ 2075 arc_buf_t *buf; 2076 2077 buf = arc_buf_alloc(spa, size, arc_onloan_tag, ARC_BUFC_DATA); 2078 2079 atomic_add_64(&arc_loaned_bytes, size); 2080 return (buf); 2081} 2082 2083/* 2084 * Return a loaned arc buffer to the arc. 2085 */ 2086void 2087arc_return_buf(arc_buf_t *buf, void *tag) 2088{ 2089 arc_buf_hdr_t *hdr = buf->b_hdr; 2090 2091 ASSERT(buf->b_data != NULL); 2092 ASSERT(HDR_HAS_L1HDR(hdr)); 2093 (void) refcount_add(&hdr->b_l1hdr.b_refcnt, tag); 2094 (void) refcount_remove(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag); 2095 2096 atomic_add_64(&arc_loaned_bytes, -hdr->b_size); 2097} 2098 2099/* Detach an arc_buf from a dbuf (tag) */ 2100void 2101arc_loan_inuse_buf(arc_buf_t *buf, void *tag) 2102{ 2103 arc_buf_hdr_t *hdr = buf->b_hdr; 2104 2105 ASSERT(buf->b_data != NULL); 2106 ASSERT(HDR_HAS_L1HDR(hdr)); 2107 (void) refcount_add(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag); 2108 (void) refcount_remove(&hdr->b_l1hdr.b_refcnt, tag); 2109 buf->b_efunc = NULL; 2110 buf->b_private = NULL; 2111 2112 atomic_add_64(&arc_loaned_bytes, hdr->b_size); 2113} 2114 2115static arc_buf_t * 2116arc_buf_clone(arc_buf_t *from) 2117{ 2118 arc_buf_t *buf; 2119 arc_buf_hdr_t *hdr = from->b_hdr; 2120 uint64_t size = hdr->b_size; 2121 2122 ASSERT(HDR_HAS_L1HDR(hdr)); 2123 ASSERT(hdr->b_l1hdr.b_state != arc_anon); 2124 2125 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 2126 buf->b_hdr = hdr; 2127 buf->b_data = NULL; 2128 buf->b_efunc = NULL; 2129 buf->b_private = NULL; 2130 buf->b_next = hdr->b_l1hdr.b_buf; 2131 hdr->b_l1hdr.b_buf = buf; 2132 arc_get_data_buf(buf); 2133 bcopy(from->b_data, buf->b_data, size); 2134 2135 /* 2136 * This buffer already exists in the arc so create a duplicate 2137 * copy for the caller. If the buffer is associated with user data 2138 * then track the size and number of duplicates. These stats will be 2139 * updated as duplicate buffers are created and destroyed. 2140 */ 2141 if (HDR_ISTYPE_DATA(hdr)) { 2142 ARCSTAT_BUMP(arcstat_duplicate_buffers); 2143 ARCSTAT_INCR(arcstat_duplicate_buffers_size, size); 2144 } 2145 hdr->b_l1hdr.b_datacnt += 1; 2146 return (buf); 2147} 2148 2149void 2150arc_buf_add_ref(arc_buf_t *buf, void* tag) 2151{ 2152 arc_buf_hdr_t *hdr; 2153 kmutex_t *hash_lock; 2154 2155 /* 2156 * Check to see if this buffer is evicted. Callers 2157 * must verify b_data != NULL to know if the add_ref 2158 * was successful. 2159 */ 2160 mutex_enter(&buf->b_evict_lock); 2161 if (buf->b_data == NULL) { 2162 mutex_exit(&buf->b_evict_lock); 2163 return; 2164 } 2165 hash_lock = HDR_LOCK(buf->b_hdr); 2166 mutex_enter(hash_lock); 2167 hdr = buf->b_hdr; 2168 ASSERT(HDR_HAS_L1HDR(hdr)); 2169 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 2170 mutex_exit(&buf->b_evict_lock); 2171 2172 ASSERT(hdr->b_l1hdr.b_state == arc_mru || 2173 hdr->b_l1hdr.b_state == arc_mfu); 2174 2175 add_reference(hdr, hash_lock, tag); 2176 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr); 2177 arc_access(hdr, hash_lock); 2178 mutex_exit(hash_lock); 2179 ARCSTAT_BUMP(arcstat_hits); 2180 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr), 2181 demand, prefetch, !HDR_ISTYPE_METADATA(hdr), 2182 data, metadata, hits); 2183} 2184 2185static void 2186arc_buf_free_on_write(void *data, size_t size, 2187 void (*free_func)(void *, size_t)) 2188{ 2189 l2arc_data_free_t *df; 2190 2191 df = kmem_alloc(sizeof (*df), KM_SLEEP); 2192 df->l2df_data = data; 2193 df->l2df_size = size; 2194 df->l2df_func = free_func; 2195 mutex_enter(&l2arc_free_on_write_mtx); 2196 list_insert_head(l2arc_free_on_write, df); 2197 mutex_exit(&l2arc_free_on_write_mtx); 2198} 2199 2200/* 2201 * Free the arc data buffer. If it is an l2arc write in progress, 2202 * the buffer is placed on l2arc_free_on_write to be freed later. 2203 */ 2204static void 2205arc_buf_data_free(arc_buf_t *buf, void (*free_func)(void *, size_t)) 2206{ 2207 arc_buf_hdr_t *hdr = buf->b_hdr; 2208 2209 if (HDR_L2_WRITING(hdr)) { 2210 arc_buf_free_on_write(buf->b_data, hdr->b_size, free_func); 2211 ARCSTAT_BUMP(arcstat_l2_free_on_write); 2212 } else { 2213 free_func(buf->b_data, hdr->b_size); 2214 } 2215} 2216 2217static void 2218arc_buf_l2_cdata_free(arc_buf_hdr_t *hdr) 2219{ 2220 ASSERT(HDR_HAS_L2HDR(hdr)); 2221 ASSERT(MUTEX_HELD(&hdr->b_l2hdr.b_dev->l2ad_mtx)); 2222 2223 /* 2224 * The b_tmp_cdata field is linked off of the b_l1hdr, so if 2225 * that doesn't exist, the header is in the arc_l2c_only state, 2226 * and there isn't anything to free (it's already been freed). 2227 */ 2228 if (!HDR_HAS_L1HDR(hdr)) 2229 return; 2230 2231 /* 2232 * The header isn't being written to the l2arc device, thus it 2233 * shouldn't have a b_tmp_cdata to free. 2234 */ 2235 if (!HDR_L2_WRITING(hdr)) { 2236 ASSERT3P(hdr->b_l1hdr.b_tmp_cdata, ==, NULL); 2237 return; 2238 } 2239 2240 /* 2241 * The header does not have compression enabled. This can be due 2242 * to the buffer not being compressible, or because we're 2243 * freeing the buffer before the second phase of 2244 * l2arc_write_buffer() has started (which does the compression 2245 * step). In either case, b_tmp_cdata does not point to a 2246 * separately compressed buffer, so there's nothing to free (it 2247 * points to the same buffer as the arc_buf_t's b_data field). 2248 */ 2249 if (hdr->b_l2hdr.b_compress == ZIO_COMPRESS_OFF) { 2250 hdr->b_l1hdr.b_tmp_cdata = NULL; 2251 return; 2252 } 2253 2254 /* 2255 * There's nothing to free since the buffer was all zero's and 2256 * compressed to a zero length buffer. 2257 */ 2258 if (hdr->b_l2hdr.b_compress == ZIO_COMPRESS_EMPTY) { 2259 ASSERT3P(hdr->b_l1hdr.b_tmp_cdata, ==, NULL); 2260 return; 2261 } 2262 2263 ASSERT(L2ARC_IS_VALID_COMPRESS(hdr->b_l2hdr.b_compress)); 2264 2265 arc_buf_free_on_write(hdr->b_l1hdr.b_tmp_cdata, 2266 hdr->b_size, zio_data_buf_free); 2267 2268 ARCSTAT_BUMP(arcstat_l2_cdata_free_on_write); 2269 hdr->b_l1hdr.b_tmp_cdata = NULL; 2270} 2271 2272/* 2273 * Free up buf->b_data and if 'remove' is set, then pull the 2274 * arc_buf_t off of the the arc_buf_hdr_t's list and free it. 2275 */ 2276static void 2277arc_buf_destroy(arc_buf_t *buf, boolean_t remove) 2278{ 2279 arc_buf_t **bufp; 2280 2281 /* free up data associated with the buf */ 2282 if (buf->b_data != NULL) { 2283 arc_state_t *state = buf->b_hdr->b_l1hdr.b_state; 2284 uint64_t size = buf->b_hdr->b_size; 2285 arc_buf_contents_t type = arc_buf_type(buf->b_hdr); 2286 2287 arc_cksum_verify(buf); 2288#ifdef illumos 2289 arc_buf_unwatch(buf); 2290#endif 2291 2292 if (type == ARC_BUFC_METADATA) { 2293 arc_buf_data_free(buf, zio_buf_free); 2294 arc_space_return(size, ARC_SPACE_META); 2295 } else { 2296 ASSERT(type == ARC_BUFC_DATA); 2297 arc_buf_data_free(buf, zio_data_buf_free); 2298 arc_space_return(size, ARC_SPACE_DATA); 2299 } 2300 2301 /* protected by hash lock, if in the hash table */ 2302 if (multilist_link_active(&buf->b_hdr->b_l1hdr.b_arc_node)) { 2303 uint64_t *cnt = &state->arcs_lsize[type]; 2304 2305 ASSERT(refcount_is_zero( 2306 &buf->b_hdr->b_l1hdr.b_refcnt)); 2307 ASSERT(state != arc_anon && state != arc_l2c_only); 2308 2309 ASSERT3U(*cnt, >=, size); 2310 atomic_add_64(cnt, -size); 2311 } 2312 2313 (void) refcount_remove_many(&state->arcs_size, size, buf); 2314 buf->b_data = NULL; 2315 2316 /* 2317 * If we're destroying a duplicate buffer make sure 2318 * that the appropriate statistics are updated. 2319 */ 2320 if (buf->b_hdr->b_l1hdr.b_datacnt > 1 && 2321 HDR_ISTYPE_DATA(buf->b_hdr)) { 2322 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers); 2323 ARCSTAT_INCR(arcstat_duplicate_buffers_size, -size); 2324 } 2325 ASSERT(buf->b_hdr->b_l1hdr.b_datacnt > 0); 2326 buf->b_hdr->b_l1hdr.b_datacnt -= 1; 2327 } 2328 2329 /* only remove the buf if requested */ 2330 if (!remove) 2331 return; 2332 2333 /* remove the buf from the hdr list */ 2334 for (bufp = &buf->b_hdr->b_l1hdr.b_buf; *bufp != buf; 2335 bufp = &(*bufp)->b_next) 2336 continue; 2337 *bufp = buf->b_next; 2338 buf->b_next = NULL; 2339 2340 ASSERT(buf->b_efunc == NULL); 2341 2342 /* clean up the buf */ 2343 buf->b_hdr = NULL; 2344 kmem_cache_free(buf_cache, buf); 2345} 2346 2347static void 2348arc_hdr_l2hdr_destroy(arc_buf_hdr_t *hdr) 2349{ 2350 l2arc_buf_hdr_t *l2hdr = &hdr->b_l2hdr; 2351 l2arc_dev_t *dev = l2hdr->b_dev; 2352 2353 ASSERT(MUTEX_HELD(&dev->l2ad_mtx)); 2354 ASSERT(HDR_HAS_L2HDR(hdr)); 2355 2356 list_remove(&dev->l2ad_buflist, hdr); 2357 2358 /* 2359 * We don't want to leak the b_tmp_cdata buffer that was 2360 * allocated in l2arc_write_buffers() 2361 */ 2362 arc_buf_l2_cdata_free(hdr); 2363 2364 /* 2365 * If the l2hdr's b_daddr is equal to L2ARC_ADDR_UNSET, then 2366 * this header is being processed by l2arc_write_buffers() (i.e. 2367 * it's in the first stage of l2arc_write_buffers()). 2368 * Re-affirming that truth here, just to serve as a reminder. If 2369 * b_daddr does not equal L2ARC_ADDR_UNSET, then the header may or 2370 * may not have its HDR_L2_WRITING flag set. (the write may have 2371 * completed, in which case HDR_L2_WRITING will be false and the 2372 * b_daddr field will point to the address of the buffer on disk). 2373 */ 2374 IMPLY(l2hdr->b_daddr == L2ARC_ADDR_UNSET, HDR_L2_WRITING(hdr)); 2375 2376 /* 2377 * If b_daddr is equal to L2ARC_ADDR_UNSET, we're racing with 2378 * l2arc_write_buffers(). Since we've just removed this header 2379 * from the l2arc buffer list, this header will never reach the 2380 * second stage of l2arc_write_buffers(), which increments the 2381 * accounting stats for this header. Thus, we must be careful 2382 * not to decrement them for this header either. 2383 */ 2384 if (l2hdr->b_daddr != L2ARC_ADDR_UNSET) { 2385 ARCSTAT_INCR(arcstat_l2_asize, -l2hdr->b_asize); 2386 ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size); 2387 2388 vdev_space_update(dev->l2ad_vdev, 2389 -l2hdr->b_asize, 0, 0); 2390 2391 (void) refcount_remove_many(&dev->l2ad_alloc, 2392 l2hdr->b_asize, hdr); 2393 } 2394 2395 hdr->b_flags &= ~ARC_FLAG_HAS_L2HDR; 2396} 2397 2398static void 2399arc_hdr_destroy(arc_buf_hdr_t *hdr) 2400{ 2401 if (HDR_HAS_L1HDR(hdr)) { 2402 ASSERT(hdr->b_l1hdr.b_buf == NULL || 2403 hdr->b_l1hdr.b_datacnt > 0); 2404 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 2405 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon); 2406 } 2407 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 2408 ASSERT(!HDR_IN_HASH_TABLE(hdr)); 2409 2410 if (HDR_HAS_L2HDR(hdr)) { 2411 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev; 2412 boolean_t buflist_held = MUTEX_HELD(&dev->l2ad_mtx); 2413 2414 if (!buflist_held) 2415 mutex_enter(&dev->l2ad_mtx); 2416 2417 /* 2418 * Even though we checked this conditional above, we 2419 * need to check this again now that we have the 2420 * l2ad_mtx. This is because we could be racing with 2421 * another thread calling l2arc_evict() which might have 2422 * destroyed this header's L2 portion as we were waiting 2423 * to acquire the l2ad_mtx. If that happens, we don't 2424 * want to re-destroy the header's L2 portion. 2425 */ 2426 if (HDR_HAS_L2HDR(hdr)) { 2427 l2arc_trim(hdr); 2428 arc_hdr_l2hdr_destroy(hdr); 2429 } 2430 2431 if (!buflist_held) 2432 mutex_exit(&dev->l2ad_mtx); 2433 } 2434 2435 if (!BUF_EMPTY(hdr)) 2436 buf_discard_identity(hdr); 2437 2438 if (hdr->b_freeze_cksum != NULL) { 2439 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 2440 hdr->b_freeze_cksum = NULL; 2441 } 2442 2443 if (HDR_HAS_L1HDR(hdr)) { 2444 while (hdr->b_l1hdr.b_buf) { 2445 arc_buf_t *buf = hdr->b_l1hdr.b_buf; 2446 2447 if (buf->b_efunc != NULL) { 2448 mutex_enter(&arc_user_evicts_lock); 2449 mutex_enter(&buf->b_evict_lock); 2450 ASSERT(buf->b_hdr != NULL); 2451 arc_buf_destroy(hdr->b_l1hdr.b_buf, FALSE); 2452 hdr->b_l1hdr.b_buf = buf->b_next; 2453 buf->b_hdr = &arc_eviction_hdr; 2454 buf->b_next = arc_eviction_list; 2455 arc_eviction_list = buf; 2456 mutex_exit(&buf->b_evict_lock); 2457 cv_signal(&arc_user_evicts_cv); 2458 mutex_exit(&arc_user_evicts_lock); 2459 } else { 2460 arc_buf_destroy(hdr->b_l1hdr.b_buf, TRUE); 2461 } 2462 } 2463#ifdef ZFS_DEBUG 2464 if (hdr->b_l1hdr.b_thawed != NULL) { 2465 kmem_free(hdr->b_l1hdr.b_thawed, 1); 2466 hdr->b_l1hdr.b_thawed = NULL; 2467 } 2468#endif 2469 } 2470 2471 ASSERT3P(hdr->b_hash_next, ==, NULL); 2472 if (HDR_HAS_L1HDR(hdr)) { 2473 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 2474 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL); 2475 kmem_cache_free(hdr_full_cache, hdr); 2476 } else { 2477 kmem_cache_free(hdr_l2only_cache, hdr); 2478 } 2479} 2480 2481void 2482arc_buf_free(arc_buf_t *buf, void *tag) 2483{ 2484 arc_buf_hdr_t *hdr = buf->b_hdr; 2485 int hashed = hdr->b_l1hdr.b_state != arc_anon; 2486 2487 ASSERT(buf->b_efunc == NULL); 2488 ASSERT(buf->b_data != NULL); 2489 2490 if (hashed) { 2491 kmutex_t *hash_lock = HDR_LOCK(hdr); 2492 2493 mutex_enter(hash_lock); 2494 hdr = buf->b_hdr; 2495 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 2496 2497 (void) remove_reference(hdr, hash_lock, tag); 2498 if (hdr->b_l1hdr.b_datacnt > 1) { 2499 arc_buf_destroy(buf, TRUE); 2500 } else { 2501 ASSERT(buf == hdr->b_l1hdr.b_buf); 2502 ASSERT(buf->b_efunc == NULL); 2503 hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE; 2504 } 2505 mutex_exit(hash_lock); 2506 } else if (HDR_IO_IN_PROGRESS(hdr)) { 2507 int destroy_hdr; 2508 /* 2509 * We are in the middle of an async write. Don't destroy 2510 * this buffer unless the write completes before we finish 2511 * decrementing the reference count. 2512 */ 2513 mutex_enter(&arc_user_evicts_lock); 2514 (void) remove_reference(hdr, NULL, tag); 2515 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 2516 destroy_hdr = !HDR_IO_IN_PROGRESS(hdr); 2517 mutex_exit(&arc_user_evicts_lock); 2518 if (destroy_hdr) 2519 arc_hdr_destroy(hdr); 2520 } else { 2521 if (remove_reference(hdr, NULL, tag) > 0) 2522 arc_buf_destroy(buf, TRUE); 2523 else 2524 arc_hdr_destroy(hdr); 2525 } 2526} 2527 2528boolean_t 2529arc_buf_remove_ref(arc_buf_t *buf, void* tag) 2530{ 2531 arc_buf_hdr_t *hdr = buf->b_hdr; 2532 kmutex_t *hash_lock = HDR_LOCK(hdr); 2533 boolean_t no_callback = (buf->b_efunc == NULL); 2534 2535 if (hdr->b_l1hdr.b_state == arc_anon) { 2536 ASSERT(hdr->b_l1hdr.b_datacnt == 1); 2537 arc_buf_free(buf, tag); 2538 return (no_callback); 2539 } 2540 2541 mutex_enter(hash_lock); 2542 hdr = buf->b_hdr; 2543 ASSERT(hdr->b_l1hdr.b_datacnt > 0); 2544 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 2545 ASSERT(hdr->b_l1hdr.b_state != arc_anon); 2546 ASSERT(buf->b_data != NULL); 2547 2548 (void) remove_reference(hdr, hash_lock, tag); 2549 if (hdr->b_l1hdr.b_datacnt > 1) { 2550 if (no_callback) 2551 arc_buf_destroy(buf, TRUE); 2552 } else if (no_callback) { 2553 ASSERT(hdr->b_l1hdr.b_buf == buf && buf->b_next == NULL); 2554 ASSERT(buf->b_efunc == NULL); 2555 hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE; 2556 } 2557 ASSERT(no_callback || hdr->b_l1hdr.b_datacnt > 1 || 2558 refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 2559 mutex_exit(hash_lock); 2560 return (no_callback); 2561} 2562 2563int32_t 2564arc_buf_size(arc_buf_t *buf) 2565{ 2566 return (buf->b_hdr->b_size); 2567} 2568 2569/* 2570 * Called from the DMU to determine if the current buffer should be 2571 * evicted. In order to ensure proper locking, the eviction must be initiated 2572 * from the DMU. Return true if the buffer is associated with user data and 2573 * duplicate buffers still exist. 2574 */ 2575boolean_t 2576arc_buf_eviction_needed(arc_buf_t *buf) 2577{ 2578 arc_buf_hdr_t *hdr; 2579 boolean_t evict_needed = B_FALSE; 2580 2581 if (zfs_disable_dup_eviction) 2582 return (B_FALSE); 2583 2584 mutex_enter(&buf->b_evict_lock); 2585 hdr = buf->b_hdr; 2586 if (hdr == NULL) { 2587 /* 2588 * We are in arc_do_user_evicts(); let that function 2589 * perform the eviction. 2590 */ 2591 ASSERT(buf->b_data == NULL); 2592 mutex_exit(&buf->b_evict_lock); 2593 return (B_FALSE); 2594 } else if (buf->b_data == NULL) { 2595 /* 2596 * We have already been added to the arc eviction list; 2597 * recommend eviction. 2598 */ 2599 ASSERT3P(hdr, ==, &arc_eviction_hdr); 2600 mutex_exit(&buf->b_evict_lock); 2601 return (B_TRUE); 2602 } 2603 2604 if (hdr->b_l1hdr.b_datacnt > 1 && HDR_ISTYPE_DATA(hdr)) 2605 evict_needed = B_TRUE; 2606 2607 mutex_exit(&buf->b_evict_lock); 2608 return (evict_needed); 2609} 2610 2611/* 2612 * Evict the arc_buf_hdr that is provided as a parameter. The resultant 2613 * state of the header is dependent on it's state prior to entering this 2614 * function. The following transitions are possible: 2615 * 2616 * - arc_mru -> arc_mru_ghost 2617 * - arc_mfu -> arc_mfu_ghost 2618 * - arc_mru_ghost -> arc_l2c_only 2619 * - arc_mru_ghost -> deleted 2620 * - arc_mfu_ghost -> arc_l2c_only 2621 * - arc_mfu_ghost -> deleted 2622 */ 2623static int64_t 2624arc_evict_hdr(arc_buf_hdr_t *hdr, kmutex_t *hash_lock) 2625{ 2626 arc_state_t *evicted_state, *state; 2627 int64_t bytes_evicted = 0; 2628 2629 ASSERT(MUTEX_HELD(hash_lock)); 2630 ASSERT(HDR_HAS_L1HDR(hdr)); 2631 2632 state = hdr->b_l1hdr.b_state; 2633 if (GHOST_STATE(state)) { 2634 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 2635 ASSERT(hdr->b_l1hdr.b_buf == NULL); 2636 2637 /* 2638 * l2arc_write_buffers() relies on a header's L1 portion 2639 * (i.e. it's b_tmp_cdata field) during it's write phase. 2640 * Thus, we cannot push a header onto the arc_l2c_only 2641 * state (removing it's L1 piece) until the header is 2642 * done being written to the l2arc. 2643 */ 2644 if (HDR_HAS_L2HDR(hdr) && HDR_L2_WRITING(hdr)) { 2645 ARCSTAT_BUMP(arcstat_evict_l2_skip); 2646 return (bytes_evicted); 2647 } 2648 2649 ARCSTAT_BUMP(arcstat_deleted); 2650 bytes_evicted += hdr->b_size; 2651 2652 DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, hdr); 2653 2654 if (HDR_HAS_L2HDR(hdr)) { 2655 /* 2656 * This buffer is cached on the 2nd Level ARC; 2657 * don't destroy the header. 2658 */ 2659 arc_change_state(arc_l2c_only, hdr, hash_lock); 2660 /* 2661 * dropping from L1+L2 cached to L2-only, 2662 * realloc to remove the L1 header. 2663 */ 2664 hdr = arc_hdr_realloc(hdr, hdr_full_cache, 2665 hdr_l2only_cache); 2666 } else { 2667 arc_change_state(arc_anon, hdr, hash_lock); 2668 arc_hdr_destroy(hdr); 2669 } 2670 return (bytes_evicted); 2671 } 2672 2673 ASSERT(state == arc_mru || state == arc_mfu); 2674 evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost; 2675 2676 /* prefetch buffers have a minimum lifespan */ 2677 if (HDR_IO_IN_PROGRESS(hdr) || 2678 ((hdr->b_flags & (ARC_FLAG_PREFETCH | ARC_FLAG_INDIRECT)) && 2679 ddi_get_lbolt() - hdr->b_l1hdr.b_arc_access < 2680 arc_min_prefetch_lifespan)) { 2681 ARCSTAT_BUMP(arcstat_evict_skip); 2682 return (bytes_evicted); 2683 } 2684 2685 ASSERT0(refcount_count(&hdr->b_l1hdr.b_refcnt)); 2686 ASSERT3U(hdr->b_l1hdr.b_datacnt, >, 0); 2687 while (hdr->b_l1hdr.b_buf) { 2688 arc_buf_t *buf = hdr->b_l1hdr.b_buf; 2689 if (!mutex_tryenter(&buf->b_evict_lock)) { 2690 ARCSTAT_BUMP(arcstat_mutex_miss); 2691 break; 2692 } 2693 if (buf->b_data != NULL) 2694 bytes_evicted += hdr->b_size; 2695 if (buf->b_efunc != NULL) { 2696 mutex_enter(&arc_user_evicts_lock); 2697 arc_buf_destroy(buf, FALSE); 2698 hdr->b_l1hdr.b_buf = buf->b_next; 2699 buf->b_hdr = &arc_eviction_hdr; 2700 buf->b_next = arc_eviction_list; 2701 arc_eviction_list = buf; 2702 cv_signal(&arc_user_evicts_cv); 2703 mutex_exit(&arc_user_evicts_lock); 2704 mutex_exit(&buf->b_evict_lock); 2705 } else { 2706 mutex_exit(&buf->b_evict_lock); 2707 arc_buf_destroy(buf, TRUE); 2708 } 2709 } 2710 2711 if (HDR_HAS_L2HDR(hdr)) { 2712 ARCSTAT_INCR(arcstat_evict_l2_cached, hdr->b_size); 2713 } else { 2714 if (l2arc_write_eligible(hdr->b_spa, hdr)) 2715 ARCSTAT_INCR(arcstat_evict_l2_eligible, hdr->b_size); 2716 else 2717 ARCSTAT_INCR(arcstat_evict_l2_ineligible, hdr->b_size); 2718 } 2719 2720 if (hdr->b_l1hdr.b_datacnt == 0) { 2721 arc_change_state(evicted_state, hdr, hash_lock); 2722 ASSERT(HDR_IN_HASH_TABLE(hdr)); 2723 hdr->b_flags |= ARC_FLAG_IN_HASH_TABLE; 2724 hdr->b_flags &= ~ARC_FLAG_BUF_AVAILABLE; 2725 DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, hdr); 2726 } 2727 2728 return (bytes_evicted); 2729} 2730 2731static uint64_t 2732arc_evict_state_impl(multilist_t *ml, int idx, arc_buf_hdr_t *marker, 2733 uint64_t spa, int64_t bytes) 2734{ 2735 multilist_sublist_t *mls; 2736 uint64_t bytes_evicted = 0; 2737 arc_buf_hdr_t *hdr; 2738 kmutex_t *hash_lock; 2739 int evict_count = 0; 2740 2741 ASSERT3P(marker, !=, NULL); 2742 IMPLY(bytes < 0, bytes == ARC_EVICT_ALL); 2743 2744 mls = multilist_sublist_lock(ml, idx); 2745 2746 for (hdr = multilist_sublist_prev(mls, marker); hdr != NULL; 2747 hdr = multilist_sublist_prev(mls, marker)) { 2748 if ((bytes != ARC_EVICT_ALL && bytes_evicted >= bytes) || 2749 (evict_count >= zfs_arc_evict_batch_limit)) 2750 break; 2751 2752 /* 2753 * To keep our iteration location, move the marker 2754 * forward. Since we're not holding hdr's hash lock, we 2755 * must be very careful and not remove 'hdr' from the 2756 * sublist. Otherwise, other consumers might mistake the 2757 * 'hdr' as not being on a sublist when they call the 2758 * multilist_link_active() function (they all rely on 2759 * the hash lock protecting concurrent insertions and 2760 * removals). multilist_sublist_move_forward() was 2761 * specifically implemented to ensure this is the case 2762 * (only 'marker' will be removed and re-inserted). 2763 */ 2764 multilist_sublist_move_forward(mls, marker); 2765 2766 /* 2767 * The only case where the b_spa field should ever be 2768 * zero, is the marker headers inserted by 2769 * arc_evict_state(). It's possible for multiple threads 2770 * to be calling arc_evict_state() concurrently (e.g. 2771 * dsl_pool_close() and zio_inject_fault()), so we must 2772 * skip any markers we see from these other threads. 2773 */ 2774 if (hdr->b_spa == 0) 2775 continue; 2776 2777 /* we're only interested in evicting buffers of a certain spa */ 2778 if (spa != 0 && hdr->b_spa != spa) { 2779 ARCSTAT_BUMP(arcstat_evict_skip); 2780 continue; 2781 } 2782 2783 hash_lock = HDR_LOCK(hdr); 2784 2785 /* 2786 * We aren't calling this function from any code path 2787 * that would already be holding a hash lock, so we're 2788 * asserting on this assumption to be defensive in case 2789 * this ever changes. Without this check, it would be 2790 * possible to incorrectly increment arcstat_mutex_miss 2791 * below (e.g. if the code changed such that we called 2792 * this function with a hash lock held). 2793 */ 2794 ASSERT(!MUTEX_HELD(hash_lock)); 2795 2796 if (mutex_tryenter(hash_lock)) { 2797 uint64_t evicted = arc_evict_hdr(hdr, hash_lock); 2798 mutex_exit(hash_lock); 2799 2800 bytes_evicted += evicted; 2801 2802 /* 2803 * If evicted is zero, arc_evict_hdr() must have 2804 * decided to skip this header, don't increment 2805 * evict_count in this case. 2806 */ 2807 if (evicted != 0) 2808 evict_count++; 2809 2810 /* 2811 * If arc_size isn't overflowing, signal any 2812 * threads that might happen to be waiting. 2813 * 2814 * For each header evicted, we wake up a single 2815 * thread. If we used cv_broadcast, we could 2816 * wake up "too many" threads causing arc_size 2817 * to significantly overflow arc_c; since 2818 * arc_get_data_buf() doesn't check for overflow 2819 * when it's woken up (it doesn't because it's 2820 * possible for the ARC to be overflowing while 2821 * full of un-evictable buffers, and the 2822 * function should proceed in this case). 2823 * 2824 * If threads are left sleeping, due to not 2825 * using cv_broadcast, they will be woken up 2826 * just before arc_reclaim_thread() sleeps. 2827 */ 2828 mutex_enter(&arc_reclaim_lock); 2829 if (!arc_is_overflowing()) 2830 cv_signal(&arc_reclaim_waiters_cv); 2831 mutex_exit(&arc_reclaim_lock); 2832 } else { 2833 ARCSTAT_BUMP(arcstat_mutex_miss); 2834 } 2835 } 2836 2837 multilist_sublist_unlock(mls); 2838 2839 return (bytes_evicted); 2840} 2841 2842/* 2843 * Evict buffers from the given arc state, until we've removed the 2844 * specified number of bytes. Move the removed buffers to the 2845 * appropriate evict state. 2846 * 2847 * This function makes a "best effort". It skips over any buffers 2848 * it can't get a hash_lock on, and so, may not catch all candidates. 2849 * It may also return without evicting as much space as requested. 2850 * 2851 * If bytes is specified using the special value ARC_EVICT_ALL, this 2852 * will evict all available (i.e. unlocked and evictable) buffers from 2853 * the given arc state; which is used by arc_flush(). 2854 */ 2855static uint64_t 2856arc_evict_state(arc_state_t *state, uint64_t spa, int64_t bytes, 2857 arc_buf_contents_t type) 2858{ 2859 uint64_t total_evicted = 0; 2860 multilist_t *ml = &state->arcs_list[type]; 2861 int num_sublists; 2862 arc_buf_hdr_t **markers; 2863 2864 IMPLY(bytes < 0, bytes == ARC_EVICT_ALL); 2865 2866 num_sublists = multilist_get_num_sublists(ml); 2867 2868 /* 2869 * If we've tried to evict from each sublist, made some 2870 * progress, but still have not hit the target number of bytes 2871 * to evict, we want to keep trying. The markers allow us to 2872 * pick up where we left off for each individual sublist, rather 2873 * than starting from the tail each time. 2874 */ 2875 markers = kmem_zalloc(sizeof (*markers) * num_sublists, KM_SLEEP); 2876 for (int i = 0; i < num_sublists; i++) { 2877 markers[i] = kmem_cache_alloc(hdr_full_cache, KM_SLEEP); 2878 2879 /* 2880 * A b_spa of 0 is used to indicate that this header is 2881 * a marker. This fact is used in arc_adjust_type() and 2882 * arc_evict_state_impl(). 2883 */ 2884 markers[i]->b_spa = 0; 2885 2886 multilist_sublist_t *mls = multilist_sublist_lock(ml, i); 2887 multilist_sublist_insert_tail(mls, markers[i]); 2888 multilist_sublist_unlock(mls); 2889 } 2890 2891 /* 2892 * While we haven't hit our target number of bytes to evict, or 2893 * we're evicting all available buffers. 2894 */ 2895 while (total_evicted < bytes || bytes == ARC_EVICT_ALL) { 2896 /* 2897 * Start eviction using a randomly selected sublist, 2898 * this is to try and evenly balance eviction across all 2899 * sublists. Always starting at the same sublist 2900 * (e.g. index 0) would cause evictions to favor certain 2901 * sublists over others. 2902 */ 2903 int sublist_idx = multilist_get_random_index(ml); 2904 uint64_t scan_evicted = 0; 2905 2906 for (int i = 0; i < num_sublists; i++) { 2907 uint64_t bytes_remaining; 2908 uint64_t bytes_evicted; 2909 2910 if (bytes == ARC_EVICT_ALL) 2911 bytes_remaining = ARC_EVICT_ALL; 2912 else if (total_evicted < bytes) 2913 bytes_remaining = bytes - total_evicted; 2914 else 2915 break; 2916 2917 bytes_evicted = arc_evict_state_impl(ml, sublist_idx, 2918 markers[sublist_idx], spa, bytes_remaining); 2919 2920 scan_evicted += bytes_evicted; 2921 total_evicted += bytes_evicted; 2922 2923 /* we've reached the end, wrap to the beginning */ 2924 if (++sublist_idx >= num_sublists) 2925 sublist_idx = 0; 2926 } 2927 2928 /* 2929 * If we didn't evict anything during this scan, we have 2930 * no reason to believe we'll evict more during another 2931 * scan, so break the loop. 2932 */ 2933 if (scan_evicted == 0) { 2934 /* This isn't possible, let's make that obvious */ 2935 ASSERT3S(bytes, !=, 0); 2936 2937 /* 2938 * When bytes is ARC_EVICT_ALL, the only way to 2939 * break the loop is when scan_evicted is zero. 2940 * In that case, we actually have evicted enough, 2941 * so we don't want to increment the kstat. 2942 */ 2943 if (bytes != ARC_EVICT_ALL) { 2944 ASSERT3S(total_evicted, <, bytes); 2945 ARCSTAT_BUMP(arcstat_evict_not_enough); 2946 } 2947 2948 break; 2949 } 2950 } 2951 2952 for (int i = 0; i < num_sublists; i++) { 2953 multilist_sublist_t *mls = multilist_sublist_lock(ml, i); 2954 multilist_sublist_remove(mls, markers[i]); 2955 multilist_sublist_unlock(mls); 2956 2957 kmem_cache_free(hdr_full_cache, markers[i]); 2958 } 2959 kmem_free(markers, sizeof (*markers) * num_sublists); 2960 2961 return (total_evicted); 2962} 2963 2964/* 2965 * Flush all "evictable" data of the given type from the arc state 2966 * specified. This will not evict any "active" buffers (i.e. referenced). 2967 * 2968 * When 'retry' is set to FALSE, the function will make a single pass 2969 * over the state and evict any buffers that it can. Since it doesn't 2970 * continually retry the eviction, it might end up leaving some buffers 2971 * in the ARC due to lock misses. 2972 * 2973 * When 'retry' is set to TRUE, the function will continually retry the 2974 * eviction until *all* evictable buffers have been removed from the 2975 * state. As a result, if concurrent insertions into the state are 2976 * allowed (e.g. if the ARC isn't shutting down), this function might 2977 * wind up in an infinite loop, continually trying to evict buffers. 2978 */ 2979static uint64_t 2980arc_flush_state(arc_state_t *state, uint64_t spa, arc_buf_contents_t type, 2981 boolean_t retry) 2982{ 2983 uint64_t evicted = 0; 2984 2985 while (state->arcs_lsize[type] != 0) { 2986 evicted += arc_evict_state(state, spa, ARC_EVICT_ALL, type); 2987 2988 if (!retry) 2989 break; 2990 } 2991 2992 return (evicted); 2993} 2994 2995/* 2996 * Evict the specified number of bytes from the state specified, 2997 * restricting eviction to the spa and type given. This function 2998 * prevents us from trying to evict more from a state's list than 2999 * is "evictable", and to skip evicting altogether when passed a 3000 * negative value for "bytes". In contrast, arc_evict_state() will 3001 * evict everything it can, when passed a negative value for "bytes". 3002 */ 3003static uint64_t 3004arc_adjust_impl(arc_state_t *state, uint64_t spa, int64_t bytes, 3005 arc_buf_contents_t type) 3006{ 3007 int64_t delta; 3008 3009 if (bytes > 0 && state->arcs_lsize[type] > 0) { 3010 delta = MIN(state->arcs_lsize[type], bytes); 3011 return (arc_evict_state(state, spa, delta, type)); 3012 } 3013 3014 return (0); 3015} 3016 3017/* 3018 * Evict metadata buffers from the cache, such that arc_meta_used is 3019 * capped by the arc_meta_limit tunable. 3020 */ 3021static uint64_t 3022arc_adjust_meta(void) 3023{ 3024 uint64_t total_evicted = 0; 3025 int64_t target; 3026 3027 /* 3028 * If we're over the meta limit, we want to evict enough 3029 * metadata to get back under the meta limit. We don't want to 3030 * evict so much that we drop the MRU below arc_p, though. If 3031 * we're over the meta limit more than we're over arc_p, we 3032 * evict some from the MRU here, and some from the MFU below. 3033 */ 3034 target = MIN((int64_t)(arc_meta_used - arc_meta_limit), 3035 (int64_t)(refcount_count(&arc_anon->arcs_size) + 3036 refcount_count(&arc_mru->arcs_size) - arc_p)); 3037 3038 total_evicted += arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA); 3039 3040 /* 3041 * Similar to the above, we want to evict enough bytes to get us 3042 * below the meta limit, but not so much as to drop us below the 3043 * space alloted to the MFU (which is defined as arc_c - arc_p). 3044 */ 3045 target = MIN((int64_t)(arc_meta_used - arc_meta_limit), 3046 (int64_t)(refcount_count(&arc_mfu->arcs_size) - (arc_c - arc_p))); 3047 3048 total_evicted += arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA); 3049 3050 return (total_evicted); 3051} 3052 3053/* 3054 * Return the type of the oldest buffer in the given arc state 3055 * 3056 * This function will select a random sublist of type ARC_BUFC_DATA and 3057 * a random sublist of type ARC_BUFC_METADATA. The tail of each sublist 3058 * is compared, and the type which contains the "older" buffer will be 3059 * returned. 3060 */ 3061static arc_buf_contents_t 3062arc_adjust_type(arc_state_t *state) 3063{ 3064 multilist_t *data_ml = &state->arcs_list[ARC_BUFC_DATA]; 3065 multilist_t *meta_ml = &state->arcs_list[ARC_BUFC_METADATA]; 3066 int data_idx = multilist_get_random_index(data_ml); 3067 int meta_idx = multilist_get_random_index(meta_ml); 3068 multilist_sublist_t *data_mls; 3069 multilist_sublist_t *meta_mls; 3070 arc_buf_contents_t type; 3071 arc_buf_hdr_t *data_hdr; 3072 arc_buf_hdr_t *meta_hdr; 3073 3074 /* 3075 * We keep the sublist lock until we're finished, to prevent 3076 * the headers from being destroyed via arc_evict_state(). 3077 */ 3078 data_mls = multilist_sublist_lock(data_ml, data_idx); 3079 meta_mls = multilist_sublist_lock(meta_ml, meta_idx); 3080 3081 /* 3082 * These two loops are to ensure we skip any markers that 3083 * might be at the tail of the lists due to arc_evict_state(). 3084 */ 3085 3086 for (data_hdr = multilist_sublist_tail(data_mls); data_hdr != NULL; 3087 data_hdr = multilist_sublist_prev(data_mls, data_hdr)) { 3088 if (data_hdr->b_spa != 0) 3089 break; 3090 } 3091 3092 for (meta_hdr = multilist_sublist_tail(meta_mls); meta_hdr != NULL; 3093 meta_hdr = multilist_sublist_prev(meta_mls, meta_hdr)) { 3094 if (meta_hdr->b_spa != 0) 3095 break; 3096 } 3097 3098 if (data_hdr == NULL && meta_hdr == NULL) { 3099 type = ARC_BUFC_DATA; 3100 } else if (data_hdr == NULL) { 3101 ASSERT3P(meta_hdr, !=, NULL); 3102 type = ARC_BUFC_METADATA; 3103 } else if (meta_hdr == NULL) { 3104 ASSERT3P(data_hdr, !=, NULL); 3105 type = ARC_BUFC_DATA; 3106 } else { 3107 ASSERT3P(data_hdr, !=, NULL); 3108 ASSERT3P(meta_hdr, !=, NULL); 3109 3110 /* The headers can't be on the sublist without an L1 header */ 3111 ASSERT(HDR_HAS_L1HDR(data_hdr)); 3112 ASSERT(HDR_HAS_L1HDR(meta_hdr)); 3113 3114 if (data_hdr->b_l1hdr.b_arc_access < 3115 meta_hdr->b_l1hdr.b_arc_access) { 3116 type = ARC_BUFC_DATA; 3117 } else { 3118 type = ARC_BUFC_METADATA; 3119 } 3120 } 3121 3122 multilist_sublist_unlock(meta_mls); 3123 multilist_sublist_unlock(data_mls); 3124 3125 return (type); 3126} 3127 3128/* 3129 * Evict buffers from the cache, such that arc_size is capped by arc_c. 3130 */ 3131static uint64_t 3132arc_adjust(void) 3133{ 3134 uint64_t total_evicted = 0; 3135 uint64_t bytes; 3136 int64_t target; 3137 3138 /* 3139 * If we're over arc_meta_limit, we want to correct that before 3140 * potentially evicting data buffers below. 3141 */ 3142 total_evicted += arc_adjust_meta(); 3143 3144 /* 3145 * Adjust MRU size 3146 * 3147 * If we're over the target cache size, we want to evict enough 3148 * from the list to get back to our target size. We don't want 3149 * to evict too much from the MRU, such that it drops below 3150 * arc_p. So, if we're over our target cache size more than 3151 * the MRU is over arc_p, we'll evict enough to get back to 3152 * arc_p here, and then evict more from the MFU below. 3153 */ 3154 target = MIN((int64_t)(arc_size - arc_c), 3155 (int64_t)(refcount_count(&arc_anon->arcs_size) + 3156 refcount_count(&arc_mru->arcs_size) + arc_meta_used - arc_p)); 3157 3158 /* 3159 * If we're below arc_meta_min, always prefer to evict data. 3160 * Otherwise, try to satisfy the requested number of bytes to 3161 * evict from the type which contains older buffers; in an 3162 * effort to keep newer buffers in the cache regardless of their 3163 * type. If we cannot satisfy the number of bytes from this 3164 * type, spill over into the next type. 3165 */ 3166 if (arc_adjust_type(arc_mru) == ARC_BUFC_METADATA && 3167 arc_meta_used > arc_meta_min) { 3168 bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA); 3169 total_evicted += bytes; 3170 3171 /* 3172 * If we couldn't evict our target number of bytes from 3173 * metadata, we try to get the rest from data. 3174 */ 3175 target -= bytes; 3176 3177 total_evicted += 3178 arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA); 3179 } else { 3180 bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA); 3181 total_evicted += bytes; 3182 3183 /* 3184 * If we couldn't evict our target number of bytes from 3185 * data, we try to get the rest from metadata. 3186 */ 3187 target -= bytes; 3188 3189 total_evicted += 3190 arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA); 3191 } 3192 3193 /* 3194 * Adjust MFU size 3195 * 3196 * Now that we've tried to evict enough from the MRU to get its 3197 * size back to arc_p, if we're still above the target cache 3198 * size, we evict the rest from the MFU. 3199 */ 3200 target = arc_size - arc_c; 3201 3202 if (arc_adjust_type(arc_mfu) == ARC_BUFC_METADATA && 3203 arc_meta_used > arc_meta_min) { 3204 bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA); 3205 total_evicted += bytes; 3206 3207 /* 3208 * If we couldn't evict our target number of bytes from 3209 * metadata, we try to get the rest from data. 3210 */ 3211 target -= bytes; 3212 3213 total_evicted += 3214 arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA); 3215 } else { 3216 bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA); 3217 total_evicted += bytes; 3218 3219 /* 3220 * If we couldn't evict our target number of bytes from 3221 * data, we try to get the rest from data. 3222 */ 3223 target -= bytes; 3224 3225 total_evicted += 3226 arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA); 3227 } 3228 3229 /* 3230 * Adjust ghost lists 3231 * 3232 * In addition to the above, the ARC also defines target values 3233 * for the ghost lists. The sum of the mru list and mru ghost 3234 * list should never exceed the target size of the cache, and 3235 * the sum of the mru list, mfu list, mru ghost list, and mfu 3236 * ghost list should never exceed twice the target size of the 3237 * cache. The following logic enforces these limits on the ghost 3238 * caches, and evicts from them as needed. 3239 */ 3240 target = refcount_count(&arc_mru->arcs_size) + 3241 refcount_count(&arc_mru_ghost->arcs_size) - arc_c; 3242 3243 bytes = arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_DATA); 3244 total_evicted += bytes; 3245 3246 target -= bytes; 3247 3248 total_evicted += 3249 arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_METADATA); 3250 3251 /* 3252 * We assume the sum of the mru list and mfu list is less than 3253 * or equal to arc_c (we enforced this above), which means we 3254 * can use the simpler of the two equations below: 3255 * 3256 * mru + mfu + mru ghost + mfu ghost <= 2 * arc_c 3257 * mru ghost + mfu ghost <= arc_c 3258 */ 3259 target = refcount_count(&arc_mru_ghost->arcs_size) + 3260 refcount_count(&arc_mfu_ghost->arcs_size) - arc_c; 3261 3262 bytes = arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_DATA); 3263 total_evicted += bytes; 3264 3265 target -= bytes; 3266 3267 total_evicted += 3268 arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_METADATA); 3269 3270 return (total_evicted); 3271} 3272 3273static void 3274arc_do_user_evicts(void) 3275{ 3276 mutex_enter(&arc_user_evicts_lock); 3277 while (arc_eviction_list != NULL) { 3278 arc_buf_t *buf = arc_eviction_list; 3279 arc_eviction_list = buf->b_next; 3280 mutex_enter(&buf->b_evict_lock); 3281 buf->b_hdr = NULL; 3282 mutex_exit(&buf->b_evict_lock); 3283 mutex_exit(&arc_user_evicts_lock); 3284 3285 if (buf->b_efunc != NULL) 3286 VERIFY0(buf->b_efunc(buf->b_private)); 3287 3288 buf->b_efunc = NULL; 3289 buf->b_private = NULL; 3290 kmem_cache_free(buf_cache, buf); 3291 mutex_enter(&arc_user_evicts_lock); 3292 } 3293 mutex_exit(&arc_user_evicts_lock); 3294} 3295 3296void 3297arc_flush(spa_t *spa, boolean_t retry) 3298{ 3299 uint64_t guid = 0; 3300 3301 /* 3302 * If retry is TRUE, a spa must not be specified since we have 3303 * no good way to determine if all of a spa's buffers have been 3304 * evicted from an arc state. 3305 */ 3306 ASSERT(!retry || spa == 0); 3307 3308 if (spa != NULL) 3309 guid = spa_load_guid(spa); 3310 3311 (void) arc_flush_state(arc_mru, guid, ARC_BUFC_DATA, retry); 3312 (void) arc_flush_state(arc_mru, guid, ARC_BUFC_METADATA, retry); 3313 3314 (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_DATA, retry); 3315 (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_METADATA, retry); 3316 3317 (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_DATA, retry); 3318 (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_METADATA, retry); 3319 3320 (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_DATA, retry); 3321 (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_METADATA, retry); 3322 3323 arc_do_user_evicts(); 3324 ASSERT(spa || arc_eviction_list == NULL); 3325} 3326 3327void 3328arc_shrink(int64_t to_free) 3329{ 3330 if (arc_c > arc_c_min) { 3331 DTRACE_PROBE4(arc__shrink, uint64_t, arc_c, uint64_t, 3332 arc_c_min, uint64_t, arc_p, uint64_t, to_free); 3333 if (arc_c > arc_c_min + to_free) 3334 atomic_add_64(&arc_c, -to_free); 3335 else 3336 arc_c = arc_c_min; 3337 3338 atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift)); 3339 if (arc_c > arc_size) 3340 arc_c = MAX(arc_size, arc_c_min); 3341 if (arc_p > arc_c) 3342 arc_p = (arc_c >> 1); 3343 3344 DTRACE_PROBE2(arc__shrunk, uint64_t, arc_c, uint64_t, 3345 arc_p); 3346 3347 ASSERT(arc_c >= arc_c_min); 3348 ASSERT((int64_t)arc_p >= 0); 3349 } 3350 3351 if (arc_size > arc_c) { 3352 DTRACE_PROBE2(arc__shrink_adjust, uint64_t, arc_size, 3353 uint64_t, arc_c); 3354 (void) arc_adjust(); 3355 } 3356} 3357 3358static long needfree = 0; 3359 3360typedef enum free_memory_reason_t { 3361 FMR_UNKNOWN, 3362 FMR_NEEDFREE, 3363 FMR_LOTSFREE, 3364 FMR_SWAPFS_MINFREE, 3365 FMR_PAGES_PP_MAXIMUM, 3366 FMR_HEAP_ARENA, 3367 FMR_ZIO_ARENA, 3368 FMR_ZIO_FRAG, 3369} free_memory_reason_t; 3370 3371int64_t last_free_memory; 3372free_memory_reason_t last_free_reason; 3373 3374/* 3375 * Additional reserve of pages for pp_reserve. 3376 */ 3377int64_t arc_pages_pp_reserve = 64; 3378 3379/* 3380 * Additional reserve of pages for swapfs. 3381 */ 3382int64_t arc_swapfs_reserve = 64; 3383 3384/* 3385 * Return the amount of memory that can be consumed before reclaim will be 3386 * needed. Positive if there is sufficient free memory, negative indicates 3387 * the amount of memory that needs to be freed up. 3388 */ 3389static int64_t 3390arc_available_memory(void) 3391{ 3392 int64_t lowest = INT64_MAX; 3393 int64_t n; 3394 free_memory_reason_t r = FMR_UNKNOWN; 3395 3396#ifdef _KERNEL 3397 if (needfree > 0) { 3398 n = PAGESIZE * (-needfree); 3399 if (n < lowest) { 3400 lowest = n; 3401 r = FMR_NEEDFREE; 3402 } 3403 } 3404 3405 /* 3406 * Cooperate with pagedaemon when it's time for it to scan 3407 * and reclaim some pages. 3408 */ 3409 n = PAGESIZE * ((int64_t)freemem - zfs_arc_free_target); 3410 if (n < lowest) { 3411 lowest = n; 3412 r = FMR_LOTSFREE; 3413 } 3414 3415#ifdef illumos 3416 /* 3417 * check that we're out of range of the pageout scanner. It starts to 3418 * schedule paging if freemem is less than lotsfree and needfree. 3419 * lotsfree is the high-water mark for pageout, and needfree is the 3420 * number of needed free pages. We add extra pages here to make sure 3421 * the scanner doesn't start up while we're freeing memory. 3422 */ 3423 n = PAGESIZE * (freemem - lotsfree - needfree - desfree); 3424 if (n < lowest) { 3425 lowest = n; 3426 r = FMR_LOTSFREE; 3427 } 3428 3429 /* 3430 * check to make sure that swapfs has enough space so that anon 3431 * reservations can still succeed. anon_resvmem() checks that the 3432 * availrmem is greater than swapfs_minfree, and the number of reserved 3433 * swap pages. We also add a bit of extra here just to prevent 3434 * circumstances from getting really dire. 3435 */ 3436 n = PAGESIZE * (availrmem - swapfs_minfree - swapfs_reserve - 3437 desfree - arc_swapfs_reserve); 3438 if (n < lowest) { 3439 lowest = n; 3440 r = FMR_SWAPFS_MINFREE; 3441 } 3442 3443 3444 /* 3445 * Check that we have enough availrmem that memory locking (e.g., via 3446 * mlock(3C) or memcntl(2)) can still succeed. (pages_pp_maximum 3447 * stores the number of pages that cannot be locked; when availrmem 3448 * drops below pages_pp_maximum, page locking mechanisms such as 3449 * page_pp_lock() will fail.) 3450 */ 3451 n = PAGESIZE * (availrmem - pages_pp_maximum - 3452 arc_pages_pp_reserve); 3453 if (n < lowest) { 3454 lowest = n; 3455 r = FMR_PAGES_PP_MAXIMUM; 3456 } 3457 3458#endif /* illumos */ 3459#if defined(__i386) || !defined(UMA_MD_SMALL_ALLOC) 3460 /* 3461 * If we're on an i386 platform, it's possible that we'll exhaust the 3462 * kernel heap space before we ever run out of available physical 3463 * memory. Most checks of the size of the heap_area compare against 3464 * tune.t_minarmem, which is the minimum available real memory that we 3465 * can have in the system. However, this is generally fixed at 25 pages 3466 * which is so low that it's useless. In this comparison, we seek to 3467 * calculate the total heap-size, and reclaim if more than 3/4ths of the 3468 * heap is allocated. (Or, in the calculation, if less than 1/4th is 3469 * free) 3470 */ 3471 n = (int64_t)vmem_size(heap_arena, VMEM_FREE) - 3472 (vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC) >> 2); 3473 if (n < lowest) { 3474 lowest = n; 3475 r = FMR_HEAP_ARENA; 3476 } 3477#define zio_arena NULL 3478#else 3479#define zio_arena heap_arena 3480#endif 3481 3482 /* 3483 * If zio data pages are being allocated out of a separate heap segment, 3484 * then enforce that the size of available vmem for this arena remains 3485 * above about 1/16th free. 3486 * 3487 * Note: The 1/16th arena free requirement was put in place 3488 * to aggressively evict memory from the arc in order to avoid 3489 * memory fragmentation issues. 3490 */ 3491 if (zio_arena != NULL) { 3492 n = (int64_t)vmem_size(zio_arena, VMEM_FREE) - 3493 (vmem_size(zio_arena, VMEM_ALLOC) >> 4); 3494 if (n < lowest) { 3495 lowest = n; 3496 r = FMR_ZIO_ARENA; 3497 } 3498 } 3499 3500 /* 3501 * Above limits know nothing about real level of KVA fragmentation. 3502 * Start aggressive reclamation if too little sequential KVA left. 3503 */ 3504 if (lowest > 0) { 3505 n = (vmem_size(heap_arena, VMEM_MAXFREE) < zfs_max_recordsize) ? 3506 -((int64_t)vmem_size(heap_arena, VMEM_ALLOC) >> 4) : 3507 INT64_MAX; 3508 if (n < lowest) { 3509 lowest = n; 3510 r = FMR_ZIO_FRAG; 3511 } 3512 } 3513 3514#else /* _KERNEL */ 3515 /* Every 100 calls, free a small amount */ 3516 if (spa_get_random(100) == 0) 3517 lowest = -1024; 3518#endif /* _KERNEL */ 3519 3520 last_free_memory = lowest; 3521 last_free_reason = r; 3522 DTRACE_PROBE2(arc__available_memory, int64_t, lowest, int, r); 3523 return (lowest); 3524} 3525 3526 3527/* 3528 * Determine if the system is under memory pressure and is asking 3529 * to reclaim memory. A return value of TRUE indicates that the system 3530 * is under memory pressure and that the arc should adjust accordingly. 3531 */ 3532static boolean_t 3533arc_reclaim_needed(void) 3534{ 3535 return (arc_available_memory() < 0); 3536} 3537 3538extern kmem_cache_t *zio_buf_cache[]; 3539extern kmem_cache_t *zio_data_buf_cache[]; 3540extern kmem_cache_t *range_seg_cache; 3541 3542static __noinline void 3543arc_kmem_reap_now(void) 3544{ 3545 size_t i; 3546 kmem_cache_t *prev_cache = NULL; 3547 kmem_cache_t *prev_data_cache = NULL; 3548 3549 DTRACE_PROBE(arc__kmem_reap_start); 3550#ifdef _KERNEL 3551 if (arc_meta_used >= arc_meta_limit) { 3552 /* 3553 * We are exceeding our meta-data cache limit. 3554 * Purge some DNLC entries to release holds on meta-data. 3555 */ 3556 dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent); 3557 } 3558#if defined(__i386) 3559 /* 3560 * Reclaim unused memory from all kmem caches. 3561 */ 3562 kmem_reap(); 3563#endif 3564#endif 3565 3566 for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) { 3567 if (zio_buf_cache[i] != prev_cache) { 3568 prev_cache = zio_buf_cache[i]; 3569 kmem_cache_reap_now(zio_buf_cache[i]); 3570 } 3571 if (zio_data_buf_cache[i] != prev_data_cache) { 3572 prev_data_cache = zio_data_buf_cache[i]; 3573 kmem_cache_reap_now(zio_data_buf_cache[i]); 3574 } 3575 } 3576 kmem_cache_reap_now(buf_cache); 3577 kmem_cache_reap_now(hdr_full_cache); 3578 kmem_cache_reap_now(hdr_l2only_cache); 3579 kmem_cache_reap_now(range_seg_cache); 3580 3581#ifdef illumos 3582 if (zio_arena != NULL) { 3583 /* 3584 * Ask the vmem arena to reclaim unused memory from its 3585 * quantum caches. 3586 */ 3587 vmem_qcache_reap(zio_arena); 3588 } 3589#endif 3590 DTRACE_PROBE(arc__kmem_reap_end); 3591} 3592 3593/* 3594 * Threads can block in arc_get_data_buf() waiting for this thread to evict 3595 * enough data and signal them to proceed. When this happens, the threads in 3596 * arc_get_data_buf() are sleeping while holding the hash lock for their 3597 * particular arc header. Thus, we must be careful to never sleep on a 3598 * hash lock in this thread. This is to prevent the following deadlock: 3599 * 3600 * - Thread A sleeps on CV in arc_get_data_buf() holding hash lock "L", 3601 * waiting for the reclaim thread to signal it. 3602 * 3603 * - arc_reclaim_thread() tries to acquire hash lock "L" using mutex_enter, 3604 * fails, and goes to sleep forever. 3605 * 3606 * This possible deadlock is avoided by always acquiring a hash lock 3607 * using mutex_tryenter() from arc_reclaim_thread(). 3608 */ 3609static void 3610arc_reclaim_thread(void *dummy __unused) 3611{ 3612 clock_t growtime = 0; 3613 callb_cpr_t cpr; 3614 3615 CALLB_CPR_INIT(&cpr, &arc_reclaim_lock, callb_generic_cpr, FTAG); 3616 3617 mutex_enter(&arc_reclaim_lock); 3618 while (!arc_reclaim_thread_exit) { 3619 int64_t free_memory = arc_available_memory(); 3620 uint64_t evicted = 0; 3621 3622 mutex_exit(&arc_reclaim_lock); 3623 3624 if (free_memory < 0) { 3625 3626 arc_no_grow = B_TRUE; 3627 arc_warm = B_TRUE; 3628 3629 /* 3630 * Wait at least zfs_grow_retry (default 60) seconds 3631 * before considering growing. 3632 */ 3633 growtime = ddi_get_lbolt() + (arc_grow_retry * hz); 3634 3635 arc_kmem_reap_now(); 3636 3637 /* 3638 * If we are still low on memory, shrink the ARC 3639 * so that we have arc_shrink_min free space. 3640 */ 3641 free_memory = arc_available_memory(); 3642 3643 int64_t to_free = 3644 (arc_c >> arc_shrink_shift) - free_memory; 3645 if (to_free > 0) { 3646#ifdef _KERNEL 3647 to_free = MAX(to_free, ptob(needfree)); 3648#endif 3649 arc_shrink(to_free); 3650 } 3651 } else if (free_memory < arc_c >> arc_no_grow_shift) { 3652 arc_no_grow = B_TRUE; 3653 } else if (ddi_get_lbolt() >= growtime) { 3654 arc_no_grow = B_FALSE; 3655 } 3656 3657 evicted = arc_adjust(); 3658 3659 mutex_enter(&arc_reclaim_lock); 3660 3661 /* 3662 * If evicted is zero, we couldn't evict anything via 3663 * arc_adjust(). This could be due to hash lock 3664 * collisions, but more likely due to the majority of 3665 * arc buffers being unevictable. Therefore, even if 3666 * arc_size is above arc_c, another pass is unlikely to 3667 * be helpful and could potentially cause us to enter an 3668 * infinite loop. 3669 */ 3670 if (arc_size <= arc_c || evicted == 0) { 3671#ifdef _KERNEL 3672 needfree = 0; 3673#endif 3674 /* 3675 * We're either no longer overflowing, or we 3676 * can't evict anything more, so we should wake 3677 * up any threads before we go to sleep. 3678 */ 3679 cv_broadcast(&arc_reclaim_waiters_cv); 3680 3681 /* 3682 * Block until signaled, or after one second (we 3683 * might need to perform arc_kmem_reap_now() 3684 * even if we aren't being signalled) 3685 */ 3686 CALLB_CPR_SAFE_BEGIN(&cpr); 3687 (void) cv_timedwait(&arc_reclaim_thread_cv, 3688 &arc_reclaim_lock, hz); 3689 CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_lock); 3690 } 3691 } 3692 3693 arc_reclaim_thread_exit = FALSE; 3694 cv_broadcast(&arc_reclaim_thread_cv); 3695 CALLB_CPR_EXIT(&cpr); /* drops arc_reclaim_lock */ 3696 thread_exit(); 3697} 3698 3699static void 3700arc_user_evicts_thread(void *dummy __unused) 3701{ 3702 callb_cpr_t cpr; 3703 3704 CALLB_CPR_INIT(&cpr, &arc_user_evicts_lock, callb_generic_cpr, FTAG); 3705 3706 mutex_enter(&arc_user_evicts_lock); 3707 while (!arc_user_evicts_thread_exit) { 3708 mutex_exit(&arc_user_evicts_lock); 3709 3710 arc_do_user_evicts(); 3711 3712 /* 3713 * This is necessary in order for the mdb ::arc dcmd to 3714 * show up to date information. Since the ::arc command 3715 * does not call the kstat's update function, without 3716 * this call, the command may show stale stats for the 3717 * anon, mru, mru_ghost, mfu, and mfu_ghost lists. Even 3718 * with this change, the data might be up to 1 second 3719 * out of date; but that should suffice. The arc_state_t 3720 * structures can be queried directly if more accurate 3721 * information is needed. 3722 */ 3723 if (arc_ksp != NULL) 3724 arc_ksp->ks_update(arc_ksp, KSTAT_READ); 3725 3726 mutex_enter(&arc_user_evicts_lock); 3727 3728 /* 3729 * Block until signaled, or after one second (we need to 3730 * call the arc's kstat update function regularly). 3731 */ 3732 CALLB_CPR_SAFE_BEGIN(&cpr); 3733 (void) cv_timedwait(&arc_user_evicts_cv, 3734 &arc_user_evicts_lock, hz); 3735 CALLB_CPR_SAFE_END(&cpr, &arc_user_evicts_lock); 3736 } 3737 3738 arc_user_evicts_thread_exit = FALSE; 3739 cv_broadcast(&arc_user_evicts_cv); 3740 CALLB_CPR_EXIT(&cpr); /* drops arc_user_evicts_lock */ 3741 thread_exit(); 3742} 3743 3744/* 3745 * Adapt arc info given the number of bytes we are trying to add and 3746 * the state that we are comming from. This function is only called 3747 * when we are adding new content to the cache. 3748 */ 3749static void 3750arc_adapt(int bytes, arc_state_t *state) 3751{ 3752 int mult; 3753 uint64_t arc_p_min = (arc_c >> arc_p_min_shift); 3754 int64_t mrug_size = refcount_count(&arc_mru_ghost->arcs_size); 3755 int64_t mfug_size = refcount_count(&arc_mfu_ghost->arcs_size); 3756 3757 if (state == arc_l2c_only) 3758 return; 3759 3760 ASSERT(bytes > 0); 3761 /* 3762 * Adapt the target size of the MRU list: 3763 * - if we just hit in the MRU ghost list, then increase 3764 * the target size of the MRU list. 3765 * - if we just hit in the MFU ghost list, then increase 3766 * the target size of the MFU list by decreasing the 3767 * target size of the MRU list. 3768 */ 3769 if (state == arc_mru_ghost) { 3770 mult = (mrug_size >= mfug_size) ? 1 : (mfug_size / mrug_size); 3771 mult = MIN(mult, 10); /* avoid wild arc_p adjustment */ 3772 3773 arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult); 3774 } else if (state == arc_mfu_ghost) { 3775 uint64_t delta; 3776 3777 mult = (mfug_size >= mrug_size) ? 1 : (mrug_size / mfug_size); 3778 mult = MIN(mult, 10); 3779 3780 delta = MIN(bytes * mult, arc_p); 3781 arc_p = MAX(arc_p_min, arc_p - delta); 3782 } 3783 ASSERT((int64_t)arc_p >= 0); 3784 3785 if (arc_reclaim_needed()) { 3786 cv_signal(&arc_reclaim_thread_cv); 3787 return; 3788 } 3789 3790 if (arc_no_grow) 3791 return; 3792 3793 if (arc_c >= arc_c_max) 3794 return; 3795 3796 /* 3797 * If we're within (2 * maxblocksize) bytes of the target 3798 * cache size, increment the target cache size 3799 */ 3800 if (arc_size > arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) { 3801 DTRACE_PROBE1(arc__inc_adapt, int, bytes); 3802 atomic_add_64(&arc_c, (int64_t)bytes); 3803 if (arc_c > arc_c_max) 3804 arc_c = arc_c_max; 3805 else if (state == arc_anon) 3806 atomic_add_64(&arc_p, (int64_t)bytes); 3807 if (arc_p > arc_c) 3808 arc_p = arc_c; 3809 } 3810 ASSERT((int64_t)arc_p >= 0); 3811} 3812 3813/* 3814 * Check if arc_size has grown past our upper threshold, determined by 3815 * zfs_arc_overflow_shift. 3816 */ 3817static boolean_t 3818arc_is_overflowing(void) 3819{ 3820 /* Always allow at least one block of overflow */ 3821 uint64_t overflow = MAX(SPA_MAXBLOCKSIZE, 3822 arc_c >> zfs_arc_overflow_shift); 3823 3824 return (arc_size >= arc_c + overflow); 3825} 3826 3827/* 3828 * The buffer, supplied as the first argument, needs a data block. If we 3829 * are hitting the hard limit for the cache size, we must sleep, waiting 3830 * for the eviction thread to catch up. If we're past the target size 3831 * but below the hard limit, we'll only signal the reclaim thread and 3832 * continue on. 3833 */ 3834static void 3835arc_get_data_buf(arc_buf_t *buf) 3836{ 3837 arc_state_t *state = buf->b_hdr->b_l1hdr.b_state; 3838 uint64_t size = buf->b_hdr->b_size; 3839 arc_buf_contents_t type = arc_buf_type(buf->b_hdr); 3840 3841 arc_adapt(size, state); 3842 3843 /* 3844 * If arc_size is currently overflowing, and has grown past our 3845 * upper limit, we must be adding data faster than the evict 3846 * thread can evict. Thus, to ensure we don't compound the 3847 * problem by adding more data and forcing arc_size to grow even 3848 * further past it's target size, we halt and wait for the 3849 * eviction thread to catch up. 3850 * 3851 * It's also possible that the reclaim thread is unable to evict 3852 * enough buffers to get arc_size below the overflow limit (e.g. 3853 * due to buffers being un-evictable, or hash lock collisions). 3854 * In this case, we want to proceed regardless if we're 3855 * overflowing; thus we don't use a while loop here. 3856 */ 3857 if (arc_is_overflowing()) { 3858 mutex_enter(&arc_reclaim_lock); 3859 3860 /* 3861 * Now that we've acquired the lock, we may no longer be 3862 * over the overflow limit, lets check. 3863 * 3864 * We're ignoring the case of spurious wake ups. If that 3865 * were to happen, it'd let this thread consume an ARC 3866 * buffer before it should have (i.e. before we're under 3867 * the overflow limit and were signalled by the reclaim 3868 * thread). As long as that is a rare occurrence, it 3869 * shouldn't cause any harm. 3870 */ 3871 if (arc_is_overflowing()) { 3872 cv_signal(&arc_reclaim_thread_cv); 3873 cv_wait(&arc_reclaim_waiters_cv, &arc_reclaim_lock); 3874 } 3875 3876 mutex_exit(&arc_reclaim_lock); 3877 } 3878 3879 if (type == ARC_BUFC_METADATA) { 3880 buf->b_data = zio_buf_alloc(size); 3881 arc_space_consume(size, ARC_SPACE_META); 3882 } else { 3883 ASSERT(type == ARC_BUFC_DATA); 3884 buf->b_data = zio_data_buf_alloc(size); 3885 arc_space_consume(size, ARC_SPACE_DATA); 3886 } 3887 3888 /* 3889 * Update the state size. Note that ghost states have a 3890 * "ghost size" and so don't need to be updated. 3891 */ 3892 if (!GHOST_STATE(buf->b_hdr->b_l1hdr.b_state)) { 3893 arc_buf_hdr_t *hdr = buf->b_hdr; 3894 arc_state_t *state = hdr->b_l1hdr.b_state; 3895 3896 (void) refcount_add_many(&state->arcs_size, size, buf); 3897 3898 /* 3899 * If this is reached via arc_read, the link is 3900 * protected by the hash lock. If reached via 3901 * arc_buf_alloc, the header should not be accessed by 3902 * any other thread. And, if reached via arc_read_done, 3903 * the hash lock will protect it if it's found in the 3904 * hash table; otherwise no other thread should be 3905 * trying to [add|remove]_reference it. 3906 */ 3907 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) { 3908 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 3909 atomic_add_64(&hdr->b_l1hdr.b_state->arcs_lsize[type], 3910 size); 3911 } 3912 /* 3913 * If we are growing the cache, and we are adding anonymous 3914 * data, and we have outgrown arc_p, update arc_p 3915 */ 3916 if (arc_size < arc_c && hdr->b_l1hdr.b_state == arc_anon && 3917 (refcount_count(&arc_anon->arcs_size) + 3918 refcount_count(&arc_mru->arcs_size) > arc_p)) 3919 arc_p = MIN(arc_c, arc_p + size); 3920 } 3921 ARCSTAT_BUMP(arcstat_allocated); 3922} 3923 3924/* 3925 * This routine is called whenever a buffer is accessed. 3926 * NOTE: the hash lock is dropped in this function. 3927 */ 3928static void 3929arc_access(arc_buf_hdr_t *hdr, kmutex_t *hash_lock) 3930{ 3931 clock_t now; 3932 3933 ASSERT(MUTEX_HELD(hash_lock)); 3934 ASSERT(HDR_HAS_L1HDR(hdr)); 3935 3936 if (hdr->b_l1hdr.b_state == arc_anon) { 3937 /* 3938 * This buffer is not in the cache, and does not 3939 * appear in our "ghost" list. Add the new buffer 3940 * to the MRU state. 3941 */ 3942 3943 ASSERT0(hdr->b_l1hdr.b_arc_access); 3944 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 3945 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr); 3946 arc_change_state(arc_mru, hdr, hash_lock); 3947 3948 } else if (hdr->b_l1hdr.b_state == arc_mru) { 3949 now = ddi_get_lbolt(); 3950 3951 /* 3952 * If this buffer is here because of a prefetch, then either: 3953 * - clear the flag if this is a "referencing" read 3954 * (any subsequent access will bump this into the MFU state). 3955 * or 3956 * - move the buffer to the head of the list if this is 3957 * another prefetch (to make it less likely to be evicted). 3958 */ 3959 if (HDR_PREFETCH(hdr)) { 3960 if (refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) { 3961 /* link protected by hash lock */ 3962 ASSERT(multilist_link_active( 3963 &hdr->b_l1hdr.b_arc_node)); 3964 } else { 3965 hdr->b_flags &= ~ARC_FLAG_PREFETCH; 3966 ARCSTAT_BUMP(arcstat_mru_hits); 3967 } 3968 hdr->b_l1hdr.b_arc_access = now; 3969 return; 3970 } 3971 3972 /* 3973 * This buffer has been "accessed" only once so far, 3974 * but it is still in the cache. Move it to the MFU 3975 * state. 3976 */ 3977 if (now > hdr->b_l1hdr.b_arc_access + ARC_MINTIME) { 3978 /* 3979 * More than 125ms have passed since we 3980 * instantiated this buffer. Move it to the 3981 * most frequently used state. 3982 */ 3983 hdr->b_l1hdr.b_arc_access = now; 3984 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 3985 arc_change_state(arc_mfu, hdr, hash_lock); 3986 } 3987 ARCSTAT_BUMP(arcstat_mru_hits); 3988 } else if (hdr->b_l1hdr.b_state == arc_mru_ghost) { 3989 arc_state_t *new_state; 3990 /* 3991 * This buffer has been "accessed" recently, but 3992 * was evicted from the cache. Move it to the 3993 * MFU state. 3994 */ 3995 3996 if (HDR_PREFETCH(hdr)) { 3997 new_state = arc_mru; 3998 if (refcount_count(&hdr->b_l1hdr.b_refcnt) > 0) 3999 hdr->b_flags &= ~ARC_FLAG_PREFETCH; 4000 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr); 4001 } else { 4002 new_state = arc_mfu; 4003 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 4004 } 4005 4006 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 4007 arc_change_state(new_state, hdr, hash_lock); 4008 4009 ARCSTAT_BUMP(arcstat_mru_ghost_hits); 4010 } else if (hdr->b_l1hdr.b_state == arc_mfu) { 4011 /* 4012 * This buffer has been accessed more than once and is 4013 * still in the cache. Keep it in the MFU state. 4014 * 4015 * NOTE: an add_reference() that occurred when we did 4016 * the arc_read() will have kicked this off the list. 4017 * If it was a prefetch, we will explicitly move it to 4018 * the head of the list now. 4019 */ 4020 if ((HDR_PREFETCH(hdr)) != 0) { 4021 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 4022 /* link protected by hash_lock */ 4023 ASSERT(multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 4024 } 4025 ARCSTAT_BUMP(arcstat_mfu_hits); 4026 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 4027 } else if (hdr->b_l1hdr.b_state == arc_mfu_ghost) { 4028 arc_state_t *new_state = arc_mfu; 4029 /* 4030 * This buffer has been accessed more than once but has 4031 * been evicted from the cache. Move it back to the 4032 * MFU state. 4033 */ 4034 4035 if (HDR_PREFETCH(hdr)) { 4036 /* 4037 * This is a prefetch access... 4038 * move this block back to the MRU state. 4039 */ 4040 ASSERT0(refcount_count(&hdr->b_l1hdr.b_refcnt)); 4041 new_state = arc_mru; 4042 } 4043 4044 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 4045 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 4046 arc_change_state(new_state, hdr, hash_lock); 4047 4048 ARCSTAT_BUMP(arcstat_mfu_ghost_hits); 4049 } else if (hdr->b_l1hdr.b_state == arc_l2c_only) { 4050 /* 4051 * This buffer is on the 2nd Level ARC. 4052 */ 4053 4054 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 4055 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 4056 arc_change_state(arc_mfu, hdr, hash_lock); 4057 } else { 4058 ASSERT(!"invalid arc state"); 4059 } 4060} 4061 4062/* a generic arc_done_func_t which you can use */ 4063/* ARGSUSED */ 4064void 4065arc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg) 4066{ 4067 if (zio == NULL || zio->io_error == 0) 4068 bcopy(buf->b_data, arg, buf->b_hdr->b_size); 4069 VERIFY(arc_buf_remove_ref(buf, arg)); 4070} 4071 4072/* a generic arc_done_func_t */ 4073void 4074arc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg) 4075{ 4076 arc_buf_t **bufp = arg; 4077 if (zio && zio->io_error) { 4078 VERIFY(arc_buf_remove_ref(buf, arg)); 4079 *bufp = NULL; 4080 } else { 4081 *bufp = buf; 4082 ASSERT(buf->b_data); 4083 } 4084} 4085 4086static void 4087arc_read_done(zio_t *zio) 4088{ 4089 arc_buf_hdr_t *hdr; 4090 arc_buf_t *buf; 4091 arc_buf_t *abuf; /* buffer we're assigning to callback */ 4092 kmutex_t *hash_lock = NULL; 4093 arc_callback_t *callback_list, *acb; 4094 int freeable = FALSE; 4095 4096 buf = zio->io_private; 4097 hdr = buf->b_hdr; 4098 4099 /* 4100 * The hdr was inserted into hash-table and removed from lists 4101 * prior to starting I/O. We should find this header, since 4102 * it's in the hash table, and it should be legit since it's 4103 * not possible to evict it during the I/O. The only possible 4104 * reason for it not to be found is if we were freed during the 4105 * read. 4106 */ 4107 if (HDR_IN_HASH_TABLE(hdr)) { 4108 ASSERT3U(hdr->b_birth, ==, BP_PHYSICAL_BIRTH(zio->io_bp)); 4109 ASSERT3U(hdr->b_dva.dva_word[0], ==, 4110 BP_IDENTITY(zio->io_bp)->dva_word[0]); 4111 ASSERT3U(hdr->b_dva.dva_word[1], ==, 4112 BP_IDENTITY(zio->io_bp)->dva_word[1]); 4113 4114 arc_buf_hdr_t *found = buf_hash_find(hdr->b_spa, zio->io_bp, 4115 &hash_lock); 4116 4117 ASSERT((found == NULL && HDR_FREED_IN_READ(hdr) && 4118 hash_lock == NULL) || 4119 (found == hdr && 4120 DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) || 4121 (found == hdr && HDR_L2_READING(hdr))); 4122 } 4123 4124 hdr->b_flags &= ~ARC_FLAG_L2_EVICTED; 4125 if (l2arc_noprefetch && HDR_PREFETCH(hdr)) 4126 hdr->b_flags &= ~ARC_FLAG_L2CACHE; 4127 4128 /* byteswap if necessary */ 4129 callback_list = hdr->b_l1hdr.b_acb; 4130 ASSERT(callback_list != NULL); 4131 if (BP_SHOULD_BYTESWAP(zio->io_bp) && zio->io_error == 0) { 4132 dmu_object_byteswap_t bswap = 4133 DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp)); 4134 arc_byteswap_func_t *func = BP_GET_LEVEL(zio->io_bp) > 0 ? 4135 byteswap_uint64_array : 4136 dmu_ot_byteswap[bswap].ob_func; 4137 func(buf->b_data, hdr->b_size); 4138 } 4139 4140 arc_cksum_compute(buf, B_FALSE); 4141#ifdef illumos 4142 arc_buf_watch(buf); 4143#endif 4144 4145 if (hash_lock && zio->io_error == 0 && 4146 hdr->b_l1hdr.b_state == arc_anon) { 4147 /* 4148 * Only call arc_access on anonymous buffers. This is because 4149 * if we've issued an I/O for an evicted buffer, we've already 4150 * called arc_access (to prevent any simultaneous readers from 4151 * getting confused). 4152 */ 4153 arc_access(hdr, hash_lock); 4154 } 4155 4156 /* create copies of the data buffer for the callers */ 4157 abuf = buf; 4158 for (acb = callback_list; acb; acb = acb->acb_next) { 4159 if (acb->acb_done) { 4160 if (abuf == NULL) { 4161 ARCSTAT_BUMP(arcstat_duplicate_reads); 4162 abuf = arc_buf_clone(buf); 4163 } 4164 acb->acb_buf = abuf; 4165 abuf = NULL; 4166 } 4167 } 4168 hdr->b_l1hdr.b_acb = NULL; 4169 hdr->b_flags &= ~ARC_FLAG_IO_IN_PROGRESS; 4170 ASSERT(!HDR_BUF_AVAILABLE(hdr)); 4171 if (abuf == buf) { 4172 ASSERT(buf->b_efunc == NULL); 4173 ASSERT(hdr->b_l1hdr.b_datacnt == 1); 4174 hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE; 4175 } 4176 4177 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt) || 4178 callback_list != NULL); 4179 4180 if (zio->io_error != 0) { 4181 hdr->b_flags |= ARC_FLAG_IO_ERROR; 4182 if (hdr->b_l1hdr.b_state != arc_anon) 4183 arc_change_state(arc_anon, hdr, hash_lock); 4184 if (HDR_IN_HASH_TABLE(hdr)) 4185 buf_hash_remove(hdr); 4186 freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt); 4187 } 4188 4189 /* 4190 * Broadcast before we drop the hash_lock to avoid the possibility 4191 * that the hdr (and hence the cv) might be freed before we get to 4192 * the cv_broadcast(). 4193 */ 4194 cv_broadcast(&hdr->b_l1hdr.b_cv); 4195 4196 if (hash_lock != NULL) { 4197 mutex_exit(hash_lock); 4198 } else { 4199 /* 4200 * This block was freed while we waited for the read to 4201 * complete. It has been removed from the hash table and 4202 * moved to the anonymous state (so that it won't show up 4203 * in the cache). 4204 */ 4205 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon); 4206 freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt); 4207 } 4208 4209 /* execute each callback and free its structure */ 4210 while ((acb = callback_list) != NULL) { 4211 if (acb->acb_done) 4212 acb->acb_done(zio, acb->acb_buf, acb->acb_private); 4213 4214 if (acb->acb_zio_dummy != NULL) { 4215 acb->acb_zio_dummy->io_error = zio->io_error; 4216 zio_nowait(acb->acb_zio_dummy); 4217 } 4218 4219 callback_list = acb->acb_next; 4220 kmem_free(acb, sizeof (arc_callback_t)); 4221 } 4222 4223 if (freeable) 4224 arc_hdr_destroy(hdr); 4225} 4226 4227/* 4228 * "Read" the block at the specified DVA (in bp) via the 4229 * cache. If the block is found in the cache, invoke the provided 4230 * callback immediately and return. Note that the `zio' parameter 4231 * in the callback will be NULL in this case, since no IO was 4232 * required. If the block is not in the cache pass the read request 4233 * on to the spa with a substitute callback function, so that the 4234 * requested block will be added to the cache. 4235 * 4236 * If a read request arrives for a block that has a read in-progress, 4237 * either wait for the in-progress read to complete (and return the 4238 * results); or, if this is a read with a "done" func, add a record 4239 * to the read to invoke the "done" func when the read completes, 4240 * and return; or just return. 4241 * 4242 * arc_read_done() will invoke all the requested "done" functions 4243 * for readers of this block. 4244 */ 4245int 4246arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_done_func_t *done, 4247 void *private, zio_priority_t priority, int zio_flags, 4248 arc_flags_t *arc_flags, const zbookmark_phys_t *zb) 4249{ 4250 arc_buf_hdr_t *hdr = NULL; 4251 arc_buf_t *buf = NULL; 4252 kmutex_t *hash_lock = NULL; 4253 zio_t *rzio; 4254 uint64_t guid = spa_load_guid(spa); 4255 4256 ASSERT(!BP_IS_EMBEDDED(bp) || 4257 BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA); 4258 4259top: 4260 if (!BP_IS_EMBEDDED(bp)) { 4261 /* 4262 * Embedded BP's have no DVA and require no I/O to "read". 4263 * Create an anonymous arc buf to back it. 4264 */ 4265 hdr = buf_hash_find(guid, bp, &hash_lock); 4266 } 4267 4268 if (hdr != NULL && HDR_HAS_L1HDR(hdr) && hdr->b_l1hdr.b_datacnt > 0) { 4269 4270 *arc_flags |= ARC_FLAG_CACHED; 4271 4272 if (HDR_IO_IN_PROGRESS(hdr)) { 4273 4274 if ((hdr->b_flags & ARC_FLAG_PRIO_ASYNC_READ) && 4275 priority == ZIO_PRIORITY_SYNC_READ) { 4276 /* 4277 * This sync read must wait for an 4278 * in-progress async read (e.g. a predictive 4279 * prefetch). Async reads are queued 4280 * separately at the vdev_queue layer, so 4281 * this is a form of priority inversion. 4282 * Ideally, we would "inherit" the demand 4283 * i/o's priority by moving the i/o from 4284 * the async queue to the synchronous queue, 4285 * but there is currently no mechanism to do 4286 * so. Track this so that we can evaluate 4287 * the magnitude of this potential performance 4288 * problem. 4289 * 4290 * Note that if the prefetch i/o is already 4291 * active (has been issued to the device), 4292 * the prefetch improved performance, because 4293 * we issued it sooner than we would have 4294 * without the prefetch. 4295 */ 4296 DTRACE_PROBE1(arc__sync__wait__for__async, 4297 arc_buf_hdr_t *, hdr); 4298 ARCSTAT_BUMP(arcstat_sync_wait_for_async); 4299 } 4300 if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) { 4301 hdr->b_flags &= ~ARC_FLAG_PREDICTIVE_PREFETCH; 4302 } 4303 4304 if (*arc_flags & ARC_FLAG_WAIT) { 4305 cv_wait(&hdr->b_l1hdr.b_cv, hash_lock); 4306 mutex_exit(hash_lock); 4307 goto top; 4308 } 4309 ASSERT(*arc_flags & ARC_FLAG_NOWAIT); 4310 4311 if (done) { 4312 arc_callback_t *acb = NULL; 4313 4314 acb = kmem_zalloc(sizeof (arc_callback_t), 4315 KM_SLEEP); 4316 acb->acb_done = done; 4317 acb->acb_private = private; 4318 if (pio != NULL) 4319 acb->acb_zio_dummy = zio_null(pio, 4320 spa, NULL, NULL, NULL, zio_flags); 4321 4322 ASSERT(acb->acb_done != NULL); 4323 acb->acb_next = hdr->b_l1hdr.b_acb; 4324 hdr->b_l1hdr.b_acb = acb; 4325 add_reference(hdr, hash_lock, private); 4326 mutex_exit(hash_lock); 4327 return (0); 4328 } 4329 mutex_exit(hash_lock); 4330 return (0); 4331 } 4332 4333 ASSERT(hdr->b_l1hdr.b_state == arc_mru || 4334 hdr->b_l1hdr.b_state == arc_mfu); 4335 4336 if (done) { 4337 if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) { 4338 /* 4339 * This is a demand read which does not have to 4340 * wait for i/o because we did a predictive 4341 * prefetch i/o for it, which has completed. 4342 */ 4343 DTRACE_PROBE1( 4344 arc__demand__hit__predictive__prefetch, 4345 arc_buf_hdr_t *, hdr); 4346 ARCSTAT_BUMP( 4347 arcstat_demand_hit_predictive_prefetch); 4348 hdr->b_flags &= ~ARC_FLAG_PREDICTIVE_PREFETCH; 4349 } 4350 add_reference(hdr, hash_lock, private); 4351 /* 4352 * If this block is already in use, create a new 4353 * copy of the data so that we will be guaranteed 4354 * that arc_release() will always succeed. 4355 */ 4356 buf = hdr->b_l1hdr.b_buf; 4357 ASSERT(buf); 4358 ASSERT(buf->b_data); 4359 if (HDR_BUF_AVAILABLE(hdr)) { 4360 ASSERT(buf->b_efunc == NULL); 4361 hdr->b_flags &= ~ARC_FLAG_BUF_AVAILABLE; 4362 } else { 4363 buf = arc_buf_clone(buf); 4364 } 4365 4366 } else if (*arc_flags & ARC_FLAG_PREFETCH && 4367 refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) { 4368 hdr->b_flags |= ARC_FLAG_PREFETCH; 4369 } 4370 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr); 4371 arc_access(hdr, hash_lock); 4372 if (*arc_flags & ARC_FLAG_L2CACHE) 4373 hdr->b_flags |= ARC_FLAG_L2CACHE; 4374 if (*arc_flags & ARC_FLAG_L2COMPRESS) 4375 hdr->b_flags |= ARC_FLAG_L2COMPRESS; 4376 mutex_exit(hash_lock); 4377 ARCSTAT_BUMP(arcstat_hits); 4378 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr), 4379 demand, prefetch, !HDR_ISTYPE_METADATA(hdr), 4380 data, metadata, hits); 4381 4382 if (done) 4383 done(NULL, buf, private); 4384 } else { 4385 uint64_t size = BP_GET_LSIZE(bp); 4386 arc_callback_t *acb; 4387 vdev_t *vd = NULL; 4388 uint64_t addr = 0; 4389 boolean_t devw = B_FALSE; 4390 enum zio_compress b_compress = ZIO_COMPRESS_OFF; 4391 int32_t b_asize = 0; 4392 4393 if (hdr == NULL) { 4394 /* this block is not in the cache */ 4395 arc_buf_hdr_t *exists = NULL; 4396 arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp); 4397 buf = arc_buf_alloc(spa, size, private, type); 4398 hdr = buf->b_hdr; 4399 if (!BP_IS_EMBEDDED(bp)) { 4400 hdr->b_dva = *BP_IDENTITY(bp); 4401 hdr->b_birth = BP_PHYSICAL_BIRTH(bp); 4402 exists = buf_hash_insert(hdr, &hash_lock); 4403 } 4404 if (exists != NULL) { 4405 /* somebody beat us to the hash insert */ 4406 mutex_exit(hash_lock); 4407 buf_discard_identity(hdr); 4408 (void) arc_buf_remove_ref(buf, private); 4409 goto top; /* restart the IO request */ 4410 } 4411 4412 /* 4413 * If there is a callback, we pass our reference to 4414 * it; otherwise we remove our reference. 4415 */ 4416 if (done == NULL) { 4417 (void) remove_reference(hdr, hash_lock, 4418 private); 4419 } 4420 if (*arc_flags & ARC_FLAG_PREFETCH) 4421 hdr->b_flags |= ARC_FLAG_PREFETCH; 4422 if (*arc_flags & ARC_FLAG_L2CACHE) 4423 hdr->b_flags |= ARC_FLAG_L2CACHE; 4424 if (*arc_flags & ARC_FLAG_L2COMPRESS) 4425 hdr->b_flags |= ARC_FLAG_L2COMPRESS; 4426 if (BP_GET_LEVEL(bp) > 0) 4427 hdr->b_flags |= ARC_FLAG_INDIRECT; 4428 } else { 4429 /* 4430 * This block is in the ghost cache. If it was L2-only 4431 * (and thus didn't have an L1 hdr), we realloc the 4432 * header to add an L1 hdr. 4433 */ 4434 if (!HDR_HAS_L1HDR(hdr)) { 4435 hdr = arc_hdr_realloc(hdr, hdr_l2only_cache, 4436 hdr_full_cache); 4437 } 4438 4439 ASSERT(GHOST_STATE(hdr->b_l1hdr.b_state)); 4440 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 4441 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 4442 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 4443 4444 /* 4445 * If there is a callback, we pass a reference to it. 4446 */ 4447 if (done != NULL) 4448 add_reference(hdr, hash_lock, private); 4449 if (*arc_flags & ARC_FLAG_PREFETCH) 4450 hdr->b_flags |= ARC_FLAG_PREFETCH; 4451 if (*arc_flags & ARC_FLAG_L2CACHE) 4452 hdr->b_flags |= ARC_FLAG_L2CACHE; 4453 if (*arc_flags & ARC_FLAG_L2COMPRESS) 4454 hdr->b_flags |= ARC_FLAG_L2COMPRESS; 4455 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 4456 buf->b_hdr = hdr; 4457 buf->b_data = NULL; 4458 buf->b_efunc = NULL; 4459 buf->b_private = NULL; 4460 buf->b_next = NULL; 4461 hdr->b_l1hdr.b_buf = buf; 4462 ASSERT0(hdr->b_l1hdr.b_datacnt); 4463 hdr->b_l1hdr.b_datacnt = 1; 4464 arc_get_data_buf(buf); 4465 arc_access(hdr, hash_lock); 4466 } 4467 4468 if (*arc_flags & ARC_FLAG_PREDICTIVE_PREFETCH) 4469 hdr->b_flags |= ARC_FLAG_PREDICTIVE_PREFETCH; 4470 ASSERT(!GHOST_STATE(hdr->b_l1hdr.b_state)); 4471 4472 acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP); 4473 acb->acb_done = done; 4474 acb->acb_private = private; 4475 4476 ASSERT(hdr->b_l1hdr.b_acb == NULL); 4477 hdr->b_l1hdr.b_acb = acb; 4478 hdr->b_flags |= ARC_FLAG_IO_IN_PROGRESS; 4479 4480 if (HDR_HAS_L2HDR(hdr) && 4481 (vd = hdr->b_l2hdr.b_dev->l2ad_vdev) != NULL) { 4482 devw = hdr->b_l2hdr.b_dev->l2ad_writing; 4483 addr = hdr->b_l2hdr.b_daddr; 4484 b_compress = hdr->b_l2hdr.b_compress; 4485 b_asize = hdr->b_l2hdr.b_asize; 4486 /* 4487 * Lock out device removal. 4488 */ 4489 if (vdev_is_dead(vd) || 4490 !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER)) 4491 vd = NULL; 4492 } 4493 4494 if (hash_lock != NULL) 4495 mutex_exit(hash_lock); 4496 4497 /* 4498 * At this point, we have a level 1 cache miss. Try again in 4499 * L2ARC if possible. 4500 */ 4501 ASSERT3U(hdr->b_size, ==, size); 4502 DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp, 4503 uint64_t, size, zbookmark_phys_t *, zb); 4504 ARCSTAT_BUMP(arcstat_misses); 4505 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr), 4506 demand, prefetch, !HDR_ISTYPE_METADATA(hdr), 4507 data, metadata, misses); 4508#ifdef _KERNEL 4509 curthread->td_ru.ru_inblock++; 4510#endif 4511 4512 if (priority == ZIO_PRIORITY_ASYNC_READ) 4513 hdr->b_flags |= ARC_FLAG_PRIO_ASYNC_READ; 4514 else 4515 hdr->b_flags &= ~ARC_FLAG_PRIO_ASYNC_READ; 4516 4517 if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) { 4518 /* 4519 * Read from the L2ARC if the following are true: 4520 * 1. The L2ARC vdev was previously cached. 4521 * 2. This buffer still has L2ARC metadata. 4522 * 3. This buffer isn't currently writing to the L2ARC. 4523 * 4. The L2ARC entry wasn't evicted, which may 4524 * also have invalidated the vdev. 4525 * 5. This isn't prefetch and l2arc_noprefetch is set. 4526 */ 4527 if (HDR_HAS_L2HDR(hdr) && 4528 !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) && 4529 !(l2arc_noprefetch && HDR_PREFETCH(hdr))) { 4530 l2arc_read_callback_t *cb; 4531 4532 DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr); 4533 ARCSTAT_BUMP(arcstat_l2_hits); 4534 4535 cb = kmem_zalloc(sizeof (l2arc_read_callback_t), 4536 KM_SLEEP); 4537 cb->l2rcb_buf = buf; 4538 cb->l2rcb_spa = spa; 4539 cb->l2rcb_bp = *bp; 4540 cb->l2rcb_zb = *zb; 4541 cb->l2rcb_flags = zio_flags; 4542 cb->l2rcb_compress = b_compress; 4543 4544 ASSERT(addr >= VDEV_LABEL_START_SIZE && 4545 addr + size < vd->vdev_psize - 4546 VDEV_LABEL_END_SIZE); 4547 4548 /* 4549 * l2arc read. The SCL_L2ARC lock will be 4550 * released by l2arc_read_done(). 4551 * Issue a null zio if the underlying buffer 4552 * was squashed to zero size by compression. 4553 */ 4554 if (b_compress == ZIO_COMPRESS_EMPTY) { 4555 rzio = zio_null(pio, spa, vd, 4556 l2arc_read_done, cb, 4557 zio_flags | ZIO_FLAG_DONT_CACHE | 4558 ZIO_FLAG_CANFAIL | 4559 ZIO_FLAG_DONT_PROPAGATE | 4560 ZIO_FLAG_DONT_RETRY); 4561 } else { 4562 rzio = zio_read_phys(pio, vd, addr, 4563 b_asize, buf->b_data, 4564 ZIO_CHECKSUM_OFF, 4565 l2arc_read_done, cb, priority, 4566 zio_flags | ZIO_FLAG_DONT_CACHE | 4567 ZIO_FLAG_CANFAIL | 4568 ZIO_FLAG_DONT_PROPAGATE | 4569 ZIO_FLAG_DONT_RETRY, B_FALSE); 4570 } 4571 DTRACE_PROBE2(l2arc__read, vdev_t *, vd, 4572 zio_t *, rzio); 4573 ARCSTAT_INCR(arcstat_l2_read_bytes, b_asize); 4574 4575 if (*arc_flags & ARC_FLAG_NOWAIT) { 4576 zio_nowait(rzio); 4577 return (0); 4578 } 4579 4580 ASSERT(*arc_flags & ARC_FLAG_WAIT); 4581 if (zio_wait(rzio) == 0) 4582 return (0); 4583 4584 /* l2arc read error; goto zio_read() */ 4585 } else { 4586 DTRACE_PROBE1(l2arc__miss, 4587 arc_buf_hdr_t *, hdr); 4588 ARCSTAT_BUMP(arcstat_l2_misses); 4589 if (HDR_L2_WRITING(hdr)) 4590 ARCSTAT_BUMP(arcstat_l2_rw_clash); 4591 spa_config_exit(spa, SCL_L2ARC, vd); 4592 } 4593 } else { 4594 if (vd != NULL) 4595 spa_config_exit(spa, SCL_L2ARC, vd); 4596 if (l2arc_ndev != 0) { 4597 DTRACE_PROBE1(l2arc__miss, 4598 arc_buf_hdr_t *, hdr); 4599 ARCSTAT_BUMP(arcstat_l2_misses); 4600 } 4601 } 4602 4603 rzio = zio_read(pio, spa, bp, buf->b_data, size, 4604 arc_read_done, buf, priority, zio_flags, zb); 4605 4606 if (*arc_flags & ARC_FLAG_WAIT) 4607 return (zio_wait(rzio)); 4608 4609 ASSERT(*arc_flags & ARC_FLAG_NOWAIT); 4610 zio_nowait(rzio); 4611 } 4612 return (0); 4613} 4614 4615void 4616arc_set_callback(arc_buf_t *buf, arc_evict_func_t *func, void *private) 4617{ 4618 ASSERT(buf->b_hdr != NULL); 4619 ASSERT(buf->b_hdr->b_l1hdr.b_state != arc_anon); 4620 ASSERT(!refcount_is_zero(&buf->b_hdr->b_l1hdr.b_refcnt) || 4621 func == NULL); 4622 ASSERT(buf->b_efunc == NULL); 4623 ASSERT(!HDR_BUF_AVAILABLE(buf->b_hdr)); 4624 4625 buf->b_efunc = func; 4626 buf->b_private = private; 4627} 4628 4629/* 4630 * Notify the arc that a block was freed, and thus will never be used again. 4631 */ 4632void 4633arc_freed(spa_t *spa, const blkptr_t *bp) 4634{ 4635 arc_buf_hdr_t *hdr; 4636 kmutex_t *hash_lock; 4637 uint64_t guid = spa_load_guid(spa); 4638 4639 ASSERT(!BP_IS_EMBEDDED(bp)); 4640 4641 hdr = buf_hash_find(guid, bp, &hash_lock); 4642 if (hdr == NULL) 4643 return; 4644 if (HDR_BUF_AVAILABLE(hdr)) { 4645 arc_buf_t *buf = hdr->b_l1hdr.b_buf; 4646 add_reference(hdr, hash_lock, FTAG); 4647 hdr->b_flags &= ~ARC_FLAG_BUF_AVAILABLE; 4648 mutex_exit(hash_lock); 4649 4650 arc_release(buf, FTAG); 4651 (void) arc_buf_remove_ref(buf, FTAG); 4652 } else { 4653 mutex_exit(hash_lock); 4654 } 4655 4656} 4657 4658/* 4659 * Clear the user eviction callback set by arc_set_callback(), first calling 4660 * it if it exists. Because the presence of a callback keeps an arc_buf cached 4661 * clearing the callback may result in the arc_buf being destroyed. However, 4662 * it will not result in the *last* arc_buf being destroyed, hence the data 4663 * will remain cached in the ARC. We make a copy of the arc buffer here so 4664 * that we can process the callback without holding any locks. 4665 * 4666 * It's possible that the callback is already in the process of being cleared 4667 * by another thread. In this case we can not clear the callback. 4668 * 4669 * Returns B_TRUE if the callback was successfully called and cleared. 4670 */ 4671boolean_t 4672arc_clear_callback(arc_buf_t *buf) 4673{ 4674 arc_buf_hdr_t *hdr; 4675 kmutex_t *hash_lock; 4676 arc_evict_func_t *efunc = buf->b_efunc; 4677 void *private = buf->b_private; 4678 4679 mutex_enter(&buf->b_evict_lock); 4680 hdr = buf->b_hdr; 4681 if (hdr == NULL) { 4682 /* 4683 * We are in arc_do_user_evicts(). 4684 */ 4685 ASSERT(buf->b_data == NULL); 4686 mutex_exit(&buf->b_evict_lock); 4687 return (B_FALSE); 4688 } else if (buf->b_data == NULL) { 4689 /* 4690 * We are on the eviction list; process this buffer now 4691 * but let arc_do_user_evicts() do the reaping. 4692 */ 4693 buf->b_efunc = NULL; 4694 mutex_exit(&buf->b_evict_lock); 4695 VERIFY0(efunc(private)); 4696 return (B_TRUE); 4697 } 4698 hash_lock = HDR_LOCK(hdr); 4699 mutex_enter(hash_lock); 4700 hdr = buf->b_hdr; 4701 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 4702 4703 ASSERT3U(refcount_count(&hdr->b_l1hdr.b_refcnt), <, 4704 hdr->b_l1hdr.b_datacnt); 4705 ASSERT(hdr->b_l1hdr.b_state == arc_mru || 4706 hdr->b_l1hdr.b_state == arc_mfu); 4707 4708 buf->b_efunc = NULL; 4709 buf->b_private = NULL; 4710 4711 if (hdr->b_l1hdr.b_datacnt > 1) { 4712 mutex_exit(&buf->b_evict_lock); 4713 arc_buf_destroy(buf, TRUE); 4714 } else { 4715 ASSERT(buf == hdr->b_l1hdr.b_buf); 4716 hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE; 4717 mutex_exit(&buf->b_evict_lock); 4718 } 4719 4720 mutex_exit(hash_lock); 4721 VERIFY0(efunc(private)); 4722 return (B_TRUE); 4723} 4724 4725/* 4726 * Release this buffer from the cache, making it an anonymous buffer. This 4727 * must be done after a read and prior to modifying the buffer contents. 4728 * If the buffer has more than one reference, we must make 4729 * a new hdr for the buffer. 4730 */ 4731void 4732arc_release(arc_buf_t *buf, void *tag) 4733{ 4734 arc_buf_hdr_t *hdr = buf->b_hdr; 4735 4736 /* 4737 * It would be nice to assert that if it's DMU metadata (level > 4738 * 0 || it's the dnode file), then it must be syncing context. 4739 * But we don't know that information at this level. 4740 */ 4741 4742 mutex_enter(&buf->b_evict_lock); 4743 4744 ASSERT(HDR_HAS_L1HDR(hdr)); 4745 4746 /* 4747 * We don't grab the hash lock prior to this check, because if 4748 * the buffer's header is in the arc_anon state, it won't be 4749 * linked into the hash table. 4750 */ 4751 if (hdr->b_l1hdr.b_state == arc_anon) { 4752 mutex_exit(&buf->b_evict_lock); 4753 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 4754 ASSERT(!HDR_IN_HASH_TABLE(hdr)); 4755 ASSERT(!HDR_HAS_L2HDR(hdr)); 4756 ASSERT(BUF_EMPTY(hdr)); 4757 ASSERT3U(hdr->b_l1hdr.b_datacnt, ==, 1); 4758 ASSERT3S(refcount_count(&hdr->b_l1hdr.b_refcnt), ==, 1); 4759 ASSERT(!list_link_active(&hdr->b_l1hdr.b_arc_node)); 4760 4761 ASSERT3P(buf->b_efunc, ==, NULL); 4762 ASSERT3P(buf->b_private, ==, NULL); 4763 4764 hdr->b_l1hdr.b_arc_access = 0; 4765 arc_buf_thaw(buf); 4766 4767 return; 4768 } 4769 4770 kmutex_t *hash_lock = HDR_LOCK(hdr); 4771 mutex_enter(hash_lock); 4772 4773 /* 4774 * This assignment is only valid as long as the hash_lock is 4775 * held, we must be careful not to reference state or the 4776 * b_state field after dropping the lock. 4777 */ 4778 arc_state_t *state = hdr->b_l1hdr.b_state; 4779 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 4780 ASSERT3P(state, !=, arc_anon); 4781 4782 /* this buffer is not on any list */ 4783 ASSERT(refcount_count(&hdr->b_l1hdr.b_refcnt) > 0); 4784 4785 if (HDR_HAS_L2HDR(hdr)) { 4786 mutex_enter(&hdr->b_l2hdr.b_dev->l2ad_mtx); 4787 4788 /* 4789 * We have to recheck this conditional again now that 4790 * we're holding the l2ad_mtx to prevent a race with 4791 * another thread which might be concurrently calling 4792 * l2arc_evict(). In that case, l2arc_evict() might have 4793 * destroyed the header's L2 portion as we were waiting 4794 * to acquire the l2ad_mtx. 4795 */ 4796 if (HDR_HAS_L2HDR(hdr)) { 4797 l2arc_trim(hdr); 4798 arc_hdr_l2hdr_destroy(hdr); 4799 } 4800 4801 mutex_exit(&hdr->b_l2hdr.b_dev->l2ad_mtx); 4802 } 4803 4804 /* 4805 * Do we have more than one buf? 4806 */ 4807 if (hdr->b_l1hdr.b_datacnt > 1) { 4808 arc_buf_hdr_t *nhdr; 4809 arc_buf_t **bufp; 4810 uint64_t blksz = hdr->b_size; 4811 uint64_t spa = hdr->b_spa; 4812 arc_buf_contents_t type = arc_buf_type(hdr); 4813 uint32_t flags = hdr->b_flags; 4814 4815 ASSERT(hdr->b_l1hdr.b_buf != buf || buf->b_next != NULL); 4816 /* 4817 * Pull the data off of this hdr and attach it to 4818 * a new anonymous hdr. 4819 */ 4820 (void) remove_reference(hdr, hash_lock, tag); 4821 bufp = &hdr->b_l1hdr.b_buf; 4822 while (*bufp != buf) 4823 bufp = &(*bufp)->b_next; 4824 *bufp = buf->b_next; 4825 buf->b_next = NULL; 4826 4827 ASSERT3P(state, !=, arc_l2c_only); 4828 4829 (void) refcount_remove_many( 4830 &state->arcs_size, hdr->b_size, buf); 4831 4832 if (refcount_is_zero(&hdr->b_l1hdr.b_refcnt)) { 4833 ASSERT3P(state, !=, arc_l2c_only); 4834 uint64_t *size = &state->arcs_lsize[type]; 4835 ASSERT3U(*size, >=, hdr->b_size); 4836 atomic_add_64(size, -hdr->b_size); 4837 } 4838 4839 /* 4840 * We're releasing a duplicate user data buffer, update 4841 * our statistics accordingly. 4842 */ 4843 if (HDR_ISTYPE_DATA(hdr)) { 4844 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers); 4845 ARCSTAT_INCR(arcstat_duplicate_buffers_size, 4846 -hdr->b_size); 4847 } 4848 hdr->b_l1hdr.b_datacnt -= 1; 4849 arc_cksum_verify(buf); 4850#ifdef illumos 4851 arc_buf_unwatch(buf); 4852#endif 4853 4854 mutex_exit(hash_lock); 4855 4856 nhdr = kmem_cache_alloc(hdr_full_cache, KM_PUSHPAGE); 4857 nhdr->b_size = blksz; 4858 nhdr->b_spa = spa; 4859 4860 nhdr->b_flags = flags & ARC_FLAG_L2_WRITING; 4861 nhdr->b_flags |= arc_bufc_to_flags(type); 4862 nhdr->b_flags |= ARC_FLAG_HAS_L1HDR; 4863 4864 nhdr->b_l1hdr.b_buf = buf; 4865 nhdr->b_l1hdr.b_datacnt = 1; 4866 nhdr->b_l1hdr.b_state = arc_anon; 4867 nhdr->b_l1hdr.b_arc_access = 0; 4868 nhdr->b_l1hdr.b_tmp_cdata = NULL; 4869 nhdr->b_freeze_cksum = NULL; 4870 4871 (void) refcount_add(&nhdr->b_l1hdr.b_refcnt, tag); 4872 buf->b_hdr = nhdr; 4873 mutex_exit(&buf->b_evict_lock); 4874 (void) refcount_add_many(&arc_anon->arcs_size, blksz, buf); 4875 } else { 4876 mutex_exit(&buf->b_evict_lock); 4877 ASSERT(refcount_count(&hdr->b_l1hdr.b_refcnt) == 1); 4878 /* protected by hash lock, or hdr is on arc_anon */ 4879 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 4880 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 4881 arc_change_state(arc_anon, hdr, hash_lock); 4882 hdr->b_l1hdr.b_arc_access = 0; 4883 mutex_exit(hash_lock); 4884 4885 buf_discard_identity(hdr); 4886 arc_buf_thaw(buf); 4887 } 4888 buf->b_efunc = NULL; 4889 buf->b_private = NULL; 4890} 4891 4892int 4893arc_released(arc_buf_t *buf) 4894{ 4895 int released; 4896 4897 mutex_enter(&buf->b_evict_lock); 4898 released = (buf->b_data != NULL && 4899 buf->b_hdr->b_l1hdr.b_state == arc_anon); 4900 mutex_exit(&buf->b_evict_lock); 4901 return (released); 4902} 4903 4904#ifdef ZFS_DEBUG 4905int 4906arc_referenced(arc_buf_t *buf) 4907{ 4908 int referenced; 4909 4910 mutex_enter(&buf->b_evict_lock); 4911 referenced = (refcount_count(&buf->b_hdr->b_l1hdr.b_refcnt)); 4912 mutex_exit(&buf->b_evict_lock); 4913 return (referenced); 4914} 4915#endif 4916 4917static void 4918arc_write_ready(zio_t *zio) 4919{ 4920 arc_write_callback_t *callback = zio->io_private; 4921 arc_buf_t *buf = callback->awcb_buf; 4922 arc_buf_hdr_t *hdr = buf->b_hdr; 4923 4924 ASSERT(HDR_HAS_L1HDR(hdr)); 4925 ASSERT(!refcount_is_zero(&buf->b_hdr->b_l1hdr.b_refcnt)); 4926 ASSERT(hdr->b_l1hdr.b_datacnt > 0); 4927 callback->awcb_ready(zio, buf, callback->awcb_private); 4928 4929 /* 4930 * If the IO is already in progress, then this is a re-write 4931 * attempt, so we need to thaw and re-compute the cksum. 4932 * It is the responsibility of the callback to handle the 4933 * accounting for any re-write attempt. 4934 */ 4935 if (HDR_IO_IN_PROGRESS(hdr)) { 4936 mutex_enter(&hdr->b_l1hdr.b_freeze_lock); 4937 if (hdr->b_freeze_cksum != NULL) { 4938 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 4939 hdr->b_freeze_cksum = NULL; 4940 } 4941 mutex_exit(&hdr->b_l1hdr.b_freeze_lock); 4942 } 4943 arc_cksum_compute(buf, B_FALSE); 4944 hdr->b_flags |= ARC_FLAG_IO_IN_PROGRESS; 4945} 4946 4947/* 4948 * The SPA calls this callback for each physical write that happens on behalf 4949 * of a logical write. See the comment in dbuf_write_physdone() for details. 4950 */ 4951static void 4952arc_write_physdone(zio_t *zio) 4953{ 4954 arc_write_callback_t *cb = zio->io_private; 4955 if (cb->awcb_physdone != NULL) 4956 cb->awcb_physdone(zio, cb->awcb_buf, cb->awcb_private); 4957} 4958 4959static void 4960arc_write_done(zio_t *zio) 4961{ 4962 arc_write_callback_t *callback = zio->io_private; 4963 arc_buf_t *buf = callback->awcb_buf; 4964 arc_buf_hdr_t *hdr = buf->b_hdr; 4965 4966 ASSERT(hdr->b_l1hdr.b_acb == NULL); 4967 4968 if (zio->io_error == 0) { 4969 if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) { 4970 buf_discard_identity(hdr); 4971 } else { 4972 hdr->b_dva = *BP_IDENTITY(zio->io_bp); 4973 hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp); 4974 } 4975 } else { 4976 ASSERT(BUF_EMPTY(hdr)); 4977 } 4978 4979 /* 4980 * If the block to be written was all-zero or compressed enough to be 4981 * embedded in the BP, no write was performed so there will be no 4982 * dva/birth/checksum. The buffer must therefore remain anonymous 4983 * (and uncached). 4984 */ 4985 if (!BUF_EMPTY(hdr)) { 4986 arc_buf_hdr_t *exists; 4987 kmutex_t *hash_lock; 4988 4989 ASSERT(zio->io_error == 0); 4990 4991 arc_cksum_verify(buf); 4992 4993 exists = buf_hash_insert(hdr, &hash_lock); 4994 if (exists != NULL) { 4995 /* 4996 * This can only happen if we overwrite for 4997 * sync-to-convergence, because we remove 4998 * buffers from the hash table when we arc_free(). 4999 */ 5000 if (zio->io_flags & ZIO_FLAG_IO_REWRITE) { 5001 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) 5002 panic("bad overwrite, hdr=%p exists=%p", 5003 (void *)hdr, (void *)exists); 5004 ASSERT(refcount_is_zero( 5005 &exists->b_l1hdr.b_refcnt)); 5006 arc_change_state(arc_anon, exists, hash_lock); 5007 mutex_exit(hash_lock); 5008 arc_hdr_destroy(exists); 5009 exists = buf_hash_insert(hdr, &hash_lock); 5010 ASSERT3P(exists, ==, NULL); 5011 } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) { 5012 /* nopwrite */ 5013 ASSERT(zio->io_prop.zp_nopwrite); 5014 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) 5015 panic("bad nopwrite, hdr=%p exists=%p", 5016 (void *)hdr, (void *)exists); 5017 } else { 5018 /* Dedup */ 5019 ASSERT(hdr->b_l1hdr.b_datacnt == 1); 5020 ASSERT(hdr->b_l1hdr.b_state == arc_anon); 5021 ASSERT(BP_GET_DEDUP(zio->io_bp)); 5022 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0); 5023 } 5024 } 5025 hdr->b_flags &= ~ARC_FLAG_IO_IN_PROGRESS; 5026 /* if it's not anon, we are doing a scrub */ 5027 if (exists == NULL && hdr->b_l1hdr.b_state == arc_anon) 5028 arc_access(hdr, hash_lock); 5029 mutex_exit(hash_lock); 5030 } else { 5031 hdr->b_flags &= ~ARC_FLAG_IO_IN_PROGRESS; 5032 } 5033 5034 ASSERT(!refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 5035 callback->awcb_done(zio, buf, callback->awcb_private); 5036 5037 kmem_free(callback, sizeof (arc_write_callback_t)); 5038} 5039 5040zio_t * 5041arc_write(zio_t *pio, spa_t *spa, uint64_t txg, 5042 blkptr_t *bp, arc_buf_t *buf, boolean_t l2arc, boolean_t l2arc_compress, 5043 const zio_prop_t *zp, arc_done_func_t *ready, arc_done_func_t *physdone, 5044 arc_done_func_t *done, void *private, zio_priority_t priority, 5045 int zio_flags, const zbookmark_phys_t *zb) 5046{ 5047 arc_buf_hdr_t *hdr = buf->b_hdr; 5048 arc_write_callback_t *callback; 5049 zio_t *zio; 5050 5051 ASSERT(ready != NULL); 5052 ASSERT(done != NULL); 5053 ASSERT(!HDR_IO_ERROR(hdr)); 5054 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 5055 ASSERT(hdr->b_l1hdr.b_acb == NULL); 5056 ASSERT(hdr->b_l1hdr.b_datacnt > 0); 5057 if (l2arc) 5058 hdr->b_flags |= ARC_FLAG_L2CACHE; 5059 if (l2arc_compress) 5060 hdr->b_flags |= ARC_FLAG_L2COMPRESS; 5061 callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP); 5062 callback->awcb_ready = ready; 5063 callback->awcb_physdone = physdone; 5064 callback->awcb_done = done; 5065 callback->awcb_private = private; 5066 callback->awcb_buf = buf; 5067 5068 zio = zio_write(pio, spa, txg, bp, buf->b_data, hdr->b_size, zp, 5069 arc_write_ready, arc_write_physdone, arc_write_done, callback, 5070 priority, zio_flags, zb); 5071 5072 return (zio); 5073} 5074 5075static int 5076arc_memory_throttle(uint64_t reserve, uint64_t txg) 5077{ 5078#ifdef _KERNEL 5079 uint64_t available_memory = ptob(freemem); 5080 static uint64_t page_load = 0; 5081 static uint64_t last_txg = 0; 5082 5083#if defined(__i386) || !defined(UMA_MD_SMALL_ALLOC) 5084 available_memory = 5085 MIN(available_memory, ptob(vmem_size(heap_arena, VMEM_FREE))); 5086#endif 5087 5088 if (freemem > (uint64_t)physmem * arc_lotsfree_percent / 100) 5089 return (0); 5090 5091 if (txg > last_txg) { 5092 last_txg = txg; 5093 page_load = 0; 5094 } 5095 /* 5096 * If we are in pageout, we know that memory is already tight, 5097 * the arc is already going to be evicting, so we just want to 5098 * continue to let page writes occur as quickly as possible. 5099 */ 5100 if (curproc == pageproc) { 5101 if (page_load > MAX(ptob(minfree), available_memory) / 4) 5102 return (SET_ERROR(ERESTART)); 5103 /* Note: reserve is inflated, so we deflate */ 5104 page_load += reserve / 8; 5105 return (0); 5106 } else if (page_load > 0 && arc_reclaim_needed()) { 5107 /* memory is low, delay before restarting */ 5108 ARCSTAT_INCR(arcstat_memory_throttle_count, 1); 5109 return (SET_ERROR(EAGAIN)); 5110 } 5111 page_load = 0; 5112#endif 5113 return (0); 5114} 5115 5116void 5117arc_tempreserve_clear(uint64_t reserve) 5118{ 5119 atomic_add_64(&arc_tempreserve, -reserve); 5120 ASSERT((int64_t)arc_tempreserve >= 0); 5121} 5122 5123int 5124arc_tempreserve_space(uint64_t reserve, uint64_t txg) 5125{ 5126 int error; 5127 uint64_t anon_size; 5128 5129 if (reserve > arc_c/4 && !arc_no_grow) { 5130 arc_c = MIN(arc_c_max, reserve * 4); 5131 DTRACE_PROBE1(arc__set_reserve, uint64_t, arc_c); 5132 } 5133 if (reserve > arc_c) 5134 return (SET_ERROR(ENOMEM)); 5135 5136 /* 5137 * Don't count loaned bufs as in flight dirty data to prevent long 5138 * network delays from blocking transactions that are ready to be 5139 * assigned to a txg. 5140 */ 5141 anon_size = MAX((int64_t)(refcount_count(&arc_anon->arcs_size) - 5142 arc_loaned_bytes), 0); 5143 5144 /* 5145 * Writes will, almost always, require additional memory allocations 5146 * in order to compress/encrypt/etc the data. We therefore need to 5147 * make sure that there is sufficient available memory for this. 5148 */ 5149 error = arc_memory_throttle(reserve, txg); 5150 if (error != 0) 5151 return (error); 5152 5153 /* 5154 * Throttle writes when the amount of dirty data in the cache 5155 * gets too large. We try to keep the cache less than half full 5156 * of dirty blocks so that our sync times don't grow too large. 5157 * Note: if two requests come in concurrently, we might let them 5158 * both succeed, when one of them should fail. Not a huge deal. 5159 */ 5160 5161 if (reserve + arc_tempreserve + anon_size > arc_c / 2 && 5162 anon_size > arc_c / 4) { 5163 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK " 5164 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n", 5165 arc_tempreserve>>10, 5166 arc_anon->arcs_lsize[ARC_BUFC_METADATA]>>10, 5167 arc_anon->arcs_lsize[ARC_BUFC_DATA]>>10, 5168 reserve>>10, arc_c>>10); 5169 return (SET_ERROR(ERESTART)); 5170 } 5171 atomic_add_64(&arc_tempreserve, reserve); 5172 return (0); 5173} 5174 5175static void 5176arc_kstat_update_state(arc_state_t *state, kstat_named_t *size, 5177 kstat_named_t *evict_data, kstat_named_t *evict_metadata) 5178{ 5179 size->value.ui64 = refcount_count(&state->arcs_size); 5180 evict_data->value.ui64 = state->arcs_lsize[ARC_BUFC_DATA]; 5181 evict_metadata->value.ui64 = state->arcs_lsize[ARC_BUFC_METADATA]; 5182} 5183 5184static int 5185arc_kstat_update(kstat_t *ksp, int rw) 5186{ 5187 arc_stats_t *as = ksp->ks_data; 5188 5189 if (rw == KSTAT_WRITE) { 5190 return (EACCES); 5191 } else { 5192 arc_kstat_update_state(arc_anon, 5193 &as->arcstat_anon_size, 5194 &as->arcstat_anon_evictable_data, 5195 &as->arcstat_anon_evictable_metadata); 5196 arc_kstat_update_state(arc_mru, 5197 &as->arcstat_mru_size, 5198 &as->arcstat_mru_evictable_data, 5199 &as->arcstat_mru_evictable_metadata); 5200 arc_kstat_update_state(arc_mru_ghost, 5201 &as->arcstat_mru_ghost_size, 5202 &as->arcstat_mru_ghost_evictable_data, 5203 &as->arcstat_mru_ghost_evictable_metadata); 5204 arc_kstat_update_state(arc_mfu, 5205 &as->arcstat_mfu_size, 5206 &as->arcstat_mfu_evictable_data, 5207 &as->arcstat_mfu_evictable_metadata); 5208 arc_kstat_update_state(arc_mfu_ghost, 5209 &as->arcstat_mfu_ghost_size, 5210 &as->arcstat_mfu_ghost_evictable_data, 5211 &as->arcstat_mfu_ghost_evictable_metadata); 5212 } 5213 5214 return (0); 5215} 5216 5217/* 5218 * This function *must* return indices evenly distributed between all 5219 * sublists of the multilist. This is needed due to how the ARC eviction 5220 * code is laid out; arc_evict_state() assumes ARC buffers are evenly 5221 * distributed between all sublists and uses this assumption when 5222 * deciding which sublist to evict from and how much to evict from it. 5223 */ 5224unsigned int 5225arc_state_multilist_index_func(multilist_t *ml, void *obj) 5226{ 5227 arc_buf_hdr_t *hdr = obj; 5228 5229 /* 5230 * We rely on b_dva to generate evenly distributed index 5231 * numbers using buf_hash below. So, as an added precaution, 5232 * let's make sure we never add empty buffers to the arc lists. 5233 */ 5234 ASSERT(!BUF_EMPTY(hdr)); 5235 5236 /* 5237 * The assumption here, is the hash value for a given 5238 * arc_buf_hdr_t will remain constant throughout it's lifetime 5239 * (i.e. it's b_spa, b_dva, and b_birth fields don't change). 5240 * Thus, we don't need to store the header's sublist index 5241 * on insertion, as this index can be recalculated on removal. 5242 * 5243 * Also, the low order bits of the hash value are thought to be 5244 * distributed evenly. Otherwise, in the case that the multilist 5245 * has a power of two number of sublists, each sublists' usage 5246 * would not be evenly distributed. 5247 */ 5248 return (buf_hash(hdr->b_spa, &hdr->b_dva, hdr->b_birth) % 5249 multilist_get_num_sublists(ml)); 5250} 5251 5252#ifdef _KERNEL 5253static eventhandler_tag arc_event_lowmem = NULL; 5254 5255static void 5256arc_lowmem(void *arg __unused, int howto __unused) 5257{ 5258 5259 mutex_enter(&arc_reclaim_lock); 5260 /* XXX: Memory deficit should be passed as argument. */ 5261 needfree = btoc(arc_c >> arc_shrink_shift); 5262 DTRACE_PROBE(arc__needfree); 5263 cv_signal(&arc_reclaim_thread_cv); 5264 5265 /* 5266 * It is unsafe to block here in arbitrary threads, because we can come 5267 * here from ARC itself and may hold ARC locks and thus risk a deadlock 5268 * with ARC reclaim thread. 5269 */ 5270 if (curproc == pageproc) 5271 (void) cv_wait(&arc_reclaim_waiters_cv, &arc_reclaim_lock); 5272 mutex_exit(&arc_reclaim_lock); 5273} 5274#endif 5275 5276void 5277arc_init(void) 5278{ 5279 int i, prefetch_tunable_set = 0; 5280 5281 mutex_init(&arc_reclaim_lock, NULL, MUTEX_DEFAULT, NULL); 5282 cv_init(&arc_reclaim_thread_cv, NULL, CV_DEFAULT, NULL); 5283 cv_init(&arc_reclaim_waiters_cv, NULL, CV_DEFAULT, NULL); 5284 5285 mutex_init(&arc_user_evicts_lock, NULL, MUTEX_DEFAULT, NULL); 5286 cv_init(&arc_user_evicts_cv, NULL, CV_DEFAULT, NULL); 5287 5288 /* Convert seconds to clock ticks */ 5289 arc_min_prefetch_lifespan = 1 * hz; 5290 5291 /* Start out with 1/8 of all memory */ 5292 arc_c = kmem_size() / 8; 5293 5294#ifdef illumos 5295#ifdef _KERNEL 5296 /* 5297 * On architectures where the physical memory can be larger 5298 * than the addressable space (intel in 32-bit mode), we may 5299 * need to limit the cache to 1/8 of VM size. 5300 */ 5301 arc_c = MIN(arc_c, vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 8); 5302#endif 5303#endif /* illumos */ 5304 /* set min cache to 1/32 of all memory, or 16MB, whichever is more */ 5305 arc_c_min = MAX(arc_c / 4, 16 << 20); 5306 /* set max to 1/2 of all memory, or all but 1GB, whichever is more */ 5307 if (arc_c * 8 >= 1 << 30) 5308 arc_c_max = (arc_c * 8) - (1 << 30); 5309 else 5310 arc_c_max = arc_c_min; 5311 arc_c_max = MAX(arc_c * 5, arc_c_max); 5312 5313 /* 5314 * In userland, there's only the memory pressure that we artificially 5315 * create (see arc_available_memory()). Don't let arc_c get too 5316 * small, because it can cause transactions to be larger than 5317 * arc_c, causing arc_tempreserve_space() to fail. 5318 */ 5319#ifndef _KERNEL 5320 arc_c_min = arc_c_max / 2; 5321#endif 5322 5323#ifdef _KERNEL 5324 /* 5325 * Allow the tunables to override our calculations if they are 5326 * reasonable (ie. over 16MB) 5327 */ 5328 if (zfs_arc_max > 16 << 20 && zfs_arc_max < kmem_size()) 5329 arc_c_max = zfs_arc_max; 5330 if (zfs_arc_min > 16 << 20 && zfs_arc_min <= arc_c_max) 5331 arc_c_min = zfs_arc_min; 5332#endif 5333 5334 arc_c = arc_c_max; 5335 arc_p = (arc_c >> 1); 5336 5337 /* limit meta-data to 1/4 of the arc capacity */ 5338 arc_meta_limit = arc_c_max / 4; 5339 5340 /* Allow the tunable to override if it is reasonable */ 5341 if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max) 5342 arc_meta_limit = zfs_arc_meta_limit; 5343 5344 if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0) 5345 arc_c_min = arc_meta_limit / 2; 5346 5347 if (zfs_arc_meta_min > 0) { 5348 arc_meta_min = zfs_arc_meta_min; 5349 } else { 5350 arc_meta_min = arc_c_min / 2; 5351 } 5352 5353 if (zfs_arc_grow_retry > 0) 5354 arc_grow_retry = zfs_arc_grow_retry; 5355 5356 if (zfs_arc_shrink_shift > 0) 5357 arc_shrink_shift = zfs_arc_shrink_shift; 5358 5359 /* 5360 * Ensure that arc_no_grow_shift is less than arc_shrink_shift. 5361 */ 5362 if (arc_no_grow_shift >= arc_shrink_shift) 5363 arc_no_grow_shift = arc_shrink_shift - 1; 5364 5365 if (zfs_arc_p_min_shift > 0) 5366 arc_p_min_shift = zfs_arc_p_min_shift; 5367 5368 if (zfs_arc_num_sublists_per_state < 1) 5369 zfs_arc_num_sublists_per_state = MAX(max_ncpus, 1); 5370 5371 /* if kmem_flags are set, lets try to use less memory */ 5372 if (kmem_debugging()) 5373 arc_c = arc_c / 2; 5374 if (arc_c < arc_c_min) 5375 arc_c = arc_c_min; 5376 5377 zfs_arc_min = arc_c_min; 5378 zfs_arc_max = arc_c_max; 5379 5380 arc_anon = &ARC_anon; 5381 arc_mru = &ARC_mru; 5382 arc_mru_ghost = &ARC_mru_ghost; 5383 arc_mfu = &ARC_mfu; 5384 arc_mfu_ghost = &ARC_mfu_ghost; 5385 arc_l2c_only = &ARC_l2c_only; 5386 arc_size = 0; 5387 5388 multilist_create(&arc_mru->arcs_list[ARC_BUFC_METADATA], 5389 sizeof (arc_buf_hdr_t), 5390 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5391 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5392 multilist_create(&arc_mru->arcs_list[ARC_BUFC_DATA], 5393 sizeof (arc_buf_hdr_t), 5394 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5395 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5396 multilist_create(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA], 5397 sizeof (arc_buf_hdr_t), 5398 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5399 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5400 multilist_create(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA], 5401 sizeof (arc_buf_hdr_t), 5402 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5403 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5404 multilist_create(&arc_mfu->arcs_list[ARC_BUFC_METADATA], 5405 sizeof (arc_buf_hdr_t), 5406 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5407 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5408 multilist_create(&arc_mfu->arcs_list[ARC_BUFC_DATA], 5409 sizeof (arc_buf_hdr_t), 5410 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5411 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5412 multilist_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA], 5413 sizeof (arc_buf_hdr_t), 5414 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5415 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5416 multilist_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA], 5417 sizeof (arc_buf_hdr_t), 5418 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5419 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5420 multilist_create(&arc_l2c_only->arcs_list[ARC_BUFC_METADATA], 5421 sizeof (arc_buf_hdr_t), 5422 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5423 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5424 multilist_create(&arc_l2c_only->arcs_list[ARC_BUFC_DATA], 5425 sizeof (arc_buf_hdr_t), 5426 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5427 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5428 5429 refcount_create(&arc_anon->arcs_size); 5430 refcount_create(&arc_mru->arcs_size); 5431 refcount_create(&arc_mru_ghost->arcs_size); 5432 refcount_create(&arc_mfu->arcs_size); 5433 refcount_create(&arc_mfu_ghost->arcs_size); 5434 refcount_create(&arc_l2c_only->arcs_size); 5435 5436 buf_init(); 5437 5438 arc_reclaim_thread_exit = FALSE; 5439 arc_user_evicts_thread_exit = FALSE; 5440 arc_eviction_list = NULL; 5441 bzero(&arc_eviction_hdr, sizeof (arc_buf_hdr_t)); 5442 5443 arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED, 5444 sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL); 5445 5446 if (arc_ksp != NULL) { 5447 arc_ksp->ks_data = &arc_stats; 5448 arc_ksp->ks_update = arc_kstat_update; 5449 kstat_install(arc_ksp); 5450 } 5451 5452 (void) thread_create(NULL, 0, arc_reclaim_thread, NULL, 0, &p0, 5453 TS_RUN, minclsyspri); 5454 5455#ifdef _KERNEL 5456 arc_event_lowmem = EVENTHANDLER_REGISTER(vm_lowmem, arc_lowmem, NULL, 5457 EVENTHANDLER_PRI_FIRST); 5458#endif 5459 5460 (void) thread_create(NULL, 0, arc_user_evicts_thread, NULL, 0, &p0, 5461 TS_RUN, minclsyspri); 5462 5463 arc_dead = FALSE; 5464 arc_warm = B_FALSE; 5465 5466 /* 5467 * Calculate maximum amount of dirty data per pool. 5468 * 5469 * If it has been set by /etc/system, take that. 5470 * Otherwise, use a percentage of physical memory defined by 5471 * zfs_dirty_data_max_percent (default 10%) with a cap at 5472 * zfs_dirty_data_max_max (default 4GB). 5473 */ 5474 if (zfs_dirty_data_max == 0) { 5475 zfs_dirty_data_max = ptob(physmem) * 5476 zfs_dirty_data_max_percent / 100; 5477 zfs_dirty_data_max = MIN(zfs_dirty_data_max, 5478 zfs_dirty_data_max_max); 5479 } 5480 5481#ifdef _KERNEL 5482 if (TUNABLE_INT_FETCH("vfs.zfs.prefetch_disable", &zfs_prefetch_disable)) 5483 prefetch_tunable_set = 1; 5484 5485#ifdef __i386__ 5486 if (prefetch_tunable_set == 0) { 5487 printf("ZFS NOTICE: Prefetch is disabled by default on i386 " 5488 "-- to enable,\n"); 5489 printf(" add \"vfs.zfs.prefetch_disable=0\" " 5490 "to /boot/loader.conf.\n"); 5491 zfs_prefetch_disable = 1; 5492 } 5493#else 5494 if ((((uint64_t)physmem * PAGESIZE) < (1ULL << 32)) && 5495 prefetch_tunable_set == 0) { 5496 printf("ZFS NOTICE: Prefetch is disabled by default if less " 5497 "than 4GB of RAM is present;\n" 5498 " to enable, add \"vfs.zfs.prefetch_disable=0\" " 5499 "to /boot/loader.conf.\n"); 5500 zfs_prefetch_disable = 1; 5501 } 5502#endif 5503 /* Warn about ZFS memory and address space requirements. */ 5504 if (((uint64_t)physmem * PAGESIZE) < (256 + 128 + 64) * (1 << 20)) { 5505 printf("ZFS WARNING: Recommended minimum RAM size is 512MB; " 5506 "expect unstable behavior.\n"); 5507 } 5508 if (kmem_size() < 512 * (1 << 20)) { 5509 printf("ZFS WARNING: Recommended minimum kmem_size is 512MB; " 5510 "expect unstable behavior.\n"); 5511 printf(" Consider tuning vm.kmem_size and " 5512 "vm.kmem_size_max\n"); 5513 printf(" in /boot/loader.conf.\n"); 5514 } 5515#endif 5516} 5517 5518void 5519arc_fini(void) 5520{ 5521 mutex_enter(&arc_reclaim_lock); 5522 arc_reclaim_thread_exit = TRUE; 5523 /* 5524 * The reclaim thread will set arc_reclaim_thread_exit back to 5525 * FALSE when it is finished exiting; we're waiting for that. 5526 */ 5527 while (arc_reclaim_thread_exit) { 5528 cv_signal(&arc_reclaim_thread_cv); 5529 cv_wait(&arc_reclaim_thread_cv, &arc_reclaim_lock); 5530 } 5531 mutex_exit(&arc_reclaim_lock); 5532 5533 mutex_enter(&arc_user_evicts_lock); 5534 arc_user_evicts_thread_exit = TRUE; 5535 /* 5536 * The user evicts thread will set arc_user_evicts_thread_exit 5537 * to FALSE when it is finished exiting; we're waiting for that. 5538 */ 5539 while (arc_user_evicts_thread_exit) { 5540 cv_signal(&arc_user_evicts_cv); 5541 cv_wait(&arc_user_evicts_cv, &arc_user_evicts_lock); 5542 } 5543 mutex_exit(&arc_user_evicts_lock); 5544 5545 /* Use TRUE to ensure *all* buffers are evicted */ 5546 arc_flush(NULL, TRUE); 5547 5548 arc_dead = TRUE; 5549 5550 if (arc_ksp != NULL) { 5551 kstat_delete(arc_ksp); 5552 arc_ksp = NULL; 5553 } 5554 5555 mutex_destroy(&arc_reclaim_lock); 5556 cv_destroy(&arc_reclaim_thread_cv); 5557 cv_destroy(&arc_reclaim_waiters_cv); 5558 5559 mutex_destroy(&arc_user_evicts_lock); 5560 cv_destroy(&arc_user_evicts_cv); 5561 5562 refcount_destroy(&arc_anon->arcs_size); 5563 refcount_destroy(&arc_mru->arcs_size); 5564 refcount_destroy(&arc_mru_ghost->arcs_size); 5565 refcount_destroy(&arc_mfu->arcs_size); 5566 refcount_destroy(&arc_mfu_ghost->arcs_size); 5567 refcount_destroy(&arc_l2c_only->arcs_size); 5568 5569 multilist_destroy(&arc_mru->arcs_list[ARC_BUFC_METADATA]); 5570 multilist_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA]); 5571 multilist_destroy(&arc_mfu->arcs_list[ARC_BUFC_METADATA]); 5572 multilist_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA]); 5573 multilist_destroy(&arc_mru->arcs_list[ARC_BUFC_DATA]); 5574 multilist_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA]); 5575 multilist_destroy(&arc_mfu->arcs_list[ARC_BUFC_DATA]); 5576 multilist_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA]); 5577 5578 buf_fini(); 5579 5580 ASSERT0(arc_loaned_bytes); 5581 5582#ifdef _KERNEL 5583 if (arc_event_lowmem != NULL) 5584 EVENTHANDLER_DEREGISTER(vm_lowmem, arc_event_lowmem); 5585#endif 5586} 5587 5588/* 5589 * Level 2 ARC 5590 * 5591 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk. 5592 * It uses dedicated storage devices to hold cached data, which are populated 5593 * using large infrequent writes. The main role of this cache is to boost 5594 * the performance of random read workloads. The intended L2ARC devices 5595 * include short-stroked disks, solid state disks, and other media with 5596 * substantially faster read latency than disk. 5597 * 5598 * +-----------------------+ 5599 * | ARC | 5600 * +-----------------------+ 5601 * | ^ ^ 5602 * | | | 5603 * l2arc_feed_thread() arc_read() 5604 * | | | 5605 * | l2arc read | 5606 * V | | 5607 * +---------------+ | 5608 * | L2ARC | | 5609 * +---------------+ | 5610 * | ^ | 5611 * l2arc_write() | | 5612 * | | | 5613 * V | | 5614 * +-------+ +-------+ 5615 * | vdev | | vdev | 5616 * | cache | | cache | 5617 * +-------+ +-------+ 5618 * +=========+ .-----. 5619 * : L2ARC : |-_____-| 5620 * : devices : | Disks | 5621 * +=========+ `-_____-' 5622 * 5623 * Read requests are satisfied from the following sources, in order: 5624 * 5625 * 1) ARC 5626 * 2) vdev cache of L2ARC devices 5627 * 3) L2ARC devices 5628 * 4) vdev cache of disks 5629 * 5) disks 5630 * 5631 * Some L2ARC device types exhibit extremely slow write performance. 5632 * To accommodate for this there are some significant differences between 5633 * the L2ARC and traditional cache design: 5634 * 5635 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from 5636 * the ARC behave as usual, freeing buffers and placing headers on ghost 5637 * lists. The ARC does not send buffers to the L2ARC during eviction as 5638 * this would add inflated write latencies for all ARC memory pressure. 5639 * 5640 * 2. The L2ARC attempts to cache data from the ARC before it is evicted. 5641 * It does this by periodically scanning buffers from the eviction-end of 5642 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are 5643 * not already there. It scans until a headroom of buffers is satisfied, 5644 * which itself is a buffer for ARC eviction. If a compressible buffer is 5645 * found during scanning and selected for writing to an L2ARC device, we 5646 * temporarily boost scanning headroom during the next scan cycle to make 5647 * sure we adapt to compression effects (which might significantly reduce 5648 * the data volume we write to L2ARC). The thread that does this is 5649 * l2arc_feed_thread(), illustrated below; example sizes are included to 5650 * provide a better sense of ratio than this diagram: 5651 * 5652 * head --> tail 5653 * +---------------------+----------+ 5654 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC 5655 * +---------------------+----------+ | o L2ARC eligible 5656 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer 5657 * +---------------------+----------+ | 5658 * 15.9 Gbytes ^ 32 Mbytes | 5659 * headroom | 5660 * l2arc_feed_thread() 5661 * | 5662 * l2arc write hand <--[oooo]--' 5663 * | 8 Mbyte 5664 * | write max 5665 * V 5666 * +==============================+ 5667 * L2ARC dev |####|#|###|###| |####| ... | 5668 * +==============================+ 5669 * 32 Gbytes 5670 * 5671 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of 5672 * evicted, then the L2ARC has cached a buffer much sooner than it probably 5673 * needed to, potentially wasting L2ARC device bandwidth and storage. It is 5674 * safe to say that this is an uncommon case, since buffers at the end of 5675 * the ARC lists have moved there due to inactivity. 5676 * 5677 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom, 5678 * then the L2ARC simply misses copying some buffers. This serves as a 5679 * pressure valve to prevent heavy read workloads from both stalling the ARC 5680 * with waits and clogging the L2ARC with writes. This also helps prevent 5681 * the potential for the L2ARC to churn if it attempts to cache content too 5682 * quickly, such as during backups of the entire pool. 5683 * 5684 * 5. After system boot and before the ARC has filled main memory, there are 5685 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru 5686 * lists can remain mostly static. Instead of searching from tail of these 5687 * lists as pictured, the l2arc_feed_thread() will search from the list heads 5688 * for eligible buffers, greatly increasing its chance of finding them. 5689 * 5690 * The L2ARC device write speed is also boosted during this time so that 5691 * the L2ARC warms up faster. Since there have been no ARC evictions yet, 5692 * there are no L2ARC reads, and no fear of degrading read performance 5693 * through increased writes. 5694 * 5695 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that 5696 * the vdev queue can aggregate them into larger and fewer writes. Each 5697 * device is written to in a rotor fashion, sweeping writes through 5698 * available space then repeating. 5699 * 5700 * 7. The L2ARC does not store dirty content. It never needs to flush 5701 * write buffers back to disk based storage. 5702 * 5703 * 8. If an ARC buffer is written (and dirtied) which also exists in the 5704 * L2ARC, the now stale L2ARC buffer is immediately dropped. 5705 * 5706 * The performance of the L2ARC can be tweaked by a number of tunables, which 5707 * may be necessary for different workloads: 5708 * 5709 * l2arc_write_max max write bytes per interval 5710 * l2arc_write_boost extra write bytes during device warmup 5711 * l2arc_noprefetch skip caching prefetched buffers 5712 * l2arc_headroom number of max device writes to precache 5713 * l2arc_headroom_boost when we find compressed buffers during ARC 5714 * scanning, we multiply headroom by this 5715 * percentage factor for the next scan cycle, 5716 * since more compressed buffers are likely to 5717 * be present 5718 * l2arc_feed_secs seconds between L2ARC writing 5719 * 5720 * Tunables may be removed or added as future performance improvements are 5721 * integrated, and also may become zpool properties. 5722 * 5723 * There are three key functions that control how the L2ARC warms up: 5724 * 5725 * l2arc_write_eligible() check if a buffer is eligible to cache 5726 * l2arc_write_size() calculate how much to write 5727 * l2arc_write_interval() calculate sleep delay between writes 5728 * 5729 * These three functions determine what to write, how much, and how quickly 5730 * to send writes. 5731 */ 5732 5733static boolean_t 5734l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *hdr) 5735{ 5736 /* 5737 * A buffer is *not* eligible for the L2ARC if it: 5738 * 1. belongs to a different spa. 5739 * 2. is already cached on the L2ARC. 5740 * 3. has an I/O in progress (it may be an incomplete read). 5741 * 4. is flagged not eligible (zfs property). 5742 */ 5743 if (hdr->b_spa != spa_guid) { 5744 ARCSTAT_BUMP(arcstat_l2_write_spa_mismatch); 5745 return (B_FALSE); 5746 } 5747 if (HDR_HAS_L2HDR(hdr)) { 5748 ARCSTAT_BUMP(arcstat_l2_write_in_l2); 5749 return (B_FALSE); 5750 } 5751 if (HDR_IO_IN_PROGRESS(hdr)) { 5752 ARCSTAT_BUMP(arcstat_l2_write_hdr_io_in_progress); 5753 return (B_FALSE); 5754 } 5755 if (!HDR_L2CACHE(hdr)) { 5756 ARCSTAT_BUMP(arcstat_l2_write_not_cacheable); 5757 return (B_FALSE); 5758 } 5759 5760 return (B_TRUE); 5761} 5762 5763static uint64_t 5764l2arc_write_size(void) 5765{ 5766 uint64_t size; 5767 5768 /* 5769 * Make sure our globals have meaningful values in case the user 5770 * altered them. 5771 */ 5772 size = l2arc_write_max; 5773 if (size == 0) { 5774 cmn_err(CE_NOTE, "Bad value for l2arc_write_max, value must " 5775 "be greater than zero, resetting it to the default (%d)", 5776 L2ARC_WRITE_SIZE); 5777 size = l2arc_write_max = L2ARC_WRITE_SIZE; 5778 } 5779 5780 if (arc_warm == B_FALSE) 5781 size += l2arc_write_boost; 5782 5783 return (size); 5784 5785} 5786 5787static clock_t 5788l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote) 5789{ 5790 clock_t interval, next, now; 5791 5792 /* 5793 * If the ARC lists are busy, increase our write rate; if the 5794 * lists are stale, idle back. This is achieved by checking 5795 * how much we previously wrote - if it was more than half of 5796 * what we wanted, schedule the next write much sooner. 5797 */ 5798 if (l2arc_feed_again && wrote > (wanted / 2)) 5799 interval = (hz * l2arc_feed_min_ms) / 1000; 5800 else 5801 interval = hz * l2arc_feed_secs; 5802 5803 now = ddi_get_lbolt(); 5804 next = MAX(now, MIN(now + interval, began + interval)); 5805 5806 return (next); 5807} 5808 5809/* 5810 * Cycle through L2ARC devices. This is how L2ARC load balances. 5811 * If a device is returned, this also returns holding the spa config lock. 5812 */ 5813static l2arc_dev_t * 5814l2arc_dev_get_next(void) 5815{ 5816 l2arc_dev_t *first, *next = NULL; 5817 5818 /* 5819 * Lock out the removal of spas (spa_namespace_lock), then removal 5820 * of cache devices (l2arc_dev_mtx). Once a device has been selected, 5821 * both locks will be dropped and a spa config lock held instead. 5822 */ 5823 mutex_enter(&spa_namespace_lock); 5824 mutex_enter(&l2arc_dev_mtx); 5825 5826 /* if there are no vdevs, there is nothing to do */ 5827 if (l2arc_ndev == 0) 5828 goto out; 5829 5830 first = NULL; 5831 next = l2arc_dev_last; 5832 do { 5833 /* loop around the list looking for a non-faulted vdev */ 5834 if (next == NULL) { 5835 next = list_head(l2arc_dev_list); 5836 } else { 5837 next = list_next(l2arc_dev_list, next); 5838 if (next == NULL) 5839 next = list_head(l2arc_dev_list); 5840 } 5841 5842 /* if we have come back to the start, bail out */ 5843 if (first == NULL) 5844 first = next; 5845 else if (next == first) 5846 break; 5847 5848 } while (vdev_is_dead(next->l2ad_vdev)); 5849 5850 /* if we were unable to find any usable vdevs, return NULL */ 5851 if (vdev_is_dead(next->l2ad_vdev)) 5852 next = NULL; 5853 5854 l2arc_dev_last = next; 5855 5856out: 5857 mutex_exit(&l2arc_dev_mtx); 5858 5859 /* 5860 * Grab the config lock to prevent the 'next' device from being 5861 * removed while we are writing to it. 5862 */ 5863 if (next != NULL) 5864 spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER); 5865 mutex_exit(&spa_namespace_lock); 5866 5867 return (next); 5868} 5869 5870/* 5871 * Free buffers that were tagged for destruction. 5872 */ 5873static void 5874l2arc_do_free_on_write() 5875{ 5876 list_t *buflist; 5877 l2arc_data_free_t *df, *df_prev; 5878 5879 mutex_enter(&l2arc_free_on_write_mtx); 5880 buflist = l2arc_free_on_write; 5881 5882 for (df = list_tail(buflist); df; df = df_prev) { 5883 df_prev = list_prev(buflist, df); 5884 ASSERT(df->l2df_data != NULL); 5885 ASSERT(df->l2df_func != NULL); 5886 df->l2df_func(df->l2df_data, df->l2df_size); 5887 list_remove(buflist, df); 5888 kmem_free(df, sizeof (l2arc_data_free_t)); 5889 } 5890 5891 mutex_exit(&l2arc_free_on_write_mtx); 5892} 5893 5894/* 5895 * A write to a cache device has completed. Update all headers to allow 5896 * reads from these buffers to begin. 5897 */ 5898static void 5899l2arc_write_done(zio_t *zio) 5900{ 5901 l2arc_write_callback_t *cb; 5902 l2arc_dev_t *dev; 5903 list_t *buflist; 5904 arc_buf_hdr_t *head, *hdr, *hdr_prev; 5905 kmutex_t *hash_lock; 5906 int64_t bytes_dropped = 0; 5907 5908 cb = zio->io_private; 5909 ASSERT(cb != NULL); 5910 dev = cb->l2wcb_dev; 5911 ASSERT(dev != NULL); 5912 head = cb->l2wcb_head; 5913 ASSERT(head != NULL); 5914 buflist = &dev->l2ad_buflist; 5915 ASSERT(buflist != NULL); 5916 DTRACE_PROBE2(l2arc__iodone, zio_t *, zio, 5917 l2arc_write_callback_t *, cb); 5918 5919 if (zio->io_error != 0) 5920 ARCSTAT_BUMP(arcstat_l2_writes_error); 5921 5922 /* 5923 * All writes completed, or an error was hit. 5924 */ 5925top: 5926 mutex_enter(&dev->l2ad_mtx); 5927 for (hdr = list_prev(buflist, head); hdr; hdr = hdr_prev) { 5928 hdr_prev = list_prev(buflist, hdr); 5929 5930 hash_lock = HDR_LOCK(hdr); 5931 5932 /* 5933 * We cannot use mutex_enter or else we can deadlock 5934 * with l2arc_write_buffers (due to swapping the order 5935 * the hash lock and l2ad_mtx are taken). 5936 */ 5937 if (!mutex_tryenter(hash_lock)) { 5938 /* 5939 * Missed the hash lock. We must retry so we 5940 * don't leave the ARC_FLAG_L2_WRITING bit set. 5941 */ 5942 ARCSTAT_BUMP(arcstat_l2_writes_lock_retry); 5943 5944 /* 5945 * We don't want to rescan the headers we've 5946 * already marked as having been written out, so 5947 * we reinsert the head node so we can pick up 5948 * where we left off. 5949 */ 5950 list_remove(buflist, head); 5951 list_insert_after(buflist, hdr, head); 5952 5953 mutex_exit(&dev->l2ad_mtx); 5954 5955 /* 5956 * We wait for the hash lock to become available 5957 * to try and prevent busy waiting, and increase 5958 * the chance we'll be able to acquire the lock 5959 * the next time around. 5960 */ 5961 mutex_enter(hash_lock); 5962 mutex_exit(hash_lock); 5963 goto top; 5964 } 5965 5966 /* 5967 * We could not have been moved into the arc_l2c_only 5968 * state while in-flight due to our ARC_FLAG_L2_WRITING 5969 * bit being set. Let's just ensure that's being enforced. 5970 */ 5971 ASSERT(HDR_HAS_L1HDR(hdr)); 5972 5973 /* 5974 * We may have allocated a buffer for L2ARC compression, 5975 * we must release it to avoid leaking this data. 5976 */ 5977 l2arc_release_cdata_buf(hdr); 5978 5979 if (zio->io_error != 0) { 5980 /* 5981 * Error - drop L2ARC entry. 5982 */ 5983 list_remove(buflist, hdr); 5984 l2arc_trim(hdr); 5985 hdr->b_flags &= ~ARC_FLAG_HAS_L2HDR; 5986 5987 ARCSTAT_INCR(arcstat_l2_asize, -hdr->b_l2hdr.b_asize); 5988 ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size); 5989 5990 bytes_dropped += hdr->b_l2hdr.b_asize; 5991 (void) refcount_remove_many(&dev->l2ad_alloc, 5992 hdr->b_l2hdr.b_asize, hdr); 5993 } 5994 5995 /* 5996 * Allow ARC to begin reads and ghost list evictions to 5997 * this L2ARC entry. 5998 */ 5999 hdr->b_flags &= ~ARC_FLAG_L2_WRITING; 6000 6001 mutex_exit(hash_lock); 6002 } 6003 6004 atomic_inc_64(&l2arc_writes_done); 6005 list_remove(buflist, head); 6006 ASSERT(!HDR_HAS_L1HDR(head)); 6007 kmem_cache_free(hdr_l2only_cache, head); 6008 mutex_exit(&dev->l2ad_mtx); 6009 6010 vdev_space_update(dev->l2ad_vdev, -bytes_dropped, 0, 0); 6011 6012 l2arc_do_free_on_write(); 6013 6014 kmem_free(cb, sizeof (l2arc_write_callback_t)); 6015} 6016 6017/* 6018 * A read to a cache device completed. Validate buffer contents before 6019 * handing over to the regular ARC routines. 6020 */ 6021static void 6022l2arc_read_done(zio_t *zio) 6023{ 6024 l2arc_read_callback_t *cb; 6025 arc_buf_hdr_t *hdr; 6026 arc_buf_t *buf; 6027 kmutex_t *hash_lock; 6028 int equal; 6029 6030 ASSERT(zio->io_vd != NULL); 6031 ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE); 6032 6033 spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd); 6034 6035 cb = zio->io_private; 6036 ASSERT(cb != NULL); 6037 buf = cb->l2rcb_buf; 6038 ASSERT(buf != NULL); 6039 6040 hash_lock = HDR_LOCK(buf->b_hdr); 6041 mutex_enter(hash_lock); 6042 hdr = buf->b_hdr; 6043 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 6044 6045 /* 6046 * If the buffer was compressed, decompress it first. 6047 */ 6048 if (cb->l2rcb_compress != ZIO_COMPRESS_OFF) 6049 l2arc_decompress_zio(zio, hdr, cb->l2rcb_compress); 6050 ASSERT(zio->io_data != NULL); 6051 ASSERT3U(zio->io_size, ==, hdr->b_size); 6052 ASSERT3U(BP_GET_LSIZE(&cb->l2rcb_bp), ==, hdr->b_size); 6053 6054 /* 6055 * Check this survived the L2ARC journey. 6056 */ 6057 equal = arc_cksum_equal(buf); 6058 if (equal && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) { 6059 mutex_exit(hash_lock); 6060 zio->io_private = buf; 6061 zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */ 6062 zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */ 6063 arc_read_done(zio); 6064 } else { 6065 mutex_exit(hash_lock); 6066 /* 6067 * Buffer didn't survive caching. Increment stats and 6068 * reissue to the original storage device. 6069 */ 6070 if (zio->io_error != 0) { 6071 ARCSTAT_BUMP(arcstat_l2_io_error); 6072 } else { 6073 zio->io_error = SET_ERROR(EIO); 6074 } 6075 if (!equal) 6076 ARCSTAT_BUMP(arcstat_l2_cksum_bad); 6077 6078 /* 6079 * If there's no waiter, issue an async i/o to the primary 6080 * storage now. If there *is* a waiter, the caller must 6081 * issue the i/o in a context where it's OK to block. 6082 */ 6083 if (zio->io_waiter == NULL) { 6084 zio_t *pio = zio_unique_parent(zio); 6085 6086 ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL); 6087 6088 zio_nowait(zio_read(pio, cb->l2rcb_spa, &cb->l2rcb_bp, 6089 buf->b_data, hdr->b_size, arc_read_done, buf, 6090 zio->io_priority, cb->l2rcb_flags, &cb->l2rcb_zb)); 6091 } 6092 } 6093 6094 kmem_free(cb, sizeof (l2arc_read_callback_t)); 6095} 6096 6097/* 6098 * This is the list priority from which the L2ARC will search for pages to 6099 * cache. This is used within loops (0..3) to cycle through lists in the 6100 * desired order. This order can have a significant effect on cache 6101 * performance. 6102 * 6103 * Currently the metadata lists are hit first, MFU then MRU, followed by 6104 * the data lists. This function returns a locked list, and also returns 6105 * the lock pointer. 6106 */ 6107static multilist_sublist_t * 6108l2arc_sublist_lock(int list_num) 6109{ 6110 multilist_t *ml = NULL; 6111 unsigned int idx; 6112 6113 ASSERT(list_num >= 0 && list_num <= 3); 6114 6115 switch (list_num) { 6116 case 0: 6117 ml = &arc_mfu->arcs_list[ARC_BUFC_METADATA]; 6118 break; 6119 case 1: 6120 ml = &arc_mru->arcs_list[ARC_BUFC_METADATA]; 6121 break; 6122 case 2: 6123 ml = &arc_mfu->arcs_list[ARC_BUFC_DATA]; 6124 break; 6125 case 3: 6126 ml = &arc_mru->arcs_list[ARC_BUFC_DATA]; 6127 break; 6128 } 6129 6130 /* 6131 * Return a randomly-selected sublist. This is acceptable 6132 * because the caller feeds only a little bit of data for each 6133 * call (8MB). Subsequent calls will result in different 6134 * sublists being selected. 6135 */ 6136 idx = multilist_get_random_index(ml); 6137 return (multilist_sublist_lock(ml, idx)); 6138} 6139 6140/* 6141 * Evict buffers from the device write hand to the distance specified in 6142 * bytes. This distance may span populated buffers, it may span nothing. 6143 * This is clearing a region on the L2ARC device ready for writing. 6144 * If the 'all' boolean is set, every buffer is evicted. 6145 */ 6146static void 6147l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all) 6148{ 6149 list_t *buflist; 6150 arc_buf_hdr_t *hdr, *hdr_prev; 6151 kmutex_t *hash_lock; 6152 uint64_t taddr; 6153 6154 buflist = &dev->l2ad_buflist; 6155 6156 if (!all && dev->l2ad_first) { 6157 /* 6158 * This is the first sweep through the device. There is 6159 * nothing to evict. 6160 */ 6161 return; 6162 } 6163 6164 if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) { 6165 /* 6166 * When nearing the end of the device, evict to the end 6167 * before the device write hand jumps to the start. 6168 */ 6169 taddr = dev->l2ad_end; 6170 } else { 6171 taddr = dev->l2ad_hand + distance; 6172 } 6173 DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist, 6174 uint64_t, taddr, boolean_t, all); 6175 6176top: 6177 mutex_enter(&dev->l2ad_mtx); 6178 for (hdr = list_tail(buflist); hdr; hdr = hdr_prev) { 6179 hdr_prev = list_prev(buflist, hdr); 6180 6181 hash_lock = HDR_LOCK(hdr); 6182 6183 /* 6184 * We cannot use mutex_enter or else we can deadlock 6185 * with l2arc_write_buffers (due to swapping the order 6186 * the hash lock and l2ad_mtx are taken). 6187 */ 6188 if (!mutex_tryenter(hash_lock)) { 6189 /* 6190 * Missed the hash lock. Retry. 6191 */ 6192 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry); 6193 mutex_exit(&dev->l2ad_mtx); 6194 mutex_enter(hash_lock); 6195 mutex_exit(hash_lock); 6196 goto top; 6197 } 6198 6199 if (HDR_L2_WRITE_HEAD(hdr)) { 6200 /* 6201 * We hit a write head node. Leave it for 6202 * l2arc_write_done(). 6203 */ 6204 list_remove(buflist, hdr); 6205 mutex_exit(hash_lock); 6206 continue; 6207 } 6208 6209 if (!all && HDR_HAS_L2HDR(hdr) && 6210 (hdr->b_l2hdr.b_daddr > taddr || 6211 hdr->b_l2hdr.b_daddr < dev->l2ad_hand)) { 6212 /* 6213 * We've evicted to the target address, 6214 * or the end of the device. 6215 */ 6216 mutex_exit(hash_lock); 6217 break; 6218 } 6219 6220 ASSERT(HDR_HAS_L2HDR(hdr)); 6221 if (!HDR_HAS_L1HDR(hdr)) { 6222 ASSERT(!HDR_L2_READING(hdr)); 6223 /* 6224 * This doesn't exist in the ARC. Destroy. 6225 * arc_hdr_destroy() will call list_remove() 6226 * and decrement arcstat_l2_size. 6227 */ 6228 arc_change_state(arc_anon, hdr, hash_lock); 6229 arc_hdr_destroy(hdr); 6230 } else { 6231 ASSERT(hdr->b_l1hdr.b_state != arc_l2c_only); 6232 ARCSTAT_BUMP(arcstat_l2_evict_l1cached); 6233 /* 6234 * Invalidate issued or about to be issued 6235 * reads, since we may be about to write 6236 * over this location. 6237 */ 6238 if (HDR_L2_READING(hdr)) { 6239 ARCSTAT_BUMP(arcstat_l2_evict_reading); 6240 hdr->b_flags |= ARC_FLAG_L2_EVICTED; 6241 } 6242 6243 /* Ensure this header has finished being written */ 6244 ASSERT(!HDR_L2_WRITING(hdr)); 6245 ASSERT3P(hdr->b_l1hdr.b_tmp_cdata, ==, NULL); 6246 6247 arc_hdr_l2hdr_destroy(hdr); 6248 } 6249 mutex_exit(hash_lock); 6250 } 6251 mutex_exit(&dev->l2ad_mtx); 6252} 6253 6254/* 6255 * Find and write ARC buffers to the L2ARC device. 6256 * 6257 * An ARC_FLAG_L2_WRITING flag is set so that the L2ARC buffers are not valid 6258 * for reading until they have completed writing. 6259 * The headroom_boost is an in-out parameter used to maintain headroom boost 6260 * state between calls to this function. 6261 * 6262 * Returns the number of bytes actually written (which may be smaller than 6263 * the delta by which the device hand has changed due to alignment). 6264 */ 6265static uint64_t 6266l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz, 6267 boolean_t *headroom_boost) 6268{ 6269 arc_buf_hdr_t *hdr, *hdr_prev, *head; 6270 uint64_t write_asize, write_sz, headroom, 6271 buf_compress_minsz; 6272 void *buf_data; 6273 boolean_t full; 6274 l2arc_write_callback_t *cb; 6275 zio_t *pio, *wzio; 6276 uint64_t guid = spa_load_guid(spa); 6277 const boolean_t do_headroom_boost = *headroom_boost; 6278 int try; 6279 6280 ASSERT(dev->l2ad_vdev != NULL); 6281 6282 /* Lower the flag now, we might want to raise it again later. */ 6283 *headroom_boost = B_FALSE; 6284 6285 pio = NULL; 6286 write_sz = write_asize = 0; 6287 full = B_FALSE; 6288 head = kmem_cache_alloc(hdr_l2only_cache, KM_PUSHPAGE); 6289 head->b_flags |= ARC_FLAG_L2_WRITE_HEAD; 6290 head->b_flags |= ARC_FLAG_HAS_L2HDR; 6291 6292 ARCSTAT_BUMP(arcstat_l2_write_buffer_iter); 6293 /* 6294 * We will want to try to compress buffers that are at least 2x the 6295 * device sector size. 6296 */ 6297 buf_compress_minsz = 2 << dev->l2ad_vdev->vdev_ashift; 6298 6299 /* 6300 * Copy buffers for L2ARC writing. 6301 */ 6302 for (try = 0; try <= 3; try++) { 6303 multilist_sublist_t *mls = l2arc_sublist_lock(try); 6304 uint64_t passed_sz = 0; 6305 6306 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_iter); 6307 6308 /* 6309 * L2ARC fast warmup. 6310 * 6311 * Until the ARC is warm and starts to evict, read from the 6312 * head of the ARC lists rather than the tail. 6313 */ 6314 if (arc_warm == B_FALSE) 6315 hdr = multilist_sublist_head(mls); 6316 else 6317 hdr = multilist_sublist_tail(mls); 6318 if (hdr == NULL) 6319 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_null_iter); 6320 6321 headroom = target_sz * l2arc_headroom; 6322 if (do_headroom_boost) 6323 headroom = (headroom * l2arc_headroom_boost) / 100; 6324 6325 for (; hdr; hdr = hdr_prev) { 6326 kmutex_t *hash_lock; 6327 uint64_t buf_sz; 6328 uint64_t buf_a_sz; 6329 6330 if (arc_warm == B_FALSE) 6331 hdr_prev = multilist_sublist_next(mls, hdr); 6332 else 6333 hdr_prev = multilist_sublist_prev(mls, hdr); 6334 ARCSTAT_INCR(arcstat_l2_write_buffer_bytes_scanned, hdr->b_size); 6335 6336 hash_lock = HDR_LOCK(hdr); 6337 if (!mutex_tryenter(hash_lock)) { 6338 ARCSTAT_BUMP(arcstat_l2_write_trylock_fail); 6339 /* 6340 * Skip this buffer rather than waiting. 6341 */ 6342 continue; 6343 } 6344 6345 passed_sz += hdr->b_size; 6346 if (passed_sz > headroom) { 6347 /* 6348 * Searched too far. 6349 */ 6350 mutex_exit(hash_lock); 6351 ARCSTAT_BUMP(arcstat_l2_write_passed_headroom); 6352 break; 6353 } 6354 6355 if (!l2arc_write_eligible(guid, hdr)) { 6356 mutex_exit(hash_lock); 6357 continue; 6358 } 6359 6360 /* 6361 * Assume that the buffer is not going to be compressed 6362 * and could take more space on disk because of a larger 6363 * disk block size. 6364 */ 6365 buf_sz = hdr->b_size; 6366 buf_a_sz = vdev_psize_to_asize(dev->l2ad_vdev, buf_sz); 6367 6368 if ((write_asize + buf_a_sz) > target_sz) { 6369 full = B_TRUE; 6370 mutex_exit(hash_lock); 6371 ARCSTAT_BUMP(arcstat_l2_write_full); 6372 break; 6373 } 6374 6375 if (pio == NULL) { 6376 /* 6377 * Insert a dummy header on the buflist so 6378 * l2arc_write_done() can find where the 6379 * write buffers begin without searching. 6380 */ 6381 mutex_enter(&dev->l2ad_mtx); 6382 list_insert_head(&dev->l2ad_buflist, head); 6383 mutex_exit(&dev->l2ad_mtx); 6384 6385 cb = kmem_alloc( 6386 sizeof (l2arc_write_callback_t), KM_SLEEP); 6387 cb->l2wcb_dev = dev; 6388 cb->l2wcb_head = head; 6389 pio = zio_root(spa, l2arc_write_done, cb, 6390 ZIO_FLAG_CANFAIL); 6391 ARCSTAT_BUMP(arcstat_l2_write_pios); 6392 } 6393 6394 /* 6395 * Create and add a new L2ARC header. 6396 */ 6397 hdr->b_l2hdr.b_dev = dev; 6398 hdr->b_flags |= ARC_FLAG_L2_WRITING; 6399 /* 6400 * Temporarily stash the data buffer in b_tmp_cdata. 6401 * The subsequent write step will pick it up from 6402 * there. This is because can't access b_l1hdr.b_buf 6403 * without holding the hash_lock, which we in turn 6404 * can't access without holding the ARC list locks 6405 * (which we want to avoid during compression/writing). 6406 */ 6407 hdr->b_l2hdr.b_compress = ZIO_COMPRESS_OFF; 6408 hdr->b_l2hdr.b_asize = hdr->b_size; 6409 hdr->b_l1hdr.b_tmp_cdata = hdr->b_l1hdr.b_buf->b_data; 6410 6411 /* 6412 * Explicitly set the b_daddr field to a known 6413 * value which means "invalid address". This 6414 * enables us to differentiate which stage of 6415 * l2arc_write_buffers() the particular header 6416 * is in (e.g. this loop, or the one below). 6417 * ARC_FLAG_L2_WRITING is not enough to make 6418 * this distinction, and we need to know in 6419 * order to do proper l2arc vdev accounting in 6420 * arc_release() and arc_hdr_destroy(). 6421 * 6422 * Note, we can't use a new flag to distinguish 6423 * the two stages because we don't hold the 6424 * header's hash_lock below, in the second stage 6425 * of this function. Thus, we can't simply 6426 * change the b_flags field to denote that the 6427 * IO has been sent. We can change the b_daddr 6428 * field of the L2 portion, though, since we'll 6429 * be holding the l2ad_mtx; which is why we're 6430 * using it to denote the header's state change. 6431 */ 6432 hdr->b_l2hdr.b_daddr = L2ARC_ADDR_UNSET; 6433 6434 hdr->b_flags |= ARC_FLAG_HAS_L2HDR; 6435 6436 mutex_enter(&dev->l2ad_mtx); 6437 list_insert_head(&dev->l2ad_buflist, hdr); 6438 mutex_exit(&dev->l2ad_mtx); 6439 6440 /* 6441 * Compute and store the buffer cksum before 6442 * writing. On debug the cksum is verified first. 6443 */ 6444 arc_cksum_verify(hdr->b_l1hdr.b_buf); 6445 arc_cksum_compute(hdr->b_l1hdr.b_buf, B_TRUE); 6446 6447 mutex_exit(hash_lock); 6448 6449 write_sz += buf_sz; 6450 write_asize += buf_a_sz; 6451 } 6452 6453 multilist_sublist_unlock(mls); 6454 6455 if (full == B_TRUE) 6456 break; 6457 } 6458 6459 /* No buffers selected for writing? */ 6460 if (pio == NULL) { 6461 ASSERT0(write_sz); 6462 ASSERT(!HDR_HAS_L1HDR(head)); 6463 kmem_cache_free(hdr_l2only_cache, head); 6464 return (0); 6465 } 6466 6467 mutex_enter(&dev->l2ad_mtx); 6468 6469 /* 6470 * Note that elsewhere in this file arcstat_l2_asize 6471 * and the used space on l2ad_vdev are updated using b_asize, 6472 * which is not necessarily rounded up to the device block size. 6473 * Too keep accounting consistent we do the same here as well: 6474 * stats_size accumulates the sum of b_asize of the written buffers, 6475 * while write_asize accumulates the sum of b_asize rounded up 6476 * to the device block size. 6477 * The latter sum is used only to validate the corectness of the code. 6478 */ 6479 uint64_t stats_size = 0; 6480 write_asize = 0; 6481 6482 /* 6483 * Now start writing the buffers. We're starting at the write head 6484 * and work backwards, retracing the course of the buffer selector 6485 * loop above. 6486 */ 6487 for (hdr = list_prev(&dev->l2ad_buflist, head); hdr; 6488 hdr = list_prev(&dev->l2ad_buflist, hdr)) { 6489 uint64_t buf_sz; 6490 6491 /* 6492 * We rely on the L1 portion of the header below, so 6493 * it's invalid for this header to have been evicted out 6494 * of the ghost cache, prior to being written out. The 6495 * ARC_FLAG_L2_WRITING bit ensures this won't happen. 6496 */ 6497 ASSERT(HDR_HAS_L1HDR(hdr)); 6498 6499 /* 6500 * We shouldn't need to lock the buffer here, since we flagged 6501 * it as ARC_FLAG_L2_WRITING in the previous step, but we must 6502 * take care to only access its L2 cache parameters. In 6503 * particular, hdr->l1hdr.b_buf may be invalid by now due to 6504 * ARC eviction. 6505 */ 6506 hdr->b_l2hdr.b_daddr = dev->l2ad_hand; 6507 6508 if ((HDR_L2COMPRESS(hdr)) && 6509 hdr->b_l2hdr.b_asize >= buf_compress_minsz) { 6510 if (l2arc_compress_buf(hdr)) { 6511 /* 6512 * If compression succeeded, enable headroom 6513 * boost on the next scan cycle. 6514 */ 6515 *headroom_boost = B_TRUE; 6516 } 6517 } 6518 6519 /* 6520 * Pick up the buffer data we had previously stashed away 6521 * (and now potentially also compressed). 6522 */ 6523 buf_data = hdr->b_l1hdr.b_tmp_cdata; 6524 buf_sz = hdr->b_l2hdr.b_asize; 6525 6526 /* 6527 * We need to do this regardless if buf_sz is zero or 6528 * not, otherwise, when this l2hdr is evicted we'll 6529 * remove a reference that was never added. 6530 */ 6531 (void) refcount_add_many(&dev->l2ad_alloc, buf_sz, hdr); 6532 6533 /* Compression may have squashed the buffer to zero length. */ 6534 if (buf_sz != 0) { 6535 uint64_t buf_a_sz; 6536 6537 wzio = zio_write_phys(pio, dev->l2ad_vdev, 6538 dev->l2ad_hand, buf_sz, buf_data, ZIO_CHECKSUM_OFF, 6539 NULL, NULL, ZIO_PRIORITY_ASYNC_WRITE, 6540 ZIO_FLAG_CANFAIL, B_FALSE); 6541 6542 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev, 6543 zio_t *, wzio); 6544 (void) zio_nowait(wzio); 6545 6546 stats_size += buf_sz; 6547 6548 /* 6549 * Keep the clock hand suitably device-aligned. 6550 */ 6551 buf_a_sz = vdev_psize_to_asize(dev->l2ad_vdev, buf_sz); 6552 write_asize += buf_a_sz; 6553 dev->l2ad_hand += buf_a_sz; 6554 } 6555 } 6556 6557 mutex_exit(&dev->l2ad_mtx); 6558 6559 ASSERT3U(write_asize, <=, target_sz); 6560 ARCSTAT_BUMP(arcstat_l2_writes_sent); 6561 ARCSTAT_INCR(arcstat_l2_write_bytes, write_asize); 6562 ARCSTAT_INCR(arcstat_l2_size, write_sz); 6563 ARCSTAT_INCR(arcstat_l2_asize, stats_size); 6564 vdev_space_update(dev->l2ad_vdev, stats_size, 0, 0); 6565 6566 /* 6567 * Bump device hand to the device start if it is approaching the end. 6568 * l2arc_evict() will already have evicted ahead for this case. 6569 */ 6570 if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) { 6571 dev->l2ad_hand = dev->l2ad_start; 6572 dev->l2ad_first = B_FALSE; 6573 } 6574 6575 dev->l2ad_writing = B_TRUE; 6576 (void) zio_wait(pio); 6577 dev->l2ad_writing = B_FALSE; 6578 6579 return (write_asize); 6580} 6581 6582/* 6583 * Compresses an L2ARC buffer. 6584 * The data to be compressed must be prefilled in l1hdr.b_tmp_cdata and its 6585 * size in l2hdr->b_asize. This routine tries to compress the data and 6586 * depending on the compression result there are three possible outcomes: 6587 * *) The buffer was incompressible. The original l2hdr contents were left 6588 * untouched and are ready for writing to an L2 device. 6589 * *) The buffer was all-zeros, so there is no need to write it to an L2 6590 * device. To indicate this situation b_tmp_cdata is NULL'ed, b_asize is 6591 * set to zero and b_compress is set to ZIO_COMPRESS_EMPTY. 6592 * *) Compression succeeded and b_tmp_cdata was replaced with a temporary 6593 * data buffer which holds the compressed data to be written, and b_asize 6594 * tells us how much data there is. b_compress is set to the appropriate 6595 * compression algorithm. Once writing is done, invoke 6596 * l2arc_release_cdata_buf on this l2hdr to free this temporary buffer. 6597 * 6598 * Returns B_TRUE if compression succeeded, or B_FALSE if it didn't (the 6599 * buffer was incompressible). 6600 */ 6601static boolean_t 6602l2arc_compress_buf(arc_buf_hdr_t *hdr) 6603{ 6604 void *cdata; 6605 size_t csize, len, rounded; 6606 ASSERT(HDR_HAS_L2HDR(hdr)); 6607 l2arc_buf_hdr_t *l2hdr = &hdr->b_l2hdr; 6608 6609 ASSERT(HDR_HAS_L1HDR(hdr)); 6610 ASSERT3S(l2hdr->b_compress, ==, ZIO_COMPRESS_OFF); 6611 ASSERT(hdr->b_l1hdr.b_tmp_cdata != NULL); 6612 6613 len = l2hdr->b_asize; 6614 cdata = zio_data_buf_alloc(len); 6615 ASSERT3P(cdata, !=, NULL); 6616 csize = zio_compress_data(ZIO_COMPRESS_LZ4, hdr->b_l1hdr.b_tmp_cdata, 6617 cdata, l2hdr->b_asize); 6618 6619 if (csize == 0) { 6620 /* zero block, indicate that there's nothing to write */ 6621 zio_data_buf_free(cdata, len); 6622 l2hdr->b_compress = ZIO_COMPRESS_EMPTY; 6623 l2hdr->b_asize = 0; 6624 hdr->b_l1hdr.b_tmp_cdata = NULL; 6625 ARCSTAT_BUMP(arcstat_l2_compress_zeros); 6626 return (B_TRUE); 6627 } 6628 6629 rounded = P2ROUNDUP(csize, 6630 (size_t)1 << l2hdr->b_dev->l2ad_vdev->vdev_ashift); 6631 if (rounded < len) { 6632 /* 6633 * Compression succeeded, we'll keep the cdata around for 6634 * writing and release it afterwards. 6635 */ 6636 if (rounded > csize) { 6637 bzero((char *)cdata + csize, rounded - csize); 6638 csize = rounded; 6639 } 6640 l2hdr->b_compress = ZIO_COMPRESS_LZ4; 6641 l2hdr->b_asize = csize; 6642 hdr->b_l1hdr.b_tmp_cdata = cdata; 6643 ARCSTAT_BUMP(arcstat_l2_compress_successes); 6644 return (B_TRUE); 6645 } else { 6646 /* 6647 * Compression failed, release the compressed buffer. 6648 * l2hdr will be left unmodified. 6649 */ 6650 zio_data_buf_free(cdata, len); 6651 ARCSTAT_BUMP(arcstat_l2_compress_failures); 6652 return (B_FALSE); 6653 } 6654} 6655 6656/* 6657 * Decompresses a zio read back from an l2arc device. On success, the 6658 * underlying zio's io_data buffer is overwritten by the uncompressed 6659 * version. On decompression error (corrupt compressed stream), the 6660 * zio->io_error value is set to signal an I/O error. 6661 * 6662 * Please note that the compressed data stream is not checksummed, so 6663 * if the underlying device is experiencing data corruption, we may feed 6664 * corrupt data to the decompressor, so the decompressor needs to be 6665 * able to handle this situation (LZ4 does). 6666 */ 6667static void 6668l2arc_decompress_zio(zio_t *zio, arc_buf_hdr_t *hdr, enum zio_compress c) 6669{ 6670 ASSERT(L2ARC_IS_VALID_COMPRESS(c)); 6671 6672 if (zio->io_error != 0) { 6673 /* 6674 * An io error has occured, just restore the original io 6675 * size in preparation for a main pool read. 6676 */ 6677 zio->io_orig_size = zio->io_size = hdr->b_size; 6678 return; 6679 } 6680 6681 if (c == ZIO_COMPRESS_EMPTY) { 6682 /* 6683 * An empty buffer results in a null zio, which means we 6684 * need to fill its io_data after we're done restoring the 6685 * buffer's contents. 6686 */ 6687 ASSERT(hdr->b_l1hdr.b_buf != NULL); 6688 bzero(hdr->b_l1hdr.b_buf->b_data, hdr->b_size); 6689 zio->io_data = zio->io_orig_data = hdr->b_l1hdr.b_buf->b_data; 6690 } else { 6691 ASSERT(zio->io_data != NULL); 6692 /* 6693 * We copy the compressed data from the start of the arc buffer 6694 * (the zio_read will have pulled in only what we need, the 6695 * rest is garbage which we will overwrite at decompression) 6696 * and then decompress back to the ARC data buffer. This way we 6697 * can minimize copying by simply decompressing back over the 6698 * original compressed data (rather than decompressing to an 6699 * aux buffer and then copying back the uncompressed buffer, 6700 * which is likely to be much larger). 6701 */ 6702 uint64_t csize; 6703 void *cdata; 6704 6705 csize = zio->io_size; 6706 cdata = zio_data_buf_alloc(csize); 6707 bcopy(zio->io_data, cdata, csize); 6708 if (zio_decompress_data(c, cdata, zio->io_data, csize, 6709 hdr->b_size) != 0) 6710 zio->io_error = EIO; 6711 zio_data_buf_free(cdata, csize); 6712 } 6713 6714 /* Restore the expected uncompressed IO size. */ 6715 zio->io_orig_size = zio->io_size = hdr->b_size; 6716} 6717 6718/* 6719 * Releases the temporary b_tmp_cdata buffer in an l2arc header structure. 6720 * This buffer serves as a temporary holder of compressed data while 6721 * the buffer entry is being written to an l2arc device. Once that is 6722 * done, we can dispose of it. 6723 */ 6724static void 6725l2arc_release_cdata_buf(arc_buf_hdr_t *hdr) 6726{ 6727 ASSERT(HDR_HAS_L2HDR(hdr)); 6728 enum zio_compress comp = hdr->b_l2hdr.b_compress; 6729 6730 ASSERT(HDR_HAS_L1HDR(hdr)); 6731 ASSERT(comp == ZIO_COMPRESS_OFF || L2ARC_IS_VALID_COMPRESS(comp)); 6732 6733 if (comp == ZIO_COMPRESS_OFF) { 6734 /* 6735 * In this case, b_tmp_cdata points to the same buffer 6736 * as the arc_buf_t's b_data field. We don't want to 6737 * free it, since the arc_buf_t will handle that. 6738 */ 6739 hdr->b_l1hdr.b_tmp_cdata = NULL; 6740 } else if (comp == ZIO_COMPRESS_EMPTY) { 6741 /* 6742 * In this case, b_tmp_cdata was compressed to an empty 6743 * buffer, thus there's nothing to free and b_tmp_cdata 6744 * should have been set to NULL in l2arc_write_buffers(). 6745 */ 6746 ASSERT3P(hdr->b_l1hdr.b_tmp_cdata, ==, NULL); 6747 } else { 6748 /* 6749 * If the data was compressed, then we've allocated a 6750 * temporary buffer for it, so now we need to release it. 6751 */ 6752 ASSERT(hdr->b_l1hdr.b_tmp_cdata != NULL); 6753 zio_data_buf_free(hdr->b_l1hdr.b_tmp_cdata, 6754 hdr->b_size); 6755 hdr->b_l1hdr.b_tmp_cdata = NULL; 6756 } 6757 6758} 6759 6760/* 6761 * This thread feeds the L2ARC at regular intervals. This is the beating 6762 * heart of the L2ARC. 6763 */ 6764static void 6765l2arc_feed_thread(void *dummy __unused) 6766{ 6767 callb_cpr_t cpr; 6768 l2arc_dev_t *dev; 6769 spa_t *spa; 6770 uint64_t size, wrote; 6771 clock_t begin, next = ddi_get_lbolt(); 6772 boolean_t headroom_boost = B_FALSE; 6773 6774 CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG); 6775 6776 mutex_enter(&l2arc_feed_thr_lock); 6777 6778 while (l2arc_thread_exit == 0) { 6779 CALLB_CPR_SAFE_BEGIN(&cpr); 6780 (void) cv_timedwait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock, 6781 next - ddi_get_lbolt()); 6782 CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock); 6783 next = ddi_get_lbolt() + hz; 6784 6785 /* 6786 * Quick check for L2ARC devices. 6787 */ 6788 mutex_enter(&l2arc_dev_mtx); 6789 if (l2arc_ndev == 0) { 6790 mutex_exit(&l2arc_dev_mtx); 6791 continue; 6792 } 6793 mutex_exit(&l2arc_dev_mtx); 6794 begin = ddi_get_lbolt(); 6795 6796 /* 6797 * This selects the next l2arc device to write to, and in 6798 * doing so the next spa to feed from: dev->l2ad_spa. This 6799 * will return NULL if there are now no l2arc devices or if 6800 * they are all faulted. 6801 * 6802 * If a device is returned, its spa's config lock is also 6803 * held to prevent device removal. l2arc_dev_get_next() 6804 * will grab and release l2arc_dev_mtx. 6805 */ 6806 if ((dev = l2arc_dev_get_next()) == NULL) 6807 continue; 6808 6809 spa = dev->l2ad_spa; 6810 ASSERT(spa != NULL); 6811 6812 /* 6813 * If the pool is read-only then force the feed thread to 6814 * sleep a little longer. 6815 */ 6816 if (!spa_writeable(spa)) { 6817 next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz; 6818 spa_config_exit(spa, SCL_L2ARC, dev); 6819 continue; 6820 } 6821 6822 /* 6823 * Avoid contributing to memory pressure. 6824 */ 6825 if (arc_reclaim_needed()) { 6826 ARCSTAT_BUMP(arcstat_l2_abort_lowmem); 6827 spa_config_exit(spa, SCL_L2ARC, dev); 6828 continue; 6829 } 6830 6831 ARCSTAT_BUMP(arcstat_l2_feeds); 6832 6833 size = l2arc_write_size(); 6834 6835 /* 6836 * Evict L2ARC buffers that will be overwritten. 6837 */ 6838 l2arc_evict(dev, size, B_FALSE); 6839 6840 /* 6841 * Write ARC buffers. 6842 */ 6843 wrote = l2arc_write_buffers(spa, dev, size, &headroom_boost); 6844 6845 /* 6846 * Calculate interval between writes. 6847 */ 6848 next = l2arc_write_interval(begin, size, wrote); 6849 spa_config_exit(spa, SCL_L2ARC, dev); 6850 } 6851 6852 l2arc_thread_exit = 0; 6853 cv_broadcast(&l2arc_feed_thr_cv); 6854 CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */ 6855 thread_exit(); 6856} 6857 6858boolean_t 6859l2arc_vdev_present(vdev_t *vd) 6860{ 6861 l2arc_dev_t *dev; 6862 6863 mutex_enter(&l2arc_dev_mtx); 6864 for (dev = list_head(l2arc_dev_list); dev != NULL; 6865 dev = list_next(l2arc_dev_list, dev)) { 6866 if (dev->l2ad_vdev == vd) 6867 break; 6868 } 6869 mutex_exit(&l2arc_dev_mtx); 6870 6871 return (dev != NULL); 6872} 6873 6874/* 6875 * Add a vdev for use by the L2ARC. By this point the spa has already 6876 * validated the vdev and opened it. 6877 */ 6878void 6879l2arc_add_vdev(spa_t *spa, vdev_t *vd) 6880{ 6881 l2arc_dev_t *adddev; 6882 6883 ASSERT(!l2arc_vdev_present(vd)); 6884 6885 vdev_ashift_optimize(vd); 6886 6887 /* 6888 * Create a new l2arc device entry. 6889 */ 6890 adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP); 6891 adddev->l2ad_spa = spa; 6892 adddev->l2ad_vdev = vd; 6893 adddev->l2ad_start = VDEV_LABEL_START_SIZE; 6894 adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd); 6895 adddev->l2ad_hand = adddev->l2ad_start; 6896 adddev->l2ad_first = B_TRUE; 6897 adddev->l2ad_writing = B_FALSE; 6898 6899 mutex_init(&adddev->l2ad_mtx, NULL, MUTEX_DEFAULT, NULL); 6900 /* 6901 * This is a list of all ARC buffers that are still valid on the 6902 * device. 6903 */ 6904 list_create(&adddev->l2ad_buflist, sizeof (arc_buf_hdr_t), 6905 offsetof(arc_buf_hdr_t, b_l2hdr.b_l2node)); 6906 6907 vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand); 6908 refcount_create(&adddev->l2ad_alloc); 6909 6910 /* 6911 * Add device to global list 6912 */ 6913 mutex_enter(&l2arc_dev_mtx); 6914 list_insert_head(l2arc_dev_list, adddev); 6915 atomic_inc_64(&l2arc_ndev); 6916 mutex_exit(&l2arc_dev_mtx); 6917} 6918 6919/* 6920 * Remove a vdev from the L2ARC. 6921 */ 6922void 6923l2arc_remove_vdev(vdev_t *vd) 6924{ 6925 l2arc_dev_t *dev, *nextdev, *remdev = NULL; 6926 6927 /* 6928 * Find the device by vdev 6929 */ 6930 mutex_enter(&l2arc_dev_mtx); 6931 for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) { 6932 nextdev = list_next(l2arc_dev_list, dev); 6933 if (vd == dev->l2ad_vdev) { 6934 remdev = dev; 6935 break; 6936 } 6937 } 6938 ASSERT(remdev != NULL); 6939 6940 /* 6941 * Remove device from global list 6942 */ 6943 list_remove(l2arc_dev_list, remdev); 6944 l2arc_dev_last = NULL; /* may have been invalidated */ 6945 atomic_dec_64(&l2arc_ndev); 6946 mutex_exit(&l2arc_dev_mtx); 6947 6948 /* 6949 * Clear all buflists and ARC references. L2ARC device flush. 6950 */ 6951 l2arc_evict(remdev, 0, B_TRUE); 6952 list_destroy(&remdev->l2ad_buflist); 6953 mutex_destroy(&remdev->l2ad_mtx); 6954 refcount_destroy(&remdev->l2ad_alloc); 6955 kmem_free(remdev, sizeof (l2arc_dev_t)); 6956} 6957 6958void 6959l2arc_init(void) 6960{ 6961 l2arc_thread_exit = 0; 6962 l2arc_ndev = 0; 6963 l2arc_writes_sent = 0; 6964 l2arc_writes_done = 0; 6965 6966 mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL); 6967 cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL); 6968 mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL); 6969 mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL); 6970 6971 l2arc_dev_list = &L2ARC_dev_list; 6972 l2arc_free_on_write = &L2ARC_free_on_write; 6973 list_create(l2arc_dev_list, sizeof (l2arc_dev_t), 6974 offsetof(l2arc_dev_t, l2ad_node)); 6975 list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t), 6976 offsetof(l2arc_data_free_t, l2df_list_node)); 6977} 6978 6979void 6980l2arc_fini(void) 6981{ 6982 /* 6983 * This is called from dmu_fini(), which is called from spa_fini(); 6984 * Because of this, we can assume that all l2arc devices have 6985 * already been removed when the pools themselves were removed. 6986 */ 6987 6988 l2arc_do_free_on_write(); 6989 6990 mutex_destroy(&l2arc_feed_thr_lock); 6991 cv_destroy(&l2arc_feed_thr_cv); 6992 mutex_destroy(&l2arc_dev_mtx); 6993 mutex_destroy(&l2arc_free_on_write_mtx); 6994 6995 list_destroy(l2arc_dev_list); 6996 list_destroy(l2arc_free_on_write); 6997} 6998 6999void 7000l2arc_start(void) 7001{ 7002 if (!(spa_mode_global & FWRITE)) 7003 return; 7004 7005 (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0, 7006 TS_RUN, minclsyspri); 7007} 7008 7009void 7010l2arc_stop(void) 7011{ 7012 if (!(spa_mode_global & FWRITE)) 7013 return; 7014 7015 mutex_enter(&l2arc_feed_thr_lock); 7016 cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */ 7017 l2arc_thread_exit = 1; 7018 while (l2arc_thread_exit != 0) 7019 cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock); 7020 mutex_exit(&l2arc_feed_thr_lock); 7021} 7022