arc.c revision 300039
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_padding_needed; 569 kstat_named_t arcstat_l2_write_trylock_fail; 570 kstat_named_t arcstat_l2_write_passed_headroom; 571 kstat_named_t arcstat_l2_write_spa_mismatch; 572 kstat_named_t arcstat_l2_write_in_l2; 573 kstat_named_t arcstat_l2_write_hdr_io_in_progress; 574 kstat_named_t arcstat_l2_write_not_cacheable; 575 kstat_named_t arcstat_l2_write_full; 576 kstat_named_t arcstat_l2_write_buffer_iter; 577 kstat_named_t arcstat_l2_write_pios; 578 kstat_named_t arcstat_l2_write_buffer_bytes_scanned; 579 kstat_named_t arcstat_l2_write_buffer_list_iter; 580 kstat_named_t arcstat_l2_write_buffer_list_null_iter; 581 kstat_named_t arcstat_memory_throttle_count; 582 kstat_named_t arcstat_duplicate_buffers; 583 kstat_named_t arcstat_duplicate_buffers_size; 584 kstat_named_t arcstat_duplicate_reads; 585 kstat_named_t arcstat_meta_used; 586 kstat_named_t arcstat_meta_limit; 587 kstat_named_t arcstat_meta_max; 588 kstat_named_t arcstat_meta_min; 589 kstat_named_t arcstat_sync_wait_for_async; 590 kstat_named_t arcstat_demand_hit_predictive_prefetch; 591} arc_stats_t; 592 593static arc_stats_t arc_stats = { 594 { "hits", KSTAT_DATA_UINT64 }, 595 { "misses", KSTAT_DATA_UINT64 }, 596 { "demand_data_hits", KSTAT_DATA_UINT64 }, 597 { "demand_data_misses", KSTAT_DATA_UINT64 }, 598 { "demand_metadata_hits", KSTAT_DATA_UINT64 }, 599 { "demand_metadata_misses", KSTAT_DATA_UINT64 }, 600 { "prefetch_data_hits", KSTAT_DATA_UINT64 }, 601 { "prefetch_data_misses", KSTAT_DATA_UINT64 }, 602 { "prefetch_metadata_hits", KSTAT_DATA_UINT64 }, 603 { "prefetch_metadata_misses", KSTAT_DATA_UINT64 }, 604 { "mru_hits", KSTAT_DATA_UINT64 }, 605 { "mru_ghost_hits", KSTAT_DATA_UINT64 }, 606 { "mfu_hits", KSTAT_DATA_UINT64 }, 607 { "mfu_ghost_hits", KSTAT_DATA_UINT64 }, 608 { "allocated", KSTAT_DATA_UINT64 }, 609 { "deleted", KSTAT_DATA_UINT64 }, 610 { "mutex_miss", KSTAT_DATA_UINT64 }, 611 { "evict_skip", KSTAT_DATA_UINT64 }, 612 { "evict_not_enough", KSTAT_DATA_UINT64 }, 613 { "evict_l2_cached", KSTAT_DATA_UINT64 }, 614 { "evict_l2_eligible", KSTAT_DATA_UINT64 }, 615 { "evict_l2_ineligible", KSTAT_DATA_UINT64 }, 616 { "evict_l2_skip", KSTAT_DATA_UINT64 }, 617 { "hash_elements", KSTAT_DATA_UINT64 }, 618 { "hash_elements_max", KSTAT_DATA_UINT64 }, 619 { "hash_collisions", KSTAT_DATA_UINT64 }, 620 { "hash_chains", KSTAT_DATA_UINT64 }, 621 { "hash_chain_max", KSTAT_DATA_UINT64 }, 622 { "p", KSTAT_DATA_UINT64 }, 623 { "c", KSTAT_DATA_UINT64 }, 624 { "c_min", KSTAT_DATA_UINT64 }, 625 { "c_max", KSTAT_DATA_UINT64 }, 626 { "size", KSTAT_DATA_UINT64 }, 627 { "hdr_size", KSTAT_DATA_UINT64 }, 628 { "data_size", KSTAT_DATA_UINT64 }, 629 { "metadata_size", KSTAT_DATA_UINT64 }, 630 { "other_size", KSTAT_DATA_UINT64 }, 631 { "anon_size", KSTAT_DATA_UINT64 }, 632 { "anon_evictable_data", KSTAT_DATA_UINT64 }, 633 { "anon_evictable_metadata", KSTAT_DATA_UINT64 }, 634 { "mru_size", KSTAT_DATA_UINT64 }, 635 { "mru_evictable_data", KSTAT_DATA_UINT64 }, 636 { "mru_evictable_metadata", KSTAT_DATA_UINT64 }, 637 { "mru_ghost_size", KSTAT_DATA_UINT64 }, 638 { "mru_ghost_evictable_data", KSTAT_DATA_UINT64 }, 639 { "mru_ghost_evictable_metadata", KSTAT_DATA_UINT64 }, 640 { "mfu_size", KSTAT_DATA_UINT64 }, 641 { "mfu_evictable_data", KSTAT_DATA_UINT64 }, 642 { "mfu_evictable_metadata", KSTAT_DATA_UINT64 }, 643 { "mfu_ghost_size", KSTAT_DATA_UINT64 }, 644 { "mfu_ghost_evictable_data", KSTAT_DATA_UINT64 }, 645 { "mfu_ghost_evictable_metadata", KSTAT_DATA_UINT64 }, 646 { "l2_hits", KSTAT_DATA_UINT64 }, 647 { "l2_misses", KSTAT_DATA_UINT64 }, 648 { "l2_feeds", KSTAT_DATA_UINT64 }, 649 { "l2_rw_clash", KSTAT_DATA_UINT64 }, 650 { "l2_read_bytes", KSTAT_DATA_UINT64 }, 651 { "l2_write_bytes", KSTAT_DATA_UINT64 }, 652 { "l2_writes_sent", KSTAT_DATA_UINT64 }, 653 { "l2_writes_done", KSTAT_DATA_UINT64 }, 654 { "l2_writes_error", KSTAT_DATA_UINT64 }, 655 { "l2_writes_lock_retry", KSTAT_DATA_UINT64 }, 656 { "l2_evict_lock_retry", KSTAT_DATA_UINT64 }, 657 { "l2_evict_reading", KSTAT_DATA_UINT64 }, 658 { "l2_evict_l1cached", KSTAT_DATA_UINT64 }, 659 { "l2_free_on_write", KSTAT_DATA_UINT64 }, 660 { "l2_cdata_free_on_write", KSTAT_DATA_UINT64 }, 661 { "l2_abort_lowmem", KSTAT_DATA_UINT64 }, 662 { "l2_cksum_bad", KSTAT_DATA_UINT64 }, 663 { "l2_io_error", KSTAT_DATA_UINT64 }, 664 { "l2_size", KSTAT_DATA_UINT64 }, 665 { "l2_asize", KSTAT_DATA_UINT64 }, 666 { "l2_hdr_size", KSTAT_DATA_UINT64 }, 667 { "l2_compress_successes", KSTAT_DATA_UINT64 }, 668 { "l2_compress_zeros", KSTAT_DATA_UINT64 }, 669 { "l2_compress_failures", KSTAT_DATA_UINT64 }, 670 { "l2_padding_needed", KSTAT_DATA_UINT64 }, 671 { "l2_write_trylock_fail", KSTAT_DATA_UINT64 }, 672 { "l2_write_passed_headroom", KSTAT_DATA_UINT64 }, 673 { "l2_write_spa_mismatch", KSTAT_DATA_UINT64 }, 674 { "l2_write_in_l2", KSTAT_DATA_UINT64 }, 675 { "l2_write_io_in_progress", KSTAT_DATA_UINT64 }, 676 { "l2_write_not_cacheable", KSTAT_DATA_UINT64 }, 677 { "l2_write_full", KSTAT_DATA_UINT64 }, 678 { "l2_write_buffer_iter", KSTAT_DATA_UINT64 }, 679 { "l2_write_pios", KSTAT_DATA_UINT64 }, 680 { "l2_write_buffer_bytes_scanned", KSTAT_DATA_UINT64 }, 681 { "l2_write_buffer_list_iter", KSTAT_DATA_UINT64 }, 682 { "l2_write_buffer_list_null_iter", KSTAT_DATA_UINT64 }, 683 { "memory_throttle_count", KSTAT_DATA_UINT64 }, 684 { "duplicate_buffers", KSTAT_DATA_UINT64 }, 685 { "duplicate_buffers_size", KSTAT_DATA_UINT64 }, 686 { "duplicate_reads", KSTAT_DATA_UINT64 }, 687 { "arc_meta_used", KSTAT_DATA_UINT64 }, 688 { "arc_meta_limit", KSTAT_DATA_UINT64 }, 689 { "arc_meta_max", KSTAT_DATA_UINT64 }, 690 { "arc_meta_min", KSTAT_DATA_UINT64 }, 691 { "sync_wait_for_async", KSTAT_DATA_UINT64 }, 692 { "demand_hit_predictive_prefetch", KSTAT_DATA_UINT64 }, 693}; 694 695#define ARCSTAT(stat) (arc_stats.stat.value.ui64) 696 697#define ARCSTAT_INCR(stat, val) \ 698 atomic_add_64(&arc_stats.stat.value.ui64, (val)) 699 700#define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1) 701#define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1) 702 703#define ARCSTAT_MAX(stat, val) { \ 704 uint64_t m; \ 705 while ((val) > (m = arc_stats.stat.value.ui64) && \ 706 (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \ 707 continue; \ 708} 709 710#define ARCSTAT_MAXSTAT(stat) \ 711 ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64) 712 713/* 714 * We define a macro to allow ARC hits/misses to be easily broken down by 715 * two separate conditions, giving a total of four different subtypes for 716 * each of hits and misses (so eight statistics total). 717 */ 718#define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \ 719 if (cond1) { \ 720 if (cond2) { \ 721 ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \ 722 } else { \ 723 ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \ 724 } \ 725 } else { \ 726 if (cond2) { \ 727 ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \ 728 } else { \ 729 ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\ 730 } \ 731 } 732 733kstat_t *arc_ksp; 734static arc_state_t *arc_anon; 735static arc_state_t *arc_mru; 736static arc_state_t *arc_mru_ghost; 737static arc_state_t *arc_mfu; 738static arc_state_t *arc_mfu_ghost; 739static arc_state_t *arc_l2c_only; 740 741/* 742 * There are several ARC variables that are critical to export as kstats -- 743 * but we don't want to have to grovel around in the kstat whenever we wish to 744 * manipulate them. For these variables, we therefore define them to be in 745 * terms of the statistic variable. This assures that we are not introducing 746 * the possibility of inconsistency by having shadow copies of the variables, 747 * while still allowing the code to be readable. 748 */ 749#define arc_size ARCSTAT(arcstat_size) /* actual total arc size */ 750#define arc_p ARCSTAT(arcstat_p) /* target size of MRU */ 751#define arc_c ARCSTAT(arcstat_c) /* target size of cache */ 752#define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */ 753#define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */ 754#define arc_meta_limit ARCSTAT(arcstat_meta_limit) /* max size for metadata */ 755#define arc_meta_min ARCSTAT(arcstat_meta_min) /* min size for metadata */ 756#define arc_meta_used ARCSTAT(arcstat_meta_used) /* size of metadata */ 757#define arc_meta_max ARCSTAT(arcstat_meta_max) /* max size of metadata */ 758 759#define L2ARC_IS_VALID_COMPRESS(_c_) \ 760 ((_c_) == ZIO_COMPRESS_LZ4 || (_c_) == ZIO_COMPRESS_EMPTY) 761 762static int arc_no_grow; /* Don't try to grow cache size */ 763static uint64_t arc_tempreserve; 764static uint64_t arc_loaned_bytes; 765 766typedef struct arc_callback arc_callback_t; 767 768struct arc_callback { 769 void *acb_private; 770 arc_done_func_t *acb_done; 771 arc_buf_t *acb_buf; 772 zio_t *acb_zio_dummy; 773 arc_callback_t *acb_next; 774}; 775 776typedef struct arc_write_callback arc_write_callback_t; 777 778struct arc_write_callback { 779 void *awcb_private; 780 arc_done_func_t *awcb_ready; 781 arc_done_func_t *awcb_physdone; 782 arc_done_func_t *awcb_done; 783 arc_buf_t *awcb_buf; 784}; 785 786/* 787 * ARC buffers are separated into multiple structs as a memory saving measure: 788 * - Common fields struct, always defined, and embedded within it: 789 * - L2-only fields, always allocated but undefined when not in L2ARC 790 * - L1-only fields, only allocated when in L1ARC 791 * 792 * Buffer in L1 Buffer only in L2 793 * +------------------------+ +------------------------+ 794 * | arc_buf_hdr_t | | arc_buf_hdr_t | 795 * | | | | 796 * | | | | 797 * | | | | 798 * +------------------------+ +------------------------+ 799 * | l2arc_buf_hdr_t | | l2arc_buf_hdr_t | 800 * | (undefined if L1-only) | | | 801 * +------------------------+ +------------------------+ 802 * | l1arc_buf_hdr_t | 803 * | | 804 * | | 805 * | | 806 * | | 807 * +------------------------+ 808 * 809 * Because it's possible for the L2ARC to become extremely large, we can wind 810 * up eating a lot of memory in L2ARC buffer headers, so the size of a header 811 * is minimized by only allocating the fields necessary for an L1-cached buffer 812 * when a header is actually in the L1 cache. The sub-headers (l1arc_buf_hdr and 813 * l2arc_buf_hdr) are embedded rather than allocated separately to save a couple 814 * words in pointers. arc_hdr_realloc() is used to switch a header between 815 * these two allocation states. 816 */ 817typedef struct l1arc_buf_hdr { 818 kmutex_t b_freeze_lock; 819#ifdef ZFS_DEBUG 820 /* 821 * used for debugging wtih kmem_flags - by allocating and freeing 822 * b_thawed when the buffer is thawed, we get a record of the stack 823 * trace that thawed it. 824 */ 825 void *b_thawed; 826#endif 827 828 arc_buf_t *b_buf; 829 uint32_t b_datacnt; 830 /* for waiting on writes to complete */ 831 kcondvar_t b_cv; 832 833 /* protected by arc state mutex */ 834 arc_state_t *b_state; 835 multilist_node_t b_arc_node; 836 837 /* updated atomically */ 838 clock_t b_arc_access; 839 840 /* self protecting */ 841 refcount_t b_refcnt; 842 843 arc_callback_t *b_acb; 844 /* temporary buffer holder for in-flight compressed or padded data */ 845 void *b_tmp_cdata; 846} l1arc_buf_hdr_t; 847 848typedef struct l2arc_dev l2arc_dev_t; 849 850typedef struct l2arc_buf_hdr { 851 /* protected by arc_buf_hdr mutex */ 852 l2arc_dev_t *b_dev; /* L2ARC device */ 853 uint64_t b_daddr; /* disk address, offset byte */ 854 /* real alloc'd buffer size depending on b_compress applied */ 855 int32_t b_asize; 856 uint8_t b_compress; 857 858 list_node_t b_l2node; 859} l2arc_buf_hdr_t; 860 861struct arc_buf_hdr { 862 /* protected by hash lock */ 863 dva_t b_dva; 864 uint64_t b_birth; 865 /* 866 * Even though this checksum is only set/verified when a buffer is in 867 * the L1 cache, it needs to be in the set of common fields because it 868 * must be preserved from the time before a buffer is written out to 869 * L2ARC until after it is read back in. 870 */ 871 zio_cksum_t *b_freeze_cksum; 872 873 arc_buf_hdr_t *b_hash_next; 874 arc_flags_t b_flags; 875 876 /* immutable */ 877 int32_t b_size; 878 uint64_t b_spa; 879 880 /* L2ARC fields. Undefined when not in L2ARC. */ 881 l2arc_buf_hdr_t b_l2hdr; 882 /* L1ARC fields. Undefined when in l2arc_only state */ 883 l1arc_buf_hdr_t b_l1hdr; 884}; 885 886#ifdef _KERNEL 887static int 888sysctl_vfs_zfs_arc_meta_limit(SYSCTL_HANDLER_ARGS) 889{ 890 uint64_t val; 891 int err; 892 893 val = arc_meta_limit; 894 err = sysctl_handle_64(oidp, &val, 0, req); 895 if (err != 0 || req->newptr == NULL) 896 return (err); 897 898 if (val <= 0 || val > arc_c_max) 899 return (EINVAL); 900 901 arc_meta_limit = val; 902 return (0); 903} 904#endif 905 906static arc_buf_t *arc_eviction_list; 907static arc_buf_hdr_t arc_eviction_hdr; 908 909#define GHOST_STATE(state) \ 910 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \ 911 (state) == arc_l2c_only) 912 913#define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_FLAG_IN_HASH_TABLE) 914#define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) 915#define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_FLAG_IO_ERROR) 916#define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_FLAG_PREFETCH) 917#define HDR_FREED_IN_READ(hdr) ((hdr)->b_flags & ARC_FLAG_FREED_IN_READ) 918#define HDR_BUF_AVAILABLE(hdr) ((hdr)->b_flags & ARC_FLAG_BUF_AVAILABLE) 919 920#define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_FLAG_L2CACHE) 921#define HDR_L2COMPRESS(hdr) ((hdr)->b_flags & ARC_FLAG_L2COMPRESS) 922#define HDR_L2_READING(hdr) \ 923 (((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) && \ 924 ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR)) 925#define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITING) 926#define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_FLAG_L2_EVICTED) 927#define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITE_HEAD) 928 929#define HDR_ISTYPE_METADATA(hdr) \ 930 ((hdr)->b_flags & ARC_FLAG_BUFC_METADATA) 931#define HDR_ISTYPE_DATA(hdr) (!HDR_ISTYPE_METADATA(hdr)) 932 933#define HDR_HAS_L1HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L1HDR) 934#define HDR_HAS_L2HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR) 935 936/* 937 * Other sizes 938 */ 939 940#define HDR_FULL_SIZE ((int64_t)sizeof (arc_buf_hdr_t)) 941#define HDR_L2ONLY_SIZE ((int64_t)offsetof(arc_buf_hdr_t, b_l1hdr)) 942 943/* 944 * Hash table routines 945 */ 946 947#define HT_LOCK_PAD CACHE_LINE_SIZE 948 949struct ht_lock { 950 kmutex_t ht_lock; 951#ifdef _KERNEL 952 unsigned char pad[(HT_LOCK_PAD - sizeof (kmutex_t))]; 953#endif 954}; 955 956#define BUF_LOCKS 256 957typedef struct buf_hash_table { 958 uint64_t ht_mask; 959 arc_buf_hdr_t **ht_table; 960 struct ht_lock ht_locks[BUF_LOCKS] __aligned(CACHE_LINE_SIZE); 961} buf_hash_table_t; 962 963static buf_hash_table_t buf_hash_table; 964 965#define BUF_HASH_INDEX(spa, dva, birth) \ 966 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask) 967#define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)]) 968#define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock)) 969#define HDR_LOCK(hdr) \ 970 (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth))) 971 972uint64_t zfs_crc64_table[256]; 973 974/* 975 * Level 2 ARC 976 */ 977 978#define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */ 979#define L2ARC_HEADROOM 2 /* num of writes */ 980/* 981 * If we discover during ARC scan any buffers to be compressed, we boost 982 * our headroom for the next scanning cycle by this percentage multiple. 983 */ 984#define L2ARC_HEADROOM_BOOST 200 985#define L2ARC_FEED_SECS 1 /* caching interval secs */ 986#define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */ 987 988/* 989 * Used to distinguish headers that are being process by 990 * l2arc_write_buffers(), but have yet to be assigned to a l2arc disk 991 * address. This can happen when the header is added to the l2arc's list 992 * of buffers to write in the first stage of l2arc_write_buffers(), but 993 * has not yet been written out which happens in the second stage of 994 * l2arc_write_buffers(). 995 */ 996#define L2ARC_ADDR_UNSET ((uint64_t)(-1)) 997 998#define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent) 999#define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done) 1000 1001/* L2ARC Performance Tunables */ 1002uint64_t l2arc_write_max = L2ARC_WRITE_SIZE; /* default max write size */ 1003uint64_t l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra write during warmup */ 1004uint64_t l2arc_headroom = L2ARC_HEADROOM; /* number of dev writes */ 1005uint64_t l2arc_headroom_boost = L2ARC_HEADROOM_BOOST; 1006uint64_t l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */ 1007uint64_t l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval milliseconds */ 1008boolean_t l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */ 1009boolean_t l2arc_feed_again = B_TRUE; /* turbo warmup */ 1010boolean_t l2arc_norw = B_TRUE; /* no reads during writes */ 1011 1012SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_max, CTLFLAG_RW, 1013 &l2arc_write_max, 0, "max write size"); 1014SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_boost, CTLFLAG_RW, 1015 &l2arc_write_boost, 0, "extra write during warmup"); 1016SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_headroom, CTLFLAG_RW, 1017 &l2arc_headroom, 0, "number of dev writes"); 1018SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_secs, CTLFLAG_RW, 1019 &l2arc_feed_secs, 0, "interval seconds"); 1020SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_min_ms, CTLFLAG_RW, 1021 &l2arc_feed_min_ms, 0, "min interval milliseconds"); 1022 1023SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_noprefetch, CTLFLAG_RW, 1024 &l2arc_noprefetch, 0, "don't cache prefetch bufs"); 1025SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_feed_again, CTLFLAG_RW, 1026 &l2arc_feed_again, 0, "turbo warmup"); 1027SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_norw, CTLFLAG_RW, 1028 &l2arc_norw, 0, "no reads during writes"); 1029 1030SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_size, CTLFLAG_RD, 1031 &ARC_anon.arcs_size.rc_count, 0, "size of anonymous state"); 1032SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_metadata_lsize, CTLFLAG_RD, 1033 &ARC_anon.arcs_lsize[ARC_BUFC_METADATA], 0, "size of anonymous state"); 1034SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_data_lsize, CTLFLAG_RD, 1035 &ARC_anon.arcs_lsize[ARC_BUFC_DATA], 0, "size of anonymous state"); 1036 1037SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_size, CTLFLAG_RD, 1038 &ARC_mru.arcs_size.rc_count, 0, "size of mru state"); 1039SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_metadata_lsize, CTLFLAG_RD, 1040 &ARC_mru.arcs_lsize[ARC_BUFC_METADATA], 0, "size of metadata in mru state"); 1041SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_data_lsize, CTLFLAG_RD, 1042 &ARC_mru.arcs_lsize[ARC_BUFC_DATA], 0, "size of data in mru state"); 1043 1044SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_size, CTLFLAG_RD, 1045 &ARC_mru_ghost.arcs_size.rc_count, 0, "size of mru ghost state"); 1046SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_metadata_lsize, CTLFLAG_RD, 1047 &ARC_mru_ghost.arcs_lsize[ARC_BUFC_METADATA], 0, 1048 "size of metadata in mru ghost state"); 1049SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_data_lsize, CTLFLAG_RD, 1050 &ARC_mru_ghost.arcs_lsize[ARC_BUFC_DATA], 0, 1051 "size of data in mru ghost state"); 1052 1053SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_size, CTLFLAG_RD, 1054 &ARC_mfu.arcs_size.rc_count, 0, "size of mfu state"); 1055SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_metadata_lsize, CTLFLAG_RD, 1056 &ARC_mfu.arcs_lsize[ARC_BUFC_METADATA], 0, "size of metadata in mfu state"); 1057SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_data_lsize, CTLFLAG_RD, 1058 &ARC_mfu.arcs_lsize[ARC_BUFC_DATA], 0, "size of data in mfu state"); 1059 1060SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_size, CTLFLAG_RD, 1061 &ARC_mfu_ghost.arcs_size.rc_count, 0, "size of mfu ghost state"); 1062SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_metadata_lsize, CTLFLAG_RD, 1063 &ARC_mfu_ghost.arcs_lsize[ARC_BUFC_METADATA], 0, 1064 "size of metadata in mfu ghost state"); 1065SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_data_lsize, CTLFLAG_RD, 1066 &ARC_mfu_ghost.arcs_lsize[ARC_BUFC_DATA], 0, 1067 "size of data in mfu ghost state"); 1068 1069SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2c_only_size, CTLFLAG_RD, 1070 &ARC_l2c_only.arcs_size.rc_count, 0, "size of mru state"); 1071 1072/* 1073 * L2ARC Internals 1074 */ 1075struct l2arc_dev { 1076 vdev_t *l2ad_vdev; /* vdev */ 1077 spa_t *l2ad_spa; /* spa */ 1078 uint64_t l2ad_hand; /* next write location */ 1079 uint64_t l2ad_start; /* first addr on device */ 1080 uint64_t l2ad_end; /* last addr on device */ 1081 boolean_t l2ad_first; /* first sweep through */ 1082 boolean_t l2ad_writing; /* currently writing */ 1083 kmutex_t l2ad_mtx; /* lock for buffer list */ 1084 list_t l2ad_buflist; /* buffer list */ 1085 list_node_t l2ad_node; /* device list node */ 1086 refcount_t l2ad_alloc; /* allocated bytes */ 1087}; 1088 1089static list_t L2ARC_dev_list; /* device list */ 1090static list_t *l2arc_dev_list; /* device list pointer */ 1091static kmutex_t l2arc_dev_mtx; /* device list mutex */ 1092static l2arc_dev_t *l2arc_dev_last; /* last device used */ 1093static list_t L2ARC_free_on_write; /* free after write buf list */ 1094static list_t *l2arc_free_on_write; /* free after write list ptr */ 1095static kmutex_t l2arc_free_on_write_mtx; /* mutex for list */ 1096static uint64_t l2arc_ndev; /* number of devices */ 1097 1098typedef struct l2arc_read_callback { 1099 arc_buf_t *l2rcb_buf; /* read buffer */ 1100 spa_t *l2rcb_spa; /* spa */ 1101 blkptr_t l2rcb_bp; /* original blkptr */ 1102 zbookmark_phys_t l2rcb_zb; /* original bookmark */ 1103 int l2rcb_flags; /* original flags */ 1104 enum zio_compress l2rcb_compress; /* applied compress */ 1105 void *l2rcb_data; /* temporary buffer */ 1106} l2arc_read_callback_t; 1107 1108typedef struct l2arc_write_callback { 1109 l2arc_dev_t *l2wcb_dev; /* device info */ 1110 arc_buf_hdr_t *l2wcb_head; /* head of write buflist */ 1111} l2arc_write_callback_t; 1112 1113typedef struct l2arc_data_free { 1114 /* protected by l2arc_free_on_write_mtx */ 1115 void *l2df_data; 1116 size_t l2df_size; 1117 void (*l2df_func)(void *, size_t); 1118 list_node_t l2df_list_node; 1119} l2arc_data_free_t; 1120 1121static kmutex_t l2arc_feed_thr_lock; 1122static kcondvar_t l2arc_feed_thr_cv; 1123static uint8_t l2arc_thread_exit; 1124 1125static void arc_get_data_buf(arc_buf_t *); 1126static void arc_access(arc_buf_hdr_t *, kmutex_t *); 1127static boolean_t arc_is_overflowing(); 1128static void arc_buf_watch(arc_buf_t *); 1129 1130static arc_buf_contents_t arc_buf_type(arc_buf_hdr_t *); 1131static uint32_t arc_bufc_to_flags(arc_buf_contents_t); 1132 1133static boolean_t l2arc_write_eligible(uint64_t, arc_buf_hdr_t *); 1134static void l2arc_read_done(zio_t *); 1135 1136static boolean_t l2arc_transform_buf(arc_buf_hdr_t *, boolean_t); 1137static void l2arc_decompress_zio(zio_t *, arc_buf_hdr_t *, enum zio_compress); 1138static void l2arc_release_cdata_buf(arc_buf_hdr_t *); 1139 1140static void 1141l2arc_trim(const arc_buf_hdr_t *hdr) 1142{ 1143 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev; 1144 1145 ASSERT(HDR_HAS_L2HDR(hdr)); 1146 ASSERT(MUTEX_HELD(&dev->l2ad_mtx)); 1147 1148 if (hdr->b_l2hdr.b_daddr == L2ARC_ADDR_UNSET) 1149 return; 1150 if (hdr->b_l2hdr.b_asize != 0) { 1151 trim_map_free(dev->l2ad_vdev, hdr->b_l2hdr.b_daddr, 1152 hdr->b_l2hdr.b_asize, 0); 1153 } else { 1154 ASSERT3U(hdr->b_l2hdr.b_compress, ==, ZIO_COMPRESS_EMPTY); 1155 } 1156} 1157 1158static uint64_t 1159buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth) 1160{ 1161 uint8_t *vdva = (uint8_t *)dva; 1162 uint64_t crc = -1ULL; 1163 int i; 1164 1165 ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY); 1166 1167 for (i = 0; i < sizeof (dva_t); i++) 1168 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ vdva[i]) & 0xFF]; 1169 1170 crc ^= (spa>>8) ^ birth; 1171 1172 return (crc); 1173} 1174 1175#define BUF_EMPTY(buf) \ 1176 ((buf)->b_dva.dva_word[0] == 0 && \ 1177 (buf)->b_dva.dva_word[1] == 0) 1178 1179#define BUF_EQUAL(spa, dva, birth, buf) \ 1180 ((buf)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \ 1181 ((buf)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \ 1182 ((buf)->b_birth == birth) && ((buf)->b_spa == spa) 1183 1184static void 1185buf_discard_identity(arc_buf_hdr_t *hdr) 1186{ 1187 hdr->b_dva.dva_word[0] = 0; 1188 hdr->b_dva.dva_word[1] = 0; 1189 hdr->b_birth = 0; 1190} 1191 1192static arc_buf_hdr_t * 1193buf_hash_find(uint64_t spa, const blkptr_t *bp, kmutex_t **lockp) 1194{ 1195 const dva_t *dva = BP_IDENTITY(bp); 1196 uint64_t birth = BP_PHYSICAL_BIRTH(bp); 1197 uint64_t idx = BUF_HASH_INDEX(spa, dva, birth); 1198 kmutex_t *hash_lock = BUF_HASH_LOCK(idx); 1199 arc_buf_hdr_t *hdr; 1200 1201 mutex_enter(hash_lock); 1202 for (hdr = buf_hash_table.ht_table[idx]; hdr != NULL; 1203 hdr = hdr->b_hash_next) { 1204 if (BUF_EQUAL(spa, dva, birth, hdr)) { 1205 *lockp = hash_lock; 1206 return (hdr); 1207 } 1208 } 1209 mutex_exit(hash_lock); 1210 *lockp = NULL; 1211 return (NULL); 1212} 1213 1214/* 1215 * Insert an entry into the hash table. If there is already an element 1216 * equal to elem in the hash table, then the already existing element 1217 * will be returned and the new element will not be inserted. 1218 * Otherwise returns NULL. 1219 * If lockp == NULL, the caller is assumed to already hold the hash lock. 1220 */ 1221static arc_buf_hdr_t * 1222buf_hash_insert(arc_buf_hdr_t *hdr, kmutex_t **lockp) 1223{ 1224 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth); 1225 kmutex_t *hash_lock = BUF_HASH_LOCK(idx); 1226 arc_buf_hdr_t *fhdr; 1227 uint32_t i; 1228 1229 ASSERT(!DVA_IS_EMPTY(&hdr->b_dva)); 1230 ASSERT(hdr->b_birth != 0); 1231 ASSERT(!HDR_IN_HASH_TABLE(hdr)); 1232 1233 if (lockp != NULL) { 1234 *lockp = hash_lock; 1235 mutex_enter(hash_lock); 1236 } else { 1237 ASSERT(MUTEX_HELD(hash_lock)); 1238 } 1239 1240 for (fhdr = buf_hash_table.ht_table[idx], i = 0; fhdr != NULL; 1241 fhdr = fhdr->b_hash_next, i++) { 1242 if (BUF_EQUAL(hdr->b_spa, &hdr->b_dva, hdr->b_birth, fhdr)) 1243 return (fhdr); 1244 } 1245 1246 hdr->b_hash_next = buf_hash_table.ht_table[idx]; 1247 buf_hash_table.ht_table[idx] = hdr; 1248 hdr->b_flags |= ARC_FLAG_IN_HASH_TABLE; 1249 1250 /* collect some hash table performance data */ 1251 if (i > 0) { 1252 ARCSTAT_BUMP(arcstat_hash_collisions); 1253 if (i == 1) 1254 ARCSTAT_BUMP(arcstat_hash_chains); 1255 1256 ARCSTAT_MAX(arcstat_hash_chain_max, i); 1257 } 1258 1259 ARCSTAT_BUMP(arcstat_hash_elements); 1260 ARCSTAT_MAXSTAT(arcstat_hash_elements); 1261 1262 return (NULL); 1263} 1264 1265static void 1266buf_hash_remove(arc_buf_hdr_t *hdr) 1267{ 1268 arc_buf_hdr_t *fhdr, **hdrp; 1269 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth); 1270 1271 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx))); 1272 ASSERT(HDR_IN_HASH_TABLE(hdr)); 1273 1274 hdrp = &buf_hash_table.ht_table[idx]; 1275 while ((fhdr = *hdrp) != hdr) { 1276 ASSERT(fhdr != NULL); 1277 hdrp = &fhdr->b_hash_next; 1278 } 1279 *hdrp = hdr->b_hash_next; 1280 hdr->b_hash_next = NULL; 1281 hdr->b_flags &= ~ARC_FLAG_IN_HASH_TABLE; 1282 1283 /* collect some hash table performance data */ 1284 ARCSTAT_BUMPDOWN(arcstat_hash_elements); 1285 1286 if (buf_hash_table.ht_table[idx] && 1287 buf_hash_table.ht_table[idx]->b_hash_next == NULL) 1288 ARCSTAT_BUMPDOWN(arcstat_hash_chains); 1289} 1290 1291/* 1292 * Global data structures and functions for the buf kmem cache. 1293 */ 1294static kmem_cache_t *hdr_full_cache; 1295static kmem_cache_t *hdr_l2only_cache; 1296static kmem_cache_t *buf_cache; 1297 1298static void 1299buf_fini(void) 1300{ 1301 int i; 1302 1303 kmem_free(buf_hash_table.ht_table, 1304 (buf_hash_table.ht_mask + 1) * sizeof (void *)); 1305 for (i = 0; i < BUF_LOCKS; i++) 1306 mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock); 1307 kmem_cache_destroy(hdr_full_cache); 1308 kmem_cache_destroy(hdr_l2only_cache); 1309 kmem_cache_destroy(buf_cache); 1310} 1311 1312/* 1313 * Constructor callback - called when the cache is empty 1314 * and a new buf is requested. 1315 */ 1316/* ARGSUSED */ 1317static int 1318hdr_full_cons(void *vbuf, void *unused, int kmflag) 1319{ 1320 arc_buf_hdr_t *hdr = vbuf; 1321 1322 bzero(hdr, HDR_FULL_SIZE); 1323 cv_init(&hdr->b_l1hdr.b_cv, NULL, CV_DEFAULT, NULL); 1324 refcount_create(&hdr->b_l1hdr.b_refcnt); 1325 mutex_init(&hdr->b_l1hdr.b_freeze_lock, NULL, MUTEX_DEFAULT, NULL); 1326 multilist_link_init(&hdr->b_l1hdr.b_arc_node); 1327 arc_space_consume(HDR_FULL_SIZE, ARC_SPACE_HDRS); 1328 1329 return (0); 1330} 1331 1332/* ARGSUSED */ 1333static int 1334hdr_l2only_cons(void *vbuf, void *unused, int kmflag) 1335{ 1336 arc_buf_hdr_t *hdr = vbuf; 1337 1338 bzero(hdr, HDR_L2ONLY_SIZE); 1339 arc_space_consume(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS); 1340 1341 return (0); 1342} 1343 1344/* ARGSUSED */ 1345static int 1346buf_cons(void *vbuf, void *unused, int kmflag) 1347{ 1348 arc_buf_t *buf = vbuf; 1349 1350 bzero(buf, sizeof (arc_buf_t)); 1351 mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL); 1352 arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS); 1353 1354 return (0); 1355} 1356 1357/* 1358 * Destructor callback - called when a cached buf is 1359 * no longer required. 1360 */ 1361/* ARGSUSED */ 1362static void 1363hdr_full_dest(void *vbuf, void *unused) 1364{ 1365 arc_buf_hdr_t *hdr = vbuf; 1366 1367 ASSERT(BUF_EMPTY(hdr)); 1368 cv_destroy(&hdr->b_l1hdr.b_cv); 1369 refcount_destroy(&hdr->b_l1hdr.b_refcnt); 1370 mutex_destroy(&hdr->b_l1hdr.b_freeze_lock); 1371 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 1372 arc_space_return(HDR_FULL_SIZE, ARC_SPACE_HDRS); 1373} 1374 1375/* ARGSUSED */ 1376static void 1377hdr_l2only_dest(void *vbuf, void *unused) 1378{ 1379 arc_buf_hdr_t *hdr = vbuf; 1380 1381 ASSERT(BUF_EMPTY(hdr)); 1382 arc_space_return(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS); 1383} 1384 1385/* ARGSUSED */ 1386static void 1387buf_dest(void *vbuf, void *unused) 1388{ 1389 arc_buf_t *buf = vbuf; 1390 1391 mutex_destroy(&buf->b_evict_lock); 1392 arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS); 1393} 1394 1395/* 1396 * Reclaim callback -- invoked when memory is low. 1397 */ 1398/* ARGSUSED */ 1399static void 1400hdr_recl(void *unused) 1401{ 1402 dprintf("hdr_recl called\n"); 1403 /* 1404 * umem calls the reclaim func when we destroy the buf cache, 1405 * which is after we do arc_fini(). 1406 */ 1407 if (!arc_dead) 1408 cv_signal(&arc_reclaim_thread_cv); 1409} 1410 1411static void 1412buf_init(void) 1413{ 1414 uint64_t *ct; 1415 uint64_t hsize = 1ULL << 12; 1416 int i, j; 1417 1418 /* 1419 * The hash table is big enough to fill all of physical memory 1420 * with an average block size of zfs_arc_average_blocksize (default 8K). 1421 * By default, the table will take up 1422 * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers). 1423 */ 1424 while (hsize * zfs_arc_average_blocksize < (uint64_t)physmem * PAGESIZE) 1425 hsize <<= 1; 1426retry: 1427 buf_hash_table.ht_mask = hsize - 1; 1428 buf_hash_table.ht_table = 1429 kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP); 1430 if (buf_hash_table.ht_table == NULL) { 1431 ASSERT(hsize > (1ULL << 8)); 1432 hsize >>= 1; 1433 goto retry; 1434 } 1435 1436 hdr_full_cache = kmem_cache_create("arc_buf_hdr_t_full", HDR_FULL_SIZE, 1437 0, hdr_full_cons, hdr_full_dest, hdr_recl, NULL, NULL, 0); 1438 hdr_l2only_cache = kmem_cache_create("arc_buf_hdr_t_l2only", 1439 HDR_L2ONLY_SIZE, 0, hdr_l2only_cons, hdr_l2only_dest, hdr_recl, 1440 NULL, NULL, 0); 1441 buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t), 1442 0, buf_cons, buf_dest, NULL, NULL, NULL, 0); 1443 1444 for (i = 0; i < 256; i++) 1445 for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--) 1446 *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY); 1447 1448 for (i = 0; i < BUF_LOCKS; i++) { 1449 mutex_init(&buf_hash_table.ht_locks[i].ht_lock, 1450 NULL, MUTEX_DEFAULT, NULL); 1451 } 1452} 1453 1454/* 1455 * Transition between the two allocation states for the arc_buf_hdr struct. 1456 * The arc_buf_hdr struct can be allocated with (hdr_full_cache) or without 1457 * (hdr_l2only_cache) the fields necessary for the L1 cache - the smaller 1458 * version is used when a cache buffer is only in the L2ARC in order to reduce 1459 * memory usage. 1460 */ 1461static arc_buf_hdr_t * 1462arc_hdr_realloc(arc_buf_hdr_t *hdr, kmem_cache_t *old, kmem_cache_t *new) 1463{ 1464 ASSERT(HDR_HAS_L2HDR(hdr)); 1465 1466 arc_buf_hdr_t *nhdr; 1467 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev; 1468 1469 ASSERT((old == hdr_full_cache && new == hdr_l2only_cache) || 1470 (old == hdr_l2only_cache && new == hdr_full_cache)); 1471 1472 nhdr = kmem_cache_alloc(new, KM_PUSHPAGE); 1473 1474 ASSERT(MUTEX_HELD(HDR_LOCK(hdr))); 1475 buf_hash_remove(hdr); 1476 1477 bcopy(hdr, nhdr, HDR_L2ONLY_SIZE); 1478 1479 if (new == hdr_full_cache) { 1480 nhdr->b_flags |= ARC_FLAG_HAS_L1HDR; 1481 /* 1482 * arc_access and arc_change_state need to be aware that a 1483 * header has just come out of L2ARC, so we set its state to 1484 * l2c_only even though it's about to change. 1485 */ 1486 nhdr->b_l1hdr.b_state = arc_l2c_only; 1487 1488 /* Verify previous threads set to NULL before freeing */ 1489 ASSERT3P(nhdr->b_l1hdr.b_tmp_cdata, ==, NULL); 1490 } else { 1491 ASSERT(hdr->b_l1hdr.b_buf == NULL); 1492 ASSERT0(hdr->b_l1hdr.b_datacnt); 1493 1494 /* 1495 * If we've reached here, We must have been called from 1496 * arc_evict_hdr(), as such we should have already been 1497 * removed from any ghost list we were previously on 1498 * (which protects us from racing with arc_evict_state), 1499 * thus no locking is needed during this check. 1500 */ 1501 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 1502 1503 /* 1504 * A buffer must not be moved into the arc_l2c_only 1505 * state if it's not finished being written out to the 1506 * l2arc device. Otherwise, the b_l1hdr.b_tmp_cdata field 1507 * might try to be accessed, even though it was removed. 1508 */ 1509 VERIFY(!HDR_L2_WRITING(hdr)); 1510 VERIFY3P(hdr->b_l1hdr.b_tmp_cdata, ==, NULL); 1511 1512#ifdef ZFS_DEBUG 1513 if (hdr->b_l1hdr.b_thawed != NULL) { 1514 kmem_free(hdr->b_l1hdr.b_thawed, 1); 1515 hdr->b_l1hdr.b_thawed = NULL; 1516 } 1517#endif 1518 1519 nhdr->b_flags &= ~ARC_FLAG_HAS_L1HDR; 1520 } 1521 /* 1522 * The header has been reallocated so we need to re-insert it into any 1523 * lists it was on. 1524 */ 1525 (void) buf_hash_insert(nhdr, NULL); 1526 1527 ASSERT(list_link_active(&hdr->b_l2hdr.b_l2node)); 1528 1529 mutex_enter(&dev->l2ad_mtx); 1530 1531 /* 1532 * We must place the realloc'ed header back into the list at 1533 * the same spot. Otherwise, if it's placed earlier in the list, 1534 * l2arc_write_buffers() could find it during the function's 1535 * write phase, and try to write it out to the l2arc. 1536 */ 1537 list_insert_after(&dev->l2ad_buflist, hdr, nhdr); 1538 list_remove(&dev->l2ad_buflist, hdr); 1539 1540 mutex_exit(&dev->l2ad_mtx); 1541 1542 /* 1543 * Since we're using the pointer address as the tag when 1544 * incrementing and decrementing the l2ad_alloc refcount, we 1545 * must remove the old pointer (that we're about to destroy) and 1546 * add the new pointer to the refcount. Otherwise we'd remove 1547 * the wrong pointer address when calling arc_hdr_destroy() later. 1548 */ 1549 1550 (void) refcount_remove_many(&dev->l2ad_alloc, 1551 hdr->b_l2hdr.b_asize, hdr); 1552 1553 (void) refcount_add_many(&dev->l2ad_alloc, 1554 nhdr->b_l2hdr.b_asize, nhdr); 1555 1556 buf_discard_identity(hdr); 1557 hdr->b_freeze_cksum = NULL; 1558 kmem_cache_free(old, hdr); 1559 1560 return (nhdr); 1561} 1562 1563 1564#define ARC_MINTIME (hz>>4) /* 62 ms */ 1565 1566static void 1567arc_cksum_verify(arc_buf_t *buf) 1568{ 1569 zio_cksum_t zc; 1570 1571 if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 1572 return; 1573 1574 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1575 if (buf->b_hdr->b_freeze_cksum == NULL || HDR_IO_ERROR(buf->b_hdr)) { 1576 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1577 return; 1578 } 1579 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, NULL, &zc); 1580 if (!ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc)) 1581 panic("buffer modified while frozen!"); 1582 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1583} 1584 1585static int 1586arc_cksum_equal(arc_buf_t *buf) 1587{ 1588 zio_cksum_t zc; 1589 int equal; 1590 1591 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1592 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, NULL, &zc); 1593 equal = ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc); 1594 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1595 1596 return (equal); 1597} 1598 1599static void 1600arc_cksum_compute(arc_buf_t *buf, boolean_t force) 1601{ 1602 if (!force && !(zfs_flags & ZFS_DEBUG_MODIFY)) 1603 return; 1604 1605 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1606 if (buf->b_hdr->b_freeze_cksum != NULL) { 1607 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1608 return; 1609 } 1610 buf->b_hdr->b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t), KM_SLEEP); 1611 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, 1612 NULL, buf->b_hdr->b_freeze_cksum); 1613 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1614#ifdef illumos 1615 arc_buf_watch(buf); 1616#endif 1617} 1618 1619#ifdef illumos 1620#ifndef _KERNEL 1621typedef struct procctl { 1622 long cmd; 1623 prwatch_t prwatch; 1624} procctl_t; 1625#endif 1626 1627/* ARGSUSED */ 1628static void 1629arc_buf_unwatch(arc_buf_t *buf) 1630{ 1631#ifndef _KERNEL 1632 if (arc_watch) { 1633 int result; 1634 procctl_t ctl; 1635 ctl.cmd = PCWATCH; 1636 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data; 1637 ctl.prwatch.pr_size = 0; 1638 ctl.prwatch.pr_wflags = 0; 1639 result = write(arc_procfd, &ctl, sizeof (ctl)); 1640 ASSERT3U(result, ==, sizeof (ctl)); 1641 } 1642#endif 1643} 1644 1645/* ARGSUSED */ 1646static void 1647arc_buf_watch(arc_buf_t *buf) 1648{ 1649#ifndef _KERNEL 1650 if (arc_watch) { 1651 int result; 1652 procctl_t ctl; 1653 ctl.cmd = PCWATCH; 1654 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data; 1655 ctl.prwatch.pr_size = buf->b_hdr->b_size; 1656 ctl.prwatch.pr_wflags = WA_WRITE; 1657 result = write(arc_procfd, &ctl, sizeof (ctl)); 1658 ASSERT3U(result, ==, sizeof (ctl)); 1659 } 1660#endif 1661} 1662#endif /* illumos */ 1663 1664static arc_buf_contents_t 1665arc_buf_type(arc_buf_hdr_t *hdr) 1666{ 1667 if (HDR_ISTYPE_METADATA(hdr)) { 1668 return (ARC_BUFC_METADATA); 1669 } else { 1670 return (ARC_BUFC_DATA); 1671 } 1672} 1673 1674static uint32_t 1675arc_bufc_to_flags(arc_buf_contents_t type) 1676{ 1677 switch (type) { 1678 case ARC_BUFC_DATA: 1679 /* metadata field is 0 if buffer contains normal data */ 1680 return (0); 1681 case ARC_BUFC_METADATA: 1682 return (ARC_FLAG_BUFC_METADATA); 1683 default: 1684 break; 1685 } 1686 panic("undefined ARC buffer type!"); 1687 return ((uint32_t)-1); 1688} 1689 1690void 1691arc_buf_thaw(arc_buf_t *buf) 1692{ 1693 if (zfs_flags & ZFS_DEBUG_MODIFY) { 1694 if (buf->b_hdr->b_l1hdr.b_state != arc_anon) 1695 panic("modifying non-anon buffer!"); 1696 if (HDR_IO_IN_PROGRESS(buf->b_hdr)) 1697 panic("modifying buffer while i/o in progress!"); 1698 arc_cksum_verify(buf); 1699 } 1700 1701 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1702 if (buf->b_hdr->b_freeze_cksum != NULL) { 1703 kmem_free(buf->b_hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 1704 buf->b_hdr->b_freeze_cksum = NULL; 1705 } 1706 1707#ifdef ZFS_DEBUG 1708 if (zfs_flags & ZFS_DEBUG_MODIFY) { 1709 if (buf->b_hdr->b_l1hdr.b_thawed != NULL) 1710 kmem_free(buf->b_hdr->b_l1hdr.b_thawed, 1); 1711 buf->b_hdr->b_l1hdr.b_thawed = kmem_alloc(1, KM_SLEEP); 1712 } 1713#endif 1714 1715 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1716 1717#ifdef illumos 1718 arc_buf_unwatch(buf); 1719#endif 1720} 1721 1722void 1723arc_buf_freeze(arc_buf_t *buf) 1724{ 1725 kmutex_t *hash_lock; 1726 1727 if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 1728 return; 1729 1730 hash_lock = HDR_LOCK(buf->b_hdr); 1731 mutex_enter(hash_lock); 1732 1733 ASSERT(buf->b_hdr->b_freeze_cksum != NULL || 1734 buf->b_hdr->b_l1hdr.b_state == arc_anon); 1735 arc_cksum_compute(buf, B_FALSE); 1736 mutex_exit(hash_lock); 1737 1738} 1739 1740static void 1741add_reference(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, void *tag) 1742{ 1743 ASSERT(HDR_HAS_L1HDR(hdr)); 1744 ASSERT(MUTEX_HELD(hash_lock)); 1745 arc_state_t *state = hdr->b_l1hdr.b_state; 1746 1747 if ((refcount_add(&hdr->b_l1hdr.b_refcnt, tag) == 1) && 1748 (state != arc_anon)) { 1749 /* We don't use the L2-only state list. */ 1750 if (state != arc_l2c_only) { 1751 arc_buf_contents_t type = arc_buf_type(hdr); 1752 uint64_t delta = hdr->b_size * hdr->b_l1hdr.b_datacnt; 1753 multilist_t *list = &state->arcs_list[type]; 1754 uint64_t *size = &state->arcs_lsize[type]; 1755 1756 multilist_remove(list, hdr); 1757 1758 if (GHOST_STATE(state)) { 1759 ASSERT0(hdr->b_l1hdr.b_datacnt); 1760 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 1761 delta = hdr->b_size; 1762 } 1763 ASSERT(delta > 0); 1764 ASSERT3U(*size, >=, delta); 1765 atomic_add_64(size, -delta); 1766 } 1767 /* remove the prefetch flag if we get a reference */ 1768 hdr->b_flags &= ~ARC_FLAG_PREFETCH; 1769 } 1770} 1771 1772static int 1773remove_reference(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, void *tag) 1774{ 1775 int cnt; 1776 arc_state_t *state = hdr->b_l1hdr.b_state; 1777 1778 ASSERT(HDR_HAS_L1HDR(hdr)); 1779 ASSERT(state == arc_anon || MUTEX_HELD(hash_lock)); 1780 ASSERT(!GHOST_STATE(state)); 1781 1782 /* 1783 * arc_l2c_only counts as a ghost state so we don't need to explicitly 1784 * check to prevent usage of the arc_l2c_only list. 1785 */ 1786 if (((cnt = refcount_remove(&hdr->b_l1hdr.b_refcnt, tag)) == 0) && 1787 (state != arc_anon)) { 1788 arc_buf_contents_t type = arc_buf_type(hdr); 1789 multilist_t *list = &state->arcs_list[type]; 1790 uint64_t *size = &state->arcs_lsize[type]; 1791 1792 multilist_insert(list, hdr); 1793 1794 ASSERT(hdr->b_l1hdr.b_datacnt > 0); 1795 atomic_add_64(size, hdr->b_size * 1796 hdr->b_l1hdr.b_datacnt); 1797 } 1798 return (cnt); 1799} 1800 1801/* 1802 * Move the supplied buffer to the indicated state. The hash lock 1803 * for the buffer must be held by the caller. 1804 */ 1805static void 1806arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *hdr, 1807 kmutex_t *hash_lock) 1808{ 1809 arc_state_t *old_state; 1810 int64_t refcnt; 1811 uint32_t datacnt; 1812 uint64_t from_delta, to_delta; 1813 arc_buf_contents_t buftype = arc_buf_type(hdr); 1814 1815 /* 1816 * We almost always have an L1 hdr here, since we call arc_hdr_realloc() 1817 * in arc_read() when bringing a buffer out of the L2ARC. However, the 1818 * L1 hdr doesn't always exist when we change state to arc_anon before 1819 * destroying a header, in which case reallocating to add the L1 hdr is 1820 * pointless. 1821 */ 1822 if (HDR_HAS_L1HDR(hdr)) { 1823 old_state = hdr->b_l1hdr.b_state; 1824 refcnt = refcount_count(&hdr->b_l1hdr.b_refcnt); 1825 datacnt = hdr->b_l1hdr.b_datacnt; 1826 } else { 1827 old_state = arc_l2c_only; 1828 refcnt = 0; 1829 datacnt = 0; 1830 } 1831 1832 ASSERT(MUTEX_HELD(hash_lock)); 1833 ASSERT3P(new_state, !=, old_state); 1834 ASSERT(refcnt == 0 || datacnt > 0); 1835 ASSERT(!GHOST_STATE(new_state) || datacnt == 0); 1836 ASSERT(old_state != arc_anon || datacnt <= 1); 1837 1838 from_delta = to_delta = datacnt * hdr->b_size; 1839 1840 /* 1841 * If this buffer is evictable, transfer it from the 1842 * old state list to the new state list. 1843 */ 1844 if (refcnt == 0) { 1845 if (old_state != arc_anon && old_state != arc_l2c_only) { 1846 uint64_t *size = &old_state->arcs_lsize[buftype]; 1847 1848 ASSERT(HDR_HAS_L1HDR(hdr)); 1849 multilist_remove(&old_state->arcs_list[buftype], hdr); 1850 1851 /* 1852 * If prefetching out of the ghost cache, 1853 * we will have a non-zero datacnt. 1854 */ 1855 if (GHOST_STATE(old_state) && datacnt == 0) { 1856 /* ghost elements have a ghost size */ 1857 ASSERT(hdr->b_l1hdr.b_buf == NULL); 1858 from_delta = hdr->b_size; 1859 } 1860 ASSERT3U(*size, >=, from_delta); 1861 atomic_add_64(size, -from_delta); 1862 } 1863 if (new_state != arc_anon && new_state != arc_l2c_only) { 1864 uint64_t *size = &new_state->arcs_lsize[buftype]; 1865 1866 /* 1867 * An L1 header always exists here, since if we're 1868 * moving to some L1-cached state (i.e. not l2c_only or 1869 * anonymous), we realloc the header to add an L1hdr 1870 * beforehand. 1871 */ 1872 ASSERT(HDR_HAS_L1HDR(hdr)); 1873 multilist_insert(&new_state->arcs_list[buftype], hdr); 1874 1875 /* ghost elements have a ghost size */ 1876 if (GHOST_STATE(new_state)) { 1877 ASSERT0(datacnt); 1878 ASSERT(hdr->b_l1hdr.b_buf == NULL); 1879 to_delta = hdr->b_size; 1880 } 1881 atomic_add_64(size, to_delta); 1882 } 1883 } 1884 1885 ASSERT(!BUF_EMPTY(hdr)); 1886 if (new_state == arc_anon && HDR_IN_HASH_TABLE(hdr)) 1887 buf_hash_remove(hdr); 1888 1889 /* adjust state sizes (ignore arc_l2c_only) */ 1890 1891 if (to_delta && new_state != arc_l2c_only) { 1892 ASSERT(HDR_HAS_L1HDR(hdr)); 1893 if (GHOST_STATE(new_state)) { 1894 ASSERT0(datacnt); 1895 1896 /* 1897 * We moving a header to a ghost state, we first 1898 * remove all arc buffers. Thus, we'll have a 1899 * datacnt of zero, and no arc buffer to use for 1900 * the reference. As a result, we use the arc 1901 * header pointer for the reference. 1902 */ 1903 (void) refcount_add_many(&new_state->arcs_size, 1904 hdr->b_size, hdr); 1905 } else { 1906 ASSERT3U(datacnt, !=, 0); 1907 1908 /* 1909 * Each individual buffer holds a unique reference, 1910 * thus we must remove each of these references one 1911 * at a time. 1912 */ 1913 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL; 1914 buf = buf->b_next) { 1915 (void) refcount_add_many(&new_state->arcs_size, 1916 hdr->b_size, buf); 1917 } 1918 } 1919 } 1920 1921 if (from_delta && old_state != arc_l2c_only) { 1922 ASSERT(HDR_HAS_L1HDR(hdr)); 1923 if (GHOST_STATE(old_state)) { 1924 /* 1925 * When moving a header off of a ghost state, 1926 * there's the possibility for datacnt to be 1927 * non-zero. This is because we first add the 1928 * arc buffer to the header prior to changing 1929 * the header's state. Since we used the header 1930 * for the reference when putting the header on 1931 * the ghost state, we must balance that and use 1932 * the header when removing off the ghost state 1933 * (even though datacnt is non zero). 1934 */ 1935 1936 IMPLY(datacnt == 0, new_state == arc_anon || 1937 new_state == arc_l2c_only); 1938 1939 (void) refcount_remove_many(&old_state->arcs_size, 1940 hdr->b_size, hdr); 1941 } else { 1942 ASSERT3P(datacnt, !=, 0); 1943 1944 /* 1945 * Each individual buffer holds a unique reference, 1946 * thus we must remove each of these references one 1947 * at a time. 1948 */ 1949 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL; 1950 buf = buf->b_next) { 1951 (void) refcount_remove_many( 1952 &old_state->arcs_size, hdr->b_size, buf); 1953 } 1954 } 1955 } 1956 1957 if (HDR_HAS_L1HDR(hdr)) 1958 hdr->b_l1hdr.b_state = new_state; 1959 1960 /* 1961 * L2 headers should never be on the L2 state list since they don't 1962 * have L1 headers allocated. 1963 */ 1964 ASSERT(multilist_is_empty(&arc_l2c_only->arcs_list[ARC_BUFC_DATA]) && 1965 multilist_is_empty(&arc_l2c_only->arcs_list[ARC_BUFC_METADATA])); 1966} 1967 1968void 1969arc_space_consume(uint64_t space, arc_space_type_t type) 1970{ 1971 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES); 1972 1973 switch (type) { 1974 case ARC_SPACE_DATA: 1975 ARCSTAT_INCR(arcstat_data_size, space); 1976 break; 1977 case ARC_SPACE_META: 1978 ARCSTAT_INCR(arcstat_metadata_size, space); 1979 break; 1980 case ARC_SPACE_OTHER: 1981 ARCSTAT_INCR(arcstat_other_size, space); 1982 break; 1983 case ARC_SPACE_HDRS: 1984 ARCSTAT_INCR(arcstat_hdr_size, space); 1985 break; 1986 case ARC_SPACE_L2HDRS: 1987 ARCSTAT_INCR(arcstat_l2_hdr_size, space); 1988 break; 1989 } 1990 1991 if (type != ARC_SPACE_DATA) 1992 ARCSTAT_INCR(arcstat_meta_used, space); 1993 1994 atomic_add_64(&arc_size, space); 1995} 1996 1997void 1998arc_space_return(uint64_t space, arc_space_type_t type) 1999{ 2000 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES); 2001 2002 switch (type) { 2003 case ARC_SPACE_DATA: 2004 ARCSTAT_INCR(arcstat_data_size, -space); 2005 break; 2006 case ARC_SPACE_META: 2007 ARCSTAT_INCR(arcstat_metadata_size, -space); 2008 break; 2009 case ARC_SPACE_OTHER: 2010 ARCSTAT_INCR(arcstat_other_size, -space); 2011 break; 2012 case ARC_SPACE_HDRS: 2013 ARCSTAT_INCR(arcstat_hdr_size, -space); 2014 break; 2015 case ARC_SPACE_L2HDRS: 2016 ARCSTAT_INCR(arcstat_l2_hdr_size, -space); 2017 break; 2018 } 2019 2020 if (type != ARC_SPACE_DATA) { 2021 ASSERT(arc_meta_used >= space); 2022 if (arc_meta_max < arc_meta_used) 2023 arc_meta_max = arc_meta_used; 2024 ARCSTAT_INCR(arcstat_meta_used, -space); 2025 } 2026 2027 ASSERT(arc_size >= space); 2028 atomic_add_64(&arc_size, -space); 2029} 2030 2031arc_buf_t * 2032arc_buf_alloc(spa_t *spa, int32_t size, void *tag, arc_buf_contents_t type) 2033{ 2034 arc_buf_hdr_t *hdr; 2035 arc_buf_t *buf; 2036 2037 ASSERT3U(size, >, 0); 2038 hdr = kmem_cache_alloc(hdr_full_cache, KM_PUSHPAGE); 2039 ASSERT(BUF_EMPTY(hdr)); 2040 ASSERT3P(hdr->b_freeze_cksum, ==, NULL); 2041 hdr->b_size = size; 2042 hdr->b_spa = spa_load_guid(spa); 2043 2044 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 2045 buf->b_hdr = hdr; 2046 buf->b_data = NULL; 2047 buf->b_efunc = NULL; 2048 buf->b_private = NULL; 2049 buf->b_next = NULL; 2050 2051 hdr->b_flags = arc_bufc_to_flags(type); 2052 hdr->b_flags |= ARC_FLAG_HAS_L1HDR; 2053 2054 hdr->b_l1hdr.b_buf = buf; 2055 hdr->b_l1hdr.b_state = arc_anon; 2056 hdr->b_l1hdr.b_arc_access = 0; 2057 hdr->b_l1hdr.b_datacnt = 1; 2058 hdr->b_l1hdr.b_tmp_cdata = NULL; 2059 2060 arc_get_data_buf(buf); 2061 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 2062 (void) refcount_add(&hdr->b_l1hdr.b_refcnt, tag); 2063 2064 return (buf); 2065} 2066 2067static char *arc_onloan_tag = "onloan"; 2068 2069/* 2070 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in 2071 * flight data by arc_tempreserve_space() until they are "returned". Loaned 2072 * buffers must be returned to the arc before they can be used by the DMU or 2073 * freed. 2074 */ 2075arc_buf_t * 2076arc_loan_buf(spa_t *spa, int size) 2077{ 2078 arc_buf_t *buf; 2079 2080 buf = arc_buf_alloc(spa, size, arc_onloan_tag, ARC_BUFC_DATA); 2081 2082 atomic_add_64(&arc_loaned_bytes, size); 2083 return (buf); 2084} 2085 2086/* 2087 * Return a loaned arc buffer to the arc. 2088 */ 2089void 2090arc_return_buf(arc_buf_t *buf, void *tag) 2091{ 2092 arc_buf_hdr_t *hdr = buf->b_hdr; 2093 2094 ASSERT(buf->b_data != NULL); 2095 ASSERT(HDR_HAS_L1HDR(hdr)); 2096 (void) refcount_add(&hdr->b_l1hdr.b_refcnt, tag); 2097 (void) refcount_remove(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag); 2098 2099 atomic_add_64(&arc_loaned_bytes, -hdr->b_size); 2100} 2101 2102/* Detach an arc_buf from a dbuf (tag) */ 2103void 2104arc_loan_inuse_buf(arc_buf_t *buf, void *tag) 2105{ 2106 arc_buf_hdr_t *hdr = buf->b_hdr; 2107 2108 ASSERT(buf->b_data != NULL); 2109 ASSERT(HDR_HAS_L1HDR(hdr)); 2110 (void) refcount_add(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag); 2111 (void) refcount_remove(&hdr->b_l1hdr.b_refcnt, tag); 2112 buf->b_efunc = NULL; 2113 buf->b_private = NULL; 2114 2115 atomic_add_64(&arc_loaned_bytes, hdr->b_size); 2116} 2117 2118static arc_buf_t * 2119arc_buf_clone(arc_buf_t *from) 2120{ 2121 arc_buf_t *buf; 2122 arc_buf_hdr_t *hdr = from->b_hdr; 2123 uint64_t size = hdr->b_size; 2124 2125 ASSERT(HDR_HAS_L1HDR(hdr)); 2126 ASSERT(hdr->b_l1hdr.b_state != arc_anon); 2127 2128 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 2129 buf->b_hdr = hdr; 2130 buf->b_data = NULL; 2131 buf->b_efunc = NULL; 2132 buf->b_private = NULL; 2133 buf->b_next = hdr->b_l1hdr.b_buf; 2134 hdr->b_l1hdr.b_buf = buf; 2135 arc_get_data_buf(buf); 2136 bcopy(from->b_data, buf->b_data, size); 2137 2138 /* 2139 * This buffer already exists in the arc so create a duplicate 2140 * copy for the caller. If the buffer is associated with user data 2141 * then track the size and number of duplicates. These stats will be 2142 * updated as duplicate buffers are created and destroyed. 2143 */ 2144 if (HDR_ISTYPE_DATA(hdr)) { 2145 ARCSTAT_BUMP(arcstat_duplicate_buffers); 2146 ARCSTAT_INCR(arcstat_duplicate_buffers_size, size); 2147 } 2148 hdr->b_l1hdr.b_datacnt += 1; 2149 return (buf); 2150} 2151 2152void 2153arc_buf_add_ref(arc_buf_t *buf, void* tag) 2154{ 2155 arc_buf_hdr_t *hdr; 2156 kmutex_t *hash_lock; 2157 2158 /* 2159 * Check to see if this buffer is evicted. Callers 2160 * must verify b_data != NULL to know if the add_ref 2161 * was successful. 2162 */ 2163 mutex_enter(&buf->b_evict_lock); 2164 if (buf->b_data == NULL) { 2165 mutex_exit(&buf->b_evict_lock); 2166 return; 2167 } 2168 hash_lock = HDR_LOCK(buf->b_hdr); 2169 mutex_enter(hash_lock); 2170 hdr = buf->b_hdr; 2171 ASSERT(HDR_HAS_L1HDR(hdr)); 2172 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 2173 mutex_exit(&buf->b_evict_lock); 2174 2175 ASSERT(hdr->b_l1hdr.b_state == arc_mru || 2176 hdr->b_l1hdr.b_state == arc_mfu); 2177 2178 add_reference(hdr, hash_lock, tag); 2179 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr); 2180 arc_access(hdr, hash_lock); 2181 mutex_exit(hash_lock); 2182 ARCSTAT_BUMP(arcstat_hits); 2183 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr), 2184 demand, prefetch, !HDR_ISTYPE_METADATA(hdr), 2185 data, metadata, hits); 2186} 2187 2188static void 2189arc_buf_free_on_write(void *data, size_t size, 2190 void (*free_func)(void *, size_t)) 2191{ 2192 l2arc_data_free_t *df; 2193 2194 df = kmem_alloc(sizeof (*df), KM_SLEEP); 2195 df->l2df_data = data; 2196 df->l2df_size = size; 2197 df->l2df_func = free_func; 2198 mutex_enter(&l2arc_free_on_write_mtx); 2199 list_insert_head(l2arc_free_on_write, df); 2200 mutex_exit(&l2arc_free_on_write_mtx); 2201} 2202 2203/* 2204 * Free the arc data buffer. If it is an l2arc write in progress, 2205 * the buffer is placed on l2arc_free_on_write to be freed later. 2206 */ 2207static void 2208arc_buf_data_free(arc_buf_t *buf, void (*free_func)(void *, size_t)) 2209{ 2210 arc_buf_hdr_t *hdr = buf->b_hdr; 2211 2212 if (HDR_L2_WRITING(hdr)) { 2213 arc_buf_free_on_write(buf->b_data, hdr->b_size, free_func); 2214 ARCSTAT_BUMP(arcstat_l2_free_on_write); 2215 } else { 2216 free_func(buf->b_data, hdr->b_size); 2217 } 2218} 2219 2220static void 2221arc_buf_l2_cdata_free(arc_buf_hdr_t *hdr) 2222{ 2223 size_t align, asize, len; 2224 2225 ASSERT(HDR_HAS_L2HDR(hdr)); 2226 ASSERT(MUTEX_HELD(&hdr->b_l2hdr.b_dev->l2ad_mtx)); 2227 2228 /* 2229 * The b_tmp_cdata field is linked off of the b_l1hdr, so if 2230 * that doesn't exist, the header is in the arc_l2c_only state, 2231 * and there isn't anything to free (it's already been freed). 2232 */ 2233 if (!HDR_HAS_L1HDR(hdr)) 2234 return; 2235 2236 /* 2237 * The header isn't being written to the l2arc device, thus it 2238 * shouldn't have a b_tmp_cdata to free. 2239 */ 2240 if (!HDR_L2_WRITING(hdr)) { 2241 ASSERT3P(hdr->b_l1hdr.b_tmp_cdata, ==, NULL); 2242 return; 2243 } 2244 2245 /* 2246 * The bufer has been chosen for writing to L2ARC, but it's 2247 * not being written just yet. In other words, 2248 * b_tmp_cdata points to exactly the same buffer as b_data, 2249 * l2arc_transform_buf hasn't been called. 2250 */ 2251 if (hdr->b_l2hdr.b_daddr == L2ARC_ADDR_UNSET) { 2252 ASSERT3P(hdr->b_l1hdr.b_tmp_cdata, ==, 2253 hdr->b_l1hdr.b_buf->b_data); 2254 ASSERT3U(hdr->b_l2hdr.b_compress, ==, ZIO_COMPRESS_OFF); 2255 return; 2256 } 2257 2258 /* 2259 * There's nothing to free since the buffer was all zero's and 2260 * compressed to a zero length buffer. 2261 */ 2262 if (hdr->b_l2hdr.b_compress == ZIO_COMPRESS_EMPTY) { 2263 ASSERT3P(hdr->b_l1hdr.b_tmp_cdata, ==, NULL); 2264 return; 2265 } 2266 2267 /* 2268 * Nothing to do if the temporary buffer was not required. 2269 */ 2270 if (hdr->b_l1hdr.b_tmp_cdata == NULL) 2271 return; 2272 2273 ARCSTAT_BUMP(arcstat_l2_cdata_free_on_write); 2274 len = hdr->b_size; 2275 align = (size_t)1 << hdr->b_l2hdr.b_dev->l2ad_vdev->vdev_ashift; 2276 asize = P2ROUNDUP(len, align); 2277 arc_buf_free_on_write(hdr->b_l1hdr.b_tmp_cdata, asize, 2278 zio_data_buf_free); 2279 hdr->b_l1hdr.b_tmp_cdata = NULL; 2280} 2281 2282/* 2283 * Free up buf->b_data and if 'remove' is set, then pull the 2284 * arc_buf_t off of the the arc_buf_hdr_t's list and free it. 2285 */ 2286static void 2287arc_buf_destroy(arc_buf_t *buf, boolean_t remove) 2288{ 2289 arc_buf_t **bufp; 2290 2291 /* free up data associated with the buf */ 2292 if (buf->b_data != NULL) { 2293 arc_state_t *state = buf->b_hdr->b_l1hdr.b_state; 2294 uint64_t size = buf->b_hdr->b_size; 2295 arc_buf_contents_t type = arc_buf_type(buf->b_hdr); 2296 2297 arc_cksum_verify(buf); 2298#ifdef illumos 2299 arc_buf_unwatch(buf); 2300#endif 2301 2302 if (type == ARC_BUFC_METADATA) { 2303 arc_buf_data_free(buf, zio_buf_free); 2304 arc_space_return(size, ARC_SPACE_META); 2305 } else { 2306 ASSERT(type == ARC_BUFC_DATA); 2307 arc_buf_data_free(buf, zio_data_buf_free); 2308 arc_space_return(size, ARC_SPACE_DATA); 2309 } 2310 2311 /* protected by hash lock, if in the hash table */ 2312 if (multilist_link_active(&buf->b_hdr->b_l1hdr.b_arc_node)) { 2313 uint64_t *cnt = &state->arcs_lsize[type]; 2314 2315 ASSERT(refcount_is_zero( 2316 &buf->b_hdr->b_l1hdr.b_refcnt)); 2317 ASSERT(state != arc_anon && state != arc_l2c_only); 2318 2319 ASSERT3U(*cnt, >=, size); 2320 atomic_add_64(cnt, -size); 2321 } 2322 2323 (void) refcount_remove_many(&state->arcs_size, size, buf); 2324 buf->b_data = NULL; 2325 2326 /* 2327 * If we're destroying a duplicate buffer make sure 2328 * that the appropriate statistics are updated. 2329 */ 2330 if (buf->b_hdr->b_l1hdr.b_datacnt > 1 && 2331 HDR_ISTYPE_DATA(buf->b_hdr)) { 2332 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers); 2333 ARCSTAT_INCR(arcstat_duplicate_buffers_size, -size); 2334 } 2335 ASSERT(buf->b_hdr->b_l1hdr.b_datacnt > 0); 2336 buf->b_hdr->b_l1hdr.b_datacnt -= 1; 2337 } 2338 2339 /* only remove the buf if requested */ 2340 if (!remove) 2341 return; 2342 2343 /* remove the buf from the hdr list */ 2344 for (bufp = &buf->b_hdr->b_l1hdr.b_buf; *bufp != buf; 2345 bufp = &(*bufp)->b_next) 2346 continue; 2347 *bufp = buf->b_next; 2348 buf->b_next = NULL; 2349 2350 ASSERT(buf->b_efunc == NULL); 2351 2352 /* clean up the buf */ 2353 buf->b_hdr = NULL; 2354 kmem_cache_free(buf_cache, buf); 2355} 2356 2357static void 2358arc_hdr_l2hdr_destroy(arc_buf_hdr_t *hdr) 2359{ 2360 l2arc_buf_hdr_t *l2hdr = &hdr->b_l2hdr; 2361 l2arc_dev_t *dev = l2hdr->b_dev; 2362 2363 ASSERT(MUTEX_HELD(&dev->l2ad_mtx)); 2364 ASSERT(HDR_HAS_L2HDR(hdr)); 2365 2366 list_remove(&dev->l2ad_buflist, hdr); 2367 2368 /* 2369 * We don't want to leak the b_tmp_cdata buffer that was 2370 * allocated in l2arc_write_buffers() 2371 */ 2372 arc_buf_l2_cdata_free(hdr); 2373 2374 /* 2375 * If the l2hdr's b_daddr is equal to L2ARC_ADDR_UNSET, then 2376 * this header is being processed by l2arc_write_buffers() (i.e. 2377 * it's in the first stage of l2arc_write_buffers()). 2378 * Re-affirming that truth here, just to serve as a reminder. If 2379 * b_daddr does not equal L2ARC_ADDR_UNSET, then the header may or 2380 * may not have its HDR_L2_WRITING flag set. (the write may have 2381 * completed, in which case HDR_L2_WRITING will be false and the 2382 * b_daddr field will point to the address of the buffer on disk). 2383 */ 2384 IMPLY(l2hdr->b_daddr == L2ARC_ADDR_UNSET, HDR_L2_WRITING(hdr)); 2385 2386 /* 2387 * If b_daddr is equal to L2ARC_ADDR_UNSET, we're racing with 2388 * l2arc_write_buffers(). Since we've just removed this header 2389 * from the l2arc buffer list, this header will never reach the 2390 * second stage of l2arc_write_buffers(), which increments the 2391 * accounting stats for this header. Thus, we must be careful 2392 * not to decrement them for this header either. 2393 */ 2394 if (l2hdr->b_daddr != L2ARC_ADDR_UNSET) { 2395 ARCSTAT_INCR(arcstat_l2_asize, -l2hdr->b_asize); 2396 ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size); 2397 2398 vdev_space_update(dev->l2ad_vdev, 2399 -l2hdr->b_asize, 0, 0); 2400 2401 (void) refcount_remove_many(&dev->l2ad_alloc, 2402 l2hdr->b_asize, hdr); 2403 } 2404 2405 hdr->b_flags &= ~ARC_FLAG_HAS_L2HDR; 2406} 2407 2408static void 2409arc_hdr_destroy(arc_buf_hdr_t *hdr) 2410{ 2411 if (HDR_HAS_L1HDR(hdr)) { 2412 ASSERT(hdr->b_l1hdr.b_buf == NULL || 2413 hdr->b_l1hdr.b_datacnt > 0); 2414 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 2415 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon); 2416 } 2417 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 2418 ASSERT(!HDR_IN_HASH_TABLE(hdr)); 2419 2420 if (HDR_HAS_L2HDR(hdr)) { 2421 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev; 2422 boolean_t buflist_held = MUTEX_HELD(&dev->l2ad_mtx); 2423 2424 if (!buflist_held) 2425 mutex_enter(&dev->l2ad_mtx); 2426 2427 /* 2428 * Even though we checked this conditional above, we 2429 * need to check this again now that we have the 2430 * l2ad_mtx. This is because we could be racing with 2431 * another thread calling l2arc_evict() which might have 2432 * destroyed this header's L2 portion as we were waiting 2433 * to acquire the l2ad_mtx. If that happens, we don't 2434 * want to re-destroy the header's L2 portion. 2435 */ 2436 if (HDR_HAS_L2HDR(hdr)) { 2437 l2arc_trim(hdr); 2438 arc_hdr_l2hdr_destroy(hdr); 2439 } 2440 2441 if (!buflist_held) 2442 mutex_exit(&dev->l2ad_mtx); 2443 } 2444 2445 if (!BUF_EMPTY(hdr)) 2446 buf_discard_identity(hdr); 2447 2448 if (hdr->b_freeze_cksum != NULL) { 2449 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 2450 hdr->b_freeze_cksum = NULL; 2451 } 2452 2453 if (HDR_HAS_L1HDR(hdr)) { 2454 while (hdr->b_l1hdr.b_buf) { 2455 arc_buf_t *buf = hdr->b_l1hdr.b_buf; 2456 2457 if (buf->b_efunc != NULL) { 2458 mutex_enter(&arc_user_evicts_lock); 2459 mutex_enter(&buf->b_evict_lock); 2460 ASSERT(buf->b_hdr != NULL); 2461 arc_buf_destroy(hdr->b_l1hdr.b_buf, FALSE); 2462 hdr->b_l1hdr.b_buf = buf->b_next; 2463 buf->b_hdr = &arc_eviction_hdr; 2464 buf->b_next = arc_eviction_list; 2465 arc_eviction_list = buf; 2466 mutex_exit(&buf->b_evict_lock); 2467 cv_signal(&arc_user_evicts_cv); 2468 mutex_exit(&arc_user_evicts_lock); 2469 } else { 2470 arc_buf_destroy(hdr->b_l1hdr.b_buf, TRUE); 2471 } 2472 } 2473#ifdef ZFS_DEBUG 2474 if (hdr->b_l1hdr.b_thawed != NULL) { 2475 kmem_free(hdr->b_l1hdr.b_thawed, 1); 2476 hdr->b_l1hdr.b_thawed = NULL; 2477 } 2478#endif 2479 } 2480 2481 ASSERT3P(hdr->b_hash_next, ==, NULL); 2482 if (HDR_HAS_L1HDR(hdr)) { 2483 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 2484 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL); 2485 kmem_cache_free(hdr_full_cache, hdr); 2486 } else { 2487 kmem_cache_free(hdr_l2only_cache, hdr); 2488 } 2489} 2490 2491void 2492arc_buf_free(arc_buf_t *buf, void *tag) 2493{ 2494 arc_buf_hdr_t *hdr = buf->b_hdr; 2495 int hashed = hdr->b_l1hdr.b_state != arc_anon; 2496 2497 ASSERT(buf->b_efunc == NULL); 2498 ASSERT(buf->b_data != NULL); 2499 2500 if (hashed) { 2501 kmutex_t *hash_lock = HDR_LOCK(hdr); 2502 2503 mutex_enter(hash_lock); 2504 hdr = buf->b_hdr; 2505 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 2506 2507 (void) remove_reference(hdr, hash_lock, tag); 2508 if (hdr->b_l1hdr.b_datacnt > 1) { 2509 arc_buf_destroy(buf, TRUE); 2510 } else { 2511 ASSERT(buf == hdr->b_l1hdr.b_buf); 2512 ASSERT(buf->b_efunc == NULL); 2513 hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE; 2514 } 2515 mutex_exit(hash_lock); 2516 } else if (HDR_IO_IN_PROGRESS(hdr)) { 2517 int destroy_hdr; 2518 /* 2519 * We are in the middle of an async write. Don't destroy 2520 * this buffer unless the write completes before we finish 2521 * decrementing the reference count. 2522 */ 2523 mutex_enter(&arc_user_evicts_lock); 2524 (void) remove_reference(hdr, NULL, tag); 2525 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 2526 destroy_hdr = !HDR_IO_IN_PROGRESS(hdr); 2527 mutex_exit(&arc_user_evicts_lock); 2528 if (destroy_hdr) 2529 arc_hdr_destroy(hdr); 2530 } else { 2531 if (remove_reference(hdr, NULL, tag) > 0) 2532 arc_buf_destroy(buf, TRUE); 2533 else 2534 arc_hdr_destroy(hdr); 2535 } 2536} 2537 2538boolean_t 2539arc_buf_remove_ref(arc_buf_t *buf, void* tag) 2540{ 2541 arc_buf_hdr_t *hdr = buf->b_hdr; 2542 kmutex_t *hash_lock = HDR_LOCK(hdr); 2543 boolean_t no_callback = (buf->b_efunc == NULL); 2544 2545 if (hdr->b_l1hdr.b_state == arc_anon) { 2546 ASSERT(hdr->b_l1hdr.b_datacnt == 1); 2547 arc_buf_free(buf, tag); 2548 return (no_callback); 2549 } 2550 2551 mutex_enter(hash_lock); 2552 hdr = buf->b_hdr; 2553 ASSERT(hdr->b_l1hdr.b_datacnt > 0); 2554 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 2555 ASSERT(hdr->b_l1hdr.b_state != arc_anon); 2556 ASSERT(buf->b_data != NULL); 2557 2558 (void) remove_reference(hdr, hash_lock, tag); 2559 if (hdr->b_l1hdr.b_datacnt > 1) { 2560 if (no_callback) 2561 arc_buf_destroy(buf, TRUE); 2562 } else if (no_callback) { 2563 ASSERT(hdr->b_l1hdr.b_buf == buf && buf->b_next == NULL); 2564 ASSERT(buf->b_efunc == NULL); 2565 hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE; 2566 } 2567 ASSERT(no_callback || hdr->b_l1hdr.b_datacnt > 1 || 2568 refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 2569 mutex_exit(hash_lock); 2570 return (no_callback); 2571} 2572 2573int32_t 2574arc_buf_size(arc_buf_t *buf) 2575{ 2576 return (buf->b_hdr->b_size); 2577} 2578 2579/* 2580 * Called from the DMU to determine if the current buffer should be 2581 * evicted. In order to ensure proper locking, the eviction must be initiated 2582 * from the DMU. Return true if the buffer is associated with user data and 2583 * duplicate buffers still exist. 2584 */ 2585boolean_t 2586arc_buf_eviction_needed(arc_buf_t *buf) 2587{ 2588 arc_buf_hdr_t *hdr; 2589 boolean_t evict_needed = B_FALSE; 2590 2591 if (zfs_disable_dup_eviction) 2592 return (B_FALSE); 2593 2594 mutex_enter(&buf->b_evict_lock); 2595 hdr = buf->b_hdr; 2596 if (hdr == NULL) { 2597 /* 2598 * We are in arc_do_user_evicts(); let that function 2599 * perform the eviction. 2600 */ 2601 ASSERT(buf->b_data == NULL); 2602 mutex_exit(&buf->b_evict_lock); 2603 return (B_FALSE); 2604 } else if (buf->b_data == NULL) { 2605 /* 2606 * We have already been added to the arc eviction list; 2607 * recommend eviction. 2608 */ 2609 ASSERT3P(hdr, ==, &arc_eviction_hdr); 2610 mutex_exit(&buf->b_evict_lock); 2611 return (B_TRUE); 2612 } 2613 2614 if (hdr->b_l1hdr.b_datacnt > 1 && HDR_ISTYPE_DATA(hdr)) 2615 evict_needed = B_TRUE; 2616 2617 mutex_exit(&buf->b_evict_lock); 2618 return (evict_needed); 2619} 2620 2621/* 2622 * Evict the arc_buf_hdr that is provided as a parameter. The resultant 2623 * state of the header is dependent on it's state prior to entering this 2624 * function. The following transitions are possible: 2625 * 2626 * - arc_mru -> arc_mru_ghost 2627 * - arc_mfu -> arc_mfu_ghost 2628 * - arc_mru_ghost -> arc_l2c_only 2629 * - arc_mru_ghost -> deleted 2630 * - arc_mfu_ghost -> arc_l2c_only 2631 * - arc_mfu_ghost -> deleted 2632 */ 2633static int64_t 2634arc_evict_hdr(arc_buf_hdr_t *hdr, kmutex_t *hash_lock) 2635{ 2636 arc_state_t *evicted_state, *state; 2637 int64_t bytes_evicted = 0; 2638 2639 ASSERT(MUTEX_HELD(hash_lock)); 2640 ASSERT(HDR_HAS_L1HDR(hdr)); 2641 2642 state = hdr->b_l1hdr.b_state; 2643 if (GHOST_STATE(state)) { 2644 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 2645 ASSERT(hdr->b_l1hdr.b_buf == NULL); 2646 2647 /* 2648 * l2arc_write_buffers() relies on a header's L1 portion 2649 * (i.e. it's b_tmp_cdata field) during it's write phase. 2650 * Thus, we cannot push a header onto the arc_l2c_only 2651 * state (removing it's L1 piece) until the header is 2652 * done being written to the l2arc. 2653 */ 2654 if (HDR_HAS_L2HDR(hdr) && HDR_L2_WRITING(hdr)) { 2655 ARCSTAT_BUMP(arcstat_evict_l2_skip); 2656 return (bytes_evicted); 2657 } 2658 2659 ARCSTAT_BUMP(arcstat_deleted); 2660 bytes_evicted += hdr->b_size; 2661 2662 DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, hdr); 2663 2664 if (HDR_HAS_L2HDR(hdr)) { 2665 /* 2666 * This buffer is cached on the 2nd Level ARC; 2667 * don't destroy the header. 2668 */ 2669 arc_change_state(arc_l2c_only, hdr, hash_lock); 2670 /* 2671 * dropping from L1+L2 cached to L2-only, 2672 * realloc to remove the L1 header. 2673 */ 2674 hdr = arc_hdr_realloc(hdr, hdr_full_cache, 2675 hdr_l2only_cache); 2676 } else { 2677 arc_change_state(arc_anon, hdr, hash_lock); 2678 arc_hdr_destroy(hdr); 2679 } 2680 return (bytes_evicted); 2681 } 2682 2683 ASSERT(state == arc_mru || state == arc_mfu); 2684 evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost; 2685 2686 /* prefetch buffers have a minimum lifespan */ 2687 if (HDR_IO_IN_PROGRESS(hdr) || 2688 ((hdr->b_flags & (ARC_FLAG_PREFETCH | ARC_FLAG_INDIRECT)) && 2689 ddi_get_lbolt() - hdr->b_l1hdr.b_arc_access < 2690 arc_min_prefetch_lifespan)) { 2691 ARCSTAT_BUMP(arcstat_evict_skip); 2692 return (bytes_evicted); 2693 } 2694 2695 ASSERT0(refcount_count(&hdr->b_l1hdr.b_refcnt)); 2696 ASSERT3U(hdr->b_l1hdr.b_datacnt, >, 0); 2697 while (hdr->b_l1hdr.b_buf) { 2698 arc_buf_t *buf = hdr->b_l1hdr.b_buf; 2699 if (!mutex_tryenter(&buf->b_evict_lock)) { 2700 ARCSTAT_BUMP(arcstat_mutex_miss); 2701 break; 2702 } 2703 if (buf->b_data != NULL) 2704 bytes_evicted += hdr->b_size; 2705 if (buf->b_efunc != NULL) { 2706 mutex_enter(&arc_user_evicts_lock); 2707 arc_buf_destroy(buf, FALSE); 2708 hdr->b_l1hdr.b_buf = buf->b_next; 2709 buf->b_hdr = &arc_eviction_hdr; 2710 buf->b_next = arc_eviction_list; 2711 arc_eviction_list = buf; 2712 cv_signal(&arc_user_evicts_cv); 2713 mutex_exit(&arc_user_evicts_lock); 2714 mutex_exit(&buf->b_evict_lock); 2715 } else { 2716 mutex_exit(&buf->b_evict_lock); 2717 arc_buf_destroy(buf, TRUE); 2718 } 2719 } 2720 2721 if (HDR_HAS_L2HDR(hdr)) { 2722 ARCSTAT_INCR(arcstat_evict_l2_cached, hdr->b_size); 2723 } else { 2724 if (l2arc_write_eligible(hdr->b_spa, hdr)) 2725 ARCSTAT_INCR(arcstat_evict_l2_eligible, hdr->b_size); 2726 else 2727 ARCSTAT_INCR(arcstat_evict_l2_ineligible, hdr->b_size); 2728 } 2729 2730 if (hdr->b_l1hdr.b_datacnt == 0) { 2731 arc_change_state(evicted_state, hdr, hash_lock); 2732 ASSERT(HDR_IN_HASH_TABLE(hdr)); 2733 hdr->b_flags |= ARC_FLAG_IN_HASH_TABLE; 2734 hdr->b_flags &= ~ARC_FLAG_BUF_AVAILABLE; 2735 DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, hdr); 2736 } 2737 2738 return (bytes_evicted); 2739} 2740 2741static uint64_t 2742arc_evict_state_impl(multilist_t *ml, int idx, arc_buf_hdr_t *marker, 2743 uint64_t spa, int64_t bytes) 2744{ 2745 multilist_sublist_t *mls; 2746 uint64_t bytes_evicted = 0; 2747 arc_buf_hdr_t *hdr; 2748 kmutex_t *hash_lock; 2749 int evict_count = 0; 2750 2751 ASSERT3P(marker, !=, NULL); 2752 IMPLY(bytes < 0, bytes == ARC_EVICT_ALL); 2753 2754 mls = multilist_sublist_lock(ml, idx); 2755 2756 for (hdr = multilist_sublist_prev(mls, marker); hdr != NULL; 2757 hdr = multilist_sublist_prev(mls, marker)) { 2758 if ((bytes != ARC_EVICT_ALL && bytes_evicted >= bytes) || 2759 (evict_count >= zfs_arc_evict_batch_limit)) 2760 break; 2761 2762 /* 2763 * To keep our iteration location, move the marker 2764 * forward. Since we're not holding hdr's hash lock, we 2765 * must be very careful and not remove 'hdr' from the 2766 * sublist. Otherwise, other consumers might mistake the 2767 * 'hdr' as not being on a sublist when they call the 2768 * multilist_link_active() function (they all rely on 2769 * the hash lock protecting concurrent insertions and 2770 * removals). multilist_sublist_move_forward() was 2771 * specifically implemented to ensure this is the case 2772 * (only 'marker' will be removed and re-inserted). 2773 */ 2774 multilist_sublist_move_forward(mls, marker); 2775 2776 /* 2777 * The only case where the b_spa field should ever be 2778 * zero, is the marker headers inserted by 2779 * arc_evict_state(). It's possible for multiple threads 2780 * to be calling arc_evict_state() concurrently (e.g. 2781 * dsl_pool_close() and zio_inject_fault()), so we must 2782 * skip any markers we see from these other threads. 2783 */ 2784 if (hdr->b_spa == 0) 2785 continue; 2786 2787 /* we're only interested in evicting buffers of a certain spa */ 2788 if (spa != 0 && hdr->b_spa != spa) { 2789 ARCSTAT_BUMP(arcstat_evict_skip); 2790 continue; 2791 } 2792 2793 hash_lock = HDR_LOCK(hdr); 2794 2795 /* 2796 * We aren't calling this function from any code path 2797 * that would already be holding a hash lock, so we're 2798 * asserting on this assumption to be defensive in case 2799 * this ever changes. Without this check, it would be 2800 * possible to incorrectly increment arcstat_mutex_miss 2801 * below (e.g. if the code changed such that we called 2802 * this function with a hash lock held). 2803 */ 2804 ASSERT(!MUTEX_HELD(hash_lock)); 2805 2806 if (mutex_tryenter(hash_lock)) { 2807 uint64_t evicted = arc_evict_hdr(hdr, hash_lock); 2808 mutex_exit(hash_lock); 2809 2810 bytes_evicted += evicted; 2811 2812 /* 2813 * If evicted is zero, arc_evict_hdr() must have 2814 * decided to skip this header, don't increment 2815 * evict_count in this case. 2816 */ 2817 if (evicted != 0) 2818 evict_count++; 2819 2820 /* 2821 * If arc_size isn't overflowing, signal any 2822 * threads that might happen to be waiting. 2823 * 2824 * For each header evicted, we wake up a single 2825 * thread. If we used cv_broadcast, we could 2826 * wake up "too many" threads causing arc_size 2827 * to significantly overflow arc_c; since 2828 * arc_get_data_buf() doesn't check for overflow 2829 * when it's woken up (it doesn't because it's 2830 * possible for the ARC to be overflowing while 2831 * full of un-evictable buffers, and the 2832 * function should proceed in this case). 2833 * 2834 * If threads are left sleeping, due to not 2835 * using cv_broadcast, they will be woken up 2836 * just before arc_reclaim_thread() sleeps. 2837 */ 2838 mutex_enter(&arc_reclaim_lock); 2839 if (!arc_is_overflowing()) 2840 cv_signal(&arc_reclaim_waiters_cv); 2841 mutex_exit(&arc_reclaim_lock); 2842 } else { 2843 ARCSTAT_BUMP(arcstat_mutex_miss); 2844 } 2845 } 2846 2847 multilist_sublist_unlock(mls); 2848 2849 return (bytes_evicted); 2850} 2851 2852/* 2853 * Evict buffers from the given arc state, until we've removed the 2854 * specified number of bytes. Move the removed buffers to the 2855 * appropriate evict state. 2856 * 2857 * This function makes a "best effort". It skips over any buffers 2858 * it can't get a hash_lock on, and so, may not catch all candidates. 2859 * It may also return without evicting as much space as requested. 2860 * 2861 * If bytes is specified using the special value ARC_EVICT_ALL, this 2862 * will evict all available (i.e. unlocked and evictable) buffers from 2863 * the given arc state; which is used by arc_flush(). 2864 */ 2865static uint64_t 2866arc_evict_state(arc_state_t *state, uint64_t spa, int64_t bytes, 2867 arc_buf_contents_t type) 2868{ 2869 uint64_t total_evicted = 0; 2870 multilist_t *ml = &state->arcs_list[type]; 2871 int num_sublists; 2872 arc_buf_hdr_t **markers; 2873 2874 IMPLY(bytes < 0, bytes == ARC_EVICT_ALL); 2875 2876 num_sublists = multilist_get_num_sublists(ml); 2877 2878 /* 2879 * If we've tried to evict from each sublist, made some 2880 * progress, but still have not hit the target number of bytes 2881 * to evict, we want to keep trying. The markers allow us to 2882 * pick up where we left off for each individual sublist, rather 2883 * than starting from the tail each time. 2884 */ 2885 markers = kmem_zalloc(sizeof (*markers) * num_sublists, KM_SLEEP); 2886 for (int i = 0; i < num_sublists; i++) { 2887 markers[i] = kmem_cache_alloc(hdr_full_cache, KM_SLEEP); 2888 2889 /* 2890 * A b_spa of 0 is used to indicate that this header is 2891 * a marker. This fact is used in arc_adjust_type() and 2892 * arc_evict_state_impl(). 2893 */ 2894 markers[i]->b_spa = 0; 2895 2896 multilist_sublist_t *mls = multilist_sublist_lock(ml, i); 2897 multilist_sublist_insert_tail(mls, markers[i]); 2898 multilist_sublist_unlock(mls); 2899 } 2900 2901 /* 2902 * While we haven't hit our target number of bytes to evict, or 2903 * we're evicting all available buffers. 2904 */ 2905 while (total_evicted < bytes || bytes == ARC_EVICT_ALL) { 2906 /* 2907 * Start eviction using a randomly selected sublist, 2908 * this is to try and evenly balance eviction across all 2909 * sublists. Always starting at the same sublist 2910 * (e.g. index 0) would cause evictions to favor certain 2911 * sublists over others. 2912 */ 2913 int sublist_idx = multilist_get_random_index(ml); 2914 uint64_t scan_evicted = 0; 2915 2916 for (int i = 0; i < num_sublists; i++) { 2917 uint64_t bytes_remaining; 2918 uint64_t bytes_evicted; 2919 2920 if (bytes == ARC_EVICT_ALL) 2921 bytes_remaining = ARC_EVICT_ALL; 2922 else if (total_evicted < bytes) 2923 bytes_remaining = bytes - total_evicted; 2924 else 2925 break; 2926 2927 bytes_evicted = arc_evict_state_impl(ml, sublist_idx, 2928 markers[sublist_idx], spa, bytes_remaining); 2929 2930 scan_evicted += bytes_evicted; 2931 total_evicted += bytes_evicted; 2932 2933 /* we've reached the end, wrap to the beginning */ 2934 if (++sublist_idx >= num_sublists) 2935 sublist_idx = 0; 2936 } 2937 2938 /* 2939 * If we didn't evict anything during this scan, we have 2940 * no reason to believe we'll evict more during another 2941 * scan, so break the loop. 2942 */ 2943 if (scan_evicted == 0) { 2944 /* This isn't possible, let's make that obvious */ 2945 ASSERT3S(bytes, !=, 0); 2946 2947 /* 2948 * When bytes is ARC_EVICT_ALL, the only way to 2949 * break the loop is when scan_evicted is zero. 2950 * In that case, we actually have evicted enough, 2951 * so we don't want to increment the kstat. 2952 */ 2953 if (bytes != ARC_EVICT_ALL) { 2954 ASSERT3S(total_evicted, <, bytes); 2955 ARCSTAT_BUMP(arcstat_evict_not_enough); 2956 } 2957 2958 break; 2959 } 2960 } 2961 2962 for (int i = 0; i < num_sublists; i++) { 2963 multilist_sublist_t *mls = multilist_sublist_lock(ml, i); 2964 multilist_sublist_remove(mls, markers[i]); 2965 multilist_sublist_unlock(mls); 2966 2967 kmem_cache_free(hdr_full_cache, markers[i]); 2968 } 2969 kmem_free(markers, sizeof (*markers) * num_sublists); 2970 2971 return (total_evicted); 2972} 2973 2974/* 2975 * Flush all "evictable" data of the given type from the arc state 2976 * specified. This will not evict any "active" buffers (i.e. referenced). 2977 * 2978 * When 'retry' is set to FALSE, the function will make a single pass 2979 * over the state and evict any buffers that it can. Since it doesn't 2980 * continually retry the eviction, it might end up leaving some buffers 2981 * in the ARC due to lock misses. 2982 * 2983 * When 'retry' is set to TRUE, the function will continually retry the 2984 * eviction until *all* evictable buffers have been removed from the 2985 * state. As a result, if concurrent insertions into the state are 2986 * allowed (e.g. if the ARC isn't shutting down), this function might 2987 * wind up in an infinite loop, continually trying to evict buffers. 2988 */ 2989static uint64_t 2990arc_flush_state(arc_state_t *state, uint64_t spa, arc_buf_contents_t type, 2991 boolean_t retry) 2992{ 2993 uint64_t evicted = 0; 2994 2995 while (state->arcs_lsize[type] != 0) { 2996 evicted += arc_evict_state(state, spa, ARC_EVICT_ALL, type); 2997 2998 if (!retry) 2999 break; 3000 } 3001 3002 return (evicted); 3003} 3004 3005/* 3006 * Evict the specified number of bytes from the state specified, 3007 * restricting eviction to the spa and type given. This function 3008 * prevents us from trying to evict more from a state's list than 3009 * is "evictable", and to skip evicting altogether when passed a 3010 * negative value for "bytes". In contrast, arc_evict_state() will 3011 * evict everything it can, when passed a negative value for "bytes". 3012 */ 3013static uint64_t 3014arc_adjust_impl(arc_state_t *state, uint64_t spa, int64_t bytes, 3015 arc_buf_contents_t type) 3016{ 3017 int64_t delta; 3018 3019 if (bytes > 0 && state->arcs_lsize[type] > 0) { 3020 delta = MIN(state->arcs_lsize[type], bytes); 3021 return (arc_evict_state(state, spa, delta, type)); 3022 } 3023 3024 return (0); 3025} 3026 3027/* 3028 * Evict metadata buffers from the cache, such that arc_meta_used is 3029 * capped by the arc_meta_limit tunable. 3030 */ 3031static uint64_t 3032arc_adjust_meta(void) 3033{ 3034 uint64_t total_evicted = 0; 3035 int64_t target; 3036 3037 /* 3038 * If we're over the meta limit, we want to evict enough 3039 * metadata to get back under the meta limit. We don't want to 3040 * evict so much that we drop the MRU below arc_p, though. If 3041 * we're over the meta limit more than we're over arc_p, we 3042 * evict some from the MRU here, and some from the MFU below. 3043 */ 3044 target = MIN((int64_t)(arc_meta_used - arc_meta_limit), 3045 (int64_t)(refcount_count(&arc_anon->arcs_size) + 3046 refcount_count(&arc_mru->arcs_size) - arc_p)); 3047 3048 total_evicted += arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA); 3049 3050 /* 3051 * Similar to the above, we want to evict enough bytes to get us 3052 * below the meta limit, but not so much as to drop us below the 3053 * space alloted to the MFU (which is defined as arc_c - arc_p). 3054 */ 3055 target = MIN((int64_t)(arc_meta_used - arc_meta_limit), 3056 (int64_t)(refcount_count(&arc_mfu->arcs_size) - (arc_c - arc_p))); 3057 3058 total_evicted += arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA); 3059 3060 return (total_evicted); 3061} 3062 3063/* 3064 * Return the type of the oldest buffer in the given arc state 3065 * 3066 * This function will select a random sublist of type ARC_BUFC_DATA and 3067 * a random sublist of type ARC_BUFC_METADATA. The tail of each sublist 3068 * is compared, and the type which contains the "older" buffer will be 3069 * returned. 3070 */ 3071static arc_buf_contents_t 3072arc_adjust_type(arc_state_t *state) 3073{ 3074 multilist_t *data_ml = &state->arcs_list[ARC_BUFC_DATA]; 3075 multilist_t *meta_ml = &state->arcs_list[ARC_BUFC_METADATA]; 3076 int data_idx = multilist_get_random_index(data_ml); 3077 int meta_idx = multilist_get_random_index(meta_ml); 3078 multilist_sublist_t *data_mls; 3079 multilist_sublist_t *meta_mls; 3080 arc_buf_contents_t type; 3081 arc_buf_hdr_t *data_hdr; 3082 arc_buf_hdr_t *meta_hdr; 3083 3084 /* 3085 * We keep the sublist lock until we're finished, to prevent 3086 * the headers from being destroyed via arc_evict_state(). 3087 */ 3088 data_mls = multilist_sublist_lock(data_ml, data_idx); 3089 meta_mls = multilist_sublist_lock(meta_ml, meta_idx); 3090 3091 /* 3092 * These two loops are to ensure we skip any markers that 3093 * might be at the tail of the lists due to arc_evict_state(). 3094 */ 3095 3096 for (data_hdr = multilist_sublist_tail(data_mls); data_hdr != NULL; 3097 data_hdr = multilist_sublist_prev(data_mls, data_hdr)) { 3098 if (data_hdr->b_spa != 0) 3099 break; 3100 } 3101 3102 for (meta_hdr = multilist_sublist_tail(meta_mls); meta_hdr != NULL; 3103 meta_hdr = multilist_sublist_prev(meta_mls, meta_hdr)) { 3104 if (meta_hdr->b_spa != 0) 3105 break; 3106 } 3107 3108 if (data_hdr == NULL && meta_hdr == NULL) { 3109 type = ARC_BUFC_DATA; 3110 } else if (data_hdr == NULL) { 3111 ASSERT3P(meta_hdr, !=, NULL); 3112 type = ARC_BUFC_METADATA; 3113 } else if (meta_hdr == NULL) { 3114 ASSERT3P(data_hdr, !=, NULL); 3115 type = ARC_BUFC_DATA; 3116 } else { 3117 ASSERT3P(data_hdr, !=, NULL); 3118 ASSERT3P(meta_hdr, !=, NULL); 3119 3120 /* The headers can't be on the sublist without an L1 header */ 3121 ASSERT(HDR_HAS_L1HDR(data_hdr)); 3122 ASSERT(HDR_HAS_L1HDR(meta_hdr)); 3123 3124 if (data_hdr->b_l1hdr.b_arc_access < 3125 meta_hdr->b_l1hdr.b_arc_access) { 3126 type = ARC_BUFC_DATA; 3127 } else { 3128 type = ARC_BUFC_METADATA; 3129 } 3130 } 3131 3132 multilist_sublist_unlock(meta_mls); 3133 multilist_sublist_unlock(data_mls); 3134 3135 return (type); 3136} 3137 3138/* 3139 * Evict buffers from the cache, such that arc_size is capped by arc_c. 3140 */ 3141static uint64_t 3142arc_adjust(void) 3143{ 3144 uint64_t total_evicted = 0; 3145 uint64_t bytes; 3146 int64_t target; 3147 3148 /* 3149 * If we're over arc_meta_limit, we want to correct that before 3150 * potentially evicting data buffers below. 3151 */ 3152 total_evicted += arc_adjust_meta(); 3153 3154 /* 3155 * Adjust MRU size 3156 * 3157 * If we're over the target cache size, we want to evict enough 3158 * from the list to get back to our target size. We don't want 3159 * to evict too much from the MRU, such that it drops below 3160 * arc_p. So, if we're over our target cache size more than 3161 * the MRU is over arc_p, we'll evict enough to get back to 3162 * arc_p here, and then evict more from the MFU below. 3163 */ 3164 target = MIN((int64_t)(arc_size - arc_c), 3165 (int64_t)(refcount_count(&arc_anon->arcs_size) + 3166 refcount_count(&arc_mru->arcs_size) + arc_meta_used - arc_p)); 3167 3168 /* 3169 * If we're below arc_meta_min, always prefer to evict data. 3170 * Otherwise, try to satisfy the requested number of bytes to 3171 * evict from the type which contains older buffers; in an 3172 * effort to keep newer buffers in the cache regardless of their 3173 * type. If we cannot satisfy the number of bytes from this 3174 * type, spill over into the next type. 3175 */ 3176 if (arc_adjust_type(arc_mru) == ARC_BUFC_METADATA && 3177 arc_meta_used > arc_meta_min) { 3178 bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA); 3179 total_evicted += bytes; 3180 3181 /* 3182 * If we couldn't evict our target number of bytes from 3183 * metadata, we try to get the rest from data. 3184 */ 3185 target -= bytes; 3186 3187 total_evicted += 3188 arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA); 3189 } else { 3190 bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA); 3191 total_evicted += bytes; 3192 3193 /* 3194 * If we couldn't evict our target number of bytes from 3195 * data, we try to get the rest from metadata. 3196 */ 3197 target -= bytes; 3198 3199 total_evicted += 3200 arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA); 3201 } 3202 3203 /* 3204 * Adjust MFU size 3205 * 3206 * Now that we've tried to evict enough from the MRU to get its 3207 * size back to arc_p, if we're still above the target cache 3208 * size, we evict the rest from the MFU. 3209 */ 3210 target = arc_size - arc_c; 3211 3212 if (arc_adjust_type(arc_mfu) == ARC_BUFC_METADATA && 3213 arc_meta_used > arc_meta_min) { 3214 bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA); 3215 total_evicted += bytes; 3216 3217 /* 3218 * If we couldn't evict our target number of bytes from 3219 * metadata, we try to get the rest from data. 3220 */ 3221 target -= bytes; 3222 3223 total_evicted += 3224 arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA); 3225 } else { 3226 bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA); 3227 total_evicted += bytes; 3228 3229 /* 3230 * If we couldn't evict our target number of bytes from 3231 * data, we try to get the rest from data. 3232 */ 3233 target -= bytes; 3234 3235 total_evicted += 3236 arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA); 3237 } 3238 3239 /* 3240 * Adjust ghost lists 3241 * 3242 * In addition to the above, the ARC also defines target values 3243 * for the ghost lists. The sum of the mru list and mru ghost 3244 * list should never exceed the target size of the cache, and 3245 * the sum of the mru list, mfu list, mru ghost list, and mfu 3246 * ghost list should never exceed twice the target size of the 3247 * cache. The following logic enforces these limits on the ghost 3248 * caches, and evicts from them as needed. 3249 */ 3250 target = refcount_count(&arc_mru->arcs_size) + 3251 refcount_count(&arc_mru_ghost->arcs_size) - arc_c; 3252 3253 bytes = arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_DATA); 3254 total_evicted += bytes; 3255 3256 target -= bytes; 3257 3258 total_evicted += 3259 arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_METADATA); 3260 3261 /* 3262 * We assume the sum of the mru list and mfu list is less than 3263 * or equal to arc_c (we enforced this above), which means we 3264 * can use the simpler of the two equations below: 3265 * 3266 * mru + mfu + mru ghost + mfu ghost <= 2 * arc_c 3267 * mru ghost + mfu ghost <= arc_c 3268 */ 3269 target = refcount_count(&arc_mru_ghost->arcs_size) + 3270 refcount_count(&arc_mfu_ghost->arcs_size) - arc_c; 3271 3272 bytes = arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_DATA); 3273 total_evicted += bytes; 3274 3275 target -= bytes; 3276 3277 total_evicted += 3278 arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_METADATA); 3279 3280 return (total_evicted); 3281} 3282 3283static void 3284arc_do_user_evicts(void) 3285{ 3286 mutex_enter(&arc_user_evicts_lock); 3287 while (arc_eviction_list != NULL) { 3288 arc_buf_t *buf = arc_eviction_list; 3289 arc_eviction_list = buf->b_next; 3290 mutex_enter(&buf->b_evict_lock); 3291 buf->b_hdr = NULL; 3292 mutex_exit(&buf->b_evict_lock); 3293 mutex_exit(&arc_user_evicts_lock); 3294 3295 if (buf->b_efunc != NULL) 3296 VERIFY0(buf->b_efunc(buf->b_private)); 3297 3298 buf->b_efunc = NULL; 3299 buf->b_private = NULL; 3300 kmem_cache_free(buf_cache, buf); 3301 mutex_enter(&arc_user_evicts_lock); 3302 } 3303 mutex_exit(&arc_user_evicts_lock); 3304} 3305 3306void 3307arc_flush(spa_t *spa, boolean_t retry) 3308{ 3309 uint64_t guid = 0; 3310 3311 /* 3312 * If retry is TRUE, a spa must not be specified since we have 3313 * no good way to determine if all of a spa's buffers have been 3314 * evicted from an arc state. 3315 */ 3316 ASSERT(!retry || spa == 0); 3317 3318 if (spa != NULL) 3319 guid = spa_load_guid(spa); 3320 3321 (void) arc_flush_state(arc_mru, guid, ARC_BUFC_DATA, retry); 3322 (void) arc_flush_state(arc_mru, guid, ARC_BUFC_METADATA, retry); 3323 3324 (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_DATA, retry); 3325 (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_METADATA, retry); 3326 3327 (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_DATA, retry); 3328 (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_METADATA, retry); 3329 3330 (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_DATA, retry); 3331 (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_METADATA, retry); 3332 3333 arc_do_user_evicts(); 3334 ASSERT(spa || arc_eviction_list == NULL); 3335} 3336 3337void 3338arc_shrink(int64_t to_free) 3339{ 3340 if (arc_c > arc_c_min) { 3341 DTRACE_PROBE4(arc__shrink, uint64_t, arc_c, uint64_t, 3342 arc_c_min, uint64_t, arc_p, uint64_t, to_free); 3343 if (arc_c > arc_c_min + to_free) 3344 atomic_add_64(&arc_c, -to_free); 3345 else 3346 arc_c = arc_c_min; 3347 3348 atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift)); 3349 if (arc_c > arc_size) 3350 arc_c = MAX(arc_size, arc_c_min); 3351 if (arc_p > arc_c) 3352 arc_p = (arc_c >> 1); 3353 3354 DTRACE_PROBE2(arc__shrunk, uint64_t, arc_c, uint64_t, 3355 arc_p); 3356 3357 ASSERT(arc_c >= arc_c_min); 3358 ASSERT((int64_t)arc_p >= 0); 3359 } 3360 3361 if (arc_size > arc_c) { 3362 DTRACE_PROBE2(arc__shrink_adjust, uint64_t, arc_size, 3363 uint64_t, arc_c); 3364 (void) arc_adjust(); 3365 } 3366} 3367 3368static long needfree = 0; 3369 3370typedef enum free_memory_reason_t { 3371 FMR_UNKNOWN, 3372 FMR_NEEDFREE, 3373 FMR_LOTSFREE, 3374 FMR_SWAPFS_MINFREE, 3375 FMR_PAGES_PP_MAXIMUM, 3376 FMR_HEAP_ARENA, 3377 FMR_ZIO_ARENA, 3378 FMR_ZIO_FRAG, 3379} free_memory_reason_t; 3380 3381int64_t last_free_memory; 3382free_memory_reason_t last_free_reason; 3383 3384/* 3385 * Additional reserve of pages for pp_reserve. 3386 */ 3387int64_t arc_pages_pp_reserve = 64; 3388 3389/* 3390 * Additional reserve of pages for swapfs. 3391 */ 3392int64_t arc_swapfs_reserve = 64; 3393 3394/* 3395 * Return the amount of memory that can be consumed before reclaim will be 3396 * needed. Positive if there is sufficient free memory, negative indicates 3397 * the amount of memory that needs to be freed up. 3398 */ 3399static int64_t 3400arc_available_memory(void) 3401{ 3402 int64_t lowest = INT64_MAX; 3403 int64_t n; 3404 free_memory_reason_t r = FMR_UNKNOWN; 3405 3406#ifdef _KERNEL 3407 if (needfree > 0) { 3408 n = PAGESIZE * (-needfree); 3409 if (n < lowest) { 3410 lowest = n; 3411 r = FMR_NEEDFREE; 3412 } 3413 } 3414 3415 /* 3416 * Cooperate with pagedaemon when it's time for it to scan 3417 * and reclaim some pages. 3418 */ 3419 n = PAGESIZE * ((int64_t)freemem - zfs_arc_free_target); 3420 if (n < lowest) { 3421 lowest = n; 3422 r = FMR_LOTSFREE; 3423 } 3424 3425#ifdef illumos 3426 /* 3427 * check that we're out of range of the pageout scanner. It starts to 3428 * schedule paging if freemem is less than lotsfree and needfree. 3429 * lotsfree is the high-water mark for pageout, and needfree is the 3430 * number of needed free pages. We add extra pages here to make sure 3431 * the scanner doesn't start up while we're freeing memory. 3432 */ 3433 n = PAGESIZE * (freemem - lotsfree - needfree - desfree); 3434 if (n < lowest) { 3435 lowest = n; 3436 r = FMR_LOTSFREE; 3437 } 3438 3439 /* 3440 * check to make sure that swapfs has enough space so that anon 3441 * reservations can still succeed. anon_resvmem() checks that the 3442 * availrmem is greater than swapfs_minfree, and the number of reserved 3443 * swap pages. We also add a bit of extra here just to prevent 3444 * circumstances from getting really dire. 3445 */ 3446 n = PAGESIZE * (availrmem - swapfs_minfree - swapfs_reserve - 3447 desfree - arc_swapfs_reserve); 3448 if (n < lowest) { 3449 lowest = n; 3450 r = FMR_SWAPFS_MINFREE; 3451 } 3452 3453 3454 /* 3455 * Check that we have enough availrmem that memory locking (e.g., via 3456 * mlock(3C) or memcntl(2)) can still succeed. (pages_pp_maximum 3457 * stores the number of pages that cannot be locked; when availrmem 3458 * drops below pages_pp_maximum, page locking mechanisms such as 3459 * page_pp_lock() will fail.) 3460 */ 3461 n = PAGESIZE * (availrmem - pages_pp_maximum - 3462 arc_pages_pp_reserve); 3463 if (n < lowest) { 3464 lowest = n; 3465 r = FMR_PAGES_PP_MAXIMUM; 3466 } 3467 3468#endif /* illumos */ 3469#if defined(__i386) || !defined(UMA_MD_SMALL_ALLOC) 3470 /* 3471 * If we're on an i386 platform, it's possible that we'll exhaust the 3472 * kernel heap space before we ever run out of available physical 3473 * memory. Most checks of the size of the heap_area compare against 3474 * tune.t_minarmem, which is the minimum available real memory that we 3475 * can have in the system. However, this is generally fixed at 25 pages 3476 * which is so low that it's useless. In this comparison, we seek to 3477 * calculate the total heap-size, and reclaim if more than 3/4ths of the 3478 * heap is allocated. (Or, in the calculation, if less than 1/4th is 3479 * free) 3480 */ 3481 n = (int64_t)vmem_size(heap_arena, VMEM_FREE) - 3482 (vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC) >> 2); 3483 if (n < lowest) { 3484 lowest = n; 3485 r = FMR_HEAP_ARENA; 3486 } 3487#define zio_arena NULL 3488#else 3489#define zio_arena heap_arena 3490#endif 3491 3492 /* 3493 * If zio data pages are being allocated out of a separate heap segment, 3494 * then enforce that the size of available vmem for this arena remains 3495 * above about 1/16th free. 3496 * 3497 * Note: The 1/16th arena free requirement was put in place 3498 * to aggressively evict memory from the arc in order to avoid 3499 * memory fragmentation issues. 3500 */ 3501 if (zio_arena != NULL) { 3502 n = (int64_t)vmem_size(zio_arena, VMEM_FREE) - 3503 (vmem_size(zio_arena, VMEM_ALLOC) >> 4); 3504 if (n < lowest) { 3505 lowest = n; 3506 r = FMR_ZIO_ARENA; 3507 } 3508 } 3509 3510 /* 3511 * Above limits know nothing about real level of KVA fragmentation. 3512 * Start aggressive reclamation if too little sequential KVA left. 3513 */ 3514 if (lowest > 0) { 3515 n = (vmem_size(heap_arena, VMEM_MAXFREE) < zfs_max_recordsize) ? 3516 -((int64_t)vmem_size(heap_arena, VMEM_ALLOC) >> 4) : 3517 INT64_MAX; 3518 if (n < lowest) { 3519 lowest = n; 3520 r = FMR_ZIO_FRAG; 3521 } 3522 } 3523 3524#else /* _KERNEL */ 3525 /* Every 100 calls, free a small amount */ 3526 if (spa_get_random(100) == 0) 3527 lowest = -1024; 3528#endif /* _KERNEL */ 3529 3530 last_free_memory = lowest; 3531 last_free_reason = r; 3532 DTRACE_PROBE2(arc__available_memory, int64_t, lowest, int, r); 3533 return (lowest); 3534} 3535 3536 3537/* 3538 * Determine if the system is under memory pressure and is asking 3539 * to reclaim memory. A return value of TRUE indicates that the system 3540 * is under memory pressure and that the arc should adjust accordingly. 3541 */ 3542static boolean_t 3543arc_reclaim_needed(void) 3544{ 3545 return (arc_available_memory() < 0); 3546} 3547 3548extern kmem_cache_t *zio_buf_cache[]; 3549extern kmem_cache_t *zio_data_buf_cache[]; 3550extern kmem_cache_t *range_seg_cache; 3551 3552static __noinline void 3553arc_kmem_reap_now(void) 3554{ 3555 size_t i; 3556 kmem_cache_t *prev_cache = NULL; 3557 kmem_cache_t *prev_data_cache = NULL; 3558 3559 DTRACE_PROBE(arc__kmem_reap_start); 3560#ifdef _KERNEL 3561 if (arc_meta_used >= arc_meta_limit) { 3562 /* 3563 * We are exceeding our meta-data cache limit. 3564 * Purge some DNLC entries to release holds on meta-data. 3565 */ 3566 dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent); 3567 } 3568#if defined(__i386) 3569 /* 3570 * Reclaim unused memory from all kmem caches. 3571 */ 3572 kmem_reap(); 3573#endif 3574#endif 3575 3576 for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) { 3577 if (zio_buf_cache[i] != prev_cache) { 3578 prev_cache = zio_buf_cache[i]; 3579 kmem_cache_reap_now(zio_buf_cache[i]); 3580 } 3581 if (zio_data_buf_cache[i] != prev_data_cache) { 3582 prev_data_cache = zio_data_buf_cache[i]; 3583 kmem_cache_reap_now(zio_data_buf_cache[i]); 3584 } 3585 } 3586 kmem_cache_reap_now(buf_cache); 3587 kmem_cache_reap_now(hdr_full_cache); 3588 kmem_cache_reap_now(hdr_l2only_cache); 3589 kmem_cache_reap_now(range_seg_cache); 3590 3591#ifdef illumos 3592 if (zio_arena != NULL) { 3593 /* 3594 * Ask the vmem arena to reclaim unused memory from its 3595 * quantum caches. 3596 */ 3597 vmem_qcache_reap(zio_arena); 3598 } 3599#endif 3600 DTRACE_PROBE(arc__kmem_reap_end); 3601} 3602 3603/* 3604 * Threads can block in arc_get_data_buf() waiting for this thread to evict 3605 * enough data and signal them to proceed. When this happens, the threads in 3606 * arc_get_data_buf() are sleeping while holding the hash lock for their 3607 * particular arc header. Thus, we must be careful to never sleep on a 3608 * hash lock in this thread. This is to prevent the following deadlock: 3609 * 3610 * - Thread A sleeps on CV in arc_get_data_buf() holding hash lock "L", 3611 * waiting for the reclaim thread to signal it. 3612 * 3613 * - arc_reclaim_thread() tries to acquire hash lock "L" using mutex_enter, 3614 * fails, and goes to sleep forever. 3615 * 3616 * This possible deadlock is avoided by always acquiring a hash lock 3617 * using mutex_tryenter() from arc_reclaim_thread(). 3618 */ 3619static void 3620arc_reclaim_thread(void *dummy __unused) 3621{ 3622 hrtime_t growtime = 0; 3623 callb_cpr_t cpr; 3624 3625 CALLB_CPR_INIT(&cpr, &arc_reclaim_lock, callb_generic_cpr, FTAG); 3626 3627 mutex_enter(&arc_reclaim_lock); 3628 while (!arc_reclaim_thread_exit) { 3629 int64_t free_memory = arc_available_memory(); 3630 uint64_t evicted = 0; 3631 3632 mutex_exit(&arc_reclaim_lock); 3633 3634 if (free_memory < 0) { 3635 3636 arc_no_grow = B_TRUE; 3637 arc_warm = B_TRUE; 3638 3639 /* 3640 * Wait at least zfs_grow_retry (default 60) seconds 3641 * before considering growing. 3642 */ 3643 growtime = gethrtime() + SEC2NSEC(arc_grow_retry); 3644 3645 arc_kmem_reap_now(); 3646 3647 /* 3648 * If we are still low on memory, shrink the ARC 3649 * so that we have arc_shrink_min free space. 3650 */ 3651 free_memory = arc_available_memory(); 3652 3653 int64_t to_free = 3654 (arc_c >> arc_shrink_shift) - free_memory; 3655 if (to_free > 0) { 3656#ifdef _KERNEL 3657 to_free = MAX(to_free, ptob(needfree)); 3658#endif 3659 arc_shrink(to_free); 3660 } 3661 } else if (free_memory < arc_c >> arc_no_grow_shift) { 3662 arc_no_grow = B_TRUE; 3663 } else if (gethrtime() >= growtime) { 3664 arc_no_grow = B_FALSE; 3665 } 3666 3667 evicted = arc_adjust(); 3668 3669 mutex_enter(&arc_reclaim_lock); 3670 3671 /* 3672 * If evicted is zero, we couldn't evict anything via 3673 * arc_adjust(). This could be due to hash lock 3674 * collisions, but more likely due to the majority of 3675 * arc buffers being unevictable. Therefore, even if 3676 * arc_size is above arc_c, another pass is unlikely to 3677 * be helpful and could potentially cause us to enter an 3678 * infinite loop. 3679 */ 3680 if (arc_size <= arc_c || evicted == 0) { 3681#ifdef _KERNEL 3682 needfree = 0; 3683#endif 3684 /* 3685 * We're either no longer overflowing, or we 3686 * can't evict anything more, so we should wake 3687 * up any threads before we go to sleep. 3688 */ 3689 cv_broadcast(&arc_reclaim_waiters_cv); 3690 3691 /* 3692 * Block until signaled, or after one second (we 3693 * might need to perform arc_kmem_reap_now() 3694 * even if we aren't being signalled) 3695 */ 3696 CALLB_CPR_SAFE_BEGIN(&cpr); 3697 (void) cv_timedwait_hires(&arc_reclaim_thread_cv, 3698 &arc_reclaim_lock, SEC2NSEC(1), MSEC2NSEC(1), 0); 3699 CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_lock); 3700 } 3701 } 3702 3703 arc_reclaim_thread_exit = FALSE; 3704 cv_broadcast(&arc_reclaim_thread_cv); 3705 CALLB_CPR_EXIT(&cpr); /* drops arc_reclaim_lock */ 3706 thread_exit(); 3707} 3708 3709static void 3710arc_user_evicts_thread(void *dummy __unused) 3711{ 3712 callb_cpr_t cpr; 3713 3714 CALLB_CPR_INIT(&cpr, &arc_user_evicts_lock, callb_generic_cpr, FTAG); 3715 3716 mutex_enter(&arc_user_evicts_lock); 3717 while (!arc_user_evicts_thread_exit) { 3718 mutex_exit(&arc_user_evicts_lock); 3719 3720 arc_do_user_evicts(); 3721 3722 /* 3723 * This is necessary in order for the mdb ::arc dcmd to 3724 * show up to date information. Since the ::arc command 3725 * does not call the kstat's update function, without 3726 * this call, the command may show stale stats for the 3727 * anon, mru, mru_ghost, mfu, and mfu_ghost lists. Even 3728 * with this change, the data might be up to 1 second 3729 * out of date; but that should suffice. The arc_state_t 3730 * structures can be queried directly if more accurate 3731 * information is needed. 3732 */ 3733 if (arc_ksp != NULL) 3734 arc_ksp->ks_update(arc_ksp, KSTAT_READ); 3735 3736 mutex_enter(&arc_user_evicts_lock); 3737 3738 /* 3739 * Block until signaled, or after one second (we need to 3740 * call the arc's kstat update function regularly). 3741 */ 3742 CALLB_CPR_SAFE_BEGIN(&cpr); 3743 (void) cv_timedwait(&arc_user_evicts_cv, 3744 &arc_user_evicts_lock, hz); 3745 CALLB_CPR_SAFE_END(&cpr, &arc_user_evicts_lock); 3746 } 3747 3748 arc_user_evicts_thread_exit = FALSE; 3749 cv_broadcast(&arc_user_evicts_cv); 3750 CALLB_CPR_EXIT(&cpr); /* drops arc_user_evicts_lock */ 3751 thread_exit(); 3752} 3753 3754/* 3755 * Adapt arc info given the number of bytes we are trying to add and 3756 * the state that we are comming from. This function is only called 3757 * when we are adding new content to the cache. 3758 */ 3759static void 3760arc_adapt(int bytes, arc_state_t *state) 3761{ 3762 int mult; 3763 uint64_t arc_p_min = (arc_c >> arc_p_min_shift); 3764 int64_t mrug_size = refcount_count(&arc_mru_ghost->arcs_size); 3765 int64_t mfug_size = refcount_count(&arc_mfu_ghost->arcs_size); 3766 3767 if (state == arc_l2c_only) 3768 return; 3769 3770 ASSERT(bytes > 0); 3771 /* 3772 * Adapt the target size of the MRU list: 3773 * - if we just hit in the MRU ghost list, then increase 3774 * the target size of the MRU list. 3775 * - if we just hit in the MFU ghost list, then increase 3776 * the target size of the MFU list by decreasing the 3777 * target size of the MRU list. 3778 */ 3779 if (state == arc_mru_ghost) { 3780 mult = (mrug_size >= mfug_size) ? 1 : (mfug_size / mrug_size); 3781 mult = MIN(mult, 10); /* avoid wild arc_p adjustment */ 3782 3783 arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult); 3784 } else if (state == arc_mfu_ghost) { 3785 uint64_t delta; 3786 3787 mult = (mfug_size >= mrug_size) ? 1 : (mrug_size / mfug_size); 3788 mult = MIN(mult, 10); 3789 3790 delta = MIN(bytes * mult, arc_p); 3791 arc_p = MAX(arc_p_min, arc_p - delta); 3792 } 3793 ASSERT((int64_t)arc_p >= 0); 3794 3795 if (arc_reclaim_needed()) { 3796 cv_signal(&arc_reclaim_thread_cv); 3797 return; 3798 } 3799 3800 if (arc_no_grow) 3801 return; 3802 3803 if (arc_c >= arc_c_max) 3804 return; 3805 3806 /* 3807 * If we're within (2 * maxblocksize) bytes of the target 3808 * cache size, increment the target cache size 3809 */ 3810 if (arc_size > arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) { 3811 DTRACE_PROBE1(arc__inc_adapt, int, bytes); 3812 atomic_add_64(&arc_c, (int64_t)bytes); 3813 if (arc_c > arc_c_max) 3814 arc_c = arc_c_max; 3815 else if (state == arc_anon) 3816 atomic_add_64(&arc_p, (int64_t)bytes); 3817 if (arc_p > arc_c) 3818 arc_p = arc_c; 3819 } 3820 ASSERT((int64_t)arc_p >= 0); 3821} 3822 3823/* 3824 * Check if arc_size has grown past our upper threshold, determined by 3825 * zfs_arc_overflow_shift. 3826 */ 3827static boolean_t 3828arc_is_overflowing(void) 3829{ 3830 /* Always allow at least one block of overflow */ 3831 uint64_t overflow = MAX(SPA_MAXBLOCKSIZE, 3832 arc_c >> zfs_arc_overflow_shift); 3833 3834 return (arc_size >= arc_c + overflow); 3835} 3836 3837/* 3838 * The buffer, supplied as the first argument, needs a data block. If we 3839 * are hitting the hard limit for the cache size, we must sleep, waiting 3840 * for the eviction thread to catch up. If we're past the target size 3841 * but below the hard limit, we'll only signal the reclaim thread and 3842 * continue on. 3843 */ 3844static void 3845arc_get_data_buf(arc_buf_t *buf) 3846{ 3847 arc_state_t *state = buf->b_hdr->b_l1hdr.b_state; 3848 uint64_t size = buf->b_hdr->b_size; 3849 arc_buf_contents_t type = arc_buf_type(buf->b_hdr); 3850 3851 arc_adapt(size, state); 3852 3853 /* 3854 * If arc_size is currently overflowing, and has grown past our 3855 * upper limit, we must be adding data faster than the evict 3856 * thread can evict. Thus, to ensure we don't compound the 3857 * problem by adding more data and forcing arc_size to grow even 3858 * further past it's target size, we halt and wait for the 3859 * eviction thread to catch up. 3860 * 3861 * It's also possible that the reclaim thread is unable to evict 3862 * enough buffers to get arc_size below the overflow limit (e.g. 3863 * due to buffers being un-evictable, or hash lock collisions). 3864 * In this case, we want to proceed regardless if we're 3865 * overflowing; thus we don't use a while loop here. 3866 */ 3867 if (arc_is_overflowing()) { 3868 mutex_enter(&arc_reclaim_lock); 3869 3870 /* 3871 * Now that we've acquired the lock, we may no longer be 3872 * over the overflow limit, lets check. 3873 * 3874 * We're ignoring the case of spurious wake ups. If that 3875 * were to happen, it'd let this thread consume an ARC 3876 * buffer before it should have (i.e. before we're under 3877 * the overflow limit and were signalled by the reclaim 3878 * thread). As long as that is a rare occurrence, it 3879 * shouldn't cause any harm. 3880 */ 3881 if (arc_is_overflowing()) { 3882 cv_signal(&arc_reclaim_thread_cv); 3883 cv_wait(&arc_reclaim_waiters_cv, &arc_reclaim_lock); 3884 } 3885 3886 mutex_exit(&arc_reclaim_lock); 3887 } 3888 3889 if (type == ARC_BUFC_METADATA) { 3890 buf->b_data = zio_buf_alloc(size); 3891 arc_space_consume(size, ARC_SPACE_META); 3892 } else { 3893 ASSERT(type == ARC_BUFC_DATA); 3894 buf->b_data = zio_data_buf_alloc(size); 3895 arc_space_consume(size, ARC_SPACE_DATA); 3896 } 3897 3898 /* 3899 * Update the state size. Note that ghost states have a 3900 * "ghost size" and so don't need to be updated. 3901 */ 3902 if (!GHOST_STATE(buf->b_hdr->b_l1hdr.b_state)) { 3903 arc_buf_hdr_t *hdr = buf->b_hdr; 3904 arc_state_t *state = hdr->b_l1hdr.b_state; 3905 3906 (void) refcount_add_many(&state->arcs_size, size, buf); 3907 3908 /* 3909 * If this is reached via arc_read, the link is 3910 * protected by the hash lock. If reached via 3911 * arc_buf_alloc, the header should not be accessed by 3912 * any other thread. And, if reached via arc_read_done, 3913 * the hash lock will protect it if it's found in the 3914 * hash table; otherwise no other thread should be 3915 * trying to [add|remove]_reference it. 3916 */ 3917 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) { 3918 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 3919 atomic_add_64(&hdr->b_l1hdr.b_state->arcs_lsize[type], 3920 size); 3921 } 3922 /* 3923 * If we are growing the cache, and we are adding anonymous 3924 * data, and we have outgrown arc_p, update arc_p 3925 */ 3926 if (arc_size < arc_c && hdr->b_l1hdr.b_state == arc_anon && 3927 (refcount_count(&arc_anon->arcs_size) + 3928 refcount_count(&arc_mru->arcs_size) > arc_p)) 3929 arc_p = MIN(arc_c, arc_p + size); 3930 } 3931 ARCSTAT_BUMP(arcstat_allocated); 3932} 3933 3934/* 3935 * This routine is called whenever a buffer is accessed. 3936 * NOTE: the hash lock is dropped in this function. 3937 */ 3938static void 3939arc_access(arc_buf_hdr_t *hdr, kmutex_t *hash_lock) 3940{ 3941 clock_t now; 3942 3943 ASSERT(MUTEX_HELD(hash_lock)); 3944 ASSERT(HDR_HAS_L1HDR(hdr)); 3945 3946 if (hdr->b_l1hdr.b_state == arc_anon) { 3947 /* 3948 * This buffer is not in the cache, and does not 3949 * appear in our "ghost" list. Add the new buffer 3950 * to the MRU state. 3951 */ 3952 3953 ASSERT0(hdr->b_l1hdr.b_arc_access); 3954 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 3955 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr); 3956 arc_change_state(arc_mru, hdr, hash_lock); 3957 3958 } else if (hdr->b_l1hdr.b_state == arc_mru) { 3959 now = ddi_get_lbolt(); 3960 3961 /* 3962 * If this buffer is here because of a prefetch, then either: 3963 * - clear the flag if this is a "referencing" read 3964 * (any subsequent access will bump this into the MFU state). 3965 * or 3966 * - move the buffer to the head of the list if this is 3967 * another prefetch (to make it less likely to be evicted). 3968 */ 3969 if (HDR_PREFETCH(hdr)) { 3970 if (refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) { 3971 /* link protected by hash lock */ 3972 ASSERT(multilist_link_active( 3973 &hdr->b_l1hdr.b_arc_node)); 3974 } else { 3975 hdr->b_flags &= ~ARC_FLAG_PREFETCH; 3976 ARCSTAT_BUMP(arcstat_mru_hits); 3977 } 3978 hdr->b_l1hdr.b_arc_access = now; 3979 return; 3980 } 3981 3982 /* 3983 * This buffer has been "accessed" only once so far, 3984 * but it is still in the cache. Move it to the MFU 3985 * state. 3986 */ 3987 if (now > hdr->b_l1hdr.b_arc_access + ARC_MINTIME) { 3988 /* 3989 * More than 125ms have passed since we 3990 * instantiated this buffer. Move it to the 3991 * most frequently used state. 3992 */ 3993 hdr->b_l1hdr.b_arc_access = now; 3994 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 3995 arc_change_state(arc_mfu, hdr, hash_lock); 3996 } 3997 ARCSTAT_BUMP(arcstat_mru_hits); 3998 } else if (hdr->b_l1hdr.b_state == arc_mru_ghost) { 3999 arc_state_t *new_state; 4000 /* 4001 * This buffer has been "accessed" recently, but 4002 * was evicted from the cache. Move it to the 4003 * MFU state. 4004 */ 4005 4006 if (HDR_PREFETCH(hdr)) { 4007 new_state = arc_mru; 4008 if (refcount_count(&hdr->b_l1hdr.b_refcnt) > 0) 4009 hdr->b_flags &= ~ARC_FLAG_PREFETCH; 4010 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr); 4011 } else { 4012 new_state = arc_mfu; 4013 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 4014 } 4015 4016 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 4017 arc_change_state(new_state, hdr, hash_lock); 4018 4019 ARCSTAT_BUMP(arcstat_mru_ghost_hits); 4020 } else if (hdr->b_l1hdr.b_state == arc_mfu) { 4021 /* 4022 * This buffer has been accessed more than once and is 4023 * still in the cache. Keep it in the MFU state. 4024 * 4025 * NOTE: an add_reference() that occurred when we did 4026 * the arc_read() will have kicked this off the list. 4027 * If it was a prefetch, we will explicitly move it to 4028 * the head of the list now. 4029 */ 4030 if ((HDR_PREFETCH(hdr)) != 0) { 4031 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 4032 /* link protected by hash_lock */ 4033 ASSERT(multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 4034 } 4035 ARCSTAT_BUMP(arcstat_mfu_hits); 4036 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 4037 } else if (hdr->b_l1hdr.b_state == arc_mfu_ghost) { 4038 arc_state_t *new_state = arc_mfu; 4039 /* 4040 * This buffer has been accessed more than once but has 4041 * been evicted from the cache. Move it back to the 4042 * MFU state. 4043 */ 4044 4045 if (HDR_PREFETCH(hdr)) { 4046 /* 4047 * This is a prefetch access... 4048 * move this block back to the MRU state. 4049 */ 4050 ASSERT0(refcount_count(&hdr->b_l1hdr.b_refcnt)); 4051 new_state = arc_mru; 4052 } 4053 4054 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 4055 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 4056 arc_change_state(new_state, hdr, hash_lock); 4057 4058 ARCSTAT_BUMP(arcstat_mfu_ghost_hits); 4059 } else if (hdr->b_l1hdr.b_state == arc_l2c_only) { 4060 /* 4061 * This buffer is on the 2nd Level ARC. 4062 */ 4063 4064 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 4065 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 4066 arc_change_state(arc_mfu, hdr, hash_lock); 4067 } else { 4068 ASSERT(!"invalid arc state"); 4069 } 4070} 4071 4072/* a generic arc_done_func_t which you can use */ 4073/* ARGSUSED */ 4074void 4075arc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg) 4076{ 4077 if (zio == NULL || zio->io_error == 0) 4078 bcopy(buf->b_data, arg, buf->b_hdr->b_size); 4079 VERIFY(arc_buf_remove_ref(buf, arg)); 4080} 4081 4082/* a generic arc_done_func_t */ 4083void 4084arc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg) 4085{ 4086 arc_buf_t **bufp = arg; 4087 if (zio && zio->io_error) { 4088 VERIFY(arc_buf_remove_ref(buf, arg)); 4089 *bufp = NULL; 4090 } else { 4091 *bufp = buf; 4092 ASSERT(buf->b_data); 4093 } 4094} 4095 4096static void 4097arc_read_done(zio_t *zio) 4098{ 4099 arc_buf_hdr_t *hdr; 4100 arc_buf_t *buf; 4101 arc_buf_t *abuf; /* buffer we're assigning to callback */ 4102 kmutex_t *hash_lock = NULL; 4103 arc_callback_t *callback_list, *acb; 4104 int freeable = FALSE; 4105 4106 buf = zio->io_private; 4107 hdr = buf->b_hdr; 4108 4109 /* 4110 * The hdr was inserted into hash-table and removed from lists 4111 * prior to starting I/O. We should find this header, since 4112 * it's in the hash table, and it should be legit since it's 4113 * not possible to evict it during the I/O. The only possible 4114 * reason for it not to be found is if we were freed during the 4115 * read. 4116 */ 4117 if (HDR_IN_HASH_TABLE(hdr)) { 4118 ASSERT3U(hdr->b_birth, ==, BP_PHYSICAL_BIRTH(zio->io_bp)); 4119 ASSERT3U(hdr->b_dva.dva_word[0], ==, 4120 BP_IDENTITY(zio->io_bp)->dva_word[0]); 4121 ASSERT3U(hdr->b_dva.dva_word[1], ==, 4122 BP_IDENTITY(zio->io_bp)->dva_word[1]); 4123 4124 arc_buf_hdr_t *found = buf_hash_find(hdr->b_spa, zio->io_bp, 4125 &hash_lock); 4126 4127 ASSERT((found == NULL && HDR_FREED_IN_READ(hdr) && 4128 hash_lock == NULL) || 4129 (found == hdr && 4130 DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) || 4131 (found == hdr && HDR_L2_READING(hdr))); 4132 } 4133 4134 hdr->b_flags &= ~ARC_FLAG_L2_EVICTED; 4135 if (l2arc_noprefetch && HDR_PREFETCH(hdr)) 4136 hdr->b_flags &= ~ARC_FLAG_L2CACHE; 4137 4138 /* byteswap if necessary */ 4139 callback_list = hdr->b_l1hdr.b_acb; 4140 ASSERT(callback_list != NULL); 4141 if (BP_SHOULD_BYTESWAP(zio->io_bp) && zio->io_error == 0) { 4142 dmu_object_byteswap_t bswap = 4143 DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp)); 4144 arc_byteswap_func_t *func = BP_GET_LEVEL(zio->io_bp) > 0 ? 4145 byteswap_uint64_array : 4146 dmu_ot_byteswap[bswap].ob_func; 4147 func(buf->b_data, hdr->b_size); 4148 } 4149 4150 arc_cksum_compute(buf, B_FALSE); 4151#ifdef illumos 4152 arc_buf_watch(buf); 4153#endif 4154 4155 if (hash_lock && zio->io_error == 0 && 4156 hdr->b_l1hdr.b_state == arc_anon) { 4157 /* 4158 * Only call arc_access on anonymous buffers. This is because 4159 * if we've issued an I/O for an evicted buffer, we've already 4160 * called arc_access (to prevent any simultaneous readers from 4161 * getting confused). 4162 */ 4163 arc_access(hdr, hash_lock); 4164 } 4165 4166 /* create copies of the data buffer for the callers */ 4167 abuf = buf; 4168 for (acb = callback_list; acb; acb = acb->acb_next) { 4169 if (acb->acb_done) { 4170 if (abuf == NULL) { 4171 ARCSTAT_BUMP(arcstat_duplicate_reads); 4172 abuf = arc_buf_clone(buf); 4173 } 4174 acb->acb_buf = abuf; 4175 abuf = NULL; 4176 } 4177 } 4178 hdr->b_l1hdr.b_acb = NULL; 4179 hdr->b_flags &= ~ARC_FLAG_IO_IN_PROGRESS; 4180 ASSERT(!HDR_BUF_AVAILABLE(hdr)); 4181 if (abuf == buf) { 4182 ASSERT(buf->b_efunc == NULL); 4183 ASSERT(hdr->b_l1hdr.b_datacnt == 1); 4184 hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE; 4185 } 4186 4187 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt) || 4188 callback_list != NULL); 4189 4190 if (zio->io_error != 0) { 4191 hdr->b_flags |= ARC_FLAG_IO_ERROR; 4192 if (hdr->b_l1hdr.b_state != arc_anon) 4193 arc_change_state(arc_anon, hdr, hash_lock); 4194 if (HDR_IN_HASH_TABLE(hdr)) 4195 buf_hash_remove(hdr); 4196 freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt); 4197 } 4198 4199 /* 4200 * Broadcast before we drop the hash_lock to avoid the possibility 4201 * that the hdr (and hence the cv) might be freed before we get to 4202 * the cv_broadcast(). 4203 */ 4204 cv_broadcast(&hdr->b_l1hdr.b_cv); 4205 4206 if (hash_lock != NULL) { 4207 mutex_exit(hash_lock); 4208 } else { 4209 /* 4210 * This block was freed while we waited for the read to 4211 * complete. It has been removed from the hash table and 4212 * moved to the anonymous state (so that it won't show up 4213 * in the cache). 4214 */ 4215 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon); 4216 freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt); 4217 } 4218 4219 /* execute each callback and free its structure */ 4220 while ((acb = callback_list) != NULL) { 4221 if (acb->acb_done) 4222 acb->acb_done(zio, acb->acb_buf, acb->acb_private); 4223 4224 if (acb->acb_zio_dummy != NULL) { 4225 acb->acb_zio_dummy->io_error = zio->io_error; 4226 zio_nowait(acb->acb_zio_dummy); 4227 } 4228 4229 callback_list = acb->acb_next; 4230 kmem_free(acb, sizeof (arc_callback_t)); 4231 } 4232 4233 if (freeable) 4234 arc_hdr_destroy(hdr); 4235} 4236 4237/* 4238 * "Read" the block at the specified DVA (in bp) via the 4239 * cache. If the block is found in the cache, invoke the provided 4240 * callback immediately and return. Note that the `zio' parameter 4241 * in the callback will be NULL in this case, since no IO was 4242 * required. If the block is not in the cache pass the read request 4243 * on to the spa with a substitute callback function, so that the 4244 * requested block will be added to the cache. 4245 * 4246 * If a read request arrives for a block that has a read in-progress, 4247 * either wait for the in-progress read to complete (and return the 4248 * results); or, if this is a read with a "done" func, add a record 4249 * to the read to invoke the "done" func when the read completes, 4250 * and return; or just return. 4251 * 4252 * arc_read_done() will invoke all the requested "done" functions 4253 * for readers of this block. 4254 */ 4255int 4256arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_done_func_t *done, 4257 void *private, zio_priority_t priority, int zio_flags, 4258 arc_flags_t *arc_flags, const zbookmark_phys_t *zb) 4259{ 4260 arc_buf_hdr_t *hdr = NULL; 4261 arc_buf_t *buf = NULL; 4262 kmutex_t *hash_lock = NULL; 4263 zio_t *rzio; 4264 uint64_t guid = spa_load_guid(spa); 4265 4266 ASSERT(!BP_IS_EMBEDDED(bp) || 4267 BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA); 4268 4269top: 4270 if (!BP_IS_EMBEDDED(bp)) { 4271 /* 4272 * Embedded BP's have no DVA and require no I/O to "read". 4273 * Create an anonymous arc buf to back it. 4274 */ 4275 hdr = buf_hash_find(guid, bp, &hash_lock); 4276 } 4277 4278 if (hdr != NULL && HDR_HAS_L1HDR(hdr) && hdr->b_l1hdr.b_datacnt > 0) { 4279 4280 *arc_flags |= ARC_FLAG_CACHED; 4281 4282 if (HDR_IO_IN_PROGRESS(hdr)) { 4283 4284 if ((hdr->b_flags & ARC_FLAG_PRIO_ASYNC_READ) && 4285 priority == ZIO_PRIORITY_SYNC_READ) { 4286 /* 4287 * This sync read must wait for an 4288 * in-progress async read (e.g. a predictive 4289 * prefetch). Async reads are queued 4290 * separately at the vdev_queue layer, so 4291 * this is a form of priority inversion. 4292 * Ideally, we would "inherit" the demand 4293 * i/o's priority by moving the i/o from 4294 * the async queue to the synchronous queue, 4295 * but there is currently no mechanism to do 4296 * so. Track this so that we can evaluate 4297 * the magnitude of this potential performance 4298 * problem. 4299 * 4300 * Note that if the prefetch i/o is already 4301 * active (has been issued to the device), 4302 * the prefetch improved performance, because 4303 * we issued it sooner than we would have 4304 * without the prefetch. 4305 */ 4306 DTRACE_PROBE1(arc__sync__wait__for__async, 4307 arc_buf_hdr_t *, hdr); 4308 ARCSTAT_BUMP(arcstat_sync_wait_for_async); 4309 } 4310 if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) { 4311 hdr->b_flags &= ~ARC_FLAG_PREDICTIVE_PREFETCH; 4312 } 4313 4314 if (*arc_flags & ARC_FLAG_WAIT) { 4315 cv_wait(&hdr->b_l1hdr.b_cv, hash_lock); 4316 mutex_exit(hash_lock); 4317 goto top; 4318 } 4319 ASSERT(*arc_flags & ARC_FLAG_NOWAIT); 4320 4321 if (done) { 4322 arc_callback_t *acb = NULL; 4323 4324 acb = kmem_zalloc(sizeof (arc_callback_t), 4325 KM_SLEEP); 4326 acb->acb_done = done; 4327 acb->acb_private = private; 4328 if (pio != NULL) 4329 acb->acb_zio_dummy = zio_null(pio, 4330 spa, NULL, NULL, NULL, zio_flags); 4331 4332 ASSERT(acb->acb_done != NULL); 4333 acb->acb_next = hdr->b_l1hdr.b_acb; 4334 hdr->b_l1hdr.b_acb = acb; 4335 add_reference(hdr, hash_lock, private); 4336 mutex_exit(hash_lock); 4337 return (0); 4338 } 4339 mutex_exit(hash_lock); 4340 return (0); 4341 } 4342 4343 ASSERT(hdr->b_l1hdr.b_state == arc_mru || 4344 hdr->b_l1hdr.b_state == arc_mfu); 4345 4346 if (done) { 4347 if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) { 4348 /* 4349 * This is a demand read which does not have to 4350 * wait for i/o because we did a predictive 4351 * prefetch i/o for it, which has completed. 4352 */ 4353 DTRACE_PROBE1( 4354 arc__demand__hit__predictive__prefetch, 4355 arc_buf_hdr_t *, hdr); 4356 ARCSTAT_BUMP( 4357 arcstat_demand_hit_predictive_prefetch); 4358 hdr->b_flags &= ~ARC_FLAG_PREDICTIVE_PREFETCH; 4359 } 4360 add_reference(hdr, hash_lock, private); 4361 /* 4362 * If this block is already in use, create a new 4363 * copy of the data so that we will be guaranteed 4364 * that arc_release() will always succeed. 4365 */ 4366 buf = hdr->b_l1hdr.b_buf; 4367 ASSERT(buf); 4368 ASSERT(buf->b_data); 4369 if (HDR_BUF_AVAILABLE(hdr)) { 4370 ASSERT(buf->b_efunc == NULL); 4371 hdr->b_flags &= ~ARC_FLAG_BUF_AVAILABLE; 4372 } else { 4373 buf = arc_buf_clone(buf); 4374 } 4375 4376 } else if (*arc_flags & ARC_FLAG_PREFETCH && 4377 refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) { 4378 hdr->b_flags |= ARC_FLAG_PREFETCH; 4379 } 4380 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr); 4381 arc_access(hdr, hash_lock); 4382 if (*arc_flags & ARC_FLAG_L2CACHE) 4383 hdr->b_flags |= ARC_FLAG_L2CACHE; 4384 if (*arc_flags & ARC_FLAG_L2COMPRESS) 4385 hdr->b_flags |= ARC_FLAG_L2COMPRESS; 4386 mutex_exit(hash_lock); 4387 ARCSTAT_BUMP(arcstat_hits); 4388 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr), 4389 demand, prefetch, !HDR_ISTYPE_METADATA(hdr), 4390 data, metadata, hits); 4391 4392 if (done) 4393 done(NULL, buf, private); 4394 } else { 4395 uint64_t size = BP_GET_LSIZE(bp); 4396 arc_callback_t *acb; 4397 vdev_t *vd = NULL; 4398 uint64_t addr = 0; 4399 boolean_t devw = B_FALSE; 4400 enum zio_compress b_compress = ZIO_COMPRESS_OFF; 4401 int32_t b_asize = 0; 4402 4403 if (hdr == NULL) { 4404 /* this block is not in the cache */ 4405 arc_buf_hdr_t *exists = NULL; 4406 arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp); 4407 buf = arc_buf_alloc(spa, size, private, type); 4408 hdr = buf->b_hdr; 4409 if (!BP_IS_EMBEDDED(bp)) { 4410 hdr->b_dva = *BP_IDENTITY(bp); 4411 hdr->b_birth = BP_PHYSICAL_BIRTH(bp); 4412 exists = buf_hash_insert(hdr, &hash_lock); 4413 } 4414 if (exists != NULL) { 4415 /* somebody beat us to the hash insert */ 4416 mutex_exit(hash_lock); 4417 buf_discard_identity(hdr); 4418 (void) arc_buf_remove_ref(buf, private); 4419 goto top; /* restart the IO request */ 4420 } 4421 4422 /* 4423 * If there is a callback, we pass our reference to 4424 * it; otherwise we remove our reference. 4425 */ 4426 if (done == NULL) { 4427 (void) remove_reference(hdr, hash_lock, 4428 private); 4429 } 4430 if (*arc_flags & ARC_FLAG_PREFETCH) 4431 hdr->b_flags |= ARC_FLAG_PREFETCH; 4432 if (*arc_flags & ARC_FLAG_L2CACHE) 4433 hdr->b_flags |= ARC_FLAG_L2CACHE; 4434 if (*arc_flags & ARC_FLAG_L2COMPRESS) 4435 hdr->b_flags |= ARC_FLAG_L2COMPRESS; 4436 if (BP_GET_LEVEL(bp) > 0) 4437 hdr->b_flags |= ARC_FLAG_INDIRECT; 4438 } else { 4439 /* 4440 * This block is in the ghost cache. If it was L2-only 4441 * (and thus didn't have an L1 hdr), we realloc the 4442 * header to add an L1 hdr. 4443 */ 4444 if (!HDR_HAS_L1HDR(hdr)) { 4445 hdr = arc_hdr_realloc(hdr, hdr_l2only_cache, 4446 hdr_full_cache); 4447 } 4448 4449 ASSERT(GHOST_STATE(hdr->b_l1hdr.b_state)); 4450 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 4451 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 4452 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 4453 4454 /* 4455 * If there is a callback, we pass a reference to it. 4456 */ 4457 if (done != NULL) 4458 add_reference(hdr, hash_lock, private); 4459 if (*arc_flags & ARC_FLAG_PREFETCH) 4460 hdr->b_flags |= ARC_FLAG_PREFETCH; 4461 if (*arc_flags & ARC_FLAG_L2CACHE) 4462 hdr->b_flags |= ARC_FLAG_L2CACHE; 4463 if (*arc_flags & ARC_FLAG_L2COMPRESS) 4464 hdr->b_flags |= ARC_FLAG_L2COMPRESS; 4465 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 4466 buf->b_hdr = hdr; 4467 buf->b_data = NULL; 4468 buf->b_efunc = NULL; 4469 buf->b_private = NULL; 4470 buf->b_next = NULL; 4471 hdr->b_l1hdr.b_buf = buf; 4472 ASSERT0(hdr->b_l1hdr.b_datacnt); 4473 hdr->b_l1hdr.b_datacnt = 1; 4474 arc_get_data_buf(buf); 4475 arc_access(hdr, hash_lock); 4476 } 4477 4478 if (*arc_flags & ARC_FLAG_PREDICTIVE_PREFETCH) 4479 hdr->b_flags |= ARC_FLAG_PREDICTIVE_PREFETCH; 4480 ASSERT(!GHOST_STATE(hdr->b_l1hdr.b_state)); 4481 4482 acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP); 4483 acb->acb_done = done; 4484 acb->acb_private = private; 4485 4486 ASSERT(hdr->b_l1hdr.b_acb == NULL); 4487 hdr->b_l1hdr.b_acb = acb; 4488 hdr->b_flags |= ARC_FLAG_IO_IN_PROGRESS; 4489 4490 if (HDR_HAS_L2HDR(hdr) && 4491 (vd = hdr->b_l2hdr.b_dev->l2ad_vdev) != NULL) { 4492 devw = hdr->b_l2hdr.b_dev->l2ad_writing; 4493 addr = hdr->b_l2hdr.b_daddr; 4494 b_compress = hdr->b_l2hdr.b_compress; 4495 b_asize = hdr->b_l2hdr.b_asize; 4496 /* 4497 * Lock out device removal. 4498 */ 4499 if (vdev_is_dead(vd) || 4500 !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER)) 4501 vd = NULL; 4502 } 4503 4504 if (hash_lock != NULL) 4505 mutex_exit(hash_lock); 4506 4507 /* 4508 * At this point, we have a level 1 cache miss. Try again in 4509 * L2ARC if possible. 4510 */ 4511 ASSERT3U(hdr->b_size, ==, size); 4512 DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp, 4513 uint64_t, size, zbookmark_phys_t *, zb); 4514 ARCSTAT_BUMP(arcstat_misses); 4515 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr), 4516 demand, prefetch, !HDR_ISTYPE_METADATA(hdr), 4517 data, metadata, misses); 4518#ifdef _KERNEL 4519 curthread->td_ru.ru_inblock++; 4520#endif 4521 4522 if (priority == ZIO_PRIORITY_ASYNC_READ) 4523 hdr->b_flags |= ARC_FLAG_PRIO_ASYNC_READ; 4524 else 4525 hdr->b_flags &= ~ARC_FLAG_PRIO_ASYNC_READ; 4526 4527 if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) { 4528 /* 4529 * Read from the L2ARC if the following are true: 4530 * 1. The L2ARC vdev was previously cached. 4531 * 2. This buffer still has L2ARC metadata. 4532 * 3. This buffer isn't currently writing to the L2ARC. 4533 * 4. The L2ARC entry wasn't evicted, which may 4534 * also have invalidated the vdev. 4535 * 5. This isn't prefetch and l2arc_noprefetch is set. 4536 */ 4537 if (HDR_HAS_L2HDR(hdr) && 4538 !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) && 4539 !(l2arc_noprefetch && HDR_PREFETCH(hdr))) { 4540 l2arc_read_callback_t *cb; 4541 void* b_data; 4542 4543 DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr); 4544 ARCSTAT_BUMP(arcstat_l2_hits); 4545 4546 cb = kmem_zalloc(sizeof (l2arc_read_callback_t), 4547 KM_SLEEP); 4548 cb->l2rcb_buf = buf; 4549 cb->l2rcb_spa = spa; 4550 cb->l2rcb_bp = *bp; 4551 cb->l2rcb_zb = *zb; 4552 cb->l2rcb_flags = zio_flags; 4553 cb->l2rcb_compress = b_compress; 4554 if (b_asize > hdr->b_size) { 4555 ASSERT3U(b_compress, ==, 4556 ZIO_COMPRESS_OFF); 4557 b_data = zio_data_buf_alloc(b_asize); 4558 cb->l2rcb_data = b_data; 4559 } else { 4560 b_data = buf->b_data; 4561 } 4562 4563 ASSERT(addr >= VDEV_LABEL_START_SIZE && 4564 addr + size < vd->vdev_psize - 4565 VDEV_LABEL_END_SIZE); 4566 4567 /* 4568 * l2arc read. The SCL_L2ARC lock will be 4569 * released by l2arc_read_done(). 4570 * Issue a null zio if the underlying buffer 4571 * was squashed to zero size by compression. 4572 */ 4573 if (b_compress == ZIO_COMPRESS_EMPTY) { 4574 ASSERT3U(b_asize, ==, 0); 4575 rzio = zio_null(pio, spa, vd, 4576 l2arc_read_done, cb, 4577 zio_flags | ZIO_FLAG_DONT_CACHE | 4578 ZIO_FLAG_CANFAIL | 4579 ZIO_FLAG_DONT_PROPAGATE | 4580 ZIO_FLAG_DONT_RETRY); 4581 } else { 4582 rzio = zio_read_phys(pio, vd, addr, 4583 b_asize, b_data, 4584 ZIO_CHECKSUM_OFF, 4585 l2arc_read_done, cb, priority, 4586 zio_flags | ZIO_FLAG_DONT_CACHE | 4587 ZIO_FLAG_CANFAIL | 4588 ZIO_FLAG_DONT_PROPAGATE | 4589 ZIO_FLAG_DONT_RETRY, B_FALSE); 4590 } 4591 DTRACE_PROBE2(l2arc__read, vdev_t *, vd, 4592 zio_t *, rzio); 4593 ARCSTAT_INCR(arcstat_l2_read_bytes, b_asize); 4594 4595 if (*arc_flags & ARC_FLAG_NOWAIT) { 4596 zio_nowait(rzio); 4597 return (0); 4598 } 4599 4600 ASSERT(*arc_flags & ARC_FLAG_WAIT); 4601 if (zio_wait(rzio) == 0) 4602 return (0); 4603 4604 /* l2arc read error; goto zio_read() */ 4605 } else { 4606 DTRACE_PROBE1(l2arc__miss, 4607 arc_buf_hdr_t *, hdr); 4608 ARCSTAT_BUMP(arcstat_l2_misses); 4609 if (HDR_L2_WRITING(hdr)) 4610 ARCSTAT_BUMP(arcstat_l2_rw_clash); 4611 spa_config_exit(spa, SCL_L2ARC, vd); 4612 } 4613 } else { 4614 if (vd != NULL) 4615 spa_config_exit(spa, SCL_L2ARC, vd); 4616 if (l2arc_ndev != 0) { 4617 DTRACE_PROBE1(l2arc__miss, 4618 arc_buf_hdr_t *, hdr); 4619 ARCSTAT_BUMP(arcstat_l2_misses); 4620 } 4621 } 4622 4623 rzio = zio_read(pio, spa, bp, buf->b_data, size, 4624 arc_read_done, buf, priority, zio_flags, zb); 4625 4626 if (*arc_flags & ARC_FLAG_WAIT) 4627 return (zio_wait(rzio)); 4628 4629 ASSERT(*arc_flags & ARC_FLAG_NOWAIT); 4630 zio_nowait(rzio); 4631 } 4632 return (0); 4633} 4634 4635void 4636arc_set_callback(arc_buf_t *buf, arc_evict_func_t *func, void *private) 4637{ 4638 ASSERT(buf->b_hdr != NULL); 4639 ASSERT(buf->b_hdr->b_l1hdr.b_state != arc_anon); 4640 ASSERT(!refcount_is_zero(&buf->b_hdr->b_l1hdr.b_refcnt) || 4641 func == NULL); 4642 ASSERT(buf->b_efunc == NULL); 4643 ASSERT(!HDR_BUF_AVAILABLE(buf->b_hdr)); 4644 4645 buf->b_efunc = func; 4646 buf->b_private = private; 4647} 4648 4649/* 4650 * Notify the arc that a block was freed, and thus will never be used again. 4651 */ 4652void 4653arc_freed(spa_t *spa, const blkptr_t *bp) 4654{ 4655 arc_buf_hdr_t *hdr; 4656 kmutex_t *hash_lock; 4657 uint64_t guid = spa_load_guid(spa); 4658 4659 ASSERT(!BP_IS_EMBEDDED(bp)); 4660 4661 hdr = buf_hash_find(guid, bp, &hash_lock); 4662 if (hdr == NULL) 4663 return; 4664 if (HDR_BUF_AVAILABLE(hdr)) { 4665 arc_buf_t *buf = hdr->b_l1hdr.b_buf; 4666 add_reference(hdr, hash_lock, FTAG); 4667 hdr->b_flags &= ~ARC_FLAG_BUF_AVAILABLE; 4668 mutex_exit(hash_lock); 4669 4670 arc_release(buf, FTAG); 4671 (void) arc_buf_remove_ref(buf, FTAG); 4672 } else { 4673 mutex_exit(hash_lock); 4674 } 4675 4676} 4677 4678/* 4679 * Clear the user eviction callback set by arc_set_callback(), first calling 4680 * it if it exists. Because the presence of a callback keeps an arc_buf cached 4681 * clearing the callback may result in the arc_buf being destroyed. However, 4682 * it will not result in the *last* arc_buf being destroyed, hence the data 4683 * will remain cached in the ARC. We make a copy of the arc buffer here so 4684 * that we can process the callback without holding any locks. 4685 * 4686 * It's possible that the callback is already in the process of being cleared 4687 * by another thread. In this case we can not clear the callback. 4688 * 4689 * Returns B_TRUE if the callback was successfully called and cleared. 4690 */ 4691boolean_t 4692arc_clear_callback(arc_buf_t *buf) 4693{ 4694 arc_buf_hdr_t *hdr; 4695 kmutex_t *hash_lock; 4696 arc_evict_func_t *efunc = buf->b_efunc; 4697 void *private = buf->b_private; 4698 4699 mutex_enter(&buf->b_evict_lock); 4700 hdr = buf->b_hdr; 4701 if (hdr == NULL) { 4702 /* 4703 * We are in arc_do_user_evicts(). 4704 */ 4705 ASSERT(buf->b_data == NULL); 4706 mutex_exit(&buf->b_evict_lock); 4707 return (B_FALSE); 4708 } else if (buf->b_data == NULL) { 4709 /* 4710 * We are on the eviction list; process this buffer now 4711 * but let arc_do_user_evicts() do the reaping. 4712 */ 4713 buf->b_efunc = NULL; 4714 mutex_exit(&buf->b_evict_lock); 4715 VERIFY0(efunc(private)); 4716 return (B_TRUE); 4717 } 4718 hash_lock = HDR_LOCK(hdr); 4719 mutex_enter(hash_lock); 4720 hdr = buf->b_hdr; 4721 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 4722 4723 ASSERT3U(refcount_count(&hdr->b_l1hdr.b_refcnt), <, 4724 hdr->b_l1hdr.b_datacnt); 4725 ASSERT(hdr->b_l1hdr.b_state == arc_mru || 4726 hdr->b_l1hdr.b_state == arc_mfu); 4727 4728 buf->b_efunc = NULL; 4729 buf->b_private = NULL; 4730 4731 if (hdr->b_l1hdr.b_datacnt > 1) { 4732 mutex_exit(&buf->b_evict_lock); 4733 arc_buf_destroy(buf, TRUE); 4734 } else { 4735 ASSERT(buf == hdr->b_l1hdr.b_buf); 4736 hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE; 4737 mutex_exit(&buf->b_evict_lock); 4738 } 4739 4740 mutex_exit(hash_lock); 4741 VERIFY0(efunc(private)); 4742 return (B_TRUE); 4743} 4744 4745/* 4746 * Release this buffer from the cache, making it an anonymous buffer. This 4747 * must be done after a read and prior to modifying the buffer contents. 4748 * If the buffer has more than one reference, we must make 4749 * a new hdr for the buffer. 4750 */ 4751void 4752arc_release(arc_buf_t *buf, void *tag) 4753{ 4754 arc_buf_hdr_t *hdr = buf->b_hdr; 4755 4756 /* 4757 * It would be nice to assert that if it's DMU metadata (level > 4758 * 0 || it's the dnode file), then it must be syncing context. 4759 * But we don't know that information at this level. 4760 */ 4761 4762 mutex_enter(&buf->b_evict_lock); 4763 4764 ASSERT(HDR_HAS_L1HDR(hdr)); 4765 4766 /* 4767 * We don't grab the hash lock prior to this check, because if 4768 * the buffer's header is in the arc_anon state, it won't be 4769 * linked into the hash table. 4770 */ 4771 if (hdr->b_l1hdr.b_state == arc_anon) { 4772 mutex_exit(&buf->b_evict_lock); 4773 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 4774 ASSERT(!HDR_IN_HASH_TABLE(hdr)); 4775 ASSERT(!HDR_HAS_L2HDR(hdr)); 4776 ASSERT(BUF_EMPTY(hdr)); 4777 ASSERT3U(hdr->b_l1hdr.b_datacnt, ==, 1); 4778 ASSERT3S(refcount_count(&hdr->b_l1hdr.b_refcnt), ==, 1); 4779 ASSERT(!list_link_active(&hdr->b_l1hdr.b_arc_node)); 4780 4781 ASSERT3P(buf->b_efunc, ==, NULL); 4782 ASSERT3P(buf->b_private, ==, NULL); 4783 4784 hdr->b_l1hdr.b_arc_access = 0; 4785 arc_buf_thaw(buf); 4786 4787 return; 4788 } 4789 4790 kmutex_t *hash_lock = HDR_LOCK(hdr); 4791 mutex_enter(hash_lock); 4792 4793 /* 4794 * This assignment is only valid as long as the hash_lock is 4795 * held, we must be careful not to reference state or the 4796 * b_state field after dropping the lock. 4797 */ 4798 arc_state_t *state = hdr->b_l1hdr.b_state; 4799 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 4800 ASSERT3P(state, !=, arc_anon); 4801 4802 /* this buffer is not on any list */ 4803 ASSERT(refcount_count(&hdr->b_l1hdr.b_refcnt) > 0); 4804 4805 if (HDR_HAS_L2HDR(hdr)) { 4806 mutex_enter(&hdr->b_l2hdr.b_dev->l2ad_mtx); 4807 4808 /* 4809 * We have to recheck this conditional again now that 4810 * we're holding the l2ad_mtx to prevent a race with 4811 * another thread which might be concurrently calling 4812 * l2arc_evict(). In that case, l2arc_evict() might have 4813 * destroyed the header's L2 portion as we were waiting 4814 * to acquire the l2ad_mtx. 4815 */ 4816 if (HDR_HAS_L2HDR(hdr)) { 4817 l2arc_trim(hdr); 4818 arc_hdr_l2hdr_destroy(hdr); 4819 } 4820 4821 mutex_exit(&hdr->b_l2hdr.b_dev->l2ad_mtx); 4822 } 4823 4824 /* 4825 * Do we have more than one buf? 4826 */ 4827 if (hdr->b_l1hdr.b_datacnt > 1) { 4828 arc_buf_hdr_t *nhdr; 4829 arc_buf_t **bufp; 4830 uint64_t blksz = hdr->b_size; 4831 uint64_t spa = hdr->b_spa; 4832 arc_buf_contents_t type = arc_buf_type(hdr); 4833 uint32_t flags = hdr->b_flags; 4834 4835 ASSERT(hdr->b_l1hdr.b_buf != buf || buf->b_next != NULL); 4836 /* 4837 * Pull the data off of this hdr and attach it to 4838 * a new anonymous hdr. 4839 */ 4840 (void) remove_reference(hdr, hash_lock, tag); 4841 bufp = &hdr->b_l1hdr.b_buf; 4842 while (*bufp != buf) 4843 bufp = &(*bufp)->b_next; 4844 *bufp = buf->b_next; 4845 buf->b_next = NULL; 4846 4847 ASSERT3P(state, !=, arc_l2c_only); 4848 4849 (void) refcount_remove_many( 4850 &state->arcs_size, hdr->b_size, buf); 4851 4852 if (refcount_is_zero(&hdr->b_l1hdr.b_refcnt)) { 4853 ASSERT3P(state, !=, arc_l2c_only); 4854 uint64_t *size = &state->arcs_lsize[type]; 4855 ASSERT3U(*size, >=, hdr->b_size); 4856 atomic_add_64(size, -hdr->b_size); 4857 } 4858 4859 /* 4860 * We're releasing a duplicate user data buffer, update 4861 * our statistics accordingly. 4862 */ 4863 if (HDR_ISTYPE_DATA(hdr)) { 4864 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers); 4865 ARCSTAT_INCR(arcstat_duplicate_buffers_size, 4866 -hdr->b_size); 4867 } 4868 hdr->b_l1hdr.b_datacnt -= 1; 4869 arc_cksum_verify(buf); 4870#ifdef illumos 4871 arc_buf_unwatch(buf); 4872#endif 4873 4874 mutex_exit(hash_lock); 4875 4876 nhdr = kmem_cache_alloc(hdr_full_cache, KM_PUSHPAGE); 4877 nhdr->b_size = blksz; 4878 nhdr->b_spa = spa; 4879 4880 nhdr->b_flags = flags & ARC_FLAG_L2_WRITING; 4881 nhdr->b_flags |= arc_bufc_to_flags(type); 4882 nhdr->b_flags |= ARC_FLAG_HAS_L1HDR; 4883 4884 nhdr->b_l1hdr.b_buf = buf; 4885 nhdr->b_l1hdr.b_datacnt = 1; 4886 nhdr->b_l1hdr.b_state = arc_anon; 4887 nhdr->b_l1hdr.b_arc_access = 0; 4888 nhdr->b_l1hdr.b_tmp_cdata = NULL; 4889 nhdr->b_freeze_cksum = NULL; 4890 4891 (void) refcount_add(&nhdr->b_l1hdr.b_refcnt, tag); 4892 buf->b_hdr = nhdr; 4893 mutex_exit(&buf->b_evict_lock); 4894 (void) refcount_add_many(&arc_anon->arcs_size, blksz, buf); 4895 } else { 4896 mutex_exit(&buf->b_evict_lock); 4897 ASSERT(refcount_count(&hdr->b_l1hdr.b_refcnt) == 1); 4898 /* protected by hash lock, or hdr is on arc_anon */ 4899 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 4900 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 4901 arc_change_state(arc_anon, hdr, hash_lock); 4902 hdr->b_l1hdr.b_arc_access = 0; 4903 mutex_exit(hash_lock); 4904 4905 buf_discard_identity(hdr); 4906 arc_buf_thaw(buf); 4907 } 4908 buf->b_efunc = NULL; 4909 buf->b_private = NULL; 4910} 4911 4912int 4913arc_released(arc_buf_t *buf) 4914{ 4915 int released; 4916 4917 mutex_enter(&buf->b_evict_lock); 4918 released = (buf->b_data != NULL && 4919 buf->b_hdr->b_l1hdr.b_state == arc_anon); 4920 mutex_exit(&buf->b_evict_lock); 4921 return (released); 4922} 4923 4924#ifdef ZFS_DEBUG 4925int 4926arc_referenced(arc_buf_t *buf) 4927{ 4928 int referenced; 4929 4930 mutex_enter(&buf->b_evict_lock); 4931 referenced = (refcount_count(&buf->b_hdr->b_l1hdr.b_refcnt)); 4932 mutex_exit(&buf->b_evict_lock); 4933 return (referenced); 4934} 4935#endif 4936 4937static void 4938arc_write_ready(zio_t *zio) 4939{ 4940 arc_write_callback_t *callback = zio->io_private; 4941 arc_buf_t *buf = callback->awcb_buf; 4942 arc_buf_hdr_t *hdr = buf->b_hdr; 4943 4944 ASSERT(HDR_HAS_L1HDR(hdr)); 4945 ASSERT(!refcount_is_zero(&buf->b_hdr->b_l1hdr.b_refcnt)); 4946 ASSERT(hdr->b_l1hdr.b_datacnt > 0); 4947 callback->awcb_ready(zio, buf, callback->awcb_private); 4948 4949 /* 4950 * If the IO is already in progress, then this is a re-write 4951 * attempt, so we need to thaw and re-compute the cksum. 4952 * It is the responsibility of the callback to handle the 4953 * accounting for any re-write attempt. 4954 */ 4955 if (HDR_IO_IN_PROGRESS(hdr)) { 4956 mutex_enter(&hdr->b_l1hdr.b_freeze_lock); 4957 if (hdr->b_freeze_cksum != NULL) { 4958 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 4959 hdr->b_freeze_cksum = NULL; 4960 } 4961 mutex_exit(&hdr->b_l1hdr.b_freeze_lock); 4962 } 4963 arc_cksum_compute(buf, B_FALSE); 4964 hdr->b_flags |= ARC_FLAG_IO_IN_PROGRESS; 4965} 4966 4967/* 4968 * The SPA calls this callback for each physical write that happens on behalf 4969 * of a logical write. See the comment in dbuf_write_physdone() for details. 4970 */ 4971static void 4972arc_write_physdone(zio_t *zio) 4973{ 4974 arc_write_callback_t *cb = zio->io_private; 4975 if (cb->awcb_physdone != NULL) 4976 cb->awcb_physdone(zio, cb->awcb_buf, cb->awcb_private); 4977} 4978 4979static void 4980arc_write_done(zio_t *zio) 4981{ 4982 arc_write_callback_t *callback = zio->io_private; 4983 arc_buf_t *buf = callback->awcb_buf; 4984 arc_buf_hdr_t *hdr = buf->b_hdr; 4985 4986 ASSERT(hdr->b_l1hdr.b_acb == NULL); 4987 4988 if (zio->io_error == 0) { 4989 if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) { 4990 buf_discard_identity(hdr); 4991 } else { 4992 hdr->b_dva = *BP_IDENTITY(zio->io_bp); 4993 hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp); 4994 } 4995 } else { 4996 ASSERT(BUF_EMPTY(hdr)); 4997 } 4998 4999 /* 5000 * If the block to be written was all-zero or compressed enough to be 5001 * embedded in the BP, no write was performed so there will be no 5002 * dva/birth/checksum. The buffer must therefore remain anonymous 5003 * (and uncached). 5004 */ 5005 if (!BUF_EMPTY(hdr)) { 5006 arc_buf_hdr_t *exists; 5007 kmutex_t *hash_lock; 5008 5009 ASSERT(zio->io_error == 0); 5010 5011 arc_cksum_verify(buf); 5012 5013 exists = buf_hash_insert(hdr, &hash_lock); 5014 if (exists != NULL) { 5015 /* 5016 * This can only happen if we overwrite for 5017 * sync-to-convergence, because we remove 5018 * buffers from the hash table when we arc_free(). 5019 */ 5020 if (zio->io_flags & ZIO_FLAG_IO_REWRITE) { 5021 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) 5022 panic("bad overwrite, hdr=%p exists=%p", 5023 (void *)hdr, (void *)exists); 5024 ASSERT(refcount_is_zero( 5025 &exists->b_l1hdr.b_refcnt)); 5026 arc_change_state(arc_anon, exists, hash_lock); 5027 mutex_exit(hash_lock); 5028 arc_hdr_destroy(exists); 5029 exists = buf_hash_insert(hdr, &hash_lock); 5030 ASSERT3P(exists, ==, NULL); 5031 } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) { 5032 /* nopwrite */ 5033 ASSERT(zio->io_prop.zp_nopwrite); 5034 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) 5035 panic("bad nopwrite, hdr=%p exists=%p", 5036 (void *)hdr, (void *)exists); 5037 } else { 5038 /* Dedup */ 5039 ASSERT(hdr->b_l1hdr.b_datacnt == 1); 5040 ASSERT(hdr->b_l1hdr.b_state == arc_anon); 5041 ASSERT(BP_GET_DEDUP(zio->io_bp)); 5042 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0); 5043 } 5044 } 5045 hdr->b_flags &= ~ARC_FLAG_IO_IN_PROGRESS; 5046 /* if it's not anon, we are doing a scrub */ 5047 if (exists == NULL && hdr->b_l1hdr.b_state == arc_anon) 5048 arc_access(hdr, hash_lock); 5049 mutex_exit(hash_lock); 5050 } else { 5051 hdr->b_flags &= ~ARC_FLAG_IO_IN_PROGRESS; 5052 } 5053 5054 ASSERT(!refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 5055 callback->awcb_done(zio, buf, callback->awcb_private); 5056 5057 kmem_free(callback, sizeof (arc_write_callback_t)); 5058} 5059 5060zio_t * 5061arc_write(zio_t *pio, spa_t *spa, uint64_t txg, 5062 blkptr_t *bp, arc_buf_t *buf, boolean_t l2arc, boolean_t l2arc_compress, 5063 const zio_prop_t *zp, arc_done_func_t *ready, arc_done_func_t *physdone, 5064 arc_done_func_t *done, void *private, zio_priority_t priority, 5065 int zio_flags, const zbookmark_phys_t *zb) 5066{ 5067 arc_buf_hdr_t *hdr = buf->b_hdr; 5068 arc_write_callback_t *callback; 5069 zio_t *zio; 5070 5071 ASSERT(ready != NULL); 5072 ASSERT(done != NULL); 5073 ASSERT(!HDR_IO_ERROR(hdr)); 5074 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 5075 ASSERT(hdr->b_l1hdr.b_acb == NULL); 5076 ASSERT(hdr->b_l1hdr.b_datacnt > 0); 5077 if (l2arc) 5078 hdr->b_flags |= ARC_FLAG_L2CACHE; 5079 if (l2arc_compress) 5080 hdr->b_flags |= ARC_FLAG_L2COMPRESS; 5081 callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP); 5082 callback->awcb_ready = ready; 5083 callback->awcb_physdone = physdone; 5084 callback->awcb_done = done; 5085 callback->awcb_private = private; 5086 callback->awcb_buf = buf; 5087 5088 zio = zio_write(pio, spa, txg, bp, buf->b_data, hdr->b_size, zp, 5089 arc_write_ready, arc_write_physdone, arc_write_done, callback, 5090 priority, zio_flags, zb); 5091 5092 return (zio); 5093} 5094 5095static int 5096arc_memory_throttle(uint64_t reserve, uint64_t txg) 5097{ 5098#ifdef _KERNEL 5099 uint64_t available_memory = ptob(freemem); 5100 static uint64_t page_load = 0; 5101 static uint64_t last_txg = 0; 5102 5103#if defined(__i386) || !defined(UMA_MD_SMALL_ALLOC) 5104 available_memory = 5105 MIN(available_memory, ptob(vmem_size(heap_arena, VMEM_FREE))); 5106#endif 5107 5108 if (freemem > (uint64_t)physmem * arc_lotsfree_percent / 100) 5109 return (0); 5110 5111 if (txg > last_txg) { 5112 last_txg = txg; 5113 page_load = 0; 5114 } 5115 /* 5116 * If we are in pageout, we know that memory is already tight, 5117 * the arc is already going to be evicting, so we just want to 5118 * continue to let page writes occur as quickly as possible. 5119 */ 5120 if (curproc == pageproc) { 5121 if (page_load > MAX(ptob(minfree), available_memory) / 4) 5122 return (SET_ERROR(ERESTART)); 5123 /* Note: reserve is inflated, so we deflate */ 5124 page_load += reserve / 8; 5125 return (0); 5126 } else if (page_load > 0 && arc_reclaim_needed()) { 5127 /* memory is low, delay before restarting */ 5128 ARCSTAT_INCR(arcstat_memory_throttle_count, 1); 5129 return (SET_ERROR(EAGAIN)); 5130 } 5131 page_load = 0; 5132#endif 5133 return (0); 5134} 5135 5136void 5137arc_tempreserve_clear(uint64_t reserve) 5138{ 5139 atomic_add_64(&arc_tempreserve, -reserve); 5140 ASSERT((int64_t)arc_tempreserve >= 0); 5141} 5142 5143int 5144arc_tempreserve_space(uint64_t reserve, uint64_t txg) 5145{ 5146 int error; 5147 uint64_t anon_size; 5148 5149 if (reserve > arc_c/4 && !arc_no_grow) { 5150 arc_c = MIN(arc_c_max, reserve * 4); 5151 DTRACE_PROBE1(arc__set_reserve, uint64_t, arc_c); 5152 } 5153 if (reserve > arc_c) 5154 return (SET_ERROR(ENOMEM)); 5155 5156 /* 5157 * Don't count loaned bufs as in flight dirty data to prevent long 5158 * network delays from blocking transactions that are ready to be 5159 * assigned to a txg. 5160 */ 5161 anon_size = MAX((int64_t)(refcount_count(&arc_anon->arcs_size) - 5162 arc_loaned_bytes), 0); 5163 5164 /* 5165 * Writes will, almost always, require additional memory allocations 5166 * in order to compress/encrypt/etc the data. We therefore need to 5167 * make sure that there is sufficient available memory for this. 5168 */ 5169 error = arc_memory_throttle(reserve, txg); 5170 if (error != 0) 5171 return (error); 5172 5173 /* 5174 * Throttle writes when the amount of dirty data in the cache 5175 * gets too large. We try to keep the cache less than half full 5176 * of dirty blocks so that our sync times don't grow too large. 5177 * Note: if two requests come in concurrently, we might let them 5178 * both succeed, when one of them should fail. Not a huge deal. 5179 */ 5180 5181 if (reserve + arc_tempreserve + anon_size > arc_c / 2 && 5182 anon_size > arc_c / 4) { 5183 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK " 5184 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n", 5185 arc_tempreserve>>10, 5186 arc_anon->arcs_lsize[ARC_BUFC_METADATA]>>10, 5187 arc_anon->arcs_lsize[ARC_BUFC_DATA]>>10, 5188 reserve>>10, arc_c>>10); 5189 return (SET_ERROR(ERESTART)); 5190 } 5191 atomic_add_64(&arc_tempreserve, reserve); 5192 return (0); 5193} 5194 5195static void 5196arc_kstat_update_state(arc_state_t *state, kstat_named_t *size, 5197 kstat_named_t *evict_data, kstat_named_t *evict_metadata) 5198{ 5199 size->value.ui64 = refcount_count(&state->arcs_size); 5200 evict_data->value.ui64 = state->arcs_lsize[ARC_BUFC_DATA]; 5201 evict_metadata->value.ui64 = state->arcs_lsize[ARC_BUFC_METADATA]; 5202} 5203 5204static int 5205arc_kstat_update(kstat_t *ksp, int rw) 5206{ 5207 arc_stats_t *as = ksp->ks_data; 5208 5209 if (rw == KSTAT_WRITE) { 5210 return (EACCES); 5211 } else { 5212 arc_kstat_update_state(arc_anon, 5213 &as->arcstat_anon_size, 5214 &as->arcstat_anon_evictable_data, 5215 &as->arcstat_anon_evictable_metadata); 5216 arc_kstat_update_state(arc_mru, 5217 &as->arcstat_mru_size, 5218 &as->arcstat_mru_evictable_data, 5219 &as->arcstat_mru_evictable_metadata); 5220 arc_kstat_update_state(arc_mru_ghost, 5221 &as->arcstat_mru_ghost_size, 5222 &as->arcstat_mru_ghost_evictable_data, 5223 &as->arcstat_mru_ghost_evictable_metadata); 5224 arc_kstat_update_state(arc_mfu, 5225 &as->arcstat_mfu_size, 5226 &as->arcstat_mfu_evictable_data, 5227 &as->arcstat_mfu_evictable_metadata); 5228 arc_kstat_update_state(arc_mfu_ghost, 5229 &as->arcstat_mfu_ghost_size, 5230 &as->arcstat_mfu_ghost_evictable_data, 5231 &as->arcstat_mfu_ghost_evictable_metadata); 5232 } 5233 5234 return (0); 5235} 5236 5237/* 5238 * This function *must* return indices evenly distributed between all 5239 * sublists of the multilist. This is needed due to how the ARC eviction 5240 * code is laid out; arc_evict_state() assumes ARC buffers are evenly 5241 * distributed between all sublists and uses this assumption when 5242 * deciding which sublist to evict from and how much to evict from it. 5243 */ 5244unsigned int 5245arc_state_multilist_index_func(multilist_t *ml, void *obj) 5246{ 5247 arc_buf_hdr_t *hdr = obj; 5248 5249 /* 5250 * We rely on b_dva to generate evenly distributed index 5251 * numbers using buf_hash below. So, as an added precaution, 5252 * let's make sure we never add empty buffers to the arc lists. 5253 */ 5254 ASSERT(!BUF_EMPTY(hdr)); 5255 5256 /* 5257 * The assumption here, is the hash value for a given 5258 * arc_buf_hdr_t will remain constant throughout it's lifetime 5259 * (i.e. it's b_spa, b_dva, and b_birth fields don't change). 5260 * Thus, we don't need to store the header's sublist index 5261 * on insertion, as this index can be recalculated on removal. 5262 * 5263 * Also, the low order bits of the hash value are thought to be 5264 * distributed evenly. Otherwise, in the case that the multilist 5265 * has a power of two number of sublists, each sublists' usage 5266 * would not be evenly distributed. 5267 */ 5268 return (buf_hash(hdr->b_spa, &hdr->b_dva, hdr->b_birth) % 5269 multilist_get_num_sublists(ml)); 5270} 5271 5272#ifdef _KERNEL 5273static eventhandler_tag arc_event_lowmem = NULL; 5274 5275static void 5276arc_lowmem(void *arg __unused, int howto __unused) 5277{ 5278 5279 mutex_enter(&arc_reclaim_lock); 5280 /* XXX: Memory deficit should be passed as argument. */ 5281 needfree = btoc(arc_c >> arc_shrink_shift); 5282 DTRACE_PROBE(arc__needfree); 5283 cv_signal(&arc_reclaim_thread_cv); 5284 5285 /* 5286 * It is unsafe to block here in arbitrary threads, because we can come 5287 * here from ARC itself and may hold ARC locks and thus risk a deadlock 5288 * with ARC reclaim thread. 5289 */ 5290 if (curproc == pageproc) 5291 (void) cv_wait(&arc_reclaim_waiters_cv, &arc_reclaim_lock); 5292 mutex_exit(&arc_reclaim_lock); 5293} 5294#endif 5295 5296void 5297arc_init(void) 5298{ 5299 int i, prefetch_tunable_set = 0; 5300 5301 mutex_init(&arc_reclaim_lock, NULL, MUTEX_DEFAULT, NULL); 5302 cv_init(&arc_reclaim_thread_cv, NULL, CV_DEFAULT, NULL); 5303 cv_init(&arc_reclaim_waiters_cv, NULL, CV_DEFAULT, NULL); 5304 5305 mutex_init(&arc_user_evicts_lock, NULL, MUTEX_DEFAULT, NULL); 5306 cv_init(&arc_user_evicts_cv, NULL, CV_DEFAULT, NULL); 5307 5308 /* Convert seconds to clock ticks */ 5309 arc_min_prefetch_lifespan = 1 * hz; 5310 5311 /* Start out with 1/8 of all memory */ 5312 arc_c = kmem_size() / 8; 5313 5314#ifdef illumos 5315#ifdef _KERNEL 5316 /* 5317 * On architectures where the physical memory can be larger 5318 * than the addressable space (intel in 32-bit mode), we may 5319 * need to limit the cache to 1/8 of VM size. 5320 */ 5321 arc_c = MIN(arc_c, vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 8); 5322#endif 5323#endif /* illumos */ 5324 /* set min cache to 1/32 of all memory, or 16MB, whichever is more */ 5325 arc_c_min = MAX(arc_c / 4, 16 << 20); 5326 /* set max to 1/2 of all memory, or all but 1GB, whichever is more */ 5327 if (arc_c * 8 >= 1 << 30) 5328 arc_c_max = (arc_c * 8) - (1 << 30); 5329 else 5330 arc_c_max = arc_c_min; 5331 arc_c_max = MAX(arc_c * 5, arc_c_max); 5332 5333 /* 5334 * In userland, there's only the memory pressure that we artificially 5335 * create (see arc_available_memory()). Don't let arc_c get too 5336 * small, because it can cause transactions to be larger than 5337 * arc_c, causing arc_tempreserve_space() to fail. 5338 */ 5339#ifndef _KERNEL 5340 arc_c_min = arc_c_max / 2; 5341#endif 5342 5343#ifdef _KERNEL 5344 /* 5345 * Allow the tunables to override our calculations if they are 5346 * reasonable (ie. over 16MB) 5347 */ 5348 if (zfs_arc_max > 16 << 20 && zfs_arc_max < kmem_size()) 5349 arc_c_max = zfs_arc_max; 5350 if (zfs_arc_min > 16 << 20 && zfs_arc_min <= arc_c_max) 5351 arc_c_min = zfs_arc_min; 5352#endif 5353 5354 arc_c = arc_c_max; 5355 arc_p = (arc_c >> 1); 5356 5357 /* limit meta-data to 1/4 of the arc capacity */ 5358 arc_meta_limit = arc_c_max / 4; 5359 5360 /* Allow the tunable to override if it is reasonable */ 5361 if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max) 5362 arc_meta_limit = zfs_arc_meta_limit; 5363 5364 if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0) 5365 arc_c_min = arc_meta_limit / 2; 5366 5367 if (zfs_arc_meta_min > 0) { 5368 arc_meta_min = zfs_arc_meta_min; 5369 } else { 5370 arc_meta_min = arc_c_min / 2; 5371 } 5372 5373 if (zfs_arc_grow_retry > 0) 5374 arc_grow_retry = zfs_arc_grow_retry; 5375 5376 if (zfs_arc_shrink_shift > 0) 5377 arc_shrink_shift = zfs_arc_shrink_shift; 5378 5379 /* 5380 * Ensure that arc_no_grow_shift is less than arc_shrink_shift. 5381 */ 5382 if (arc_no_grow_shift >= arc_shrink_shift) 5383 arc_no_grow_shift = arc_shrink_shift - 1; 5384 5385 if (zfs_arc_p_min_shift > 0) 5386 arc_p_min_shift = zfs_arc_p_min_shift; 5387 5388 if (zfs_arc_num_sublists_per_state < 1) 5389 zfs_arc_num_sublists_per_state = MAX(max_ncpus, 1); 5390 5391 /* if kmem_flags are set, lets try to use less memory */ 5392 if (kmem_debugging()) 5393 arc_c = arc_c / 2; 5394 if (arc_c < arc_c_min) 5395 arc_c = arc_c_min; 5396 5397 zfs_arc_min = arc_c_min; 5398 zfs_arc_max = arc_c_max; 5399 5400 arc_anon = &ARC_anon; 5401 arc_mru = &ARC_mru; 5402 arc_mru_ghost = &ARC_mru_ghost; 5403 arc_mfu = &ARC_mfu; 5404 arc_mfu_ghost = &ARC_mfu_ghost; 5405 arc_l2c_only = &ARC_l2c_only; 5406 arc_size = 0; 5407 5408 multilist_create(&arc_mru->arcs_list[ARC_BUFC_METADATA], 5409 sizeof (arc_buf_hdr_t), 5410 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5411 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5412 multilist_create(&arc_mru->arcs_list[ARC_BUFC_DATA], 5413 sizeof (arc_buf_hdr_t), 5414 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5415 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5416 multilist_create(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA], 5417 sizeof (arc_buf_hdr_t), 5418 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5419 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5420 multilist_create(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA], 5421 sizeof (arc_buf_hdr_t), 5422 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5423 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5424 multilist_create(&arc_mfu->arcs_list[ARC_BUFC_METADATA], 5425 sizeof (arc_buf_hdr_t), 5426 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5427 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5428 multilist_create(&arc_mfu->arcs_list[ARC_BUFC_DATA], 5429 sizeof (arc_buf_hdr_t), 5430 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5431 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5432 multilist_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA], 5433 sizeof (arc_buf_hdr_t), 5434 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5435 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5436 multilist_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA], 5437 sizeof (arc_buf_hdr_t), 5438 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5439 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5440 multilist_create(&arc_l2c_only->arcs_list[ARC_BUFC_METADATA], 5441 sizeof (arc_buf_hdr_t), 5442 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5443 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5444 multilist_create(&arc_l2c_only->arcs_list[ARC_BUFC_DATA], 5445 sizeof (arc_buf_hdr_t), 5446 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5447 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5448 5449 refcount_create(&arc_anon->arcs_size); 5450 refcount_create(&arc_mru->arcs_size); 5451 refcount_create(&arc_mru_ghost->arcs_size); 5452 refcount_create(&arc_mfu->arcs_size); 5453 refcount_create(&arc_mfu_ghost->arcs_size); 5454 refcount_create(&arc_l2c_only->arcs_size); 5455 5456 buf_init(); 5457 5458 arc_reclaim_thread_exit = FALSE; 5459 arc_user_evicts_thread_exit = FALSE; 5460 arc_eviction_list = NULL; 5461 bzero(&arc_eviction_hdr, sizeof (arc_buf_hdr_t)); 5462 5463 arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED, 5464 sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL); 5465 5466 if (arc_ksp != NULL) { 5467 arc_ksp->ks_data = &arc_stats; 5468 arc_ksp->ks_update = arc_kstat_update; 5469 kstat_install(arc_ksp); 5470 } 5471 5472 (void) thread_create(NULL, 0, arc_reclaim_thread, NULL, 0, &p0, 5473 TS_RUN, minclsyspri); 5474 5475#ifdef _KERNEL 5476 arc_event_lowmem = EVENTHANDLER_REGISTER(vm_lowmem, arc_lowmem, NULL, 5477 EVENTHANDLER_PRI_FIRST); 5478#endif 5479 5480 (void) thread_create(NULL, 0, arc_user_evicts_thread, NULL, 0, &p0, 5481 TS_RUN, minclsyspri); 5482 5483 arc_dead = FALSE; 5484 arc_warm = B_FALSE; 5485 5486 /* 5487 * Calculate maximum amount of dirty data per pool. 5488 * 5489 * If it has been set by /etc/system, take that. 5490 * Otherwise, use a percentage of physical memory defined by 5491 * zfs_dirty_data_max_percent (default 10%) with a cap at 5492 * zfs_dirty_data_max_max (default 4GB). 5493 */ 5494 if (zfs_dirty_data_max == 0) { 5495 zfs_dirty_data_max = ptob(physmem) * 5496 zfs_dirty_data_max_percent / 100; 5497 zfs_dirty_data_max = MIN(zfs_dirty_data_max, 5498 zfs_dirty_data_max_max); 5499 } 5500 5501#ifdef _KERNEL 5502 if (TUNABLE_INT_FETCH("vfs.zfs.prefetch_disable", &zfs_prefetch_disable)) 5503 prefetch_tunable_set = 1; 5504 5505#ifdef __i386__ 5506 if (prefetch_tunable_set == 0) { 5507 printf("ZFS NOTICE: Prefetch is disabled by default on i386 " 5508 "-- to enable,\n"); 5509 printf(" add \"vfs.zfs.prefetch_disable=0\" " 5510 "to /boot/loader.conf.\n"); 5511 zfs_prefetch_disable = 1; 5512 } 5513#else 5514 if ((((uint64_t)physmem * PAGESIZE) < (1ULL << 32)) && 5515 prefetch_tunable_set == 0) { 5516 printf("ZFS NOTICE: Prefetch is disabled by default if less " 5517 "than 4GB of RAM is present;\n" 5518 " to enable, add \"vfs.zfs.prefetch_disable=0\" " 5519 "to /boot/loader.conf.\n"); 5520 zfs_prefetch_disable = 1; 5521 } 5522#endif 5523 /* Warn about ZFS memory and address space requirements. */ 5524 if (((uint64_t)physmem * PAGESIZE) < (256 + 128 + 64) * (1 << 20)) { 5525 printf("ZFS WARNING: Recommended minimum RAM size is 512MB; " 5526 "expect unstable behavior.\n"); 5527 } 5528 if (kmem_size() < 512 * (1 << 20)) { 5529 printf("ZFS WARNING: Recommended minimum kmem_size is 512MB; " 5530 "expect unstable behavior.\n"); 5531 printf(" Consider tuning vm.kmem_size and " 5532 "vm.kmem_size_max\n"); 5533 printf(" in /boot/loader.conf.\n"); 5534 } 5535#endif 5536} 5537 5538void 5539arc_fini(void) 5540{ 5541 mutex_enter(&arc_reclaim_lock); 5542 arc_reclaim_thread_exit = TRUE; 5543 /* 5544 * The reclaim thread will set arc_reclaim_thread_exit back to 5545 * FALSE when it is finished exiting; we're waiting for that. 5546 */ 5547 while (arc_reclaim_thread_exit) { 5548 cv_signal(&arc_reclaim_thread_cv); 5549 cv_wait(&arc_reclaim_thread_cv, &arc_reclaim_lock); 5550 } 5551 mutex_exit(&arc_reclaim_lock); 5552 5553 mutex_enter(&arc_user_evicts_lock); 5554 arc_user_evicts_thread_exit = TRUE; 5555 /* 5556 * The user evicts thread will set arc_user_evicts_thread_exit 5557 * to FALSE when it is finished exiting; we're waiting for that. 5558 */ 5559 while (arc_user_evicts_thread_exit) { 5560 cv_signal(&arc_user_evicts_cv); 5561 cv_wait(&arc_user_evicts_cv, &arc_user_evicts_lock); 5562 } 5563 mutex_exit(&arc_user_evicts_lock); 5564 5565 /* Use TRUE to ensure *all* buffers are evicted */ 5566 arc_flush(NULL, TRUE); 5567 5568 arc_dead = TRUE; 5569 5570 if (arc_ksp != NULL) { 5571 kstat_delete(arc_ksp); 5572 arc_ksp = NULL; 5573 } 5574 5575 mutex_destroy(&arc_reclaim_lock); 5576 cv_destroy(&arc_reclaim_thread_cv); 5577 cv_destroy(&arc_reclaim_waiters_cv); 5578 5579 mutex_destroy(&arc_user_evicts_lock); 5580 cv_destroy(&arc_user_evicts_cv); 5581 5582 refcount_destroy(&arc_anon->arcs_size); 5583 refcount_destroy(&arc_mru->arcs_size); 5584 refcount_destroy(&arc_mru_ghost->arcs_size); 5585 refcount_destroy(&arc_mfu->arcs_size); 5586 refcount_destroy(&arc_mfu_ghost->arcs_size); 5587 refcount_destroy(&arc_l2c_only->arcs_size); 5588 5589 multilist_destroy(&arc_mru->arcs_list[ARC_BUFC_METADATA]); 5590 multilist_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA]); 5591 multilist_destroy(&arc_mfu->arcs_list[ARC_BUFC_METADATA]); 5592 multilist_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA]); 5593 multilist_destroy(&arc_l2c_only->arcs_list[ARC_BUFC_METADATA]); 5594 multilist_destroy(&arc_mru->arcs_list[ARC_BUFC_DATA]); 5595 multilist_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA]); 5596 multilist_destroy(&arc_mfu->arcs_list[ARC_BUFC_DATA]); 5597 multilist_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA]); 5598 multilist_destroy(&arc_l2c_only->arcs_list[ARC_BUFC_DATA]); 5599 5600 buf_fini(); 5601 5602 ASSERT0(arc_loaned_bytes); 5603 5604#ifdef _KERNEL 5605 if (arc_event_lowmem != NULL) 5606 EVENTHANDLER_DEREGISTER(vm_lowmem, arc_event_lowmem); 5607#endif 5608} 5609 5610/* 5611 * Level 2 ARC 5612 * 5613 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk. 5614 * It uses dedicated storage devices to hold cached data, which are populated 5615 * using large infrequent writes. The main role of this cache is to boost 5616 * the performance of random read workloads. The intended L2ARC devices 5617 * include short-stroked disks, solid state disks, and other media with 5618 * substantially faster read latency than disk. 5619 * 5620 * +-----------------------+ 5621 * | ARC | 5622 * +-----------------------+ 5623 * | ^ ^ 5624 * | | | 5625 * l2arc_feed_thread() arc_read() 5626 * | | | 5627 * | l2arc read | 5628 * V | | 5629 * +---------------+ | 5630 * | L2ARC | | 5631 * +---------------+ | 5632 * | ^ | 5633 * l2arc_write() | | 5634 * | | | 5635 * V | | 5636 * +-------+ +-------+ 5637 * | vdev | | vdev | 5638 * | cache | | cache | 5639 * +-------+ +-------+ 5640 * +=========+ .-----. 5641 * : L2ARC : |-_____-| 5642 * : devices : | Disks | 5643 * +=========+ `-_____-' 5644 * 5645 * Read requests are satisfied from the following sources, in order: 5646 * 5647 * 1) ARC 5648 * 2) vdev cache of L2ARC devices 5649 * 3) L2ARC devices 5650 * 4) vdev cache of disks 5651 * 5) disks 5652 * 5653 * Some L2ARC device types exhibit extremely slow write performance. 5654 * To accommodate for this there are some significant differences between 5655 * the L2ARC and traditional cache design: 5656 * 5657 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from 5658 * the ARC behave as usual, freeing buffers and placing headers on ghost 5659 * lists. The ARC does not send buffers to the L2ARC during eviction as 5660 * this would add inflated write latencies for all ARC memory pressure. 5661 * 5662 * 2. The L2ARC attempts to cache data from the ARC before it is evicted. 5663 * It does this by periodically scanning buffers from the eviction-end of 5664 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are 5665 * not already there. It scans until a headroom of buffers is satisfied, 5666 * which itself is a buffer for ARC eviction. If a compressible buffer is 5667 * found during scanning and selected for writing to an L2ARC device, we 5668 * temporarily boost scanning headroom during the next scan cycle to make 5669 * sure we adapt to compression effects (which might significantly reduce 5670 * the data volume we write to L2ARC). The thread that does this is 5671 * l2arc_feed_thread(), illustrated below; example sizes are included to 5672 * provide a better sense of ratio than this diagram: 5673 * 5674 * head --> tail 5675 * +---------------------+----------+ 5676 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC 5677 * +---------------------+----------+ | o L2ARC eligible 5678 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer 5679 * +---------------------+----------+ | 5680 * 15.9 Gbytes ^ 32 Mbytes | 5681 * headroom | 5682 * l2arc_feed_thread() 5683 * | 5684 * l2arc write hand <--[oooo]--' 5685 * | 8 Mbyte 5686 * | write max 5687 * V 5688 * +==============================+ 5689 * L2ARC dev |####|#|###|###| |####| ... | 5690 * +==============================+ 5691 * 32 Gbytes 5692 * 5693 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of 5694 * evicted, then the L2ARC has cached a buffer much sooner than it probably 5695 * needed to, potentially wasting L2ARC device bandwidth and storage. It is 5696 * safe to say that this is an uncommon case, since buffers at the end of 5697 * the ARC lists have moved there due to inactivity. 5698 * 5699 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom, 5700 * then the L2ARC simply misses copying some buffers. This serves as a 5701 * pressure valve to prevent heavy read workloads from both stalling the ARC 5702 * with waits and clogging the L2ARC with writes. This also helps prevent 5703 * the potential for the L2ARC to churn if it attempts to cache content too 5704 * quickly, such as during backups of the entire pool. 5705 * 5706 * 5. After system boot and before the ARC has filled main memory, there are 5707 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru 5708 * lists can remain mostly static. Instead of searching from tail of these 5709 * lists as pictured, the l2arc_feed_thread() will search from the list heads 5710 * for eligible buffers, greatly increasing its chance of finding them. 5711 * 5712 * The L2ARC device write speed is also boosted during this time so that 5713 * the L2ARC warms up faster. Since there have been no ARC evictions yet, 5714 * there are no L2ARC reads, and no fear of degrading read performance 5715 * through increased writes. 5716 * 5717 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that 5718 * the vdev queue can aggregate them into larger and fewer writes. Each 5719 * device is written to in a rotor fashion, sweeping writes through 5720 * available space then repeating. 5721 * 5722 * 7. The L2ARC does not store dirty content. It never needs to flush 5723 * write buffers back to disk based storage. 5724 * 5725 * 8. If an ARC buffer is written (and dirtied) which also exists in the 5726 * L2ARC, the now stale L2ARC buffer is immediately dropped. 5727 * 5728 * The performance of the L2ARC can be tweaked by a number of tunables, which 5729 * may be necessary for different workloads: 5730 * 5731 * l2arc_write_max max write bytes per interval 5732 * l2arc_write_boost extra write bytes during device warmup 5733 * l2arc_noprefetch skip caching prefetched buffers 5734 * l2arc_headroom number of max device writes to precache 5735 * l2arc_headroom_boost when we find compressed buffers during ARC 5736 * scanning, we multiply headroom by this 5737 * percentage factor for the next scan cycle, 5738 * since more compressed buffers are likely to 5739 * be present 5740 * l2arc_feed_secs seconds between L2ARC writing 5741 * 5742 * Tunables may be removed or added as future performance improvements are 5743 * integrated, and also may become zpool properties. 5744 * 5745 * There are three key functions that control how the L2ARC warms up: 5746 * 5747 * l2arc_write_eligible() check if a buffer is eligible to cache 5748 * l2arc_write_size() calculate how much to write 5749 * l2arc_write_interval() calculate sleep delay between writes 5750 * 5751 * These three functions determine what to write, how much, and how quickly 5752 * to send writes. 5753 */ 5754 5755static boolean_t 5756l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *hdr) 5757{ 5758 /* 5759 * A buffer is *not* eligible for the L2ARC if it: 5760 * 1. belongs to a different spa. 5761 * 2. is already cached on the L2ARC. 5762 * 3. has an I/O in progress (it may be an incomplete read). 5763 * 4. is flagged not eligible (zfs property). 5764 */ 5765 if (hdr->b_spa != spa_guid) { 5766 ARCSTAT_BUMP(arcstat_l2_write_spa_mismatch); 5767 return (B_FALSE); 5768 } 5769 if (HDR_HAS_L2HDR(hdr)) { 5770 ARCSTAT_BUMP(arcstat_l2_write_in_l2); 5771 return (B_FALSE); 5772 } 5773 if (HDR_IO_IN_PROGRESS(hdr)) { 5774 ARCSTAT_BUMP(arcstat_l2_write_hdr_io_in_progress); 5775 return (B_FALSE); 5776 } 5777 if (!HDR_L2CACHE(hdr)) { 5778 ARCSTAT_BUMP(arcstat_l2_write_not_cacheable); 5779 return (B_FALSE); 5780 } 5781 5782 return (B_TRUE); 5783} 5784 5785static uint64_t 5786l2arc_write_size(void) 5787{ 5788 uint64_t size; 5789 5790 /* 5791 * Make sure our globals have meaningful values in case the user 5792 * altered them. 5793 */ 5794 size = l2arc_write_max; 5795 if (size == 0) { 5796 cmn_err(CE_NOTE, "Bad value for l2arc_write_max, value must " 5797 "be greater than zero, resetting it to the default (%d)", 5798 L2ARC_WRITE_SIZE); 5799 size = l2arc_write_max = L2ARC_WRITE_SIZE; 5800 } 5801 5802 if (arc_warm == B_FALSE) 5803 size += l2arc_write_boost; 5804 5805 return (size); 5806 5807} 5808 5809static clock_t 5810l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote) 5811{ 5812 clock_t interval, next, now; 5813 5814 /* 5815 * If the ARC lists are busy, increase our write rate; if the 5816 * lists are stale, idle back. This is achieved by checking 5817 * how much we previously wrote - if it was more than half of 5818 * what we wanted, schedule the next write much sooner. 5819 */ 5820 if (l2arc_feed_again && wrote > (wanted / 2)) 5821 interval = (hz * l2arc_feed_min_ms) / 1000; 5822 else 5823 interval = hz * l2arc_feed_secs; 5824 5825 now = ddi_get_lbolt(); 5826 next = MAX(now, MIN(now + interval, began + interval)); 5827 5828 return (next); 5829} 5830 5831/* 5832 * Cycle through L2ARC devices. This is how L2ARC load balances. 5833 * If a device is returned, this also returns holding the spa config lock. 5834 */ 5835static l2arc_dev_t * 5836l2arc_dev_get_next(void) 5837{ 5838 l2arc_dev_t *first, *next = NULL; 5839 5840 /* 5841 * Lock out the removal of spas (spa_namespace_lock), then removal 5842 * of cache devices (l2arc_dev_mtx). Once a device has been selected, 5843 * both locks will be dropped and a spa config lock held instead. 5844 */ 5845 mutex_enter(&spa_namespace_lock); 5846 mutex_enter(&l2arc_dev_mtx); 5847 5848 /* if there are no vdevs, there is nothing to do */ 5849 if (l2arc_ndev == 0) 5850 goto out; 5851 5852 first = NULL; 5853 next = l2arc_dev_last; 5854 do { 5855 /* loop around the list looking for a non-faulted vdev */ 5856 if (next == NULL) { 5857 next = list_head(l2arc_dev_list); 5858 } else { 5859 next = list_next(l2arc_dev_list, next); 5860 if (next == NULL) 5861 next = list_head(l2arc_dev_list); 5862 } 5863 5864 /* if we have come back to the start, bail out */ 5865 if (first == NULL) 5866 first = next; 5867 else if (next == first) 5868 break; 5869 5870 } while (vdev_is_dead(next->l2ad_vdev)); 5871 5872 /* if we were unable to find any usable vdevs, return NULL */ 5873 if (vdev_is_dead(next->l2ad_vdev)) 5874 next = NULL; 5875 5876 l2arc_dev_last = next; 5877 5878out: 5879 mutex_exit(&l2arc_dev_mtx); 5880 5881 /* 5882 * Grab the config lock to prevent the 'next' device from being 5883 * removed while we are writing to it. 5884 */ 5885 if (next != NULL) 5886 spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER); 5887 mutex_exit(&spa_namespace_lock); 5888 5889 return (next); 5890} 5891 5892/* 5893 * Free buffers that were tagged for destruction. 5894 */ 5895static void 5896l2arc_do_free_on_write() 5897{ 5898 list_t *buflist; 5899 l2arc_data_free_t *df, *df_prev; 5900 5901 mutex_enter(&l2arc_free_on_write_mtx); 5902 buflist = l2arc_free_on_write; 5903 5904 for (df = list_tail(buflist); df; df = df_prev) { 5905 df_prev = list_prev(buflist, df); 5906 ASSERT(df->l2df_data != NULL); 5907 ASSERT(df->l2df_func != NULL); 5908 df->l2df_func(df->l2df_data, df->l2df_size); 5909 list_remove(buflist, df); 5910 kmem_free(df, sizeof (l2arc_data_free_t)); 5911 } 5912 5913 mutex_exit(&l2arc_free_on_write_mtx); 5914} 5915 5916/* 5917 * A write to a cache device has completed. Update all headers to allow 5918 * reads from these buffers to begin. 5919 */ 5920static void 5921l2arc_write_done(zio_t *zio) 5922{ 5923 l2arc_write_callback_t *cb; 5924 l2arc_dev_t *dev; 5925 list_t *buflist; 5926 arc_buf_hdr_t *head, *hdr, *hdr_prev; 5927 kmutex_t *hash_lock; 5928 int64_t bytes_dropped = 0; 5929 5930 cb = zio->io_private; 5931 ASSERT(cb != NULL); 5932 dev = cb->l2wcb_dev; 5933 ASSERT(dev != NULL); 5934 head = cb->l2wcb_head; 5935 ASSERT(head != NULL); 5936 buflist = &dev->l2ad_buflist; 5937 ASSERT(buflist != NULL); 5938 DTRACE_PROBE2(l2arc__iodone, zio_t *, zio, 5939 l2arc_write_callback_t *, cb); 5940 5941 if (zio->io_error != 0) 5942 ARCSTAT_BUMP(arcstat_l2_writes_error); 5943 5944 /* 5945 * All writes completed, or an error was hit. 5946 */ 5947top: 5948 mutex_enter(&dev->l2ad_mtx); 5949 for (hdr = list_prev(buflist, head); hdr; hdr = hdr_prev) { 5950 hdr_prev = list_prev(buflist, hdr); 5951 5952 hash_lock = HDR_LOCK(hdr); 5953 5954 /* 5955 * We cannot use mutex_enter or else we can deadlock 5956 * with l2arc_write_buffers (due to swapping the order 5957 * the hash lock and l2ad_mtx are taken). 5958 */ 5959 if (!mutex_tryenter(hash_lock)) { 5960 /* 5961 * Missed the hash lock. We must retry so we 5962 * don't leave the ARC_FLAG_L2_WRITING bit set. 5963 */ 5964 ARCSTAT_BUMP(arcstat_l2_writes_lock_retry); 5965 5966 /* 5967 * We don't want to rescan the headers we've 5968 * already marked as having been written out, so 5969 * we reinsert the head node so we can pick up 5970 * where we left off. 5971 */ 5972 list_remove(buflist, head); 5973 list_insert_after(buflist, hdr, head); 5974 5975 mutex_exit(&dev->l2ad_mtx); 5976 5977 /* 5978 * We wait for the hash lock to become available 5979 * to try and prevent busy waiting, and increase 5980 * the chance we'll be able to acquire the lock 5981 * the next time around. 5982 */ 5983 mutex_enter(hash_lock); 5984 mutex_exit(hash_lock); 5985 goto top; 5986 } 5987 5988 /* 5989 * We could not have been moved into the arc_l2c_only 5990 * state while in-flight due to our ARC_FLAG_L2_WRITING 5991 * bit being set. Let's just ensure that's being enforced. 5992 */ 5993 ASSERT(HDR_HAS_L1HDR(hdr)); 5994 5995 /* 5996 * We may have allocated a buffer for L2ARC compression, 5997 * we must release it to avoid leaking this data. 5998 */ 5999 l2arc_release_cdata_buf(hdr); 6000 6001 if (zio->io_error != 0) { 6002 /* 6003 * Error - drop L2ARC entry. 6004 */ 6005 list_remove(buflist, hdr); 6006 l2arc_trim(hdr); 6007 hdr->b_flags &= ~ARC_FLAG_HAS_L2HDR; 6008 6009 ARCSTAT_INCR(arcstat_l2_asize, -hdr->b_l2hdr.b_asize); 6010 ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size); 6011 6012 bytes_dropped += hdr->b_l2hdr.b_asize; 6013 (void) refcount_remove_many(&dev->l2ad_alloc, 6014 hdr->b_l2hdr.b_asize, hdr); 6015 } 6016 6017 /* 6018 * Allow ARC to begin reads and ghost list evictions to 6019 * this L2ARC entry. 6020 */ 6021 hdr->b_flags &= ~ARC_FLAG_L2_WRITING; 6022 6023 mutex_exit(hash_lock); 6024 } 6025 6026 atomic_inc_64(&l2arc_writes_done); 6027 list_remove(buflist, head); 6028 ASSERT(!HDR_HAS_L1HDR(head)); 6029 kmem_cache_free(hdr_l2only_cache, head); 6030 mutex_exit(&dev->l2ad_mtx); 6031 6032 vdev_space_update(dev->l2ad_vdev, -bytes_dropped, 0, 0); 6033 6034 l2arc_do_free_on_write(); 6035 6036 kmem_free(cb, sizeof (l2arc_write_callback_t)); 6037} 6038 6039/* 6040 * A read to a cache device completed. Validate buffer contents before 6041 * handing over to the regular ARC routines. 6042 */ 6043static void 6044l2arc_read_done(zio_t *zio) 6045{ 6046 l2arc_read_callback_t *cb; 6047 arc_buf_hdr_t *hdr; 6048 arc_buf_t *buf; 6049 kmutex_t *hash_lock; 6050 int equal; 6051 6052 ASSERT(zio->io_vd != NULL); 6053 ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE); 6054 6055 spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd); 6056 6057 cb = zio->io_private; 6058 ASSERT(cb != NULL); 6059 buf = cb->l2rcb_buf; 6060 ASSERT(buf != NULL); 6061 6062 hash_lock = HDR_LOCK(buf->b_hdr); 6063 mutex_enter(hash_lock); 6064 hdr = buf->b_hdr; 6065 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 6066 6067 /* 6068 * If the data was read into a temporary buffer, 6069 * move it and free the buffer. 6070 */ 6071 if (cb->l2rcb_data != NULL) { 6072 ASSERT3U(hdr->b_size, <, zio->io_size); 6073 ASSERT3U(cb->l2rcb_compress, ==, ZIO_COMPRESS_OFF); 6074 if (zio->io_error == 0) 6075 bcopy(cb->l2rcb_data, buf->b_data, hdr->b_size); 6076 6077 /* 6078 * The following must be done regardless of whether 6079 * there was an error: 6080 * - free the temporary buffer 6081 * - point zio to the real ARC buffer 6082 * - set zio size accordingly 6083 * These are required because zio is either re-used for 6084 * an I/O of the block in the case of the error 6085 * or the zio is passed to arc_read_done() and it 6086 * needs real data. 6087 */ 6088 zio_data_buf_free(cb->l2rcb_data, zio->io_size); 6089 zio->io_size = zio->io_orig_size = hdr->b_size; 6090 zio->io_data = zio->io_orig_data = buf->b_data; 6091 } 6092 6093 /* 6094 * If the buffer was compressed, decompress it first. 6095 */ 6096 if (cb->l2rcb_compress != ZIO_COMPRESS_OFF) 6097 l2arc_decompress_zio(zio, hdr, cb->l2rcb_compress); 6098 ASSERT(zio->io_data != NULL); 6099 ASSERT3U(zio->io_size, ==, hdr->b_size); 6100 ASSERT3U(BP_GET_LSIZE(&cb->l2rcb_bp), ==, hdr->b_size); 6101 6102 /* 6103 * Check this survived the L2ARC journey. 6104 */ 6105 equal = arc_cksum_equal(buf); 6106 if (equal && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) { 6107 mutex_exit(hash_lock); 6108 zio->io_private = buf; 6109 zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */ 6110 zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */ 6111 arc_read_done(zio); 6112 } else { 6113 mutex_exit(hash_lock); 6114 /* 6115 * Buffer didn't survive caching. Increment stats and 6116 * reissue to the original storage device. 6117 */ 6118 if (zio->io_error != 0) { 6119 ARCSTAT_BUMP(arcstat_l2_io_error); 6120 } else { 6121 zio->io_error = SET_ERROR(EIO); 6122 } 6123 if (!equal) 6124 ARCSTAT_BUMP(arcstat_l2_cksum_bad); 6125 6126 /* 6127 * If there's no waiter, issue an async i/o to the primary 6128 * storage now. If there *is* a waiter, the caller must 6129 * issue the i/o in a context where it's OK to block. 6130 */ 6131 if (zio->io_waiter == NULL) { 6132 zio_t *pio = zio_unique_parent(zio); 6133 6134 ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL); 6135 6136 zio_nowait(zio_read(pio, cb->l2rcb_spa, &cb->l2rcb_bp, 6137 buf->b_data, hdr->b_size, arc_read_done, buf, 6138 zio->io_priority, cb->l2rcb_flags, &cb->l2rcb_zb)); 6139 } 6140 } 6141 6142 kmem_free(cb, sizeof (l2arc_read_callback_t)); 6143} 6144 6145/* 6146 * This is the list priority from which the L2ARC will search for pages to 6147 * cache. This is used within loops (0..3) to cycle through lists in the 6148 * desired order. This order can have a significant effect on cache 6149 * performance. 6150 * 6151 * Currently the metadata lists are hit first, MFU then MRU, followed by 6152 * the data lists. This function returns a locked list, and also returns 6153 * the lock pointer. 6154 */ 6155static multilist_sublist_t * 6156l2arc_sublist_lock(int list_num) 6157{ 6158 multilist_t *ml = NULL; 6159 unsigned int idx; 6160 6161 ASSERT(list_num >= 0 && list_num <= 3); 6162 6163 switch (list_num) { 6164 case 0: 6165 ml = &arc_mfu->arcs_list[ARC_BUFC_METADATA]; 6166 break; 6167 case 1: 6168 ml = &arc_mru->arcs_list[ARC_BUFC_METADATA]; 6169 break; 6170 case 2: 6171 ml = &arc_mfu->arcs_list[ARC_BUFC_DATA]; 6172 break; 6173 case 3: 6174 ml = &arc_mru->arcs_list[ARC_BUFC_DATA]; 6175 break; 6176 } 6177 6178 /* 6179 * Return a randomly-selected sublist. This is acceptable 6180 * because the caller feeds only a little bit of data for each 6181 * call (8MB). Subsequent calls will result in different 6182 * sublists being selected. 6183 */ 6184 idx = multilist_get_random_index(ml); 6185 return (multilist_sublist_lock(ml, idx)); 6186} 6187 6188/* 6189 * Evict buffers from the device write hand to the distance specified in 6190 * bytes. This distance may span populated buffers, it may span nothing. 6191 * This is clearing a region on the L2ARC device ready for writing. 6192 * If the 'all' boolean is set, every buffer is evicted. 6193 */ 6194static void 6195l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all) 6196{ 6197 list_t *buflist; 6198 arc_buf_hdr_t *hdr, *hdr_prev; 6199 kmutex_t *hash_lock; 6200 uint64_t taddr; 6201 6202 buflist = &dev->l2ad_buflist; 6203 6204 if (!all && dev->l2ad_first) { 6205 /* 6206 * This is the first sweep through the device. There is 6207 * nothing to evict. 6208 */ 6209 return; 6210 } 6211 6212 if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) { 6213 /* 6214 * When nearing the end of the device, evict to the end 6215 * before the device write hand jumps to the start. 6216 */ 6217 taddr = dev->l2ad_end; 6218 } else { 6219 taddr = dev->l2ad_hand + distance; 6220 } 6221 DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist, 6222 uint64_t, taddr, boolean_t, all); 6223 6224top: 6225 mutex_enter(&dev->l2ad_mtx); 6226 for (hdr = list_tail(buflist); hdr; hdr = hdr_prev) { 6227 hdr_prev = list_prev(buflist, hdr); 6228 6229 hash_lock = HDR_LOCK(hdr); 6230 6231 /* 6232 * We cannot use mutex_enter or else we can deadlock 6233 * with l2arc_write_buffers (due to swapping the order 6234 * the hash lock and l2ad_mtx are taken). 6235 */ 6236 if (!mutex_tryenter(hash_lock)) { 6237 /* 6238 * Missed the hash lock. Retry. 6239 */ 6240 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry); 6241 mutex_exit(&dev->l2ad_mtx); 6242 mutex_enter(hash_lock); 6243 mutex_exit(hash_lock); 6244 goto top; 6245 } 6246 6247 if (HDR_L2_WRITE_HEAD(hdr)) { 6248 /* 6249 * We hit a write head node. Leave it for 6250 * l2arc_write_done(). 6251 */ 6252 list_remove(buflist, hdr); 6253 mutex_exit(hash_lock); 6254 continue; 6255 } 6256 6257 if (!all && HDR_HAS_L2HDR(hdr) && 6258 (hdr->b_l2hdr.b_daddr > taddr || 6259 hdr->b_l2hdr.b_daddr < dev->l2ad_hand)) { 6260 /* 6261 * We've evicted to the target address, 6262 * or the end of the device. 6263 */ 6264 mutex_exit(hash_lock); 6265 break; 6266 } 6267 6268 ASSERT(HDR_HAS_L2HDR(hdr)); 6269 if (!HDR_HAS_L1HDR(hdr)) { 6270 ASSERT(!HDR_L2_READING(hdr)); 6271 /* 6272 * This doesn't exist in the ARC. Destroy. 6273 * arc_hdr_destroy() will call list_remove() 6274 * and decrement arcstat_l2_size. 6275 */ 6276 arc_change_state(arc_anon, hdr, hash_lock); 6277 arc_hdr_destroy(hdr); 6278 } else { 6279 ASSERT(hdr->b_l1hdr.b_state != arc_l2c_only); 6280 ARCSTAT_BUMP(arcstat_l2_evict_l1cached); 6281 /* 6282 * Invalidate issued or about to be issued 6283 * reads, since we may be about to write 6284 * over this location. 6285 */ 6286 if (HDR_L2_READING(hdr)) { 6287 ARCSTAT_BUMP(arcstat_l2_evict_reading); 6288 hdr->b_flags |= ARC_FLAG_L2_EVICTED; 6289 } 6290 6291 /* Ensure this header has finished being written */ 6292 ASSERT(!HDR_L2_WRITING(hdr)); 6293 ASSERT3P(hdr->b_l1hdr.b_tmp_cdata, ==, NULL); 6294 6295 arc_hdr_l2hdr_destroy(hdr); 6296 } 6297 mutex_exit(hash_lock); 6298 } 6299 mutex_exit(&dev->l2ad_mtx); 6300} 6301 6302/* 6303 * Find and write ARC buffers to the L2ARC device. 6304 * 6305 * An ARC_FLAG_L2_WRITING flag is set so that the L2ARC buffers are not valid 6306 * for reading until they have completed writing. 6307 * The headroom_boost is an in-out parameter used to maintain headroom boost 6308 * state between calls to this function. 6309 * 6310 * Returns the number of bytes actually written (which may be smaller than 6311 * the delta by which the device hand has changed due to alignment). 6312 */ 6313static uint64_t 6314l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz, 6315 boolean_t *headroom_boost) 6316{ 6317 arc_buf_hdr_t *hdr, *hdr_prev, *head; 6318 uint64_t write_asize, write_sz, headroom, 6319 buf_compress_minsz; 6320 void *buf_data; 6321 boolean_t full; 6322 l2arc_write_callback_t *cb; 6323 zio_t *pio, *wzio; 6324 uint64_t guid = spa_load_guid(spa); 6325 const boolean_t do_headroom_boost = *headroom_boost; 6326 int try; 6327 6328 ASSERT(dev->l2ad_vdev != NULL); 6329 6330 /* Lower the flag now, we might want to raise it again later. */ 6331 *headroom_boost = B_FALSE; 6332 6333 pio = NULL; 6334 write_sz = write_asize = 0; 6335 full = B_FALSE; 6336 head = kmem_cache_alloc(hdr_l2only_cache, KM_PUSHPAGE); 6337 head->b_flags |= ARC_FLAG_L2_WRITE_HEAD; 6338 head->b_flags |= ARC_FLAG_HAS_L2HDR; 6339 6340 ARCSTAT_BUMP(arcstat_l2_write_buffer_iter); 6341 /* 6342 * We will want to try to compress buffers that are at least 2x the 6343 * device sector size. 6344 */ 6345 buf_compress_minsz = 2 << dev->l2ad_vdev->vdev_ashift; 6346 6347 /* 6348 * Copy buffers for L2ARC writing. 6349 */ 6350 for (try = 0; try <= 3; try++) { 6351 multilist_sublist_t *mls = l2arc_sublist_lock(try); 6352 uint64_t passed_sz = 0; 6353 6354 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_iter); 6355 6356 /* 6357 * L2ARC fast warmup. 6358 * 6359 * Until the ARC is warm and starts to evict, read from the 6360 * head of the ARC lists rather than the tail. 6361 */ 6362 if (arc_warm == B_FALSE) 6363 hdr = multilist_sublist_head(mls); 6364 else 6365 hdr = multilist_sublist_tail(mls); 6366 if (hdr == NULL) 6367 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_null_iter); 6368 6369 headroom = target_sz * l2arc_headroom; 6370 if (do_headroom_boost) 6371 headroom = (headroom * l2arc_headroom_boost) / 100; 6372 6373 for (; hdr; hdr = hdr_prev) { 6374 kmutex_t *hash_lock; 6375 uint64_t buf_sz; 6376 uint64_t buf_a_sz; 6377 size_t align; 6378 6379 if (arc_warm == B_FALSE) 6380 hdr_prev = multilist_sublist_next(mls, hdr); 6381 else 6382 hdr_prev = multilist_sublist_prev(mls, hdr); 6383 ARCSTAT_INCR(arcstat_l2_write_buffer_bytes_scanned, hdr->b_size); 6384 6385 hash_lock = HDR_LOCK(hdr); 6386 if (!mutex_tryenter(hash_lock)) { 6387 ARCSTAT_BUMP(arcstat_l2_write_trylock_fail); 6388 /* 6389 * Skip this buffer rather than waiting. 6390 */ 6391 continue; 6392 } 6393 6394 passed_sz += hdr->b_size; 6395 if (passed_sz > headroom) { 6396 /* 6397 * Searched too far. 6398 */ 6399 mutex_exit(hash_lock); 6400 ARCSTAT_BUMP(arcstat_l2_write_passed_headroom); 6401 break; 6402 } 6403 6404 if (!l2arc_write_eligible(guid, hdr)) { 6405 mutex_exit(hash_lock); 6406 continue; 6407 } 6408 6409 /* 6410 * Assume that the buffer is not going to be compressed 6411 * and could take more space on disk because of a larger 6412 * disk block size. 6413 */ 6414 buf_sz = hdr->b_size; 6415 align = (size_t)1 << dev->l2ad_vdev->vdev_ashift; 6416 buf_a_sz = P2ROUNDUP(buf_sz, align); 6417 6418 if ((write_asize + buf_a_sz) > target_sz) { 6419 full = B_TRUE; 6420 mutex_exit(hash_lock); 6421 ARCSTAT_BUMP(arcstat_l2_write_full); 6422 break; 6423 } 6424 6425 if (pio == NULL) { 6426 /* 6427 * Insert a dummy header on the buflist so 6428 * l2arc_write_done() can find where the 6429 * write buffers begin without searching. 6430 */ 6431 mutex_enter(&dev->l2ad_mtx); 6432 list_insert_head(&dev->l2ad_buflist, head); 6433 mutex_exit(&dev->l2ad_mtx); 6434 6435 cb = kmem_alloc( 6436 sizeof (l2arc_write_callback_t), KM_SLEEP); 6437 cb->l2wcb_dev = dev; 6438 cb->l2wcb_head = head; 6439 pio = zio_root(spa, l2arc_write_done, cb, 6440 ZIO_FLAG_CANFAIL); 6441 ARCSTAT_BUMP(arcstat_l2_write_pios); 6442 } 6443 6444 /* 6445 * Create and add a new L2ARC header. 6446 */ 6447 hdr->b_l2hdr.b_dev = dev; 6448 hdr->b_flags |= ARC_FLAG_L2_WRITING; 6449 /* 6450 * Temporarily stash the data buffer in b_tmp_cdata. 6451 * The subsequent write step will pick it up from 6452 * there. This is because can't access b_l1hdr.b_buf 6453 * without holding the hash_lock, which we in turn 6454 * can't access without holding the ARC list locks 6455 * (which we want to avoid during compression/writing). 6456 */ 6457 hdr->b_l2hdr.b_compress = ZIO_COMPRESS_OFF; 6458 hdr->b_l2hdr.b_asize = hdr->b_size; 6459 hdr->b_l1hdr.b_tmp_cdata = hdr->b_l1hdr.b_buf->b_data; 6460 6461 /* 6462 * Explicitly set the b_daddr field to a known 6463 * value which means "invalid address". This 6464 * enables us to differentiate which stage of 6465 * l2arc_write_buffers() the particular header 6466 * is in (e.g. this loop, or the one below). 6467 * ARC_FLAG_L2_WRITING is not enough to make 6468 * this distinction, and we need to know in 6469 * order to do proper l2arc vdev accounting in 6470 * arc_release() and arc_hdr_destroy(). 6471 * 6472 * Note, we can't use a new flag to distinguish 6473 * the two stages because we don't hold the 6474 * header's hash_lock below, in the second stage 6475 * of this function. Thus, we can't simply 6476 * change the b_flags field to denote that the 6477 * IO has been sent. We can change the b_daddr 6478 * field of the L2 portion, though, since we'll 6479 * be holding the l2ad_mtx; which is why we're 6480 * using it to denote the header's state change. 6481 */ 6482 hdr->b_l2hdr.b_daddr = L2ARC_ADDR_UNSET; 6483 6484 hdr->b_flags |= ARC_FLAG_HAS_L2HDR; 6485 6486 mutex_enter(&dev->l2ad_mtx); 6487 list_insert_head(&dev->l2ad_buflist, hdr); 6488 mutex_exit(&dev->l2ad_mtx); 6489 6490 /* 6491 * Compute and store the buffer cksum before 6492 * writing. On debug the cksum is verified first. 6493 */ 6494 arc_cksum_verify(hdr->b_l1hdr.b_buf); 6495 arc_cksum_compute(hdr->b_l1hdr.b_buf, B_TRUE); 6496 6497 mutex_exit(hash_lock); 6498 6499 write_sz += buf_sz; 6500 write_asize += buf_a_sz; 6501 } 6502 6503 multilist_sublist_unlock(mls); 6504 6505 if (full == B_TRUE) 6506 break; 6507 } 6508 6509 /* No buffers selected for writing? */ 6510 if (pio == NULL) { 6511 ASSERT0(write_sz); 6512 ASSERT(!HDR_HAS_L1HDR(head)); 6513 kmem_cache_free(hdr_l2only_cache, head); 6514 return (0); 6515 } 6516 6517 mutex_enter(&dev->l2ad_mtx); 6518 6519 /* 6520 * Now start writing the buffers. We're starting at the write head 6521 * and work backwards, retracing the course of the buffer selector 6522 * loop above. 6523 */ 6524 write_asize = 0; 6525 for (hdr = list_prev(&dev->l2ad_buflist, head); hdr; 6526 hdr = list_prev(&dev->l2ad_buflist, hdr)) { 6527 uint64_t buf_sz; 6528 boolean_t compress; 6529 6530 /* 6531 * We rely on the L1 portion of the header below, so 6532 * it's invalid for this header to have been evicted out 6533 * of the ghost cache, prior to being written out. The 6534 * ARC_FLAG_L2_WRITING bit ensures this won't happen. 6535 */ 6536 ASSERT(HDR_HAS_L1HDR(hdr)); 6537 6538 /* 6539 * We shouldn't need to lock the buffer here, since we flagged 6540 * it as ARC_FLAG_L2_WRITING in the previous step, but we must 6541 * take care to only access its L2 cache parameters. In 6542 * particular, hdr->l1hdr.b_buf may be invalid by now due to 6543 * ARC eviction. 6544 */ 6545 hdr->b_l2hdr.b_daddr = dev->l2ad_hand; 6546 6547 /* 6548 * Save a pointer to the original buffer data we had previously 6549 * stashed away. 6550 */ 6551 buf_data = hdr->b_l1hdr.b_tmp_cdata; 6552 6553 compress = HDR_L2COMPRESS(hdr) && 6554 hdr->b_l2hdr.b_asize >= buf_compress_minsz; 6555 if (l2arc_transform_buf(hdr, compress)) { 6556 /* 6557 * If compression succeeded, enable headroom 6558 * boost on the next scan cycle. 6559 */ 6560 *headroom_boost = B_TRUE; 6561 } 6562 6563 /* 6564 * Get the new buffer size that accounts for compression 6565 * and padding. 6566 */ 6567 buf_sz = hdr->b_l2hdr.b_asize; 6568 6569 /* 6570 * We need to do this regardless if buf_sz is zero or 6571 * not, otherwise, when this l2hdr is evicted we'll 6572 * remove a reference that was never added. 6573 */ 6574 (void) refcount_add_many(&dev->l2ad_alloc, buf_sz, hdr); 6575 6576 /* Compression may have squashed the buffer to zero length. */ 6577 if (buf_sz != 0) { 6578 /* 6579 * If the data was padded or compressed, then it 6580 * it is in a new buffer. 6581 */ 6582 if (hdr->b_l1hdr.b_tmp_cdata != NULL) 6583 buf_data = hdr->b_l1hdr.b_tmp_cdata; 6584 wzio = zio_write_phys(pio, dev->l2ad_vdev, 6585 dev->l2ad_hand, buf_sz, buf_data, ZIO_CHECKSUM_OFF, 6586 NULL, NULL, ZIO_PRIORITY_ASYNC_WRITE, 6587 ZIO_FLAG_CANFAIL, B_FALSE); 6588 6589 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev, 6590 zio_t *, wzio); 6591 (void) zio_nowait(wzio); 6592 6593 write_asize += buf_sz; 6594 dev->l2ad_hand += buf_sz; 6595 } 6596 } 6597 6598 mutex_exit(&dev->l2ad_mtx); 6599 6600 ASSERT3U(write_asize, <=, target_sz); 6601 ARCSTAT_BUMP(arcstat_l2_writes_sent); 6602 ARCSTAT_INCR(arcstat_l2_write_bytes, write_asize); 6603 ARCSTAT_INCR(arcstat_l2_size, write_sz); 6604 ARCSTAT_INCR(arcstat_l2_asize, write_asize); 6605 vdev_space_update(dev->l2ad_vdev, write_asize, 0, 0); 6606 6607 /* 6608 * Bump device hand to the device start if it is approaching the end. 6609 * l2arc_evict() will already have evicted ahead for this case. 6610 */ 6611 if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) { 6612 dev->l2ad_hand = dev->l2ad_start; 6613 dev->l2ad_first = B_FALSE; 6614 } 6615 6616 dev->l2ad_writing = B_TRUE; 6617 (void) zio_wait(pio); 6618 dev->l2ad_writing = B_FALSE; 6619 6620 return (write_asize); 6621} 6622 6623/* 6624 * Transforms, possibly compresses and pads, an L2ARC buffer. 6625 * The data to be compressed must be prefilled in l1hdr.b_tmp_cdata and its 6626 * size in l2hdr->b_asize. This routine tries to compress the data and 6627 * depending on the compression result there are three possible outcomes: 6628 * *) The buffer was incompressible. The buffer size was already ashift aligned. 6629 * The original hdr contents were left untouched except for b_tmp_cdata, 6630 * which is reset to NULL. The caller must keep a pointer to the original 6631 * data. 6632 * *) The buffer was incompressible. The buffer size was not ashift aligned. 6633 * b_tmp_cdata was replaced with a temporary data buffer which holds a padded 6634 * (aligned) copy of the data. Once writing is done, invoke 6635 * l2arc_release_cdata_buf on this hdr to free the temporary buffer. 6636 * *) The buffer was all-zeros, so there is no need to write it to an L2 6637 * device. To indicate this situation b_tmp_cdata is NULL'ed, b_asize is 6638 * set to zero and b_compress is set to ZIO_COMPRESS_EMPTY. 6639 * *) Compression succeeded and b_tmp_cdata was replaced with a temporary 6640 * data buffer which holds the compressed data to be written, and b_asize 6641 * tells us how much data there is. b_compress is set to the appropriate 6642 * compression algorithm. Once writing is done, invoke 6643 * l2arc_release_cdata_buf on this l2hdr to free this temporary buffer. 6644 * 6645 * Returns B_TRUE if compression succeeded, or B_FALSE if it didn't (the 6646 * buffer was incompressible). 6647 */ 6648static boolean_t 6649l2arc_transform_buf(arc_buf_hdr_t *hdr, boolean_t compress) 6650{ 6651 void *cdata; 6652 size_t align, asize, csize, len, rounded; 6653 6654 ASSERT(HDR_HAS_L2HDR(hdr)); 6655 l2arc_buf_hdr_t *l2hdr = &hdr->b_l2hdr; 6656 6657 ASSERT(HDR_HAS_L1HDR(hdr)); 6658 ASSERT3S(l2hdr->b_compress, ==, ZIO_COMPRESS_OFF); 6659 ASSERT(hdr->b_l1hdr.b_tmp_cdata != NULL); 6660 6661 len = l2hdr->b_asize; 6662 align = (size_t)1 << l2hdr->b_dev->l2ad_vdev->vdev_ashift; 6663 asize = P2ROUNDUP(len, align); 6664 cdata = zio_data_buf_alloc(asize); 6665 ASSERT3P(cdata, !=, NULL); 6666 if (compress) 6667 csize = zio_compress_data(ZIO_COMPRESS_LZ4, 6668 hdr->b_l1hdr.b_tmp_cdata, cdata, len); 6669 else 6670 csize = len; 6671 6672 if (csize == 0) { 6673 /* zero block, indicate that there's nothing to write */ 6674 zio_data_buf_free(cdata, asize); 6675 l2hdr->b_compress = ZIO_COMPRESS_EMPTY; 6676 l2hdr->b_asize = 0; 6677 hdr->b_l1hdr.b_tmp_cdata = NULL; 6678 ARCSTAT_BUMP(arcstat_l2_compress_zeros); 6679 return (B_TRUE); 6680 } 6681 6682 rounded = P2ROUNDUP(csize, align); 6683 ASSERT3U(rounded, <=, asize); 6684 if (rounded < len) { 6685 /* 6686 * Compression succeeded, we'll keep the cdata around for 6687 * writing and release it afterwards. 6688 */ 6689 if (rounded > csize) { 6690 bzero((char *)cdata + csize, rounded - csize); 6691 csize = rounded; 6692 } 6693 l2hdr->b_compress = ZIO_COMPRESS_LZ4; 6694 l2hdr->b_asize = csize; 6695 hdr->b_l1hdr.b_tmp_cdata = cdata; 6696 ARCSTAT_BUMP(arcstat_l2_compress_successes); 6697 return (B_TRUE); 6698 } else { 6699 /* 6700 * Compression did not save space. 6701 */ 6702 if (P2PHASE(len, align) != 0) { 6703 /* 6704 * Use compression buffer for a copy of data padded to 6705 * the proper size. Compression algorithm remains set 6706 * to ZIO_COMPRESS_OFF. 6707 */ 6708 ASSERT3U(len, <, asize); 6709 bcopy(hdr->b_l1hdr.b_tmp_cdata, cdata, len); 6710 bzero((char *)cdata + len, asize - len); 6711 l2hdr->b_asize = asize; 6712 hdr->b_l1hdr.b_tmp_cdata = cdata; 6713 ARCSTAT_BUMP(arcstat_l2_padding_needed); 6714 } else { 6715 ASSERT3U(len, ==, asize); 6716 /* 6717 * The original buffer is good as is, 6718 * release the compressed buffer. 6719 * l2hdr will be left unmodified except for b_tmp_cdata. 6720 */ 6721 zio_data_buf_free(cdata, asize); 6722 hdr->b_l1hdr.b_tmp_cdata = NULL; 6723 } 6724 if (compress) 6725 ARCSTAT_BUMP(arcstat_l2_compress_failures); 6726 return (B_FALSE); 6727 } 6728} 6729 6730/* 6731 * Decompresses a zio read back from an l2arc device. On success, the 6732 * underlying zio's io_data buffer is overwritten by the uncompressed 6733 * version. On decompression error (corrupt compressed stream), the 6734 * zio->io_error value is set to signal an I/O error. 6735 * 6736 * Please note that the compressed data stream is not checksummed, so 6737 * if the underlying device is experiencing data corruption, we may feed 6738 * corrupt data to the decompressor, so the decompressor needs to be 6739 * able to handle this situation (LZ4 does). 6740 */ 6741static void 6742l2arc_decompress_zio(zio_t *zio, arc_buf_hdr_t *hdr, enum zio_compress c) 6743{ 6744 ASSERT(L2ARC_IS_VALID_COMPRESS(c)); 6745 6746 if (zio->io_error != 0) { 6747 /* 6748 * An io error has occured, just restore the original io 6749 * size in preparation for a main pool read. 6750 */ 6751 zio->io_orig_size = zio->io_size = hdr->b_size; 6752 return; 6753 } 6754 6755 if (c == ZIO_COMPRESS_EMPTY) { 6756 /* 6757 * An empty buffer results in a null zio, which means we 6758 * need to fill its io_data after we're done restoring the 6759 * buffer's contents. 6760 */ 6761 ASSERT(hdr->b_l1hdr.b_buf != NULL); 6762 bzero(hdr->b_l1hdr.b_buf->b_data, hdr->b_size); 6763 zio->io_data = zio->io_orig_data = hdr->b_l1hdr.b_buf->b_data; 6764 } else { 6765 ASSERT(zio->io_data != NULL); 6766 /* 6767 * We copy the compressed data from the start of the arc buffer 6768 * (the zio_read will have pulled in only what we need, the 6769 * rest is garbage which we will overwrite at decompression) 6770 * and then decompress back to the ARC data buffer. This way we 6771 * can minimize copying by simply decompressing back over the 6772 * original compressed data (rather than decompressing to an 6773 * aux buffer and then copying back the uncompressed buffer, 6774 * which is likely to be much larger). 6775 */ 6776 uint64_t csize; 6777 void *cdata; 6778 6779 csize = zio->io_size; 6780 cdata = zio_data_buf_alloc(csize); 6781 bcopy(zio->io_data, cdata, csize); 6782 if (zio_decompress_data(c, cdata, zio->io_data, csize, 6783 hdr->b_size) != 0) 6784 zio->io_error = EIO; 6785 zio_data_buf_free(cdata, csize); 6786 } 6787 6788 /* Restore the expected uncompressed IO size. */ 6789 zio->io_orig_size = zio->io_size = hdr->b_size; 6790} 6791 6792/* 6793 * Releases the temporary b_tmp_cdata buffer in an l2arc header structure. 6794 * This buffer serves as a temporary holder of compressed or padded data while 6795 * the buffer entry is being written to an l2arc device. Once that is 6796 * done, we can dispose of it. 6797 */ 6798static void 6799l2arc_release_cdata_buf(arc_buf_hdr_t *hdr) 6800{ 6801 size_t align, asize, len; 6802 enum zio_compress comp = hdr->b_l2hdr.b_compress; 6803 6804 ASSERT(HDR_HAS_L2HDR(hdr)); 6805 ASSERT(HDR_HAS_L1HDR(hdr)); 6806 ASSERT(comp == ZIO_COMPRESS_OFF || L2ARC_IS_VALID_COMPRESS(comp)); 6807 6808 if (hdr->b_l1hdr.b_tmp_cdata != NULL) { 6809 ASSERT(comp != ZIO_COMPRESS_EMPTY); 6810 len = hdr->b_size; 6811 align = (size_t)1 << hdr->b_l2hdr.b_dev->l2ad_vdev->vdev_ashift; 6812 asize = P2ROUNDUP(len, align); 6813 zio_data_buf_free(hdr->b_l1hdr.b_tmp_cdata, asize); 6814 hdr->b_l1hdr.b_tmp_cdata = NULL; 6815 } else { 6816 ASSERT(comp == ZIO_COMPRESS_OFF || comp == ZIO_COMPRESS_EMPTY); 6817 } 6818} 6819 6820/* 6821 * This thread feeds the L2ARC at regular intervals. This is the beating 6822 * heart of the L2ARC. 6823 */ 6824static void 6825l2arc_feed_thread(void *dummy __unused) 6826{ 6827 callb_cpr_t cpr; 6828 l2arc_dev_t *dev; 6829 spa_t *spa; 6830 uint64_t size, wrote; 6831 clock_t begin, next = ddi_get_lbolt(); 6832 boolean_t headroom_boost = B_FALSE; 6833 6834 CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG); 6835 6836 mutex_enter(&l2arc_feed_thr_lock); 6837 6838 while (l2arc_thread_exit == 0) { 6839 CALLB_CPR_SAFE_BEGIN(&cpr); 6840 (void) cv_timedwait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock, 6841 next - ddi_get_lbolt()); 6842 CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock); 6843 next = ddi_get_lbolt() + hz; 6844 6845 /* 6846 * Quick check for L2ARC devices. 6847 */ 6848 mutex_enter(&l2arc_dev_mtx); 6849 if (l2arc_ndev == 0) { 6850 mutex_exit(&l2arc_dev_mtx); 6851 continue; 6852 } 6853 mutex_exit(&l2arc_dev_mtx); 6854 begin = ddi_get_lbolt(); 6855 6856 /* 6857 * This selects the next l2arc device to write to, and in 6858 * doing so the next spa to feed from: dev->l2ad_spa. This 6859 * will return NULL if there are now no l2arc devices or if 6860 * they are all faulted. 6861 * 6862 * If a device is returned, its spa's config lock is also 6863 * held to prevent device removal. l2arc_dev_get_next() 6864 * will grab and release l2arc_dev_mtx. 6865 */ 6866 if ((dev = l2arc_dev_get_next()) == NULL) 6867 continue; 6868 6869 spa = dev->l2ad_spa; 6870 ASSERT(spa != NULL); 6871 6872 /* 6873 * If the pool is read-only then force the feed thread to 6874 * sleep a little longer. 6875 */ 6876 if (!spa_writeable(spa)) { 6877 next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz; 6878 spa_config_exit(spa, SCL_L2ARC, dev); 6879 continue; 6880 } 6881 6882 /* 6883 * Avoid contributing to memory pressure. 6884 */ 6885 if (arc_reclaim_needed()) { 6886 ARCSTAT_BUMP(arcstat_l2_abort_lowmem); 6887 spa_config_exit(spa, SCL_L2ARC, dev); 6888 continue; 6889 } 6890 6891 ARCSTAT_BUMP(arcstat_l2_feeds); 6892 6893 size = l2arc_write_size(); 6894 6895 /* 6896 * Evict L2ARC buffers that will be overwritten. 6897 */ 6898 l2arc_evict(dev, size, B_FALSE); 6899 6900 /* 6901 * Write ARC buffers. 6902 */ 6903 wrote = l2arc_write_buffers(spa, dev, size, &headroom_boost); 6904 6905 /* 6906 * Calculate interval between writes. 6907 */ 6908 next = l2arc_write_interval(begin, size, wrote); 6909 spa_config_exit(spa, SCL_L2ARC, dev); 6910 } 6911 6912 l2arc_thread_exit = 0; 6913 cv_broadcast(&l2arc_feed_thr_cv); 6914 CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */ 6915 thread_exit(); 6916} 6917 6918boolean_t 6919l2arc_vdev_present(vdev_t *vd) 6920{ 6921 l2arc_dev_t *dev; 6922 6923 mutex_enter(&l2arc_dev_mtx); 6924 for (dev = list_head(l2arc_dev_list); dev != NULL; 6925 dev = list_next(l2arc_dev_list, dev)) { 6926 if (dev->l2ad_vdev == vd) 6927 break; 6928 } 6929 mutex_exit(&l2arc_dev_mtx); 6930 6931 return (dev != NULL); 6932} 6933 6934/* 6935 * Add a vdev for use by the L2ARC. By this point the spa has already 6936 * validated the vdev and opened it. 6937 */ 6938void 6939l2arc_add_vdev(spa_t *spa, vdev_t *vd) 6940{ 6941 l2arc_dev_t *adddev; 6942 6943 ASSERT(!l2arc_vdev_present(vd)); 6944 6945 vdev_ashift_optimize(vd); 6946 6947 /* 6948 * Create a new l2arc device entry. 6949 */ 6950 adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP); 6951 adddev->l2ad_spa = spa; 6952 adddev->l2ad_vdev = vd; 6953 adddev->l2ad_start = VDEV_LABEL_START_SIZE; 6954 adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd); 6955 adddev->l2ad_hand = adddev->l2ad_start; 6956 adddev->l2ad_first = B_TRUE; 6957 adddev->l2ad_writing = B_FALSE; 6958 6959 mutex_init(&adddev->l2ad_mtx, NULL, MUTEX_DEFAULT, NULL); 6960 /* 6961 * This is a list of all ARC buffers that are still valid on the 6962 * device. 6963 */ 6964 list_create(&adddev->l2ad_buflist, sizeof (arc_buf_hdr_t), 6965 offsetof(arc_buf_hdr_t, b_l2hdr.b_l2node)); 6966 6967 vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand); 6968 refcount_create(&adddev->l2ad_alloc); 6969 6970 /* 6971 * Add device to global list 6972 */ 6973 mutex_enter(&l2arc_dev_mtx); 6974 list_insert_head(l2arc_dev_list, adddev); 6975 atomic_inc_64(&l2arc_ndev); 6976 mutex_exit(&l2arc_dev_mtx); 6977} 6978 6979/* 6980 * Remove a vdev from the L2ARC. 6981 */ 6982void 6983l2arc_remove_vdev(vdev_t *vd) 6984{ 6985 l2arc_dev_t *dev, *nextdev, *remdev = NULL; 6986 6987 /* 6988 * Find the device by vdev 6989 */ 6990 mutex_enter(&l2arc_dev_mtx); 6991 for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) { 6992 nextdev = list_next(l2arc_dev_list, dev); 6993 if (vd == dev->l2ad_vdev) { 6994 remdev = dev; 6995 break; 6996 } 6997 } 6998 ASSERT(remdev != NULL); 6999 7000 /* 7001 * Remove device from global list 7002 */ 7003 list_remove(l2arc_dev_list, remdev); 7004 l2arc_dev_last = NULL; /* may have been invalidated */ 7005 atomic_dec_64(&l2arc_ndev); 7006 mutex_exit(&l2arc_dev_mtx); 7007 7008 /* 7009 * Clear all buflists and ARC references. L2ARC device flush. 7010 */ 7011 l2arc_evict(remdev, 0, B_TRUE); 7012 list_destroy(&remdev->l2ad_buflist); 7013 mutex_destroy(&remdev->l2ad_mtx); 7014 refcount_destroy(&remdev->l2ad_alloc); 7015 kmem_free(remdev, sizeof (l2arc_dev_t)); 7016} 7017 7018void 7019l2arc_init(void) 7020{ 7021 l2arc_thread_exit = 0; 7022 l2arc_ndev = 0; 7023 l2arc_writes_sent = 0; 7024 l2arc_writes_done = 0; 7025 7026 mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL); 7027 cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL); 7028 mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL); 7029 mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL); 7030 7031 l2arc_dev_list = &L2ARC_dev_list; 7032 l2arc_free_on_write = &L2ARC_free_on_write; 7033 list_create(l2arc_dev_list, sizeof (l2arc_dev_t), 7034 offsetof(l2arc_dev_t, l2ad_node)); 7035 list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t), 7036 offsetof(l2arc_data_free_t, l2df_list_node)); 7037} 7038 7039void 7040l2arc_fini(void) 7041{ 7042 /* 7043 * This is called from dmu_fini(), which is called from spa_fini(); 7044 * Because of this, we can assume that all l2arc devices have 7045 * already been removed when the pools themselves were removed. 7046 */ 7047 7048 l2arc_do_free_on_write(); 7049 7050 mutex_destroy(&l2arc_feed_thr_lock); 7051 cv_destroy(&l2arc_feed_thr_cv); 7052 mutex_destroy(&l2arc_dev_mtx); 7053 mutex_destroy(&l2arc_free_on_write_mtx); 7054 7055 list_destroy(l2arc_dev_list); 7056 list_destroy(l2arc_free_on_write); 7057} 7058 7059void 7060l2arc_start(void) 7061{ 7062 if (!(spa_mode_global & FWRITE)) 7063 return; 7064 7065 (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0, 7066 TS_RUN, minclsyspri); 7067} 7068 7069void 7070l2arc_stop(void) 7071{ 7072 if (!(spa_mode_global & FWRITE)) 7073 return; 7074 7075 mutex_enter(&l2arc_feed_thr_lock); 7076 cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */ 7077 l2arc_thread_exit = 1; 7078 while (l2arc_thread_exit != 0) 7079 cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock); 7080 mutex_exit(&l2arc_feed_thr_lock); 7081} 7082