arc.c revision 288596
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 uint64_t 1138buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth) 1139{ 1140 uint8_t *vdva = (uint8_t *)dva; 1141 uint64_t crc = -1ULL; 1142 int i; 1143 1144 ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY); 1145 1146 for (i = 0; i < sizeof (dva_t); i++) 1147 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ vdva[i]) & 0xFF]; 1148 1149 crc ^= (spa>>8) ^ birth; 1150 1151 return (crc); 1152} 1153 1154#define BUF_EMPTY(buf) \ 1155 ((buf)->b_dva.dva_word[0] == 0 && \ 1156 (buf)->b_dva.dva_word[1] == 0) 1157 1158#define BUF_EQUAL(spa, dva, birth, buf) \ 1159 ((buf)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \ 1160 ((buf)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \ 1161 ((buf)->b_birth == birth) && ((buf)->b_spa == spa) 1162 1163static void 1164buf_discard_identity(arc_buf_hdr_t *hdr) 1165{ 1166 hdr->b_dva.dva_word[0] = 0; 1167 hdr->b_dva.dva_word[1] = 0; 1168 hdr->b_birth = 0; 1169} 1170 1171static arc_buf_hdr_t * 1172buf_hash_find(uint64_t spa, const blkptr_t *bp, kmutex_t **lockp) 1173{ 1174 const dva_t *dva = BP_IDENTITY(bp); 1175 uint64_t birth = BP_PHYSICAL_BIRTH(bp); 1176 uint64_t idx = BUF_HASH_INDEX(spa, dva, birth); 1177 kmutex_t *hash_lock = BUF_HASH_LOCK(idx); 1178 arc_buf_hdr_t *hdr; 1179 1180 mutex_enter(hash_lock); 1181 for (hdr = buf_hash_table.ht_table[idx]; hdr != NULL; 1182 hdr = hdr->b_hash_next) { 1183 if (BUF_EQUAL(spa, dva, birth, hdr)) { 1184 *lockp = hash_lock; 1185 return (hdr); 1186 } 1187 } 1188 mutex_exit(hash_lock); 1189 *lockp = NULL; 1190 return (NULL); 1191} 1192 1193/* 1194 * Insert an entry into the hash table. If there is already an element 1195 * equal to elem in the hash table, then the already existing element 1196 * will be returned and the new element will not be inserted. 1197 * Otherwise returns NULL. 1198 * If lockp == NULL, the caller is assumed to already hold the hash lock. 1199 */ 1200static arc_buf_hdr_t * 1201buf_hash_insert(arc_buf_hdr_t *hdr, kmutex_t **lockp) 1202{ 1203 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth); 1204 kmutex_t *hash_lock = BUF_HASH_LOCK(idx); 1205 arc_buf_hdr_t *fhdr; 1206 uint32_t i; 1207 1208 ASSERT(!DVA_IS_EMPTY(&hdr->b_dva)); 1209 ASSERT(hdr->b_birth != 0); 1210 ASSERT(!HDR_IN_HASH_TABLE(hdr)); 1211 1212 if (lockp != NULL) { 1213 *lockp = hash_lock; 1214 mutex_enter(hash_lock); 1215 } else { 1216 ASSERT(MUTEX_HELD(hash_lock)); 1217 } 1218 1219 for (fhdr = buf_hash_table.ht_table[idx], i = 0; fhdr != NULL; 1220 fhdr = fhdr->b_hash_next, i++) { 1221 if (BUF_EQUAL(hdr->b_spa, &hdr->b_dva, hdr->b_birth, fhdr)) 1222 return (fhdr); 1223 } 1224 1225 hdr->b_hash_next = buf_hash_table.ht_table[idx]; 1226 buf_hash_table.ht_table[idx] = hdr; 1227 hdr->b_flags |= ARC_FLAG_IN_HASH_TABLE; 1228 1229 /* collect some hash table performance data */ 1230 if (i > 0) { 1231 ARCSTAT_BUMP(arcstat_hash_collisions); 1232 if (i == 1) 1233 ARCSTAT_BUMP(arcstat_hash_chains); 1234 1235 ARCSTAT_MAX(arcstat_hash_chain_max, i); 1236 } 1237 1238 ARCSTAT_BUMP(arcstat_hash_elements); 1239 ARCSTAT_MAXSTAT(arcstat_hash_elements); 1240 1241 return (NULL); 1242} 1243 1244static void 1245buf_hash_remove(arc_buf_hdr_t *hdr) 1246{ 1247 arc_buf_hdr_t *fhdr, **hdrp; 1248 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth); 1249 1250 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx))); 1251 ASSERT(HDR_IN_HASH_TABLE(hdr)); 1252 1253 hdrp = &buf_hash_table.ht_table[idx]; 1254 while ((fhdr = *hdrp) != hdr) { 1255 ASSERT(fhdr != NULL); 1256 hdrp = &fhdr->b_hash_next; 1257 } 1258 *hdrp = hdr->b_hash_next; 1259 hdr->b_hash_next = NULL; 1260 hdr->b_flags &= ~ARC_FLAG_IN_HASH_TABLE; 1261 1262 /* collect some hash table performance data */ 1263 ARCSTAT_BUMPDOWN(arcstat_hash_elements); 1264 1265 if (buf_hash_table.ht_table[idx] && 1266 buf_hash_table.ht_table[idx]->b_hash_next == NULL) 1267 ARCSTAT_BUMPDOWN(arcstat_hash_chains); 1268} 1269 1270/* 1271 * Global data structures and functions for the buf kmem cache. 1272 */ 1273static kmem_cache_t *hdr_full_cache; 1274static kmem_cache_t *hdr_l2only_cache; 1275static kmem_cache_t *buf_cache; 1276 1277static void 1278buf_fini(void) 1279{ 1280 int i; 1281 1282 kmem_free(buf_hash_table.ht_table, 1283 (buf_hash_table.ht_mask + 1) * sizeof (void *)); 1284 for (i = 0; i < BUF_LOCKS; i++) 1285 mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock); 1286 kmem_cache_destroy(hdr_full_cache); 1287 kmem_cache_destroy(hdr_l2only_cache); 1288 kmem_cache_destroy(buf_cache); 1289} 1290 1291/* 1292 * Constructor callback - called when the cache is empty 1293 * and a new buf is requested. 1294 */ 1295/* ARGSUSED */ 1296static int 1297hdr_full_cons(void *vbuf, void *unused, int kmflag) 1298{ 1299 arc_buf_hdr_t *hdr = vbuf; 1300 1301 bzero(hdr, HDR_FULL_SIZE); 1302 cv_init(&hdr->b_l1hdr.b_cv, NULL, CV_DEFAULT, NULL); 1303 refcount_create(&hdr->b_l1hdr.b_refcnt); 1304 mutex_init(&hdr->b_l1hdr.b_freeze_lock, NULL, MUTEX_DEFAULT, NULL); 1305 multilist_link_init(&hdr->b_l1hdr.b_arc_node); 1306 arc_space_consume(HDR_FULL_SIZE, ARC_SPACE_HDRS); 1307 1308 return (0); 1309} 1310 1311/* ARGSUSED */ 1312static int 1313hdr_l2only_cons(void *vbuf, void *unused, int kmflag) 1314{ 1315 arc_buf_hdr_t *hdr = vbuf; 1316 1317 bzero(hdr, HDR_L2ONLY_SIZE); 1318 arc_space_consume(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS); 1319 1320 return (0); 1321} 1322 1323/* ARGSUSED */ 1324static int 1325buf_cons(void *vbuf, void *unused, int kmflag) 1326{ 1327 arc_buf_t *buf = vbuf; 1328 1329 bzero(buf, sizeof (arc_buf_t)); 1330 mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL); 1331 arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS); 1332 1333 return (0); 1334} 1335 1336/* 1337 * Destructor callback - called when a cached buf is 1338 * no longer required. 1339 */ 1340/* ARGSUSED */ 1341static void 1342hdr_full_dest(void *vbuf, void *unused) 1343{ 1344 arc_buf_hdr_t *hdr = vbuf; 1345 1346 ASSERT(BUF_EMPTY(hdr)); 1347 cv_destroy(&hdr->b_l1hdr.b_cv); 1348 refcount_destroy(&hdr->b_l1hdr.b_refcnt); 1349 mutex_destroy(&hdr->b_l1hdr.b_freeze_lock); 1350 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 1351 arc_space_return(HDR_FULL_SIZE, ARC_SPACE_HDRS); 1352} 1353 1354/* ARGSUSED */ 1355static void 1356hdr_l2only_dest(void *vbuf, void *unused) 1357{ 1358 arc_buf_hdr_t *hdr = vbuf; 1359 1360 ASSERT(BUF_EMPTY(hdr)); 1361 arc_space_return(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS); 1362} 1363 1364/* ARGSUSED */ 1365static void 1366buf_dest(void *vbuf, void *unused) 1367{ 1368 arc_buf_t *buf = vbuf; 1369 1370 mutex_destroy(&buf->b_evict_lock); 1371 arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS); 1372} 1373 1374/* 1375 * Reclaim callback -- invoked when memory is low. 1376 */ 1377/* ARGSUSED */ 1378static void 1379hdr_recl(void *unused) 1380{ 1381 dprintf("hdr_recl called\n"); 1382 /* 1383 * umem calls the reclaim func when we destroy the buf cache, 1384 * which is after we do arc_fini(). 1385 */ 1386 if (!arc_dead) 1387 cv_signal(&arc_reclaim_thread_cv); 1388} 1389 1390static void 1391buf_init(void) 1392{ 1393 uint64_t *ct; 1394 uint64_t hsize = 1ULL << 12; 1395 int i, j; 1396 1397 /* 1398 * The hash table is big enough to fill all of physical memory 1399 * with an average block size of zfs_arc_average_blocksize (default 8K). 1400 * By default, the table will take up 1401 * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers). 1402 */ 1403 while (hsize * zfs_arc_average_blocksize < (uint64_t)physmem * PAGESIZE) 1404 hsize <<= 1; 1405retry: 1406 buf_hash_table.ht_mask = hsize - 1; 1407 buf_hash_table.ht_table = 1408 kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP); 1409 if (buf_hash_table.ht_table == NULL) { 1410 ASSERT(hsize > (1ULL << 8)); 1411 hsize >>= 1; 1412 goto retry; 1413 } 1414 1415 hdr_full_cache = kmem_cache_create("arc_buf_hdr_t_full", HDR_FULL_SIZE, 1416 0, hdr_full_cons, hdr_full_dest, hdr_recl, NULL, NULL, 0); 1417 hdr_l2only_cache = kmem_cache_create("arc_buf_hdr_t_l2only", 1418 HDR_L2ONLY_SIZE, 0, hdr_l2only_cons, hdr_l2only_dest, hdr_recl, 1419 NULL, NULL, 0); 1420 buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t), 1421 0, buf_cons, buf_dest, NULL, NULL, NULL, 0); 1422 1423 for (i = 0; i < 256; i++) 1424 for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--) 1425 *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY); 1426 1427 for (i = 0; i < BUF_LOCKS; i++) { 1428 mutex_init(&buf_hash_table.ht_locks[i].ht_lock, 1429 NULL, MUTEX_DEFAULT, NULL); 1430 } 1431} 1432 1433/* 1434 * Transition between the two allocation states for the arc_buf_hdr struct. 1435 * The arc_buf_hdr struct can be allocated with (hdr_full_cache) or without 1436 * (hdr_l2only_cache) the fields necessary for the L1 cache - the smaller 1437 * version is used when a cache buffer is only in the L2ARC in order to reduce 1438 * memory usage. 1439 */ 1440static arc_buf_hdr_t * 1441arc_hdr_realloc(arc_buf_hdr_t *hdr, kmem_cache_t *old, kmem_cache_t *new) 1442{ 1443 ASSERT(HDR_HAS_L2HDR(hdr)); 1444 1445 arc_buf_hdr_t *nhdr; 1446 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev; 1447 1448 ASSERT((old == hdr_full_cache && new == hdr_l2only_cache) || 1449 (old == hdr_l2only_cache && new == hdr_full_cache)); 1450 1451 nhdr = kmem_cache_alloc(new, KM_PUSHPAGE); 1452 1453 ASSERT(MUTEX_HELD(HDR_LOCK(hdr))); 1454 buf_hash_remove(hdr); 1455 1456 bcopy(hdr, nhdr, HDR_L2ONLY_SIZE); 1457 1458 if (new == hdr_full_cache) { 1459 nhdr->b_flags |= ARC_FLAG_HAS_L1HDR; 1460 /* 1461 * arc_access and arc_change_state need to be aware that a 1462 * header has just come out of L2ARC, so we set its state to 1463 * l2c_only even though it's about to change. 1464 */ 1465 nhdr->b_l1hdr.b_state = arc_l2c_only; 1466 1467 /* Verify previous threads set to NULL before freeing */ 1468 ASSERT3P(nhdr->b_l1hdr.b_tmp_cdata, ==, NULL); 1469 } else { 1470 ASSERT(hdr->b_l1hdr.b_buf == NULL); 1471 ASSERT0(hdr->b_l1hdr.b_datacnt); 1472 1473 /* 1474 * If we've reached here, We must have been called from 1475 * arc_evict_hdr(), as such we should have already been 1476 * removed from any ghost list we were previously on 1477 * (which protects us from racing with arc_evict_state), 1478 * thus no locking is needed during this check. 1479 */ 1480 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 1481 1482 /* 1483 * A buffer must not be moved into the arc_l2c_only 1484 * state if it's not finished being written out to the 1485 * l2arc device. Otherwise, the b_l1hdr.b_tmp_cdata field 1486 * might try to be accessed, even though it was removed. 1487 */ 1488 VERIFY(!HDR_L2_WRITING(hdr)); 1489 VERIFY3P(hdr->b_l1hdr.b_tmp_cdata, ==, NULL); 1490 1491 nhdr->b_flags &= ~ARC_FLAG_HAS_L1HDR; 1492 } 1493 /* 1494 * The header has been reallocated so we need to re-insert it into any 1495 * lists it was on. 1496 */ 1497 (void) buf_hash_insert(nhdr, NULL); 1498 1499 ASSERT(list_link_active(&hdr->b_l2hdr.b_l2node)); 1500 1501 mutex_enter(&dev->l2ad_mtx); 1502 1503 /* 1504 * We must place the realloc'ed header back into the list at 1505 * the same spot. Otherwise, if it's placed earlier in the list, 1506 * l2arc_write_buffers() could find it during the function's 1507 * write phase, and try to write it out to the l2arc. 1508 */ 1509 list_insert_after(&dev->l2ad_buflist, hdr, nhdr); 1510 list_remove(&dev->l2ad_buflist, hdr); 1511 1512 mutex_exit(&dev->l2ad_mtx); 1513 1514 /* 1515 * Since we're using the pointer address as the tag when 1516 * incrementing and decrementing the l2ad_alloc refcount, we 1517 * must remove the old pointer (that we're about to destroy) and 1518 * add the new pointer to the refcount. Otherwise we'd remove 1519 * the wrong pointer address when calling arc_hdr_destroy() later. 1520 */ 1521 1522 (void) refcount_remove_many(&dev->l2ad_alloc, 1523 hdr->b_l2hdr.b_asize, hdr); 1524 1525 (void) refcount_add_many(&dev->l2ad_alloc, 1526 nhdr->b_l2hdr.b_asize, nhdr); 1527 1528 buf_discard_identity(hdr); 1529 hdr->b_freeze_cksum = NULL; 1530 kmem_cache_free(old, hdr); 1531 1532 return (nhdr); 1533} 1534 1535 1536#define ARC_MINTIME (hz>>4) /* 62 ms */ 1537 1538static void 1539arc_cksum_verify(arc_buf_t *buf) 1540{ 1541 zio_cksum_t zc; 1542 1543 if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 1544 return; 1545 1546 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1547 if (buf->b_hdr->b_freeze_cksum == NULL || HDR_IO_ERROR(buf->b_hdr)) { 1548 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1549 return; 1550 } 1551 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc); 1552 if (!ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc)) 1553 panic("buffer modified while frozen!"); 1554 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1555} 1556 1557static int 1558arc_cksum_equal(arc_buf_t *buf) 1559{ 1560 zio_cksum_t zc; 1561 int equal; 1562 1563 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1564 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc); 1565 equal = ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc); 1566 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1567 1568 return (equal); 1569} 1570 1571static void 1572arc_cksum_compute(arc_buf_t *buf, boolean_t force) 1573{ 1574 if (!force && !(zfs_flags & ZFS_DEBUG_MODIFY)) 1575 return; 1576 1577 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1578 if (buf->b_hdr->b_freeze_cksum != NULL) { 1579 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1580 return; 1581 } 1582 buf->b_hdr->b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t), KM_SLEEP); 1583 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, 1584 buf->b_hdr->b_freeze_cksum); 1585 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1586#ifdef illumos 1587 arc_buf_watch(buf); 1588#endif /* illumos */ 1589} 1590 1591#ifdef illumos 1592#ifndef _KERNEL 1593typedef struct procctl { 1594 long cmd; 1595 prwatch_t prwatch; 1596} procctl_t; 1597#endif 1598 1599/* ARGSUSED */ 1600static void 1601arc_buf_unwatch(arc_buf_t *buf) 1602{ 1603#ifndef _KERNEL 1604 if (arc_watch) { 1605 int result; 1606 procctl_t ctl; 1607 ctl.cmd = PCWATCH; 1608 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data; 1609 ctl.prwatch.pr_size = 0; 1610 ctl.prwatch.pr_wflags = 0; 1611 result = write(arc_procfd, &ctl, sizeof (ctl)); 1612 ASSERT3U(result, ==, sizeof (ctl)); 1613 } 1614#endif 1615} 1616 1617/* ARGSUSED */ 1618static void 1619arc_buf_watch(arc_buf_t *buf) 1620{ 1621#ifndef _KERNEL 1622 if (arc_watch) { 1623 int result; 1624 procctl_t ctl; 1625 ctl.cmd = PCWATCH; 1626 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data; 1627 ctl.prwatch.pr_size = buf->b_hdr->b_size; 1628 ctl.prwatch.pr_wflags = WA_WRITE; 1629 result = write(arc_procfd, &ctl, sizeof (ctl)); 1630 ASSERT3U(result, ==, sizeof (ctl)); 1631 } 1632#endif 1633} 1634#endif /* illumos */ 1635 1636static arc_buf_contents_t 1637arc_buf_type(arc_buf_hdr_t *hdr) 1638{ 1639 if (HDR_ISTYPE_METADATA(hdr)) { 1640 return (ARC_BUFC_METADATA); 1641 } else { 1642 return (ARC_BUFC_DATA); 1643 } 1644} 1645 1646static uint32_t 1647arc_bufc_to_flags(arc_buf_contents_t type) 1648{ 1649 switch (type) { 1650 case ARC_BUFC_DATA: 1651 /* metadata field is 0 if buffer contains normal data */ 1652 return (0); 1653 case ARC_BUFC_METADATA: 1654 return (ARC_FLAG_BUFC_METADATA); 1655 default: 1656 break; 1657 } 1658 panic("undefined ARC buffer type!"); 1659 return ((uint32_t)-1); 1660} 1661 1662void 1663arc_buf_thaw(arc_buf_t *buf) 1664{ 1665 if (zfs_flags & ZFS_DEBUG_MODIFY) { 1666 if (buf->b_hdr->b_l1hdr.b_state != arc_anon) 1667 panic("modifying non-anon buffer!"); 1668 if (HDR_IO_IN_PROGRESS(buf->b_hdr)) 1669 panic("modifying buffer while i/o in progress!"); 1670 arc_cksum_verify(buf); 1671 } 1672 1673 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1674 if (buf->b_hdr->b_freeze_cksum != NULL) { 1675 kmem_free(buf->b_hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 1676 buf->b_hdr->b_freeze_cksum = NULL; 1677 } 1678 1679#ifdef ZFS_DEBUG 1680 if (zfs_flags & ZFS_DEBUG_MODIFY) { 1681 if (buf->b_hdr->b_l1hdr.b_thawed != NULL) 1682 kmem_free(buf->b_hdr->b_l1hdr.b_thawed, 1); 1683 buf->b_hdr->b_l1hdr.b_thawed = kmem_alloc(1, KM_SLEEP); 1684 } 1685#endif 1686 1687 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1688 1689#ifdef illumos 1690 arc_buf_unwatch(buf); 1691#endif /* illumos */ 1692} 1693 1694void 1695arc_buf_freeze(arc_buf_t *buf) 1696{ 1697 kmutex_t *hash_lock; 1698 1699 if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 1700 return; 1701 1702 hash_lock = HDR_LOCK(buf->b_hdr); 1703 mutex_enter(hash_lock); 1704 1705 ASSERT(buf->b_hdr->b_freeze_cksum != NULL || 1706 buf->b_hdr->b_l1hdr.b_state == arc_anon); 1707 arc_cksum_compute(buf, B_FALSE); 1708 mutex_exit(hash_lock); 1709 1710} 1711 1712static void 1713add_reference(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, void *tag) 1714{ 1715 ASSERT(HDR_HAS_L1HDR(hdr)); 1716 ASSERT(MUTEX_HELD(hash_lock)); 1717 arc_state_t *state = hdr->b_l1hdr.b_state; 1718 1719 if ((refcount_add(&hdr->b_l1hdr.b_refcnt, tag) == 1) && 1720 (state != arc_anon)) { 1721 /* We don't use the L2-only state list. */ 1722 if (state != arc_l2c_only) { 1723 arc_buf_contents_t type = arc_buf_type(hdr); 1724 uint64_t delta = hdr->b_size * hdr->b_l1hdr.b_datacnt; 1725 multilist_t *list = &state->arcs_list[type]; 1726 uint64_t *size = &state->arcs_lsize[type]; 1727 1728 multilist_remove(list, hdr); 1729 1730 if (GHOST_STATE(state)) { 1731 ASSERT0(hdr->b_l1hdr.b_datacnt); 1732 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 1733 delta = hdr->b_size; 1734 } 1735 ASSERT(delta > 0); 1736 ASSERT3U(*size, >=, delta); 1737 atomic_add_64(size, -delta); 1738 } 1739 /* remove the prefetch flag if we get a reference */ 1740 hdr->b_flags &= ~ARC_FLAG_PREFETCH; 1741 } 1742} 1743 1744static int 1745remove_reference(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, void *tag) 1746{ 1747 int cnt; 1748 arc_state_t *state = hdr->b_l1hdr.b_state; 1749 1750 ASSERT(HDR_HAS_L1HDR(hdr)); 1751 ASSERT(state == arc_anon || MUTEX_HELD(hash_lock)); 1752 ASSERT(!GHOST_STATE(state)); 1753 1754 /* 1755 * arc_l2c_only counts as a ghost state so we don't need to explicitly 1756 * check to prevent usage of the arc_l2c_only list. 1757 */ 1758 if (((cnt = refcount_remove(&hdr->b_l1hdr.b_refcnt, tag)) == 0) && 1759 (state != arc_anon)) { 1760 arc_buf_contents_t type = arc_buf_type(hdr); 1761 multilist_t *list = &state->arcs_list[type]; 1762 uint64_t *size = &state->arcs_lsize[type]; 1763 1764 multilist_insert(list, hdr); 1765 1766 ASSERT(hdr->b_l1hdr.b_datacnt > 0); 1767 atomic_add_64(size, hdr->b_size * 1768 hdr->b_l1hdr.b_datacnt); 1769 } 1770 return (cnt); 1771} 1772 1773/* 1774 * Move the supplied buffer to the indicated state. The hash lock 1775 * for the buffer must be held by the caller. 1776 */ 1777static void 1778arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *hdr, 1779 kmutex_t *hash_lock) 1780{ 1781 arc_state_t *old_state; 1782 int64_t refcnt; 1783 uint32_t datacnt; 1784 uint64_t from_delta, to_delta; 1785 arc_buf_contents_t buftype = arc_buf_type(hdr); 1786 1787 /* 1788 * We almost always have an L1 hdr here, since we call arc_hdr_realloc() 1789 * in arc_read() when bringing a buffer out of the L2ARC. However, the 1790 * L1 hdr doesn't always exist when we change state to arc_anon before 1791 * destroying a header, in which case reallocating to add the L1 hdr is 1792 * pointless. 1793 */ 1794 if (HDR_HAS_L1HDR(hdr)) { 1795 old_state = hdr->b_l1hdr.b_state; 1796 refcnt = refcount_count(&hdr->b_l1hdr.b_refcnt); 1797 datacnt = hdr->b_l1hdr.b_datacnt; 1798 } else { 1799 old_state = arc_l2c_only; 1800 refcnt = 0; 1801 datacnt = 0; 1802 } 1803 1804 ASSERT(MUTEX_HELD(hash_lock)); 1805 ASSERT3P(new_state, !=, old_state); 1806 ASSERT(refcnt == 0 || datacnt > 0); 1807 ASSERT(!GHOST_STATE(new_state) || datacnt == 0); 1808 ASSERT(old_state != arc_anon || datacnt <= 1); 1809 1810 from_delta = to_delta = datacnt * hdr->b_size; 1811 1812 /* 1813 * If this buffer is evictable, transfer it from the 1814 * old state list to the new state list. 1815 */ 1816 if (refcnt == 0) { 1817 if (old_state != arc_anon && old_state != arc_l2c_only) { 1818 uint64_t *size = &old_state->arcs_lsize[buftype]; 1819 1820 ASSERT(HDR_HAS_L1HDR(hdr)); 1821 multilist_remove(&old_state->arcs_list[buftype], hdr); 1822 1823 /* 1824 * If prefetching out of the ghost cache, 1825 * we will have a non-zero datacnt. 1826 */ 1827 if (GHOST_STATE(old_state) && datacnt == 0) { 1828 /* ghost elements have a ghost size */ 1829 ASSERT(hdr->b_l1hdr.b_buf == NULL); 1830 from_delta = hdr->b_size; 1831 } 1832 ASSERT3U(*size, >=, from_delta); 1833 atomic_add_64(size, -from_delta); 1834 } 1835 if (new_state != arc_anon && new_state != arc_l2c_only) { 1836 uint64_t *size = &new_state->arcs_lsize[buftype]; 1837 1838 /* 1839 * An L1 header always exists here, since if we're 1840 * moving to some L1-cached state (i.e. not l2c_only or 1841 * anonymous), we realloc the header to add an L1hdr 1842 * beforehand. 1843 */ 1844 ASSERT(HDR_HAS_L1HDR(hdr)); 1845 multilist_insert(&new_state->arcs_list[buftype], hdr); 1846 1847 /* ghost elements have a ghost size */ 1848 if (GHOST_STATE(new_state)) { 1849 ASSERT0(datacnt); 1850 ASSERT(hdr->b_l1hdr.b_buf == NULL); 1851 to_delta = hdr->b_size; 1852 } 1853 atomic_add_64(size, to_delta); 1854 } 1855 } 1856 1857 ASSERT(!BUF_EMPTY(hdr)); 1858 if (new_state == arc_anon && HDR_IN_HASH_TABLE(hdr)) 1859 buf_hash_remove(hdr); 1860 1861 /* adjust state sizes (ignore arc_l2c_only) */ 1862 1863 if (to_delta && new_state != arc_l2c_only) { 1864 ASSERT(HDR_HAS_L1HDR(hdr)); 1865 if (GHOST_STATE(new_state)) { 1866 ASSERT0(datacnt); 1867 1868 /* 1869 * We moving a header to a ghost state, we first 1870 * remove all arc buffers. Thus, we'll have a 1871 * datacnt of zero, and no arc buffer to use for 1872 * the reference. As a result, we use the arc 1873 * header pointer for the reference. 1874 */ 1875 (void) refcount_add_many(&new_state->arcs_size, 1876 hdr->b_size, hdr); 1877 } else { 1878 ASSERT3U(datacnt, !=, 0); 1879 1880 /* 1881 * Each individual buffer holds a unique reference, 1882 * thus we must remove each of these references one 1883 * at a time. 1884 */ 1885 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL; 1886 buf = buf->b_next) { 1887 (void) refcount_add_many(&new_state->arcs_size, 1888 hdr->b_size, buf); 1889 } 1890 } 1891 } 1892 1893 if (from_delta && old_state != arc_l2c_only) { 1894 ASSERT(HDR_HAS_L1HDR(hdr)); 1895 if (GHOST_STATE(old_state)) { 1896 /* 1897 * When moving a header off of a ghost state, 1898 * there's the possibility for datacnt to be 1899 * non-zero. This is because we first add the 1900 * arc buffer to the header prior to changing 1901 * the header's state. Since we used the header 1902 * for the reference when putting the header on 1903 * the ghost state, we must balance that and use 1904 * the header when removing off the ghost state 1905 * (even though datacnt is non zero). 1906 */ 1907 1908 IMPLY(datacnt == 0, new_state == arc_anon || 1909 new_state == arc_l2c_only); 1910 1911 (void) refcount_remove_many(&old_state->arcs_size, 1912 hdr->b_size, hdr); 1913 } else { 1914 ASSERT3P(datacnt, !=, 0); 1915 1916 /* 1917 * Each individual buffer holds a unique reference, 1918 * thus we must remove each of these references one 1919 * at a time. 1920 */ 1921 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL; 1922 buf = buf->b_next) { 1923 (void) refcount_remove_many( 1924 &old_state->arcs_size, hdr->b_size, buf); 1925 } 1926 } 1927 } 1928 1929 if (HDR_HAS_L1HDR(hdr)) 1930 hdr->b_l1hdr.b_state = new_state; 1931 1932 /* 1933 * L2 headers should never be on the L2 state list since they don't 1934 * have L1 headers allocated. 1935 */ 1936 ASSERT(multilist_is_empty(&arc_l2c_only->arcs_list[ARC_BUFC_DATA]) && 1937 multilist_is_empty(&arc_l2c_only->arcs_list[ARC_BUFC_METADATA])); 1938} 1939 1940void 1941arc_space_consume(uint64_t space, arc_space_type_t type) 1942{ 1943 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES); 1944 1945 switch (type) { 1946 case ARC_SPACE_DATA: 1947 ARCSTAT_INCR(arcstat_data_size, space); 1948 break; 1949 case ARC_SPACE_META: 1950 ARCSTAT_INCR(arcstat_metadata_size, space); 1951 break; 1952 case ARC_SPACE_OTHER: 1953 ARCSTAT_INCR(arcstat_other_size, space); 1954 break; 1955 case ARC_SPACE_HDRS: 1956 ARCSTAT_INCR(arcstat_hdr_size, space); 1957 break; 1958 case ARC_SPACE_L2HDRS: 1959 ARCSTAT_INCR(arcstat_l2_hdr_size, space); 1960 break; 1961 } 1962 1963 if (type != ARC_SPACE_DATA) 1964 ARCSTAT_INCR(arcstat_meta_used, space); 1965 1966 atomic_add_64(&arc_size, space); 1967} 1968 1969void 1970arc_space_return(uint64_t space, arc_space_type_t type) 1971{ 1972 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES); 1973 1974 switch (type) { 1975 case ARC_SPACE_DATA: 1976 ARCSTAT_INCR(arcstat_data_size, -space); 1977 break; 1978 case ARC_SPACE_META: 1979 ARCSTAT_INCR(arcstat_metadata_size, -space); 1980 break; 1981 case ARC_SPACE_OTHER: 1982 ARCSTAT_INCR(arcstat_other_size, -space); 1983 break; 1984 case ARC_SPACE_HDRS: 1985 ARCSTAT_INCR(arcstat_hdr_size, -space); 1986 break; 1987 case ARC_SPACE_L2HDRS: 1988 ARCSTAT_INCR(arcstat_l2_hdr_size, -space); 1989 break; 1990 } 1991 1992 if (type != ARC_SPACE_DATA) { 1993 ASSERT(arc_meta_used >= space); 1994 if (arc_meta_max < arc_meta_used) 1995 arc_meta_max = arc_meta_used; 1996 ARCSTAT_INCR(arcstat_meta_used, -space); 1997 } 1998 1999 ASSERT(arc_size >= space); 2000 atomic_add_64(&arc_size, -space); 2001} 2002 2003arc_buf_t * 2004arc_buf_alloc(spa_t *spa, int32_t size, void *tag, arc_buf_contents_t type) 2005{ 2006 arc_buf_hdr_t *hdr; 2007 arc_buf_t *buf; 2008 2009 ASSERT3U(size, >, 0); 2010 hdr = kmem_cache_alloc(hdr_full_cache, KM_PUSHPAGE); 2011 ASSERT(BUF_EMPTY(hdr)); 2012 ASSERT3P(hdr->b_freeze_cksum, ==, NULL); 2013 hdr->b_size = size; 2014 hdr->b_spa = spa_load_guid(spa); 2015 2016 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 2017 buf->b_hdr = hdr; 2018 buf->b_data = NULL; 2019 buf->b_efunc = NULL; 2020 buf->b_private = NULL; 2021 buf->b_next = NULL; 2022 2023 hdr->b_flags = arc_bufc_to_flags(type); 2024 hdr->b_flags |= ARC_FLAG_HAS_L1HDR; 2025 2026 hdr->b_l1hdr.b_buf = buf; 2027 hdr->b_l1hdr.b_state = arc_anon; 2028 hdr->b_l1hdr.b_arc_access = 0; 2029 hdr->b_l1hdr.b_datacnt = 1; 2030 hdr->b_l1hdr.b_tmp_cdata = NULL; 2031 2032 arc_get_data_buf(buf); 2033 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 2034 (void) refcount_add(&hdr->b_l1hdr.b_refcnt, tag); 2035 2036 return (buf); 2037} 2038 2039static char *arc_onloan_tag = "onloan"; 2040 2041/* 2042 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in 2043 * flight data by arc_tempreserve_space() until they are "returned". Loaned 2044 * buffers must be returned to the arc before they can be used by the DMU or 2045 * freed. 2046 */ 2047arc_buf_t * 2048arc_loan_buf(spa_t *spa, int size) 2049{ 2050 arc_buf_t *buf; 2051 2052 buf = arc_buf_alloc(spa, size, arc_onloan_tag, ARC_BUFC_DATA); 2053 2054 atomic_add_64(&arc_loaned_bytes, size); 2055 return (buf); 2056} 2057 2058/* 2059 * Return a loaned arc buffer to the arc. 2060 */ 2061void 2062arc_return_buf(arc_buf_t *buf, void *tag) 2063{ 2064 arc_buf_hdr_t *hdr = buf->b_hdr; 2065 2066 ASSERT(buf->b_data != NULL); 2067 ASSERT(HDR_HAS_L1HDR(hdr)); 2068 (void) refcount_add(&hdr->b_l1hdr.b_refcnt, tag); 2069 (void) refcount_remove(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag); 2070 2071 atomic_add_64(&arc_loaned_bytes, -hdr->b_size); 2072} 2073 2074/* Detach an arc_buf from a dbuf (tag) */ 2075void 2076arc_loan_inuse_buf(arc_buf_t *buf, void *tag) 2077{ 2078 arc_buf_hdr_t *hdr = buf->b_hdr; 2079 2080 ASSERT(buf->b_data != NULL); 2081 ASSERT(HDR_HAS_L1HDR(hdr)); 2082 (void) refcount_add(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag); 2083 (void) refcount_remove(&hdr->b_l1hdr.b_refcnt, tag); 2084 buf->b_efunc = NULL; 2085 buf->b_private = NULL; 2086 2087 atomic_add_64(&arc_loaned_bytes, hdr->b_size); 2088} 2089 2090static arc_buf_t * 2091arc_buf_clone(arc_buf_t *from) 2092{ 2093 arc_buf_t *buf; 2094 arc_buf_hdr_t *hdr = from->b_hdr; 2095 uint64_t size = hdr->b_size; 2096 2097 ASSERT(HDR_HAS_L1HDR(hdr)); 2098 ASSERT(hdr->b_l1hdr.b_state != arc_anon); 2099 2100 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 2101 buf->b_hdr = hdr; 2102 buf->b_data = NULL; 2103 buf->b_efunc = NULL; 2104 buf->b_private = NULL; 2105 buf->b_next = hdr->b_l1hdr.b_buf; 2106 hdr->b_l1hdr.b_buf = buf; 2107 arc_get_data_buf(buf); 2108 bcopy(from->b_data, buf->b_data, size); 2109 2110 /* 2111 * This buffer already exists in the arc so create a duplicate 2112 * copy for the caller. If the buffer is associated with user data 2113 * then track the size and number of duplicates. These stats will be 2114 * updated as duplicate buffers are created and destroyed. 2115 */ 2116 if (HDR_ISTYPE_DATA(hdr)) { 2117 ARCSTAT_BUMP(arcstat_duplicate_buffers); 2118 ARCSTAT_INCR(arcstat_duplicate_buffers_size, size); 2119 } 2120 hdr->b_l1hdr.b_datacnt += 1; 2121 return (buf); 2122} 2123 2124void 2125arc_buf_add_ref(arc_buf_t *buf, void* tag) 2126{ 2127 arc_buf_hdr_t *hdr; 2128 kmutex_t *hash_lock; 2129 2130 /* 2131 * Check to see if this buffer is evicted. Callers 2132 * must verify b_data != NULL to know if the add_ref 2133 * was successful. 2134 */ 2135 mutex_enter(&buf->b_evict_lock); 2136 if (buf->b_data == NULL) { 2137 mutex_exit(&buf->b_evict_lock); 2138 return; 2139 } 2140 hash_lock = HDR_LOCK(buf->b_hdr); 2141 mutex_enter(hash_lock); 2142 hdr = buf->b_hdr; 2143 ASSERT(HDR_HAS_L1HDR(hdr)); 2144 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 2145 mutex_exit(&buf->b_evict_lock); 2146 2147 ASSERT(hdr->b_l1hdr.b_state == arc_mru || 2148 hdr->b_l1hdr.b_state == arc_mfu); 2149 2150 add_reference(hdr, hash_lock, tag); 2151 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr); 2152 arc_access(hdr, hash_lock); 2153 mutex_exit(hash_lock); 2154 ARCSTAT_BUMP(arcstat_hits); 2155 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr), 2156 demand, prefetch, !HDR_ISTYPE_METADATA(hdr), 2157 data, metadata, hits); 2158} 2159 2160static void 2161arc_buf_free_on_write(void *data, size_t size, 2162 void (*free_func)(void *, size_t)) 2163{ 2164 l2arc_data_free_t *df; 2165 2166 df = kmem_alloc(sizeof (*df), KM_SLEEP); 2167 df->l2df_data = data; 2168 df->l2df_size = size; 2169 df->l2df_func = free_func; 2170 mutex_enter(&l2arc_free_on_write_mtx); 2171 list_insert_head(l2arc_free_on_write, df); 2172 mutex_exit(&l2arc_free_on_write_mtx); 2173} 2174 2175/* 2176 * Free the arc data buffer. If it is an l2arc write in progress, 2177 * the buffer is placed on l2arc_free_on_write to be freed later. 2178 */ 2179static void 2180arc_buf_data_free(arc_buf_t *buf, void (*free_func)(void *, size_t)) 2181{ 2182 arc_buf_hdr_t *hdr = buf->b_hdr; 2183 2184 if (HDR_L2_WRITING(hdr)) { 2185 arc_buf_free_on_write(buf->b_data, hdr->b_size, free_func); 2186 ARCSTAT_BUMP(arcstat_l2_free_on_write); 2187 } else { 2188 free_func(buf->b_data, hdr->b_size); 2189 } 2190} 2191 2192static void 2193arc_buf_l2_cdata_free(arc_buf_hdr_t *hdr) 2194{ 2195 ASSERT(HDR_HAS_L2HDR(hdr)); 2196 ASSERT(MUTEX_HELD(&hdr->b_l2hdr.b_dev->l2ad_mtx)); 2197 2198 /* 2199 * The b_tmp_cdata field is linked off of the b_l1hdr, so if 2200 * that doesn't exist, the header is in the arc_l2c_only state, 2201 * and there isn't anything to free (it's already been freed). 2202 */ 2203 if (!HDR_HAS_L1HDR(hdr)) 2204 return; 2205 2206 /* 2207 * The header isn't being written to the l2arc device, thus it 2208 * shouldn't have a b_tmp_cdata to free. 2209 */ 2210 if (!HDR_L2_WRITING(hdr)) { 2211 ASSERT3P(hdr->b_l1hdr.b_tmp_cdata, ==, NULL); 2212 return; 2213 } 2214 2215 /* 2216 * The header does not have compression enabled. This can be due 2217 * to the buffer not being compressible, or because we're 2218 * freeing the buffer before the second phase of 2219 * l2arc_write_buffer() has started (which does the compression 2220 * step). In either case, b_tmp_cdata does not point to a 2221 * separately compressed buffer, so there's nothing to free (it 2222 * points to the same buffer as the arc_buf_t's b_data field). 2223 */ 2224 if (hdr->b_l2hdr.b_compress == ZIO_COMPRESS_OFF) { 2225 hdr->b_l1hdr.b_tmp_cdata = NULL; 2226 return; 2227 } 2228 2229 /* 2230 * There's nothing to free since the buffer was all zero's and 2231 * compressed to a zero length buffer. 2232 */ 2233 if (hdr->b_l2hdr.b_compress == ZIO_COMPRESS_EMPTY) { 2234 ASSERT3P(hdr->b_l1hdr.b_tmp_cdata, ==, NULL); 2235 return; 2236 } 2237 2238 ASSERT(L2ARC_IS_VALID_COMPRESS(hdr->b_l2hdr.b_compress)); 2239 2240 arc_buf_free_on_write(hdr->b_l1hdr.b_tmp_cdata, 2241 hdr->b_size, zio_data_buf_free); 2242 2243 ARCSTAT_BUMP(arcstat_l2_cdata_free_on_write); 2244 hdr->b_l1hdr.b_tmp_cdata = NULL; 2245} 2246 2247/* 2248 * Free up buf->b_data and if 'remove' is set, then pull the 2249 * arc_buf_t off of the the arc_buf_hdr_t's list and free it. 2250 */ 2251static void 2252arc_buf_destroy(arc_buf_t *buf, boolean_t remove) 2253{ 2254 arc_buf_t **bufp; 2255 2256 /* free up data associated with the buf */ 2257 if (buf->b_data != NULL) { 2258 arc_state_t *state = buf->b_hdr->b_l1hdr.b_state; 2259 uint64_t size = buf->b_hdr->b_size; 2260 arc_buf_contents_t type = arc_buf_type(buf->b_hdr); 2261 2262 arc_cksum_verify(buf); 2263#ifdef illumos 2264 arc_buf_unwatch(buf); 2265#endif /* illumos */ 2266 2267 if (type == ARC_BUFC_METADATA) { 2268 arc_buf_data_free(buf, zio_buf_free); 2269 arc_space_return(size, ARC_SPACE_META); 2270 } else { 2271 ASSERT(type == ARC_BUFC_DATA); 2272 arc_buf_data_free(buf, zio_data_buf_free); 2273 arc_space_return(size, ARC_SPACE_DATA); 2274 } 2275 2276 /* protected by hash lock, if in the hash table */ 2277 if (multilist_link_active(&buf->b_hdr->b_l1hdr.b_arc_node)) { 2278 uint64_t *cnt = &state->arcs_lsize[type]; 2279 2280 ASSERT(refcount_is_zero( 2281 &buf->b_hdr->b_l1hdr.b_refcnt)); 2282 ASSERT(state != arc_anon && state != arc_l2c_only); 2283 2284 ASSERT3U(*cnt, >=, size); 2285 atomic_add_64(cnt, -size); 2286 } 2287 2288 (void) refcount_remove_many(&state->arcs_size, size, buf); 2289 buf->b_data = NULL; 2290 2291 /* 2292 * If we're destroying a duplicate buffer make sure 2293 * that the appropriate statistics are updated. 2294 */ 2295 if (buf->b_hdr->b_l1hdr.b_datacnt > 1 && 2296 HDR_ISTYPE_DATA(buf->b_hdr)) { 2297 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers); 2298 ARCSTAT_INCR(arcstat_duplicate_buffers_size, -size); 2299 } 2300 ASSERT(buf->b_hdr->b_l1hdr.b_datacnt > 0); 2301 buf->b_hdr->b_l1hdr.b_datacnt -= 1; 2302 } 2303 2304 /* only remove the buf if requested */ 2305 if (!remove) 2306 return; 2307 2308 /* remove the buf from the hdr list */ 2309 for (bufp = &buf->b_hdr->b_l1hdr.b_buf; *bufp != buf; 2310 bufp = &(*bufp)->b_next) 2311 continue; 2312 *bufp = buf->b_next; 2313 buf->b_next = NULL; 2314 2315 ASSERT(buf->b_efunc == NULL); 2316 2317 /* clean up the buf */ 2318 buf->b_hdr = NULL; 2319 kmem_cache_free(buf_cache, buf); 2320} 2321 2322static void 2323arc_hdr_l2hdr_destroy(arc_buf_hdr_t *hdr) 2324{ 2325 l2arc_buf_hdr_t *l2hdr = &hdr->b_l2hdr; 2326 l2arc_dev_t *dev = l2hdr->b_dev; 2327 2328 ASSERT(MUTEX_HELD(&dev->l2ad_mtx)); 2329 ASSERT(HDR_HAS_L2HDR(hdr)); 2330 2331 list_remove(&dev->l2ad_buflist, hdr); 2332 2333 /* 2334 * We don't want to leak the b_tmp_cdata buffer that was 2335 * allocated in l2arc_write_buffers() 2336 */ 2337 arc_buf_l2_cdata_free(hdr); 2338 2339 /* 2340 * If the l2hdr's b_daddr is equal to L2ARC_ADDR_UNSET, then 2341 * this header is being processed by l2arc_write_buffers() (i.e. 2342 * it's in the first stage of l2arc_write_buffers()). 2343 * Re-affirming that truth here, just to serve as a reminder. If 2344 * b_daddr does not equal L2ARC_ADDR_UNSET, then the header may or 2345 * may not have its HDR_L2_WRITING flag set. (the write may have 2346 * completed, in which case HDR_L2_WRITING will be false and the 2347 * b_daddr field will point to the address of the buffer on disk). 2348 */ 2349 IMPLY(l2hdr->b_daddr == L2ARC_ADDR_UNSET, HDR_L2_WRITING(hdr)); 2350 2351 /* 2352 * If b_daddr is equal to L2ARC_ADDR_UNSET, we're racing with 2353 * l2arc_write_buffers(). Since we've just removed this header 2354 * from the l2arc buffer list, this header will never reach the 2355 * second stage of l2arc_write_buffers(), which increments the 2356 * accounting stats for this header. Thus, we must be careful 2357 * not to decrement them for this header either. 2358 */ 2359 if (l2hdr->b_daddr != L2ARC_ADDR_UNSET) { 2360 ARCSTAT_INCR(arcstat_l2_asize, -l2hdr->b_asize); 2361 ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size); 2362 2363 vdev_space_update(dev->l2ad_vdev, 2364 -l2hdr->b_asize, 0, 0); 2365 2366 (void) refcount_remove_many(&dev->l2ad_alloc, 2367 l2hdr->b_asize, hdr); 2368 } 2369 2370 hdr->b_flags &= ~ARC_FLAG_HAS_L2HDR; 2371} 2372 2373static void 2374arc_hdr_destroy(arc_buf_hdr_t *hdr) 2375{ 2376 if (HDR_HAS_L1HDR(hdr)) { 2377 ASSERT(hdr->b_l1hdr.b_buf == NULL || 2378 hdr->b_l1hdr.b_datacnt > 0); 2379 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 2380 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon); 2381 } 2382 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 2383 ASSERT(!HDR_IN_HASH_TABLE(hdr)); 2384 2385 if (HDR_HAS_L2HDR(hdr)) { 2386 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev; 2387 boolean_t buflist_held = MUTEX_HELD(&dev->l2ad_mtx); 2388 2389 if (!buflist_held) 2390 mutex_enter(&dev->l2ad_mtx); 2391 2392 /* 2393 * Even though we checked this conditional above, we 2394 * need to check this again now that we have the 2395 * l2ad_mtx. This is because we could be racing with 2396 * another thread calling l2arc_evict() which might have 2397 * destroyed this header's L2 portion as we were waiting 2398 * to acquire the l2ad_mtx. If that happens, we don't 2399 * want to re-destroy the header's L2 portion. 2400 */ 2401 if (HDR_HAS_L2HDR(hdr)) { 2402 if (hdr->b_l2hdr.b_daddr != L2ARC_ADDR_UNSET) 2403 trim_map_free(dev->l2ad_vdev, 2404 hdr->b_l2hdr.b_daddr, 2405 hdr->b_l2hdr.b_asize, 0); 2406 arc_hdr_l2hdr_destroy(hdr); 2407 } 2408 2409 if (!buflist_held) 2410 mutex_exit(&dev->l2ad_mtx); 2411 } 2412 2413 if (!BUF_EMPTY(hdr)) 2414 buf_discard_identity(hdr); 2415 2416 if (hdr->b_freeze_cksum != NULL) { 2417 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 2418 hdr->b_freeze_cksum = NULL; 2419 } 2420 2421 if (HDR_HAS_L1HDR(hdr)) { 2422 while (hdr->b_l1hdr.b_buf) { 2423 arc_buf_t *buf = hdr->b_l1hdr.b_buf; 2424 2425 if (buf->b_efunc != NULL) { 2426 mutex_enter(&arc_user_evicts_lock); 2427 mutex_enter(&buf->b_evict_lock); 2428 ASSERT(buf->b_hdr != NULL); 2429 arc_buf_destroy(hdr->b_l1hdr.b_buf, FALSE); 2430 hdr->b_l1hdr.b_buf = buf->b_next; 2431 buf->b_hdr = &arc_eviction_hdr; 2432 buf->b_next = arc_eviction_list; 2433 arc_eviction_list = buf; 2434 mutex_exit(&buf->b_evict_lock); 2435 cv_signal(&arc_user_evicts_cv); 2436 mutex_exit(&arc_user_evicts_lock); 2437 } else { 2438 arc_buf_destroy(hdr->b_l1hdr.b_buf, TRUE); 2439 } 2440 } 2441#ifdef ZFS_DEBUG 2442 if (hdr->b_l1hdr.b_thawed != NULL) { 2443 kmem_free(hdr->b_l1hdr.b_thawed, 1); 2444 hdr->b_l1hdr.b_thawed = NULL; 2445 } 2446#endif 2447 } 2448 2449 ASSERT3P(hdr->b_hash_next, ==, NULL); 2450 if (HDR_HAS_L1HDR(hdr)) { 2451 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 2452 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL); 2453 kmem_cache_free(hdr_full_cache, hdr); 2454 } else { 2455 kmem_cache_free(hdr_l2only_cache, hdr); 2456 } 2457} 2458 2459void 2460arc_buf_free(arc_buf_t *buf, void *tag) 2461{ 2462 arc_buf_hdr_t *hdr = buf->b_hdr; 2463 int hashed = hdr->b_l1hdr.b_state != arc_anon; 2464 2465 ASSERT(buf->b_efunc == NULL); 2466 ASSERT(buf->b_data != NULL); 2467 2468 if (hashed) { 2469 kmutex_t *hash_lock = HDR_LOCK(hdr); 2470 2471 mutex_enter(hash_lock); 2472 hdr = buf->b_hdr; 2473 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 2474 2475 (void) remove_reference(hdr, hash_lock, tag); 2476 if (hdr->b_l1hdr.b_datacnt > 1) { 2477 arc_buf_destroy(buf, TRUE); 2478 } else { 2479 ASSERT(buf == hdr->b_l1hdr.b_buf); 2480 ASSERT(buf->b_efunc == NULL); 2481 hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE; 2482 } 2483 mutex_exit(hash_lock); 2484 } else if (HDR_IO_IN_PROGRESS(hdr)) { 2485 int destroy_hdr; 2486 /* 2487 * We are in the middle of an async write. Don't destroy 2488 * this buffer unless the write completes before we finish 2489 * decrementing the reference count. 2490 */ 2491 mutex_enter(&arc_user_evicts_lock); 2492 (void) remove_reference(hdr, NULL, tag); 2493 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 2494 destroy_hdr = !HDR_IO_IN_PROGRESS(hdr); 2495 mutex_exit(&arc_user_evicts_lock); 2496 if (destroy_hdr) 2497 arc_hdr_destroy(hdr); 2498 } else { 2499 if (remove_reference(hdr, NULL, tag) > 0) 2500 arc_buf_destroy(buf, TRUE); 2501 else 2502 arc_hdr_destroy(hdr); 2503 } 2504} 2505 2506boolean_t 2507arc_buf_remove_ref(arc_buf_t *buf, void* tag) 2508{ 2509 arc_buf_hdr_t *hdr = buf->b_hdr; 2510 kmutex_t *hash_lock = HDR_LOCK(hdr); 2511 boolean_t no_callback = (buf->b_efunc == NULL); 2512 2513 if (hdr->b_l1hdr.b_state == arc_anon) { 2514 ASSERT(hdr->b_l1hdr.b_datacnt == 1); 2515 arc_buf_free(buf, tag); 2516 return (no_callback); 2517 } 2518 2519 mutex_enter(hash_lock); 2520 hdr = buf->b_hdr; 2521 ASSERT(hdr->b_l1hdr.b_datacnt > 0); 2522 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 2523 ASSERT(hdr->b_l1hdr.b_state != arc_anon); 2524 ASSERT(buf->b_data != NULL); 2525 2526 (void) remove_reference(hdr, hash_lock, tag); 2527 if (hdr->b_l1hdr.b_datacnt > 1) { 2528 if (no_callback) 2529 arc_buf_destroy(buf, TRUE); 2530 } else if (no_callback) { 2531 ASSERT(hdr->b_l1hdr.b_buf == buf && buf->b_next == NULL); 2532 ASSERT(buf->b_efunc == NULL); 2533 hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE; 2534 } 2535 ASSERT(no_callback || hdr->b_l1hdr.b_datacnt > 1 || 2536 refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 2537 mutex_exit(hash_lock); 2538 return (no_callback); 2539} 2540 2541int32_t 2542arc_buf_size(arc_buf_t *buf) 2543{ 2544 return (buf->b_hdr->b_size); 2545} 2546 2547/* 2548 * Called from the DMU to determine if the current buffer should be 2549 * evicted. In order to ensure proper locking, the eviction must be initiated 2550 * from the DMU. Return true if the buffer is associated with user data and 2551 * duplicate buffers still exist. 2552 */ 2553boolean_t 2554arc_buf_eviction_needed(arc_buf_t *buf) 2555{ 2556 arc_buf_hdr_t *hdr; 2557 boolean_t evict_needed = B_FALSE; 2558 2559 if (zfs_disable_dup_eviction) 2560 return (B_FALSE); 2561 2562 mutex_enter(&buf->b_evict_lock); 2563 hdr = buf->b_hdr; 2564 if (hdr == NULL) { 2565 /* 2566 * We are in arc_do_user_evicts(); let that function 2567 * perform the eviction. 2568 */ 2569 ASSERT(buf->b_data == NULL); 2570 mutex_exit(&buf->b_evict_lock); 2571 return (B_FALSE); 2572 } else if (buf->b_data == NULL) { 2573 /* 2574 * We have already been added to the arc eviction list; 2575 * recommend eviction. 2576 */ 2577 ASSERT3P(hdr, ==, &arc_eviction_hdr); 2578 mutex_exit(&buf->b_evict_lock); 2579 return (B_TRUE); 2580 } 2581 2582 if (hdr->b_l1hdr.b_datacnt > 1 && HDR_ISTYPE_DATA(hdr)) 2583 evict_needed = B_TRUE; 2584 2585 mutex_exit(&buf->b_evict_lock); 2586 return (evict_needed); 2587} 2588 2589/* 2590 * Evict the arc_buf_hdr that is provided as a parameter. The resultant 2591 * state of the header is dependent on it's state prior to entering this 2592 * function. The following transitions are possible: 2593 * 2594 * - arc_mru -> arc_mru_ghost 2595 * - arc_mfu -> arc_mfu_ghost 2596 * - arc_mru_ghost -> arc_l2c_only 2597 * - arc_mru_ghost -> deleted 2598 * - arc_mfu_ghost -> arc_l2c_only 2599 * - arc_mfu_ghost -> deleted 2600 */ 2601static int64_t 2602arc_evict_hdr(arc_buf_hdr_t *hdr, kmutex_t *hash_lock) 2603{ 2604 arc_state_t *evicted_state, *state; 2605 int64_t bytes_evicted = 0; 2606 2607 ASSERT(MUTEX_HELD(hash_lock)); 2608 ASSERT(HDR_HAS_L1HDR(hdr)); 2609 2610 state = hdr->b_l1hdr.b_state; 2611 if (GHOST_STATE(state)) { 2612 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 2613 ASSERT(hdr->b_l1hdr.b_buf == NULL); 2614 2615 /* 2616 * l2arc_write_buffers() relies on a header's L1 portion 2617 * (i.e. it's b_tmp_cdata field) during it's write phase. 2618 * Thus, we cannot push a header onto the arc_l2c_only 2619 * state (removing it's L1 piece) until the header is 2620 * done being written to the l2arc. 2621 */ 2622 if (HDR_HAS_L2HDR(hdr) && HDR_L2_WRITING(hdr)) { 2623 ARCSTAT_BUMP(arcstat_evict_l2_skip); 2624 return (bytes_evicted); 2625 } 2626 2627 ARCSTAT_BUMP(arcstat_deleted); 2628 bytes_evicted += hdr->b_size; 2629 2630 DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, hdr); 2631 2632 if (HDR_HAS_L2HDR(hdr)) { 2633 /* 2634 * This buffer is cached on the 2nd Level ARC; 2635 * don't destroy the header. 2636 */ 2637 arc_change_state(arc_l2c_only, hdr, hash_lock); 2638 /* 2639 * dropping from L1+L2 cached to L2-only, 2640 * realloc to remove the L1 header. 2641 */ 2642 hdr = arc_hdr_realloc(hdr, hdr_full_cache, 2643 hdr_l2only_cache); 2644 } else { 2645 arc_change_state(arc_anon, hdr, hash_lock); 2646 arc_hdr_destroy(hdr); 2647 } 2648 return (bytes_evicted); 2649 } 2650 2651 ASSERT(state == arc_mru || state == arc_mfu); 2652 evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost; 2653 2654 /* prefetch buffers have a minimum lifespan */ 2655 if (HDR_IO_IN_PROGRESS(hdr) || 2656 ((hdr->b_flags & (ARC_FLAG_PREFETCH | ARC_FLAG_INDIRECT)) && 2657 ddi_get_lbolt() - hdr->b_l1hdr.b_arc_access < 2658 arc_min_prefetch_lifespan)) { 2659 ARCSTAT_BUMP(arcstat_evict_skip); 2660 return (bytes_evicted); 2661 } 2662 2663 ASSERT0(refcount_count(&hdr->b_l1hdr.b_refcnt)); 2664 ASSERT3U(hdr->b_l1hdr.b_datacnt, >, 0); 2665 while (hdr->b_l1hdr.b_buf) { 2666 arc_buf_t *buf = hdr->b_l1hdr.b_buf; 2667 if (!mutex_tryenter(&buf->b_evict_lock)) { 2668 ARCSTAT_BUMP(arcstat_mutex_miss); 2669 break; 2670 } 2671 if (buf->b_data != NULL) 2672 bytes_evicted += hdr->b_size; 2673 if (buf->b_efunc != NULL) { 2674 mutex_enter(&arc_user_evicts_lock); 2675 arc_buf_destroy(buf, FALSE); 2676 hdr->b_l1hdr.b_buf = buf->b_next; 2677 buf->b_hdr = &arc_eviction_hdr; 2678 buf->b_next = arc_eviction_list; 2679 arc_eviction_list = buf; 2680 cv_signal(&arc_user_evicts_cv); 2681 mutex_exit(&arc_user_evicts_lock); 2682 mutex_exit(&buf->b_evict_lock); 2683 } else { 2684 mutex_exit(&buf->b_evict_lock); 2685 arc_buf_destroy(buf, TRUE); 2686 } 2687 } 2688 2689 if (HDR_HAS_L2HDR(hdr)) { 2690 ARCSTAT_INCR(arcstat_evict_l2_cached, hdr->b_size); 2691 } else { 2692 if (l2arc_write_eligible(hdr->b_spa, hdr)) 2693 ARCSTAT_INCR(arcstat_evict_l2_eligible, hdr->b_size); 2694 else 2695 ARCSTAT_INCR(arcstat_evict_l2_ineligible, hdr->b_size); 2696 } 2697 2698 if (hdr->b_l1hdr.b_datacnt == 0) { 2699 arc_change_state(evicted_state, hdr, hash_lock); 2700 ASSERT(HDR_IN_HASH_TABLE(hdr)); 2701 hdr->b_flags |= ARC_FLAG_IN_HASH_TABLE; 2702 hdr->b_flags &= ~ARC_FLAG_BUF_AVAILABLE; 2703 DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, hdr); 2704 } 2705 2706 return (bytes_evicted); 2707} 2708 2709static uint64_t 2710arc_evict_state_impl(multilist_t *ml, int idx, arc_buf_hdr_t *marker, 2711 uint64_t spa, int64_t bytes) 2712{ 2713 multilist_sublist_t *mls; 2714 uint64_t bytes_evicted = 0; 2715 arc_buf_hdr_t *hdr; 2716 kmutex_t *hash_lock; 2717 int evict_count = 0; 2718 2719 ASSERT3P(marker, !=, NULL); 2720 IMPLY(bytes < 0, bytes == ARC_EVICT_ALL); 2721 2722 mls = multilist_sublist_lock(ml, idx); 2723 2724 for (hdr = multilist_sublist_prev(mls, marker); hdr != NULL; 2725 hdr = multilist_sublist_prev(mls, marker)) { 2726 if ((bytes != ARC_EVICT_ALL && bytes_evicted >= bytes) || 2727 (evict_count >= zfs_arc_evict_batch_limit)) 2728 break; 2729 2730 /* 2731 * To keep our iteration location, move the marker 2732 * forward. Since we're not holding hdr's hash lock, we 2733 * must be very careful and not remove 'hdr' from the 2734 * sublist. Otherwise, other consumers might mistake the 2735 * 'hdr' as not being on a sublist when they call the 2736 * multilist_link_active() function (they all rely on 2737 * the hash lock protecting concurrent insertions and 2738 * removals). multilist_sublist_move_forward() was 2739 * specifically implemented to ensure this is the case 2740 * (only 'marker' will be removed and re-inserted). 2741 */ 2742 multilist_sublist_move_forward(mls, marker); 2743 2744 /* 2745 * The only case where the b_spa field should ever be 2746 * zero, is the marker headers inserted by 2747 * arc_evict_state(). It's possible for multiple threads 2748 * to be calling arc_evict_state() concurrently (e.g. 2749 * dsl_pool_close() and zio_inject_fault()), so we must 2750 * skip any markers we see from these other threads. 2751 */ 2752 if (hdr->b_spa == 0) 2753 continue; 2754 2755 /* we're only interested in evicting buffers of a certain spa */ 2756 if (spa != 0 && hdr->b_spa != spa) { 2757 ARCSTAT_BUMP(arcstat_evict_skip); 2758 continue; 2759 } 2760 2761 hash_lock = HDR_LOCK(hdr); 2762 2763 /* 2764 * We aren't calling this function from any code path 2765 * that would already be holding a hash lock, so we're 2766 * asserting on this assumption to be defensive in case 2767 * this ever changes. Without this check, it would be 2768 * possible to incorrectly increment arcstat_mutex_miss 2769 * below (e.g. if the code changed such that we called 2770 * this function with a hash lock held). 2771 */ 2772 ASSERT(!MUTEX_HELD(hash_lock)); 2773 2774 if (mutex_tryenter(hash_lock)) { 2775 uint64_t evicted = arc_evict_hdr(hdr, hash_lock); 2776 mutex_exit(hash_lock); 2777 2778 bytes_evicted += evicted; 2779 2780 /* 2781 * If evicted is zero, arc_evict_hdr() must have 2782 * decided to skip this header, don't increment 2783 * evict_count in this case. 2784 */ 2785 if (evicted != 0) 2786 evict_count++; 2787 2788 /* 2789 * If arc_size isn't overflowing, signal any 2790 * threads that might happen to be waiting. 2791 * 2792 * For each header evicted, we wake up a single 2793 * thread. If we used cv_broadcast, we could 2794 * wake up "too many" threads causing arc_size 2795 * to significantly overflow arc_c; since 2796 * arc_get_data_buf() doesn't check for overflow 2797 * when it's woken up (it doesn't because it's 2798 * possible for the ARC to be overflowing while 2799 * full of un-evictable buffers, and the 2800 * function should proceed in this case). 2801 * 2802 * If threads are left sleeping, due to not 2803 * using cv_broadcast, they will be woken up 2804 * just before arc_reclaim_thread() sleeps. 2805 */ 2806 mutex_enter(&arc_reclaim_lock); 2807 if (!arc_is_overflowing()) 2808 cv_signal(&arc_reclaim_waiters_cv); 2809 mutex_exit(&arc_reclaim_lock); 2810 } else { 2811 ARCSTAT_BUMP(arcstat_mutex_miss); 2812 } 2813 } 2814 2815 multilist_sublist_unlock(mls); 2816 2817 return (bytes_evicted); 2818} 2819 2820/* 2821 * Evict buffers from the given arc state, until we've removed the 2822 * specified number of bytes. Move the removed buffers to the 2823 * appropriate evict state. 2824 * 2825 * This function makes a "best effort". It skips over any buffers 2826 * it can't get a hash_lock on, and so, may not catch all candidates. 2827 * It may also return without evicting as much space as requested. 2828 * 2829 * If bytes is specified using the special value ARC_EVICT_ALL, this 2830 * will evict all available (i.e. unlocked and evictable) buffers from 2831 * the given arc state; which is used by arc_flush(). 2832 */ 2833static uint64_t 2834arc_evict_state(arc_state_t *state, uint64_t spa, int64_t bytes, 2835 arc_buf_contents_t type) 2836{ 2837 uint64_t total_evicted = 0; 2838 multilist_t *ml = &state->arcs_list[type]; 2839 int num_sublists; 2840 arc_buf_hdr_t **markers; 2841 2842 IMPLY(bytes < 0, bytes == ARC_EVICT_ALL); 2843 2844 num_sublists = multilist_get_num_sublists(ml); 2845 2846 /* 2847 * If we've tried to evict from each sublist, made some 2848 * progress, but still have not hit the target number of bytes 2849 * to evict, we want to keep trying. The markers allow us to 2850 * pick up where we left off for each individual sublist, rather 2851 * than starting from the tail each time. 2852 */ 2853 markers = kmem_zalloc(sizeof (*markers) * num_sublists, KM_SLEEP); 2854 for (int i = 0; i < num_sublists; i++) { 2855 markers[i] = kmem_cache_alloc(hdr_full_cache, KM_SLEEP); 2856 2857 /* 2858 * A b_spa of 0 is used to indicate that this header is 2859 * a marker. This fact is used in arc_adjust_type() and 2860 * arc_evict_state_impl(). 2861 */ 2862 markers[i]->b_spa = 0; 2863 2864 multilist_sublist_t *mls = multilist_sublist_lock(ml, i); 2865 multilist_sublist_insert_tail(mls, markers[i]); 2866 multilist_sublist_unlock(mls); 2867 } 2868 2869 /* 2870 * While we haven't hit our target number of bytes to evict, or 2871 * we're evicting all available buffers. 2872 */ 2873 while (total_evicted < bytes || bytes == ARC_EVICT_ALL) { 2874 /* 2875 * Start eviction using a randomly selected sublist, 2876 * this is to try and evenly balance eviction across all 2877 * sublists. Always starting at the same sublist 2878 * (e.g. index 0) would cause evictions to favor certain 2879 * sublists over others. 2880 */ 2881 int sublist_idx = multilist_get_random_index(ml); 2882 uint64_t scan_evicted = 0; 2883 2884 for (int i = 0; i < num_sublists; i++) { 2885 uint64_t bytes_remaining; 2886 uint64_t bytes_evicted; 2887 2888 if (bytes == ARC_EVICT_ALL) 2889 bytes_remaining = ARC_EVICT_ALL; 2890 else if (total_evicted < bytes) 2891 bytes_remaining = bytes - total_evicted; 2892 else 2893 break; 2894 2895 bytes_evicted = arc_evict_state_impl(ml, sublist_idx, 2896 markers[sublist_idx], spa, bytes_remaining); 2897 2898 scan_evicted += bytes_evicted; 2899 total_evicted += bytes_evicted; 2900 2901 /* we've reached the end, wrap to the beginning */ 2902 if (++sublist_idx >= num_sublists) 2903 sublist_idx = 0; 2904 } 2905 2906 /* 2907 * If we didn't evict anything during this scan, we have 2908 * no reason to believe we'll evict more during another 2909 * scan, so break the loop. 2910 */ 2911 if (scan_evicted == 0) { 2912 /* This isn't possible, let's make that obvious */ 2913 ASSERT3S(bytes, !=, 0); 2914 2915 /* 2916 * When bytes is ARC_EVICT_ALL, the only way to 2917 * break the loop is when scan_evicted is zero. 2918 * In that case, we actually have evicted enough, 2919 * so we don't want to increment the kstat. 2920 */ 2921 if (bytes != ARC_EVICT_ALL) { 2922 ASSERT3S(total_evicted, <, bytes); 2923 ARCSTAT_BUMP(arcstat_evict_not_enough); 2924 } 2925 2926 break; 2927 } 2928 } 2929 2930 for (int i = 0; i < num_sublists; i++) { 2931 multilist_sublist_t *mls = multilist_sublist_lock(ml, i); 2932 multilist_sublist_remove(mls, markers[i]); 2933 multilist_sublist_unlock(mls); 2934 2935 kmem_cache_free(hdr_full_cache, markers[i]); 2936 } 2937 kmem_free(markers, sizeof (*markers) * num_sublists); 2938 2939 return (total_evicted); 2940} 2941 2942/* 2943 * Flush all "evictable" data of the given type from the arc state 2944 * specified. This will not evict any "active" buffers (i.e. referenced). 2945 * 2946 * When 'retry' is set to FALSE, the function will make a single pass 2947 * over the state and evict any buffers that it can. Since it doesn't 2948 * continually retry the eviction, it might end up leaving some buffers 2949 * in the ARC due to lock misses. 2950 * 2951 * When 'retry' is set to TRUE, the function will continually retry the 2952 * eviction until *all* evictable buffers have been removed from the 2953 * state. As a result, if concurrent insertions into the state are 2954 * allowed (e.g. if the ARC isn't shutting down), this function might 2955 * wind up in an infinite loop, continually trying to evict buffers. 2956 */ 2957static uint64_t 2958arc_flush_state(arc_state_t *state, uint64_t spa, arc_buf_contents_t type, 2959 boolean_t retry) 2960{ 2961 uint64_t evicted = 0; 2962 2963 while (state->arcs_lsize[type] != 0) { 2964 evicted += arc_evict_state(state, spa, ARC_EVICT_ALL, type); 2965 2966 if (!retry) 2967 break; 2968 } 2969 2970 return (evicted); 2971} 2972 2973/* 2974 * Evict the specified number of bytes from the state specified, 2975 * restricting eviction to the spa and type given. This function 2976 * prevents us from trying to evict more from a state's list than 2977 * is "evictable", and to skip evicting altogether when passed a 2978 * negative value for "bytes". In contrast, arc_evict_state() will 2979 * evict everything it can, when passed a negative value for "bytes". 2980 */ 2981static uint64_t 2982arc_adjust_impl(arc_state_t *state, uint64_t spa, int64_t bytes, 2983 arc_buf_contents_t type) 2984{ 2985 int64_t delta; 2986 2987 if (bytes > 0 && state->arcs_lsize[type] > 0) { 2988 delta = MIN(state->arcs_lsize[type], bytes); 2989 return (arc_evict_state(state, spa, delta, type)); 2990 } 2991 2992 return (0); 2993} 2994 2995/* 2996 * Evict metadata buffers from the cache, such that arc_meta_used is 2997 * capped by the arc_meta_limit tunable. 2998 */ 2999static uint64_t 3000arc_adjust_meta(void) 3001{ 3002 uint64_t total_evicted = 0; 3003 int64_t target; 3004 3005 /* 3006 * If we're over the meta limit, we want to evict enough 3007 * metadata to get back under the meta limit. We don't want to 3008 * evict so much that we drop the MRU below arc_p, though. If 3009 * we're over the meta limit more than we're over arc_p, we 3010 * evict some from the MRU here, and some from the MFU below. 3011 */ 3012 target = MIN((int64_t)(arc_meta_used - arc_meta_limit), 3013 (int64_t)(refcount_count(&arc_anon->arcs_size) + 3014 refcount_count(&arc_mru->arcs_size) - arc_p)); 3015 3016 total_evicted += arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA); 3017 3018 /* 3019 * Similar to the above, we want to evict enough bytes to get us 3020 * below the meta limit, but not so much as to drop us below the 3021 * space alloted to the MFU (which is defined as arc_c - arc_p). 3022 */ 3023 target = MIN((int64_t)(arc_meta_used - arc_meta_limit), 3024 (int64_t)(refcount_count(&arc_mfu->arcs_size) - (arc_c - arc_p))); 3025 3026 total_evicted += arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA); 3027 3028 return (total_evicted); 3029} 3030 3031/* 3032 * Return the type of the oldest buffer in the given arc state 3033 * 3034 * This function will select a random sublist of type ARC_BUFC_DATA and 3035 * a random sublist of type ARC_BUFC_METADATA. The tail of each sublist 3036 * is compared, and the type which contains the "older" buffer will be 3037 * returned. 3038 */ 3039static arc_buf_contents_t 3040arc_adjust_type(arc_state_t *state) 3041{ 3042 multilist_t *data_ml = &state->arcs_list[ARC_BUFC_DATA]; 3043 multilist_t *meta_ml = &state->arcs_list[ARC_BUFC_METADATA]; 3044 int data_idx = multilist_get_random_index(data_ml); 3045 int meta_idx = multilist_get_random_index(meta_ml); 3046 multilist_sublist_t *data_mls; 3047 multilist_sublist_t *meta_mls; 3048 arc_buf_contents_t type; 3049 arc_buf_hdr_t *data_hdr; 3050 arc_buf_hdr_t *meta_hdr; 3051 3052 /* 3053 * We keep the sublist lock until we're finished, to prevent 3054 * the headers from being destroyed via arc_evict_state(). 3055 */ 3056 data_mls = multilist_sublist_lock(data_ml, data_idx); 3057 meta_mls = multilist_sublist_lock(meta_ml, meta_idx); 3058 3059 /* 3060 * These two loops are to ensure we skip any markers that 3061 * might be at the tail of the lists due to arc_evict_state(). 3062 */ 3063 3064 for (data_hdr = multilist_sublist_tail(data_mls); data_hdr != NULL; 3065 data_hdr = multilist_sublist_prev(data_mls, data_hdr)) { 3066 if (data_hdr->b_spa != 0) 3067 break; 3068 } 3069 3070 for (meta_hdr = multilist_sublist_tail(meta_mls); meta_hdr != NULL; 3071 meta_hdr = multilist_sublist_prev(meta_mls, meta_hdr)) { 3072 if (meta_hdr->b_spa != 0) 3073 break; 3074 } 3075 3076 if (data_hdr == NULL && meta_hdr == NULL) { 3077 type = ARC_BUFC_DATA; 3078 } else if (data_hdr == NULL) { 3079 ASSERT3P(meta_hdr, !=, NULL); 3080 type = ARC_BUFC_METADATA; 3081 } else if (meta_hdr == NULL) { 3082 ASSERT3P(data_hdr, !=, NULL); 3083 type = ARC_BUFC_DATA; 3084 } else { 3085 ASSERT3P(data_hdr, !=, NULL); 3086 ASSERT3P(meta_hdr, !=, NULL); 3087 3088 /* The headers can't be on the sublist without an L1 header */ 3089 ASSERT(HDR_HAS_L1HDR(data_hdr)); 3090 ASSERT(HDR_HAS_L1HDR(meta_hdr)); 3091 3092 if (data_hdr->b_l1hdr.b_arc_access < 3093 meta_hdr->b_l1hdr.b_arc_access) { 3094 type = ARC_BUFC_DATA; 3095 } else { 3096 type = ARC_BUFC_METADATA; 3097 } 3098 } 3099 3100 multilist_sublist_unlock(meta_mls); 3101 multilist_sublist_unlock(data_mls); 3102 3103 return (type); 3104} 3105 3106/* 3107 * Evict buffers from the cache, such that arc_size is capped by arc_c. 3108 */ 3109static uint64_t 3110arc_adjust(void) 3111{ 3112 uint64_t total_evicted = 0; 3113 uint64_t bytes; 3114 int64_t target; 3115 3116 /* 3117 * If we're over arc_meta_limit, we want to correct that before 3118 * potentially evicting data buffers below. 3119 */ 3120 total_evicted += arc_adjust_meta(); 3121 3122 /* 3123 * Adjust MRU size 3124 * 3125 * If we're over the target cache size, we want to evict enough 3126 * from the list to get back to our target size. We don't want 3127 * to evict too much from the MRU, such that it drops below 3128 * arc_p. So, if we're over our target cache size more than 3129 * the MRU is over arc_p, we'll evict enough to get back to 3130 * arc_p here, and then evict more from the MFU below. 3131 */ 3132 target = MIN((int64_t)(arc_size - arc_c), 3133 (int64_t)(refcount_count(&arc_anon->arcs_size) + 3134 refcount_count(&arc_mru->arcs_size) + arc_meta_used - arc_p)); 3135 3136 /* 3137 * If we're below arc_meta_min, always prefer to evict data. 3138 * Otherwise, try to satisfy the requested number of bytes to 3139 * evict from the type which contains older buffers; in an 3140 * effort to keep newer buffers in the cache regardless of their 3141 * type. If we cannot satisfy the number of bytes from this 3142 * type, spill over into the next type. 3143 */ 3144 if (arc_adjust_type(arc_mru) == ARC_BUFC_METADATA && 3145 arc_meta_used > arc_meta_min) { 3146 bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA); 3147 total_evicted += bytes; 3148 3149 /* 3150 * If we couldn't evict our target number of bytes from 3151 * metadata, we try to get the rest from data. 3152 */ 3153 target -= bytes; 3154 3155 total_evicted += 3156 arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA); 3157 } else { 3158 bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA); 3159 total_evicted += bytes; 3160 3161 /* 3162 * If we couldn't evict our target number of bytes from 3163 * data, we try to get the rest from metadata. 3164 */ 3165 target -= bytes; 3166 3167 total_evicted += 3168 arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA); 3169 } 3170 3171 /* 3172 * Adjust MFU size 3173 * 3174 * Now that we've tried to evict enough from the MRU to get its 3175 * size back to arc_p, if we're still above the target cache 3176 * size, we evict the rest from the MFU. 3177 */ 3178 target = arc_size - arc_c; 3179 3180 if (arc_adjust_type(arc_mfu) == ARC_BUFC_METADATA && 3181 arc_meta_used > arc_meta_min) { 3182 bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA); 3183 total_evicted += bytes; 3184 3185 /* 3186 * If we couldn't evict our target number of bytes from 3187 * metadata, we try to get the rest from data. 3188 */ 3189 target -= bytes; 3190 3191 total_evicted += 3192 arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA); 3193 } else { 3194 bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA); 3195 total_evicted += bytes; 3196 3197 /* 3198 * If we couldn't evict our target number of bytes from 3199 * data, we try to get the rest from data. 3200 */ 3201 target -= bytes; 3202 3203 total_evicted += 3204 arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA); 3205 } 3206 3207 /* 3208 * Adjust ghost lists 3209 * 3210 * In addition to the above, the ARC also defines target values 3211 * for the ghost lists. The sum of the mru list and mru ghost 3212 * list should never exceed the target size of the cache, and 3213 * the sum of the mru list, mfu list, mru ghost list, and mfu 3214 * ghost list should never exceed twice the target size of the 3215 * cache. The following logic enforces these limits on the ghost 3216 * caches, and evicts from them as needed. 3217 */ 3218 target = refcount_count(&arc_mru->arcs_size) + 3219 refcount_count(&arc_mru_ghost->arcs_size) - arc_c; 3220 3221 bytes = arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_DATA); 3222 total_evicted += bytes; 3223 3224 target -= bytes; 3225 3226 total_evicted += 3227 arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_METADATA); 3228 3229 /* 3230 * We assume the sum of the mru list and mfu list is less than 3231 * or equal to arc_c (we enforced this above), which means we 3232 * can use the simpler of the two equations below: 3233 * 3234 * mru + mfu + mru ghost + mfu ghost <= 2 * arc_c 3235 * mru ghost + mfu ghost <= arc_c 3236 */ 3237 target = refcount_count(&arc_mru_ghost->arcs_size) + 3238 refcount_count(&arc_mfu_ghost->arcs_size) - arc_c; 3239 3240 bytes = arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_DATA); 3241 total_evicted += bytes; 3242 3243 target -= bytes; 3244 3245 total_evicted += 3246 arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_METADATA); 3247 3248 return (total_evicted); 3249} 3250 3251static void 3252arc_do_user_evicts(void) 3253{ 3254 mutex_enter(&arc_user_evicts_lock); 3255 while (arc_eviction_list != NULL) { 3256 arc_buf_t *buf = arc_eviction_list; 3257 arc_eviction_list = buf->b_next; 3258 mutex_enter(&buf->b_evict_lock); 3259 buf->b_hdr = NULL; 3260 mutex_exit(&buf->b_evict_lock); 3261 mutex_exit(&arc_user_evicts_lock); 3262 3263 if (buf->b_efunc != NULL) 3264 VERIFY0(buf->b_efunc(buf->b_private)); 3265 3266 buf->b_efunc = NULL; 3267 buf->b_private = NULL; 3268 kmem_cache_free(buf_cache, buf); 3269 mutex_enter(&arc_user_evicts_lock); 3270 } 3271 mutex_exit(&arc_user_evicts_lock); 3272} 3273 3274void 3275arc_flush(spa_t *spa, boolean_t retry) 3276{ 3277 uint64_t guid = 0; 3278 3279 /* 3280 * If retry is TRUE, a spa must not be specified since we have 3281 * no good way to determine if all of a spa's buffers have been 3282 * evicted from an arc state. 3283 */ 3284 ASSERT(!retry || spa == 0); 3285 3286 if (spa != NULL) 3287 guid = spa_load_guid(spa); 3288 3289 (void) arc_flush_state(arc_mru, guid, ARC_BUFC_DATA, retry); 3290 (void) arc_flush_state(arc_mru, guid, ARC_BUFC_METADATA, retry); 3291 3292 (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_DATA, retry); 3293 (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_METADATA, retry); 3294 3295 (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_DATA, retry); 3296 (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_METADATA, retry); 3297 3298 (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_DATA, retry); 3299 (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_METADATA, retry); 3300 3301 arc_do_user_evicts(); 3302 ASSERT(spa || arc_eviction_list == NULL); 3303} 3304 3305void 3306arc_shrink(int64_t to_free) 3307{ 3308 if (arc_c > arc_c_min) { 3309 DTRACE_PROBE4(arc__shrink, uint64_t, arc_c, uint64_t, 3310 arc_c_min, uint64_t, arc_p, uint64_t, to_free); 3311 if (arc_c > arc_c_min + to_free) 3312 atomic_add_64(&arc_c, -to_free); 3313 else 3314 arc_c = arc_c_min; 3315 3316 atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift)); 3317 if (arc_c > arc_size) 3318 arc_c = MAX(arc_size, arc_c_min); 3319 if (arc_p > arc_c) 3320 arc_p = (arc_c >> 1); 3321 3322 DTRACE_PROBE2(arc__shrunk, uint64_t, arc_c, uint64_t, 3323 arc_p); 3324 3325 ASSERT(arc_c >= arc_c_min); 3326 ASSERT((int64_t)arc_p >= 0); 3327 } 3328 3329 if (arc_size > arc_c) { 3330 DTRACE_PROBE2(arc__shrink_adjust, uint64_t, arc_size, 3331 uint64_t, arc_c); 3332 (void) arc_adjust(); 3333 } 3334} 3335 3336static long needfree = 0; 3337 3338typedef enum free_memory_reason_t { 3339 FMR_UNKNOWN, 3340 FMR_NEEDFREE, 3341 FMR_LOTSFREE, 3342 FMR_SWAPFS_MINFREE, 3343 FMR_PAGES_PP_MAXIMUM, 3344 FMR_HEAP_ARENA, 3345 FMR_ZIO_ARENA, 3346 FMR_ZIO_FRAG, 3347} free_memory_reason_t; 3348 3349int64_t last_free_memory; 3350free_memory_reason_t last_free_reason; 3351 3352/* 3353 * Additional reserve of pages for pp_reserve. 3354 */ 3355int64_t arc_pages_pp_reserve = 64; 3356 3357/* 3358 * Additional reserve of pages for swapfs. 3359 */ 3360int64_t arc_swapfs_reserve = 64; 3361 3362/* 3363 * Return the amount of memory that can be consumed before reclaim will be 3364 * needed. Positive if there is sufficient free memory, negative indicates 3365 * the amount of memory that needs to be freed up. 3366 */ 3367static int64_t 3368arc_available_memory(void) 3369{ 3370 int64_t lowest = INT64_MAX; 3371 int64_t n; 3372 free_memory_reason_t r = FMR_UNKNOWN; 3373 3374#ifdef _KERNEL 3375 if (needfree > 0) { 3376 n = PAGESIZE * (-needfree); 3377 if (n < lowest) { 3378 lowest = n; 3379 r = FMR_NEEDFREE; 3380 } 3381 } 3382 3383 /* 3384 * Cooperate with pagedaemon when it's time for it to scan 3385 * and reclaim some pages. 3386 */ 3387 n = PAGESIZE * ((int64_t)freemem - zfs_arc_free_target); 3388 if (n < lowest) { 3389 lowest = n; 3390 r = FMR_LOTSFREE; 3391 } 3392 3393#ifdef sun 3394 /* 3395 * check that we're out of range of the pageout scanner. It starts to 3396 * schedule paging if freemem is less than lotsfree and needfree. 3397 * lotsfree is the high-water mark for pageout, and needfree is the 3398 * number of needed free pages. We add extra pages here to make sure 3399 * the scanner doesn't start up while we're freeing memory. 3400 */ 3401 n = PAGESIZE * (freemem - lotsfree - needfree - desfree); 3402 if (n < lowest) { 3403 lowest = n; 3404 r = FMR_LOTSFREE; 3405 } 3406 3407 /* 3408 * check to make sure that swapfs has enough space so that anon 3409 * reservations can still succeed. anon_resvmem() checks that the 3410 * availrmem is greater than swapfs_minfree, and the number of reserved 3411 * swap pages. We also add a bit of extra here just to prevent 3412 * circumstances from getting really dire. 3413 */ 3414 n = PAGESIZE * (availrmem - swapfs_minfree - swapfs_reserve - 3415 desfree - arc_swapfs_reserve); 3416 if (n < lowest) { 3417 lowest = n; 3418 r = FMR_SWAPFS_MINFREE; 3419 } 3420 3421 3422 /* 3423 * Check that we have enough availrmem that memory locking (e.g., via 3424 * mlock(3C) or memcntl(2)) can still succeed. (pages_pp_maximum 3425 * stores the number of pages that cannot be locked; when availrmem 3426 * drops below pages_pp_maximum, page locking mechanisms such as 3427 * page_pp_lock() will fail.) 3428 */ 3429 n = PAGESIZE * (availrmem - pages_pp_maximum - 3430 arc_pages_pp_reserve); 3431 if (n < lowest) { 3432 lowest = n; 3433 r = FMR_PAGES_PP_MAXIMUM; 3434 } 3435 3436#endif /* sun */ 3437#if defined(__i386) || !defined(UMA_MD_SMALL_ALLOC) 3438 /* 3439 * If we're on an i386 platform, it's possible that we'll exhaust the 3440 * kernel heap space before we ever run out of available physical 3441 * memory. Most checks of the size of the heap_area compare against 3442 * tune.t_minarmem, which is the minimum available real memory that we 3443 * can have in the system. However, this is generally fixed at 25 pages 3444 * which is so low that it's useless. In this comparison, we seek to 3445 * calculate the total heap-size, and reclaim if more than 3/4ths of the 3446 * heap is allocated. (Or, in the calculation, if less than 1/4th is 3447 * free) 3448 */ 3449 n = (int64_t)vmem_size(heap_arena, VMEM_FREE) - 3450 (vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC) >> 2); 3451 if (n < lowest) { 3452 lowest = n; 3453 r = FMR_HEAP_ARENA; 3454 } 3455#define zio_arena NULL 3456#else 3457#define zio_arena heap_arena 3458#endif 3459 3460 /* 3461 * If zio data pages are being allocated out of a separate heap segment, 3462 * then enforce that the size of available vmem for this arena remains 3463 * above about 1/16th free. 3464 * 3465 * Note: The 1/16th arena free requirement was put in place 3466 * to aggressively evict memory from the arc in order to avoid 3467 * memory fragmentation issues. 3468 */ 3469 if (zio_arena != NULL) { 3470 n = (int64_t)vmem_size(zio_arena, VMEM_FREE) - 3471 (vmem_size(zio_arena, VMEM_ALLOC) >> 4); 3472 if (n < lowest) { 3473 lowest = n; 3474 r = FMR_ZIO_ARENA; 3475 } 3476 } 3477 3478 /* 3479 * Above limits know nothing about real level of KVA fragmentation. 3480 * Start aggressive reclamation if too little sequential KVA left. 3481 */ 3482 if (lowest > 0) { 3483 n = (vmem_size(heap_arena, VMEM_MAXFREE) < zfs_max_recordsize) ? 3484 -((int64_t)vmem_size(heap_arena, VMEM_ALLOC) >> 4) : 3485 INT64_MAX; 3486 if (n < lowest) { 3487 lowest = n; 3488 r = FMR_ZIO_FRAG; 3489 } 3490 } 3491 3492#else /* _KERNEL */ 3493 /* Every 100 calls, free a small amount */ 3494 if (spa_get_random(100) == 0) 3495 lowest = -1024; 3496#endif /* _KERNEL */ 3497 3498 last_free_memory = lowest; 3499 last_free_reason = r; 3500 DTRACE_PROBE2(arc__available_memory, int64_t, lowest, int, r); 3501 return (lowest); 3502} 3503 3504 3505/* 3506 * Determine if the system is under memory pressure and is asking 3507 * to reclaim memory. A return value of TRUE indicates that the system 3508 * is under memory pressure and that the arc should adjust accordingly. 3509 */ 3510static boolean_t 3511arc_reclaim_needed(void) 3512{ 3513 return (arc_available_memory() < 0); 3514} 3515 3516extern kmem_cache_t *zio_buf_cache[]; 3517extern kmem_cache_t *zio_data_buf_cache[]; 3518extern kmem_cache_t *range_seg_cache; 3519 3520static __noinline void 3521arc_kmem_reap_now(void) 3522{ 3523 size_t i; 3524 kmem_cache_t *prev_cache = NULL; 3525 kmem_cache_t *prev_data_cache = NULL; 3526 3527 DTRACE_PROBE(arc__kmem_reap_start); 3528#ifdef _KERNEL 3529 if (arc_meta_used >= arc_meta_limit) { 3530 /* 3531 * We are exceeding our meta-data cache limit. 3532 * Purge some DNLC entries to release holds on meta-data. 3533 */ 3534 dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent); 3535 } 3536#if defined(__i386) 3537 /* 3538 * Reclaim unused memory from all kmem caches. 3539 */ 3540 kmem_reap(); 3541#endif 3542#endif 3543 3544 for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) { 3545 if (zio_buf_cache[i] != prev_cache) { 3546 prev_cache = zio_buf_cache[i]; 3547 kmem_cache_reap_now(zio_buf_cache[i]); 3548 } 3549 if (zio_data_buf_cache[i] != prev_data_cache) { 3550 prev_data_cache = zio_data_buf_cache[i]; 3551 kmem_cache_reap_now(zio_data_buf_cache[i]); 3552 } 3553 } 3554 kmem_cache_reap_now(buf_cache); 3555 kmem_cache_reap_now(hdr_full_cache); 3556 kmem_cache_reap_now(hdr_l2only_cache); 3557 kmem_cache_reap_now(range_seg_cache); 3558 3559#ifdef sun 3560 if (zio_arena != NULL) { 3561 /* 3562 * Ask the vmem arena to reclaim unused memory from its 3563 * quantum caches. 3564 */ 3565 vmem_qcache_reap(zio_arena); 3566 } 3567#endif 3568 DTRACE_PROBE(arc__kmem_reap_end); 3569} 3570 3571/* 3572 * Threads can block in arc_get_data_buf() waiting for this thread to evict 3573 * enough data and signal them to proceed. When this happens, the threads in 3574 * arc_get_data_buf() are sleeping while holding the hash lock for their 3575 * particular arc header. Thus, we must be careful to never sleep on a 3576 * hash lock in this thread. This is to prevent the following deadlock: 3577 * 3578 * - Thread A sleeps on CV in arc_get_data_buf() holding hash lock "L", 3579 * waiting for the reclaim thread to signal it. 3580 * 3581 * - arc_reclaim_thread() tries to acquire hash lock "L" using mutex_enter, 3582 * fails, and goes to sleep forever. 3583 * 3584 * This possible deadlock is avoided by always acquiring a hash lock 3585 * using mutex_tryenter() from arc_reclaim_thread(). 3586 */ 3587static void 3588arc_reclaim_thread(void *dummy __unused) 3589{ 3590 clock_t growtime = 0; 3591 callb_cpr_t cpr; 3592 3593 CALLB_CPR_INIT(&cpr, &arc_reclaim_lock, callb_generic_cpr, FTAG); 3594 3595 mutex_enter(&arc_reclaim_lock); 3596 while (!arc_reclaim_thread_exit) { 3597 int64_t free_memory = arc_available_memory(); 3598 uint64_t evicted = 0; 3599 3600 mutex_exit(&arc_reclaim_lock); 3601 3602 if (free_memory < 0) { 3603 3604 arc_no_grow = B_TRUE; 3605 arc_warm = B_TRUE; 3606 3607 /* 3608 * Wait at least zfs_grow_retry (default 60) seconds 3609 * before considering growing. 3610 */ 3611 growtime = ddi_get_lbolt() + (arc_grow_retry * hz); 3612 3613 arc_kmem_reap_now(); 3614 3615 /* 3616 * If we are still low on memory, shrink the ARC 3617 * so that we have arc_shrink_min free space. 3618 */ 3619 free_memory = arc_available_memory(); 3620 3621 int64_t to_free = 3622 (arc_c >> arc_shrink_shift) - free_memory; 3623 if (to_free > 0) { 3624#ifdef _KERNEL 3625 to_free = MAX(to_free, ptob(needfree)); 3626#endif 3627 arc_shrink(to_free); 3628 } 3629 } else if (free_memory < arc_c >> arc_no_grow_shift) { 3630 arc_no_grow = B_TRUE; 3631 } else if (ddi_get_lbolt() >= growtime) { 3632 arc_no_grow = B_FALSE; 3633 } 3634 3635 evicted = arc_adjust(); 3636 3637 mutex_enter(&arc_reclaim_lock); 3638 3639 /* 3640 * If evicted is zero, we couldn't evict anything via 3641 * arc_adjust(). This could be due to hash lock 3642 * collisions, but more likely due to the majority of 3643 * arc buffers being unevictable. Therefore, even if 3644 * arc_size is above arc_c, another pass is unlikely to 3645 * be helpful and could potentially cause us to enter an 3646 * infinite loop. 3647 */ 3648 if (arc_size <= arc_c || evicted == 0) { 3649#ifdef _KERNEL 3650 needfree = 0; 3651#endif 3652 /* 3653 * We're either no longer overflowing, or we 3654 * can't evict anything more, so we should wake 3655 * up any threads before we go to sleep. 3656 */ 3657 cv_broadcast(&arc_reclaim_waiters_cv); 3658 3659 /* 3660 * Block until signaled, or after one second (we 3661 * might need to perform arc_kmem_reap_now() 3662 * even if we aren't being signalled) 3663 */ 3664 CALLB_CPR_SAFE_BEGIN(&cpr); 3665 (void) cv_timedwait(&arc_reclaim_thread_cv, 3666 &arc_reclaim_lock, hz); 3667 CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_lock); 3668 } 3669 } 3670 3671 arc_reclaim_thread_exit = FALSE; 3672 cv_broadcast(&arc_reclaim_thread_cv); 3673 CALLB_CPR_EXIT(&cpr); /* drops arc_reclaim_lock */ 3674 thread_exit(); 3675} 3676 3677static void 3678arc_user_evicts_thread(void *dummy __unused) 3679{ 3680 callb_cpr_t cpr; 3681 3682 CALLB_CPR_INIT(&cpr, &arc_user_evicts_lock, callb_generic_cpr, FTAG); 3683 3684 mutex_enter(&arc_user_evicts_lock); 3685 while (!arc_user_evicts_thread_exit) { 3686 mutex_exit(&arc_user_evicts_lock); 3687 3688 arc_do_user_evicts(); 3689 3690 /* 3691 * This is necessary in order for the mdb ::arc dcmd to 3692 * show up to date information. Since the ::arc command 3693 * does not call the kstat's update function, without 3694 * this call, the command may show stale stats for the 3695 * anon, mru, mru_ghost, mfu, and mfu_ghost lists. Even 3696 * with this change, the data might be up to 1 second 3697 * out of date; but that should suffice. The arc_state_t 3698 * structures can be queried directly if more accurate 3699 * information is needed. 3700 */ 3701 if (arc_ksp != NULL) 3702 arc_ksp->ks_update(arc_ksp, KSTAT_READ); 3703 3704 mutex_enter(&arc_user_evicts_lock); 3705 3706 /* 3707 * Block until signaled, or after one second (we need to 3708 * call the arc's kstat update function regularly). 3709 */ 3710 CALLB_CPR_SAFE_BEGIN(&cpr); 3711 (void) cv_timedwait(&arc_user_evicts_cv, 3712 &arc_user_evicts_lock, hz); 3713 CALLB_CPR_SAFE_END(&cpr, &arc_user_evicts_lock); 3714 } 3715 3716 arc_user_evicts_thread_exit = FALSE; 3717 cv_broadcast(&arc_user_evicts_cv); 3718 CALLB_CPR_EXIT(&cpr); /* drops arc_user_evicts_lock */ 3719 thread_exit(); 3720} 3721 3722/* 3723 * Adapt arc info given the number of bytes we are trying to add and 3724 * the state that we are comming from. This function is only called 3725 * when we are adding new content to the cache. 3726 */ 3727static void 3728arc_adapt(int bytes, arc_state_t *state) 3729{ 3730 int mult; 3731 uint64_t arc_p_min = (arc_c >> arc_p_min_shift); 3732 int64_t mrug_size = refcount_count(&arc_mru_ghost->arcs_size); 3733 int64_t mfug_size = refcount_count(&arc_mfu_ghost->arcs_size); 3734 3735 if (state == arc_l2c_only) 3736 return; 3737 3738 ASSERT(bytes > 0); 3739 /* 3740 * Adapt the target size of the MRU list: 3741 * - if we just hit in the MRU ghost list, then increase 3742 * the target size of the MRU list. 3743 * - if we just hit in the MFU ghost list, then increase 3744 * the target size of the MFU list by decreasing the 3745 * target size of the MRU list. 3746 */ 3747 if (state == arc_mru_ghost) { 3748 mult = (mrug_size >= mfug_size) ? 1 : (mfug_size / mrug_size); 3749 mult = MIN(mult, 10); /* avoid wild arc_p adjustment */ 3750 3751 arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult); 3752 } else if (state == arc_mfu_ghost) { 3753 uint64_t delta; 3754 3755 mult = (mfug_size >= mrug_size) ? 1 : (mrug_size / mfug_size); 3756 mult = MIN(mult, 10); 3757 3758 delta = MIN(bytes * mult, arc_p); 3759 arc_p = MAX(arc_p_min, arc_p - delta); 3760 } 3761 ASSERT((int64_t)arc_p >= 0); 3762 3763 if (arc_reclaim_needed()) { 3764 cv_signal(&arc_reclaim_thread_cv); 3765 return; 3766 } 3767 3768 if (arc_no_grow) 3769 return; 3770 3771 if (arc_c >= arc_c_max) 3772 return; 3773 3774 /* 3775 * If we're within (2 * maxblocksize) bytes of the target 3776 * cache size, increment the target cache size 3777 */ 3778 if (arc_size > arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) { 3779 DTRACE_PROBE1(arc__inc_adapt, int, bytes); 3780 atomic_add_64(&arc_c, (int64_t)bytes); 3781 if (arc_c > arc_c_max) 3782 arc_c = arc_c_max; 3783 else if (state == arc_anon) 3784 atomic_add_64(&arc_p, (int64_t)bytes); 3785 if (arc_p > arc_c) 3786 arc_p = arc_c; 3787 } 3788 ASSERT((int64_t)arc_p >= 0); 3789} 3790 3791/* 3792 * Check if arc_size has grown past our upper threshold, determined by 3793 * zfs_arc_overflow_shift. 3794 */ 3795static boolean_t 3796arc_is_overflowing(void) 3797{ 3798 /* Always allow at least one block of overflow */ 3799 uint64_t overflow = MAX(SPA_MAXBLOCKSIZE, 3800 arc_c >> zfs_arc_overflow_shift); 3801 3802 return (arc_size >= arc_c + overflow); 3803} 3804 3805/* 3806 * The buffer, supplied as the first argument, needs a data block. If we 3807 * are hitting the hard limit for the cache size, we must sleep, waiting 3808 * for the eviction thread to catch up. If we're past the target size 3809 * but below the hard limit, we'll only signal the reclaim thread and 3810 * continue on. 3811 */ 3812static void 3813arc_get_data_buf(arc_buf_t *buf) 3814{ 3815 arc_state_t *state = buf->b_hdr->b_l1hdr.b_state; 3816 uint64_t size = buf->b_hdr->b_size; 3817 arc_buf_contents_t type = arc_buf_type(buf->b_hdr); 3818 3819 arc_adapt(size, state); 3820 3821 /* 3822 * If arc_size is currently overflowing, and has grown past our 3823 * upper limit, we must be adding data faster than the evict 3824 * thread can evict. Thus, to ensure we don't compound the 3825 * problem by adding more data and forcing arc_size to grow even 3826 * further past it's target size, we halt and wait for the 3827 * eviction thread to catch up. 3828 * 3829 * It's also possible that the reclaim thread is unable to evict 3830 * enough buffers to get arc_size below the overflow limit (e.g. 3831 * due to buffers being un-evictable, or hash lock collisions). 3832 * In this case, we want to proceed regardless if we're 3833 * overflowing; thus we don't use a while loop here. 3834 */ 3835 if (arc_is_overflowing()) { 3836 mutex_enter(&arc_reclaim_lock); 3837 3838 /* 3839 * Now that we've acquired the lock, we may no longer be 3840 * over the overflow limit, lets check. 3841 * 3842 * We're ignoring the case of spurious wake ups. If that 3843 * were to happen, it'd let this thread consume an ARC 3844 * buffer before it should have (i.e. before we're under 3845 * the overflow limit and were signalled by the reclaim 3846 * thread). As long as that is a rare occurrence, it 3847 * shouldn't cause any harm. 3848 */ 3849 if (arc_is_overflowing()) { 3850 cv_signal(&arc_reclaim_thread_cv); 3851 cv_wait(&arc_reclaim_waiters_cv, &arc_reclaim_lock); 3852 } 3853 3854 mutex_exit(&arc_reclaim_lock); 3855 } 3856 3857 if (type == ARC_BUFC_METADATA) { 3858 buf->b_data = zio_buf_alloc(size); 3859 arc_space_consume(size, ARC_SPACE_META); 3860 } else { 3861 ASSERT(type == ARC_BUFC_DATA); 3862 buf->b_data = zio_data_buf_alloc(size); 3863 arc_space_consume(size, ARC_SPACE_DATA); 3864 } 3865 3866 /* 3867 * Update the state size. Note that ghost states have a 3868 * "ghost size" and so don't need to be updated. 3869 */ 3870 if (!GHOST_STATE(buf->b_hdr->b_l1hdr.b_state)) { 3871 arc_buf_hdr_t *hdr = buf->b_hdr; 3872 arc_state_t *state = hdr->b_l1hdr.b_state; 3873 3874 (void) refcount_add_many(&state->arcs_size, size, buf); 3875 3876 /* 3877 * If this is reached via arc_read, the link is 3878 * protected by the hash lock. If reached via 3879 * arc_buf_alloc, the header should not be accessed by 3880 * any other thread. And, if reached via arc_read_done, 3881 * the hash lock will protect it if it's found in the 3882 * hash table; otherwise no other thread should be 3883 * trying to [add|remove]_reference it. 3884 */ 3885 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) { 3886 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 3887 atomic_add_64(&hdr->b_l1hdr.b_state->arcs_lsize[type], 3888 size); 3889 } 3890 /* 3891 * If we are growing the cache, and we are adding anonymous 3892 * data, and we have outgrown arc_p, update arc_p 3893 */ 3894 if (arc_size < arc_c && hdr->b_l1hdr.b_state == arc_anon && 3895 (refcount_count(&arc_anon->arcs_size) + 3896 refcount_count(&arc_mru->arcs_size) > arc_p)) 3897 arc_p = MIN(arc_c, arc_p + size); 3898 } 3899 ARCSTAT_BUMP(arcstat_allocated); 3900} 3901 3902/* 3903 * This routine is called whenever a buffer is accessed. 3904 * NOTE: the hash lock is dropped in this function. 3905 */ 3906static void 3907arc_access(arc_buf_hdr_t *hdr, kmutex_t *hash_lock) 3908{ 3909 clock_t now; 3910 3911 ASSERT(MUTEX_HELD(hash_lock)); 3912 ASSERT(HDR_HAS_L1HDR(hdr)); 3913 3914 if (hdr->b_l1hdr.b_state == arc_anon) { 3915 /* 3916 * This buffer is not in the cache, and does not 3917 * appear in our "ghost" list. Add the new buffer 3918 * to the MRU state. 3919 */ 3920 3921 ASSERT0(hdr->b_l1hdr.b_arc_access); 3922 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 3923 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr); 3924 arc_change_state(arc_mru, hdr, hash_lock); 3925 3926 } else if (hdr->b_l1hdr.b_state == arc_mru) { 3927 now = ddi_get_lbolt(); 3928 3929 /* 3930 * If this buffer is here because of a prefetch, then either: 3931 * - clear the flag if this is a "referencing" read 3932 * (any subsequent access will bump this into the MFU state). 3933 * or 3934 * - move the buffer to the head of the list if this is 3935 * another prefetch (to make it less likely to be evicted). 3936 */ 3937 if (HDR_PREFETCH(hdr)) { 3938 if (refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) { 3939 /* link protected by hash lock */ 3940 ASSERT(multilist_link_active( 3941 &hdr->b_l1hdr.b_arc_node)); 3942 } else { 3943 hdr->b_flags &= ~ARC_FLAG_PREFETCH; 3944 ARCSTAT_BUMP(arcstat_mru_hits); 3945 } 3946 hdr->b_l1hdr.b_arc_access = now; 3947 return; 3948 } 3949 3950 /* 3951 * This buffer has been "accessed" only once so far, 3952 * but it is still in the cache. Move it to the MFU 3953 * state. 3954 */ 3955 if (now > hdr->b_l1hdr.b_arc_access + ARC_MINTIME) { 3956 /* 3957 * More than 125ms have passed since we 3958 * instantiated this buffer. Move it to the 3959 * most frequently used state. 3960 */ 3961 hdr->b_l1hdr.b_arc_access = now; 3962 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 3963 arc_change_state(arc_mfu, hdr, hash_lock); 3964 } 3965 ARCSTAT_BUMP(arcstat_mru_hits); 3966 } else if (hdr->b_l1hdr.b_state == arc_mru_ghost) { 3967 arc_state_t *new_state; 3968 /* 3969 * This buffer has been "accessed" recently, but 3970 * was evicted from the cache. Move it to the 3971 * MFU state. 3972 */ 3973 3974 if (HDR_PREFETCH(hdr)) { 3975 new_state = arc_mru; 3976 if (refcount_count(&hdr->b_l1hdr.b_refcnt) > 0) 3977 hdr->b_flags &= ~ARC_FLAG_PREFETCH; 3978 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr); 3979 } else { 3980 new_state = arc_mfu; 3981 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 3982 } 3983 3984 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 3985 arc_change_state(new_state, hdr, hash_lock); 3986 3987 ARCSTAT_BUMP(arcstat_mru_ghost_hits); 3988 } else if (hdr->b_l1hdr.b_state == arc_mfu) { 3989 /* 3990 * This buffer has been accessed more than once and is 3991 * still in the cache. Keep it in the MFU state. 3992 * 3993 * NOTE: an add_reference() that occurred when we did 3994 * the arc_read() will have kicked this off the list. 3995 * If it was a prefetch, we will explicitly move it to 3996 * the head of the list now. 3997 */ 3998 if ((HDR_PREFETCH(hdr)) != 0) { 3999 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 4000 /* link protected by hash_lock */ 4001 ASSERT(multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 4002 } 4003 ARCSTAT_BUMP(arcstat_mfu_hits); 4004 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 4005 } else if (hdr->b_l1hdr.b_state == arc_mfu_ghost) { 4006 arc_state_t *new_state = arc_mfu; 4007 /* 4008 * This buffer has been accessed more than once but has 4009 * been evicted from the cache. Move it back to the 4010 * MFU state. 4011 */ 4012 4013 if (HDR_PREFETCH(hdr)) { 4014 /* 4015 * This is a prefetch access... 4016 * move this block back to the MRU state. 4017 */ 4018 ASSERT0(refcount_count(&hdr->b_l1hdr.b_refcnt)); 4019 new_state = arc_mru; 4020 } 4021 4022 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 4023 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 4024 arc_change_state(new_state, hdr, hash_lock); 4025 4026 ARCSTAT_BUMP(arcstat_mfu_ghost_hits); 4027 } else if (hdr->b_l1hdr.b_state == arc_l2c_only) { 4028 /* 4029 * This buffer is on the 2nd Level ARC. 4030 */ 4031 4032 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 4033 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 4034 arc_change_state(arc_mfu, hdr, hash_lock); 4035 } else { 4036 ASSERT(!"invalid arc state"); 4037 } 4038} 4039 4040/* a generic arc_done_func_t which you can use */ 4041/* ARGSUSED */ 4042void 4043arc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg) 4044{ 4045 if (zio == NULL || zio->io_error == 0) 4046 bcopy(buf->b_data, arg, buf->b_hdr->b_size); 4047 VERIFY(arc_buf_remove_ref(buf, arg)); 4048} 4049 4050/* a generic arc_done_func_t */ 4051void 4052arc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg) 4053{ 4054 arc_buf_t **bufp = arg; 4055 if (zio && zio->io_error) { 4056 VERIFY(arc_buf_remove_ref(buf, arg)); 4057 *bufp = NULL; 4058 } else { 4059 *bufp = buf; 4060 ASSERT(buf->b_data); 4061 } 4062} 4063 4064static void 4065arc_read_done(zio_t *zio) 4066{ 4067 arc_buf_hdr_t *hdr; 4068 arc_buf_t *buf; 4069 arc_buf_t *abuf; /* buffer we're assigning to callback */ 4070 kmutex_t *hash_lock = NULL; 4071 arc_callback_t *callback_list, *acb; 4072 int freeable = FALSE; 4073 4074 buf = zio->io_private; 4075 hdr = buf->b_hdr; 4076 4077 /* 4078 * The hdr was inserted into hash-table and removed from lists 4079 * prior to starting I/O. We should find this header, since 4080 * it's in the hash table, and it should be legit since it's 4081 * not possible to evict it during the I/O. The only possible 4082 * reason for it not to be found is if we were freed during the 4083 * read. 4084 */ 4085 if (HDR_IN_HASH_TABLE(hdr)) { 4086 ASSERT3U(hdr->b_birth, ==, BP_PHYSICAL_BIRTH(zio->io_bp)); 4087 ASSERT3U(hdr->b_dva.dva_word[0], ==, 4088 BP_IDENTITY(zio->io_bp)->dva_word[0]); 4089 ASSERT3U(hdr->b_dva.dva_word[1], ==, 4090 BP_IDENTITY(zio->io_bp)->dva_word[1]); 4091 4092 arc_buf_hdr_t *found = buf_hash_find(hdr->b_spa, zio->io_bp, 4093 &hash_lock); 4094 4095 ASSERT((found == NULL && HDR_FREED_IN_READ(hdr) && 4096 hash_lock == NULL) || 4097 (found == hdr && 4098 DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) || 4099 (found == hdr && HDR_L2_READING(hdr))); 4100 } 4101 4102 hdr->b_flags &= ~ARC_FLAG_L2_EVICTED; 4103 if (l2arc_noprefetch && HDR_PREFETCH(hdr)) 4104 hdr->b_flags &= ~ARC_FLAG_L2CACHE; 4105 4106 /* byteswap if necessary */ 4107 callback_list = hdr->b_l1hdr.b_acb; 4108 ASSERT(callback_list != NULL); 4109 if (BP_SHOULD_BYTESWAP(zio->io_bp) && zio->io_error == 0) { 4110 dmu_object_byteswap_t bswap = 4111 DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp)); 4112 arc_byteswap_func_t *func = BP_GET_LEVEL(zio->io_bp) > 0 ? 4113 byteswap_uint64_array : 4114 dmu_ot_byteswap[bswap].ob_func; 4115 func(buf->b_data, hdr->b_size); 4116 } 4117 4118 arc_cksum_compute(buf, B_FALSE); 4119#ifdef illumos 4120 arc_buf_watch(buf); 4121#endif /* illumos */ 4122 4123 if (hash_lock && zio->io_error == 0 && 4124 hdr->b_l1hdr.b_state == arc_anon) { 4125 /* 4126 * Only call arc_access on anonymous buffers. This is because 4127 * if we've issued an I/O for an evicted buffer, we've already 4128 * called arc_access (to prevent any simultaneous readers from 4129 * getting confused). 4130 */ 4131 arc_access(hdr, hash_lock); 4132 } 4133 4134 /* create copies of the data buffer for the callers */ 4135 abuf = buf; 4136 for (acb = callback_list; acb; acb = acb->acb_next) { 4137 if (acb->acb_done) { 4138 if (abuf == NULL) { 4139 ARCSTAT_BUMP(arcstat_duplicate_reads); 4140 abuf = arc_buf_clone(buf); 4141 } 4142 acb->acb_buf = abuf; 4143 abuf = NULL; 4144 } 4145 } 4146 hdr->b_l1hdr.b_acb = NULL; 4147 hdr->b_flags &= ~ARC_FLAG_IO_IN_PROGRESS; 4148 ASSERT(!HDR_BUF_AVAILABLE(hdr)); 4149 if (abuf == buf) { 4150 ASSERT(buf->b_efunc == NULL); 4151 ASSERT(hdr->b_l1hdr.b_datacnt == 1); 4152 hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE; 4153 } 4154 4155 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt) || 4156 callback_list != NULL); 4157 4158 if (zio->io_error != 0) { 4159 hdr->b_flags |= ARC_FLAG_IO_ERROR; 4160 if (hdr->b_l1hdr.b_state != arc_anon) 4161 arc_change_state(arc_anon, hdr, hash_lock); 4162 if (HDR_IN_HASH_TABLE(hdr)) 4163 buf_hash_remove(hdr); 4164 freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt); 4165 } 4166 4167 /* 4168 * Broadcast before we drop the hash_lock to avoid the possibility 4169 * that the hdr (and hence the cv) might be freed before we get to 4170 * the cv_broadcast(). 4171 */ 4172 cv_broadcast(&hdr->b_l1hdr.b_cv); 4173 4174 if (hash_lock != NULL) { 4175 mutex_exit(hash_lock); 4176 } else { 4177 /* 4178 * This block was freed while we waited for the read to 4179 * complete. It has been removed from the hash table and 4180 * moved to the anonymous state (so that it won't show up 4181 * in the cache). 4182 */ 4183 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon); 4184 freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt); 4185 } 4186 4187 /* execute each callback and free its structure */ 4188 while ((acb = callback_list) != NULL) { 4189 if (acb->acb_done) 4190 acb->acb_done(zio, acb->acb_buf, acb->acb_private); 4191 4192 if (acb->acb_zio_dummy != NULL) { 4193 acb->acb_zio_dummy->io_error = zio->io_error; 4194 zio_nowait(acb->acb_zio_dummy); 4195 } 4196 4197 callback_list = acb->acb_next; 4198 kmem_free(acb, sizeof (arc_callback_t)); 4199 } 4200 4201 if (freeable) 4202 arc_hdr_destroy(hdr); 4203} 4204 4205/* 4206 * "Read" the block at the specified DVA (in bp) via the 4207 * cache. If the block is found in the cache, invoke the provided 4208 * callback immediately and return. Note that the `zio' parameter 4209 * in the callback will be NULL in this case, since no IO was 4210 * required. If the block is not in the cache pass the read request 4211 * on to the spa with a substitute callback function, so that the 4212 * requested block will be added to the cache. 4213 * 4214 * If a read request arrives for a block that has a read in-progress, 4215 * either wait for the in-progress read to complete (and return the 4216 * results); or, if this is a read with a "done" func, add a record 4217 * to the read to invoke the "done" func when the read completes, 4218 * and return; or just return. 4219 * 4220 * arc_read_done() will invoke all the requested "done" functions 4221 * for readers of this block. 4222 */ 4223int 4224arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_done_func_t *done, 4225 void *private, zio_priority_t priority, int zio_flags, 4226 arc_flags_t *arc_flags, const zbookmark_phys_t *zb) 4227{ 4228 arc_buf_hdr_t *hdr = NULL; 4229 arc_buf_t *buf = NULL; 4230 kmutex_t *hash_lock = NULL; 4231 zio_t *rzio; 4232 uint64_t guid = spa_load_guid(spa); 4233 4234 ASSERT(!BP_IS_EMBEDDED(bp) || 4235 BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA); 4236 4237top: 4238 if (!BP_IS_EMBEDDED(bp)) { 4239 /* 4240 * Embedded BP's have no DVA and require no I/O to "read". 4241 * Create an anonymous arc buf to back it. 4242 */ 4243 hdr = buf_hash_find(guid, bp, &hash_lock); 4244 } 4245 4246 if (hdr != NULL && HDR_HAS_L1HDR(hdr) && hdr->b_l1hdr.b_datacnt > 0) { 4247 4248 *arc_flags |= ARC_FLAG_CACHED; 4249 4250 if (HDR_IO_IN_PROGRESS(hdr)) { 4251 4252 if ((hdr->b_flags & ARC_FLAG_PRIO_ASYNC_READ) && 4253 priority == ZIO_PRIORITY_SYNC_READ) { 4254 /* 4255 * This sync read must wait for an 4256 * in-progress async read (e.g. a predictive 4257 * prefetch). Async reads are queued 4258 * separately at the vdev_queue layer, so 4259 * this is a form of priority inversion. 4260 * Ideally, we would "inherit" the demand 4261 * i/o's priority by moving the i/o from 4262 * the async queue to the synchronous queue, 4263 * but there is currently no mechanism to do 4264 * so. Track this so that we can evaluate 4265 * the magnitude of this potential performance 4266 * problem. 4267 * 4268 * Note that if the prefetch i/o is already 4269 * active (has been issued to the device), 4270 * the prefetch improved performance, because 4271 * we issued it sooner than we would have 4272 * without the prefetch. 4273 */ 4274 DTRACE_PROBE1(arc__sync__wait__for__async, 4275 arc_buf_hdr_t *, hdr); 4276 ARCSTAT_BUMP(arcstat_sync_wait_for_async); 4277 } 4278 if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) { 4279 hdr->b_flags &= ~ARC_FLAG_PREDICTIVE_PREFETCH; 4280 } 4281 4282 if (*arc_flags & ARC_FLAG_WAIT) { 4283 cv_wait(&hdr->b_l1hdr.b_cv, hash_lock); 4284 mutex_exit(hash_lock); 4285 goto top; 4286 } 4287 ASSERT(*arc_flags & ARC_FLAG_NOWAIT); 4288 4289 if (done) { 4290 arc_callback_t *acb = NULL; 4291 4292 acb = kmem_zalloc(sizeof (arc_callback_t), 4293 KM_SLEEP); 4294 acb->acb_done = done; 4295 acb->acb_private = private; 4296 if (pio != NULL) 4297 acb->acb_zio_dummy = zio_null(pio, 4298 spa, NULL, NULL, NULL, zio_flags); 4299 4300 ASSERT(acb->acb_done != NULL); 4301 acb->acb_next = hdr->b_l1hdr.b_acb; 4302 hdr->b_l1hdr.b_acb = acb; 4303 add_reference(hdr, hash_lock, private); 4304 mutex_exit(hash_lock); 4305 return (0); 4306 } 4307 mutex_exit(hash_lock); 4308 return (0); 4309 } 4310 4311 ASSERT(hdr->b_l1hdr.b_state == arc_mru || 4312 hdr->b_l1hdr.b_state == arc_mfu); 4313 4314 if (done) { 4315 if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) { 4316 /* 4317 * This is a demand read which does not have to 4318 * wait for i/o because we did a predictive 4319 * prefetch i/o for it, which has completed. 4320 */ 4321 DTRACE_PROBE1( 4322 arc__demand__hit__predictive__prefetch, 4323 arc_buf_hdr_t *, hdr); 4324 ARCSTAT_BUMP( 4325 arcstat_demand_hit_predictive_prefetch); 4326 hdr->b_flags &= ~ARC_FLAG_PREDICTIVE_PREFETCH; 4327 } 4328 add_reference(hdr, hash_lock, private); 4329 /* 4330 * If this block is already in use, create a new 4331 * copy of the data so that we will be guaranteed 4332 * that arc_release() will always succeed. 4333 */ 4334 buf = hdr->b_l1hdr.b_buf; 4335 ASSERT(buf); 4336 ASSERT(buf->b_data); 4337 if (HDR_BUF_AVAILABLE(hdr)) { 4338 ASSERT(buf->b_efunc == NULL); 4339 hdr->b_flags &= ~ARC_FLAG_BUF_AVAILABLE; 4340 } else { 4341 buf = arc_buf_clone(buf); 4342 } 4343 4344 } else if (*arc_flags & ARC_FLAG_PREFETCH && 4345 refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) { 4346 hdr->b_flags |= ARC_FLAG_PREFETCH; 4347 } 4348 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr); 4349 arc_access(hdr, hash_lock); 4350 if (*arc_flags & ARC_FLAG_L2CACHE) 4351 hdr->b_flags |= ARC_FLAG_L2CACHE; 4352 if (*arc_flags & ARC_FLAG_L2COMPRESS) 4353 hdr->b_flags |= ARC_FLAG_L2COMPRESS; 4354 mutex_exit(hash_lock); 4355 ARCSTAT_BUMP(arcstat_hits); 4356 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr), 4357 demand, prefetch, !HDR_ISTYPE_METADATA(hdr), 4358 data, metadata, hits); 4359 4360 if (done) 4361 done(NULL, buf, private); 4362 } else { 4363 uint64_t size = BP_GET_LSIZE(bp); 4364 arc_callback_t *acb; 4365 vdev_t *vd = NULL; 4366 uint64_t addr = 0; 4367 boolean_t devw = B_FALSE; 4368 enum zio_compress b_compress = ZIO_COMPRESS_OFF; 4369 int32_t b_asize = 0; 4370 4371 if (hdr == NULL) { 4372 /* this block is not in the cache */ 4373 arc_buf_hdr_t *exists = NULL; 4374 arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp); 4375 buf = arc_buf_alloc(spa, size, private, type); 4376 hdr = buf->b_hdr; 4377 if (!BP_IS_EMBEDDED(bp)) { 4378 hdr->b_dva = *BP_IDENTITY(bp); 4379 hdr->b_birth = BP_PHYSICAL_BIRTH(bp); 4380 exists = buf_hash_insert(hdr, &hash_lock); 4381 } 4382 if (exists != NULL) { 4383 /* somebody beat us to the hash insert */ 4384 mutex_exit(hash_lock); 4385 buf_discard_identity(hdr); 4386 (void) arc_buf_remove_ref(buf, private); 4387 goto top; /* restart the IO request */ 4388 } 4389 4390 /* 4391 * If there is a callback, we pass our reference to 4392 * it; otherwise we remove our reference. 4393 */ 4394 if (done == NULL) { 4395 (void) remove_reference(hdr, hash_lock, 4396 private); 4397 } 4398 if (*arc_flags & ARC_FLAG_PREFETCH) 4399 hdr->b_flags |= ARC_FLAG_PREFETCH; 4400 if (*arc_flags & ARC_FLAG_L2CACHE) 4401 hdr->b_flags |= ARC_FLAG_L2CACHE; 4402 if (*arc_flags & ARC_FLAG_L2COMPRESS) 4403 hdr->b_flags |= ARC_FLAG_L2COMPRESS; 4404 if (BP_GET_LEVEL(bp) > 0) 4405 hdr->b_flags |= ARC_FLAG_INDIRECT; 4406 } else { 4407 /* 4408 * This block is in the ghost cache. If it was L2-only 4409 * (and thus didn't have an L1 hdr), we realloc the 4410 * header to add an L1 hdr. 4411 */ 4412 if (!HDR_HAS_L1HDR(hdr)) { 4413 hdr = arc_hdr_realloc(hdr, hdr_l2only_cache, 4414 hdr_full_cache); 4415 } 4416 4417 ASSERT(GHOST_STATE(hdr->b_l1hdr.b_state)); 4418 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 4419 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 4420 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 4421 4422 /* 4423 * If there is a callback, we pass a reference to it. 4424 */ 4425 if (done != NULL) 4426 add_reference(hdr, hash_lock, private); 4427 if (*arc_flags & ARC_FLAG_PREFETCH) 4428 hdr->b_flags |= ARC_FLAG_PREFETCH; 4429 if (*arc_flags & ARC_FLAG_L2CACHE) 4430 hdr->b_flags |= ARC_FLAG_L2CACHE; 4431 if (*arc_flags & ARC_FLAG_L2COMPRESS) 4432 hdr->b_flags |= ARC_FLAG_L2COMPRESS; 4433 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 4434 buf->b_hdr = hdr; 4435 buf->b_data = NULL; 4436 buf->b_efunc = NULL; 4437 buf->b_private = NULL; 4438 buf->b_next = NULL; 4439 hdr->b_l1hdr.b_buf = buf; 4440 ASSERT0(hdr->b_l1hdr.b_datacnt); 4441 hdr->b_l1hdr.b_datacnt = 1; 4442 arc_get_data_buf(buf); 4443 arc_access(hdr, hash_lock); 4444 } 4445 4446 if (*arc_flags & ARC_FLAG_PREDICTIVE_PREFETCH) 4447 hdr->b_flags |= ARC_FLAG_PREDICTIVE_PREFETCH; 4448 ASSERT(!GHOST_STATE(hdr->b_l1hdr.b_state)); 4449 4450 acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP); 4451 acb->acb_done = done; 4452 acb->acb_private = private; 4453 4454 ASSERT(hdr->b_l1hdr.b_acb == NULL); 4455 hdr->b_l1hdr.b_acb = acb; 4456 hdr->b_flags |= ARC_FLAG_IO_IN_PROGRESS; 4457 4458 if (HDR_HAS_L2HDR(hdr) && 4459 (vd = hdr->b_l2hdr.b_dev->l2ad_vdev) != NULL) { 4460 devw = hdr->b_l2hdr.b_dev->l2ad_writing; 4461 addr = hdr->b_l2hdr.b_daddr; 4462 b_compress = hdr->b_l2hdr.b_compress; 4463 b_asize = hdr->b_l2hdr.b_asize; 4464 /* 4465 * Lock out device removal. 4466 */ 4467 if (vdev_is_dead(vd) || 4468 !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER)) 4469 vd = NULL; 4470 } 4471 4472 if (hash_lock != NULL) 4473 mutex_exit(hash_lock); 4474 4475 /* 4476 * At this point, we have a level 1 cache miss. Try again in 4477 * L2ARC if possible. 4478 */ 4479 ASSERT3U(hdr->b_size, ==, size); 4480 DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp, 4481 uint64_t, size, zbookmark_phys_t *, zb); 4482 ARCSTAT_BUMP(arcstat_misses); 4483 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr), 4484 demand, prefetch, !HDR_ISTYPE_METADATA(hdr), 4485 data, metadata, misses); 4486#ifdef _KERNEL 4487 curthread->td_ru.ru_inblock++; 4488#endif 4489 4490 if (priority == ZIO_PRIORITY_ASYNC_READ) 4491 hdr->b_flags |= ARC_FLAG_PRIO_ASYNC_READ; 4492 else 4493 hdr->b_flags &= ~ARC_FLAG_PRIO_ASYNC_READ; 4494 4495 if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) { 4496 /* 4497 * Read from the L2ARC if the following are true: 4498 * 1. The L2ARC vdev was previously cached. 4499 * 2. This buffer still has L2ARC metadata. 4500 * 3. This buffer isn't currently writing to the L2ARC. 4501 * 4. The L2ARC entry wasn't evicted, which may 4502 * also have invalidated the vdev. 4503 * 5. This isn't prefetch and l2arc_noprefetch is set. 4504 */ 4505 if (HDR_HAS_L2HDR(hdr) && 4506 !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) && 4507 !(l2arc_noprefetch && HDR_PREFETCH(hdr))) { 4508 l2arc_read_callback_t *cb; 4509 4510 DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr); 4511 ARCSTAT_BUMP(arcstat_l2_hits); 4512 4513 cb = kmem_zalloc(sizeof (l2arc_read_callback_t), 4514 KM_SLEEP); 4515 cb->l2rcb_buf = buf; 4516 cb->l2rcb_spa = spa; 4517 cb->l2rcb_bp = *bp; 4518 cb->l2rcb_zb = *zb; 4519 cb->l2rcb_flags = zio_flags; 4520 cb->l2rcb_compress = b_compress; 4521 4522 ASSERT(addr >= VDEV_LABEL_START_SIZE && 4523 addr + size < vd->vdev_psize - 4524 VDEV_LABEL_END_SIZE); 4525 4526 /* 4527 * l2arc read. The SCL_L2ARC lock will be 4528 * released by l2arc_read_done(). 4529 * Issue a null zio if the underlying buffer 4530 * was squashed to zero size by compression. 4531 */ 4532 if (b_compress == ZIO_COMPRESS_EMPTY) { 4533 rzio = zio_null(pio, spa, vd, 4534 l2arc_read_done, cb, 4535 zio_flags | ZIO_FLAG_DONT_CACHE | 4536 ZIO_FLAG_CANFAIL | 4537 ZIO_FLAG_DONT_PROPAGATE | 4538 ZIO_FLAG_DONT_RETRY); 4539 } else { 4540 rzio = zio_read_phys(pio, vd, addr, 4541 b_asize, buf->b_data, 4542 ZIO_CHECKSUM_OFF, 4543 l2arc_read_done, cb, priority, 4544 zio_flags | ZIO_FLAG_DONT_CACHE | 4545 ZIO_FLAG_CANFAIL | 4546 ZIO_FLAG_DONT_PROPAGATE | 4547 ZIO_FLAG_DONT_RETRY, B_FALSE); 4548 } 4549 DTRACE_PROBE2(l2arc__read, vdev_t *, vd, 4550 zio_t *, rzio); 4551 ARCSTAT_INCR(arcstat_l2_read_bytes, b_asize); 4552 4553 if (*arc_flags & ARC_FLAG_NOWAIT) { 4554 zio_nowait(rzio); 4555 return (0); 4556 } 4557 4558 ASSERT(*arc_flags & ARC_FLAG_WAIT); 4559 if (zio_wait(rzio) == 0) 4560 return (0); 4561 4562 /* l2arc read error; goto zio_read() */ 4563 } else { 4564 DTRACE_PROBE1(l2arc__miss, 4565 arc_buf_hdr_t *, hdr); 4566 ARCSTAT_BUMP(arcstat_l2_misses); 4567 if (HDR_L2_WRITING(hdr)) 4568 ARCSTAT_BUMP(arcstat_l2_rw_clash); 4569 spa_config_exit(spa, SCL_L2ARC, vd); 4570 } 4571 } else { 4572 if (vd != NULL) 4573 spa_config_exit(spa, SCL_L2ARC, vd); 4574 if (l2arc_ndev != 0) { 4575 DTRACE_PROBE1(l2arc__miss, 4576 arc_buf_hdr_t *, hdr); 4577 ARCSTAT_BUMP(arcstat_l2_misses); 4578 } 4579 } 4580 4581 rzio = zio_read(pio, spa, bp, buf->b_data, size, 4582 arc_read_done, buf, priority, zio_flags, zb); 4583 4584 if (*arc_flags & ARC_FLAG_WAIT) 4585 return (zio_wait(rzio)); 4586 4587 ASSERT(*arc_flags & ARC_FLAG_NOWAIT); 4588 zio_nowait(rzio); 4589 } 4590 return (0); 4591} 4592 4593void 4594arc_set_callback(arc_buf_t *buf, arc_evict_func_t *func, void *private) 4595{ 4596 ASSERT(buf->b_hdr != NULL); 4597 ASSERT(buf->b_hdr->b_l1hdr.b_state != arc_anon); 4598 ASSERT(!refcount_is_zero(&buf->b_hdr->b_l1hdr.b_refcnt) || 4599 func == NULL); 4600 ASSERT(buf->b_efunc == NULL); 4601 ASSERT(!HDR_BUF_AVAILABLE(buf->b_hdr)); 4602 4603 buf->b_efunc = func; 4604 buf->b_private = private; 4605} 4606 4607/* 4608 * Notify the arc that a block was freed, and thus will never be used again. 4609 */ 4610void 4611arc_freed(spa_t *spa, const blkptr_t *bp) 4612{ 4613 arc_buf_hdr_t *hdr; 4614 kmutex_t *hash_lock; 4615 uint64_t guid = spa_load_guid(spa); 4616 4617 ASSERT(!BP_IS_EMBEDDED(bp)); 4618 4619 hdr = buf_hash_find(guid, bp, &hash_lock); 4620 if (hdr == NULL) 4621 return; 4622 if (HDR_BUF_AVAILABLE(hdr)) { 4623 arc_buf_t *buf = hdr->b_l1hdr.b_buf; 4624 add_reference(hdr, hash_lock, FTAG); 4625 hdr->b_flags &= ~ARC_FLAG_BUF_AVAILABLE; 4626 mutex_exit(hash_lock); 4627 4628 arc_release(buf, FTAG); 4629 (void) arc_buf_remove_ref(buf, FTAG); 4630 } else { 4631 mutex_exit(hash_lock); 4632 } 4633 4634} 4635 4636/* 4637 * Clear the user eviction callback set by arc_set_callback(), first calling 4638 * it if it exists. Because the presence of a callback keeps an arc_buf cached 4639 * clearing the callback may result in the arc_buf being destroyed. However, 4640 * it will not result in the *last* arc_buf being destroyed, hence the data 4641 * will remain cached in the ARC. We make a copy of the arc buffer here so 4642 * that we can process the callback without holding any locks. 4643 * 4644 * It's possible that the callback is already in the process of being cleared 4645 * by another thread. In this case we can not clear the callback. 4646 * 4647 * Returns B_TRUE if the callback was successfully called and cleared. 4648 */ 4649boolean_t 4650arc_clear_callback(arc_buf_t *buf) 4651{ 4652 arc_buf_hdr_t *hdr; 4653 kmutex_t *hash_lock; 4654 arc_evict_func_t *efunc = buf->b_efunc; 4655 void *private = buf->b_private; 4656 4657 mutex_enter(&buf->b_evict_lock); 4658 hdr = buf->b_hdr; 4659 if (hdr == NULL) { 4660 /* 4661 * We are in arc_do_user_evicts(). 4662 */ 4663 ASSERT(buf->b_data == NULL); 4664 mutex_exit(&buf->b_evict_lock); 4665 return (B_FALSE); 4666 } else if (buf->b_data == NULL) { 4667 /* 4668 * We are on the eviction list; process this buffer now 4669 * but let arc_do_user_evicts() do the reaping. 4670 */ 4671 buf->b_efunc = NULL; 4672 mutex_exit(&buf->b_evict_lock); 4673 VERIFY0(efunc(private)); 4674 return (B_TRUE); 4675 } 4676 hash_lock = HDR_LOCK(hdr); 4677 mutex_enter(hash_lock); 4678 hdr = buf->b_hdr; 4679 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 4680 4681 ASSERT3U(refcount_count(&hdr->b_l1hdr.b_refcnt), <, 4682 hdr->b_l1hdr.b_datacnt); 4683 ASSERT(hdr->b_l1hdr.b_state == arc_mru || 4684 hdr->b_l1hdr.b_state == arc_mfu); 4685 4686 buf->b_efunc = NULL; 4687 buf->b_private = NULL; 4688 4689 if (hdr->b_l1hdr.b_datacnt > 1) { 4690 mutex_exit(&buf->b_evict_lock); 4691 arc_buf_destroy(buf, TRUE); 4692 } else { 4693 ASSERT(buf == hdr->b_l1hdr.b_buf); 4694 hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE; 4695 mutex_exit(&buf->b_evict_lock); 4696 } 4697 4698 mutex_exit(hash_lock); 4699 VERIFY0(efunc(private)); 4700 return (B_TRUE); 4701} 4702 4703/* 4704 * Release this buffer from the cache, making it an anonymous buffer. This 4705 * must be done after a read and prior to modifying the buffer contents. 4706 * If the buffer has more than one reference, we must make 4707 * a new hdr for the buffer. 4708 */ 4709void 4710arc_release(arc_buf_t *buf, void *tag) 4711{ 4712 arc_buf_hdr_t *hdr = buf->b_hdr; 4713 4714 /* 4715 * It would be nice to assert that if it's DMU metadata (level > 4716 * 0 || it's the dnode file), then it must be syncing context. 4717 * But we don't know that information at this level. 4718 */ 4719 4720 mutex_enter(&buf->b_evict_lock); 4721 4722 ASSERT(HDR_HAS_L1HDR(hdr)); 4723 4724 /* 4725 * We don't grab the hash lock prior to this check, because if 4726 * the buffer's header is in the arc_anon state, it won't be 4727 * linked into the hash table. 4728 */ 4729 if (hdr->b_l1hdr.b_state == arc_anon) { 4730 mutex_exit(&buf->b_evict_lock); 4731 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 4732 ASSERT(!HDR_IN_HASH_TABLE(hdr)); 4733 ASSERT(!HDR_HAS_L2HDR(hdr)); 4734 ASSERT(BUF_EMPTY(hdr)); 4735 ASSERT3U(hdr->b_l1hdr.b_datacnt, ==, 1); 4736 ASSERT3S(refcount_count(&hdr->b_l1hdr.b_refcnt), ==, 1); 4737 ASSERT(!list_link_active(&hdr->b_l1hdr.b_arc_node)); 4738 4739 ASSERT3P(buf->b_efunc, ==, NULL); 4740 ASSERT3P(buf->b_private, ==, NULL); 4741 4742 hdr->b_l1hdr.b_arc_access = 0; 4743 arc_buf_thaw(buf); 4744 4745 return; 4746 } 4747 4748 kmutex_t *hash_lock = HDR_LOCK(hdr); 4749 mutex_enter(hash_lock); 4750 4751 /* 4752 * This assignment is only valid as long as the hash_lock is 4753 * held, we must be careful not to reference state or the 4754 * b_state field after dropping the lock. 4755 */ 4756 arc_state_t *state = hdr->b_l1hdr.b_state; 4757 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 4758 ASSERT3P(state, !=, arc_anon); 4759 4760 /* this buffer is not on any list */ 4761 ASSERT(refcount_count(&hdr->b_l1hdr.b_refcnt) > 0); 4762 4763 if (HDR_HAS_L2HDR(hdr)) { 4764 mutex_enter(&hdr->b_l2hdr.b_dev->l2ad_mtx); 4765 4766 /* 4767 * We have to recheck this conditional again now that 4768 * we're holding the l2ad_mtx to prevent a race with 4769 * another thread which might be concurrently calling 4770 * l2arc_evict(). In that case, l2arc_evict() might have 4771 * destroyed the header's L2 portion as we were waiting 4772 * to acquire the l2ad_mtx. 4773 */ 4774 if (HDR_HAS_L2HDR(hdr)) { 4775 if (hdr->b_l2hdr.b_daddr != L2ARC_ADDR_UNSET) 4776 trim_map_free(hdr->b_l2hdr.b_dev->l2ad_vdev, 4777 hdr->b_l2hdr.b_daddr, 4778 hdr->b_l2hdr.b_asize, 0); 4779 arc_hdr_l2hdr_destroy(hdr); 4780 } 4781 4782 mutex_exit(&hdr->b_l2hdr.b_dev->l2ad_mtx); 4783 } 4784 4785 /* 4786 * Do we have more than one buf? 4787 */ 4788 if (hdr->b_l1hdr.b_datacnt > 1) { 4789 arc_buf_hdr_t *nhdr; 4790 arc_buf_t **bufp; 4791 uint64_t blksz = hdr->b_size; 4792 uint64_t spa = hdr->b_spa; 4793 arc_buf_contents_t type = arc_buf_type(hdr); 4794 uint32_t flags = hdr->b_flags; 4795 4796 ASSERT(hdr->b_l1hdr.b_buf != buf || buf->b_next != NULL); 4797 /* 4798 * Pull the data off of this hdr and attach it to 4799 * a new anonymous hdr. 4800 */ 4801 (void) remove_reference(hdr, hash_lock, tag); 4802 bufp = &hdr->b_l1hdr.b_buf; 4803 while (*bufp != buf) 4804 bufp = &(*bufp)->b_next; 4805 *bufp = buf->b_next; 4806 buf->b_next = NULL; 4807 4808 ASSERT3P(state, !=, arc_l2c_only); 4809 4810 (void) refcount_remove_many( 4811 &state->arcs_size, hdr->b_size, buf); 4812 4813 if (refcount_is_zero(&hdr->b_l1hdr.b_refcnt)) { 4814 ASSERT3P(state, !=, arc_l2c_only); 4815 uint64_t *size = &state->arcs_lsize[type]; 4816 ASSERT3U(*size, >=, hdr->b_size); 4817 atomic_add_64(size, -hdr->b_size); 4818 } 4819 4820 /* 4821 * We're releasing a duplicate user data buffer, update 4822 * our statistics accordingly. 4823 */ 4824 if (HDR_ISTYPE_DATA(hdr)) { 4825 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers); 4826 ARCSTAT_INCR(arcstat_duplicate_buffers_size, 4827 -hdr->b_size); 4828 } 4829 hdr->b_l1hdr.b_datacnt -= 1; 4830 arc_cksum_verify(buf); 4831#ifdef illumos 4832 arc_buf_unwatch(buf); 4833#endif /* illumos */ 4834 4835 mutex_exit(hash_lock); 4836 4837 nhdr = kmem_cache_alloc(hdr_full_cache, KM_PUSHPAGE); 4838 nhdr->b_size = blksz; 4839 nhdr->b_spa = spa; 4840 4841 nhdr->b_flags = flags & ARC_FLAG_L2_WRITING; 4842 nhdr->b_flags |= arc_bufc_to_flags(type); 4843 nhdr->b_flags |= ARC_FLAG_HAS_L1HDR; 4844 4845 nhdr->b_l1hdr.b_buf = buf; 4846 nhdr->b_l1hdr.b_datacnt = 1; 4847 nhdr->b_l1hdr.b_state = arc_anon; 4848 nhdr->b_l1hdr.b_arc_access = 0; 4849 nhdr->b_l1hdr.b_tmp_cdata = NULL; 4850 nhdr->b_freeze_cksum = NULL; 4851 4852 (void) refcount_add(&nhdr->b_l1hdr.b_refcnt, tag); 4853 buf->b_hdr = nhdr; 4854 mutex_exit(&buf->b_evict_lock); 4855 (void) refcount_add_many(&arc_anon->arcs_size, blksz, buf); 4856 } else { 4857 mutex_exit(&buf->b_evict_lock); 4858 ASSERT(refcount_count(&hdr->b_l1hdr.b_refcnt) == 1); 4859 /* protected by hash lock, or hdr is on arc_anon */ 4860 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 4861 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 4862 arc_change_state(arc_anon, hdr, hash_lock); 4863 hdr->b_l1hdr.b_arc_access = 0; 4864 mutex_exit(hash_lock); 4865 4866 buf_discard_identity(hdr); 4867 arc_buf_thaw(buf); 4868 } 4869 buf->b_efunc = NULL; 4870 buf->b_private = NULL; 4871} 4872 4873int 4874arc_released(arc_buf_t *buf) 4875{ 4876 int released; 4877 4878 mutex_enter(&buf->b_evict_lock); 4879 released = (buf->b_data != NULL && 4880 buf->b_hdr->b_l1hdr.b_state == arc_anon); 4881 mutex_exit(&buf->b_evict_lock); 4882 return (released); 4883} 4884 4885#ifdef ZFS_DEBUG 4886int 4887arc_referenced(arc_buf_t *buf) 4888{ 4889 int referenced; 4890 4891 mutex_enter(&buf->b_evict_lock); 4892 referenced = (refcount_count(&buf->b_hdr->b_l1hdr.b_refcnt)); 4893 mutex_exit(&buf->b_evict_lock); 4894 return (referenced); 4895} 4896#endif 4897 4898static void 4899arc_write_ready(zio_t *zio) 4900{ 4901 arc_write_callback_t *callback = zio->io_private; 4902 arc_buf_t *buf = callback->awcb_buf; 4903 arc_buf_hdr_t *hdr = buf->b_hdr; 4904 4905 ASSERT(HDR_HAS_L1HDR(hdr)); 4906 ASSERT(!refcount_is_zero(&buf->b_hdr->b_l1hdr.b_refcnt)); 4907 ASSERT(hdr->b_l1hdr.b_datacnt > 0); 4908 callback->awcb_ready(zio, buf, callback->awcb_private); 4909 4910 /* 4911 * If the IO is already in progress, then this is a re-write 4912 * attempt, so we need to thaw and re-compute the cksum. 4913 * It is the responsibility of the callback to handle the 4914 * accounting for any re-write attempt. 4915 */ 4916 if (HDR_IO_IN_PROGRESS(hdr)) { 4917 mutex_enter(&hdr->b_l1hdr.b_freeze_lock); 4918 if (hdr->b_freeze_cksum != NULL) { 4919 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 4920 hdr->b_freeze_cksum = NULL; 4921 } 4922 mutex_exit(&hdr->b_l1hdr.b_freeze_lock); 4923 } 4924 arc_cksum_compute(buf, B_FALSE); 4925 hdr->b_flags |= ARC_FLAG_IO_IN_PROGRESS; 4926} 4927 4928/* 4929 * The SPA calls this callback for each physical write that happens on behalf 4930 * of a logical write. See the comment in dbuf_write_physdone() for details. 4931 */ 4932static void 4933arc_write_physdone(zio_t *zio) 4934{ 4935 arc_write_callback_t *cb = zio->io_private; 4936 if (cb->awcb_physdone != NULL) 4937 cb->awcb_physdone(zio, cb->awcb_buf, cb->awcb_private); 4938} 4939 4940static void 4941arc_write_done(zio_t *zio) 4942{ 4943 arc_write_callback_t *callback = zio->io_private; 4944 arc_buf_t *buf = callback->awcb_buf; 4945 arc_buf_hdr_t *hdr = buf->b_hdr; 4946 4947 ASSERT(hdr->b_l1hdr.b_acb == NULL); 4948 4949 if (zio->io_error == 0) { 4950 if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) { 4951 buf_discard_identity(hdr); 4952 } else { 4953 hdr->b_dva = *BP_IDENTITY(zio->io_bp); 4954 hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp); 4955 } 4956 } else { 4957 ASSERT(BUF_EMPTY(hdr)); 4958 } 4959 4960 /* 4961 * If the block to be written was all-zero or compressed enough to be 4962 * embedded in the BP, no write was performed so there will be no 4963 * dva/birth/checksum. The buffer must therefore remain anonymous 4964 * (and uncached). 4965 */ 4966 if (!BUF_EMPTY(hdr)) { 4967 arc_buf_hdr_t *exists; 4968 kmutex_t *hash_lock; 4969 4970 ASSERT(zio->io_error == 0); 4971 4972 arc_cksum_verify(buf); 4973 4974 exists = buf_hash_insert(hdr, &hash_lock); 4975 if (exists != NULL) { 4976 /* 4977 * This can only happen if we overwrite for 4978 * sync-to-convergence, because we remove 4979 * buffers from the hash table when we arc_free(). 4980 */ 4981 if (zio->io_flags & ZIO_FLAG_IO_REWRITE) { 4982 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) 4983 panic("bad overwrite, hdr=%p exists=%p", 4984 (void *)hdr, (void *)exists); 4985 ASSERT(refcount_is_zero( 4986 &exists->b_l1hdr.b_refcnt)); 4987 arc_change_state(arc_anon, exists, hash_lock); 4988 mutex_exit(hash_lock); 4989 arc_hdr_destroy(exists); 4990 exists = buf_hash_insert(hdr, &hash_lock); 4991 ASSERT3P(exists, ==, NULL); 4992 } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) { 4993 /* nopwrite */ 4994 ASSERT(zio->io_prop.zp_nopwrite); 4995 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) 4996 panic("bad nopwrite, hdr=%p exists=%p", 4997 (void *)hdr, (void *)exists); 4998 } else { 4999 /* Dedup */ 5000 ASSERT(hdr->b_l1hdr.b_datacnt == 1); 5001 ASSERT(hdr->b_l1hdr.b_state == arc_anon); 5002 ASSERT(BP_GET_DEDUP(zio->io_bp)); 5003 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0); 5004 } 5005 } 5006 hdr->b_flags &= ~ARC_FLAG_IO_IN_PROGRESS; 5007 /* if it's not anon, we are doing a scrub */ 5008 if (exists == NULL && hdr->b_l1hdr.b_state == arc_anon) 5009 arc_access(hdr, hash_lock); 5010 mutex_exit(hash_lock); 5011 } else { 5012 hdr->b_flags &= ~ARC_FLAG_IO_IN_PROGRESS; 5013 } 5014 5015 ASSERT(!refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 5016 callback->awcb_done(zio, buf, callback->awcb_private); 5017 5018 kmem_free(callback, sizeof (arc_write_callback_t)); 5019} 5020 5021zio_t * 5022arc_write(zio_t *pio, spa_t *spa, uint64_t txg, 5023 blkptr_t *bp, arc_buf_t *buf, boolean_t l2arc, boolean_t l2arc_compress, 5024 const zio_prop_t *zp, arc_done_func_t *ready, arc_done_func_t *physdone, 5025 arc_done_func_t *done, void *private, zio_priority_t priority, 5026 int zio_flags, const zbookmark_phys_t *zb) 5027{ 5028 arc_buf_hdr_t *hdr = buf->b_hdr; 5029 arc_write_callback_t *callback; 5030 zio_t *zio; 5031 5032 ASSERT(ready != NULL); 5033 ASSERT(done != NULL); 5034 ASSERT(!HDR_IO_ERROR(hdr)); 5035 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 5036 ASSERT(hdr->b_l1hdr.b_acb == NULL); 5037 ASSERT(hdr->b_l1hdr.b_datacnt > 0); 5038 if (l2arc) 5039 hdr->b_flags |= ARC_FLAG_L2CACHE; 5040 if (l2arc_compress) 5041 hdr->b_flags |= ARC_FLAG_L2COMPRESS; 5042 callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP); 5043 callback->awcb_ready = ready; 5044 callback->awcb_physdone = physdone; 5045 callback->awcb_done = done; 5046 callback->awcb_private = private; 5047 callback->awcb_buf = buf; 5048 5049 zio = zio_write(pio, spa, txg, bp, buf->b_data, hdr->b_size, zp, 5050 arc_write_ready, arc_write_physdone, arc_write_done, callback, 5051 priority, zio_flags, zb); 5052 5053 return (zio); 5054} 5055 5056static int 5057arc_memory_throttle(uint64_t reserve, uint64_t txg) 5058{ 5059#ifdef _KERNEL 5060 uint64_t available_memory = ptob(freemem); 5061 static uint64_t page_load = 0; 5062 static uint64_t last_txg = 0; 5063 5064#if defined(__i386) || !defined(UMA_MD_SMALL_ALLOC) 5065 available_memory = 5066 MIN(available_memory, ptob(vmem_size(heap_arena, VMEM_FREE))); 5067#endif 5068 5069 if (freemem > (uint64_t)physmem * arc_lotsfree_percent / 100) 5070 return (0); 5071 5072 if (txg > last_txg) { 5073 last_txg = txg; 5074 page_load = 0; 5075 } 5076 /* 5077 * If we are in pageout, we know that memory is already tight, 5078 * the arc is already going to be evicting, so we just want to 5079 * continue to let page writes occur as quickly as possible. 5080 */ 5081 if (curproc == pageproc) { 5082 if (page_load > MAX(ptob(minfree), available_memory) / 4) 5083 return (SET_ERROR(ERESTART)); 5084 /* Note: reserve is inflated, so we deflate */ 5085 page_load += reserve / 8; 5086 return (0); 5087 } else if (page_load > 0 && arc_reclaim_needed()) { 5088 /* memory is low, delay before restarting */ 5089 ARCSTAT_INCR(arcstat_memory_throttle_count, 1); 5090 return (SET_ERROR(EAGAIN)); 5091 } 5092 page_load = 0; 5093#endif 5094 return (0); 5095} 5096 5097void 5098arc_tempreserve_clear(uint64_t reserve) 5099{ 5100 atomic_add_64(&arc_tempreserve, -reserve); 5101 ASSERT((int64_t)arc_tempreserve >= 0); 5102} 5103 5104int 5105arc_tempreserve_space(uint64_t reserve, uint64_t txg) 5106{ 5107 int error; 5108 uint64_t anon_size; 5109 5110 if (reserve > arc_c/4 && !arc_no_grow) { 5111 arc_c = MIN(arc_c_max, reserve * 4); 5112 DTRACE_PROBE1(arc__set_reserve, uint64_t, arc_c); 5113 } 5114 if (reserve > arc_c) 5115 return (SET_ERROR(ENOMEM)); 5116 5117 /* 5118 * Don't count loaned bufs as in flight dirty data to prevent long 5119 * network delays from blocking transactions that are ready to be 5120 * assigned to a txg. 5121 */ 5122 anon_size = MAX((int64_t)(refcount_count(&arc_anon->arcs_size) - 5123 arc_loaned_bytes), 0); 5124 5125 /* 5126 * Writes will, almost always, require additional memory allocations 5127 * in order to compress/encrypt/etc the data. We therefore need to 5128 * make sure that there is sufficient available memory for this. 5129 */ 5130 error = arc_memory_throttle(reserve, txg); 5131 if (error != 0) 5132 return (error); 5133 5134 /* 5135 * Throttle writes when the amount of dirty data in the cache 5136 * gets too large. We try to keep the cache less than half full 5137 * of dirty blocks so that our sync times don't grow too large. 5138 * Note: if two requests come in concurrently, we might let them 5139 * both succeed, when one of them should fail. Not a huge deal. 5140 */ 5141 5142 if (reserve + arc_tempreserve + anon_size > arc_c / 2 && 5143 anon_size > arc_c / 4) { 5144 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK " 5145 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n", 5146 arc_tempreserve>>10, 5147 arc_anon->arcs_lsize[ARC_BUFC_METADATA]>>10, 5148 arc_anon->arcs_lsize[ARC_BUFC_DATA]>>10, 5149 reserve>>10, arc_c>>10); 5150 return (SET_ERROR(ERESTART)); 5151 } 5152 atomic_add_64(&arc_tempreserve, reserve); 5153 return (0); 5154} 5155 5156static void 5157arc_kstat_update_state(arc_state_t *state, kstat_named_t *size, 5158 kstat_named_t *evict_data, kstat_named_t *evict_metadata) 5159{ 5160 size->value.ui64 = refcount_count(&state->arcs_size); 5161 evict_data->value.ui64 = state->arcs_lsize[ARC_BUFC_DATA]; 5162 evict_metadata->value.ui64 = state->arcs_lsize[ARC_BUFC_METADATA]; 5163} 5164 5165static int 5166arc_kstat_update(kstat_t *ksp, int rw) 5167{ 5168 arc_stats_t *as = ksp->ks_data; 5169 5170 if (rw == KSTAT_WRITE) { 5171 return (EACCES); 5172 } else { 5173 arc_kstat_update_state(arc_anon, 5174 &as->arcstat_anon_size, 5175 &as->arcstat_anon_evictable_data, 5176 &as->arcstat_anon_evictable_metadata); 5177 arc_kstat_update_state(arc_mru, 5178 &as->arcstat_mru_size, 5179 &as->arcstat_mru_evictable_data, 5180 &as->arcstat_mru_evictable_metadata); 5181 arc_kstat_update_state(arc_mru_ghost, 5182 &as->arcstat_mru_ghost_size, 5183 &as->arcstat_mru_ghost_evictable_data, 5184 &as->arcstat_mru_ghost_evictable_metadata); 5185 arc_kstat_update_state(arc_mfu, 5186 &as->arcstat_mfu_size, 5187 &as->arcstat_mfu_evictable_data, 5188 &as->arcstat_mfu_evictable_metadata); 5189 arc_kstat_update_state(arc_mfu_ghost, 5190 &as->arcstat_mfu_ghost_size, 5191 &as->arcstat_mfu_ghost_evictable_data, 5192 &as->arcstat_mfu_ghost_evictable_metadata); 5193 } 5194 5195 return (0); 5196} 5197 5198/* 5199 * This function *must* return indices evenly distributed between all 5200 * sublists of the multilist. This is needed due to how the ARC eviction 5201 * code is laid out; arc_evict_state() assumes ARC buffers are evenly 5202 * distributed between all sublists and uses this assumption when 5203 * deciding which sublist to evict from and how much to evict from it. 5204 */ 5205unsigned int 5206arc_state_multilist_index_func(multilist_t *ml, void *obj) 5207{ 5208 arc_buf_hdr_t *hdr = obj; 5209 5210 /* 5211 * We rely on b_dva to generate evenly distributed index 5212 * numbers using buf_hash below. So, as an added precaution, 5213 * let's make sure we never add empty buffers to the arc lists. 5214 */ 5215 ASSERT(!BUF_EMPTY(hdr)); 5216 5217 /* 5218 * The assumption here, is the hash value for a given 5219 * arc_buf_hdr_t will remain constant throughout it's lifetime 5220 * (i.e. it's b_spa, b_dva, and b_birth fields don't change). 5221 * Thus, we don't need to store the header's sublist index 5222 * on insertion, as this index can be recalculated on removal. 5223 * 5224 * Also, the low order bits of the hash value are thought to be 5225 * distributed evenly. Otherwise, in the case that the multilist 5226 * has a power of two number of sublists, each sublists' usage 5227 * would not be evenly distributed. 5228 */ 5229 return (buf_hash(hdr->b_spa, &hdr->b_dva, hdr->b_birth) % 5230 multilist_get_num_sublists(ml)); 5231} 5232 5233#ifdef _KERNEL 5234static eventhandler_tag arc_event_lowmem = NULL; 5235 5236static void 5237arc_lowmem(void *arg __unused, int howto __unused) 5238{ 5239 5240 mutex_enter(&arc_reclaim_lock); 5241 /* XXX: Memory deficit should be passed as argument. */ 5242 needfree = btoc(arc_c >> arc_shrink_shift); 5243 DTRACE_PROBE(arc__needfree); 5244 cv_signal(&arc_reclaim_thread_cv); 5245 5246 /* 5247 * It is unsafe to block here in arbitrary threads, because we can come 5248 * here from ARC itself and may hold ARC locks and thus risk a deadlock 5249 * with ARC reclaim thread. 5250 */ 5251 if (curproc == pageproc) 5252 (void) cv_wait(&arc_reclaim_waiters_cv, &arc_reclaim_lock); 5253 mutex_exit(&arc_reclaim_lock); 5254} 5255#endif 5256 5257void 5258arc_init(void) 5259{ 5260 int i, prefetch_tunable_set = 0; 5261 5262 mutex_init(&arc_reclaim_lock, NULL, MUTEX_DEFAULT, NULL); 5263 cv_init(&arc_reclaim_thread_cv, NULL, CV_DEFAULT, NULL); 5264 cv_init(&arc_reclaim_waiters_cv, NULL, CV_DEFAULT, NULL); 5265 5266 mutex_init(&arc_user_evicts_lock, NULL, MUTEX_DEFAULT, NULL); 5267 cv_init(&arc_user_evicts_cv, NULL, CV_DEFAULT, NULL); 5268 5269 /* Convert seconds to clock ticks */ 5270 arc_min_prefetch_lifespan = 1 * hz; 5271 5272 /* Start out with 1/8 of all memory */ 5273 arc_c = kmem_size() / 8; 5274 5275#ifdef sun 5276#ifdef _KERNEL 5277 /* 5278 * On architectures where the physical memory can be larger 5279 * than the addressable space (intel in 32-bit mode), we may 5280 * need to limit the cache to 1/8 of VM size. 5281 */ 5282 arc_c = MIN(arc_c, vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 8); 5283#endif 5284#endif /* sun */ 5285 /* set min cache to 1/32 of all memory, or 16MB, whichever is more */ 5286 arc_c_min = MAX(arc_c / 4, 16 << 20); 5287 /* set max to 1/2 of all memory, or all but 1GB, whichever is more */ 5288 if (arc_c * 8 >= 1 << 30) 5289 arc_c_max = (arc_c * 8) - (1 << 30); 5290 else 5291 arc_c_max = arc_c_min; 5292 arc_c_max = MAX(arc_c * 5, arc_c_max); 5293 5294#ifdef _KERNEL 5295 /* 5296 * Allow the tunables to override our calculations if they are 5297 * reasonable (ie. over 16MB) 5298 */ 5299 if (zfs_arc_max > 16 << 20 && zfs_arc_max < kmem_size()) 5300 arc_c_max = zfs_arc_max; 5301 if (zfs_arc_min > 16 << 20 && zfs_arc_min <= arc_c_max) 5302 arc_c_min = zfs_arc_min; 5303#endif 5304 5305 arc_c = arc_c_max; 5306 arc_p = (arc_c >> 1); 5307 5308 /* limit meta-data to 1/4 of the arc capacity */ 5309 arc_meta_limit = arc_c_max / 4; 5310 5311 /* Allow the tunable to override if it is reasonable */ 5312 if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max) 5313 arc_meta_limit = zfs_arc_meta_limit; 5314 5315 if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0) 5316 arc_c_min = arc_meta_limit / 2; 5317 5318 if (zfs_arc_meta_min > 0) { 5319 arc_meta_min = zfs_arc_meta_min; 5320 } else { 5321 arc_meta_min = arc_c_min / 2; 5322 } 5323 5324 if (zfs_arc_grow_retry > 0) 5325 arc_grow_retry = zfs_arc_grow_retry; 5326 5327 if (zfs_arc_shrink_shift > 0) 5328 arc_shrink_shift = zfs_arc_shrink_shift; 5329 5330 /* 5331 * Ensure that arc_no_grow_shift is less than arc_shrink_shift. 5332 */ 5333 if (arc_no_grow_shift >= arc_shrink_shift) 5334 arc_no_grow_shift = arc_shrink_shift - 1; 5335 5336 if (zfs_arc_p_min_shift > 0) 5337 arc_p_min_shift = zfs_arc_p_min_shift; 5338 5339 if (zfs_arc_num_sublists_per_state < 1) 5340 zfs_arc_num_sublists_per_state = MAX(max_ncpus, 1); 5341 5342 /* if kmem_flags are set, lets try to use less memory */ 5343 if (kmem_debugging()) 5344 arc_c = arc_c / 2; 5345 if (arc_c < arc_c_min) 5346 arc_c = arc_c_min; 5347 5348 zfs_arc_min = arc_c_min; 5349 zfs_arc_max = arc_c_max; 5350 5351 arc_anon = &ARC_anon; 5352 arc_mru = &ARC_mru; 5353 arc_mru_ghost = &ARC_mru_ghost; 5354 arc_mfu = &ARC_mfu; 5355 arc_mfu_ghost = &ARC_mfu_ghost; 5356 arc_l2c_only = &ARC_l2c_only; 5357 arc_size = 0; 5358 5359 multilist_create(&arc_mru->arcs_list[ARC_BUFC_METADATA], 5360 sizeof (arc_buf_hdr_t), 5361 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5362 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5363 multilist_create(&arc_mru->arcs_list[ARC_BUFC_DATA], 5364 sizeof (arc_buf_hdr_t), 5365 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5366 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5367 multilist_create(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA], 5368 sizeof (arc_buf_hdr_t), 5369 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5370 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5371 multilist_create(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA], 5372 sizeof (arc_buf_hdr_t), 5373 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5374 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5375 multilist_create(&arc_mfu->arcs_list[ARC_BUFC_METADATA], 5376 sizeof (arc_buf_hdr_t), 5377 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5378 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5379 multilist_create(&arc_mfu->arcs_list[ARC_BUFC_DATA], 5380 sizeof (arc_buf_hdr_t), 5381 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5382 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5383 multilist_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA], 5384 sizeof (arc_buf_hdr_t), 5385 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5386 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5387 multilist_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA], 5388 sizeof (arc_buf_hdr_t), 5389 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5390 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5391 multilist_create(&arc_l2c_only->arcs_list[ARC_BUFC_METADATA], 5392 sizeof (arc_buf_hdr_t), 5393 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5394 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5395 multilist_create(&arc_l2c_only->arcs_list[ARC_BUFC_DATA], 5396 sizeof (arc_buf_hdr_t), 5397 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5398 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5399 5400 refcount_create(&arc_anon->arcs_size); 5401 refcount_create(&arc_mru->arcs_size); 5402 refcount_create(&arc_mru_ghost->arcs_size); 5403 refcount_create(&arc_mfu->arcs_size); 5404 refcount_create(&arc_mfu_ghost->arcs_size); 5405 refcount_create(&arc_l2c_only->arcs_size); 5406 5407 buf_init(); 5408 5409 arc_reclaim_thread_exit = FALSE; 5410 arc_user_evicts_thread_exit = FALSE; 5411 arc_eviction_list = NULL; 5412 bzero(&arc_eviction_hdr, sizeof (arc_buf_hdr_t)); 5413 5414 arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED, 5415 sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL); 5416 5417 if (arc_ksp != NULL) { 5418 arc_ksp->ks_data = &arc_stats; 5419 arc_ksp->ks_update = arc_kstat_update; 5420 kstat_install(arc_ksp); 5421 } 5422 5423 (void) thread_create(NULL, 0, arc_reclaim_thread, NULL, 0, &p0, 5424 TS_RUN, minclsyspri); 5425 5426#ifdef _KERNEL 5427 arc_event_lowmem = EVENTHANDLER_REGISTER(vm_lowmem, arc_lowmem, NULL, 5428 EVENTHANDLER_PRI_FIRST); 5429#endif 5430 5431 (void) thread_create(NULL, 0, arc_user_evicts_thread, NULL, 0, &p0, 5432 TS_RUN, minclsyspri); 5433 5434 arc_dead = FALSE; 5435 arc_warm = B_FALSE; 5436 5437 /* 5438 * Calculate maximum amount of dirty data per pool. 5439 * 5440 * If it has been set by /etc/system, take that. 5441 * Otherwise, use a percentage of physical memory defined by 5442 * zfs_dirty_data_max_percent (default 10%) with a cap at 5443 * zfs_dirty_data_max_max (default 4GB). 5444 */ 5445 if (zfs_dirty_data_max == 0) { 5446 zfs_dirty_data_max = ptob(physmem) * 5447 zfs_dirty_data_max_percent / 100; 5448 zfs_dirty_data_max = MIN(zfs_dirty_data_max, 5449 zfs_dirty_data_max_max); 5450 } 5451 5452#ifdef _KERNEL 5453 if (TUNABLE_INT_FETCH("vfs.zfs.prefetch_disable", &zfs_prefetch_disable)) 5454 prefetch_tunable_set = 1; 5455 5456#ifdef __i386__ 5457 if (prefetch_tunable_set == 0) { 5458 printf("ZFS NOTICE: Prefetch is disabled by default on i386 " 5459 "-- to enable,\n"); 5460 printf(" add \"vfs.zfs.prefetch_disable=0\" " 5461 "to /boot/loader.conf.\n"); 5462 zfs_prefetch_disable = 1; 5463 } 5464#else 5465 if ((((uint64_t)physmem * PAGESIZE) < (1ULL << 32)) && 5466 prefetch_tunable_set == 0) { 5467 printf("ZFS NOTICE: Prefetch is disabled by default if less " 5468 "than 4GB of RAM is present;\n" 5469 " to enable, add \"vfs.zfs.prefetch_disable=0\" " 5470 "to /boot/loader.conf.\n"); 5471 zfs_prefetch_disable = 1; 5472 } 5473#endif 5474 /* Warn about ZFS memory and address space requirements. */ 5475 if (((uint64_t)physmem * PAGESIZE) < (256 + 128 + 64) * (1 << 20)) { 5476 printf("ZFS WARNING: Recommended minimum RAM size is 512MB; " 5477 "expect unstable behavior.\n"); 5478 } 5479 if (kmem_size() < 512 * (1 << 20)) { 5480 printf("ZFS WARNING: Recommended minimum kmem_size is 512MB; " 5481 "expect unstable behavior.\n"); 5482 printf(" Consider tuning vm.kmem_size and " 5483 "vm.kmem_size_max\n"); 5484 printf(" in /boot/loader.conf.\n"); 5485 } 5486#endif 5487} 5488 5489void 5490arc_fini(void) 5491{ 5492 mutex_enter(&arc_reclaim_lock); 5493 arc_reclaim_thread_exit = TRUE; 5494 /* 5495 * The reclaim thread will set arc_reclaim_thread_exit back to 5496 * FALSE when it is finished exiting; we're waiting for that. 5497 */ 5498 while (arc_reclaim_thread_exit) { 5499 cv_signal(&arc_reclaim_thread_cv); 5500 cv_wait(&arc_reclaim_thread_cv, &arc_reclaim_lock); 5501 } 5502 mutex_exit(&arc_reclaim_lock); 5503 5504 mutex_enter(&arc_user_evicts_lock); 5505 arc_user_evicts_thread_exit = TRUE; 5506 /* 5507 * The user evicts thread will set arc_user_evicts_thread_exit 5508 * to FALSE when it is finished exiting; we're waiting for that. 5509 */ 5510 while (arc_user_evicts_thread_exit) { 5511 cv_signal(&arc_user_evicts_cv); 5512 cv_wait(&arc_user_evicts_cv, &arc_user_evicts_lock); 5513 } 5514 mutex_exit(&arc_user_evicts_lock); 5515 5516 /* Use TRUE to ensure *all* buffers are evicted */ 5517 arc_flush(NULL, TRUE); 5518 5519 arc_dead = TRUE; 5520 5521 if (arc_ksp != NULL) { 5522 kstat_delete(arc_ksp); 5523 arc_ksp = NULL; 5524 } 5525 5526 mutex_destroy(&arc_reclaim_lock); 5527 cv_destroy(&arc_reclaim_thread_cv); 5528 cv_destroy(&arc_reclaim_waiters_cv); 5529 5530 mutex_destroy(&arc_user_evicts_lock); 5531 cv_destroy(&arc_user_evicts_cv); 5532 5533 refcount_destroy(&arc_anon->arcs_size); 5534 refcount_destroy(&arc_mru->arcs_size); 5535 refcount_destroy(&arc_mru_ghost->arcs_size); 5536 refcount_destroy(&arc_mfu->arcs_size); 5537 refcount_destroy(&arc_mfu_ghost->arcs_size); 5538 refcount_destroy(&arc_l2c_only->arcs_size); 5539 5540 multilist_destroy(&arc_mru->arcs_list[ARC_BUFC_METADATA]); 5541 multilist_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA]); 5542 multilist_destroy(&arc_mfu->arcs_list[ARC_BUFC_METADATA]); 5543 multilist_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA]); 5544 multilist_destroy(&arc_mru->arcs_list[ARC_BUFC_DATA]); 5545 multilist_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA]); 5546 multilist_destroy(&arc_mfu->arcs_list[ARC_BUFC_DATA]); 5547 multilist_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA]); 5548 5549 buf_fini(); 5550 5551 ASSERT0(arc_loaned_bytes); 5552 5553#ifdef _KERNEL 5554 if (arc_event_lowmem != NULL) 5555 EVENTHANDLER_DEREGISTER(vm_lowmem, arc_event_lowmem); 5556#endif 5557} 5558 5559/* 5560 * Level 2 ARC 5561 * 5562 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk. 5563 * It uses dedicated storage devices to hold cached data, which are populated 5564 * using large infrequent writes. The main role of this cache is to boost 5565 * the performance of random read workloads. The intended L2ARC devices 5566 * include short-stroked disks, solid state disks, and other media with 5567 * substantially faster read latency than disk. 5568 * 5569 * +-----------------------+ 5570 * | ARC | 5571 * +-----------------------+ 5572 * | ^ ^ 5573 * | | | 5574 * l2arc_feed_thread() arc_read() 5575 * | | | 5576 * | l2arc read | 5577 * V | | 5578 * +---------------+ | 5579 * | L2ARC | | 5580 * +---------------+ | 5581 * | ^ | 5582 * l2arc_write() | | 5583 * | | | 5584 * V | | 5585 * +-------+ +-------+ 5586 * | vdev | | vdev | 5587 * | cache | | cache | 5588 * +-------+ +-------+ 5589 * +=========+ .-----. 5590 * : L2ARC : |-_____-| 5591 * : devices : | Disks | 5592 * +=========+ `-_____-' 5593 * 5594 * Read requests are satisfied from the following sources, in order: 5595 * 5596 * 1) ARC 5597 * 2) vdev cache of L2ARC devices 5598 * 3) L2ARC devices 5599 * 4) vdev cache of disks 5600 * 5) disks 5601 * 5602 * Some L2ARC device types exhibit extremely slow write performance. 5603 * To accommodate for this there are some significant differences between 5604 * the L2ARC and traditional cache design: 5605 * 5606 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from 5607 * the ARC behave as usual, freeing buffers and placing headers on ghost 5608 * lists. The ARC does not send buffers to the L2ARC during eviction as 5609 * this would add inflated write latencies for all ARC memory pressure. 5610 * 5611 * 2. The L2ARC attempts to cache data from the ARC before it is evicted. 5612 * It does this by periodically scanning buffers from the eviction-end of 5613 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are 5614 * not already there. It scans until a headroom of buffers is satisfied, 5615 * which itself is a buffer for ARC eviction. If a compressible buffer is 5616 * found during scanning and selected for writing to an L2ARC device, we 5617 * temporarily boost scanning headroom during the next scan cycle to make 5618 * sure we adapt to compression effects (which might significantly reduce 5619 * the data volume we write to L2ARC). The thread that does this is 5620 * l2arc_feed_thread(), illustrated below; example sizes are included to 5621 * provide a better sense of ratio than this diagram: 5622 * 5623 * head --> tail 5624 * +---------------------+----------+ 5625 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC 5626 * +---------------------+----------+ | o L2ARC eligible 5627 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer 5628 * +---------------------+----------+ | 5629 * 15.9 Gbytes ^ 32 Mbytes | 5630 * headroom | 5631 * l2arc_feed_thread() 5632 * | 5633 * l2arc write hand <--[oooo]--' 5634 * | 8 Mbyte 5635 * | write max 5636 * V 5637 * +==============================+ 5638 * L2ARC dev |####|#|###|###| |####| ... | 5639 * +==============================+ 5640 * 32 Gbytes 5641 * 5642 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of 5643 * evicted, then the L2ARC has cached a buffer much sooner than it probably 5644 * needed to, potentially wasting L2ARC device bandwidth and storage. It is 5645 * safe to say that this is an uncommon case, since buffers at the end of 5646 * the ARC lists have moved there due to inactivity. 5647 * 5648 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom, 5649 * then the L2ARC simply misses copying some buffers. This serves as a 5650 * pressure valve to prevent heavy read workloads from both stalling the ARC 5651 * with waits and clogging the L2ARC with writes. This also helps prevent 5652 * the potential for the L2ARC to churn if it attempts to cache content too 5653 * quickly, such as during backups of the entire pool. 5654 * 5655 * 5. After system boot and before the ARC has filled main memory, there are 5656 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru 5657 * lists can remain mostly static. Instead of searching from tail of these 5658 * lists as pictured, the l2arc_feed_thread() will search from the list heads 5659 * for eligible buffers, greatly increasing its chance of finding them. 5660 * 5661 * The L2ARC device write speed is also boosted during this time so that 5662 * the L2ARC warms up faster. Since there have been no ARC evictions yet, 5663 * there are no L2ARC reads, and no fear of degrading read performance 5664 * through increased writes. 5665 * 5666 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that 5667 * the vdev queue can aggregate them into larger and fewer writes. Each 5668 * device is written to in a rotor fashion, sweeping writes through 5669 * available space then repeating. 5670 * 5671 * 7. The L2ARC does not store dirty content. It never needs to flush 5672 * write buffers back to disk based storage. 5673 * 5674 * 8. If an ARC buffer is written (and dirtied) which also exists in the 5675 * L2ARC, the now stale L2ARC buffer is immediately dropped. 5676 * 5677 * The performance of the L2ARC can be tweaked by a number of tunables, which 5678 * may be necessary for different workloads: 5679 * 5680 * l2arc_write_max max write bytes per interval 5681 * l2arc_write_boost extra write bytes during device warmup 5682 * l2arc_noprefetch skip caching prefetched buffers 5683 * l2arc_headroom number of max device writes to precache 5684 * l2arc_headroom_boost when we find compressed buffers during ARC 5685 * scanning, we multiply headroom by this 5686 * percentage factor for the next scan cycle, 5687 * since more compressed buffers are likely to 5688 * be present 5689 * l2arc_feed_secs seconds between L2ARC writing 5690 * 5691 * Tunables may be removed or added as future performance improvements are 5692 * integrated, and also may become zpool properties. 5693 * 5694 * There are three key functions that control how the L2ARC warms up: 5695 * 5696 * l2arc_write_eligible() check if a buffer is eligible to cache 5697 * l2arc_write_size() calculate how much to write 5698 * l2arc_write_interval() calculate sleep delay between writes 5699 * 5700 * These three functions determine what to write, how much, and how quickly 5701 * to send writes. 5702 */ 5703 5704static boolean_t 5705l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *hdr) 5706{ 5707 /* 5708 * A buffer is *not* eligible for the L2ARC if it: 5709 * 1. belongs to a different spa. 5710 * 2. is already cached on the L2ARC. 5711 * 3. has an I/O in progress (it may be an incomplete read). 5712 * 4. is flagged not eligible (zfs property). 5713 */ 5714 if (hdr->b_spa != spa_guid) { 5715 ARCSTAT_BUMP(arcstat_l2_write_spa_mismatch); 5716 return (B_FALSE); 5717 } 5718 if (HDR_HAS_L2HDR(hdr)) { 5719 ARCSTAT_BUMP(arcstat_l2_write_in_l2); 5720 return (B_FALSE); 5721 } 5722 if (HDR_IO_IN_PROGRESS(hdr)) { 5723 ARCSTAT_BUMP(arcstat_l2_write_hdr_io_in_progress); 5724 return (B_FALSE); 5725 } 5726 if (!HDR_L2CACHE(hdr)) { 5727 ARCSTAT_BUMP(arcstat_l2_write_not_cacheable); 5728 return (B_FALSE); 5729 } 5730 5731 return (B_TRUE); 5732} 5733 5734static uint64_t 5735l2arc_write_size(void) 5736{ 5737 uint64_t size; 5738 5739 /* 5740 * Make sure our globals have meaningful values in case the user 5741 * altered them. 5742 */ 5743 size = l2arc_write_max; 5744 if (size == 0) { 5745 cmn_err(CE_NOTE, "Bad value for l2arc_write_max, value must " 5746 "be greater than zero, resetting it to the default (%d)", 5747 L2ARC_WRITE_SIZE); 5748 size = l2arc_write_max = L2ARC_WRITE_SIZE; 5749 } 5750 5751 if (arc_warm == B_FALSE) 5752 size += l2arc_write_boost; 5753 5754 return (size); 5755 5756} 5757 5758static clock_t 5759l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote) 5760{ 5761 clock_t interval, next, now; 5762 5763 /* 5764 * If the ARC lists are busy, increase our write rate; if the 5765 * lists are stale, idle back. This is achieved by checking 5766 * how much we previously wrote - if it was more than half of 5767 * what we wanted, schedule the next write much sooner. 5768 */ 5769 if (l2arc_feed_again && wrote > (wanted / 2)) 5770 interval = (hz * l2arc_feed_min_ms) / 1000; 5771 else 5772 interval = hz * l2arc_feed_secs; 5773 5774 now = ddi_get_lbolt(); 5775 next = MAX(now, MIN(now + interval, began + interval)); 5776 5777 return (next); 5778} 5779 5780/* 5781 * Cycle through L2ARC devices. This is how L2ARC load balances. 5782 * If a device is returned, this also returns holding the spa config lock. 5783 */ 5784static l2arc_dev_t * 5785l2arc_dev_get_next(void) 5786{ 5787 l2arc_dev_t *first, *next = NULL; 5788 5789 /* 5790 * Lock out the removal of spas (spa_namespace_lock), then removal 5791 * of cache devices (l2arc_dev_mtx). Once a device has been selected, 5792 * both locks will be dropped and a spa config lock held instead. 5793 */ 5794 mutex_enter(&spa_namespace_lock); 5795 mutex_enter(&l2arc_dev_mtx); 5796 5797 /* if there are no vdevs, there is nothing to do */ 5798 if (l2arc_ndev == 0) 5799 goto out; 5800 5801 first = NULL; 5802 next = l2arc_dev_last; 5803 do { 5804 /* loop around the list looking for a non-faulted vdev */ 5805 if (next == NULL) { 5806 next = list_head(l2arc_dev_list); 5807 } else { 5808 next = list_next(l2arc_dev_list, next); 5809 if (next == NULL) 5810 next = list_head(l2arc_dev_list); 5811 } 5812 5813 /* if we have come back to the start, bail out */ 5814 if (first == NULL) 5815 first = next; 5816 else if (next == first) 5817 break; 5818 5819 } while (vdev_is_dead(next->l2ad_vdev)); 5820 5821 /* if we were unable to find any usable vdevs, return NULL */ 5822 if (vdev_is_dead(next->l2ad_vdev)) 5823 next = NULL; 5824 5825 l2arc_dev_last = next; 5826 5827out: 5828 mutex_exit(&l2arc_dev_mtx); 5829 5830 /* 5831 * Grab the config lock to prevent the 'next' device from being 5832 * removed while we are writing to it. 5833 */ 5834 if (next != NULL) 5835 spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER); 5836 mutex_exit(&spa_namespace_lock); 5837 5838 return (next); 5839} 5840 5841/* 5842 * Free buffers that were tagged for destruction. 5843 */ 5844static void 5845l2arc_do_free_on_write() 5846{ 5847 list_t *buflist; 5848 l2arc_data_free_t *df, *df_prev; 5849 5850 mutex_enter(&l2arc_free_on_write_mtx); 5851 buflist = l2arc_free_on_write; 5852 5853 for (df = list_tail(buflist); df; df = df_prev) { 5854 df_prev = list_prev(buflist, df); 5855 ASSERT(df->l2df_data != NULL); 5856 ASSERT(df->l2df_func != NULL); 5857 df->l2df_func(df->l2df_data, df->l2df_size); 5858 list_remove(buflist, df); 5859 kmem_free(df, sizeof (l2arc_data_free_t)); 5860 } 5861 5862 mutex_exit(&l2arc_free_on_write_mtx); 5863} 5864 5865/* 5866 * A write to a cache device has completed. Update all headers to allow 5867 * reads from these buffers to begin. 5868 */ 5869static void 5870l2arc_write_done(zio_t *zio) 5871{ 5872 l2arc_write_callback_t *cb; 5873 l2arc_dev_t *dev; 5874 list_t *buflist; 5875 arc_buf_hdr_t *head, *hdr, *hdr_prev; 5876 kmutex_t *hash_lock; 5877 int64_t bytes_dropped = 0; 5878 5879 cb = zio->io_private; 5880 ASSERT(cb != NULL); 5881 dev = cb->l2wcb_dev; 5882 ASSERT(dev != NULL); 5883 head = cb->l2wcb_head; 5884 ASSERT(head != NULL); 5885 buflist = &dev->l2ad_buflist; 5886 ASSERT(buflist != NULL); 5887 DTRACE_PROBE2(l2arc__iodone, zio_t *, zio, 5888 l2arc_write_callback_t *, cb); 5889 5890 if (zio->io_error != 0) 5891 ARCSTAT_BUMP(arcstat_l2_writes_error); 5892 5893 /* 5894 * All writes completed, or an error was hit. 5895 */ 5896top: 5897 mutex_enter(&dev->l2ad_mtx); 5898 for (hdr = list_prev(buflist, head); hdr; hdr = hdr_prev) { 5899 hdr_prev = list_prev(buflist, hdr); 5900 5901 hash_lock = HDR_LOCK(hdr); 5902 5903 /* 5904 * We cannot use mutex_enter or else we can deadlock 5905 * with l2arc_write_buffers (due to swapping the order 5906 * the hash lock and l2ad_mtx are taken). 5907 */ 5908 if (!mutex_tryenter(hash_lock)) { 5909 /* 5910 * Missed the hash lock. We must retry so we 5911 * don't leave the ARC_FLAG_L2_WRITING bit set. 5912 */ 5913 ARCSTAT_BUMP(arcstat_l2_writes_lock_retry); 5914 5915 /* 5916 * We don't want to rescan the headers we've 5917 * already marked as having been written out, so 5918 * we reinsert the head node so we can pick up 5919 * where we left off. 5920 */ 5921 list_remove(buflist, head); 5922 list_insert_after(buflist, hdr, head); 5923 5924 mutex_exit(&dev->l2ad_mtx); 5925 5926 /* 5927 * We wait for the hash lock to become available 5928 * to try and prevent busy waiting, and increase 5929 * the chance we'll be able to acquire the lock 5930 * the next time around. 5931 */ 5932 mutex_enter(hash_lock); 5933 mutex_exit(hash_lock); 5934 goto top; 5935 } 5936 5937 /* 5938 * We could not have been moved into the arc_l2c_only 5939 * state while in-flight due to our ARC_FLAG_L2_WRITING 5940 * bit being set. Let's just ensure that's being enforced. 5941 */ 5942 ASSERT(HDR_HAS_L1HDR(hdr)); 5943 5944 /* 5945 * We may have allocated a buffer for L2ARC compression, 5946 * we must release it to avoid leaking this data. 5947 */ 5948 l2arc_release_cdata_buf(hdr); 5949 5950 if (zio->io_error != 0) { 5951 /* 5952 * Error - drop L2ARC entry. 5953 */ 5954 list_remove(buflist, hdr); 5955 trim_map_free(hdr->b_l2hdr.b_dev->l2ad_vdev, 5956 hdr->b_l2hdr.b_daddr, hdr->b_l2hdr.b_asize, 0); 5957 hdr->b_flags &= ~ARC_FLAG_HAS_L2HDR; 5958 5959 ARCSTAT_INCR(arcstat_l2_asize, -hdr->b_l2hdr.b_asize); 5960 ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size); 5961 5962 bytes_dropped += hdr->b_l2hdr.b_asize; 5963 (void) refcount_remove_many(&dev->l2ad_alloc, 5964 hdr->b_l2hdr.b_asize, hdr); 5965 } 5966 5967 /* 5968 * Allow ARC to begin reads and ghost list evictions to 5969 * this L2ARC entry. 5970 */ 5971 hdr->b_flags &= ~ARC_FLAG_L2_WRITING; 5972 5973 mutex_exit(hash_lock); 5974 } 5975 5976 atomic_inc_64(&l2arc_writes_done); 5977 list_remove(buflist, head); 5978 ASSERT(!HDR_HAS_L1HDR(head)); 5979 kmem_cache_free(hdr_l2only_cache, head); 5980 mutex_exit(&dev->l2ad_mtx); 5981 5982 vdev_space_update(dev->l2ad_vdev, -bytes_dropped, 0, 0); 5983 5984 l2arc_do_free_on_write(); 5985 5986 kmem_free(cb, sizeof (l2arc_write_callback_t)); 5987} 5988 5989/* 5990 * A read to a cache device completed. Validate buffer contents before 5991 * handing over to the regular ARC routines. 5992 */ 5993static void 5994l2arc_read_done(zio_t *zio) 5995{ 5996 l2arc_read_callback_t *cb; 5997 arc_buf_hdr_t *hdr; 5998 arc_buf_t *buf; 5999 kmutex_t *hash_lock; 6000 int equal; 6001 6002 ASSERT(zio->io_vd != NULL); 6003 ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE); 6004 6005 spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd); 6006 6007 cb = zio->io_private; 6008 ASSERT(cb != NULL); 6009 buf = cb->l2rcb_buf; 6010 ASSERT(buf != NULL); 6011 6012 hash_lock = HDR_LOCK(buf->b_hdr); 6013 mutex_enter(hash_lock); 6014 hdr = buf->b_hdr; 6015 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 6016 6017 /* 6018 * If the buffer was compressed, decompress it first. 6019 */ 6020 if (cb->l2rcb_compress != ZIO_COMPRESS_OFF) 6021 l2arc_decompress_zio(zio, hdr, cb->l2rcb_compress); 6022 ASSERT(zio->io_data != NULL); 6023 ASSERT3U(zio->io_size, ==, hdr->b_size); 6024 ASSERT3U(BP_GET_LSIZE(&cb->l2rcb_bp), ==, hdr->b_size); 6025 6026 /* 6027 * Check this survived the L2ARC journey. 6028 */ 6029 equal = arc_cksum_equal(buf); 6030 if (equal && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) { 6031 mutex_exit(hash_lock); 6032 zio->io_private = buf; 6033 zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */ 6034 zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */ 6035 arc_read_done(zio); 6036 } else { 6037 mutex_exit(hash_lock); 6038 /* 6039 * Buffer didn't survive caching. Increment stats and 6040 * reissue to the original storage device. 6041 */ 6042 if (zio->io_error != 0) { 6043 ARCSTAT_BUMP(arcstat_l2_io_error); 6044 } else { 6045 zio->io_error = SET_ERROR(EIO); 6046 } 6047 if (!equal) 6048 ARCSTAT_BUMP(arcstat_l2_cksum_bad); 6049 6050 /* 6051 * If there's no waiter, issue an async i/o to the primary 6052 * storage now. If there *is* a waiter, the caller must 6053 * issue the i/o in a context where it's OK to block. 6054 */ 6055 if (zio->io_waiter == NULL) { 6056 zio_t *pio = zio_unique_parent(zio); 6057 6058 ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL); 6059 6060 zio_nowait(zio_read(pio, cb->l2rcb_spa, &cb->l2rcb_bp, 6061 buf->b_data, hdr->b_size, arc_read_done, buf, 6062 zio->io_priority, cb->l2rcb_flags, &cb->l2rcb_zb)); 6063 } 6064 } 6065 6066 kmem_free(cb, sizeof (l2arc_read_callback_t)); 6067} 6068 6069/* 6070 * This is the list priority from which the L2ARC will search for pages to 6071 * cache. This is used within loops (0..3) to cycle through lists in the 6072 * desired order. This order can have a significant effect on cache 6073 * performance. 6074 * 6075 * Currently the metadata lists are hit first, MFU then MRU, followed by 6076 * the data lists. This function returns a locked list, and also returns 6077 * the lock pointer. 6078 */ 6079static multilist_sublist_t * 6080l2arc_sublist_lock(int list_num) 6081{ 6082 multilist_t *ml = NULL; 6083 unsigned int idx; 6084 6085 ASSERT(list_num >= 0 && list_num <= 3); 6086 6087 switch (list_num) { 6088 case 0: 6089 ml = &arc_mfu->arcs_list[ARC_BUFC_METADATA]; 6090 break; 6091 case 1: 6092 ml = &arc_mru->arcs_list[ARC_BUFC_METADATA]; 6093 break; 6094 case 2: 6095 ml = &arc_mfu->arcs_list[ARC_BUFC_DATA]; 6096 break; 6097 case 3: 6098 ml = &arc_mru->arcs_list[ARC_BUFC_DATA]; 6099 break; 6100 } 6101 6102 /* 6103 * Return a randomly-selected sublist. This is acceptable 6104 * because the caller feeds only a little bit of data for each 6105 * call (8MB). Subsequent calls will result in different 6106 * sublists being selected. 6107 */ 6108 idx = multilist_get_random_index(ml); 6109 return (multilist_sublist_lock(ml, idx)); 6110} 6111 6112/* 6113 * Evict buffers from the device write hand to the distance specified in 6114 * bytes. This distance may span populated buffers, it may span nothing. 6115 * This is clearing a region on the L2ARC device ready for writing. 6116 * If the 'all' boolean is set, every buffer is evicted. 6117 */ 6118static void 6119l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all) 6120{ 6121 list_t *buflist; 6122 arc_buf_hdr_t *hdr, *hdr_prev; 6123 kmutex_t *hash_lock; 6124 uint64_t taddr; 6125 6126 buflist = &dev->l2ad_buflist; 6127 6128 if (!all && dev->l2ad_first) { 6129 /* 6130 * This is the first sweep through the device. There is 6131 * nothing to evict. 6132 */ 6133 return; 6134 } 6135 6136 if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) { 6137 /* 6138 * When nearing the end of the device, evict to the end 6139 * before the device write hand jumps to the start. 6140 */ 6141 taddr = dev->l2ad_end; 6142 } else { 6143 taddr = dev->l2ad_hand + distance; 6144 } 6145 DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist, 6146 uint64_t, taddr, boolean_t, all); 6147 6148top: 6149 mutex_enter(&dev->l2ad_mtx); 6150 for (hdr = list_tail(buflist); hdr; hdr = hdr_prev) { 6151 hdr_prev = list_prev(buflist, hdr); 6152 6153 hash_lock = HDR_LOCK(hdr); 6154 6155 /* 6156 * We cannot use mutex_enter or else we can deadlock 6157 * with l2arc_write_buffers (due to swapping the order 6158 * the hash lock and l2ad_mtx are taken). 6159 */ 6160 if (!mutex_tryenter(hash_lock)) { 6161 /* 6162 * Missed the hash lock. Retry. 6163 */ 6164 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry); 6165 mutex_exit(&dev->l2ad_mtx); 6166 mutex_enter(hash_lock); 6167 mutex_exit(hash_lock); 6168 goto top; 6169 } 6170 6171 if (HDR_L2_WRITE_HEAD(hdr)) { 6172 /* 6173 * We hit a write head node. Leave it for 6174 * l2arc_write_done(). 6175 */ 6176 list_remove(buflist, hdr); 6177 mutex_exit(hash_lock); 6178 continue; 6179 } 6180 6181 if (!all && HDR_HAS_L2HDR(hdr) && 6182 (hdr->b_l2hdr.b_daddr > taddr || 6183 hdr->b_l2hdr.b_daddr < dev->l2ad_hand)) { 6184 /* 6185 * We've evicted to the target address, 6186 * or the end of the device. 6187 */ 6188 mutex_exit(hash_lock); 6189 break; 6190 } 6191 6192 ASSERT(HDR_HAS_L2HDR(hdr)); 6193 if (!HDR_HAS_L1HDR(hdr)) { 6194 ASSERT(!HDR_L2_READING(hdr)); 6195 /* 6196 * This doesn't exist in the ARC. Destroy. 6197 * arc_hdr_destroy() will call list_remove() 6198 * and decrement arcstat_l2_size. 6199 */ 6200 arc_change_state(arc_anon, hdr, hash_lock); 6201 arc_hdr_destroy(hdr); 6202 } else { 6203 ASSERT(hdr->b_l1hdr.b_state != arc_l2c_only); 6204 ARCSTAT_BUMP(arcstat_l2_evict_l1cached); 6205 /* 6206 * Invalidate issued or about to be issued 6207 * reads, since we may be about to write 6208 * over this location. 6209 */ 6210 if (HDR_L2_READING(hdr)) { 6211 ARCSTAT_BUMP(arcstat_l2_evict_reading); 6212 hdr->b_flags |= ARC_FLAG_L2_EVICTED; 6213 } 6214 6215 /* Ensure this header has finished being written */ 6216 ASSERT(!HDR_L2_WRITING(hdr)); 6217 ASSERT3P(hdr->b_l1hdr.b_tmp_cdata, ==, NULL); 6218 6219 arc_hdr_l2hdr_destroy(hdr); 6220 } 6221 mutex_exit(hash_lock); 6222 } 6223 mutex_exit(&dev->l2ad_mtx); 6224} 6225 6226/* 6227 * Find and write ARC buffers to the L2ARC device. 6228 * 6229 * An ARC_FLAG_L2_WRITING flag is set so that the L2ARC buffers are not valid 6230 * for reading until they have completed writing. 6231 * The headroom_boost is an in-out parameter used to maintain headroom boost 6232 * state between calls to this function. 6233 * 6234 * Returns the number of bytes actually written (which may be smaller than 6235 * the delta by which the device hand has changed due to alignment). 6236 */ 6237static uint64_t 6238l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz, 6239 boolean_t *headroom_boost) 6240{ 6241 arc_buf_hdr_t *hdr, *hdr_prev, *head; 6242 uint64_t write_asize, write_sz, headroom, buf_compress_minsz; 6243 void *buf_data; 6244 boolean_t full; 6245 l2arc_write_callback_t *cb; 6246 zio_t *pio, *wzio; 6247 uint64_t guid = spa_load_guid(spa); 6248 const boolean_t do_headroom_boost = *headroom_boost; 6249 int try; 6250 6251 ASSERT(dev->l2ad_vdev != NULL); 6252 6253 /* Lower the flag now, we might want to raise it again later. */ 6254 *headroom_boost = B_FALSE; 6255 6256 pio = NULL; 6257 write_sz = write_asize = 0; 6258 full = B_FALSE; 6259 head = kmem_cache_alloc(hdr_l2only_cache, KM_PUSHPAGE); 6260 head->b_flags |= ARC_FLAG_L2_WRITE_HEAD; 6261 head->b_flags |= ARC_FLAG_HAS_L2HDR; 6262 6263 ARCSTAT_BUMP(arcstat_l2_write_buffer_iter); 6264 /* 6265 * We will want to try to compress buffers that are at least 2x the 6266 * device sector size. 6267 */ 6268 buf_compress_minsz = 2 << dev->l2ad_vdev->vdev_ashift; 6269 6270 /* 6271 * Copy buffers for L2ARC writing. 6272 */ 6273 for (try = 0; try <= 3; try++) { 6274 multilist_sublist_t *mls = l2arc_sublist_lock(try); 6275 uint64_t passed_sz = 0; 6276 6277 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_iter); 6278 6279 /* 6280 * L2ARC fast warmup. 6281 * 6282 * Until the ARC is warm and starts to evict, read from the 6283 * head of the ARC lists rather than the tail. 6284 */ 6285 if (arc_warm == B_FALSE) 6286 hdr = multilist_sublist_head(mls); 6287 else 6288 hdr = multilist_sublist_tail(mls); 6289 if (hdr == NULL) 6290 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_null_iter); 6291 6292 headroom = target_sz * l2arc_headroom; 6293 if (do_headroom_boost) 6294 headroom = (headroom * l2arc_headroom_boost) / 100; 6295 6296 for (; hdr; hdr = hdr_prev) { 6297 kmutex_t *hash_lock; 6298 uint64_t buf_sz; 6299 uint64_t buf_a_sz; 6300 6301 if (arc_warm == B_FALSE) 6302 hdr_prev = multilist_sublist_next(mls, hdr); 6303 else 6304 hdr_prev = multilist_sublist_prev(mls, hdr); 6305 ARCSTAT_INCR(arcstat_l2_write_buffer_bytes_scanned, hdr->b_size); 6306 6307 hash_lock = HDR_LOCK(hdr); 6308 if (!mutex_tryenter(hash_lock)) { 6309 ARCSTAT_BUMP(arcstat_l2_write_trylock_fail); 6310 /* 6311 * Skip this buffer rather than waiting. 6312 */ 6313 continue; 6314 } 6315 6316 passed_sz += hdr->b_size; 6317 if (passed_sz > headroom) { 6318 /* 6319 * Searched too far. 6320 */ 6321 mutex_exit(hash_lock); 6322 ARCSTAT_BUMP(arcstat_l2_write_passed_headroom); 6323 break; 6324 } 6325 6326 if (!l2arc_write_eligible(guid, hdr)) { 6327 mutex_exit(hash_lock); 6328 continue; 6329 } 6330 6331 /* 6332 * Assume that the buffer is not going to be compressed 6333 * and could take more space on disk because of a larger 6334 * disk block size. 6335 */ 6336 buf_sz = hdr->b_size; 6337 buf_a_sz = vdev_psize_to_asize(dev->l2ad_vdev, buf_sz); 6338 6339 if ((write_asize + buf_a_sz) > target_sz) { 6340 full = B_TRUE; 6341 mutex_exit(hash_lock); 6342 ARCSTAT_BUMP(arcstat_l2_write_full); 6343 break; 6344 } 6345 6346 if (pio == NULL) { 6347 /* 6348 * Insert a dummy header on the buflist so 6349 * l2arc_write_done() can find where the 6350 * write buffers begin without searching. 6351 */ 6352 mutex_enter(&dev->l2ad_mtx); 6353 list_insert_head(&dev->l2ad_buflist, head); 6354 mutex_exit(&dev->l2ad_mtx); 6355 6356 cb = kmem_alloc( 6357 sizeof (l2arc_write_callback_t), KM_SLEEP); 6358 cb->l2wcb_dev = dev; 6359 cb->l2wcb_head = head; 6360 pio = zio_root(spa, l2arc_write_done, cb, 6361 ZIO_FLAG_CANFAIL); 6362 ARCSTAT_BUMP(arcstat_l2_write_pios); 6363 } 6364 6365 /* 6366 * Create and add a new L2ARC header. 6367 */ 6368 hdr->b_l2hdr.b_dev = dev; 6369 hdr->b_flags |= ARC_FLAG_L2_WRITING; 6370 /* 6371 * Temporarily stash the data buffer in b_tmp_cdata. 6372 * The subsequent write step will pick it up from 6373 * there. This is because can't access b_l1hdr.b_buf 6374 * without holding the hash_lock, which we in turn 6375 * can't access without holding the ARC list locks 6376 * (which we want to avoid during compression/writing). 6377 */ 6378 hdr->b_l2hdr.b_compress = ZIO_COMPRESS_OFF; 6379 hdr->b_l2hdr.b_asize = hdr->b_size; 6380 hdr->b_l1hdr.b_tmp_cdata = hdr->b_l1hdr.b_buf->b_data; 6381 6382 /* 6383 * Explicitly set the b_daddr field to a known 6384 * value which means "invalid address". This 6385 * enables us to differentiate which stage of 6386 * l2arc_write_buffers() the particular header 6387 * is in (e.g. this loop, or the one below). 6388 * ARC_FLAG_L2_WRITING is not enough to make 6389 * this distinction, and we need to know in 6390 * order to do proper l2arc vdev accounting in 6391 * arc_release() and arc_hdr_destroy(). 6392 * 6393 * Note, we can't use a new flag to distinguish 6394 * the two stages because we don't hold the 6395 * header's hash_lock below, in the second stage 6396 * of this function. Thus, we can't simply 6397 * change the b_flags field to denote that the 6398 * IO has been sent. We can change the b_daddr 6399 * field of the L2 portion, though, since we'll 6400 * be holding the l2ad_mtx; which is why we're 6401 * using it to denote the header's state change. 6402 */ 6403 hdr->b_l2hdr.b_daddr = L2ARC_ADDR_UNSET; 6404 hdr->b_flags |= ARC_FLAG_HAS_L2HDR; 6405 6406 mutex_enter(&dev->l2ad_mtx); 6407 list_insert_head(&dev->l2ad_buflist, hdr); 6408 mutex_exit(&dev->l2ad_mtx); 6409 6410 /* 6411 * Compute and store the buffer cksum before 6412 * writing. On debug the cksum is verified first. 6413 */ 6414 arc_cksum_verify(hdr->b_l1hdr.b_buf); 6415 arc_cksum_compute(hdr->b_l1hdr.b_buf, B_TRUE); 6416 6417 mutex_exit(hash_lock); 6418 6419 write_sz += buf_sz; 6420 write_asize += buf_a_sz; 6421 } 6422 6423 multilist_sublist_unlock(mls); 6424 6425 if (full == B_TRUE) 6426 break; 6427 } 6428 6429 /* No buffers selected for writing? */ 6430 if (pio == NULL) { 6431 ASSERT0(write_sz); 6432 ASSERT(!HDR_HAS_L1HDR(head)); 6433 kmem_cache_free(hdr_l2only_cache, head); 6434 return (0); 6435 } 6436 6437 mutex_enter(&dev->l2ad_mtx); 6438 6439 /* 6440 * Note that elsewhere in this file arcstat_l2_asize 6441 * and the used space on l2ad_vdev are updated using b_asize, 6442 * which is not necessarily rounded up to the device block size. 6443 * Too keep accounting consistent we do the same here as well: 6444 * stats_size accumulates the sum of b_asize of the written buffers, 6445 * while write_asize accumulates the sum of b_asize rounded up 6446 * to the device block size. 6447 * The latter sum is used only to validate the corectness of the code. 6448 */ 6449 uint64_t stats_size = 0; 6450 write_asize = 0; 6451 6452 /* 6453 * Now start writing the buffers. We're starting at the write head 6454 * and work backwards, retracing the course of the buffer selector 6455 * loop above. 6456 */ 6457 for (hdr = list_prev(&dev->l2ad_buflist, head); hdr; 6458 hdr = list_prev(&dev->l2ad_buflist, hdr)) { 6459 uint64_t buf_sz; 6460 6461 /* 6462 * We rely on the L1 portion of the header below, so 6463 * it's invalid for this header to have been evicted out 6464 * of the ghost cache, prior to being written out. The 6465 * ARC_FLAG_L2_WRITING bit ensures this won't happen. 6466 */ 6467 ASSERT(HDR_HAS_L1HDR(hdr)); 6468 6469 /* 6470 * We shouldn't need to lock the buffer here, since we flagged 6471 * it as ARC_FLAG_L2_WRITING in the previous step, but we must 6472 * take care to only access its L2 cache parameters. In 6473 * particular, hdr->l1hdr.b_buf may be invalid by now due to 6474 * ARC eviction. 6475 */ 6476 hdr->b_l2hdr.b_daddr = dev->l2ad_hand; 6477 6478 if ((HDR_L2COMPRESS(hdr)) && 6479 hdr->b_l2hdr.b_asize >= buf_compress_minsz) { 6480 if (l2arc_compress_buf(hdr)) { 6481 /* 6482 * If compression succeeded, enable headroom 6483 * boost on the next scan cycle. 6484 */ 6485 *headroom_boost = B_TRUE; 6486 } 6487 } 6488 6489 /* 6490 * Pick up the buffer data we had previously stashed away 6491 * (and now potentially also compressed). 6492 */ 6493 buf_data = hdr->b_l1hdr.b_tmp_cdata; 6494 buf_sz = hdr->b_l2hdr.b_asize; 6495 6496 /* 6497 * We need to do this regardless if buf_sz is zero or 6498 * not, otherwise, when this l2hdr is evicted we'll 6499 * remove a reference that was never added. 6500 */ 6501 (void) refcount_add_many(&dev->l2ad_alloc, buf_sz, hdr); 6502 6503 /* Compression may have squashed the buffer to zero length. */ 6504 if (buf_sz != 0) { 6505 uint64_t buf_a_sz; 6506 6507 wzio = zio_write_phys(pio, dev->l2ad_vdev, 6508 dev->l2ad_hand, buf_sz, buf_data, ZIO_CHECKSUM_OFF, 6509 NULL, NULL, ZIO_PRIORITY_ASYNC_WRITE, 6510 ZIO_FLAG_CANFAIL, B_FALSE); 6511 6512 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev, 6513 zio_t *, wzio); 6514 (void) zio_nowait(wzio); 6515 6516 stats_size += buf_sz; 6517 6518 /* 6519 * Keep the clock hand suitably device-aligned. 6520 */ 6521 buf_a_sz = vdev_psize_to_asize(dev->l2ad_vdev, buf_sz); 6522 write_asize += buf_a_sz; 6523 dev->l2ad_hand += buf_a_sz; 6524 } 6525 } 6526 6527 mutex_exit(&dev->l2ad_mtx); 6528 6529 ASSERT3U(write_asize, <=, target_sz); 6530 ARCSTAT_BUMP(arcstat_l2_writes_sent); 6531 ARCSTAT_INCR(arcstat_l2_write_bytes, write_asize); 6532 ARCSTAT_INCR(arcstat_l2_size, write_sz); 6533 ARCSTAT_INCR(arcstat_l2_asize, stats_size); 6534 vdev_space_update(dev->l2ad_vdev, stats_size, 0, 0); 6535 6536 /* 6537 * Bump device hand to the device start if it is approaching the end. 6538 * l2arc_evict() will already have evicted ahead for this case. 6539 */ 6540 if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) { 6541 dev->l2ad_hand = dev->l2ad_start; 6542 dev->l2ad_first = B_FALSE; 6543 } 6544 6545 dev->l2ad_writing = B_TRUE; 6546 (void) zio_wait(pio); 6547 dev->l2ad_writing = B_FALSE; 6548 6549 return (write_asize); 6550} 6551 6552/* 6553 * Compresses an L2ARC buffer. 6554 * The data to be compressed must be prefilled in l1hdr.b_tmp_cdata and its 6555 * size in l2hdr->b_asize. This routine tries to compress the data and 6556 * depending on the compression result there are three possible outcomes: 6557 * *) The buffer was incompressible. The original l2hdr contents were left 6558 * untouched and are ready for writing to an L2 device. 6559 * *) The buffer was all-zeros, so there is no need to write it to an L2 6560 * device. To indicate this situation b_tmp_cdata is NULL'ed, b_asize is 6561 * set to zero and b_compress is set to ZIO_COMPRESS_EMPTY. 6562 * *) Compression succeeded and b_tmp_cdata was replaced with a temporary 6563 * data buffer which holds the compressed data to be written, and b_asize 6564 * tells us how much data there is. b_compress is set to the appropriate 6565 * compression algorithm. Once writing is done, invoke 6566 * l2arc_release_cdata_buf on this l2hdr to free this temporary buffer. 6567 * 6568 * Returns B_TRUE if compression succeeded, or B_FALSE if it didn't (the 6569 * buffer was incompressible). 6570 */ 6571static boolean_t 6572l2arc_compress_buf(arc_buf_hdr_t *hdr) 6573{ 6574 void *cdata; 6575 size_t csize, len, rounded; 6576 ASSERT(HDR_HAS_L2HDR(hdr)); 6577 l2arc_buf_hdr_t *l2hdr = &hdr->b_l2hdr; 6578 6579 ASSERT(HDR_HAS_L1HDR(hdr)); 6580 ASSERT3S(l2hdr->b_compress, ==, ZIO_COMPRESS_OFF); 6581 ASSERT(hdr->b_l1hdr.b_tmp_cdata != NULL); 6582 6583 len = l2hdr->b_asize; 6584 cdata = zio_data_buf_alloc(len); 6585 ASSERT3P(cdata, !=, NULL); 6586 csize = zio_compress_data(ZIO_COMPRESS_LZ4, hdr->b_l1hdr.b_tmp_cdata, 6587 cdata, l2hdr->b_asize); 6588 6589 if (csize == 0) { 6590 /* zero block, indicate that there's nothing to write */ 6591 zio_data_buf_free(cdata, len); 6592 l2hdr->b_compress = ZIO_COMPRESS_EMPTY; 6593 l2hdr->b_asize = 0; 6594 hdr->b_l1hdr.b_tmp_cdata = NULL; 6595 ARCSTAT_BUMP(arcstat_l2_compress_zeros); 6596 return (B_TRUE); 6597 } 6598 6599 rounded = P2ROUNDUP(csize, 6600 (size_t)1 << l2hdr->b_dev->l2ad_vdev->vdev_ashift); 6601 if (rounded < len) { 6602 /* 6603 * Compression succeeded, we'll keep the cdata around for 6604 * writing and release it afterwards. 6605 */ 6606 if (rounded > csize) { 6607 bzero((char *)cdata + csize, rounded - csize); 6608 csize = rounded; 6609 } 6610 l2hdr->b_compress = ZIO_COMPRESS_LZ4; 6611 l2hdr->b_asize = csize; 6612 hdr->b_l1hdr.b_tmp_cdata = cdata; 6613 ARCSTAT_BUMP(arcstat_l2_compress_successes); 6614 return (B_TRUE); 6615 } else { 6616 /* 6617 * Compression failed, release the compressed buffer. 6618 * l2hdr will be left unmodified. 6619 */ 6620 zio_data_buf_free(cdata, len); 6621 ARCSTAT_BUMP(arcstat_l2_compress_failures); 6622 return (B_FALSE); 6623 } 6624} 6625 6626/* 6627 * Decompresses a zio read back from an l2arc device. On success, the 6628 * underlying zio's io_data buffer is overwritten by the uncompressed 6629 * version. On decompression error (corrupt compressed stream), the 6630 * zio->io_error value is set to signal an I/O error. 6631 * 6632 * Please note that the compressed data stream is not checksummed, so 6633 * if the underlying device is experiencing data corruption, we may feed 6634 * corrupt data to the decompressor, so the decompressor needs to be 6635 * able to handle this situation (LZ4 does). 6636 */ 6637static void 6638l2arc_decompress_zio(zio_t *zio, arc_buf_hdr_t *hdr, enum zio_compress c) 6639{ 6640 ASSERT(L2ARC_IS_VALID_COMPRESS(c)); 6641 6642 if (zio->io_error != 0) { 6643 /* 6644 * An io error has occured, just restore the original io 6645 * size in preparation for a main pool read. 6646 */ 6647 zio->io_orig_size = zio->io_size = hdr->b_size; 6648 return; 6649 } 6650 6651 if (c == ZIO_COMPRESS_EMPTY) { 6652 /* 6653 * An empty buffer results in a null zio, which means we 6654 * need to fill its io_data after we're done restoring the 6655 * buffer's contents. 6656 */ 6657 ASSERT(hdr->b_l1hdr.b_buf != NULL); 6658 bzero(hdr->b_l1hdr.b_buf->b_data, hdr->b_size); 6659 zio->io_data = zio->io_orig_data = hdr->b_l1hdr.b_buf->b_data; 6660 } else { 6661 ASSERT(zio->io_data != NULL); 6662 /* 6663 * We copy the compressed data from the start of the arc buffer 6664 * (the zio_read will have pulled in only what we need, the 6665 * rest is garbage which we will overwrite at decompression) 6666 * and then decompress back to the ARC data buffer. This way we 6667 * can minimize copying by simply decompressing back over the 6668 * original compressed data (rather than decompressing to an 6669 * aux buffer and then copying back the uncompressed buffer, 6670 * which is likely to be much larger). 6671 */ 6672 uint64_t csize; 6673 void *cdata; 6674 6675 csize = zio->io_size; 6676 cdata = zio_data_buf_alloc(csize); 6677 bcopy(zio->io_data, cdata, csize); 6678 if (zio_decompress_data(c, cdata, zio->io_data, csize, 6679 hdr->b_size) != 0) 6680 zio->io_error = EIO; 6681 zio_data_buf_free(cdata, csize); 6682 } 6683 6684 /* Restore the expected uncompressed IO size. */ 6685 zio->io_orig_size = zio->io_size = hdr->b_size; 6686} 6687 6688/* 6689 * Releases the temporary b_tmp_cdata buffer in an l2arc header structure. 6690 * This buffer serves as a temporary holder of compressed data while 6691 * the buffer entry is being written to an l2arc device. Once that is 6692 * done, we can dispose of it. 6693 */ 6694static void 6695l2arc_release_cdata_buf(arc_buf_hdr_t *hdr) 6696{ 6697 ASSERT(HDR_HAS_L2HDR(hdr)); 6698 enum zio_compress comp = hdr->b_l2hdr.b_compress; 6699 6700 ASSERT(HDR_HAS_L1HDR(hdr)); 6701 ASSERT(comp == ZIO_COMPRESS_OFF || L2ARC_IS_VALID_COMPRESS(comp)); 6702 6703 if (comp == ZIO_COMPRESS_OFF) { 6704 /* 6705 * In this case, b_tmp_cdata points to the same buffer 6706 * as the arc_buf_t's b_data field. We don't want to 6707 * free it, since the arc_buf_t will handle that. 6708 */ 6709 hdr->b_l1hdr.b_tmp_cdata = NULL; 6710 } else if (comp == ZIO_COMPRESS_EMPTY) { 6711 /* 6712 * In this case, b_tmp_cdata was compressed to an empty 6713 * buffer, thus there's nothing to free and b_tmp_cdata 6714 * should have been set to NULL in l2arc_write_buffers(). 6715 */ 6716 ASSERT3P(hdr->b_l1hdr.b_tmp_cdata, ==, NULL); 6717 } else { 6718 /* 6719 * If the data was compressed, then we've allocated a 6720 * temporary buffer for it, so now we need to release it. 6721 */ 6722 ASSERT(hdr->b_l1hdr.b_tmp_cdata != NULL); 6723 zio_data_buf_free(hdr->b_l1hdr.b_tmp_cdata, 6724 hdr->b_size); 6725 hdr->b_l1hdr.b_tmp_cdata = NULL; 6726 } 6727 6728} 6729 6730/* 6731 * This thread feeds the L2ARC at regular intervals. This is the beating 6732 * heart of the L2ARC. 6733 */ 6734static void 6735l2arc_feed_thread(void *dummy __unused) 6736{ 6737 callb_cpr_t cpr; 6738 l2arc_dev_t *dev; 6739 spa_t *spa; 6740 uint64_t size, wrote; 6741 clock_t begin, next = ddi_get_lbolt(); 6742 boolean_t headroom_boost = B_FALSE; 6743 6744 CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG); 6745 6746 mutex_enter(&l2arc_feed_thr_lock); 6747 6748 while (l2arc_thread_exit == 0) { 6749 CALLB_CPR_SAFE_BEGIN(&cpr); 6750 (void) cv_timedwait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock, 6751 next - ddi_get_lbolt()); 6752 CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock); 6753 next = ddi_get_lbolt() + hz; 6754 6755 /* 6756 * Quick check for L2ARC devices. 6757 */ 6758 mutex_enter(&l2arc_dev_mtx); 6759 if (l2arc_ndev == 0) { 6760 mutex_exit(&l2arc_dev_mtx); 6761 continue; 6762 } 6763 mutex_exit(&l2arc_dev_mtx); 6764 begin = ddi_get_lbolt(); 6765 6766 /* 6767 * This selects the next l2arc device to write to, and in 6768 * doing so the next spa to feed from: dev->l2ad_spa. This 6769 * will return NULL if there are now no l2arc devices or if 6770 * they are all faulted. 6771 * 6772 * If a device is returned, its spa's config lock is also 6773 * held to prevent device removal. l2arc_dev_get_next() 6774 * will grab and release l2arc_dev_mtx. 6775 */ 6776 if ((dev = l2arc_dev_get_next()) == NULL) 6777 continue; 6778 6779 spa = dev->l2ad_spa; 6780 ASSERT(spa != NULL); 6781 6782 /* 6783 * If the pool is read-only then force the feed thread to 6784 * sleep a little longer. 6785 */ 6786 if (!spa_writeable(spa)) { 6787 next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz; 6788 spa_config_exit(spa, SCL_L2ARC, dev); 6789 continue; 6790 } 6791 6792 /* 6793 * Avoid contributing to memory pressure. 6794 */ 6795 if (arc_reclaim_needed()) { 6796 ARCSTAT_BUMP(arcstat_l2_abort_lowmem); 6797 spa_config_exit(spa, SCL_L2ARC, dev); 6798 continue; 6799 } 6800 6801 ARCSTAT_BUMP(arcstat_l2_feeds); 6802 6803 size = l2arc_write_size(); 6804 6805 /* 6806 * Evict L2ARC buffers that will be overwritten. 6807 */ 6808 l2arc_evict(dev, size, B_FALSE); 6809 6810 /* 6811 * Write ARC buffers. 6812 */ 6813 wrote = l2arc_write_buffers(spa, dev, size, &headroom_boost); 6814 6815 /* 6816 * Calculate interval between writes. 6817 */ 6818 next = l2arc_write_interval(begin, size, wrote); 6819 spa_config_exit(spa, SCL_L2ARC, dev); 6820 } 6821 6822 l2arc_thread_exit = 0; 6823 cv_broadcast(&l2arc_feed_thr_cv); 6824 CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */ 6825 thread_exit(); 6826} 6827 6828boolean_t 6829l2arc_vdev_present(vdev_t *vd) 6830{ 6831 l2arc_dev_t *dev; 6832 6833 mutex_enter(&l2arc_dev_mtx); 6834 for (dev = list_head(l2arc_dev_list); dev != NULL; 6835 dev = list_next(l2arc_dev_list, dev)) { 6836 if (dev->l2ad_vdev == vd) 6837 break; 6838 } 6839 mutex_exit(&l2arc_dev_mtx); 6840 6841 return (dev != NULL); 6842} 6843 6844/* 6845 * Add a vdev for use by the L2ARC. By this point the spa has already 6846 * validated the vdev and opened it. 6847 */ 6848void 6849l2arc_add_vdev(spa_t *spa, vdev_t *vd) 6850{ 6851 l2arc_dev_t *adddev; 6852 6853 ASSERT(!l2arc_vdev_present(vd)); 6854 6855 vdev_ashift_optimize(vd); 6856 6857 /* 6858 * Create a new l2arc device entry. 6859 */ 6860 adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP); 6861 adddev->l2ad_spa = spa; 6862 adddev->l2ad_vdev = vd; 6863 adddev->l2ad_start = VDEV_LABEL_START_SIZE; 6864 adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd); 6865 adddev->l2ad_hand = adddev->l2ad_start; 6866 adddev->l2ad_first = B_TRUE; 6867 adddev->l2ad_writing = B_FALSE; 6868 6869 mutex_init(&adddev->l2ad_mtx, NULL, MUTEX_DEFAULT, NULL); 6870 /* 6871 * This is a list of all ARC buffers that are still valid on the 6872 * device. 6873 */ 6874 list_create(&adddev->l2ad_buflist, sizeof (arc_buf_hdr_t), 6875 offsetof(arc_buf_hdr_t, b_l2hdr.b_l2node)); 6876 6877 vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand); 6878 refcount_create(&adddev->l2ad_alloc); 6879 6880 /* 6881 * Add device to global list 6882 */ 6883 mutex_enter(&l2arc_dev_mtx); 6884 list_insert_head(l2arc_dev_list, adddev); 6885 atomic_inc_64(&l2arc_ndev); 6886 mutex_exit(&l2arc_dev_mtx); 6887} 6888 6889/* 6890 * Remove a vdev from the L2ARC. 6891 */ 6892void 6893l2arc_remove_vdev(vdev_t *vd) 6894{ 6895 l2arc_dev_t *dev, *nextdev, *remdev = NULL; 6896 6897 /* 6898 * Find the device by vdev 6899 */ 6900 mutex_enter(&l2arc_dev_mtx); 6901 for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) { 6902 nextdev = list_next(l2arc_dev_list, dev); 6903 if (vd == dev->l2ad_vdev) { 6904 remdev = dev; 6905 break; 6906 } 6907 } 6908 ASSERT(remdev != NULL); 6909 6910 /* 6911 * Remove device from global list 6912 */ 6913 list_remove(l2arc_dev_list, remdev); 6914 l2arc_dev_last = NULL; /* may have been invalidated */ 6915 atomic_dec_64(&l2arc_ndev); 6916 mutex_exit(&l2arc_dev_mtx); 6917 6918 /* 6919 * Clear all buflists and ARC references. L2ARC device flush. 6920 */ 6921 l2arc_evict(remdev, 0, B_TRUE); 6922 list_destroy(&remdev->l2ad_buflist); 6923 mutex_destroy(&remdev->l2ad_mtx); 6924 refcount_destroy(&remdev->l2ad_alloc); 6925 kmem_free(remdev, sizeof (l2arc_dev_t)); 6926} 6927 6928void 6929l2arc_init(void) 6930{ 6931 l2arc_thread_exit = 0; 6932 l2arc_ndev = 0; 6933 l2arc_writes_sent = 0; 6934 l2arc_writes_done = 0; 6935 6936 mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL); 6937 cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL); 6938 mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL); 6939 mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL); 6940 6941 l2arc_dev_list = &L2ARC_dev_list; 6942 l2arc_free_on_write = &L2ARC_free_on_write; 6943 list_create(l2arc_dev_list, sizeof (l2arc_dev_t), 6944 offsetof(l2arc_dev_t, l2ad_node)); 6945 list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t), 6946 offsetof(l2arc_data_free_t, l2df_list_node)); 6947} 6948 6949void 6950l2arc_fini(void) 6951{ 6952 /* 6953 * This is called from dmu_fini(), which is called from spa_fini(); 6954 * Because of this, we can assume that all l2arc devices have 6955 * already been removed when the pools themselves were removed. 6956 */ 6957 6958 l2arc_do_free_on_write(); 6959 6960 mutex_destroy(&l2arc_feed_thr_lock); 6961 cv_destroy(&l2arc_feed_thr_cv); 6962 mutex_destroy(&l2arc_dev_mtx); 6963 mutex_destroy(&l2arc_free_on_write_mtx); 6964 6965 list_destroy(l2arc_dev_list); 6966 list_destroy(l2arc_free_on_write); 6967} 6968 6969void 6970l2arc_start(void) 6971{ 6972 if (!(spa_mode_global & FWRITE)) 6973 return; 6974 6975 (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0, 6976 TS_RUN, minclsyspri); 6977} 6978 6979void 6980l2arc_stop(void) 6981{ 6982 if (!(spa_mode_global & FWRITE)) 6983 return; 6984 6985 mutex_enter(&l2arc_feed_thr_lock); 6986 cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */ 6987 l2arc_thread_exit = 1; 6988 while (l2arc_thread_exit != 0) 6989 cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock); 6990 mutex_exit(&l2arc_feed_thr_lock); 6991} 6992