arc.c revision 290752
1/* 2 * CDDL HEADER START 3 * 4 * The contents of this file are subject to the terms of the 5 * Common Development and Distribution License (the "License"). 6 * You may not use this file except in compliance with the License. 7 * 8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE 9 * or http://www.opensolaris.org/os/licensing. 10 * See the License for the specific language governing permissions 11 * and limitations under the License. 12 * 13 * When distributing Covered Code, include this CDDL HEADER in each 14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE. 15 * If applicable, add the following below this CDDL HEADER, with the 16 * fields enclosed by brackets "[]" replaced with your own identifying 17 * information: Portions Copyright [yyyy] [name of copyright owner] 18 * 19 * CDDL HEADER END 20 */ 21/* 22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved. 23 * Copyright (c) 2012, Joyent, Inc. All rights reserved. 24 * Copyright (c) 2011, 2015 by Delphix. All rights reserved. 25 * Copyright (c) 2014 by Saso Kiselkov. All rights reserved. 26 * Copyright 2015 Nexenta Systems, Inc. All rights reserved. 27 */ 28 29/* 30 * DVA-based Adjustable Replacement Cache 31 * 32 * While much of the theory of operation used here is 33 * based on the self-tuning, low overhead replacement cache 34 * presented by Megiddo and Modha at FAST 2003, there are some 35 * significant differences: 36 * 37 * 1. The Megiddo and Modha model assumes any page is evictable. 38 * Pages in its cache cannot be "locked" into memory. This makes 39 * the eviction algorithm simple: evict the last page in the list. 40 * This also make the performance characteristics easy to reason 41 * about. Our cache is not so simple. At any given moment, some 42 * subset of the blocks in the cache are un-evictable because we 43 * have handed out a reference to them. Blocks are only evictable 44 * when there are no external references active. This makes 45 * eviction far more problematic: we choose to evict the evictable 46 * blocks that are the "lowest" in the list. 47 * 48 * There are times when it is not possible to evict the requested 49 * space. In these circumstances we are unable to adjust the cache 50 * size. To prevent the cache growing unbounded at these times we 51 * implement a "cache throttle" that slows the flow of new data 52 * into the cache until we can make space available. 53 * 54 * 2. The Megiddo and Modha model assumes a fixed cache size. 55 * Pages are evicted when the cache is full and there is a cache 56 * miss. Our model has a variable sized cache. It grows with 57 * high use, but also tries to react to memory pressure from the 58 * operating system: decreasing its size when system memory is 59 * tight. 60 * 61 * 3. The Megiddo and Modha model assumes a fixed page size. All 62 * elements of the cache are therefore exactly the same size. So 63 * when adjusting the cache size following a cache miss, its simply 64 * a matter of choosing a single page to evict. In our model, we 65 * have variable sized cache blocks (rangeing from 512 bytes to 66 * 128K bytes). We therefore choose a set of blocks to evict to make 67 * space for a cache miss that approximates as closely as possible 68 * the space used by the new block. 69 * 70 * See also: "ARC: A Self-Tuning, Low Overhead Replacement Cache" 71 * by N. Megiddo & D. Modha, FAST 2003 72 */ 73 74/* 75 * The locking model: 76 * 77 * A new reference to a cache buffer can be obtained in two 78 * ways: 1) via a hash table lookup using the DVA as a key, 79 * or 2) via one of the ARC lists. The arc_read() interface 80 * uses method 1, while the internal arc algorithms for 81 * adjusting the cache use method 2. We therefore provide two 82 * types of locks: 1) the hash table lock array, and 2) the 83 * arc list locks. 84 * 85 * Buffers do not have their own mutexes, rather they rely on the 86 * hash table mutexes for the bulk of their protection (i.e. most 87 * fields in the arc_buf_hdr_t are protected by these mutexes). 88 * 89 * buf_hash_find() returns the appropriate mutex (held) when it 90 * locates the requested buffer in the hash table. It returns 91 * NULL for the mutex if the buffer was not in the table. 92 * 93 * buf_hash_remove() expects the appropriate hash mutex to be 94 * already held before it is invoked. 95 * 96 * Each arc state also has a mutex which is used to protect the 97 * buffer list associated with the state. When attempting to 98 * obtain a hash table lock while holding an arc list lock you 99 * must use: mutex_tryenter() to avoid deadlock. Also note that 100 * the active state mutex must be held before the ghost state mutex. 101 * 102 * Arc buffers may have an associated eviction callback function. 103 * This function will be invoked prior to removing the buffer (e.g. 104 * in arc_do_user_evicts()). Note however that the data associated 105 * with the buffer may be evicted prior to the callback. The callback 106 * must be made with *no locks held* (to prevent deadlock). Additionally, 107 * the users of callbacks must ensure that their private data is 108 * protected from simultaneous callbacks from arc_clear_callback() 109 * and arc_do_user_evicts(). 110 * 111 * Note that the majority of the performance stats are manipulated 112 * with atomic operations. 113 * 114 * The L2ARC uses the l2ad_mtx on each vdev for the following: 115 * 116 * - L2ARC buflist creation 117 * - L2ARC buflist eviction 118 * - L2ARC write completion, which walks L2ARC buflists 119 * - ARC header destruction, as it removes from L2ARC buflists 120 * - ARC header release, as it removes from L2ARC buflists 121 */ 122 123#include <sys/spa.h> 124#include <sys/zio.h> 125#include <sys/zio_compress.h> 126#include <sys/zfs_context.h> 127#include <sys/arc.h> 128#include <sys/refcount.h> 129#include <sys/vdev.h> 130#include <sys/vdev_impl.h> 131#include <sys/dsl_pool.h> 132#include <sys/multilist.h> 133#ifdef _KERNEL 134#include <sys/dnlc.h> 135#endif 136#include <sys/callb.h> 137#include <sys/kstat.h> 138#include <sys/trim_map.h> 139#include <zfs_fletcher.h> 140#include <sys/sdt.h> 141 142#include <vm/vm_pageout.h> 143#include <machine/vmparam.h> 144 145#ifdef illumos 146#ifndef _KERNEL 147/* set with ZFS_DEBUG=watch, to enable watchpoints on frozen buffers */ 148boolean_t arc_watch = B_FALSE; 149int arc_procfd; 150#endif 151#endif /* illumos */ 152 153static kmutex_t arc_reclaim_lock; 154static kcondvar_t arc_reclaim_thread_cv; 155static boolean_t arc_reclaim_thread_exit; 156static kcondvar_t arc_reclaim_waiters_cv; 157 158static kmutex_t arc_user_evicts_lock; 159static kcondvar_t arc_user_evicts_cv; 160static boolean_t arc_user_evicts_thread_exit; 161 162uint_t arc_reduce_dnlc_percent = 3; 163 164/* 165 * The number of headers to evict in arc_evict_state_impl() before 166 * dropping the sublist lock and evicting from another sublist. A lower 167 * value means we're more likely to evict the "correct" header (i.e. the 168 * oldest header in the arc state), but comes with higher overhead 169 * (i.e. more invocations of arc_evict_state_impl()). 170 */ 171int zfs_arc_evict_batch_limit = 10; 172 173/* 174 * The number of sublists used for each of the arc state lists. If this 175 * is not set to a suitable value by the user, it will be configured to 176 * the number of CPUs on the system in arc_init(). 177 */ 178int zfs_arc_num_sublists_per_state = 0; 179 180/* number of seconds before growing cache again */ 181static int arc_grow_retry = 60; 182 183/* shift of arc_c for calculating overflow limit in arc_get_data_buf */ 184int zfs_arc_overflow_shift = 8; 185 186/* shift of arc_c for calculating both min and max arc_p */ 187static int arc_p_min_shift = 4; 188 189/* log2(fraction of arc to reclaim) */ 190static int arc_shrink_shift = 7; 191 192/* 193 * log2(fraction of ARC which must be free to allow growing). 194 * I.e. If there is less than arc_c >> arc_no_grow_shift free memory, 195 * when reading a new block into the ARC, we will evict an equal-sized block 196 * from the ARC. 197 * 198 * This must be less than arc_shrink_shift, so that when we shrink the ARC, 199 * we will still not allow it to grow. 200 */ 201int arc_no_grow_shift = 5; 202 203 204/* 205 * minimum lifespan of a prefetch block in clock ticks 206 * (initialized in arc_init()) 207 */ 208static int arc_min_prefetch_lifespan; 209 210/* 211 * If this percent of memory is free, don't throttle. 212 */ 213int arc_lotsfree_percent = 10; 214 215static int arc_dead; 216extern boolean_t zfs_prefetch_disable; 217 218/* 219 * The arc has filled available memory and has now warmed up. 220 */ 221static boolean_t arc_warm; 222 223/* 224 * These tunables are for performance analysis. 225 */ 226uint64_t zfs_arc_max; 227uint64_t zfs_arc_min; 228uint64_t zfs_arc_meta_limit = 0; 229uint64_t zfs_arc_meta_min = 0; 230int zfs_arc_grow_retry = 0; 231int zfs_arc_shrink_shift = 0; 232int zfs_arc_p_min_shift = 0; 233int zfs_disable_dup_eviction = 0; 234uint64_t zfs_arc_average_blocksize = 8 * 1024; /* 8KB */ 235u_int zfs_arc_free_target = 0; 236 237static int sysctl_vfs_zfs_arc_free_target(SYSCTL_HANDLER_ARGS); 238static int sysctl_vfs_zfs_arc_meta_limit(SYSCTL_HANDLER_ARGS); 239 240#ifdef _KERNEL 241static void 242arc_free_target_init(void *unused __unused) 243{ 244 245 zfs_arc_free_target = vm_pageout_wakeup_thresh; 246} 247SYSINIT(arc_free_target_init, SI_SUB_KTHREAD_PAGE, SI_ORDER_ANY, 248 arc_free_target_init, NULL); 249 250TUNABLE_QUAD("vfs.zfs.arc_max", &zfs_arc_max); 251TUNABLE_QUAD("vfs.zfs.arc_min", &zfs_arc_min); 252TUNABLE_QUAD("vfs.zfs.arc_meta_limit", &zfs_arc_meta_limit); 253TUNABLE_QUAD("vfs.zfs.arc_meta_min", &zfs_arc_meta_min); 254TUNABLE_QUAD("vfs.zfs.arc_average_blocksize", &zfs_arc_average_blocksize); 255TUNABLE_INT("vfs.zfs.arc_shrink_shift", &zfs_arc_shrink_shift); 256SYSCTL_DECL(_vfs_zfs); 257SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_max, CTLFLAG_RDTUN, &zfs_arc_max, 0, 258 "Maximum ARC size"); 259SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_min, CTLFLAG_RDTUN, &zfs_arc_min, 0, 260 "Minimum ARC size"); 261SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_average_blocksize, CTLFLAG_RDTUN, 262 &zfs_arc_average_blocksize, 0, 263 "ARC average blocksize"); 264SYSCTL_INT(_vfs_zfs, OID_AUTO, arc_shrink_shift, CTLFLAG_RW, 265 &arc_shrink_shift, 0, 266 "log2(fraction of arc to reclaim)"); 267 268/* 269 * We don't have a tunable for arc_free_target due to the dependency on 270 * pagedaemon initialisation. 271 */ 272SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_free_target, 273 CTLTYPE_UINT | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, sizeof(u_int), 274 sysctl_vfs_zfs_arc_free_target, "IU", 275 "Desired number of free pages below which ARC triggers reclaim"); 276 277static int 278sysctl_vfs_zfs_arc_free_target(SYSCTL_HANDLER_ARGS) 279{ 280 u_int val; 281 int err; 282 283 val = zfs_arc_free_target; 284 err = sysctl_handle_int(oidp, &val, 0, req); 285 if (err != 0 || req->newptr == NULL) 286 return (err); 287 288 if (val < minfree) 289 return (EINVAL); 290 if (val > cnt.v_page_count) 291 return (EINVAL); 292 293 zfs_arc_free_target = val; 294 295 return (0); 296} 297 298/* 299 * Must be declared here, before the definition of corresponding kstat 300 * macro which uses the same names will confuse the compiler. 301 */ 302SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_meta_limit, 303 CTLTYPE_U64 | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, sizeof(uint64_t), 304 sysctl_vfs_zfs_arc_meta_limit, "QU", 305 "ARC metadata limit"); 306#endif 307 308/* 309 * Note that buffers can be in one of 6 states: 310 * ARC_anon - anonymous (discussed below) 311 * ARC_mru - recently used, currently cached 312 * ARC_mru_ghost - recentely used, no longer in cache 313 * ARC_mfu - frequently used, currently cached 314 * ARC_mfu_ghost - frequently used, no longer in cache 315 * ARC_l2c_only - exists in L2ARC but not other states 316 * When there are no active references to the buffer, they are 317 * are linked onto a list in one of these arc states. These are 318 * the only buffers that can be evicted or deleted. Within each 319 * state there are multiple lists, one for meta-data and one for 320 * non-meta-data. Meta-data (indirect blocks, blocks of dnodes, 321 * etc.) is tracked separately so that it can be managed more 322 * explicitly: favored over data, limited explicitly. 323 * 324 * Anonymous buffers are buffers that are not associated with 325 * a DVA. These are buffers that hold dirty block copies 326 * before they are written to stable storage. By definition, 327 * they are "ref'd" and are considered part of arc_mru 328 * that cannot be freed. Generally, they will aquire a DVA 329 * as they are written and migrate onto the arc_mru list. 330 * 331 * The ARC_l2c_only state is for buffers that are in the second 332 * level ARC but no longer in any of the ARC_m* lists. The second 333 * level ARC itself may also contain buffers that are in any of 334 * the ARC_m* states - meaning that a buffer can exist in two 335 * places. The reason for the ARC_l2c_only state is to keep the 336 * buffer header in the hash table, so that reads that hit the 337 * second level ARC benefit from these fast lookups. 338 */ 339 340typedef struct arc_state { 341 /* 342 * list of evictable buffers 343 */ 344 multilist_t arcs_list[ARC_BUFC_NUMTYPES]; 345 /* 346 * total amount of evictable data in this state 347 */ 348 uint64_t arcs_lsize[ARC_BUFC_NUMTYPES]; 349 /* 350 * total amount of data in this state; this includes: evictable, 351 * non-evictable, ARC_BUFC_DATA, and ARC_BUFC_METADATA. 352 */ 353 refcount_t arcs_size; 354} arc_state_t; 355 356/* The 6 states: */ 357static arc_state_t ARC_anon; 358static arc_state_t ARC_mru; 359static arc_state_t ARC_mru_ghost; 360static arc_state_t ARC_mfu; 361static arc_state_t ARC_mfu_ghost; 362static arc_state_t ARC_l2c_only; 363 364typedef struct arc_stats { 365 kstat_named_t arcstat_hits; 366 kstat_named_t arcstat_misses; 367 kstat_named_t arcstat_demand_data_hits; 368 kstat_named_t arcstat_demand_data_misses; 369 kstat_named_t arcstat_demand_metadata_hits; 370 kstat_named_t arcstat_demand_metadata_misses; 371 kstat_named_t arcstat_prefetch_data_hits; 372 kstat_named_t arcstat_prefetch_data_misses; 373 kstat_named_t arcstat_prefetch_metadata_hits; 374 kstat_named_t arcstat_prefetch_metadata_misses; 375 kstat_named_t arcstat_mru_hits; 376 kstat_named_t arcstat_mru_ghost_hits; 377 kstat_named_t arcstat_mfu_hits; 378 kstat_named_t arcstat_mfu_ghost_hits; 379 kstat_named_t arcstat_allocated; 380 kstat_named_t arcstat_deleted; 381 /* 382 * Number of buffers that could not be evicted because the hash lock 383 * was held by another thread. The lock may not necessarily be held 384 * by something using the same buffer, since hash locks are shared 385 * by multiple buffers. 386 */ 387 kstat_named_t arcstat_mutex_miss; 388 /* 389 * Number of buffers skipped because they have I/O in progress, are 390 * indrect prefetch buffers that have not lived long enough, or are 391 * not from the spa we're trying to evict from. 392 */ 393 kstat_named_t arcstat_evict_skip; 394 /* 395 * Number of times arc_evict_state() was unable to evict enough 396 * buffers to reach it's target amount. 397 */ 398 kstat_named_t arcstat_evict_not_enough; 399 kstat_named_t arcstat_evict_l2_cached; 400 kstat_named_t arcstat_evict_l2_eligible; 401 kstat_named_t arcstat_evict_l2_ineligible; 402 kstat_named_t arcstat_evict_l2_skip; 403 kstat_named_t arcstat_hash_elements; 404 kstat_named_t arcstat_hash_elements_max; 405 kstat_named_t arcstat_hash_collisions; 406 kstat_named_t arcstat_hash_chains; 407 kstat_named_t arcstat_hash_chain_max; 408 kstat_named_t arcstat_p; 409 kstat_named_t arcstat_c; 410 kstat_named_t arcstat_c_min; 411 kstat_named_t arcstat_c_max; 412 kstat_named_t arcstat_size; 413 /* 414 * Number of bytes consumed by internal ARC structures necessary 415 * for tracking purposes; these structures are not actually 416 * backed by ARC buffers. This includes arc_buf_hdr_t structures 417 * (allocated via arc_buf_hdr_t_full and arc_buf_hdr_t_l2only 418 * caches), and arc_buf_t structures (allocated via arc_buf_t 419 * cache). 420 */ 421 kstat_named_t arcstat_hdr_size; 422 /* 423 * Number of bytes consumed by ARC buffers of type equal to 424 * ARC_BUFC_DATA. This is generally consumed by buffers backing 425 * on disk user data (e.g. plain file contents). 426 */ 427 kstat_named_t arcstat_data_size; 428 /* 429 * Number of bytes consumed by ARC buffers of type equal to 430 * ARC_BUFC_METADATA. This is generally consumed by buffers 431 * backing on disk data that is used for internal ZFS 432 * structures (e.g. ZAP, dnode, indirect blocks, etc). 433 */ 434 kstat_named_t arcstat_metadata_size; 435 /* 436 * Number of bytes consumed by various buffers and structures 437 * not actually backed with ARC buffers. This includes bonus 438 * buffers (allocated directly via zio_buf_* functions), 439 * dmu_buf_impl_t structures (allocated via dmu_buf_impl_t 440 * cache), and dnode_t structures (allocated via dnode_t cache). 441 */ 442 kstat_named_t arcstat_other_size; 443 /* 444 * Total number of bytes consumed by ARC buffers residing in the 445 * arc_anon state. This includes *all* buffers in the arc_anon 446 * state; e.g. data, metadata, evictable, and unevictable buffers 447 * are all included in this value. 448 */ 449 kstat_named_t arcstat_anon_size; 450 /* 451 * Number of bytes consumed by ARC buffers that meet the 452 * following criteria: backing buffers of type ARC_BUFC_DATA, 453 * residing in the arc_anon state, and are eligible for eviction 454 * (e.g. have no outstanding holds on the buffer). 455 */ 456 kstat_named_t arcstat_anon_evictable_data; 457 /* 458 * Number of bytes consumed by ARC buffers that meet the 459 * following criteria: backing buffers of type ARC_BUFC_METADATA, 460 * residing in the arc_anon state, and are eligible for eviction 461 * (e.g. have no outstanding holds on the buffer). 462 */ 463 kstat_named_t arcstat_anon_evictable_metadata; 464 /* 465 * Total number of bytes consumed by ARC buffers residing in the 466 * arc_mru state. This includes *all* buffers in the arc_mru 467 * state; e.g. data, metadata, evictable, and unevictable buffers 468 * are all included in this value. 469 */ 470 kstat_named_t arcstat_mru_size; 471 /* 472 * Number of bytes consumed by ARC buffers that meet the 473 * following criteria: backing buffers of type ARC_BUFC_DATA, 474 * residing in the arc_mru state, and are eligible for eviction 475 * (e.g. have no outstanding holds on the buffer). 476 */ 477 kstat_named_t arcstat_mru_evictable_data; 478 /* 479 * Number of bytes consumed by ARC buffers that meet the 480 * following criteria: backing buffers of type ARC_BUFC_METADATA, 481 * residing in the arc_mru state, and are eligible for eviction 482 * (e.g. have no outstanding holds on the buffer). 483 */ 484 kstat_named_t arcstat_mru_evictable_metadata; 485 /* 486 * Total number of bytes that *would have been* consumed by ARC 487 * buffers in the arc_mru_ghost state. The key thing to note 488 * here, is the fact that this size doesn't actually indicate 489 * RAM consumption. The ghost lists only consist of headers and 490 * don't actually have ARC buffers linked off of these headers. 491 * Thus, *if* the headers had associated ARC buffers, these 492 * buffers *would have* consumed this number of bytes. 493 */ 494 kstat_named_t arcstat_mru_ghost_size; 495 /* 496 * Number of bytes that *would have been* consumed by ARC 497 * buffers that are eligible for eviction, of type 498 * ARC_BUFC_DATA, and linked off the arc_mru_ghost state. 499 */ 500 kstat_named_t arcstat_mru_ghost_evictable_data; 501 /* 502 * Number of bytes that *would have been* consumed by ARC 503 * buffers that are eligible for eviction, of type 504 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state. 505 */ 506 kstat_named_t arcstat_mru_ghost_evictable_metadata; 507 /* 508 * Total number of bytes consumed by ARC buffers residing in the 509 * arc_mfu state. This includes *all* buffers in the arc_mfu 510 * state; e.g. data, metadata, evictable, and unevictable buffers 511 * are all included in this value. 512 */ 513 kstat_named_t arcstat_mfu_size; 514 /* 515 * Number of bytes consumed by ARC buffers that are eligible for 516 * eviction, of type ARC_BUFC_DATA, and reside in the arc_mfu 517 * state. 518 */ 519 kstat_named_t arcstat_mfu_evictable_data; 520 /* 521 * Number of bytes consumed by ARC buffers that are eligible for 522 * eviction, of type ARC_BUFC_METADATA, and reside in the 523 * arc_mfu state. 524 */ 525 kstat_named_t arcstat_mfu_evictable_metadata; 526 /* 527 * Total number of bytes that *would have been* consumed by ARC 528 * buffers in the arc_mfu_ghost state. See the comment above 529 * arcstat_mru_ghost_size for more details. 530 */ 531 kstat_named_t arcstat_mfu_ghost_size; 532 /* 533 * Number of bytes that *would have been* consumed by ARC 534 * buffers that are eligible for eviction, of type 535 * ARC_BUFC_DATA, and linked off the arc_mfu_ghost state. 536 */ 537 kstat_named_t arcstat_mfu_ghost_evictable_data; 538 /* 539 * Number of bytes that *would have been* consumed by ARC 540 * buffers that are eligible for eviction, of type 541 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state. 542 */ 543 kstat_named_t arcstat_mfu_ghost_evictable_metadata; 544 kstat_named_t arcstat_l2_hits; 545 kstat_named_t arcstat_l2_misses; 546 kstat_named_t arcstat_l2_feeds; 547 kstat_named_t arcstat_l2_rw_clash; 548 kstat_named_t arcstat_l2_read_bytes; 549 kstat_named_t arcstat_l2_write_bytes; 550 kstat_named_t arcstat_l2_writes_sent; 551 kstat_named_t arcstat_l2_writes_done; 552 kstat_named_t arcstat_l2_writes_error; 553 kstat_named_t arcstat_l2_writes_lock_retry; 554 kstat_named_t arcstat_l2_evict_lock_retry; 555 kstat_named_t arcstat_l2_evict_reading; 556 kstat_named_t arcstat_l2_evict_l1cached; 557 kstat_named_t arcstat_l2_free_on_write; 558 kstat_named_t arcstat_l2_cdata_free_on_write; 559 kstat_named_t arcstat_l2_abort_lowmem; 560 kstat_named_t arcstat_l2_cksum_bad; 561 kstat_named_t arcstat_l2_io_error; 562 kstat_named_t arcstat_l2_size; 563 kstat_named_t arcstat_l2_asize; 564 kstat_named_t arcstat_l2_hdr_size; 565 kstat_named_t arcstat_l2_compress_successes; 566 kstat_named_t arcstat_l2_compress_zeros; 567 kstat_named_t arcstat_l2_compress_failures; 568 kstat_named_t arcstat_l2_write_trylock_fail; 569 kstat_named_t arcstat_l2_write_passed_headroom; 570 kstat_named_t arcstat_l2_write_spa_mismatch; 571 kstat_named_t arcstat_l2_write_in_l2; 572 kstat_named_t arcstat_l2_write_hdr_io_in_progress; 573 kstat_named_t arcstat_l2_write_not_cacheable; 574 kstat_named_t arcstat_l2_write_full; 575 kstat_named_t arcstat_l2_write_buffer_iter; 576 kstat_named_t arcstat_l2_write_pios; 577 kstat_named_t arcstat_l2_write_buffer_bytes_scanned; 578 kstat_named_t arcstat_l2_write_buffer_list_iter; 579 kstat_named_t arcstat_l2_write_buffer_list_null_iter; 580 kstat_named_t arcstat_memory_throttle_count; 581 kstat_named_t arcstat_duplicate_buffers; 582 kstat_named_t arcstat_duplicate_buffers_size; 583 kstat_named_t arcstat_duplicate_reads; 584 kstat_named_t arcstat_meta_used; 585 kstat_named_t arcstat_meta_limit; 586 kstat_named_t arcstat_meta_max; 587 kstat_named_t arcstat_meta_min; 588 kstat_named_t arcstat_sync_wait_for_async; 589 kstat_named_t arcstat_demand_hit_predictive_prefetch; 590} arc_stats_t; 591 592static arc_stats_t arc_stats = { 593 { "hits", KSTAT_DATA_UINT64 }, 594 { "misses", KSTAT_DATA_UINT64 }, 595 { "demand_data_hits", KSTAT_DATA_UINT64 }, 596 { "demand_data_misses", KSTAT_DATA_UINT64 }, 597 { "demand_metadata_hits", KSTAT_DATA_UINT64 }, 598 { "demand_metadata_misses", KSTAT_DATA_UINT64 }, 599 { "prefetch_data_hits", KSTAT_DATA_UINT64 }, 600 { "prefetch_data_misses", KSTAT_DATA_UINT64 }, 601 { "prefetch_metadata_hits", KSTAT_DATA_UINT64 }, 602 { "prefetch_metadata_misses", KSTAT_DATA_UINT64 }, 603 { "mru_hits", KSTAT_DATA_UINT64 }, 604 { "mru_ghost_hits", KSTAT_DATA_UINT64 }, 605 { "mfu_hits", KSTAT_DATA_UINT64 }, 606 { "mfu_ghost_hits", KSTAT_DATA_UINT64 }, 607 { "allocated", KSTAT_DATA_UINT64 }, 608 { "deleted", KSTAT_DATA_UINT64 }, 609 { "mutex_miss", KSTAT_DATA_UINT64 }, 610 { "evict_skip", KSTAT_DATA_UINT64 }, 611 { "evict_not_enough", KSTAT_DATA_UINT64 }, 612 { "evict_l2_cached", KSTAT_DATA_UINT64 }, 613 { "evict_l2_eligible", KSTAT_DATA_UINT64 }, 614 { "evict_l2_ineligible", KSTAT_DATA_UINT64 }, 615 { "evict_l2_skip", KSTAT_DATA_UINT64 }, 616 { "hash_elements", KSTAT_DATA_UINT64 }, 617 { "hash_elements_max", KSTAT_DATA_UINT64 }, 618 { "hash_collisions", KSTAT_DATA_UINT64 }, 619 { "hash_chains", KSTAT_DATA_UINT64 }, 620 { "hash_chain_max", KSTAT_DATA_UINT64 }, 621 { "p", KSTAT_DATA_UINT64 }, 622 { "c", KSTAT_DATA_UINT64 }, 623 { "c_min", KSTAT_DATA_UINT64 }, 624 { "c_max", KSTAT_DATA_UINT64 }, 625 { "size", KSTAT_DATA_UINT64 }, 626 { "hdr_size", KSTAT_DATA_UINT64 }, 627 { "data_size", KSTAT_DATA_UINT64 }, 628 { "metadata_size", KSTAT_DATA_UINT64 }, 629 { "other_size", KSTAT_DATA_UINT64 }, 630 { "anon_size", KSTAT_DATA_UINT64 }, 631 { "anon_evictable_data", KSTAT_DATA_UINT64 }, 632 { "anon_evictable_metadata", KSTAT_DATA_UINT64 }, 633 { "mru_size", KSTAT_DATA_UINT64 }, 634 { "mru_evictable_data", KSTAT_DATA_UINT64 }, 635 { "mru_evictable_metadata", KSTAT_DATA_UINT64 }, 636 { "mru_ghost_size", KSTAT_DATA_UINT64 }, 637 { "mru_ghost_evictable_data", KSTAT_DATA_UINT64 }, 638 { "mru_ghost_evictable_metadata", KSTAT_DATA_UINT64 }, 639 { "mfu_size", KSTAT_DATA_UINT64 }, 640 { "mfu_evictable_data", KSTAT_DATA_UINT64 }, 641 { "mfu_evictable_metadata", KSTAT_DATA_UINT64 }, 642 { "mfu_ghost_size", KSTAT_DATA_UINT64 }, 643 { "mfu_ghost_evictable_data", KSTAT_DATA_UINT64 }, 644 { "mfu_ghost_evictable_metadata", KSTAT_DATA_UINT64 }, 645 { "l2_hits", KSTAT_DATA_UINT64 }, 646 { "l2_misses", KSTAT_DATA_UINT64 }, 647 { "l2_feeds", KSTAT_DATA_UINT64 }, 648 { "l2_rw_clash", KSTAT_DATA_UINT64 }, 649 { "l2_read_bytes", KSTAT_DATA_UINT64 }, 650 { "l2_write_bytes", KSTAT_DATA_UINT64 }, 651 { "l2_writes_sent", KSTAT_DATA_UINT64 }, 652 { "l2_writes_done", KSTAT_DATA_UINT64 }, 653 { "l2_writes_error", KSTAT_DATA_UINT64 }, 654 { "l2_writes_lock_retry", KSTAT_DATA_UINT64 }, 655 { "l2_evict_lock_retry", KSTAT_DATA_UINT64 }, 656 { "l2_evict_reading", KSTAT_DATA_UINT64 }, 657 { "l2_evict_l1cached", KSTAT_DATA_UINT64 }, 658 { "l2_free_on_write", KSTAT_DATA_UINT64 }, 659 { "l2_cdata_free_on_write", KSTAT_DATA_UINT64 }, 660 { "l2_abort_lowmem", KSTAT_DATA_UINT64 }, 661 { "l2_cksum_bad", KSTAT_DATA_UINT64 }, 662 { "l2_io_error", KSTAT_DATA_UINT64 }, 663 { "l2_size", KSTAT_DATA_UINT64 }, 664 { "l2_asize", KSTAT_DATA_UINT64 }, 665 { "l2_hdr_size", KSTAT_DATA_UINT64 }, 666 { "l2_compress_successes", KSTAT_DATA_UINT64 }, 667 { "l2_compress_zeros", KSTAT_DATA_UINT64 }, 668 { "l2_compress_failures", KSTAT_DATA_UINT64 }, 669 { "l2_write_trylock_fail", KSTAT_DATA_UINT64 }, 670 { "l2_write_passed_headroom", KSTAT_DATA_UINT64 }, 671 { "l2_write_spa_mismatch", KSTAT_DATA_UINT64 }, 672 { "l2_write_in_l2", KSTAT_DATA_UINT64 }, 673 { "l2_write_io_in_progress", KSTAT_DATA_UINT64 }, 674 { "l2_write_not_cacheable", KSTAT_DATA_UINT64 }, 675 { "l2_write_full", KSTAT_DATA_UINT64 }, 676 { "l2_write_buffer_iter", KSTAT_DATA_UINT64 }, 677 { "l2_write_pios", KSTAT_DATA_UINT64 }, 678 { "l2_write_buffer_bytes_scanned", KSTAT_DATA_UINT64 }, 679 { "l2_write_buffer_list_iter", KSTAT_DATA_UINT64 }, 680 { "l2_write_buffer_list_null_iter", KSTAT_DATA_UINT64 }, 681 { "memory_throttle_count", KSTAT_DATA_UINT64 }, 682 { "duplicate_buffers", KSTAT_DATA_UINT64 }, 683 { "duplicate_buffers_size", KSTAT_DATA_UINT64 }, 684 { "duplicate_reads", KSTAT_DATA_UINT64 }, 685 { "arc_meta_used", KSTAT_DATA_UINT64 }, 686 { "arc_meta_limit", KSTAT_DATA_UINT64 }, 687 { "arc_meta_max", KSTAT_DATA_UINT64 }, 688 { "arc_meta_min", KSTAT_DATA_UINT64 }, 689 { "sync_wait_for_async", KSTAT_DATA_UINT64 }, 690 { "demand_hit_predictive_prefetch", KSTAT_DATA_UINT64 }, 691}; 692 693#define ARCSTAT(stat) (arc_stats.stat.value.ui64) 694 695#define ARCSTAT_INCR(stat, val) \ 696 atomic_add_64(&arc_stats.stat.value.ui64, (val)) 697 698#define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1) 699#define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1) 700 701#define ARCSTAT_MAX(stat, val) { \ 702 uint64_t m; \ 703 while ((val) > (m = arc_stats.stat.value.ui64) && \ 704 (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \ 705 continue; \ 706} 707 708#define ARCSTAT_MAXSTAT(stat) \ 709 ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64) 710 711/* 712 * We define a macro to allow ARC hits/misses to be easily broken down by 713 * two separate conditions, giving a total of four different subtypes for 714 * each of hits and misses (so eight statistics total). 715 */ 716#define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \ 717 if (cond1) { \ 718 if (cond2) { \ 719 ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \ 720 } else { \ 721 ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \ 722 } \ 723 } else { \ 724 if (cond2) { \ 725 ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \ 726 } else { \ 727 ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\ 728 } \ 729 } 730 731kstat_t *arc_ksp; 732static arc_state_t *arc_anon; 733static arc_state_t *arc_mru; 734static arc_state_t *arc_mru_ghost; 735static arc_state_t *arc_mfu; 736static arc_state_t *arc_mfu_ghost; 737static arc_state_t *arc_l2c_only; 738 739/* 740 * There are several ARC variables that are critical to export as kstats -- 741 * but we don't want to have to grovel around in the kstat whenever we wish to 742 * manipulate them. For these variables, we therefore define them to be in 743 * terms of the statistic variable. This assures that we are not introducing 744 * the possibility of inconsistency by having shadow copies of the variables, 745 * while still allowing the code to be readable. 746 */ 747#define arc_size ARCSTAT(arcstat_size) /* actual total arc size */ 748#define arc_p ARCSTAT(arcstat_p) /* target size of MRU */ 749#define arc_c ARCSTAT(arcstat_c) /* target size of cache */ 750#define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */ 751#define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */ 752#define arc_meta_limit ARCSTAT(arcstat_meta_limit) /* max size for metadata */ 753#define arc_meta_min ARCSTAT(arcstat_meta_min) /* min size for metadata */ 754#define arc_meta_used ARCSTAT(arcstat_meta_used) /* size of metadata */ 755#define arc_meta_max ARCSTAT(arcstat_meta_max) /* max size of metadata */ 756 757#define L2ARC_IS_VALID_COMPRESS(_c_) \ 758 ((_c_) == ZIO_COMPRESS_LZ4 || (_c_) == ZIO_COMPRESS_EMPTY) 759 760static int arc_no_grow; /* Don't try to grow cache size */ 761static uint64_t arc_tempreserve; 762static uint64_t arc_loaned_bytes; 763 764typedef struct arc_callback arc_callback_t; 765 766struct arc_callback { 767 void *acb_private; 768 arc_done_func_t *acb_done; 769 arc_buf_t *acb_buf; 770 zio_t *acb_zio_dummy; 771 arc_callback_t *acb_next; 772}; 773 774typedef struct arc_write_callback arc_write_callback_t; 775 776struct arc_write_callback { 777 void *awcb_private; 778 arc_done_func_t *awcb_ready; 779 arc_done_func_t *awcb_physdone; 780 arc_done_func_t *awcb_done; 781 arc_buf_t *awcb_buf; 782}; 783 784/* 785 * ARC buffers are separated into multiple structs as a memory saving measure: 786 * - Common fields struct, always defined, and embedded within it: 787 * - L2-only fields, always allocated but undefined when not in L2ARC 788 * - L1-only fields, only allocated when in L1ARC 789 * 790 * Buffer in L1 Buffer only in L2 791 * +------------------------+ +------------------------+ 792 * | arc_buf_hdr_t | | arc_buf_hdr_t | 793 * | | | | 794 * | | | | 795 * | | | | 796 * +------------------------+ +------------------------+ 797 * | l2arc_buf_hdr_t | | l2arc_buf_hdr_t | 798 * | (undefined if L1-only) | | | 799 * +------------------------+ +------------------------+ 800 * | l1arc_buf_hdr_t | 801 * | | 802 * | | 803 * | | 804 * | | 805 * +------------------------+ 806 * 807 * Because it's possible for the L2ARC to become extremely large, we can wind 808 * up eating a lot of memory in L2ARC buffer headers, so the size of a header 809 * is minimized by only allocating the fields necessary for an L1-cached buffer 810 * when a header is actually in the L1 cache. The sub-headers (l1arc_buf_hdr and 811 * l2arc_buf_hdr) are embedded rather than allocated separately to save a couple 812 * words in pointers. arc_hdr_realloc() is used to switch a header between 813 * these two allocation states. 814 */ 815typedef struct l1arc_buf_hdr { 816 kmutex_t b_freeze_lock; 817#ifdef ZFS_DEBUG 818 /* 819 * used for debugging wtih kmem_flags - by allocating and freeing 820 * b_thawed when the buffer is thawed, we get a record of the stack 821 * trace that thawed it. 822 */ 823 void *b_thawed; 824#endif 825 826 arc_buf_t *b_buf; 827 uint32_t b_datacnt; 828 /* for waiting on writes to complete */ 829 kcondvar_t b_cv; 830 831 /* protected by arc state mutex */ 832 arc_state_t *b_state; 833 multilist_node_t b_arc_node; 834 835 /* updated atomically */ 836 clock_t b_arc_access; 837 838 /* self protecting */ 839 refcount_t b_refcnt; 840 841 arc_callback_t *b_acb; 842 /* temporary buffer holder for in-flight compressed data */ 843 void *b_tmp_cdata; 844} l1arc_buf_hdr_t; 845 846typedef struct l2arc_dev l2arc_dev_t; 847 848typedef struct l2arc_buf_hdr { 849 /* protected by arc_buf_hdr mutex */ 850 l2arc_dev_t *b_dev; /* L2ARC device */ 851 uint64_t b_daddr; /* disk address, offset byte */ 852 /* real alloc'd buffer size depending on b_compress applied */ 853 int32_t b_asize; 854 uint8_t b_compress; 855 856 list_node_t b_l2node; 857} l2arc_buf_hdr_t; 858 859struct arc_buf_hdr { 860 /* protected by hash lock */ 861 dva_t b_dva; 862 uint64_t b_birth; 863 /* 864 * Even though this checksum is only set/verified when a buffer is in 865 * the L1 cache, it needs to be in the set of common fields because it 866 * must be preserved from the time before a buffer is written out to 867 * L2ARC until after it is read back in. 868 */ 869 zio_cksum_t *b_freeze_cksum; 870 871 arc_buf_hdr_t *b_hash_next; 872 arc_flags_t b_flags; 873 874 /* immutable */ 875 int32_t b_size; 876 uint64_t b_spa; 877 878 /* L2ARC fields. Undefined when not in L2ARC. */ 879 l2arc_buf_hdr_t b_l2hdr; 880 /* L1ARC fields. Undefined when in l2arc_only state */ 881 l1arc_buf_hdr_t b_l1hdr; 882}; 883 884#ifdef _KERNEL 885static int 886sysctl_vfs_zfs_arc_meta_limit(SYSCTL_HANDLER_ARGS) 887{ 888 uint64_t val; 889 int err; 890 891 val = arc_meta_limit; 892 err = sysctl_handle_64(oidp, &val, 0, req); 893 if (err != 0 || req->newptr == NULL) 894 return (err); 895 896 if (val <= 0 || val > arc_c_max) 897 return (EINVAL); 898 899 arc_meta_limit = val; 900 return (0); 901} 902#endif 903 904static arc_buf_t *arc_eviction_list; 905static arc_buf_hdr_t arc_eviction_hdr; 906 907#define GHOST_STATE(state) \ 908 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \ 909 (state) == arc_l2c_only) 910 911#define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_FLAG_IN_HASH_TABLE) 912#define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) 913#define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_FLAG_IO_ERROR) 914#define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_FLAG_PREFETCH) 915#define HDR_FREED_IN_READ(hdr) ((hdr)->b_flags & ARC_FLAG_FREED_IN_READ) 916#define HDR_BUF_AVAILABLE(hdr) ((hdr)->b_flags & ARC_FLAG_BUF_AVAILABLE) 917 918#define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_FLAG_L2CACHE) 919#define HDR_L2COMPRESS(hdr) ((hdr)->b_flags & ARC_FLAG_L2COMPRESS) 920#define HDR_L2_READING(hdr) \ 921 (((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) && \ 922 ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR)) 923#define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITING) 924#define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_FLAG_L2_EVICTED) 925#define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITE_HEAD) 926 927#define HDR_ISTYPE_METADATA(hdr) \ 928 ((hdr)->b_flags & ARC_FLAG_BUFC_METADATA) 929#define HDR_ISTYPE_DATA(hdr) (!HDR_ISTYPE_METADATA(hdr)) 930 931#define HDR_HAS_L1HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L1HDR) 932#define HDR_HAS_L2HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR) 933 934/* 935 * Other sizes 936 */ 937 938#define HDR_FULL_SIZE ((int64_t)sizeof (arc_buf_hdr_t)) 939#define HDR_L2ONLY_SIZE ((int64_t)offsetof(arc_buf_hdr_t, b_l1hdr)) 940 941/* 942 * Hash table routines 943 */ 944 945#define HT_LOCK_PAD CACHE_LINE_SIZE 946 947struct ht_lock { 948 kmutex_t ht_lock; 949#ifdef _KERNEL 950 unsigned char pad[(HT_LOCK_PAD - sizeof (kmutex_t))]; 951#endif 952}; 953 954#define BUF_LOCKS 256 955typedef struct buf_hash_table { 956 uint64_t ht_mask; 957 arc_buf_hdr_t **ht_table; 958 struct ht_lock ht_locks[BUF_LOCKS] __aligned(CACHE_LINE_SIZE); 959} buf_hash_table_t; 960 961static buf_hash_table_t buf_hash_table; 962 963#define BUF_HASH_INDEX(spa, dva, birth) \ 964 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask) 965#define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)]) 966#define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock)) 967#define HDR_LOCK(hdr) \ 968 (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth))) 969 970uint64_t zfs_crc64_table[256]; 971 972/* 973 * Level 2 ARC 974 */ 975 976#define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */ 977#define L2ARC_HEADROOM 2 /* num of writes */ 978/* 979 * If we discover during ARC scan any buffers to be compressed, we boost 980 * our headroom for the next scanning cycle by this percentage multiple. 981 */ 982#define L2ARC_HEADROOM_BOOST 200 983#define L2ARC_FEED_SECS 1 /* caching interval secs */ 984#define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */ 985 986/* 987 * Used to distinguish headers that are being process by 988 * l2arc_write_buffers(), but have yet to be assigned to a l2arc disk 989 * address. This can happen when the header is added to the l2arc's list 990 * of buffers to write in the first stage of l2arc_write_buffers(), but 991 * has not yet been written out which happens in the second stage of 992 * l2arc_write_buffers(). 993 */ 994#define L2ARC_ADDR_UNSET ((uint64_t)(-1)) 995 996#define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent) 997#define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done) 998 999/* L2ARC Performance Tunables */ 1000uint64_t l2arc_write_max = L2ARC_WRITE_SIZE; /* default max write size */ 1001uint64_t l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra write during warmup */ 1002uint64_t l2arc_headroom = L2ARC_HEADROOM; /* number of dev writes */ 1003uint64_t l2arc_headroom_boost = L2ARC_HEADROOM_BOOST; 1004uint64_t l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */ 1005uint64_t l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval milliseconds */ 1006boolean_t l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */ 1007boolean_t l2arc_feed_again = B_TRUE; /* turbo warmup */ 1008boolean_t l2arc_norw = B_TRUE; /* no reads during writes */ 1009 1010SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_max, CTLFLAG_RW, 1011 &l2arc_write_max, 0, "max write size"); 1012SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_boost, CTLFLAG_RW, 1013 &l2arc_write_boost, 0, "extra write during warmup"); 1014SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_headroom, CTLFLAG_RW, 1015 &l2arc_headroom, 0, "number of dev writes"); 1016SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_secs, CTLFLAG_RW, 1017 &l2arc_feed_secs, 0, "interval seconds"); 1018SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_min_ms, CTLFLAG_RW, 1019 &l2arc_feed_min_ms, 0, "min interval milliseconds"); 1020 1021SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_noprefetch, CTLFLAG_RW, 1022 &l2arc_noprefetch, 0, "don't cache prefetch bufs"); 1023SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_feed_again, CTLFLAG_RW, 1024 &l2arc_feed_again, 0, "turbo warmup"); 1025SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_norw, CTLFLAG_RW, 1026 &l2arc_norw, 0, "no reads during writes"); 1027 1028SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_size, CTLFLAG_RD, 1029 &ARC_anon.arcs_size.rc_count, 0, "size of anonymous state"); 1030SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_metadata_lsize, CTLFLAG_RD, 1031 &ARC_anon.arcs_lsize[ARC_BUFC_METADATA], 0, "size of anonymous state"); 1032SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_data_lsize, CTLFLAG_RD, 1033 &ARC_anon.arcs_lsize[ARC_BUFC_DATA], 0, "size of anonymous state"); 1034 1035SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_size, CTLFLAG_RD, 1036 &ARC_mru.arcs_size.rc_count, 0, "size of mru state"); 1037SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_metadata_lsize, CTLFLAG_RD, 1038 &ARC_mru.arcs_lsize[ARC_BUFC_METADATA], 0, "size of metadata in mru state"); 1039SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_data_lsize, CTLFLAG_RD, 1040 &ARC_mru.arcs_lsize[ARC_BUFC_DATA], 0, "size of data in mru state"); 1041 1042SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_size, CTLFLAG_RD, 1043 &ARC_mru_ghost.arcs_size.rc_count, 0, "size of mru ghost state"); 1044SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_metadata_lsize, CTLFLAG_RD, 1045 &ARC_mru_ghost.arcs_lsize[ARC_BUFC_METADATA], 0, 1046 "size of metadata in mru ghost state"); 1047SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_data_lsize, CTLFLAG_RD, 1048 &ARC_mru_ghost.arcs_lsize[ARC_BUFC_DATA], 0, 1049 "size of data in mru ghost state"); 1050 1051SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_size, CTLFLAG_RD, 1052 &ARC_mfu.arcs_size.rc_count, 0, "size of mfu state"); 1053SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_metadata_lsize, CTLFLAG_RD, 1054 &ARC_mfu.arcs_lsize[ARC_BUFC_METADATA], 0, "size of metadata in mfu state"); 1055SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_data_lsize, CTLFLAG_RD, 1056 &ARC_mfu.arcs_lsize[ARC_BUFC_DATA], 0, "size of data in mfu state"); 1057 1058SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_size, CTLFLAG_RD, 1059 &ARC_mfu_ghost.arcs_size.rc_count, 0, "size of mfu ghost state"); 1060SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_metadata_lsize, CTLFLAG_RD, 1061 &ARC_mfu_ghost.arcs_lsize[ARC_BUFC_METADATA], 0, 1062 "size of metadata in mfu ghost state"); 1063SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_data_lsize, CTLFLAG_RD, 1064 &ARC_mfu_ghost.arcs_lsize[ARC_BUFC_DATA], 0, 1065 "size of data in mfu ghost state"); 1066 1067SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2c_only_size, CTLFLAG_RD, 1068 &ARC_l2c_only.arcs_size.rc_count, 0, "size of mru state"); 1069 1070/* 1071 * L2ARC Internals 1072 */ 1073struct l2arc_dev { 1074 vdev_t *l2ad_vdev; /* vdev */ 1075 spa_t *l2ad_spa; /* spa */ 1076 uint64_t l2ad_hand; /* next write location */ 1077 uint64_t l2ad_start; /* first addr on device */ 1078 uint64_t l2ad_end; /* last addr on device */ 1079 boolean_t l2ad_first; /* first sweep through */ 1080 boolean_t l2ad_writing; /* currently writing */ 1081 kmutex_t l2ad_mtx; /* lock for buffer list */ 1082 list_t l2ad_buflist; /* buffer list */ 1083 list_node_t l2ad_node; /* device list node */ 1084 refcount_t l2ad_alloc; /* allocated bytes */ 1085}; 1086 1087static list_t L2ARC_dev_list; /* device list */ 1088static list_t *l2arc_dev_list; /* device list pointer */ 1089static kmutex_t l2arc_dev_mtx; /* device list mutex */ 1090static l2arc_dev_t *l2arc_dev_last; /* last device used */ 1091static list_t L2ARC_free_on_write; /* free after write buf list */ 1092static list_t *l2arc_free_on_write; /* free after write list ptr */ 1093static kmutex_t l2arc_free_on_write_mtx; /* mutex for list */ 1094static uint64_t l2arc_ndev; /* number of devices */ 1095 1096typedef struct l2arc_read_callback { 1097 arc_buf_t *l2rcb_buf; /* read buffer */ 1098 spa_t *l2rcb_spa; /* spa */ 1099 blkptr_t l2rcb_bp; /* original blkptr */ 1100 zbookmark_phys_t l2rcb_zb; /* original bookmark */ 1101 int l2rcb_flags; /* original flags */ 1102 enum zio_compress l2rcb_compress; /* applied compress */ 1103} l2arc_read_callback_t; 1104 1105typedef struct l2arc_write_callback { 1106 l2arc_dev_t *l2wcb_dev; /* device info */ 1107 arc_buf_hdr_t *l2wcb_head; /* head of write buflist */ 1108} l2arc_write_callback_t; 1109 1110typedef struct l2arc_data_free { 1111 /* protected by l2arc_free_on_write_mtx */ 1112 void *l2df_data; 1113 size_t l2df_size; 1114 void (*l2df_func)(void *, size_t); 1115 list_node_t l2df_list_node; 1116} l2arc_data_free_t; 1117 1118static kmutex_t l2arc_feed_thr_lock; 1119static kcondvar_t l2arc_feed_thr_cv; 1120static uint8_t l2arc_thread_exit; 1121 1122static void arc_get_data_buf(arc_buf_t *); 1123static void arc_access(arc_buf_hdr_t *, kmutex_t *); 1124static boolean_t arc_is_overflowing(); 1125static void arc_buf_watch(arc_buf_t *); 1126 1127static arc_buf_contents_t arc_buf_type(arc_buf_hdr_t *); 1128static uint32_t arc_bufc_to_flags(arc_buf_contents_t); 1129 1130static boolean_t l2arc_write_eligible(uint64_t, arc_buf_hdr_t *); 1131static void l2arc_read_done(zio_t *); 1132 1133static boolean_t l2arc_compress_buf(arc_buf_hdr_t *); 1134static void l2arc_decompress_zio(zio_t *, arc_buf_hdr_t *, enum zio_compress); 1135static void l2arc_release_cdata_buf(arc_buf_hdr_t *); 1136 1137static uint64_t 1138buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth) 1139{ 1140 uint8_t *vdva = (uint8_t *)dva; 1141 uint64_t crc = -1ULL; 1142 int i; 1143 1144 ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY); 1145 1146 for (i = 0; i < sizeof (dva_t); i++) 1147 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ vdva[i]) & 0xFF]; 1148 1149 crc ^= (spa>>8) ^ birth; 1150 1151 return (crc); 1152} 1153 1154#define BUF_EMPTY(buf) \ 1155 ((buf)->b_dva.dva_word[0] == 0 && \ 1156 (buf)->b_dva.dva_word[1] == 0) 1157 1158#define BUF_EQUAL(spa, dva, birth, buf) \ 1159 ((buf)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \ 1160 ((buf)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \ 1161 ((buf)->b_birth == birth) && ((buf)->b_spa == spa) 1162 1163static void 1164buf_discard_identity(arc_buf_hdr_t *hdr) 1165{ 1166 hdr->b_dva.dva_word[0] = 0; 1167 hdr->b_dva.dva_word[1] = 0; 1168 hdr->b_birth = 0; 1169} 1170 1171static arc_buf_hdr_t * 1172buf_hash_find(uint64_t spa, const blkptr_t *bp, kmutex_t **lockp) 1173{ 1174 const dva_t *dva = BP_IDENTITY(bp); 1175 uint64_t birth = BP_PHYSICAL_BIRTH(bp); 1176 uint64_t idx = BUF_HASH_INDEX(spa, dva, birth); 1177 kmutex_t *hash_lock = BUF_HASH_LOCK(idx); 1178 arc_buf_hdr_t *hdr; 1179 1180 mutex_enter(hash_lock); 1181 for (hdr = buf_hash_table.ht_table[idx]; hdr != NULL; 1182 hdr = hdr->b_hash_next) { 1183 if (BUF_EQUAL(spa, dva, birth, hdr)) { 1184 *lockp = hash_lock; 1185 return (hdr); 1186 } 1187 } 1188 mutex_exit(hash_lock); 1189 *lockp = NULL; 1190 return (NULL); 1191} 1192 1193/* 1194 * Insert an entry into the hash table. If there is already an element 1195 * equal to elem in the hash table, then the already existing element 1196 * will be returned and the new element will not be inserted. 1197 * Otherwise returns NULL. 1198 * If lockp == NULL, the caller is assumed to already hold the hash lock. 1199 */ 1200static arc_buf_hdr_t * 1201buf_hash_insert(arc_buf_hdr_t *hdr, kmutex_t **lockp) 1202{ 1203 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth); 1204 kmutex_t *hash_lock = BUF_HASH_LOCK(idx); 1205 arc_buf_hdr_t *fhdr; 1206 uint32_t i; 1207 1208 ASSERT(!DVA_IS_EMPTY(&hdr->b_dva)); 1209 ASSERT(hdr->b_birth != 0); 1210 ASSERT(!HDR_IN_HASH_TABLE(hdr)); 1211 1212 if (lockp != NULL) { 1213 *lockp = hash_lock; 1214 mutex_enter(hash_lock); 1215 } else { 1216 ASSERT(MUTEX_HELD(hash_lock)); 1217 } 1218 1219 for (fhdr = buf_hash_table.ht_table[idx], i = 0; fhdr != NULL; 1220 fhdr = fhdr->b_hash_next, i++) { 1221 if (BUF_EQUAL(hdr->b_spa, &hdr->b_dva, hdr->b_birth, fhdr)) 1222 return (fhdr); 1223 } 1224 1225 hdr->b_hash_next = buf_hash_table.ht_table[idx]; 1226 buf_hash_table.ht_table[idx] = hdr; 1227 hdr->b_flags |= ARC_FLAG_IN_HASH_TABLE; 1228 1229 /* collect some hash table performance data */ 1230 if (i > 0) { 1231 ARCSTAT_BUMP(arcstat_hash_collisions); 1232 if (i == 1) 1233 ARCSTAT_BUMP(arcstat_hash_chains); 1234 1235 ARCSTAT_MAX(arcstat_hash_chain_max, i); 1236 } 1237 1238 ARCSTAT_BUMP(arcstat_hash_elements); 1239 ARCSTAT_MAXSTAT(arcstat_hash_elements); 1240 1241 return (NULL); 1242} 1243 1244static void 1245buf_hash_remove(arc_buf_hdr_t *hdr) 1246{ 1247 arc_buf_hdr_t *fhdr, **hdrp; 1248 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth); 1249 1250 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx))); 1251 ASSERT(HDR_IN_HASH_TABLE(hdr)); 1252 1253 hdrp = &buf_hash_table.ht_table[idx]; 1254 while ((fhdr = *hdrp) != hdr) { 1255 ASSERT(fhdr != NULL); 1256 hdrp = &fhdr->b_hash_next; 1257 } 1258 *hdrp = hdr->b_hash_next; 1259 hdr->b_hash_next = NULL; 1260 hdr->b_flags &= ~ARC_FLAG_IN_HASH_TABLE; 1261 1262 /* collect some hash table performance data */ 1263 ARCSTAT_BUMPDOWN(arcstat_hash_elements); 1264 1265 if (buf_hash_table.ht_table[idx] && 1266 buf_hash_table.ht_table[idx]->b_hash_next == NULL) 1267 ARCSTAT_BUMPDOWN(arcstat_hash_chains); 1268} 1269 1270/* 1271 * Global data structures and functions for the buf kmem cache. 1272 */ 1273static kmem_cache_t *hdr_full_cache; 1274static kmem_cache_t *hdr_l2only_cache; 1275static kmem_cache_t *buf_cache; 1276 1277static void 1278buf_fini(void) 1279{ 1280 int i; 1281 1282 kmem_free(buf_hash_table.ht_table, 1283 (buf_hash_table.ht_mask + 1) * sizeof (void *)); 1284 for (i = 0; i < BUF_LOCKS; i++) 1285 mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock); 1286 kmem_cache_destroy(hdr_full_cache); 1287 kmem_cache_destroy(hdr_l2only_cache); 1288 kmem_cache_destroy(buf_cache); 1289} 1290 1291/* 1292 * Constructor callback - called when the cache is empty 1293 * and a new buf is requested. 1294 */ 1295/* ARGSUSED */ 1296static int 1297hdr_full_cons(void *vbuf, void *unused, int kmflag) 1298{ 1299 arc_buf_hdr_t *hdr = vbuf; 1300 1301 bzero(hdr, HDR_FULL_SIZE); 1302 cv_init(&hdr->b_l1hdr.b_cv, NULL, CV_DEFAULT, NULL); 1303 refcount_create(&hdr->b_l1hdr.b_refcnt); 1304 mutex_init(&hdr->b_l1hdr.b_freeze_lock, NULL, MUTEX_DEFAULT, NULL); 1305 multilist_link_init(&hdr->b_l1hdr.b_arc_node); 1306 arc_space_consume(HDR_FULL_SIZE, ARC_SPACE_HDRS); 1307 1308 return (0); 1309} 1310 1311/* ARGSUSED */ 1312static int 1313hdr_l2only_cons(void *vbuf, void *unused, int kmflag) 1314{ 1315 arc_buf_hdr_t *hdr = vbuf; 1316 1317 bzero(hdr, HDR_L2ONLY_SIZE); 1318 arc_space_consume(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS); 1319 1320 return (0); 1321} 1322 1323/* ARGSUSED */ 1324static int 1325buf_cons(void *vbuf, void *unused, int kmflag) 1326{ 1327 arc_buf_t *buf = vbuf; 1328 1329 bzero(buf, sizeof (arc_buf_t)); 1330 mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL); 1331 arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS); 1332 1333 return (0); 1334} 1335 1336/* 1337 * Destructor callback - called when a cached buf is 1338 * no longer required. 1339 */ 1340/* ARGSUSED */ 1341static void 1342hdr_full_dest(void *vbuf, void *unused) 1343{ 1344 arc_buf_hdr_t *hdr = vbuf; 1345 1346 ASSERT(BUF_EMPTY(hdr)); 1347 cv_destroy(&hdr->b_l1hdr.b_cv); 1348 refcount_destroy(&hdr->b_l1hdr.b_refcnt); 1349 mutex_destroy(&hdr->b_l1hdr.b_freeze_lock); 1350 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 1351 arc_space_return(HDR_FULL_SIZE, ARC_SPACE_HDRS); 1352} 1353 1354/* ARGSUSED */ 1355static void 1356hdr_l2only_dest(void *vbuf, void *unused) 1357{ 1358 arc_buf_hdr_t *hdr = vbuf; 1359 1360 ASSERT(BUF_EMPTY(hdr)); 1361 arc_space_return(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS); 1362} 1363 1364/* ARGSUSED */ 1365static void 1366buf_dest(void *vbuf, void *unused) 1367{ 1368 arc_buf_t *buf = vbuf; 1369 1370 mutex_destroy(&buf->b_evict_lock); 1371 arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS); 1372} 1373 1374/* 1375 * Reclaim callback -- invoked when memory is low. 1376 */ 1377/* ARGSUSED */ 1378static void 1379hdr_recl(void *unused) 1380{ 1381 dprintf("hdr_recl called\n"); 1382 /* 1383 * umem calls the reclaim func when we destroy the buf cache, 1384 * which is after we do arc_fini(). 1385 */ 1386 if (!arc_dead) 1387 cv_signal(&arc_reclaim_thread_cv); 1388} 1389 1390static void 1391buf_init(void) 1392{ 1393 uint64_t *ct; 1394 uint64_t hsize = 1ULL << 12; 1395 int i, j; 1396 1397 /* 1398 * The hash table is big enough to fill all of physical memory 1399 * with an average block size of zfs_arc_average_blocksize (default 8K). 1400 * By default, the table will take up 1401 * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers). 1402 */ 1403 while (hsize * zfs_arc_average_blocksize < (uint64_t)physmem * PAGESIZE) 1404 hsize <<= 1; 1405retry: 1406 buf_hash_table.ht_mask = hsize - 1; 1407 buf_hash_table.ht_table = 1408 kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP); 1409 if (buf_hash_table.ht_table == NULL) { 1410 ASSERT(hsize > (1ULL << 8)); 1411 hsize >>= 1; 1412 goto retry; 1413 } 1414 1415 hdr_full_cache = kmem_cache_create("arc_buf_hdr_t_full", HDR_FULL_SIZE, 1416 0, hdr_full_cons, hdr_full_dest, hdr_recl, NULL, NULL, 0); 1417 hdr_l2only_cache = kmem_cache_create("arc_buf_hdr_t_l2only", 1418 HDR_L2ONLY_SIZE, 0, hdr_l2only_cons, hdr_l2only_dest, hdr_recl, 1419 NULL, NULL, 0); 1420 buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t), 1421 0, buf_cons, buf_dest, NULL, NULL, NULL, 0); 1422 1423 for (i = 0; i < 256; i++) 1424 for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--) 1425 *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY); 1426 1427 for (i = 0; i < BUF_LOCKS; i++) { 1428 mutex_init(&buf_hash_table.ht_locks[i].ht_lock, 1429 NULL, MUTEX_DEFAULT, NULL); 1430 } 1431} 1432 1433/* 1434 * Transition between the two allocation states for the arc_buf_hdr struct. 1435 * The arc_buf_hdr struct can be allocated with (hdr_full_cache) or without 1436 * (hdr_l2only_cache) the fields necessary for the L1 cache - the smaller 1437 * version is used when a cache buffer is only in the L2ARC in order to reduce 1438 * memory usage. 1439 */ 1440static arc_buf_hdr_t * 1441arc_hdr_realloc(arc_buf_hdr_t *hdr, kmem_cache_t *old, kmem_cache_t *new) 1442{ 1443 ASSERT(HDR_HAS_L2HDR(hdr)); 1444 1445 arc_buf_hdr_t *nhdr; 1446 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev; 1447 1448 ASSERT((old == hdr_full_cache && new == hdr_l2only_cache) || 1449 (old == hdr_l2only_cache && new == hdr_full_cache)); 1450 1451 nhdr = kmem_cache_alloc(new, KM_PUSHPAGE); 1452 1453 ASSERT(MUTEX_HELD(HDR_LOCK(hdr))); 1454 buf_hash_remove(hdr); 1455 1456 bcopy(hdr, nhdr, HDR_L2ONLY_SIZE); 1457 1458 if (new == hdr_full_cache) { 1459 nhdr->b_flags |= ARC_FLAG_HAS_L1HDR; 1460 /* 1461 * arc_access and arc_change_state need to be aware that a 1462 * header has just come out of L2ARC, so we set its state to 1463 * l2c_only even though it's about to change. 1464 */ 1465 nhdr->b_l1hdr.b_state = arc_l2c_only; 1466 1467 /* Verify previous threads set to NULL before freeing */ 1468 ASSERT3P(nhdr->b_l1hdr.b_tmp_cdata, ==, NULL); 1469 } else { 1470 ASSERT(hdr->b_l1hdr.b_buf == NULL); 1471 ASSERT0(hdr->b_l1hdr.b_datacnt); 1472 1473 /* 1474 * If we've reached here, We must have been called from 1475 * arc_evict_hdr(), as such we should have already been 1476 * removed from any ghost list we were previously on 1477 * (which protects us from racing with arc_evict_state), 1478 * thus no locking is needed during this check. 1479 */ 1480 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 1481 1482 /* 1483 * A buffer must not be moved into the arc_l2c_only 1484 * state if it's not finished being written out to the 1485 * l2arc device. Otherwise, the b_l1hdr.b_tmp_cdata field 1486 * might try to be accessed, even though it was removed. 1487 */ 1488 VERIFY(!HDR_L2_WRITING(hdr)); 1489 VERIFY3P(hdr->b_l1hdr.b_tmp_cdata, ==, NULL); 1490 1491#ifdef ZFS_DEBUG 1492 if (hdr->b_l1hdr.b_thawed != NULL) { 1493 kmem_free(hdr->b_l1hdr.b_thawed, 1); 1494 hdr->b_l1hdr.b_thawed = NULL; 1495 } 1496#endif 1497 1498 nhdr->b_flags &= ~ARC_FLAG_HAS_L1HDR; 1499 } 1500 /* 1501 * The header has been reallocated so we need to re-insert it into any 1502 * lists it was on. 1503 */ 1504 (void) buf_hash_insert(nhdr, NULL); 1505 1506 ASSERT(list_link_active(&hdr->b_l2hdr.b_l2node)); 1507 1508 mutex_enter(&dev->l2ad_mtx); 1509 1510 /* 1511 * We must place the realloc'ed header back into the list at 1512 * the same spot. Otherwise, if it's placed earlier in the list, 1513 * l2arc_write_buffers() could find it during the function's 1514 * write phase, and try to write it out to the l2arc. 1515 */ 1516 list_insert_after(&dev->l2ad_buflist, hdr, nhdr); 1517 list_remove(&dev->l2ad_buflist, hdr); 1518 1519 mutex_exit(&dev->l2ad_mtx); 1520 1521 /* 1522 * Since we're using the pointer address as the tag when 1523 * incrementing and decrementing the l2ad_alloc refcount, we 1524 * must remove the old pointer (that we're about to destroy) and 1525 * add the new pointer to the refcount. Otherwise we'd remove 1526 * the wrong pointer address when calling arc_hdr_destroy() later. 1527 */ 1528 1529 (void) refcount_remove_many(&dev->l2ad_alloc, 1530 hdr->b_l2hdr.b_asize, hdr); 1531 1532 (void) refcount_add_many(&dev->l2ad_alloc, 1533 nhdr->b_l2hdr.b_asize, nhdr); 1534 1535 buf_discard_identity(hdr); 1536 hdr->b_freeze_cksum = NULL; 1537 kmem_cache_free(old, hdr); 1538 1539 return (nhdr); 1540} 1541 1542 1543#define ARC_MINTIME (hz>>4) /* 62 ms */ 1544 1545static void 1546arc_cksum_verify(arc_buf_t *buf) 1547{ 1548 zio_cksum_t zc; 1549 1550 if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 1551 return; 1552 1553 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1554 if (buf->b_hdr->b_freeze_cksum == NULL || HDR_IO_ERROR(buf->b_hdr)) { 1555 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1556 return; 1557 } 1558 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc); 1559 if (!ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc)) 1560 panic("buffer modified while frozen!"); 1561 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1562} 1563 1564static int 1565arc_cksum_equal(arc_buf_t *buf) 1566{ 1567 zio_cksum_t zc; 1568 int equal; 1569 1570 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1571 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc); 1572 equal = ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc); 1573 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1574 1575 return (equal); 1576} 1577 1578static void 1579arc_cksum_compute(arc_buf_t *buf, boolean_t force) 1580{ 1581 if (!force && !(zfs_flags & ZFS_DEBUG_MODIFY)) 1582 return; 1583 1584 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1585 if (buf->b_hdr->b_freeze_cksum != NULL) { 1586 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1587 return; 1588 } 1589 buf->b_hdr->b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t), KM_SLEEP); 1590 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, 1591 buf->b_hdr->b_freeze_cksum); 1592 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1593#ifdef illumos 1594 arc_buf_watch(buf); 1595#endif /* illumos */ 1596} 1597 1598#ifdef illumos 1599#ifndef _KERNEL 1600typedef struct procctl { 1601 long cmd; 1602 prwatch_t prwatch; 1603} procctl_t; 1604#endif 1605 1606/* ARGSUSED */ 1607static void 1608arc_buf_unwatch(arc_buf_t *buf) 1609{ 1610#ifndef _KERNEL 1611 if (arc_watch) { 1612 int result; 1613 procctl_t ctl; 1614 ctl.cmd = PCWATCH; 1615 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data; 1616 ctl.prwatch.pr_size = 0; 1617 ctl.prwatch.pr_wflags = 0; 1618 result = write(arc_procfd, &ctl, sizeof (ctl)); 1619 ASSERT3U(result, ==, sizeof (ctl)); 1620 } 1621#endif 1622} 1623 1624/* ARGSUSED */ 1625static void 1626arc_buf_watch(arc_buf_t *buf) 1627{ 1628#ifndef _KERNEL 1629 if (arc_watch) { 1630 int result; 1631 procctl_t ctl; 1632 ctl.cmd = PCWATCH; 1633 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data; 1634 ctl.prwatch.pr_size = buf->b_hdr->b_size; 1635 ctl.prwatch.pr_wflags = WA_WRITE; 1636 result = write(arc_procfd, &ctl, sizeof (ctl)); 1637 ASSERT3U(result, ==, sizeof (ctl)); 1638 } 1639#endif 1640} 1641#endif /* illumos */ 1642 1643static arc_buf_contents_t 1644arc_buf_type(arc_buf_hdr_t *hdr) 1645{ 1646 if (HDR_ISTYPE_METADATA(hdr)) { 1647 return (ARC_BUFC_METADATA); 1648 } else { 1649 return (ARC_BUFC_DATA); 1650 } 1651} 1652 1653static uint32_t 1654arc_bufc_to_flags(arc_buf_contents_t type) 1655{ 1656 switch (type) { 1657 case ARC_BUFC_DATA: 1658 /* metadata field is 0 if buffer contains normal data */ 1659 return (0); 1660 case ARC_BUFC_METADATA: 1661 return (ARC_FLAG_BUFC_METADATA); 1662 default: 1663 break; 1664 } 1665 panic("undefined ARC buffer type!"); 1666 return ((uint32_t)-1); 1667} 1668 1669void 1670arc_buf_thaw(arc_buf_t *buf) 1671{ 1672 if (zfs_flags & ZFS_DEBUG_MODIFY) { 1673 if (buf->b_hdr->b_l1hdr.b_state != arc_anon) 1674 panic("modifying non-anon buffer!"); 1675 if (HDR_IO_IN_PROGRESS(buf->b_hdr)) 1676 panic("modifying buffer while i/o in progress!"); 1677 arc_cksum_verify(buf); 1678 } 1679 1680 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1681 if (buf->b_hdr->b_freeze_cksum != NULL) { 1682 kmem_free(buf->b_hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 1683 buf->b_hdr->b_freeze_cksum = NULL; 1684 } 1685 1686#ifdef ZFS_DEBUG 1687 if (zfs_flags & ZFS_DEBUG_MODIFY) { 1688 if (buf->b_hdr->b_l1hdr.b_thawed != NULL) 1689 kmem_free(buf->b_hdr->b_l1hdr.b_thawed, 1); 1690 buf->b_hdr->b_l1hdr.b_thawed = kmem_alloc(1, KM_SLEEP); 1691 } 1692#endif 1693 1694 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1695 1696#ifdef illumos 1697 arc_buf_unwatch(buf); 1698#endif /* illumos */ 1699} 1700 1701void 1702arc_buf_freeze(arc_buf_t *buf) 1703{ 1704 kmutex_t *hash_lock; 1705 1706 if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 1707 return; 1708 1709 hash_lock = HDR_LOCK(buf->b_hdr); 1710 mutex_enter(hash_lock); 1711 1712 ASSERT(buf->b_hdr->b_freeze_cksum != NULL || 1713 buf->b_hdr->b_l1hdr.b_state == arc_anon); 1714 arc_cksum_compute(buf, B_FALSE); 1715 mutex_exit(hash_lock); 1716 1717} 1718 1719static void 1720add_reference(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, void *tag) 1721{ 1722 ASSERT(HDR_HAS_L1HDR(hdr)); 1723 ASSERT(MUTEX_HELD(hash_lock)); 1724 arc_state_t *state = hdr->b_l1hdr.b_state; 1725 1726 if ((refcount_add(&hdr->b_l1hdr.b_refcnt, tag) == 1) && 1727 (state != arc_anon)) { 1728 /* We don't use the L2-only state list. */ 1729 if (state != arc_l2c_only) { 1730 arc_buf_contents_t type = arc_buf_type(hdr); 1731 uint64_t delta = hdr->b_size * hdr->b_l1hdr.b_datacnt; 1732 multilist_t *list = &state->arcs_list[type]; 1733 uint64_t *size = &state->arcs_lsize[type]; 1734 1735 multilist_remove(list, hdr); 1736 1737 if (GHOST_STATE(state)) { 1738 ASSERT0(hdr->b_l1hdr.b_datacnt); 1739 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 1740 delta = hdr->b_size; 1741 } 1742 ASSERT(delta > 0); 1743 ASSERT3U(*size, >=, delta); 1744 atomic_add_64(size, -delta); 1745 } 1746 /* remove the prefetch flag if we get a reference */ 1747 hdr->b_flags &= ~ARC_FLAG_PREFETCH; 1748 } 1749} 1750 1751static int 1752remove_reference(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, void *tag) 1753{ 1754 int cnt; 1755 arc_state_t *state = hdr->b_l1hdr.b_state; 1756 1757 ASSERT(HDR_HAS_L1HDR(hdr)); 1758 ASSERT(state == arc_anon || MUTEX_HELD(hash_lock)); 1759 ASSERT(!GHOST_STATE(state)); 1760 1761 /* 1762 * arc_l2c_only counts as a ghost state so we don't need to explicitly 1763 * check to prevent usage of the arc_l2c_only list. 1764 */ 1765 if (((cnt = refcount_remove(&hdr->b_l1hdr.b_refcnt, tag)) == 0) && 1766 (state != arc_anon)) { 1767 arc_buf_contents_t type = arc_buf_type(hdr); 1768 multilist_t *list = &state->arcs_list[type]; 1769 uint64_t *size = &state->arcs_lsize[type]; 1770 1771 multilist_insert(list, hdr); 1772 1773 ASSERT(hdr->b_l1hdr.b_datacnt > 0); 1774 atomic_add_64(size, hdr->b_size * 1775 hdr->b_l1hdr.b_datacnt); 1776 } 1777 return (cnt); 1778} 1779 1780/* 1781 * Move the supplied buffer to the indicated state. The hash lock 1782 * for the buffer must be held by the caller. 1783 */ 1784static void 1785arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *hdr, 1786 kmutex_t *hash_lock) 1787{ 1788 arc_state_t *old_state; 1789 int64_t refcnt; 1790 uint32_t datacnt; 1791 uint64_t from_delta, to_delta; 1792 arc_buf_contents_t buftype = arc_buf_type(hdr); 1793 1794 /* 1795 * We almost always have an L1 hdr here, since we call arc_hdr_realloc() 1796 * in arc_read() when bringing a buffer out of the L2ARC. However, the 1797 * L1 hdr doesn't always exist when we change state to arc_anon before 1798 * destroying a header, in which case reallocating to add the L1 hdr is 1799 * pointless. 1800 */ 1801 if (HDR_HAS_L1HDR(hdr)) { 1802 old_state = hdr->b_l1hdr.b_state; 1803 refcnt = refcount_count(&hdr->b_l1hdr.b_refcnt); 1804 datacnt = hdr->b_l1hdr.b_datacnt; 1805 } else { 1806 old_state = arc_l2c_only; 1807 refcnt = 0; 1808 datacnt = 0; 1809 } 1810 1811 ASSERT(MUTEX_HELD(hash_lock)); 1812 ASSERT3P(new_state, !=, old_state); 1813 ASSERT(refcnt == 0 || datacnt > 0); 1814 ASSERT(!GHOST_STATE(new_state) || datacnt == 0); 1815 ASSERT(old_state != arc_anon || datacnt <= 1); 1816 1817 from_delta = to_delta = datacnt * hdr->b_size; 1818 1819 /* 1820 * If this buffer is evictable, transfer it from the 1821 * old state list to the new state list. 1822 */ 1823 if (refcnt == 0) { 1824 if (old_state != arc_anon && old_state != arc_l2c_only) { 1825 uint64_t *size = &old_state->arcs_lsize[buftype]; 1826 1827 ASSERT(HDR_HAS_L1HDR(hdr)); 1828 multilist_remove(&old_state->arcs_list[buftype], hdr); 1829 1830 /* 1831 * If prefetching out of the ghost cache, 1832 * we will have a non-zero datacnt. 1833 */ 1834 if (GHOST_STATE(old_state) && datacnt == 0) { 1835 /* ghost elements have a ghost size */ 1836 ASSERT(hdr->b_l1hdr.b_buf == NULL); 1837 from_delta = hdr->b_size; 1838 } 1839 ASSERT3U(*size, >=, from_delta); 1840 atomic_add_64(size, -from_delta); 1841 } 1842 if (new_state != arc_anon && new_state != arc_l2c_only) { 1843 uint64_t *size = &new_state->arcs_lsize[buftype]; 1844 1845 /* 1846 * An L1 header always exists here, since if we're 1847 * moving to some L1-cached state (i.e. not l2c_only or 1848 * anonymous), we realloc the header to add an L1hdr 1849 * beforehand. 1850 */ 1851 ASSERT(HDR_HAS_L1HDR(hdr)); 1852 multilist_insert(&new_state->arcs_list[buftype], hdr); 1853 1854 /* ghost elements have a ghost size */ 1855 if (GHOST_STATE(new_state)) { 1856 ASSERT0(datacnt); 1857 ASSERT(hdr->b_l1hdr.b_buf == NULL); 1858 to_delta = hdr->b_size; 1859 } 1860 atomic_add_64(size, to_delta); 1861 } 1862 } 1863 1864 ASSERT(!BUF_EMPTY(hdr)); 1865 if (new_state == arc_anon && HDR_IN_HASH_TABLE(hdr)) 1866 buf_hash_remove(hdr); 1867 1868 /* adjust state sizes (ignore arc_l2c_only) */ 1869 1870 if (to_delta && new_state != arc_l2c_only) { 1871 ASSERT(HDR_HAS_L1HDR(hdr)); 1872 if (GHOST_STATE(new_state)) { 1873 ASSERT0(datacnt); 1874 1875 /* 1876 * We moving a header to a ghost state, we first 1877 * remove all arc buffers. Thus, we'll have a 1878 * datacnt of zero, and no arc buffer to use for 1879 * the reference. As a result, we use the arc 1880 * header pointer for the reference. 1881 */ 1882 (void) refcount_add_many(&new_state->arcs_size, 1883 hdr->b_size, hdr); 1884 } else { 1885 ASSERT3U(datacnt, !=, 0); 1886 1887 /* 1888 * Each individual buffer holds a unique reference, 1889 * thus we must remove each of these references one 1890 * at a time. 1891 */ 1892 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL; 1893 buf = buf->b_next) { 1894 (void) refcount_add_many(&new_state->arcs_size, 1895 hdr->b_size, buf); 1896 } 1897 } 1898 } 1899 1900 if (from_delta && old_state != arc_l2c_only) { 1901 ASSERT(HDR_HAS_L1HDR(hdr)); 1902 if (GHOST_STATE(old_state)) { 1903 /* 1904 * When moving a header off of a ghost state, 1905 * there's the possibility for datacnt to be 1906 * non-zero. This is because we first add the 1907 * arc buffer to the header prior to changing 1908 * the header's state. Since we used the header 1909 * for the reference when putting the header on 1910 * the ghost state, we must balance that and use 1911 * the header when removing off the ghost state 1912 * (even though datacnt is non zero). 1913 */ 1914 1915 IMPLY(datacnt == 0, new_state == arc_anon || 1916 new_state == arc_l2c_only); 1917 1918 (void) refcount_remove_many(&old_state->arcs_size, 1919 hdr->b_size, hdr); 1920 } else { 1921 ASSERT3P(datacnt, !=, 0); 1922 1923 /* 1924 * Each individual buffer holds a unique reference, 1925 * thus we must remove each of these references one 1926 * at a time. 1927 */ 1928 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL; 1929 buf = buf->b_next) { 1930 (void) refcount_remove_many( 1931 &old_state->arcs_size, hdr->b_size, buf); 1932 } 1933 } 1934 } 1935 1936 if (HDR_HAS_L1HDR(hdr)) 1937 hdr->b_l1hdr.b_state = new_state; 1938 1939 /* 1940 * L2 headers should never be on the L2 state list since they don't 1941 * have L1 headers allocated. 1942 */ 1943 ASSERT(multilist_is_empty(&arc_l2c_only->arcs_list[ARC_BUFC_DATA]) && 1944 multilist_is_empty(&arc_l2c_only->arcs_list[ARC_BUFC_METADATA])); 1945} 1946 1947void 1948arc_space_consume(uint64_t space, arc_space_type_t type) 1949{ 1950 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES); 1951 1952 switch (type) { 1953 case ARC_SPACE_DATA: 1954 ARCSTAT_INCR(arcstat_data_size, space); 1955 break; 1956 case ARC_SPACE_META: 1957 ARCSTAT_INCR(arcstat_metadata_size, space); 1958 break; 1959 case ARC_SPACE_OTHER: 1960 ARCSTAT_INCR(arcstat_other_size, space); 1961 break; 1962 case ARC_SPACE_HDRS: 1963 ARCSTAT_INCR(arcstat_hdr_size, space); 1964 break; 1965 case ARC_SPACE_L2HDRS: 1966 ARCSTAT_INCR(arcstat_l2_hdr_size, space); 1967 break; 1968 } 1969 1970 if (type != ARC_SPACE_DATA) 1971 ARCSTAT_INCR(arcstat_meta_used, space); 1972 1973 atomic_add_64(&arc_size, space); 1974} 1975 1976void 1977arc_space_return(uint64_t space, arc_space_type_t type) 1978{ 1979 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES); 1980 1981 switch (type) { 1982 case ARC_SPACE_DATA: 1983 ARCSTAT_INCR(arcstat_data_size, -space); 1984 break; 1985 case ARC_SPACE_META: 1986 ARCSTAT_INCR(arcstat_metadata_size, -space); 1987 break; 1988 case ARC_SPACE_OTHER: 1989 ARCSTAT_INCR(arcstat_other_size, -space); 1990 break; 1991 case ARC_SPACE_HDRS: 1992 ARCSTAT_INCR(arcstat_hdr_size, -space); 1993 break; 1994 case ARC_SPACE_L2HDRS: 1995 ARCSTAT_INCR(arcstat_l2_hdr_size, -space); 1996 break; 1997 } 1998 1999 if (type != ARC_SPACE_DATA) { 2000 ASSERT(arc_meta_used >= space); 2001 if (arc_meta_max < arc_meta_used) 2002 arc_meta_max = arc_meta_used; 2003 ARCSTAT_INCR(arcstat_meta_used, -space); 2004 } 2005 2006 ASSERT(arc_size >= space); 2007 atomic_add_64(&arc_size, -space); 2008} 2009 2010arc_buf_t * 2011arc_buf_alloc(spa_t *spa, int32_t size, void *tag, arc_buf_contents_t type) 2012{ 2013 arc_buf_hdr_t *hdr; 2014 arc_buf_t *buf; 2015 2016 ASSERT3U(size, >, 0); 2017 hdr = kmem_cache_alloc(hdr_full_cache, KM_PUSHPAGE); 2018 ASSERT(BUF_EMPTY(hdr)); 2019 ASSERT3P(hdr->b_freeze_cksum, ==, NULL); 2020 hdr->b_size = size; 2021 hdr->b_spa = spa_load_guid(spa); 2022 2023 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 2024 buf->b_hdr = hdr; 2025 buf->b_data = NULL; 2026 buf->b_efunc = NULL; 2027 buf->b_private = NULL; 2028 buf->b_next = NULL; 2029 2030 hdr->b_flags = arc_bufc_to_flags(type); 2031 hdr->b_flags |= ARC_FLAG_HAS_L1HDR; 2032 2033 hdr->b_l1hdr.b_buf = buf; 2034 hdr->b_l1hdr.b_state = arc_anon; 2035 hdr->b_l1hdr.b_arc_access = 0; 2036 hdr->b_l1hdr.b_datacnt = 1; 2037 hdr->b_l1hdr.b_tmp_cdata = NULL; 2038 2039 arc_get_data_buf(buf); 2040 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 2041 (void) refcount_add(&hdr->b_l1hdr.b_refcnt, tag); 2042 2043 return (buf); 2044} 2045 2046static char *arc_onloan_tag = "onloan"; 2047 2048/* 2049 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in 2050 * flight data by arc_tempreserve_space() until they are "returned". Loaned 2051 * buffers must be returned to the arc before they can be used by the DMU or 2052 * freed. 2053 */ 2054arc_buf_t * 2055arc_loan_buf(spa_t *spa, int size) 2056{ 2057 arc_buf_t *buf; 2058 2059 buf = arc_buf_alloc(spa, size, arc_onloan_tag, ARC_BUFC_DATA); 2060 2061 atomic_add_64(&arc_loaned_bytes, size); 2062 return (buf); 2063} 2064 2065/* 2066 * Return a loaned arc buffer to the arc. 2067 */ 2068void 2069arc_return_buf(arc_buf_t *buf, void *tag) 2070{ 2071 arc_buf_hdr_t *hdr = buf->b_hdr; 2072 2073 ASSERT(buf->b_data != NULL); 2074 ASSERT(HDR_HAS_L1HDR(hdr)); 2075 (void) refcount_add(&hdr->b_l1hdr.b_refcnt, tag); 2076 (void) refcount_remove(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag); 2077 2078 atomic_add_64(&arc_loaned_bytes, -hdr->b_size); 2079} 2080 2081/* Detach an arc_buf from a dbuf (tag) */ 2082void 2083arc_loan_inuse_buf(arc_buf_t *buf, void *tag) 2084{ 2085 arc_buf_hdr_t *hdr = buf->b_hdr; 2086 2087 ASSERT(buf->b_data != NULL); 2088 ASSERT(HDR_HAS_L1HDR(hdr)); 2089 (void) refcount_add(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag); 2090 (void) refcount_remove(&hdr->b_l1hdr.b_refcnt, tag); 2091 buf->b_efunc = NULL; 2092 buf->b_private = NULL; 2093 2094 atomic_add_64(&arc_loaned_bytes, hdr->b_size); 2095} 2096 2097static arc_buf_t * 2098arc_buf_clone(arc_buf_t *from) 2099{ 2100 arc_buf_t *buf; 2101 arc_buf_hdr_t *hdr = from->b_hdr; 2102 uint64_t size = hdr->b_size; 2103 2104 ASSERT(HDR_HAS_L1HDR(hdr)); 2105 ASSERT(hdr->b_l1hdr.b_state != arc_anon); 2106 2107 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 2108 buf->b_hdr = hdr; 2109 buf->b_data = NULL; 2110 buf->b_efunc = NULL; 2111 buf->b_private = NULL; 2112 buf->b_next = hdr->b_l1hdr.b_buf; 2113 hdr->b_l1hdr.b_buf = buf; 2114 arc_get_data_buf(buf); 2115 bcopy(from->b_data, buf->b_data, size); 2116 2117 /* 2118 * This buffer already exists in the arc so create a duplicate 2119 * copy for the caller. If the buffer is associated with user data 2120 * then track the size and number of duplicates. These stats will be 2121 * updated as duplicate buffers are created and destroyed. 2122 */ 2123 if (HDR_ISTYPE_DATA(hdr)) { 2124 ARCSTAT_BUMP(arcstat_duplicate_buffers); 2125 ARCSTAT_INCR(arcstat_duplicate_buffers_size, size); 2126 } 2127 hdr->b_l1hdr.b_datacnt += 1; 2128 return (buf); 2129} 2130 2131void 2132arc_buf_add_ref(arc_buf_t *buf, void* tag) 2133{ 2134 arc_buf_hdr_t *hdr; 2135 kmutex_t *hash_lock; 2136 2137 /* 2138 * Check to see if this buffer is evicted. Callers 2139 * must verify b_data != NULL to know if the add_ref 2140 * was successful. 2141 */ 2142 mutex_enter(&buf->b_evict_lock); 2143 if (buf->b_data == NULL) { 2144 mutex_exit(&buf->b_evict_lock); 2145 return; 2146 } 2147 hash_lock = HDR_LOCK(buf->b_hdr); 2148 mutex_enter(hash_lock); 2149 hdr = buf->b_hdr; 2150 ASSERT(HDR_HAS_L1HDR(hdr)); 2151 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 2152 mutex_exit(&buf->b_evict_lock); 2153 2154 ASSERT(hdr->b_l1hdr.b_state == arc_mru || 2155 hdr->b_l1hdr.b_state == arc_mfu); 2156 2157 add_reference(hdr, hash_lock, tag); 2158 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr); 2159 arc_access(hdr, hash_lock); 2160 mutex_exit(hash_lock); 2161 ARCSTAT_BUMP(arcstat_hits); 2162 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr), 2163 demand, prefetch, !HDR_ISTYPE_METADATA(hdr), 2164 data, metadata, hits); 2165} 2166 2167static void 2168arc_buf_free_on_write(void *data, size_t size, 2169 void (*free_func)(void *, size_t)) 2170{ 2171 l2arc_data_free_t *df; 2172 2173 df = kmem_alloc(sizeof (*df), KM_SLEEP); 2174 df->l2df_data = data; 2175 df->l2df_size = size; 2176 df->l2df_func = free_func; 2177 mutex_enter(&l2arc_free_on_write_mtx); 2178 list_insert_head(l2arc_free_on_write, df); 2179 mutex_exit(&l2arc_free_on_write_mtx); 2180} 2181 2182/* 2183 * Free the arc data buffer. If it is an l2arc write in progress, 2184 * the buffer is placed on l2arc_free_on_write to be freed later. 2185 */ 2186static void 2187arc_buf_data_free(arc_buf_t *buf, void (*free_func)(void *, size_t)) 2188{ 2189 arc_buf_hdr_t *hdr = buf->b_hdr; 2190 2191 if (HDR_L2_WRITING(hdr)) { 2192 arc_buf_free_on_write(buf->b_data, hdr->b_size, free_func); 2193 ARCSTAT_BUMP(arcstat_l2_free_on_write); 2194 } else { 2195 free_func(buf->b_data, hdr->b_size); 2196 } 2197} 2198 2199static void 2200arc_buf_l2_cdata_free(arc_buf_hdr_t *hdr) 2201{ 2202 ASSERT(HDR_HAS_L2HDR(hdr)); 2203 ASSERT(MUTEX_HELD(&hdr->b_l2hdr.b_dev->l2ad_mtx)); 2204 2205 /* 2206 * The b_tmp_cdata field is linked off of the b_l1hdr, so if 2207 * that doesn't exist, the header is in the arc_l2c_only state, 2208 * and there isn't anything to free (it's already been freed). 2209 */ 2210 if (!HDR_HAS_L1HDR(hdr)) 2211 return; 2212 2213 /* 2214 * The header isn't being written to the l2arc device, thus it 2215 * shouldn't have a b_tmp_cdata to free. 2216 */ 2217 if (!HDR_L2_WRITING(hdr)) { 2218 ASSERT3P(hdr->b_l1hdr.b_tmp_cdata, ==, NULL); 2219 return; 2220 } 2221 2222 /* 2223 * The header does not have compression enabled. This can be due 2224 * to the buffer not being compressible, or because we're 2225 * freeing the buffer before the second phase of 2226 * l2arc_write_buffer() has started (which does the compression 2227 * step). In either case, b_tmp_cdata does not point to a 2228 * separately compressed buffer, so there's nothing to free (it 2229 * points to the same buffer as the arc_buf_t's b_data field). 2230 */ 2231 if (hdr->b_l2hdr.b_compress == ZIO_COMPRESS_OFF) { 2232 hdr->b_l1hdr.b_tmp_cdata = NULL; 2233 return; 2234 } 2235 2236 /* 2237 * There's nothing to free since the buffer was all zero's and 2238 * compressed to a zero length buffer. 2239 */ 2240 if (hdr->b_l2hdr.b_compress == ZIO_COMPRESS_EMPTY) { 2241 ASSERT3P(hdr->b_l1hdr.b_tmp_cdata, ==, NULL); 2242 return; 2243 } 2244 2245 ASSERT(L2ARC_IS_VALID_COMPRESS(hdr->b_l2hdr.b_compress)); 2246 2247 arc_buf_free_on_write(hdr->b_l1hdr.b_tmp_cdata, 2248 hdr->b_size, zio_data_buf_free); 2249 2250 ARCSTAT_BUMP(arcstat_l2_cdata_free_on_write); 2251 hdr->b_l1hdr.b_tmp_cdata = NULL; 2252} 2253 2254/* 2255 * Free up buf->b_data and if 'remove' is set, then pull the 2256 * arc_buf_t off of the the arc_buf_hdr_t's list and free it. 2257 */ 2258static void 2259arc_buf_destroy(arc_buf_t *buf, boolean_t remove) 2260{ 2261 arc_buf_t **bufp; 2262 2263 /* free up data associated with the buf */ 2264 if (buf->b_data != NULL) { 2265 arc_state_t *state = buf->b_hdr->b_l1hdr.b_state; 2266 uint64_t size = buf->b_hdr->b_size; 2267 arc_buf_contents_t type = arc_buf_type(buf->b_hdr); 2268 2269 arc_cksum_verify(buf); 2270#ifdef illumos 2271 arc_buf_unwatch(buf); 2272#endif /* illumos */ 2273 2274 if (type == ARC_BUFC_METADATA) { 2275 arc_buf_data_free(buf, zio_buf_free); 2276 arc_space_return(size, ARC_SPACE_META); 2277 } else { 2278 ASSERT(type == ARC_BUFC_DATA); 2279 arc_buf_data_free(buf, zio_data_buf_free); 2280 arc_space_return(size, ARC_SPACE_DATA); 2281 } 2282 2283 /* protected by hash lock, if in the hash table */ 2284 if (multilist_link_active(&buf->b_hdr->b_l1hdr.b_arc_node)) { 2285 uint64_t *cnt = &state->arcs_lsize[type]; 2286 2287 ASSERT(refcount_is_zero( 2288 &buf->b_hdr->b_l1hdr.b_refcnt)); 2289 ASSERT(state != arc_anon && state != arc_l2c_only); 2290 2291 ASSERT3U(*cnt, >=, size); 2292 atomic_add_64(cnt, -size); 2293 } 2294 2295 (void) refcount_remove_many(&state->arcs_size, size, buf); 2296 buf->b_data = NULL; 2297 2298 /* 2299 * If we're destroying a duplicate buffer make sure 2300 * that the appropriate statistics are updated. 2301 */ 2302 if (buf->b_hdr->b_l1hdr.b_datacnt > 1 && 2303 HDR_ISTYPE_DATA(buf->b_hdr)) { 2304 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers); 2305 ARCSTAT_INCR(arcstat_duplicate_buffers_size, -size); 2306 } 2307 ASSERT(buf->b_hdr->b_l1hdr.b_datacnt > 0); 2308 buf->b_hdr->b_l1hdr.b_datacnt -= 1; 2309 } 2310 2311 /* only remove the buf if requested */ 2312 if (!remove) 2313 return; 2314 2315 /* remove the buf from the hdr list */ 2316 for (bufp = &buf->b_hdr->b_l1hdr.b_buf; *bufp != buf; 2317 bufp = &(*bufp)->b_next) 2318 continue; 2319 *bufp = buf->b_next; 2320 buf->b_next = NULL; 2321 2322 ASSERT(buf->b_efunc == NULL); 2323 2324 /* clean up the buf */ 2325 buf->b_hdr = NULL; 2326 kmem_cache_free(buf_cache, buf); 2327} 2328 2329static void 2330arc_hdr_l2hdr_destroy(arc_buf_hdr_t *hdr) 2331{ 2332 l2arc_buf_hdr_t *l2hdr = &hdr->b_l2hdr; 2333 l2arc_dev_t *dev = l2hdr->b_dev; 2334 2335 ASSERT(MUTEX_HELD(&dev->l2ad_mtx)); 2336 ASSERT(HDR_HAS_L2HDR(hdr)); 2337 2338 list_remove(&dev->l2ad_buflist, hdr); 2339 2340 /* 2341 * We don't want to leak the b_tmp_cdata buffer that was 2342 * allocated in l2arc_write_buffers() 2343 */ 2344 arc_buf_l2_cdata_free(hdr); 2345 2346 /* 2347 * If the l2hdr's b_daddr is equal to L2ARC_ADDR_UNSET, then 2348 * this header is being processed by l2arc_write_buffers() (i.e. 2349 * it's in the first stage of l2arc_write_buffers()). 2350 * Re-affirming that truth here, just to serve as a reminder. If 2351 * b_daddr does not equal L2ARC_ADDR_UNSET, then the header may or 2352 * may not have its HDR_L2_WRITING flag set. (the write may have 2353 * completed, in which case HDR_L2_WRITING will be false and the 2354 * b_daddr field will point to the address of the buffer on disk). 2355 */ 2356 IMPLY(l2hdr->b_daddr == L2ARC_ADDR_UNSET, HDR_L2_WRITING(hdr)); 2357 2358 /* 2359 * If b_daddr is equal to L2ARC_ADDR_UNSET, we're racing with 2360 * l2arc_write_buffers(). Since we've just removed this header 2361 * from the l2arc buffer list, this header will never reach the 2362 * second stage of l2arc_write_buffers(), which increments the 2363 * accounting stats for this header. Thus, we must be careful 2364 * not to decrement them for this header either. 2365 */ 2366 if (l2hdr->b_daddr != L2ARC_ADDR_UNSET) { 2367 ARCSTAT_INCR(arcstat_l2_asize, -l2hdr->b_asize); 2368 ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size); 2369 2370 vdev_space_update(dev->l2ad_vdev, 2371 -l2hdr->b_asize, 0, 0); 2372 2373 (void) refcount_remove_many(&dev->l2ad_alloc, 2374 l2hdr->b_asize, hdr); 2375 } 2376 2377 hdr->b_flags &= ~ARC_FLAG_HAS_L2HDR; 2378} 2379 2380static void 2381arc_hdr_destroy(arc_buf_hdr_t *hdr) 2382{ 2383 if (HDR_HAS_L1HDR(hdr)) { 2384 ASSERT(hdr->b_l1hdr.b_buf == NULL || 2385 hdr->b_l1hdr.b_datacnt > 0); 2386 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 2387 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon); 2388 } 2389 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 2390 ASSERT(!HDR_IN_HASH_TABLE(hdr)); 2391 2392 if (HDR_HAS_L2HDR(hdr)) { 2393 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev; 2394 boolean_t buflist_held = MUTEX_HELD(&dev->l2ad_mtx); 2395 2396 if (!buflist_held) 2397 mutex_enter(&dev->l2ad_mtx); 2398 2399 /* 2400 * Even though we checked this conditional above, we 2401 * need to check this again now that we have the 2402 * l2ad_mtx. This is because we could be racing with 2403 * another thread calling l2arc_evict() which might have 2404 * destroyed this header's L2 portion as we were waiting 2405 * to acquire the l2ad_mtx. If that happens, we don't 2406 * want to re-destroy the header's L2 portion. 2407 */ 2408 if (HDR_HAS_L2HDR(hdr)) { 2409 if (hdr->b_l2hdr.b_daddr != L2ARC_ADDR_UNSET) 2410 trim_map_free(dev->l2ad_vdev, 2411 hdr->b_l2hdr.b_daddr, 2412 hdr->b_l2hdr.b_asize, 0); 2413 arc_hdr_l2hdr_destroy(hdr); 2414 } 2415 2416 if (!buflist_held) 2417 mutex_exit(&dev->l2ad_mtx); 2418 } 2419 2420 if (!BUF_EMPTY(hdr)) 2421 buf_discard_identity(hdr); 2422 2423 if (hdr->b_freeze_cksum != NULL) { 2424 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 2425 hdr->b_freeze_cksum = NULL; 2426 } 2427 2428 if (HDR_HAS_L1HDR(hdr)) { 2429 while (hdr->b_l1hdr.b_buf) { 2430 arc_buf_t *buf = hdr->b_l1hdr.b_buf; 2431 2432 if (buf->b_efunc != NULL) { 2433 mutex_enter(&arc_user_evicts_lock); 2434 mutex_enter(&buf->b_evict_lock); 2435 ASSERT(buf->b_hdr != NULL); 2436 arc_buf_destroy(hdr->b_l1hdr.b_buf, FALSE); 2437 hdr->b_l1hdr.b_buf = buf->b_next; 2438 buf->b_hdr = &arc_eviction_hdr; 2439 buf->b_next = arc_eviction_list; 2440 arc_eviction_list = buf; 2441 mutex_exit(&buf->b_evict_lock); 2442 cv_signal(&arc_user_evicts_cv); 2443 mutex_exit(&arc_user_evicts_lock); 2444 } else { 2445 arc_buf_destroy(hdr->b_l1hdr.b_buf, TRUE); 2446 } 2447 } 2448#ifdef ZFS_DEBUG 2449 if (hdr->b_l1hdr.b_thawed != NULL) { 2450 kmem_free(hdr->b_l1hdr.b_thawed, 1); 2451 hdr->b_l1hdr.b_thawed = NULL; 2452 } 2453#endif 2454 } 2455 2456 ASSERT3P(hdr->b_hash_next, ==, NULL); 2457 if (HDR_HAS_L1HDR(hdr)) { 2458 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 2459 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL); 2460 kmem_cache_free(hdr_full_cache, hdr); 2461 } else { 2462 kmem_cache_free(hdr_l2only_cache, hdr); 2463 } 2464} 2465 2466void 2467arc_buf_free(arc_buf_t *buf, void *tag) 2468{ 2469 arc_buf_hdr_t *hdr = buf->b_hdr; 2470 int hashed = hdr->b_l1hdr.b_state != arc_anon; 2471 2472 ASSERT(buf->b_efunc == NULL); 2473 ASSERT(buf->b_data != NULL); 2474 2475 if (hashed) { 2476 kmutex_t *hash_lock = HDR_LOCK(hdr); 2477 2478 mutex_enter(hash_lock); 2479 hdr = buf->b_hdr; 2480 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 2481 2482 (void) remove_reference(hdr, hash_lock, tag); 2483 if (hdr->b_l1hdr.b_datacnt > 1) { 2484 arc_buf_destroy(buf, TRUE); 2485 } else { 2486 ASSERT(buf == hdr->b_l1hdr.b_buf); 2487 ASSERT(buf->b_efunc == NULL); 2488 hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE; 2489 } 2490 mutex_exit(hash_lock); 2491 } else if (HDR_IO_IN_PROGRESS(hdr)) { 2492 int destroy_hdr; 2493 /* 2494 * We are in the middle of an async write. Don't destroy 2495 * this buffer unless the write completes before we finish 2496 * decrementing the reference count. 2497 */ 2498 mutex_enter(&arc_user_evicts_lock); 2499 (void) remove_reference(hdr, NULL, tag); 2500 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 2501 destroy_hdr = !HDR_IO_IN_PROGRESS(hdr); 2502 mutex_exit(&arc_user_evicts_lock); 2503 if (destroy_hdr) 2504 arc_hdr_destroy(hdr); 2505 } else { 2506 if (remove_reference(hdr, NULL, tag) > 0) 2507 arc_buf_destroy(buf, TRUE); 2508 else 2509 arc_hdr_destroy(hdr); 2510 } 2511} 2512 2513boolean_t 2514arc_buf_remove_ref(arc_buf_t *buf, void* tag) 2515{ 2516 arc_buf_hdr_t *hdr = buf->b_hdr; 2517 kmutex_t *hash_lock = HDR_LOCK(hdr); 2518 boolean_t no_callback = (buf->b_efunc == NULL); 2519 2520 if (hdr->b_l1hdr.b_state == arc_anon) { 2521 ASSERT(hdr->b_l1hdr.b_datacnt == 1); 2522 arc_buf_free(buf, tag); 2523 return (no_callback); 2524 } 2525 2526 mutex_enter(hash_lock); 2527 hdr = buf->b_hdr; 2528 ASSERT(hdr->b_l1hdr.b_datacnt > 0); 2529 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 2530 ASSERT(hdr->b_l1hdr.b_state != arc_anon); 2531 ASSERT(buf->b_data != NULL); 2532 2533 (void) remove_reference(hdr, hash_lock, tag); 2534 if (hdr->b_l1hdr.b_datacnt > 1) { 2535 if (no_callback) 2536 arc_buf_destroy(buf, TRUE); 2537 } else if (no_callback) { 2538 ASSERT(hdr->b_l1hdr.b_buf == buf && buf->b_next == NULL); 2539 ASSERT(buf->b_efunc == NULL); 2540 hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE; 2541 } 2542 ASSERT(no_callback || hdr->b_l1hdr.b_datacnt > 1 || 2543 refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 2544 mutex_exit(hash_lock); 2545 return (no_callback); 2546} 2547 2548int32_t 2549arc_buf_size(arc_buf_t *buf) 2550{ 2551 return (buf->b_hdr->b_size); 2552} 2553 2554/* 2555 * Called from the DMU to determine if the current buffer should be 2556 * evicted. In order to ensure proper locking, the eviction must be initiated 2557 * from the DMU. Return true if the buffer is associated with user data and 2558 * duplicate buffers still exist. 2559 */ 2560boolean_t 2561arc_buf_eviction_needed(arc_buf_t *buf) 2562{ 2563 arc_buf_hdr_t *hdr; 2564 boolean_t evict_needed = B_FALSE; 2565 2566 if (zfs_disable_dup_eviction) 2567 return (B_FALSE); 2568 2569 mutex_enter(&buf->b_evict_lock); 2570 hdr = buf->b_hdr; 2571 if (hdr == NULL) { 2572 /* 2573 * We are in arc_do_user_evicts(); let that function 2574 * perform the eviction. 2575 */ 2576 ASSERT(buf->b_data == NULL); 2577 mutex_exit(&buf->b_evict_lock); 2578 return (B_FALSE); 2579 } else if (buf->b_data == NULL) { 2580 /* 2581 * We have already been added to the arc eviction list; 2582 * recommend eviction. 2583 */ 2584 ASSERT3P(hdr, ==, &arc_eviction_hdr); 2585 mutex_exit(&buf->b_evict_lock); 2586 return (B_TRUE); 2587 } 2588 2589 if (hdr->b_l1hdr.b_datacnt > 1 && HDR_ISTYPE_DATA(hdr)) 2590 evict_needed = B_TRUE; 2591 2592 mutex_exit(&buf->b_evict_lock); 2593 return (evict_needed); 2594} 2595 2596/* 2597 * Evict the arc_buf_hdr that is provided as a parameter. The resultant 2598 * state of the header is dependent on it's state prior to entering this 2599 * function. The following transitions are possible: 2600 * 2601 * - arc_mru -> arc_mru_ghost 2602 * - arc_mfu -> arc_mfu_ghost 2603 * - arc_mru_ghost -> arc_l2c_only 2604 * - arc_mru_ghost -> deleted 2605 * - arc_mfu_ghost -> arc_l2c_only 2606 * - arc_mfu_ghost -> deleted 2607 */ 2608static int64_t 2609arc_evict_hdr(arc_buf_hdr_t *hdr, kmutex_t *hash_lock) 2610{ 2611 arc_state_t *evicted_state, *state; 2612 int64_t bytes_evicted = 0; 2613 2614 ASSERT(MUTEX_HELD(hash_lock)); 2615 ASSERT(HDR_HAS_L1HDR(hdr)); 2616 2617 state = hdr->b_l1hdr.b_state; 2618 if (GHOST_STATE(state)) { 2619 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 2620 ASSERT(hdr->b_l1hdr.b_buf == NULL); 2621 2622 /* 2623 * l2arc_write_buffers() relies on a header's L1 portion 2624 * (i.e. it's b_tmp_cdata field) during it's write phase. 2625 * Thus, we cannot push a header onto the arc_l2c_only 2626 * state (removing it's L1 piece) until the header is 2627 * done being written to the l2arc. 2628 */ 2629 if (HDR_HAS_L2HDR(hdr) && HDR_L2_WRITING(hdr)) { 2630 ARCSTAT_BUMP(arcstat_evict_l2_skip); 2631 return (bytes_evicted); 2632 } 2633 2634 ARCSTAT_BUMP(arcstat_deleted); 2635 bytes_evicted += hdr->b_size; 2636 2637 DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, hdr); 2638 2639 if (HDR_HAS_L2HDR(hdr)) { 2640 /* 2641 * This buffer is cached on the 2nd Level ARC; 2642 * don't destroy the header. 2643 */ 2644 arc_change_state(arc_l2c_only, hdr, hash_lock); 2645 /* 2646 * dropping from L1+L2 cached to L2-only, 2647 * realloc to remove the L1 header. 2648 */ 2649 hdr = arc_hdr_realloc(hdr, hdr_full_cache, 2650 hdr_l2only_cache); 2651 } else { 2652 arc_change_state(arc_anon, hdr, hash_lock); 2653 arc_hdr_destroy(hdr); 2654 } 2655 return (bytes_evicted); 2656 } 2657 2658 ASSERT(state == arc_mru || state == arc_mfu); 2659 evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost; 2660 2661 /* prefetch buffers have a minimum lifespan */ 2662 if (HDR_IO_IN_PROGRESS(hdr) || 2663 ((hdr->b_flags & (ARC_FLAG_PREFETCH | ARC_FLAG_INDIRECT)) && 2664 ddi_get_lbolt() - hdr->b_l1hdr.b_arc_access < 2665 arc_min_prefetch_lifespan)) { 2666 ARCSTAT_BUMP(arcstat_evict_skip); 2667 return (bytes_evicted); 2668 } 2669 2670 ASSERT0(refcount_count(&hdr->b_l1hdr.b_refcnt)); 2671 ASSERT3U(hdr->b_l1hdr.b_datacnt, >, 0); 2672 while (hdr->b_l1hdr.b_buf) { 2673 arc_buf_t *buf = hdr->b_l1hdr.b_buf; 2674 if (!mutex_tryenter(&buf->b_evict_lock)) { 2675 ARCSTAT_BUMP(arcstat_mutex_miss); 2676 break; 2677 } 2678 if (buf->b_data != NULL) 2679 bytes_evicted += hdr->b_size; 2680 if (buf->b_efunc != NULL) { 2681 mutex_enter(&arc_user_evicts_lock); 2682 arc_buf_destroy(buf, FALSE); 2683 hdr->b_l1hdr.b_buf = buf->b_next; 2684 buf->b_hdr = &arc_eviction_hdr; 2685 buf->b_next = arc_eviction_list; 2686 arc_eviction_list = buf; 2687 cv_signal(&arc_user_evicts_cv); 2688 mutex_exit(&arc_user_evicts_lock); 2689 mutex_exit(&buf->b_evict_lock); 2690 } else { 2691 mutex_exit(&buf->b_evict_lock); 2692 arc_buf_destroy(buf, TRUE); 2693 } 2694 } 2695 2696 if (HDR_HAS_L2HDR(hdr)) { 2697 ARCSTAT_INCR(arcstat_evict_l2_cached, hdr->b_size); 2698 } else { 2699 if (l2arc_write_eligible(hdr->b_spa, hdr)) 2700 ARCSTAT_INCR(arcstat_evict_l2_eligible, hdr->b_size); 2701 else 2702 ARCSTAT_INCR(arcstat_evict_l2_ineligible, hdr->b_size); 2703 } 2704 2705 if (hdr->b_l1hdr.b_datacnt == 0) { 2706 arc_change_state(evicted_state, hdr, hash_lock); 2707 ASSERT(HDR_IN_HASH_TABLE(hdr)); 2708 hdr->b_flags |= ARC_FLAG_IN_HASH_TABLE; 2709 hdr->b_flags &= ~ARC_FLAG_BUF_AVAILABLE; 2710 DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, hdr); 2711 } 2712 2713 return (bytes_evicted); 2714} 2715 2716static uint64_t 2717arc_evict_state_impl(multilist_t *ml, int idx, arc_buf_hdr_t *marker, 2718 uint64_t spa, int64_t bytes) 2719{ 2720 multilist_sublist_t *mls; 2721 uint64_t bytes_evicted = 0; 2722 arc_buf_hdr_t *hdr; 2723 kmutex_t *hash_lock; 2724 int evict_count = 0; 2725 2726 ASSERT3P(marker, !=, NULL); 2727 IMPLY(bytes < 0, bytes == ARC_EVICT_ALL); 2728 2729 mls = multilist_sublist_lock(ml, idx); 2730 2731 for (hdr = multilist_sublist_prev(mls, marker); hdr != NULL; 2732 hdr = multilist_sublist_prev(mls, marker)) { 2733 if ((bytes != ARC_EVICT_ALL && bytes_evicted >= bytes) || 2734 (evict_count >= zfs_arc_evict_batch_limit)) 2735 break; 2736 2737 /* 2738 * To keep our iteration location, move the marker 2739 * forward. Since we're not holding hdr's hash lock, we 2740 * must be very careful and not remove 'hdr' from the 2741 * sublist. Otherwise, other consumers might mistake the 2742 * 'hdr' as not being on a sublist when they call the 2743 * multilist_link_active() function (they all rely on 2744 * the hash lock protecting concurrent insertions and 2745 * removals). multilist_sublist_move_forward() was 2746 * specifically implemented to ensure this is the case 2747 * (only 'marker' will be removed and re-inserted). 2748 */ 2749 multilist_sublist_move_forward(mls, marker); 2750 2751 /* 2752 * The only case where the b_spa field should ever be 2753 * zero, is the marker headers inserted by 2754 * arc_evict_state(). It's possible for multiple threads 2755 * to be calling arc_evict_state() concurrently (e.g. 2756 * dsl_pool_close() and zio_inject_fault()), so we must 2757 * skip any markers we see from these other threads. 2758 */ 2759 if (hdr->b_spa == 0) 2760 continue; 2761 2762 /* we're only interested in evicting buffers of a certain spa */ 2763 if (spa != 0 && hdr->b_spa != spa) { 2764 ARCSTAT_BUMP(arcstat_evict_skip); 2765 continue; 2766 } 2767 2768 hash_lock = HDR_LOCK(hdr); 2769 2770 /* 2771 * We aren't calling this function from any code path 2772 * that would already be holding a hash lock, so we're 2773 * asserting on this assumption to be defensive in case 2774 * this ever changes. Without this check, it would be 2775 * possible to incorrectly increment arcstat_mutex_miss 2776 * below (e.g. if the code changed such that we called 2777 * this function with a hash lock held). 2778 */ 2779 ASSERT(!MUTEX_HELD(hash_lock)); 2780 2781 if (mutex_tryenter(hash_lock)) { 2782 uint64_t evicted = arc_evict_hdr(hdr, hash_lock); 2783 mutex_exit(hash_lock); 2784 2785 bytes_evicted += evicted; 2786 2787 /* 2788 * If evicted is zero, arc_evict_hdr() must have 2789 * decided to skip this header, don't increment 2790 * evict_count in this case. 2791 */ 2792 if (evicted != 0) 2793 evict_count++; 2794 2795 /* 2796 * If arc_size isn't overflowing, signal any 2797 * threads that might happen to be waiting. 2798 * 2799 * For each header evicted, we wake up a single 2800 * thread. If we used cv_broadcast, we could 2801 * wake up "too many" threads causing arc_size 2802 * to significantly overflow arc_c; since 2803 * arc_get_data_buf() doesn't check for overflow 2804 * when it's woken up (it doesn't because it's 2805 * possible for the ARC to be overflowing while 2806 * full of un-evictable buffers, and the 2807 * function should proceed in this case). 2808 * 2809 * If threads are left sleeping, due to not 2810 * using cv_broadcast, they will be woken up 2811 * just before arc_reclaim_thread() sleeps. 2812 */ 2813 mutex_enter(&arc_reclaim_lock); 2814 if (!arc_is_overflowing()) 2815 cv_signal(&arc_reclaim_waiters_cv); 2816 mutex_exit(&arc_reclaim_lock); 2817 } else { 2818 ARCSTAT_BUMP(arcstat_mutex_miss); 2819 } 2820 } 2821 2822 multilist_sublist_unlock(mls); 2823 2824 return (bytes_evicted); 2825} 2826 2827/* 2828 * Evict buffers from the given arc state, until we've removed the 2829 * specified number of bytes. Move the removed buffers to the 2830 * appropriate evict state. 2831 * 2832 * This function makes a "best effort". It skips over any buffers 2833 * it can't get a hash_lock on, and so, may not catch all candidates. 2834 * It may also return without evicting as much space as requested. 2835 * 2836 * If bytes is specified using the special value ARC_EVICT_ALL, this 2837 * will evict all available (i.e. unlocked and evictable) buffers from 2838 * the given arc state; which is used by arc_flush(). 2839 */ 2840static uint64_t 2841arc_evict_state(arc_state_t *state, uint64_t spa, int64_t bytes, 2842 arc_buf_contents_t type) 2843{ 2844 uint64_t total_evicted = 0; 2845 multilist_t *ml = &state->arcs_list[type]; 2846 int num_sublists; 2847 arc_buf_hdr_t **markers; 2848 2849 IMPLY(bytes < 0, bytes == ARC_EVICT_ALL); 2850 2851 num_sublists = multilist_get_num_sublists(ml); 2852 2853 /* 2854 * If we've tried to evict from each sublist, made some 2855 * progress, but still have not hit the target number of bytes 2856 * to evict, we want to keep trying. The markers allow us to 2857 * pick up where we left off for each individual sublist, rather 2858 * than starting from the tail each time. 2859 */ 2860 markers = kmem_zalloc(sizeof (*markers) * num_sublists, KM_SLEEP); 2861 for (int i = 0; i < num_sublists; i++) { 2862 markers[i] = kmem_cache_alloc(hdr_full_cache, KM_SLEEP); 2863 2864 /* 2865 * A b_spa of 0 is used to indicate that this header is 2866 * a marker. This fact is used in arc_adjust_type() and 2867 * arc_evict_state_impl(). 2868 */ 2869 markers[i]->b_spa = 0; 2870 2871 multilist_sublist_t *mls = multilist_sublist_lock(ml, i); 2872 multilist_sublist_insert_tail(mls, markers[i]); 2873 multilist_sublist_unlock(mls); 2874 } 2875 2876 /* 2877 * While we haven't hit our target number of bytes to evict, or 2878 * we're evicting all available buffers. 2879 */ 2880 while (total_evicted < bytes || bytes == ARC_EVICT_ALL) { 2881 /* 2882 * Start eviction using a randomly selected sublist, 2883 * this is to try and evenly balance eviction across all 2884 * sublists. Always starting at the same sublist 2885 * (e.g. index 0) would cause evictions to favor certain 2886 * sublists over others. 2887 */ 2888 int sublist_idx = multilist_get_random_index(ml); 2889 uint64_t scan_evicted = 0; 2890 2891 for (int i = 0; i < num_sublists; i++) { 2892 uint64_t bytes_remaining; 2893 uint64_t bytes_evicted; 2894 2895 if (bytes == ARC_EVICT_ALL) 2896 bytes_remaining = ARC_EVICT_ALL; 2897 else if (total_evicted < bytes) 2898 bytes_remaining = bytes - total_evicted; 2899 else 2900 break; 2901 2902 bytes_evicted = arc_evict_state_impl(ml, sublist_idx, 2903 markers[sublist_idx], spa, bytes_remaining); 2904 2905 scan_evicted += bytes_evicted; 2906 total_evicted += bytes_evicted; 2907 2908 /* we've reached the end, wrap to the beginning */ 2909 if (++sublist_idx >= num_sublists) 2910 sublist_idx = 0; 2911 } 2912 2913 /* 2914 * If we didn't evict anything during this scan, we have 2915 * no reason to believe we'll evict more during another 2916 * scan, so break the loop. 2917 */ 2918 if (scan_evicted == 0) { 2919 /* This isn't possible, let's make that obvious */ 2920 ASSERT3S(bytes, !=, 0); 2921 2922 /* 2923 * When bytes is ARC_EVICT_ALL, the only way to 2924 * break the loop is when scan_evicted is zero. 2925 * In that case, we actually have evicted enough, 2926 * so we don't want to increment the kstat. 2927 */ 2928 if (bytes != ARC_EVICT_ALL) { 2929 ASSERT3S(total_evicted, <, bytes); 2930 ARCSTAT_BUMP(arcstat_evict_not_enough); 2931 } 2932 2933 break; 2934 } 2935 } 2936 2937 for (int i = 0; i < num_sublists; i++) { 2938 multilist_sublist_t *mls = multilist_sublist_lock(ml, i); 2939 multilist_sublist_remove(mls, markers[i]); 2940 multilist_sublist_unlock(mls); 2941 2942 kmem_cache_free(hdr_full_cache, markers[i]); 2943 } 2944 kmem_free(markers, sizeof (*markers) * num_sublists); 2945 2946 return (total_evicted); 2947} 2948 2949/* 2950 * Flush all "evictable" data of the given type from the arc state 2951 * specified. This will not evict any "active" buffers (i.e. referenced). 2952 * 2953 * When 'retry' is set to FALSE, the function will make a single pass 2954 * over the state and evict any buffers that it can. Since it doesn't 2955 * continually retry the eviction, it might end up leaving some buffers 2956 * in the ARC due to lock misses. 2957 * 2958 * When 'retry' is set to TRUE, the function will continually retry the 2959 * eviction until *all* evictable buffers have been removed from the 2960 * state. As a result, if concurrent insertions into the state are 2961 * allowed (e.g. if the ARC isn't shutting down), this function might 2962 * wind up in an infinite loop, continually trying to evict buffers. 2963 */ 2964static uint64_t 2965arc_flush_state(arc_state_t *state, uint64_t spa, arc_buf_contents_t type, 2966 boolean_t retry) 2967{ 2968 uint64_t evicted = 0; 2969 2970 while (state->arcs_lsize[type] != 0) { 2971 evicted += arc_evict_state(state, spa, ARC_EVICT_ALL, type); 2972 2973 if (!retry) 2974 break; 2975 } 2976 2977 return (evicted); 2978} 2979 2980/* 2981 * Evict the specified number of bytes from the state specified, 2982 * restricting eviction to the spa and type given. This function 2983 * prevents us from trying to evict more from a state's list than 2984 * is "evictable", and to skip evicting altogether when passed a 2985 * negative value for "bytes". In contrast, arc_evict_state() will 2986 * evict everything it can, when passed a negative value for "bytes". 2987 */ 2988static uint64_t 2989arc_adjust_impl(arc_state_t *state, uint64_t spa, int64_t bytes, 2990 arc_buf_contents_t type) 2991{ 2992 int64_t delta; 2993 2994 if (bytes > 0 && state->arcs_lsize[type] > 0) { 2995 delta = MIN(state->arcs_lsize[type], bytes); 2996 return (arc_evict_state(state, spa, delta, type)); 2997 } 2998 2999 return (0); 3000} 3001 3002/* 3003 * Evict metadata buffers from the cache, such that arc_meta_used is 3004 * capped by the arc_meta_limit tunable. 3005 */ 3006static uint64_t 3007arc_adjust_meta(void) 3008{ 3009 uint64_t total_evicted = 0; 3010 int64_t target; 3011 3012 /* 3013 * If we're over the meta limit, we want to evict enough 3014 * metadata to get back under the meta limit. We don't want to 3015 * evict so much that we drop the MRU below arc_p, though. If 3016 * we're over the meta limit more than we're over arc_p, we 3017 * evict some from the MRU here, and some from the MFU below. 3018 */ 3019 target = MIN((int64_t)(arc_meta_used - arc_meta_limit), 3020 (int64_t)(refcount_count(&arc_anon->arcs_size) + 3021 refcount_count(&arc_mru->arcs_size) - arc_p)); 3022 3023 total_evicted += arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA); 3024 3025 /* 3026 * Similar to the above, we want to evict enough bytes to get us 3027 * below the meta limit, but not so much as to drop us below the 3028 * space alloted to the MFU (which is defined as arc_c - arc_p). 3029 */ 3030 target = MIN((int64_t)(arc_meta_used - arc_meta_limit), 3031 (int64_t)(refcount_count(&arc_mfu->arcs_size) - (arc_c - arc_p))); 3032 3033 total_evicted += arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA); 3034 3035 return (total_evicted); 3036} 3037 3038/* 3039 * Return the type of the oldest buffer in the given arc state 3040 * 3041 * This function will select a random sublist of type ARC_BUFC_DATA and 3042 * a random sublist of type ARC_BUFC_METADATA. The tail of each sublist 3043 * is compared, and the type which contains the "older" buffer will be 3044 * returned. 3045 */ 3046static arc_buf_contents_t 3047arc_adjust_type(arc_state_t *state) 3048{ 3049 multilist_t *data_ml = &state->arcs_list[ARC_BUFC_DATA]; 3050 multilist_t *meta_ml = &state->arcs_list[ARC_BUFC_METADATA]; 3051 int data_idx = multilist_get_random_index(data_ml); 3052 int meta_idx = multilist_get_random_index(meta_ml); 3053 multilist_sublist_t *data_mls; 3054 multilist_sublist_t *meta_mls; 3055 arc_buf_contents_t type; 3056 arc_buf_hdr_t *data_hdr; 3057 arc_buf_hdr_t *meta_hdr; 3058 3059 /* 3060 * We keep the sublist lock until we're finished, to prevent 3061 * the headers from being destroyed via arc_evict_state(). 3062 */ 3063 data_mls = multilist_sublist_lock(data_ml, data_idx); 3064 meta_mls = multilist_sublist_lock(meta_ml, meta_idx); 3065 3066 /* 3067 * These two loops are to ensure we skip any markers that 3068 * might be at the tail of the lists due to arc_evict_state(). 3069 */ 3070 3071 for (data_hdr = multilist_sublist_tail(data_mls); data_hdr != NULL; 3072 data_hdr = multilist_sublist_prev(data_mls, data_hdr)) { 3073 if (data_hdr->b_spa != 0) 3074 break; 3075 } 3076 3077 for (meta_hdr = multilist_sublist_tail(meta_mls); meta_hdr != NULL; 3078 meta_hdr = multilist_sublist_prev(meta_mls, meta_hdr)) { 3079 if (meta_hdr->b_spa != 0) 3080 break; 3081 } 3082 3083 if (data_hdr == NULL && meta_hdr == NULL) { 3084 type = ARC_BUFC_DATA; 3085 } else if (data_hdr == NULL) { 3086 ASSERT3P(meta_hdr, !=, NULL); 3087 type = ARC_BUFC_METADATA; 3088 } else if (meta_hdr == NULL) { 3089 ASSERT3P(data_hdr, !=, NULL); 3090 type = ARC_BUFC_DATA; 3091 } else { 3092 ASSERT3P(data_hdr, !=, NULL); 3093 ASSERT3P(meta_hdr, !=, NULL); 3094 3095 /* The headers can't be on the sublist without an L1 header */ 3096 ASSERT(HDR_HAS_L1HDR(data_hdr)); 3097 ASSERT(HDR_HAS_L1HDR(meta_hdr)); 3098 3099 if (data_hdr->b_l1hdr.b_arc_access < 3100 meta_hdr->b_l1hdr.b_arc_access) { 3101 type = ARC_BUFC_DATA; 3102 } else { 3103 type = ARC_BUFC_METADATA; 3104 } 3105 } 3106 3107 multilist_sublist_unlock(meta_mls); 3108 multilist_sublist_unlock(data_mls); 3109 3110 return (type); 3111} 3112 3113/* 3114 * Evict buffers from the cache, such that arc_size is capped by arc_c. 3115 */ 3116static uint64_t 3117arc_adjust(void) 3118{ 3119 uint64_t total_evicted = 0; 3120 uint64_t bytes; 3121 int64_t target; 3122 3123 /* 3124 * If we're over arc_meta_limit, we want to correct that before 3125 * potentially evicting data buffers below. 3126 */ 3127 total_evicted += arc_adjust_meta(); 3128 3129 /* 3130 * Adjust MRU size 3131 * 3132 * If we're over the target cache size, we want to evict enough 3133 * from the list to get back to our target size. We don't want 3134 * to evict too much from the MRU, such that it drops below 3135 * arc_p. So, if we're over our target cache size more than 3136 * the MRU is over arc_p, we'll evict enough to get back to 3137 * arc_p here, and then evict more from the MFU below. 3138 */ 3139 target = MIN((int64_t)(arc_size - arc_c), 3140 (int64_t)(refcount_count(&arc_anon->arcs_size) + 3141 refcount_count(&arc_mru->arcs_size) + arc_meta_used - arc_p)); 3142 3143 /* 3144 * If we're below arc_meta_min, always prefer to evict data. 3145 * Otherwise, try to satisfy the requested number of bytes to 3146 * evict from the type which contains older buffers; in an 3147 * effort to keep newer buffers in the cache regardless of their 3148 * type. If we cannot satisfy the number of bytes from this 3149 * type, spill over into the next type. 3150 */ 3151 if (arc_adjust_type(arc_mru) == ARC_BUFC_METADATA && 3152 arc_meta_used > arc_meta_min) { 3153 bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA); 3154 total_evicted += bytes; 3155 3156 /* 3157 * If we couldn't evict our target number of bytes from 3158 * metadata, we try to get the rest from data. 3159 */ 3160 target -= bytes; 3161 3162 total_evicted += 3163 arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA); 3164 } else { 3165 bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA); 3166 total_evicted += bytes; 3167 3168 /* 3169 * If we couldn't evict our target number of bytes from 3170 * data, we try to get the rest from metadata. 3171 */ 3172 target -= bytes; 3173 3174 total_evicted += 3175 arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA); 3176 } 3177 3178 /* 3179 * Adjust MFU size 3180 * 3181 * Now that we've tried to evict enough from the MRU to get its 3182 * size back to arc_p, if we're still above the target cache 3183 * size, we evict the rest from the MFU. 3184 */ 3185 target = arc_size - arc_c; 3186 3187 if (arc_adjust_type(arc_mfu) == ARC_BUFC_METADATA && 3188 arc_meta_used > arc_meta_min) { 3189 bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA); 3190 total_evicted += bytes; 3191 3192 /* 3193 * If we couldn't evict our target number of bytes from 3194 * metadata, we try to get the rest from data. 3195 */ 3196 target -= bytes; 3197 3198 total_evicted += 3199 arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA); 3200 } else { 3201 bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA); 3202 total_evicted += bytes; 3203 3204 /* 3205 * If we couldn't evict our target number of bytes from 3206 * data, we try to get the rest from data. 3207 */ 3208 target -= bytes; 3209 3210 total_evicted += 3211 arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA); 3212 } 3213 3214 /* 3215 * Adjust ghost lists 3216 * 3217 * In addition to the above, the ARC also defines target values 3218 * for the ghost lists. The sum of the mru list and mru ghost 3219 * list should never exceed the target size of the cache, and 3220 * the sum of the mru list, mfu list, mru ghost list, and mfu 3221 * ghost list should never exceed twice the target size of the 3222 * cache. The following logic enforces these limits on the ghost 3223 * caches, and evicts from them as needed. 3224 */ 3225 target = refcount_count(&arc_mru->arcs_size) + 3226 refcount_count(&arc_mru_ghost->arcs_size) - arc_c; 3227 3228 bytes = arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_DATA); 3229 total_evicted += bytes; 3230 3231 target -= bytes; 3232 3233 total_evicted += 3234 arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_METADATA); 3235 3236 /* 3237 * We assume the sum of the mru list and mfu list is less than 3238 * or equal to arc_c (we enforced this above), which means we 3239 * can use the simpler of the two equations below: 3240 * 3241 * mru + mfu + mru ghost + mfu ghost <= 2 * arc_c 3242 * mru ghost + mfu ghost <= arc_c 3243 */ 3244 target = refcount_count(&arc_mru_ghost->arcs_size) + 3245 refcount_count(&arc_mfu_ghost->arcs_size) - arc_c; 3246 3247 bytes = arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_DATA); 3248 total_evicted += bytes; 3249 3250 target -= bytes; 3251 3252 total_evicted += 3253 arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_METADATA); 3254 3255 return (total_evicted); 3256} 3257 3258static void 3259arc_do_user_evicts(void) 3260{ 3261 mutex_enter(&arc_user_evicts_lock); 3262 while (arc_eviction_list != NULL) { 3263 arc_buf_t *buf = arc_eviction_list; 3264 arc_eviction_list = buf->b_next; 3265 mutex_enter(&buf->b_evict_lock); 3266 buf->b_hdr = NULL; 3267 mutex_exit(&buf->b_evict_lock); 3268 mutex_exit(&arc_user_evicts_lock); 3269 3270 if (buf->b_efunc != NULL) 3271 VERIFY0(buf->b_efunc(buf->b_private)); 3272 3273 buf->b_efunc = NULL; 3274 buf->b_private = NULL; 3275 kmem_cache_free(buf_cache, buf); 3276 mutex_enter(&arc_user_evicts_lock); 3277 } 3278 mutex_exit(&arc_user_evicts_lock); 3279} 3280 3281void 3282arc_flush(spa_t *spa, boolean_t retry) 3283{ 3284 uint64_t guid = 0; 3285 3286 /* 3287 * If retry is TRUE, a spa must not be specified since we have 3288 * no good way to determine if all of a spa's buffers have been 3289 * evicted from an arc state. 3290 */ 3291 ASSERT(!retry || spa == 0); 3292 3293 if (spa != NULL) 3294 guid = spa_load_guid(spa); 3295 3296 (void) arc_flush_state(arc_mru, guid, ARC_BUFC_DATA, retry); 3297 (void) arc_flush_state(arc_mru, guid, ARC_BUFC_METADATA, retry); 3298 3299 (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_DATA, retry); 3300 (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_METADATA, retry); 3301 3302 (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_DATA, retry); 3303 (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_METADATA, retry); 3304 3305 (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_DATA, retry); 3306 (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_METADATA, retry); 3307 3308 arc_do_user_evicts(); 3309 ASSERT(spa || arc_eviction_list == NULL); 3310} 3311 3312void 3313arc_shrink(int64_t to_free) 3314{ 3315 if (arc_c > arc_c_min) { 3316 DTRACE_PROBE4(arc__shrink, uint64_t, arc_c, uint64_t, 3317 arc_c_min, uint64_t, arc_p, uint64_t, to_free); 3318 if (arc_c > arc_c_min + to_free) 3319 atomic_add_64(&arc_c, -to_free); 3320 else 3321 arc_c = arc_c_min; 3322 3323 atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift)); 3324 if (arc_c > arc_size) 3325 arc_c = MAX(arc_size, arc_c_min); 3326 if (arc_p > arc_c) 3327 arc_p = (arc_c >> 1); 3328 3329 DTRACE_PROBE2(arc__shrunk, uint64_t, arc_c, uint64_t, 3330 arc_p); 3331 3332 ASSERT(arc_c >= arc_c_min); 3333 ASSERT((int64_t)arc_p >= 0); 3334 } 3335 3336 if (arc_size > arc_c) { 3337 DTRACE_PROBE2(arc__shrink_adjust, uint64_t, arc_size, 3338 uint64_t, arc_c); 3339 (void) arc_adjust(); 3340 } 3341} 3342 3343static long needfree = 0; 3344 3345typedef enum free_memory_reason_t { 3346 FMR_UNKNOWN, 3347 FMR_NEEDFREE, 3348 FMR_LOTSFREE, 3349 FMR_SWAPFS_MINFREE, 3350 FMR_PAGES_PP_MAXIMUM, 3351 FMR_HEAP_ARENA, 3352 FMR_ZIO_ARENA, 3353 FMR_ZIO_FRAG, 3354} free_memory_reason_t; 3355 3356int64_t last_free_memory; 3357free_memory_reason_t last_free_reason; 3358 3359/* 3360 * Additional reserve of pages for pp_reserve. 3361 */ 3362int64_t arc_pages_pp_reserve = 64; 3363 3364/* 3365 * Additional reserve of pages for swapfs. 3366 */ 3367int64_t arc_swapfs_reserve = 64; 3368 3369/* 3370 * Return the amount of memory that can be consumed before reclaim will be 3371 * needed. Positive if there is sufficient free memory, negative indicates 3372 * the amount of memory that needs to be freed up. 3373 */ 3374static int64_t 3375arc_available_memory(void) 3376{ 3377 int64_t lowest = INT64_MAX; 3378 int64_t n; 3379 free_memory_reason_t r = FMR_UNKNOWN; 3380 3381#ifdef _KERNEL 3382 if (needfree > 0) { 3383 n = PAGESIZE * (-needfree); 3384 if (n < lowest) { 3385 lowest = n; 3386 r = FMR_NEEDFREE; 3387 } 3388 } 3389 3390 /* 3391 * Cooperate with pagedaemon when it's time for it to scan 3392 * and reclaim some pages. 3393 */ 3394 n = PAGESIZE * ((int64_t)freemem - zfs_arc_free_target); 3395 if (n < lowest) { 3396 lowest = n; 3397 r = FMR_LOTSFREE; 3398 } 3399 3400#ifdef sun 3401 /* 3402 * check that we're out of range of the pageout scanner. It starts to 3403 * schedule paging if freemem is less than lotsfree and needfree. 3404 * lotsfree is the high-water mark for pageout, and needfree is the 3405 * number of needed free pages. We add extra pages here to make sure 3406 * the scanner doesn't start up while we're freeing memory. 3407 */ 3408 n = PAGESIZE * (freemem - lotsfree - needfree - desfree); 3409 if (n < lowest) { 3410 lowest = n; 3411 r = FMR_LOTSFREE; 3412 } 3413 3414 /* 3415 * check to make sure that swapfs has enough space so that anon 3416 * reservations can still succeed. anon_resvmem() checks that the 3417 * availrmem is greater than swapfs_minfree, and the number of reserved 3418 * swap pages. We also add a bit of extra here just to prevent 3419 * circumstances from getting really dire. 3420 */ 3421 n = PAGESIZE * (availrmem - swapfs_minfree - swapfs_reserve - 3422 desfree - arc_swapfs_reserve); 3423 if (n < lowest) { 3424 lowest = n; 3425 r = FMR_SWAPFS_MINFREE; 3426 } 3427 3428 3429 /* 3430 * Check that we have enough availrmem that memory locking (e.g., via 3431 * mlock(3C) or memcntl(2)) can still succeed. (pages_pp_maximum 3432 * stores the number of pages that cannot be locked; when availrmem 3433 * drops below pages_pp_maximum, page locking mechanisms such as 3434 * page_pp_lock() will fail.) 3435 */ 3436 n = PAGESIZE * (availrmem - pages_pp_maximum - 3437 arc_pages_pp_reserve); 3438 if (n < lowest) { 3439 lowest = n; 3440 r = FMR_PAGES_PP_MAXIMUM; 3441 } 3442 3443#endif /* sun */ 3444#if defined(__i386) || !defined(UMA_MD_SMALL_ALLOC) 3445 /* 3446 * If we're on an i386 platform, it's possible that we'll exhaust the 3447 * kernel heap space before we ever run out of available physical 3448 * memory. Most checks of the size of the heap_area compare against 3449 * tune.t_minarmem, which is the minimum available real memory that we 3450 * can have in the system. However, this is generally fixed at 25 pages 3451 * which is so low that it's useless. In this comparison, we seek to 3452 * calculate the total heap-size, and reclaim if more than 3/4ths of the 3453 * heap is allocated. (Or, in the calculation, if less than 1/4th is 3454 * free) 3455 */ 3456 n = (int64_t)vmem_size(heap_arena, VMEM_FREE) - 3457 (vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC) >> 2); 3458 if (n < lowest) { 3459 lowest = n; 3460 r = FMR_HEAP_ARENA; 3461 } 3462#define zio_arena NULL 3463#else 3464#define zio_arena heap_arena 3465#endif 3466 3467 /* 3468 * If zio data pages are being allocated out of a separate heap segment, 3469 * then enforce that the size of available vmem for this arena remains 3470 * above about 1/16th free. 3471 * 3472 * Note: The 1/16th arena free requirement was put in place 3473 * to aggressively evict memory from the arc in order to avoid 3474 * memory fragmentation issues. 3475 */ 3476 if (zio_arena != NULL) { 3477 n = (int64_t)vmem_size(zio_arena, VMEM_FREE) - 3478 (vmem_size(zio_arena, VMEM_ALLOC) >> 4); 3479 if (n < lowest) { 3480 lowest = n; 3481 r = FMR_ZIO_ARENA; 3482 } 3483 } 3484 3485 /* 3486 * Above limits know nothing about real level of KVA fragmentation. 3487 * Start aggressive reclamation if too little sequential KVA left. 3488 */ 3489 if (lowest > 0) { 3490 n = (vmem_size(heap_arena, VMEM_MAXFREE) < zfs_max_recordsize) ? 3491 -((int64_t)vmem_size(heap_arena, VMEM_ALLOC) >> 4) : 3492 INT64_MAX; 3493 if (n < lowest) { 3494 lowest = n; 3495 r = FMR_ZIO_FRAG; 3496 } 3497 } 3498 3499#else /* _KERNEL */ 3500 /* Every 100 calls, free a small amount */ 3501 if (spa_get_random(100) == 0) 3502 lowest = -1024; 3503#endif /* _KERNEL */ 3504 3505 last_free_memory = lowest; 3506 last_free_reason = r; 3507 DTRACE_PROBE2(arc__available_memory, int64_t, lowest, int, r); 3508 return (lowest); 3509} 3510 3511 3512/* 3513 * Determine if the system is under memory pressure and is asking 3514 * to reclaim memory. A return value of TRUE indicates that the system 3515 * is under memory pressure and that the arc should adjust accordingly. 3516 */ 3517static boolean_t 3518arc_reclaim_needed(void) 3519{ 3520 return (arc_available_memory() < 0); 3521} 3522 3523extern kmem_cache_t *zio_buf_cache[]; 3524extern kmem_cache_t *zio_data_buf_cache[]; 3525extern kmem_cache_t *range_seg_cache; 3526 3527static __noinline void 3528arc_kmem_reap_now(void) 3529{ 3530 size_t i; 3531 kmem_cache_t *prev_cache = NULL; 3532 kmem_cache_t *prev_data_cache = NULL; 3533 3534 DTRACE_PROBE(arc__kmem_reap_start); 3535#ifdef _KERNEL 3536 if (arc_meta_used >= arc_meta_limit) { 3537 /* 3538 * We are exceeding our meta-data cache limit. 3539 * Purge some DNLC entries to release holds on meta-data. 3540 */ 3541 dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent); 3542 } 3543#if defined(__i386) 3544 /* 3545 * Reclaim unused memory from all kmem caches. 3546 */ 3547 kmem_reap(); 3548#endif 3549#endif 3550 3551 for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) { 3552 if (zio_buf_cache[i] != prev_cache) { 3553 prev_cache = zio_buf_cache[i]; 3554 kmem_cache_reap_now(zio_buf_cache[i]); 3555 } 3556 if (zio_data_buf_cache[i] != prev_data_cache) { 3557 prev_data_cache = zio_data_buf_cache[i]; 3558 kmem_cache_reap_now(zio_data_buf_cache[i]); 3559 } 3560 } 3561 kmem_cache_reap_now(buf_cache); 3562 kmem_cache_reap_now(hdr_full_cache); 3563 kmem_cache_reap_now(hdr_l2only_cache); 3564 kmem_cache_reap_now(range_seg_cache); 3565 3566#ifdef sun 3567 if (zio_arena != NULL) { 3568 /* 3569 * Ask the vmem arena to reclaim unused memory from its 3570 * quantum caches. 3571 */ 3572 vmem_qcache_reap(zio_arena); 3573 } 3574#endif 3575 DTRACE_PROBE(arc__kmem_reap_end); 3576} 3577 3578/* 3579 * Threads can block in arc_get_data_buf() waiting for this thread to evict 3580 * enough data and signal them to proceed. When this happens, the threads in 3581 * arc_get_data_buf() are sleeping while holding the hash lock for their 3582 * particular arc header. Thus, we must be careful to never sleep on a 3583 * hash lock in this thread. This is to prevent the following deadlock: 3584 * 3585 * - Thread A sleeps on CV in arc_get_data_buf() holding hash lock "L", 3586 * waiting for the reclaim thread to signal it. 3587 * 3588 * - arc_reclaim_thread() tries to acquire hash lock "L" using mutex_enter, 3589 * fails, and goes to sleep forever. 3590 * 3591 * This possible deadlock is avoided by always acquiring a hash lock 3592 * using mutex_tryenter() from arc_reclaim_thread(). 3593 */ 3594static void 3595arc_reclaim_thread(void *dummy __unused) 3596{ 3597 clock_t growtime = 0; 3598 callb_cpr_t cpr; 3599 3600 CALLB_CPR_INIT(&cpr, &arc_reclaim_lock, callb_generic_cpr, FTAG); 3601 3602 mutex_enter(&arc_reclaim_lock); 3603 while (!arc_reclaim_thread_exit) { 3604 int64_t free_memory = arc_available_memory(); 3605 uint64_t evicted = 0; 3606 3607 mutex_exit(&arc_reclaim_lock); 3608 3609 if (free_memory < 0) { 3610 3611 arc_no_grow = B_TRUE; 3612 arc_warm = B_TRUE; 3613 3614 /* 3615 * Wait at least zfs_grow_retry (default 60) seconds 3616 * before considering growing. 3617 */ 3618 growtime = ddi_get_lbolt() + (arc_grow_retry * hz); 3619 3620 arc_kmem_reap_now(); 3621 3622 /* 3623 * If we are still low on memory, shrink the ARC 3624 * so that we have arc_shrink_min free space. 3625 */ 3626 free_memory = arc_available_memory(); 3627 3628 int64_t to_free = 3629 (arc_c >> arc_shrink_shift) - free_memory; 3630 if (to_free > 0) { 3631#ifdef _KERNEL 3632 to_free = MAX(to_free, ptob(needfree)); 3633#endif 3634 arc_shrink(to_free); 3635 } 3636 } else if (free_memory < arc_c >> arc_no_grow_shift) { 3637 arc_no_grow = B_TRUE; 3638 } else if (ddi_get_lbolt() >= growtime) { 3639 arc_no_grow = B_FALSE; 3640 } 3641 3642 evicted = arc_adjust(); 3643 3644 mutex_enter(&arc_reclaim_lock); 3645 3646 /* 3647 * If evicted is zero, we couldn't evict anything via 3648 * arc_adjust(). This could be due to hash lock 3649 * collisions, but more likely due to the majority of 3650 * arc buffers being unevictable. Therefore, even if 3651 * arc_size is above arc_c, another pass is unlikely to 3652 * be helpful and could potentially cause us to enter an 3653 * infinite loop. 3654 */ 3655 if (arc_size <= arc_c || evicted == 0) { 3656#ifdef _KERNEL 3657 needfree = 0; 3658#endif 3659 /* 3660 * We're either no longer overflowing, or we 3661 * can't evict anything more, so we should wake 3662 * up any threads before we go to sleep. 3663 */ 3664 cv_broadcast(&arc_reclaim_waiters_cv); 3665 3666 /* 3667 * Block until signaled, or after one second (we 3668 * might need to perform arc_kmem_reap_now() 3669 * even if we aren't being signalled) 3670 */ 3671 CALLB_CPR_SAFE_BEGIN(&cpr); 3672 (void) cv_timedwait(&arc_reclaim_thread_cv, 3673 &arc_reclaim_lock, hz); 3674 CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_lock); 3675 } 3676 } 3677 3678 arc_reclaim_thread_exit = FALSE; 3679 cv_broadcast(&arc_reclaim_thread_cv); 3680 CALLB_CPR_EXIT(&cpr); /* drops arc_reclaim_lock */ 3681 thread_exit(); 3682} 3683 3684static void 3685arc_user_evicts_thread(void *dummy __unused) 3686{ 3687 callb_cpr_t cpr; 3688 3689 CALLB_CPR_INIT(&cpr, &arc_user_evicts_lock, callb_generic_cpr, FTAG); 3690 3691 mutex_enter(&arc_user_evicts_lock); 3692 while (!arc_user_evicts_thread_exit) { 3693 mutex_exit(&arc_user_evicts_lock); 3694 3695 arc_do_user_evicts(); 3696 3697 /* 3698 * This is necessary in order for the mdb ::arc dcmd to 3699 * show up to date information. Since the ::arc command 3700 * does not call the kstat's update function, without 3701 * this call, the command may show stale stats for the 3702 * anon, mru, mru_ghost, mfu, and mfu_ghost lists. Even 3703 * with this change, the data might be up to 1 second 3704 * out of date; but that should suffice. The arc_state_t 3705 * structures can be queried directly if more accurate 3706 * information is needed. 3707 */ 3708 if (arc_ksp != NULL) 3709 arc_ksp->ks_update(arc_ksp, KSTAT_READ); 3710 3711 mutex_enter(&arc_user_evicts_lock); 3712 3713 /* 3714 * Block until signaled, or after one second (we need to 3715 * call the arc's kstat update function regularly). 3716 */ 3717 CALLB_CPR_SAFE_BEGIN(&cpr); 3718 (void) cv_timedwait(&arc_user_evicts_cv, 3719 &arc_user_evicts_lock, hz); 3720 CALLB_CPR_SAFE_END(&cpr, &arc_user_evicts_lock); 3721 } 3722 3723 arc_user_evicts_thread_exit = FALSE; 3724 cv_broadcast(&arc_user_evicts_cv); 3725 CALLB_CPR_EXIT(&cpr); /* drops arc_user_evicts_lock */ 3726 thread_exit(); 3727} 3728 3729/* 3730 * Adapt arc info given the number of bytes we are trying to add and 3731 * the state that we are comming from. This function is only called 3732 * when we are adding new content to the cache. 3733 */ 3734static void 3735arc_adapt(int bytes, arc_state_t *state) 3736{ 3737 int mult; 3738 uint64_t arc_p_min = (arc_c >> arc_p_min_shift); 3739 int64_t mrug_size = refcount_count(&arc_mru_ghost->arcs_size); 3740 int64_t mfug_size = refcount_count(&arc_mfu_ghost->arcs_size); 3741 3742 if (state == arc_l2c_only) 3743 return; 3744 3745 ASSERT(bytes > 0); 3746 /* 3747 * Adapt the target size of the MRU list: 3748 * - if we just hit in the MRU ghost list, then increase 3749 * the target size of the MRU list. 3750 * - if we just hit in the MFU ghost list, then increase 3751 * the target size of the MFU list by decreasing the 3752 * target size of the MRU list. 3753 */ 3754 if (state == arc_mru_ghost) { 3755 mult = (mrug_size >= mfug_size) ? 1 : (mfug_size / mrug_size); 3756 mult = MIN(mult, 10); /* avoid wild arc_p adjustment */ 3757 3758 arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult); 3759 } else if (state == arc_mfu_ghost) { 3760 uint64_t delta; 3761 3762 mult = (mfug_size >= mrug_size) ? 1 : (mrug_size / mfug_size); 3763 mult = MIN(mult, 10); 3764 3765 delta = MIN(bytes * mult, arc_p); 3766 arc_p = MAX(arc_p_min, arc_p - delta); 3767 } 3768 ASSERT((int64_t)arc_p >= 0); 3769 3770 if (arc_reclaim_needed()) { 3771 cv_signal(&arc_reclaim_thread_cv); 3772 return; 3773 } 3774 3775 if (arc_no_grow) 3776 return; 3777 3778 if (arc_c >= arc_c_max) 3779 return; 3780 3781 /* 3782 * If we're within (2 * maxblocksize) bytes of the target 3783 * cache size, increment the target cache size 3784 */ 3785 if (arc_size > arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) { 3786 DTRACE_PROBE1(arc__inc_adapt, int, bytes); 3787 atomic_add_64(&arc_c, (int64_t)bytes); 3788 if (arc_c > arc_c_max) 3789 arc_c = arc_c_max; 3790 else if (state == arc_anon) 3791 atomic_add_64(&arc_p, (int64_t)bytes); 3792 if (arc_p > arc_c) 3793 arc_p = arc_c; 3794 } 3795 ASSERT((int64_t)arc_p >= 0); 3796} 3797 3798/* 3799 * Check if arc_size has grown past our upper threshold, determined by 3800 * zfs_arc_overflow_shift. 3801 */ 3802static boolean_t 3803arc_is_overflowing(void) 3804{ 3805 /* Always allow at least one block of overflow */ 3806 uint64_t overflow = MAX(SPA_MAXBLOCKSIZE, 3807 arc_c >> zfs_arc_overflow_shift); 3808 3809 return (arc_size >= arc_c + overflow); 3810} 3811 3812/* 3813 * The buffer, supplied as the first argument, needs a data block. If we 3814 * are hitting the hard limit for the cache size, we must sleep, waiting 3815 * for the eviction thread to catch up. If we're past the target size 3816 * but below the hard limit, we'll only signal the reclaim thread and 3817 * continue on. 3818 */ 3819static void 3820arc_get_data_buf(arc_buf_t *buf) 3821{ 3822 arc_state_t *state = buf->b_hdr->b_l1hdr.b_state; 3823 uint64_t size = buf->b_hdr->b_size; 3824 arc_buf_contents_t type = arc_buf_type(buf->b_hdr); 3825 3826 arc_adapt(size, state); 3827 3828 /* 3829 * If arc_size is currently overflowing, and has grown past our 3830 * upper limit, we must be adding data faster than the evict 3831 * thread can evict. Thus, to ensure we don't compound the 3832 * problem by adding more data and forcing arc_size to grow even 3833 * further past it's target size, we halt and wait for the 3834 * eviction thread to catch up. 3835 * 3836 * It's also possible that the reclaim thread is unable to evict 3837 * enough buffers to get arc_size below the overflow limit (e.g. 3838 * due to buffers being un-evictable, or hash lock collisions). 3839 * In this case, we want to proceed regardless if we're 3840 * overflowing; thus we don't use a while loop here. 3841 */ 3842 if (arc_is_overflowing()) { 3843 mutex_enter(&arc_reclaim_lock); 3844 3845 /* 3846 * Now that we've acquired the lock, we may no longer be 3847 * over the overflow limit, lets check. 3848 * 3849 * We're ignoring the case of spurious wake ups. If that 3850 * were to happen, it'd let this thread consume an ARC 3851 * buffer before it should have (i.e. before we're under 3852 * the overflow limit and were signalled by the reclaim 3853 * thread). As long as that is a rare occurrence, it 3854 * shouldn't cause any harm. 3855 */ 3856 if (arc_is_overflowing()) { 3857 cv_signal(&arc_reclaim_thread_cv); 3858 cv_wait(&arc_reclaim_waiters_cv, &arc_reclaim_lock); 3859 } 3860 3861 mutex_exit(&arc_reclaim_lock); 3862 } 3863 3864 if (type == ARC_BUFC_METADATA) { 3865 buf->b_data = zio_buf_alloc(size); 3866 arc_space_consume(size, ARC_SPACE_META); 3867 } else { 3868 ASSERT(type == ARC_BUFC_DATA); 3869 buf->b_data = zio_data_buf_alloc(size); 3870 arc_space_consume(size, ARC_SPACE_DATA); 3871 } 3872 3873 /* 3874 * Update the state size. Note that ghost states have a 3875 * "ghost size" and so don't need to be updated. 3876 */ 3877 if (!GHOST_STATE(buf->b_hdr->b_l1hdr.b_state)) { 3878 arc_buf_hdr_t *hdr = buf->b_hdr; 3879 arc_state_t *state = hdr->b_l1hdr.b_state; 3880 3881 (void) refcount_add_many(&state->arcs_size, size, buf); 3882 3883 /* 3884 * If this is reached via arc_read, the link is 3885 * protected by the hash lock. If reached via 3886 * arc_buf_alloc, the header should not be accessed by 3887 * any other thread. And, if reached via arc_read_done, 3888 * the hash lock will protect it if it's found in the 3889 * hash table; otherwise no other thread should be 3890 * trying to [add|remove]_reference it. 3891 */ 3892 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) { 3893 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 3894 atomic_add_64(&hdr->b_l1hdr.b_state->arcs_lsize[type], 3895 size); 3896 } 3897 /* 3898 * If we are growing the cache, and we are adding anonymous 3899 * data, and we have outgrown arc_p, update arc_p 3900 */ 3901 if (arc_size < arc_c && hdr->b_l1hdr.b_state == arc_anon && 3902 (refcount_count(&arc_anon->arcs_size) + 3903 refcount_count(&arc_mru->arcs_size) > arc_p)) 3904 arc_p = MIN(arc_c, arc_p + size); 3905 } 3906 ARCSTAT_BUMP(arcstat_allocated); 3907} 3908 3909/* 3910 * This routine is called whenever a buffer is accessed. 3911 * NOTE: the hash lock is dropped in this function. 3912 */ 3913static void 3914arc_access(arc_buf_hdr_t *hdr, kmutex_t *hash_lock) 3915{ 3916 clock_t now; 3917 3918 ASSERT(MUTEX_HELD(hash_lock)); 3919 ASSERT(HDR_HAS_L1HDR(hdr)); 3920 3921 if (hdr->b_l1hdr.b_state == arc_anon) { 3922 /* 3923 * This buffer is not in the cache, and does not 3924 * appear in our "ghost" list. Add the new buffer 3925 * to the MRU state. 3926 */ 3927 3928 ASSERT0(hdr->b_l1hdr.b_arc_access); 3929 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 3930 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr); 3931 arc_change_state(arc_mru, hdr, hash_lock); 3932 3933 } else if (hdr->b_l1hdr.b_state == arc_mru) { 3934 now = ddi_get_lbolt(); 3935 3936 /* 3937 * If this buffer is here because of a prefetch, then either: 3938 * - clear the flag if this is a "referencing" read 3939 * (any subsequent access will bump this into the MFU state). 3940 * or 3941 * - move the buffer to the head of the list if this is 3942 * another prefetch (to make it less likely to be evicted). 3943 */ 3944 if (HDR_PREFETCH(hdr)) { 3945 if (refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) { 3946 /* link protected by hash lock */ 3947 ASSERT(multilist_link_active( 3948 &hdr->b_l1hdr.b_arc_node)); 3949 } else { 3950 hdr->b_flags &= ~ARC_FLAG_PREFETCH; 3951 ARCSTAT_BUMP(arcstat_mru_hits); 3952 } 3953 hdr->b_l1hdr.b_arc_access = now; 3954 return; 3955 } 3956 3957 /* 3958 * This buffer has been "accessed" only once so far, 3959 * but it is still in the cache. Move it to the MFU 3960 * state. 3961 */ 3962 if (now > hdr->b_l1hdr.b_arc_access + ARC_MINTIME) { 3963 /* 3964 * More than 125ms have passed since we 3965 * instantiated this buffer. Move it to the 3966 * most frequently used state. 3967 */ 3968 hdr->b_l1hdr.b_arc_access = now; 3969 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 3970 arc_change_state(arc_mfu, hdr, hash_lock); 3971 } 3972 ARCSTAT_BUMP(arcstat_mru_hits); 3973 } else if (hdr->b_l1hdr.b_state == arc_mru_ghost) { 3974 arc_state_t *new_state; 3975 /* 3976 * This buffer has been "accessed" recently, but 3977 * was evicted from the cache. Move it to the 3978 * MFU state. 3979 */ 3980 3981 if (HDR_PREFETCH(hdr)) { 3982 new_state = arc_mru; 3983 if (refcount_count(&hdr->b_l1hdr.b_refcnt) > 0) 3984 hdr->b_flags &= ~ARC_FLAG_PREFETCH; 3985 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr); 3986 } else { 3987 new_state = arc_mfu; 3988 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 3989 } 3990 3991 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 3992 arc_change_state(new_state, hdr, hash_lock); 3993 3994 ARCSTAT_BUMP(arcstat_mru_ghost_hits); 3995 } else if (hdr->b_l1hdr.b_state == arc_mfu) { 3996 /* 3997 * This buffer has been accessed more than once and is 3998 * still in the cache. Keep it in the MFU state. 3999 * 4000 * NOTE: an add_reference() that occurred when we did 4001 * the arc_read() will have kicked this off the list. 4002 * If it was a prefetch, we will explicitly move it to 4003 * the head of the list now. 4004 */ 4005 if ((HDR_PREFETCH(hdr)) != 0) { 4006 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 4007 /* link protected by hash_lock */ 4008 ASSERT(multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 4009 } 4010 ARCSTAT_BUMP(arcstat_mfu_hits); 4011 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 4012 } else if (hdr->b_l1hdr.b_state == arc_mfu_ghost) { 4013 arc_state_t *new_state = arc_mfu; 4014 /* 4015 * This buffer has been accessed more than once but has 4016 * been evicted from the cache. Move it back to the 4017 * MFU state. 4018 */ 4019 4020 if (HDR_PREFETCH(hdr)) { 4021 /* 4022 * This is a prefetch access... 4023 * move this block back to the MRU state. 4024 */ 4025 ASSERT0(refcount_count(&hdr->b_l1hdr.b_refcnt)); 4026 new_state = arc_mru; 4027 } 4028 4029 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 4030 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 4031 arc_change_state(new_state, hdr, hash_lock); 4032 4033 ARCSTAT_BUMP(arcstat_mfu_ghost_hits); 4034 } else if (hdr->b_l1hdr.b_state == arc_l2c_only) { 4035 /* 4036 * This buffer is on the 2nd Level ARC. 4037 */ 4038 4039 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 4040 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 4041 arc_change_state(arc_mfu, hdr, hash_lock); 4042 } else { 4043 ASSERT(!"invalid arc state"); 4044 } 4045} 4046 4047/* a generic arc_done_func_t which you can use */ 4048/* ARGSUSED */ 4049void 4050arc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg) 4051{ 4052 if (zio == NULL || zio->io_error == 0) 4053 bcopy(buf->b_data, arg, buf->b_hdr->b_size); 4054 VERIFY(arc_buf_remove_ref(buf, arg)); 4055} 4056 4057/* a generic arc_done_func_t */ 4058void 4059arc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg) 4060{ 4061 arc_buf_t **bufp = arg; 4062 if (zio && zio->io_error) { 4063 VERIFY(arc_buf_remove_ref(buf, arg)); 4064 *bufp = NULL; 4065 } else { 4066 *bufp = buf; 4067 ASSERT(buf->b_data); 4068 } 4069} 4070 4071static void 4072arc_read_done(zio_t *zio) 4073{ 4074 arc_buf_hdr_t *hdr; 4075 arc_buf_t *buf; 4076 arc_buf_t *abuf; /* buffer we're assigning to callback */ 4077 kmutex_t *hash_lock = NULL; 4078 arc_callback_t *callback_list, *acb; 4079 int freeable = FALSE; 4080 4081 buf = zio->io_private; 4082 hdr = buf->b_hdr; 4083 4084 /* 4085 * The hdr was inserted into hash-table and removed from lists 4086 * prior to starting I/O. We should find this header, since 4087 * it's in the hash table, and it should be legit since it's 4088 * not possible to evict it during the I/O. The only possible 4089 * reason for it not to be found is if we were freed during the 4090 * read. 4091 */ 4092 if (HDR_IN_HASH_TABLE(hdr)) { 4093 ASSERT3U(hdr->b_birth, ==, BP_PHYSICAL_BIRTH(zio->io_bp)); 4094 ASSERT3U(hdr->b_dva.dva_word[0], ==, 4095 BP_IDENTITY(zio->io_bp)->dva_word[0]); 4096 ASSERT3U(hdr->b_dva.dva_word[1], ==, 4097 BP_IDENTITY(zio->io_bp)->dva_word[1]); 4098 4099 arc_buf_hdr_t *found = buf_hash_find(hdr->b_spa, zio->io_bp, 4100 &hash_lock); 4101 4102 ASSERT((found == NULL && HDR_FREED_IN_READ(hdr) && 4103 hash_lock == NULL) || 4104 (found == hdr && 4105 DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) || 4106 (found == hdr && HDR_L2_READING(hdr))); 4107 } 4108 4109 hdr->b_flags &= ~ARC_FLAG_L2_EVICTED; 4110 if (l2arc_noprefetch && HDR_PREFETCH(hdr)) 4111 hdr->b_flags &= ~ARC_FLAG_L2CACHE; 4112 4113 /* byteswap if necessary */ 4114 callback_list = hdr->b_l1hdr.b_acb; 4115 ASSERT(callback_list != NULL); 4116 if (BP_SHOULD_BYTESWAP(zio->io_bp) && zio->io_error == 0) { 4117 dmu_object_byteswap_t bswap = 4118 DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp)); 4119 arc_byteswap_func_t *func = BP_GET_LEVEL(zio->io_bp) > 0 ? 4120 byteswap_uint64_array : 4121 dmu_ot_byteswap[bswap].ob_func; 4122 func(buf->b_data, hdr->b_size); 4123 } 4124 4125 arc_cksum_compute(buf, B_FALSE); 4126#ifdef illumos 4127 arc_buf_watch(buf); 4128#endif /* illumos */ 4129 4130 if (hash_lock && zio->io_error == 0 && 4131 hdr->b_l1hdr.b_state == arc_anon) { 4132 /* 4133 * Only call arc_access on anonymous buffers. This is because 4134 * if we've issued an I/O for an evicted buffer, we've already 4135 * called arc_access (to prevent any simultaneous readers from 4136 * getting confused). 4137 */ 4138 arc_access(hdr, hash_lock); 4139 } 4140 4141 /* create copies of the data buffer for the callers */ 4142 abuf = buf; 4143 for (acb = callback_list; acb; acb = acb->acb_next) { 4144 if (acb->acb_done) { 4145 if (abuf == NULL) { 4146 ARCSTAT_BUMP(arcstat_duplicate_reads); 4147 abuf = arc_buf_clone(buf); 4148 } 4149 acb->acb_buf = abuf; 4150 abuf = NULL; 4151 } 4152 } 4153 hdr->b_l1hdr.b_acb = NULL; 4154 hdr->b_flags &= ~ARC_FLAG_IO_IN_PROGRESS; 4155 ASSERT(!HDR_BUF_AVAILABLE(hdr)); 4156 if (abuf == buf) { 4157 ASSERT(buf->b_efunc == NULL); 4158 ASSERT(hdr->b_l1hdr.b_datacnt == 1); 4159 hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE; 4160 } 4161 4162 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt) || 4163 callback_list != NULL); 4164 4165 if (zio->io_error != 0) { 4166 hdr->b_flags |= ARC_FLAG_IO_ERROR; 4167 if (hdr->b_l1hdr.b_state != arc_anon) 4168 arc_change_state(arc_anon, hdr, hash_lock); 4169 if (HDR_IN_HASH_TABLE(hdr)) 4170 buf_hash_remove(hdr); 4171 freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt); 4172 } 4173 4174 /* 4175 * Broadcast before we drop the hash_lock to avoid the possibility 4176 * that the hdr (and hence the cv) might be freed before we get to 4177 * the cv_broadcast(). 4178 */ 4179 cv_broadcast(&hdr->b_l1hdr.b_cv); 4180 4181 if (hash_lock != NULL) { 4182 mutex_exit(hash_lock); 4183 } else { 4184 /* 4185 * This block was freed while we waited for the read to 4186 * complete. It has been removed from the hash table and 4187 * moved to the anonymous state (so that it won't show up 4188 * in the cache). 4189 */ 4190 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon); 4191 freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt); 4192 } 4193 4194 /* execute each callback and free its structure */ 4195 while ((acb = callback_list) != NULL) { 4196 if (acb->acb_done) 4197 acb->acb_done(zio, acb->acb_buf, acb->acb_private); 4198 4199 if (acb->acb_zio_dummy != NULL) { 4200 acb->acb_zio_dummy->io_error = zio->io_error; 4201 zio_nowait(acb->acb_zio_dummy); 4202 } 4203 4204 callback_list = acb->acb_next; 4205 kmem_free(acb, sizeof (arc_callback_t)); 4206 } 4207 4208 if (freeable) 4209 arc_hdr_destroy(hdr); 4210} 4211 4212/* 4213 * "Read" the block at the specified DVA (in bp) via the 4214 * cache. If the block is found in the cache, invoke the provided 4215 * callback immediately and return. Note that the `zio' parameter 4216 * in the callback will be NULL in this case, since no IO was 4217 * required. If the block is not in the cache pass the read request 4218 * on to the spa with a substitute callback function, so that the 4219 * requested block will be added to the cache. 4220 * 4221 * If a read request arrives for a block that has a read in-progress, 4222 * either wait for the in-progress read to complete (and return the 4223 * results); or, if this is a read with a "done" func, add a record 4224 * to the read to invoke the "done" func when the read completes, 4225 * and return; or just return. 4226 * 4227 * arc_read_done() will invoke all the requested "done" functions 4228 * for readers of this block. 4229 */ 4230int 4231arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_done_func_t *done, 4232 void *private, zio_priority_t priority, int zio_flags, 4233 arc_flags_t *arc_flags, const zbookmark_phys_t *zb) 4234{ 4235 arc_buf_hdr_t *hdr = NULL; 4236 arc_buf_t *buf = NULL; 4237 kmutex_t *hash_lock = NULL; 4238 zio_t *rzio; 4239 uint64_t guid = spa_load_guid(spa); 4240 4241 ASSERT(!BP_IS_EMBEDDED(bp) || 4242 BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA); 4243 4244top: 4245 if (!BP_IS_EMBEDDED(bp)) { 4246 /* 4247 * Embedded BP's have no DVA and require no I/O to "read". 4248 * Create an anonymous arc buf to back it. 4249 */ 4250 hdr = buf_hash_find(guid, bp, &hash_lock); 4251 } 4252 4253 if (hdr != NULL && HDR_HAS_L1HDR(hdr) && hdr->b_l1hdr.b_datacnt > 0) { 4254 4255 *arc_flags |= ARC_FLAG_CACHED; 4256 4257 if (HDR_IO_IN_PROGRESS(hdr)) { 4258 4259 if ((hdr->b_flags & ARC_FLAG_PRIO_ASYNC_READ) && 4260 priority == ZIO_PRIORITY_SYNC_READ) { 4261 /* 4262 * This sync read must wait for an 4263 * in-progress async read (e.g. a predictive 4264 * prefetch). Async reads are queued 4265 * separately at the vdev_queue layer, so 4266 * this is a form of priority inversion. 4267 * Ideally, we would "inherit" the demand 4268 * i/o's priority by moving the i/o from 4269 * the async queue to the synchronous queue, 4270 * but there is currently no mechanism to do 4271 * so. Track this so that we can evaluate 4272 * the magnitude of this potential performance 4273 * problem. 4274 * 4275 * Note that if the prefetch i/o is already 4276 * active (has been issued to the device), 4277 * the prefetch improved performance, because 4278 * we issued it sooner than we would have 4279 * without the prefetch. 4280 */ 4281 DTRACE_PROBE1(arc__sync__wait__for__async, 4282 arc_buf_hdr_t *, hdr); 4283 ARCSTAT_BUMP(arcstat_sync_wait_for_async); 4284 } 4285 if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) { 4286 hdr->b_flags &= ~ARC_FLAG_PREDICTIVE_PREFETCH; 4287 } 4288 4289 if (*arc_flags & ARC_FLAG_WAIT) { 4290 cv_wait(&hdr->b_l1hdr.b_cv, hash_lock); 4291 mutex_exit(hash_lock); 4292 goto top; 4293 } 4294 ASSERT(*arc_flags & ARC_FLAG_NOWAIT); 4295 4296 if (done) { 4297 arc_callback_t *acb = NULL; 4298 4299 acb = kmem_zalloc(sizeof (arc_callback_t), 4300 KM_SLEEP); 4301 acb->acb_done = done; 4302 acb->acb_private = private; 4303 if (pio != NULL) 4304 acb->acb_zio_dummy = zio_null(pio, 4305 spa, NULL, NULL, NULL, zio_flags); 4306 4307 ASSERT(acb->acb_done != NULL); 4308 acb->acb_next = hdr->b_l1hdr.b_acb; 4309 hdr->b_l1hdr.b_acb = acb; 4310 add_reference(hdr, hash_lock, private); 4311 mutex_exit(hash_lock); 4312 return (0); 4313 } 4314 mutex_exit(hash_lock); 4315 return (0); 4316 } 4317 4318 ASSERT(hdr->b_l1hdr.b_state == arc_mru || 4319 hdr->b_l1hdr.b_state == arc_mfu); 4320 4321 if (done) { 4322 if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) { 4323 /* 4324 * This is a demand read which does not have to 4325 * wait for i/o because we did a predictive 4326 * prefetch i/o for it, which has completed. 4327 */ 4328 DTRACE_PROBE1( 4329 arc__demand__hit__predictive__prefetch, 4330 arc_buf_hdr_t *, hdr); 4331 ARCSTAT_BUMP( 4332 arcstat_demand_hit_predictive_prefetch); 4333 hdr->b_flags &= ~ARC_FLAG_PREDICTIVE_PREFETCH; 4334 } 4335 add_reference(hdr, hash_lock, private); 4336 /* 4337 * If this block is already in use, create a new 4338 * copy of the data so that we will be guaranteed 4339 * that arc_release() will always succeed. 4340 */ 4341 buf = hdr->b_l1hdr.b_buf; 4342 ASSERT(buf); 4343 ASSERT(buf->b_data); 4344 if (HDR_BUF_AVAILABLE(hdr)) { 4345 ASSERT(buf->b_efunc == NULL); 4346 hdr->b_flags &= ~ARC_FLAG_BUF_AVAILABLE; 4347 } else { 4348 buf = arc_buf_clone(buf); 4349 } 4350 4351 } else if (*arc_flags & ARC_FLAG_PREFETCH && 4352 refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) { 4353 hdr->b_flags |= ARC_FLAG_PREFETCH; 4354 } 4355 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr); 4356 arc_access(hdr, hash_lock); 4357 if (*arc_flags & ARC_FLAG_L2CACHE) 4358 hdr->b_flags |= ARC_FLAG_L2CACHE; 4359 if (*arc_flags & ARC_FLAG_L2COMPRESS) 4360 hdr->b_flags |= ARC_FLAG_L2COMPRESS; 4361 mutex_exit(hash_lock); 4362 ARCSTAT_BUMP(arcstat_hits); 4363 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr), 4364 demand, prefetch, !HDR_ISTYPE_METADATA(hdr), 4365 data, metadata, hits); 4366 4367 if (done) 4368 done(NULL, buf, private); 4369 } else { 4370 uint64_t size = BP_GET_LSIZE(bp); 4371 arc_callback_t *acb; 4372 vdev_t *vd = NULL; 4373 uint64_t addr = 0; 4374 boolean_t devw = B_FALSE; 4375 enum zio_compress b_compress = ZIO_COMPRESS_OFF; 4376 int32_t b_asize = 0; 4377 4378 if (hdr == NULL) { 4379 /* this block is not in the cache */ 4380 arc_buf_hdr_t *exists = NULL; 4381 arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp); 4382 buf = arc_buf_alloc(spa, size, private, type); 4383 hdr = buf->b_hdr; 4384 if (!BP_IS_EMBEDDED(bp)) { 4385 hdr->b_dva = *BP_IDENTITY(bp); 4386 hdr->b_birth = BP_PHYSICAL_BIRTH(bp); 4387 exists = buf_hash_insert(hdr, &hash_lock); 4388 } 4389 if (exists != NULL) { 4390 /* somebody beat us to the hash insert */ 4391 mutex_exit(hash_lock); 4392 buf_discard_identity(hdr); 4393 (void) arc_buf_remove_ref(buf, private); 4394 goto top; /* restart the IO request */ 4395 } 4396 4397 /* 4398 * If there is a callback, we pass our reference to 4399 * it; otherwise we remove our reference. 4400 */ 4401 if (done == NULL) { 4402 (void) remove_reference(hdr, hash_lock, 4403 private); 4404 } 4405 if (*arc_flags & ARC_FLAG_PREFETCH) 4406 hdr->b_flags |= ARC_FLAG_PREFETCH; 4407 if (*arc_flags & ARC_FLAG_L2CACHE) 4408 hdr->b_flags |= ARC_FLAG_L2CACHE; 4409 if (*arc_flags & ARC_FLAG_L2COMPRESS) 4410 hdr->b_flags |= ARC_FLAG_L2COMPRESS; 4411 if (BP_GET_LEVEL(bp) > 0) 4412 hdr->b_flags |= ARC_FLAG_INDIRECT; 4413 } else { 4414 /* 4415 * This block is in the ghost cache. If it was L2-only 4416 * (and thus didn't have an L1 hdr), we realloc the 4417 * header to add an L1 hdr. 4418 */ 4419 if (!HDR_HAS_L1HDR(hdr)) { 4420 hdr = arc_hdr_realloc(hdr, hdr_l2only_cache, 4421 hdr_full_cache); 4422 } 4423 4424 ASSERT(GHOST_STATE(hdr->b_l1hdr.b_state)); 4425 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 4426 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 4427 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 4428 4429 /* 4430 * If there is a callback, we pass a reference to it. 4431 */ 4432 if (done != NULL) 4433 add_reference(hdr, hash_lock, private); 4434 if (*arc_flags & ARC_FLAG_PREFETCH) 4435 hdr->b_flags |= ARC_FLAG_PREFETCH; 4436 if (*arc_flags & ARC_FLAG_L2CACHE) 4437 hdr->b_flags |= ARC_FLAG_L2CACHE; 4438 if (*arc_flags & ARC_FLAG_L2COMPRESS) 4439 hdr->b_flags |= ARC_FLAG_L2COMPRESS; 4440 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 4441 buf->b_hdr = hdr; 4442 buf->b_data = NULL; 4443 buf->b_efunc = NULL; 4444 buf->b_private = NULL; 4445 buf->b_next = NULL; 4446 hdr->b_l1hdr.b_buf = buf; 4447 ASSERT0(hdr->b_l1hdr.b_datacnt); 4448 hdr->b_l1hdr.b_datacnt = 1; 4449 arc_get_data_buf(buf); 4450 arc_access(hdr, hash_lock); 4451 } 4452 4453 if (*arc_flags & ARC_FLAG_PREDICTIVE_PREFETCH) 4454 hdr->b_flags |= ARC_FLAG_PREDICTIVE_PREFETCH; 4455 ASSERT(!GHOST_STATE(hdr->b_l1hdr.b_state)); 4456 4457 acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP); 4458 acb->acb_done = done; 4459 acb->acb_private = private; 4460 4461 ASSERT(hdr->b_l1hdr.b_acb == NULL); 4462 hdr->b_l1hdr.b_acb = acb; 4463 hdr->b_flags |= ARC_FLAG_IO_IN_PROGRESS; 4464 4465 if (HDR_HAS_L2HDR(hdr) && 4466 (vd = hdr->b_l2hdr.b_dev->l2ad_vdev) != NULL) { 4467 devw = hdr->b_l2hdr.b_dev->l2ad_writing; 4468 addr = hdr->b_l2hdr.b_daddr; 4469 b_compress = hdr->b_l2hdr.b_compress; 4470 b_asize = hdr->b_l2hdr.b_asize; 4471 /* 4472 * Lock out device removal. 4473 */ 4474 if (vdev_is_dead(vd) || 4475 !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER)) 4476 vd = NULL; 4477 } 4478 4479 if (hash_lock != NULL) 4480 mutex_exit(hash_lock); 4481 4482 /* 4483 * At this point, we have a level 1 cache miss. Try again in 4484 * L2ARC if possible. 4485 */ 4486 ASSERT3U(hdr->b_size, ==, size); 4487 DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp, 4488 uint64_t, size, zbookmark_phys_t *, zb); 4489 ARCSTAT_BUMP(arcstat_misses); 4490 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr), 4491 demand, prefetch, !HDR_ISTYPE_METADATA(hdr), 4492 data, metadata, misses); 4493#ifdef _KERNEL 4494 curthread->td_ru.ru_inblock++; 4495#endif 4496 4497 if (priority == ZIO_PRIORITY_ASYNC_READ) 4498 hdr->b_flags |= ARC_FLAG_PRIO_ASYNC_READ; 4499 else 4500 hdr->b_flags &= ~ARC_FLAG_PRIO_ASYNC_READ; 4501 4502 if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) { 4503 /* 4504 * Read from the L2ARC if the following are true: 4505 * 1. The L2ARC vdev was previously cached. 4506 * 2. This buffer still has L2ARC metadata. 4507 * 3. This buffer isn't currently writing to the L2ARC. 4508 * 4. The L2ARC entry wasn't evicted, which may 4509 * also have invalidated the vdev. 4510 * 5. This isn't prefetch and l2arc_noprefetch is set. 4511 */ 4512 if (HDR_HAS_L2HDR(hdr) && 4513 !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) && 4514 !(l2arc_noprefetch && HDR_PREFETCH(hdr))) { 4515 l2arc_read_callback_t *cb; 4516 4517 DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr); 4518 ARCSTAT_BUMP(arcstat_l2_hits); 4519 4520 cb = kmem_zalloc(sizeof (l2arc_read_callback_t), 4521 KM_SLEEP); 4522 cb->l2rcb_buf = buf; 4523 cb->l2rcb_spa = spa; 4524 cb->l2rcb_bp = *bp; 4525 cb->l2rcb_zb = *zb; 4526 cb->l2rcb_flags = zio_flags; 4527 cb->l2rcb_compress = b_compress; 4528 4529 ASSERT(addr >= VDEV_LABEL_START_SIZE && 4530 addr + size < vd->vdev_psize - 4531 VDEV_LABEL_END_SIZE); 4532 4533 /* 4534 * l2arc read. The SCL_L2ARC lock will be 4535 * released by l2arc_read_done(). 4536 * Issue a null zio if the underlying buffer 4537 * was squashed to zero size by compression. 4538 */ 4539 if (b_compress == ZIO_COMPRESS_EMPTY) { 4540 rzio = zio_null(pio, spa, vd, 4541 l2arc_read_done, cb, 4542 zio_flags | ZIO_FLAG_DONT_CACHE | 4543 ZIO_FLAG_CANFAIL | 4544 ZIO_FLAG_DONT_PROPAGATE | 4545 ZIO_FLAG_DONT_RETRY); 4546 } else { 4547 rzio = zio_read_phys(pio, vd, addr, 4548 b_asize, buf->b_data, 4549 ZIO_CHECKSUM_OFF, 4550 l2arc_read_done, cb, priority, 4551 zio_flags | ZIO_FLAG_DONT_CACHE | 4552 ZIO_FLAG_CANFAIL | 4553 ZIO_FLAG_DONT_PROPAGATE | 4554 ZIO_FLAG_DONT_RETRY, B_FALSE); 4555 } 4556 DTRACE_PROBE2(l2arc__read, vdev_t *, vd, 4557 zio_t *, rzio); 4558 ARCSTAT_INCR(arcstat_l2_read_bytes, b_asize); 4559 4560 if (*arc_flags & ARC_FLAG_NOWAIT) { 4561 zio_nowait(rzio); 4562 return (0); 4563 } 4564 4565 ASSERT(*arc_flags & ARC_FLAG_WAIT); 4566 if (zio_wait(rzio) == 0) 4567 return (0); 4568 4569 /* l2arc read error; goto zio_read() */ 4570 } else { 4571 DTRACE_PROBE1(l2arc__miss, 4572 arc_buf_hdr_t *, hdr); 4573 ARCSTAT_BUMP(arcstat_l2_misses); 4574 if (HDR_L2_WRITING(hdr)) 4575 ARCSTAT_BUMP(arcstat_l2_rw_clash); 4576 spa_config_exit(spa, SCL_L2ARC, vd); 4577 } 4578 } else { 4579 if (vd != NULL) 4580 spa_config_exit(spa, SCL_L2ARC, vd); 4581 if (l2arc_ndev != 0) { 4582 DTRACE_PROBE1(l2arc__miss, 4583 arc_buf_hdr_t *, hdr); 4584 ARCSTAT_BUMP(arcstat_l2_misses); 4585 } 4586 } 4587 4588 rzio = zio_read(pio, spa, bp, buf->b_data, size, 4589 arc_read_done, buf, priority, zio_flags, zb); 4590 4591 if (*arc_flags & ARC_FLAG_WAIT) 4592 return (zio_wait(rzio)); 4593 4594 ASSERT(*arc_flags & ARC_FLAG_NOWAIT); 4595 zio_nowait(rzio); 4596 } 4597 return (0); 4598} 4599 4600void 4601arc_set_callback(arc_buf_t *buf, arc_evict_func_t *func, void *private) 4602{ 4603 ASSERT(buf->b_hdr != NULL); 4604 ASSERT(buf->b_hdr->b_l1hdr.b_state != arc_anon); 4605 ASSERT(!refcount_is_zero(&buf->b_hdr->b_l1hdr.b_refcnt) || 4606 func == NULL); 4607 ASSERT(buf->b_efunc == NULL); 4608 ASSERT(!HDR_BUF_AVAILABLE(buf->b_hdr)); 4609 4610 buf->b_efunc = func; 4611 buf->b_private = private; 4612} 4613 4614/* 4615 * Notify the arc that a block was freed, and thus will never be used again. 4616 */ 4617void 4618arc_freed(spa_t *spa, const blkptr_t *bp) 4619{ 4620 arc_buf_hdr_t *hdr; 4621 kmutex_t *hash_lock; 4622 uint64_t guid = spa_load_guid(spa); 4623 4624 ASSERT(!BP_IS_EMBEDDED(bp)); 4625 4626 hdr = buf_hash_find(guid, bp, &hash_lock); 4627 if (hdr == NULL) 4628 return; 4629 if (HDR_BUF_AVAILABLE(hdr)) { 4630 arc_buf_t *buf = hdr->b_l1hdr.b_buf; 4631 add_reference(hdr, hash_lock, FTAG); 4632 hdr->b_flags &= ~ARC_FLAG_BUF_AVAILABLE; 4633 mutex_exit(hash_lock); 4634 4635 arc_release(buf, FTAG); 4636 (void) arc_buf_remove_ref(buf, FTAG); 4637 } else { 4638 mutex_exit(hash_lock); 4639 } 4640 4641} 4642 4643/* 4644 * Clear the user eviction callback set by arc_set_callback(), first calling 4645 * it if it exists. Because the presence of a callback keeps an arc_buf cached 4646 * clearing the callback may result in the arc_buf being destroyed. However, 4647 * it will not result in the *last* arc_buf being destroyed, hence the data 4648 * will remain cached in the ARC. We make a copy of the arc buffer here so 4649 * that we can process the callback without holding any locks. 4650 * 4651 * It's possible that the callback is already in the process of being cleared 4652 * by another thread. In this case we can not clear the callback. 4653 * 4654 * Returns B_TRUE if the callback was successfully called and cleared. 4655 */ 4656boolean_t 4657arc_clear_callback(arc_buf_t *buf) 4658{ 4659 arc_buf_hdr_t *hdr; 4660 kmutex_t *hash_lock; 4661 arc_evict_func_t *efunc = buf->b_efunc; 4662 void *private = buf->b_private; 4663 4664 mutex_enter(&buf->b_evict_lock); 4665 hdr = buf->b_hdr; 4666 if (hdr == NULL) { 4667 /* 4668 * We are in arc_do_user_evicts(). 4669 */ 4670 ASSERT(buf->b_data == NULL); 4671 mutex_exit(&buf->b_evict_lock); 4672 return (B_FALSE); 4673 } else if (buf->b_data == NULL) { 4674 /* 4675 * We are on the eviction list; process this buffer now 4676 * but let arc_do_user_evicts() do the reaping. 4677 */ 4678 buf->b_efunc = NULL; 4679 mutex_exit(&buf->b_evict_lock); 4680 VERIFY0(efunc(private)); 4681 return (B_TRUE); 4682 } 4683 hash_lock = HDR_LOCK(hdr); 4684 mutex_enter(hash_lock); 4685 hdr = buf->b_hdr; 4686 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 4687 4688 ASSERT3U(refcount_count(&hdr->b_l1hdr.b_refcnt), <, 4689 hdr->b_l1hdr.b_datacnt); 4690 ASSERT(hdr->b_l1hdr.b_state == arc_mru || 4691 hdr->b_l1hdr.b_state == arc_mfu); 4692 4693 buf->b_efunc = NULL; 4694 buf->b_private = NULL; 4695 4696 if (hdr->b_l1hdr.b_datacnt > 1) { 4697 mutex_exit(&buf->b_evict_lock); 4698 arc_buf_destroy(buf, TRUE); 4699 } else { 4700 ASSERT(buf == hdr->b_l1hdr.b_buf); 4701 hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE; 4702 mutex_exit(&buf->b_evict_lock); 4703 } 4704 4705 mutex_exit(hash_lock); 4706 VERIFY0(efunc(private)); 4707 return (B_TRUE); 4708} 4709 4710/* 4711 * Release this buffer from the cache, making it an anonymous buffer. This 4712 * must be done after a read and prior to modifying the buffer contents. 4713 * If the buffer has more than one reference, we must make 4714 * a new hdr for the buffer. 4715 */ 4716void 4717arc_release(arc_buf_t *buf, void *tag) 4718{ 4719 arc_buf_hdr_t *hdr = buf->b_hdr; 4720 4721 /* 4722 * It would be nice to assert that if it's DMU metadata (level > 4723 * 0 || it's the dnode file), then it must be syncing context. 4724 * But we don't know that information at this level. 4725 */ 4726 4727 mutex_enter(&buf->b_evict_lock); 4728 4729 ASSERT(HDR_HAS_L1HDR(hdr)); 4730 4731 /* 4732 * We don't grab the hash lock prior to this check, because if 4733 * the buffer's header is in the arc_anon state, it won't be 4734 * linked into the hash table. 4735 */ 4736 if (hdr->b_l1hdr.b_state == arc_anon) { 4737 mutex_exit(&buf->b_evict_lock); 4738 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 4739 ASSERT(!HDR_IN_HASH_TABLE(hdr)); 4740 ASSERT(!HDR_HAS_L2HDR(hdr)); 4741 ASSERT(BUF_EMPTY(hdr)); 4742 ASSERT3U(hdr->b_l1hdr.b_datacnt, ==, 1); 4743 ASSERT3S(refcount_count(&hdr->b_l1hdr.b_refcnt), ==, 1); 4744 ASSERT(!list_link_active(&hdr->b_l1hdr.b_arc_node)); 4745 4746 ASSERT3P(buf->b_efunc, ==, NULL); 4747 ASSERT3P(buf->b_private, ==, NULL); 4748 4749 hdr->b_l1hdr.b_arc_access = 0; 4750 arc_buf_thaw(buf); 4751 4752 return; 4753 } 4754 4755 kmutex_t *hash_lock = HDR_LOCK(hdr); 4756 mutex_enter(hash_lock); 4757 4758 /* 4759 * This assignment is only valid as long as the hash_lock is 4760 * held, we must be careful not to reference state or the 4761 * b_state field after dropping the lock. 4762 */ 4763 arc_state_t *state = hdr->b_l1hdr.b_state; 4764 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 4765 ASSERT3P(state, !=, arc_anon); 4766 4767 /* this buffer is not on any list */ 4768 ASSERT(refcount_count(&hdr->b_l1hdr.b_refcnt) > 0); 4769 4770 if (HDR_HAS_L2HDR(hdr)) { 4771 mutex_enter(&hdr->b_l2hdr.b_dev->l2ad_mtx); 4772 4773 /* 4774 * We have to recheck this conditional again now that 4775 * we're holding the l2ad_mtx to prevent a race with 4776 * another thread which might be concurrently calling 4777 * l2arc_evict(). In that case, l2arc_evict() might have 4778 * destroyed the header's L2 portion as we were waiting 4779 * to acquire the l2ad_mtx. 4780 */ 4781 if (HDR_HAS_L2HDR(hdr)) { 4782 if (hdr->b_l2hdr.b_daddr != L2ARC_ADDR_UNSET) 4783 trim_map_free(hdr->b_l2hdr.b_dev->l2ad_vdev, 4784 hdr->b_l2hdr.b_daddr, 4785 hdr->b_l2hdr.b_asize, 0); 4786 arc_hdr_l2hdr_destroy(hdr); 4787 } 4788 4789 mutex_exit(&hdr->b_l2hdr.b_dev->l2ad_mtx); 4790 } 4791 4792 /* 4793 * Do we have more than one buf? 4794 */ 4795 if (hdr->b_l1hdr.b_datacnt > 1) { 4796 arc_buf_hdr_t *nhdr; 4797 arc_buf_t **bufp; 4798 uint64_t blksz = hdr->b_size; 4799 uint64_t spa = hdr->b_spa; 4800 arc_buf_contents_t type = arc_buf_type(hdr); 4801 uint32_t flags = hdr->b_flags; 4802 4803 ASSERT(hdr->b_l1hdr.b_buf != buf || buf->b_next != NULL); 4804 /* 4805 * Pull the data off of this hdr and attach it to 4806 * a new anonymous hdr. 4807 */ 4808 (void) remove_reference(hdr, hash_lock, tag); 4809 bufp = &hdr->b_l1hdr.b_buf; 4810 while (*bufp != buf) 4811 bufp = &(*bufp)->b_next; 4812 *bufp = buf->b_next; 4813 buf->b_next = NULL; 4814 4815 ASSERT3P(state, !=, arc_l2c_only); 4816 4817 (void) refcount_remove_many( 4818 &state->arcs_size, hdr->b_size, buf); 4819 4820 if (refcount_is_zero(&hdr->b_l1hdr.b_refcnt)) { 4821 ASSERT3P(state, !=, arc_l2c_only); 4822 uint64_t *size = &state->arcs_lsize[type]; 4823 ASSERT3U(*size, >=, hdr->b_size); 4824 atomic_add_64(size, -hdr->b_size); 4825 } 4826 4827 /* 4828 * We're releasing a duplicate user data buffer, update 4829 * our statistics accordingly. 4830 */ 4831 if (HDR_ISTYPE_DATA(hdr)) { 4832 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers); 4833 ARCSTAT_INCR(arcstat_duplicate_buffers_size, 4834 -hdr->b_size); 4835 } 4836 hdr->b_l1hdr.b_datacnt -= 1; 4837 arc_cksum_verify(buf); 4838#ifdef illumos 4839 arc_buf_unwatch(buf); 4840#endif /* illumos */ 4841 4842 mutex_exit(hash_lock); 4843 4844 nhdr = kmem_cache_alloc(hdr_full_cache, KM_PUSHPAGE); 4845 nhdr->b_size = blksz; 4846 nhdr->b_spa = spa; 4847 4848 nhdr->b_flags = flags & ARC_FLAG_L2_WRITING; 4849 nhdr->b_flags |= arc_bufc_to_flags(type); 4850 nhdr->b_flags |= ARC_FLAG_HAS_L1HDR; 4851 4852 nhdr->b_l1hdr.b_buf = buf; 4853 nhdr->b_l1hdr.b_datacnt = 1; 4854 nhdr->b_l1hdr.b_state = arc_anon; 4855 nhdr->b_l1hdr.b_arc_access = 0; 4856 nhdr->b_l1hdr.b_tmp_cdata = NULL; 4857 nhdr->b_freeze_cksum = NULL; 4858 4859 (void) refcount_add(&nhdr->b_l1hdr.b_refcnt, tag); 4860 buf->b_hdr = nhdr; 4861 mutex_exit(&buf->b_evict_lock); 4862 (void) refcount_add_many(&arc_anon->arcs_size, blksz, buf); 4863 } else { 4864 mutex_exit(&buf->b_evict_lock); 4865 ASSERT(refcount_count(&hdr->b_l1hdr.b_refcnt) == 1); 4866 /* protected by hash lock, or hdr is on arc_anon */ 4867 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 4868 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 4869 arc_change_state(arc_anon, hdr, hash_lock); 4870 hdr->b_l1hdr.b_arc_access = 0; 4871 mutex_exit(hash_lock); 4872 4873 buf_discard_identity(hdr); 4874 arc_buf_thaw(buf); 4875 } 4876 buf->b_efunc = NULL; 4877 buf->b_private = NULL; 4878} 4879 4880int 4881arc_released(arc_buf_t *buf) 4882{ 4883 int released; 4884 4885 mutex_enter(&buf->b_evict_lock); 4886 released = (buf->b_data != NULL && 4887 buf->b_hdr->b_l1hdr.b_state == arc_anon); 4888 mutex_exit(&buf->b_evict_lock); 4889 return (released); 4890} 4891 4892#ifdef ZFS_DEBUG 4893int 4894arc_referenced(arc_buf_t *buf) 4895{ 4896 int referenced; 4897 4898 mutex_enter(&buf->b_evict_lock); 4899 referenced = (refcount_count(&buf->b_hdr->b_l1hdr.b_refcnt)); 4900 mutex_exit(&buf->b_evict_lock); 4901 return (referenced); 4902} 4903#endif 4904 4905static void 4906arc_write_ready(zio_t *zio) 4907{ 4908 arc_write_callback_t *callback = zio->io_private; 4909 arc_buf_t *buf = callback->awcb_buf; 4910 arc_buf_hdr_t *hdr = buf->b_hdr; 4911 4912 ASSERT(HDR_HAS_L1HDR(hdr)); 4913 ASSERT(!refcount_is_zero(&buf->b_hdr->b_l1hdr.b_refcnt)); 4914 ASSERT(hdr->b_l1hdr.b_datacnt > 0); 4915 callback->awcb_ready(zio, buf, callback->awcb_private); 4916 4917 /* 4918 * If the IO is already in progress, then this is a re-write 4919 * attempt, so we need to thaw and re-compute the cksum. 4920 * It is the responsibility of the callback to handle the 4921 * accounting for any re-write attempt. 4922 */ 4923 if (HDR_IO_IN_PROGRESS(hdr)) { 4924 mutex_enter(&hdr->b_l1hdr.b_freeze_lock); 4925 if (hdr->b_freeze_cksum != NULL) { 4926 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 4927 hdr->b_freeze_cksum = NULL; 4928 } 4929 mutex_exit(&hdr->b_l1hdr.b_freeze_lock); 4930 } 4931 arc_cksum_compute(buf, B_FALSE); 4932 hdr->b_flags |= ARC_FLAG_IO_IN_PROGRESS; 4933} 4934 4935/* 4936 * The SPA calls this callback for each physical write that happens on behalf 4937 * of a logical write. See the comment in dbuf_write_physdone() for details. 4938 */ 4939static void 4940arc_write_physdone(zio_t *zio) 4941{ 4942 arc_write_callback_t *cb = zio->io_private; 4943 if (cb->awcb_physdone != NULL) 4944 cb->awcb_physdone(zio, cb->awcb_buf, cb->awcb_private); 4945} 4946 4947static void 4948arc_write_done(zio_t *zio) 4949{ 4950 arc_write_callback_t *callback = zio->io_private; 4951 arc_buf_t *buf = callback->awcb_buf; 4952 arc_buf_hdr_t *hdr = buf->b_hdr; 4953 4954 ASSERT(hdr->b_l1hdr.b_acb == NULL); 4955 4956 if (zio->io_error == 0) { 4957 if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) { 4958 buf_discard_identity(hdr); 4959 } else { 4960 hdr->b_dva = *BP_IDENTITY(zio->io_bp); 4961 hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp); 4962 } 4963 } else { 4964 ASSERT(BUF_EMPTY(hdr)); 4965 } 4966 4967 /* 4968 * If the block to be written was all-zero or compressed enough to be 4969 * embedded in the BP, no write was performed so there will be no 4970 * dva/birth/checksum. The buffer must therefore remain anonymous 4971 * (and uncached). 4972 */ 4973 if (!BUF_EMPTY(hdr)) { 4974 arc_buf_hdr_t *exists; 4975 kmutex_t *hash_lock; 4976 4977 ASSERT(zio->io_error == 0); 4978 4979 arc_cksum_verify(buf); 4980 4981 exists = buf_hash_insert(hdr, &hash_lock); 4982 if (exists != NULL) { 4983 /* 4984 * This can only happen if we overwrite for 4985 * sync-to-convergence, because we remove 4986 * buffers from the hash table when we arc_free(). 4987 */ 4988 if (zio->io_flags & ZIO_FLAG_IO_REWRITE) { 4989 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) 4990 panic("bad overwrite, hdr=%p exists=%p", 4991 (void *)hdr, (void *)exists); 4992 ASSERT(refcount_is_zero( 4993 &exists->b_l1hdr.b_refcnt)); 4994 arc_change_state(arc_anon, exists, hash_lock); 4995 mutex_exit(hash_lock); 4996 arc_hdr_destroy(exists); 4997 exists = buf_hash_insert(hdr, &hash_lock); 4998 ASSERT3P(exists, ==, NULL); 4999 } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) { 5000 /* nopwrite */ 5001 ASSERT(zio->io_prop.zp_nopwrite); 5002 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) 5003 panic("bad nopwrite, hdr=%p exists=%p", 5004 (void *)hdr, (void *)exists); 5005 } else { 5006 /* Dedup */ 5007 ASSERT(hdr->b_l1hdr.b_datacnt == 1); 5008 ASSERT(hdr->b_l1hdr.b_state == arc_anon); 5009 ASSERT(BP_GET_DEDUP(zio->io_bp)); 5010 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0); 5011 } 5012 } 5013 hdr->b_flags &= ~ARC_FLAG_IO_IN_PROGRESS; 5014 /* if it's not anon, we are doing a scrub */ 5015 if (exists == NULL && hdr->b_l1hdr.b_state == arc_anon) 5016 arc_access(hdr, hash_lock); 5017 mutex_exit(hash_lock); 5018 } else { 5019 hdr->b_flags &= ~ARC_FLAG_IO_IN_PROGRESS; 5020 } 5021 5022 ASSERT(!refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 5023 callback->awcb_done(zio, buf, callback->awcb_private); 5024 5025 kmem_free(callback, sizeof (arc_write_callback_t)); 5026} 5027 5028zio_t * 5029arc_write(zio_t *pio, spa_t *spa, uint64_t txg, 5030 blkptr_t *bp, arc_buf_t *buf, boolean_t l2arc, boolean_t l2arc_compress, 5031 const zio_prop_t *zp, arc_done_func_t *ready, arc_done_func_t *physdone, 5032 arc_done_func_t *done, void *private, zio_priority_t priority, 5033 int zio_flags, const zbookmark_phys_t *zb) 5034{ 5035 arc_buf_hdr_t *hdr = buf->b_hdr; 5036 arc_write_callback_t *callback; 5037 zio_t *zio; 5038 5039 ASSERT(ready != NULL); 5040 ASSERT(done != NULL); 5041 ASSERT(!HDR_IO_ERROR(hdr)); 5042 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 5043 ASSERT(hdr->b_l1hdr.b_acb == NULL); 5044 ASSERT(hdr->b_l1hdr.b_datacnt > 0); 5045 if (l2arc) 5046 hdr->b_flags |= ARC_FLAG_L2CACHE; 5047 if (l2arc_compress) 5048 hdr->b_flags |= ARC_FLAG_L2COMPRESS; 5049 callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP); 5050 callback->awcb_ready = ready; 5051 callback->awcb_physdone = physdone; 5052 callback->awcb_done = done; 5053 callback->awcb_private = private; 5054 callback->awcb_buf = buf; 5055 5056 zio = zio_write(pio, spa, txg, bp, buf->b_data, hdr->b_size, zp, 5057 arc_write_ready, arc_write_physdone, arc_write_done, callback, 5058 priority, zio_flags, zb); 5059 5060 return (zio); 5061} 5062 5063static int 5064arc_memory_throttle(uint64_t reserve, uint64_t txg) 5065{ 5066#ifdef _KERNEL 5067 uint64_t available_memory = ptob(freemem); 5068 static uint64_t page_load = 0; 5069 static uint64_t last_txg = 0; 5070 5071#if defined(__i386) || !defined(UMA_MD_SMALL_ALLOC) 5072 available_memory = 5073 MIN(available_memory, ptob(vmem_size(heap_arena, VMEM_FREE))); 5074#endif 5075 5076 if (freemem > (uint64_t)physmem * arc_lotsfree_percent / 100) 5077 return (0); 5078 5079 if (txg > last_txg) { 5080 last_txg = txg; 5081 page_load = 0; 5082 } 5083 /* 5084 * If we are in pageout, we know that memory is already tight, 5085 * the arc is already going to be evicting, so we just want to 5086 * continue to let page writes occur as quickly as possible. 5087 */ 5088 if (curproc == pageproc) { 5089 if (page_load > MAX(ptob(minfree), available_memory) / 4) 5090 return (SET_ERROR(ERESTART)); 5091 /* Note: reserve is inflated, so we deflate */ 5092 page_load += reserve / 8; 5093 return (0); 5094 } else if (page_load > 0 && arc_reclaim_needed()) { 5095 /* memory is low, delay before restarting */ 5096 ARCSTAT_INCR(arcstat_memory_throttle_count, 1); 5097 return (SET_ERROR(EAGAIN)); 5098 } 5099 page_load = 0; 5100#endif 5101 return (0); 5102} 5103 5104void 5105arc_tempreserve_clear(uint64_t reserve) 5106{ 5107 atomic_add_64(&arc_tempreserve, -reserve); 5108 ASSERT((int64_t)arc_tempreserve >= 0); 5109} 5110 5111int 5112arc_tempreserve_space(uint64_t reserve, uint64_t txg) 5113{ 5114 int error; 5115 uint64_t anon_size; 5116 5117 if (reserve > arc_c/4 && !arc_no_grow) { 5118 arc_c = MIN(arc_c_max, reserve * 4); 5119 DTRACE_PROBE1(arc__set_reserve, uint64_t, arc_c); 5120 } 5121 if (reserve > arc_c) 5122 return (SET_ERROR(ENOMEM)); 5123 5124 /* 5125 * Don't count loaned bufs as in flight dirty data to prevent long 5126 * network delays from blocking transactions that are ready to be 5127 * assigned to a txg. 5128 */ 5129 anon_size = MAX((int64_t)(refcount_count(&arc_anon->arcs_size) - 5130 arc_loaned_bytes), 0); 5131 5132 /* 5133 * Writes will, almost always, require additional memory allocations 5134 * in order to compress/encrypt/etc the data. We therefore need to 5135 * make sure that there is sufficient available memory for this. 5136 */ 5137 error = arc_memory_throttle(reserve, txg); 5138 if (error != 0) 5139 return (error); 5140 5141 /* 5142 * Throttle writes when the amount of dirty data in the cache 5143 * gets too large. We try to keep the cache less than half full 5144 * of dirty blocks so that our sync times don't grow too large. 5145 * Note: if two requests come in concurrently, we might let them 5146 * both succeed, when one of them should fail. Not a huge deal. 5147 */ 5148 5149 if (reserve + arc_tempreserve + anon_size > arc_c / 2 && 5150 anon_size > arc_c / 4) { 5151 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK " 5152 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n", 5153 arc_tempreserve>>10, 5154 arc_anon->arcs_lsize[ARC_BUFC_METADATA]>>10, 5155 arc_anon->arcs_lsize[ARC_BUFC_DATA]>>10, 5156 reserve>>10, arc_c>>10); 5157 return (SET_ERROR(ERESTART)); 5158 } 5159 atomic_add_64(&arc_tempreserve, reserve); 5160 return (0); 5161} 5162 5163static void 5164arc_kstat_update_state(arc_state_t *state, kstat_named_t *size, 5165 kstat_named_t *evict_data, kstat_named_t *evict_metadata) 5166{ 5167 size->value.ui64 = refcount_count(&state->arcs_size); 5168 evict_data->value.ui64 = state->arcs_lsize[ARC_BUFC_DATA]; 5169 evict_metadata->value.ui64 = state->arcs_lsize[ARC_BUFC_METADATA]; 5170} 5171 5172static int 5173arc_kstat_update(kstat_t *ksp, int rw) 5174{ 5175 arc_stats_t *as = ksp->ks_data; 5176 5177 if (rw == KSTAT_WRITE) { 5178 return (EACCES); 5179 } else { 5180 arc_kstat_update_state(arc_anon, 5181 &as->arcstat_anon_size, 5182 &as->arcstat_anon_evictable_data, 5183 &as->arcstat_anon_evictable_metadata); 5184 arc_kstat_update_state(arc_mru, 5185 &as->arcstat_mru_size, 5186 &as->arcstat_mru_evictable_data, 5187 &as->arcstat_mru_evictable_metadata); 5188 arc_kstat_update_state(arc_mru_ghost, 5189 &as->arcstat_mru_ghost_size, 5190 &as->arcstat_mru_ghost_evictable_data, 5191 &as->arcstat_mru_ghost_evictable_metadata); 5192 arc_kstat_update_state(arc_mfu, 5193 &as->arcstat_mfu_size, 5194 &as->arcstat_mfu_evictable_data, 5195 &as->arcstat_mfu_evictable_metadata); 5196 arc_kstat_update_state(arc_mfu_ghost, 5197 &as->arcstat_mfu_ghost_size, 5198 &as->arcstat_mfu_ghost_evictable_data, 5199 &as->arcstat_mfu_ghost_evictable_metadata); 5200 } 5201 5202 return (0); 5203} 5204 5205/* 5206 * This function *must* return indices evenly distributed between all 5207 * sublists of the multilist. This is needed due to how the ARC eviction 5208 * code is laid out; arc_evict_state() assumes ARC buffers are evenly 5209 * distributed between all sublists and uses this assumption when 5210 * deciding which sublist to evict from and how much to evict from it. 5211 */ 5212unsigned int 5213arc_state_multilist_index_func(multilist_t *ml, void *obj) 5214{ 5215 arc_buf_hdr_t *hdr = obj; 5216 5217 /* 5218 * We rely on b_dva to generate evenly distributed index 5219 * numbers using buf_hash below. So, as an added precaution, 5220 * let's make sure we never add empty buffers to the arc lists. 5221 */ 5222 ASSERT(!BUF_EMPTY(hdr)); 5223 5224 /* 5225 * The assumption here, is the hash value for a given 5226 * arc_buf_hdr_t will remain constant throughout it's lifetime 5227 * (i.e. it's b_spa, b_dva, and b_birth fields don't change). 5228 * Thus, we don't need to store the header's sublist index 5229 * on insertion, as this index can be recalculated on removal. 5230 * 5231 * Also, the low order bits of the hash value are thought to be 5232 * distributed evenly. Otherwise, in the case that the multilist 5233 * has a power of two number of sublists, each sublists' usage 5234 * would not be evenly distributed. 5235 */ 5236 return (buf_hash(hdr->b_spa, &hdr->b_dva, hdr->b_birth) % 5237 multilist_get_num_sublists(ml)); 5238} 5239 5240#ifdef _KERNEL 5241static eventhandler_tag arc_event_lowmem = NULL; 5242 5243static void 5244arc_lowmem(void *arg __unused, int howto __unused) 5245{ 5246 5247 mutex_enter(&arc_reclaim_lock); 5248 /* XXX: Memory deficit should be passed as argument. */ 5249 needfree = btoc(arc_c >> arc_shrink_shift); 5250 DTRACE_PROBE(arc__needfree); 5251 cv_signal(&arc_reclaim_thread_cv); 5252 5253 /* 5254 * It is unsafe to block here in arbitrary threads, because we can come 5255 * here from ARC itself and may hold ARC locks and thus risk a deadlock 5256 * with ARC reclaim thread. 5257 */ 5258 if (curproc == pageproc) 5259 (void) cv_wait(&arc_reclaim_waiters_cv, &arc_reclaim_lock); 5260 mutex_exit(&arc_reclaim_lock); 5261} 5262#endif 5263 5264void 5265arc_init(void) 5266{ 5267 int i, prefetch_tunable_set = 0; 5268 5269 mutex_init(&arc_reclaim_lock, NULL, MUTEX_DEFAULT, NULL); 5270 cv_init(&arc_reclaim_thread_cv, NULL, CV_DEFAULT, NULL); 5271 cv_init(&arc_reclaim_waiters_cv, NULL, CV_DEFAULT, NULL); 5272 5273 mutex_init(&arc_user_evicts_lock, NULL, MUTEX_DEFAULT, NULL); 5274 cv_init(&arc_user_evicts_cv, NULL, CV_DEFAULT, NULL); 5275 5276 /* Convert seconds to clock ticks */ 5277 arc_min_prefetch_lifespan = 1 * hz; 5278 5279 /* Start out with 1/8 of all memory */ 5280 arc_c = kmem_size() / 8; 5281 5282#ifdef sun 5283#ifdef _KERNEL 5284 /* 5285 * On architectures where the physical memory can be larger 5286 * than the addressable space (intel in 32-bit mode), we may 5287 * need to limit the cache to 1/8 of VM size. 5288 */ 5289 arc_c = MIN(arc_c, vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 8); 5290#endif 5291#endif /* sun */ 5292 /* set min cache to 1/32 of all memory, or 16MB, whichever is more */ 5293 arc_c_min = MAX(arc_c / 4, 16 << 20); 5294 /* set max to 1/2 of all memory, or all but 1GB, whichever is more */ 5295 if (arc_c * 8 >= 1 << 30) 5296 arc_c_max = (arc_c * 8) - (1 << 30); 5297 else 5298 arc_c_max = arc_c_min; 5299 arc_c_max = MAX(arc_c * 5, arc_c_max); 5300 5301 /* 5302 * In userland, there's only the memory pressure that we artificially 5303 * create (see arc_available_memory()). Don't let arc_c get too 5304 * small, because it can cause transactions to be larger than 5305 * arc_c, causing arc_tempreserve_space() to fail. 5306 */ 5307#ifndef _KERNEL 5308 arc_c_min = arc_c_max / 2; 5309#endif 5310 5311#ifdef _KERNEL 5312 /* 5313 * Allow the tunables to override our calculations if they are 5314 * reasonable (ie. over 16MB) 5315 */ 5316 if (zfs_arc_max > 16 << 20 && zfs_arc_max < kmem_size()) 5317 arc_c_max = zfs_arc_max; 5318 if (zfs_arc_min > 16 << 20 && zfs_arc_min <= arc_c_max) 5319 arc_c_min = zfs_arc_min; 5320#endif 5321 5322 arc_c = arc_c_max; 5323 arc_p = (arc_c >> 1); 5324 5325 /* limit meta-data to 1/4 of the arc capacity */ 5326 arc_meta_limit = arc_c_max / 4; 5327 5328 /* Allow the tunable to override if it is reasonable */ 5329 if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max) 5330 arc_meta_limit = zfs_arc_meta_limit; 5331 5332 if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0) 5333 arc_c_min = arc_meta_limit / 2; 5334 5335 if (zfs_arc_meta_min > 0) { 5336 arc_meta_min = zfs_arc_meta_min; 5337 } else { 5338 arc_meta_min = arc_c_min / 2; 5339 } 5340 5341 if (zfs_arc_grow_retry > 0) 5342 arc_grow_retry = zfs_arc_grow_retry; 5343 5344 if (zfs_arc_shrink_shift > 0) 5345 arc_shrink_shift = zfs_arc_shrink_shift; 5346 5347 /* 5348 * Ensure that arc_no_grow_shift is less than arc_shrink_shift. 5349 */ 5350 if (arc_no_grow_shift >= arc_shrink_shift) 5351 arc_no_grow_shift = arc_shrink_shift - 1; 5352 5353 if (zfs_arc_p_min_shift > 0) 5354 arc_p_min_shift = zfs_arc_p_min_shift; 5355 5356 if (zfs_arc_num_sublists_per_state < 1) 5357 zfs_arc_num_sublists_per_state = MAX(max_ncpus, 1); 5358 5359 /* if kmem_flags are set, lets try to use less memory */ 5360 if (kmem_debugging()) 5361 arc_c = arc_c / 2; 5362 if (arc_c < arc_c_min) 5363 arc_c = arc_c_min; 5364 5365 zfs_arc_min = arc_c_min; 5366 zfs_arc_max = arc_c_max; 5367 5368 arc_anon = &ARC_anon; 5369 arc_mru = &ARC_mru; 5370 arc_mru_ghost = &ARC_mru_ghost; 5371 arc_mfu = &ARC_mfu; 5372 arc_mfu_ghost = &ARC_mfu_ghost; 5373 arc_l2c_only = &ARC_l2c_only; 5374 arc_size = 0; 5375 5376 multilist_create(&arc_mru->arcs_list[ARC_BUFC_METADATA], 5377 sizeof (arc_buf_hdr_t), 5378 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5379 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5380 multilist_create(&arc_mru->arcs_list[ARC_BUFC_DATA], 5381 sizeof (arc_buf_hdr_t), 5382 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5383 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5384 multilist_create(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA], 5385 sizeof (arc_buf_hdr_t), 5386 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5387 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5388 multilist_create(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA], 5389 sizeof (arc_buf_hdr_t), 5390 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5391 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5392 multilist_create(&arc_mfu->arcs_list[ARC_BUFC_METADATA], 5393 sizeof (arc_buf_hdr_t), 5394 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5395 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5396 multilist_create(&arc_mfu->arcs_list[ARC_BUFC_DATA], 5397 sizeof (arc_buf_hdr_t), 5398 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5399 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5400 multilist_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA], 5401 sizeof (arc_buf_hdr_t), 5402 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5403 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5404 multilist_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA], 5405 sizeof (arc_buf_hdr_t), 5406 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5407 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5408 multilist_create(&arc_l2c_only->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_l2c_only->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 5417 refcount_create(&arc_anon->arcs_size); 5418 refcount_create(&arc_mru->arcs_size); 5419 refcount_create(&arc_mru_ghost->arcs_size); 5420 refcount_create(&arc_mfu->arcs_size); 5421 refcount_create(&arc_mfu_ghost->arcs_size); 5422 refcount_create(&arc_l2c_only->arcs_size); 5423 5424 buf_init(); 5425 5426 arc_reclaim_thread_exit = FALSE; 5427 arc_user_evicts_thread_exit = FALSE; 5428 arc_eviction_list = NULL; 5429 bzero(&arc_eviction_hdr, sizeof (arc_buf_hdr_t)); 5430 5431 arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED, 5432 sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL); 5433 5434 if (arc_ksp != NULL) { 5435 arc_ksp->ks_data = &arc_stats; 5436 arc_ksp->ks_update = arc_kstat_update; 5437 kstat_install(arc_ksp); 5438 } 5439 5440 (void) thread_create(NULL, 0, arc_reclaim_thread, NULL, 0, &p0, 5441 TS_RUN, minclsyspri); 5442 5443#ifdef _KERNEL 5444 arc_event_lowmem = EVENTHANDLER_REGISTER(vm_lowmem, arc_lowmem, NULL, 5445 EVENTHANDLER_PRI_FIRST); 5446#endif 5447 5448 (void) thread_create(NULL, 0, arc_user_evicts_thread, NULL, 0, &p0, 5449 TS_RUN, minclsyspri); 5450 5451 arc_dead = FALSE; 5452 arc_warm = B_FALSE; 5453 5454 /* 5455 * Calculate maximum amount of dirty data per pool. 5456 * 5457 * If it has been set by /etc/system, take that. 5458 * Otherwise, use a percentage of physical memory defined by 5459 * zfs_dirty_data_max_percent (default 10%) with a cap at 5460 * zfs_dirty_data_max_max (default 4GB). 5461 */ 5462 if (zfs_dirty_data_max == 0) { 5463 zfs_dirty_data_max = ptob(physmem) * 5464 zfs_dirty_data_max_percent / 100; 5465 zfs_dirty_data_max = MIN(zfs_dirty_data_max, 5466 zfs_dirty_data_max_max); 5467 } 5468 5469#ifdef _KERNEL 5470 if (TUNABLE_INT_FETCH("vfs.zfs.prefetch_disable", &zfs_prefetch_disable)) 5471 prefetch_tunable_set = 1; 5472 5473#ifdef __i386__ 5474 if (prefetch_tunable_set == 0) { 5475 printf("ZFS NOTICE: Prefetch is disabled by default on i386 " 5476 "-- to enable,\n"); 5477 printf(" add \"vfs.zfs.prefetch_disable=0\" " 5478 "to /boot/loader.conf.\n"); 5479 zfs_prefetch_disable = 1; 5480 } 5481#else 5482 if ((((uint64_t)physmem * PAGESIZE) < (1ULL << 32)) && 5483 prefetch_tunable_set == 0) { 5484 printf("ZFS NOTICE: Prefetch is disabled by default if less " 5485 "than 4GB of RAM is present;\n" 5486 " to enable, add \"vfs.zfs.prefetch_disable=0\" " 5487 "to /boot/loader.conf.\n"); 5488 zfs_prefetch_disable = 1; 5489 } 5490#endif 5491 /* Warn about ZFS memory and address space requirements. */ 5492 if (((uint64_t)physmem * PAGESIZE) < (256 + 128 + 64) * (1 << 20)) { 5493 printf("ZFS WARNING: Recommended minimum RAM size is 512MB; " 5494 "expect unstable behavior.\n"); 5495 } 5496 if (kmem_size() < 512 * (1 << 20)) { 5497 printf("ZFS WARNING: Recommended minimum kmem_size is 512MB; " 5498 "expect unstable behavior.\n"); 5499 printf(" Consider tuning vm.kmem_size and " 5500 "vm.kmem_size_max\n"); 5501 printf(" in /boot/loader.conf.\n"); 5502 } 5503#endif 5504} 5505 5506void 5507arc_fini(void) 5508{ 5509 mutex_enter(&arc_reclaim_lock); 5510 arc_reclaim_thread_exit = TRUE; 5511 /* 5512 * The reclaim thread will set arc_reclaim_thread_exit back to 5513 * FALSE when it is finished exiting; we're waiting for that. 5514 */ 5515 while (arc_reclaim_thread_exit) { 5516 cv_signal(&arc_reclaim_thread_cv); 5517 cv_wait(&arc_reclaim_thread_cv, &arc_reclaim_lock); 5518 } 5519 mutex_exit(&arc_reclaim_lock); 5520 5521 mutex_enter(&arc_user_evicts_lock); 5522 arc_user_evicts_thread_exit = TRUE; 5523 /* 5524 * The user evicts thread will set arc_user_evicts_thread_exit 5525 * to FALSE when it is finished exiting; we're waiting for that. 5526 */ 5527 while (arc_user_evicts_thread_exit) { 5528 cv_signal(&arc_user_evicts_cv); 5529 cv_wait(&arc_user_evicts_cv, &arc_user_evicts_lock); 5530 } 5531 mutex_exit(&arc_user_evicts_lock); 5532 5533 /* Use TRUE to ensure *all* buffers are evicted */ 5534 arc_flush(NULL, TRUE); 5535 5536 arc_dead = TRUE; 5537 5538 if (arc_ksp != NULL) { 5539 kstat_delete(arc_ksp); 5540 arc_ksp = NULL; 5541 } 5542 5543 mutex_destroy(&arc_reclaim_lock); 5544 cv_destroy(&arc_reclaim_thread_cv); 5545 cv_destroy(&arc_reclaim_waiters_cv); 5546 5547 mutex_destroy(&arc_user_evicts_lock); 5548 cv_destroy(&arc_user_evicts_cv); 5549 5550 refcount_destroy(&arc_anon->arcs_size); 5551 refcount_destroy(&arc_mru->arcs_size); 5552 refcount_destroy(&arc_mru_ghost->arcs_size); 5553 refcount_destroy(&arc_mfu->arcs_size); 5554 refcount_destroy(&arc_mfu_ghost->arcs_size); 5555 refcount_destroy(&arc_l2c_only->arcs_size); 5556 5557 multilist_destroy(&arc_mru->arcs_list[ARC_BUFC_METADATA]); 5558 multilist_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA]); 5559 multilist_destroy(&arc_mfu->arcs_list[ARC_BUFC_METADATA]); 5560 multilist_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA]); 5561 multilist_destroy(&arc_mru->arcs_list[ARC_BUFC_DATA]); 5562 multilist_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA]); 5563 multilist_destroy(&arc_mfu->arcs_list[ARC_BUFC_DATA]); 5564 multilist_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA]); 5565 5566 buf_fini(); 5567 5568 ASSERT0(arc_loaned_bytes); 5569 5570#ifdef _KERNEL 5571 if (arc_event_lowmem != NULL) 5572 EVENTHANDLER_DEREGISTER(vm_lowmem, arc_event_lowmem); 5573#endif 5574} 5575 5576/* 5577 * Level 2 ARC 5578 * 5579 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk. 5580 * It uses dedicated storage devices to hold cached data, which are populated 5581 * using large infrequent writes. The main role of this cache is to boost 5582 * the performance of random read workloads. The intended L2ARC devices 5583 * include short-stroked disks, solid state disks, and other media with 5584 * substantially faster read latency than disk. 5585 * 5586 * +-----------------------+ 5587 * | ARC | 5588 * +-----------------------+ 5589 * | ^ ^ 5590 * | | | 5591 * l2arc_feed_thread() arc_read() 5592 * | | | 5593 * | l2arc read | 5594 * V | | 5595 * +---------------+ | 5596 * | L2ARC | | 5597 * +---------------+ | 5598 * | ^ | 5599 * l2arc_write() | | 5600 * | | | 5601 * V | | 5602 * +-------+ +-------+ 5603 * | vdev | | vdev | 5604 * | cache | | cache | 5605 * +-------+ +-------+ 5606 * +=========+ .-----. 5607 * : L2ARC : |-_____-| 5608 * : devices : | Disks | 5609 * +=========+ `-_____-' 5610 * 5611 * Read requests are satisfied from the following sources, in order: 5612 * 5613 * 1) ARC 5614 * 2) vdev cache of L2ARC devices 5615 * 3) L2ARC devices 5616 * 4) vdev cache of disks 5617 * 5) disks 5618 * 5619 * Some L2ARC device types exhibit extremely slow write performance. 5620 * To accommodate for this there are some significant differences between 5621 * the L2ARC and traditional cache design: 5622 * 5623 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from 5624 * the ARC behave as usual, freeing buffers and placing headers on ghost 5625 * lists. The ARC does not send buffers to the L2ARC during eviction as 5626 * this would add inflated write latencies for all ARC memory pressure. 5627 * 5628 * 2. The L2ARC attempts to cache data from the ARC before it is evicted. 5629 * It does this by periodically scanning buffers from the eviction-end of 5630 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are 5631 * not already there. It scans until a headroom of buffers is satisfied, 5632 * which itself is a buffer for ARC eviction. If a compressible buffer is 5633 * found during scanning and selected for writing to an L2ARC device, we 5634 * temporarily boost scanning headroom during the next scan cycle to make 5635 * sure we adapt to compression effects (which might significantly reduce 5636 * the data volume we write to L2ARC). The thread that does this is 5637 * l2arc_feed_thread(), illustrated below; example sizes are included to 5638 * provide a better sense of ratio than this diagram: 5639 * 5640 * head --> tail 5641 * +---------------------+----------+ 5642 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC 5643 * +---------------------+----------+ | o L2ARC eligible 5644 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer 5645 * +---------------------+----------+ | 5646 * 15.9 Gbytes ^ 32 Mbytes | 5647 * headroom | 5648 * l2arc_feed_thread() 5649 * | 5650 * l2arc write hand <--[oooo]--' 5651 * | 8 Mbyte 5652 * | write max 5653 * V 5654 * +==============================+ 5655 * L2ARC dev |####|#|###|###| |####| ... | 5656 * +==============================+ 5657 * 32 Gbytes 5658 * 5659 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of 5660 * evicted, then the L2ARC has cached a buffer much sooner than it probably 5661 * needed to, potentially wasting L2ARC device bandwidth and storage. It is 5662 * safe to say that this is an uncommon case, since buffers at the end of 5663 * the ARC lists have moved there due to inactivity. 5664 * 5665 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom, 5666 * then the L2ARC simply misses copying some buffers. This serves as a 5667 * pressure valve to prevent heavy read workloads from both stalling the ARC 5668 * with waits and clogging the L2ARC with writes. This also helps prevent 5669 * the potential for the L2ARC to churn if it attempts to cache content too 5670 * quickly, such as during backups of the entire pool. 5671 * 5672 * 5. After system boot and before the ARC has filled main memory, there are 5673 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru 5674 * lists can remain mostly static. Instead of searching from tail of these 5675 * lists as pictured, the l2arc_feed_thread() will search from the list heads 5676 * for eligible buffers, greatly increasing its chance of finding them. 5677 * 5678 * The L2ARC device write speed is also boosted during this time so that 5679 * the L2ARC warms up faster. Since there have been no ARC evictions yet, 5680 * there are no L2ARC reads, and no fear of degrading read performance 5681 * through increased writes. 5682 * 5683 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that 5684 * the vdev queue can aggregate them into larger and fewer writes. Each 5685 * device is written to in a rotor fashion, sweeping writes through 5686 * available space then repeating. 5687 * 5688 * 7. The L2ARC does not store dirty content. It never needs to flush 5689 * write buffers back to disk based storage. 5690 * 5691 * 8. If an ARC buffer is written (and dirtied) which also exists in the 5692 * L2ARC, the now stale L2ARC buffer is immediately dropped. 5693 * 5694 * The performance of the L2ARC can be tweaked by a number of tunables, which 5695 * may be necessary for different workloads: 5696 * 5697 * l2arc_write_max max write bytes per interval 5698 * l2arc_write_boost extra write bytes during device warmup 5699 * l2arc_noprefetch skip caching prefetched buffers 5700 * l2arc_headroom number of max device writes to precache 5701 * l2arc_headroom_boost when we find compressed buffers during ARC 5702 * scanning, we multiply headroom by this 5703 * percentage factor for the next scan cycle, 5704 * since more compressed buffers are likely to 5705 * be present 5706 * l2arc_feed_secs seconds between L2ARC writing 5707 * 5708 * Tunables may be removed or added as future performance improvements are 5709 * integrated, and also may become zpool properties. 5710 * 5711 * There are three key functions that control how the L2ARC warms up: 5712 * 5713 * l2arc_write_eligible() check if a buffer is eligible to cache 5714 * l2arc_write_size() calculate how much to write 5715 * l2arc_write_interval() calculate sleep delay between writes 5716 * 5717 * These three functions determine what to write, how much, and how quickly 5718 * to send writes. 5719 */ 5720 5721static boolean_t 5722l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *hdr) 5723{ 5724 /* 5725 * A buffer is *not* eligible for the L2ARC if it: 5726 * 1. belongs to a different spa. 5727 * 2. is already cached on the L2ARC. 5728 * 3. has an I/O in progress (it may be an incomplete read). 5729 * 4. is flagged not eligible (zfs property). 5730 */ 5731 if (hdr->b_spa != spa_guid) { 5732 ARCSTAT_BUMP(arcstat_l2_write_spa_mismatch); 5733 return (B_FALSE); 5734 } 5735 if (HDR_HAS_L2HDR(hdr)) { 5736 ARCSTAT_BUMP(arcstat_l2_write_in_l2); 5737 return (B_FALSE); 5738 } 5739 if (HDR_IO_IN_PROGRESS(hdr)) { 5740 ARCSTAT_BUMP(arcstat_l2_write_hdr_io_in_progress); 5741 return (B_FALSE); 5742 } 5743 if (!HDR_L2CACHE(hdr)) { 5744 ARCSTAT_BUMP(arcstat_l2_write_not_cacheable); 5745 return (B_FALSE); 5746 } 5747 5748 return (B_TRUE); 5749} 5750 5751static uint64_t 5752l2arc_write_size(void) 5753{ 5754 uint64_t size; 5755 5756 /* 5757 * Make sure our globals have meaningful values in case the user 5758 * altered them. 5759 */ 5760 size = l2arc_write_max; 5761 if (size == 0) { 5762 cmn_err(CE_NOTE, "Bad value for l2arc_write_max, value must " 5763 "be greater than zero, resetting it to the default (%d)", 5764 L2ARC_WRITE_SIZE); 5765 size = l2arc_write_max = L2ARC_WRITE_SIZE; 5766 } 5767 5768 if (arc_warm == B_FALSE) 5769 size += l2arc_write_boost; 5770 5771 return (size); 5772 5773} 5774 5775static clock_t 5776l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote) 5777{ 5778 clock_t interval, next, now; 5779 5780 /* 5781 * If the ARC lists are busy, increase our write rate; if the 5782 * lists are stale, idle back. This is achieved by checking 5783 * how much we previously wrote - if it was more than half of 5784 * what we wanted, schedule the next write much sooner. 5785 */ 5786 if (l2arc_feed_again && wrote > (wanted / 2)) 5787 interval = (hz * l2arc_feed_min_ms) / 1000; 5788 else 5789 interval = hz * l2arc_feed_secs; 5790 5791 now = ddi_get_lbolt(); 5792 next = MAX(now, MIN(now + interval, began + interval)); 5793 5794 return (next); 5795} 5796 5797/* 5798 * Cycle through L2ARC devices. This is how L2ARC load balances. 5799 * If a device is returned, this also returns holding the spa config lock. 5800 */ 5801static l2arc_dev_t * 5802l2arc_dev_get_next(void) 5803{ 5804 l2arc_dev_t *first, *next = NULL; 5805 5806 /* 5807 * Lock out the removal of spas (spa_namespace_lock), then removal 5808 * of cache devices (l2arc_dev_mtx). Once a device has been selected, 5809 * both locks will be dropped and a spa config lock held instead. 5810 */ 5811 mutex_enter(&spa_namespace_lock); 5812 mutex_enter(&l2arc_dev_mtx); 5813 5814 /* if there are no vdevs, there is nothing to do */ 5815 if (l2arc_ndev == 0) 5816 goto out; 5817 5818 first = NULL; 5819 next = l2arc_dev_last; 5820 do { 5821 /* loop around the list looking for a non-faulted vdev */ 5822 if (next == NULL) { 5823 next = list_head(l2arc_dev_list); 5824 } else { 5825 next = list_next(l2arc_dev_list, next); 5826 if (next == NULL) 5827 next = list_head(l2arc_dev_list); 5828 } 5829 5830 /* if we have come back to the start, bail out */ 5831 if (first == NULL) 5832 first = next; 5833 else if (next == first) 5834 break; 5835 5836 } while (vdev_is_dead(next->l2ad_vdev)); 5837 5838 /* if we were unable to find any usable vdevs, return NULL */ 5839 if (vdev_is_dead(next->l2ad_vdev)) 5840 next = NULL; 5841 5842 l2arc_dev_last = next; 5843 5844out: 5845 mutex_exit(&l2arc_dev_mtx); 5846 5847 /* 5848 * Grab the config lock to prevent the 'next' device from being 5849 * removed while we are writing to it. 5850 */ 5851 if (next != NULL) 5852 spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER); 5853 mutex_exit(&spa_namespace_lock); 5854 5855 return (next); 5856} 5857 5858/* 5859 * Free buffers that were tagged for destruction. 5860 */ 5861static void 5862l2arc_do_free_on_write() 5863{ 5864 list_t *buflist; 5865 l2arc_data_free_t *df, *df_prev; 5866 5867 mutex_enter(&l2arc_free_on_write_mtx); 5868 buflist = l2arc_free_on_write; 5869 5870 for (df = list_tail(buflist); df; df = df_prev) { 5871 df_prev = list_prev(buflist, df); 5872 ASSERT(df->l2df_data != NULL); 5873 ASSERT(df->l2df_func != NULL); 5874 df->l2df_func(df->l2df_data, df->l2df_size); 5875 list_remove(buflist, df); 5876 kmem_free(df, sizeof (l2arc_data_free_t)); 5877 } 5878 5879 mutex_exit(&l2arc_free_on_write_mtx); 5880} 5881 5882/* 5883 * A write to a cache device has completed. Update all headers to allow 5884 * reads from these buffers to begin. 5885 */ 5886static void 5887l2arc_write_done(zio_t *zio) 5888{ 5889 l2arc_write_callback_t *cb; 5890 l2arc_dev_t *dev; 5891 list_t *buflist; 5892 arc_buf_hdr_t *head, *hdr, *hdr_prev; 5893 kmutex_t *hash_lock; 5894 int64_t bytes_dropped = 0; 5895 5896 cb = zio->io_private; 5897 ASSERT(cb != NULL); 5898 dev = cb->l2wcb_dev; 5899 ASSERT(dev != NULL); 5900 head = cb->l2wcb_head; 5901 ASSERT(head != NULL); 5902 buflist = &dev->l2ad_buflist; 5903 ASSERT(buflist != NULL); 5904 DTRACE_PROBE2(l2arc__iodone, zio_t *, zio, 5905 l2arc_write_callback_t *, cb); 5906 5907 if (zio->io_error != 0) 5908 ARCSTAT_BUMP(arcstat_l2_writes_error); 5909 5910 /* 5911 * All writes completed, or an error was hit. 5912 */ 5913top: 5914 mutex_enter(&dev->l2ad_mtx); 5915 for (hdr = list_prev(buflist, head); hdr; hdr = hdr_prev) { 5916 hdr_prev = list_prev(buflist, hdr); 5917 5918 hash_lock = HDR_LOCK(hdr); 5919 5920 /* 5921 * We cannot use mutex_enter or else we can deadlock 5922 * with l2arc_write_buffers (due to swapping the order 5923 * the hash lock and l2ad_mtx are taken). 5924 */ 5925 if (!mutex_tryenter(hash_lock)) { 5926 /* 5927 * Missed the hash lock. We must retry so we 5928 * don't leave the ARC_FLAG_L2_WRITING bit set. 5929 */ 5930 ARCSTAT_BUMP(arcstat_l2_writes_lock_retry); 5931 5932 /* 5933 * We don't want to rescan the headers we've 5934 * already marked as having been written out, so 5935 * we reinsert the head node so we can pick up 5936 * where we left off. 5937 */ 5938 list_remove(buflist, head); 5939 list_insert_after(buflist, hdr, head); 5940 5941 mutex_exit(&dev->l2ad_mtx); 5942 5943 /* 5944 * We wait for the hash lock to become available 5945 * to try and prevent busy waiting, and increase 5946 * the chance we'll be able to acquire the lock 5947 * the next time around. 5948 */ 5949 mutex_enter(hash_lock); 5950 mutex_exit(hash_lock); 5951 goto top; 5952 } 5953 5954 /* 5955 * We could not have been moved into the arc_l2c_only 5956 * state while in-flight due to our ARC_FLAG_L2_WRITING 5957 * bit being set. Let's just ensure that's being enforced. 5958 */ 5959 ASSERT(HDR_HAS_L1HDR(hdr)); 5960 5961 /* 5962 * We may have allocated a buffer for L2ARC compression, 5963 * we must release it to avoid leaking this data. 5964 */ 5965 l2arc_release_cdata_buf(hdr); 5966 5967 if (zio->io_error != 0) { 5968 /* 5969 * Error - drop L2ARC entry. 5970 */ 5971 list_remove(buflist, hdr); 5972 trim_map_free(hdr->b_l2hdr.b_dev->l2ad_vdev, 5973 hdr->b_l2hdr.b_daddr, hdr->b_l2hdr.b_asize, 0); 5974 hdr->b_flags &= ~ARC_FLAG_HAS_L2HDR; 5975 5976 ARCSTAT_INCR(arcstat_l2_asize, -hdr->b_l2hdr.b_asize); 5977 ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size); 5978 5979 bytes_dropped += hdr->b_l2hdr.b_asize; 5980 (void) refcount_remove_many(&dev->l2ad_alloc, 5981 hdr->b_l2hdr.b_asize, hdr); 5982 } 5983 5984 /* 5985 * Allow ARC to begin reads and ghost list evictions to 5986 * this L2ARC entry. 5987 */ 5988 hdr->b_flags &= ~ARC_FLAG_L2_WRITING; 5989 5990 mutex_exit(hash_lock); 5991 } 5992 5993 atomic_inc_64(&l2arc_writes_done); 5994 list_remove(buflist, head); 5995 ASSERT(!HDR_HAS_L1HDR(head)); 5996 kmem_cache_free(hdr_l2only_cache, head); 5997 mutex_exit(&dev->l2ad_mtx); 5998 5999 vdev_space_update(dev->l2ad_vdev, -bytes_dropped, 0, 0); 6000 6001 l2arc_do_free_on_write(); 6002 6003 kmem_free(cb, sizeof (l2arc_write_callback_t)); 6004} 6005 6006/* 6007 * A read to a cache device completed. Validate buffer contents before 6008 * handing over to the regular ARC routines. 6009 */ 6010static void 6011l2arc_read_done(zio_t *zio) 6012{ 6013 l2arc_read_callback_t *cb; 6014 arc_buf_hdr_t *hdr; 6015 arc_buf_t *buf; 6016 kmutex_t *hash_lock; 6017 int equal; 6018 6019 ASSERT(zio->io_vd != NULL); 6020 ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE); 6021 6022 spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd); 6023 6024 cb = zio->io_private; 6025 ASSERT(cb != NULL); 6026 buf = cb->l2rcb_buf; 6027 ASSERT(buf != NULL); 6028 6029 hash_lock = HDR_LOCK(buf->b_hdr); 6030 mutex_enter(hash_lock); 6031 hdr = buf->b_hdr; 6032 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 6033 6034 /* 6035 * If the buffer was compressed, decompress it first. 6036 */ 6037 if (cb->l2rcb_compress != ZIO_COMPRESS_OFF) 6038 l2arc_decompress_zio(zio, hdr, cb->l2rcb_compress); 6039 ASSERT(zio->io_data != NULL); 6040 ASSERT3U(zio->io_size, ==, hdr->b_size); 6041 ASSERT3U(BP_GET_LSIZE(&cb->l2rcb_bp), ==, hdr->b_size); 6042 6043 /* 6044 * Check this survived the L2ARC journey. 6045 */ 6046 equal = arc_cksum_equal(buf); 6047 if (equal && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) { 6048 mutex_exit(hash_lock); 6049 zio->io_private = buf; 6050 zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */ 6051 zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */ 6052 arc_read_done(zio); 6053 } else { 6054 mutex_exit(hash_lock); 6055 /* 6056 * Buffer didn't survive caching. Increment stats and 6057 * reissue to the original storage device. 6058 */ 6059 if (zio->io_error != 0) { 6060 ARCSTAT_BUMP(arcstat_l2_io_error); 6061 } else { 6062 zio->io_error = SET_ERROR(EIO); 6063 } 6064 if (!equal) 6065 ARCSTAT_BUMP(arcstat_l2_cksum_bad); 6066 6067 /* 6068 * If there's no waiter, issue an async i/o to the primary 6069 * storage now. If there *is* a waiter, the caller must 6070 * issue the i/o in a context where it's OK to block. 6071 */ 6072 if (zio->io_waiter == NULL) { 6073 zio_t *pio = zio_unique_parent(zio); 6074 6075 ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL); 6076 6077 zio_nowait(zio_read(pio, cb->l2rcb_spa, &cb->l2rcb_bp, 6078 buf->b_data, hdr->b_size, arc_read_done, buf, 6079 zio->io_priority, cb->l2rcb_flags, &cb->l2rcb_zb)); 6080 } 6081 } 6082 6083 kmem_free(cb, sizeof (l2arc_read_callback_t)); 6084} 6085 6086/* 6087 * This is the list priority from which the L2ARC will search for pages to 6088 * cache. This is used within loops (0..3) to cycle through lists in the 6089 * desired order. This order can have a significant effect on cache 6090 * performance. 6091 * 6092 * Currently the metadata lists are hit first, MFU then MRU, followed by 6093 * the data lists. This function returns a locked list, and also returns 6094 * the lock pointer. 6095 */ 6096static multilist_sublist_t * 6097l2arc_sublist_lock(int list_num) 6098{ 6099 multilist_t *ml = NULL; 6100 unsigned int idx; 6101 6102 ASSERT(list_num >= 0 && list_num <= 3); 6103 6104 switch (list_num) { 6105 case 0: 6106 ml = &arc_mfu->arcs_list[ARC_BUFC_METADATA]; 6107 break; 6108 case 1: 6109 ml = &arc_mru->arcs_list[ARC_BUFC_METADATA]; 6110 break; 6111 case 2: 6112 ml = &arc_mfu->arcs_list[ARC_BUFC_DATA]; 6113 break; 6114 case 3: 6115 ml = &arc_mru->arcs_list[ARC_BUFC_DATA]; 6116 break; 6117 } 6118 6119 /* 6120 * Return a randomly-selected sublist. This is acceptable 6121 * because the caller feeds only a little bit of data for each 6122 * call (8MB). Subsequent calls will result in different 6123 * sublists being selected. 6124 */ 6125 idx = multilist_get_random_index(ml); 6126 return (multilist_sublist_lock(ml, idx)); 6127} 6128 6129/* 6130 * Evict buffers from the device write hand to the distance specified in 6131 * bytes. This distance may span populated buffers, it may span nothing. 6132 * This is clearing a region on the L2ARC device ready for writing. 6133 * If the 'all' boolean is set, every buffer is evicted. 6134 */ 6135static void 6136l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all) 6137{ 6138 list_t *buflist; 6139 arc_buf_hdr_t *hdr, *hdr_prev; 6140 kmutex_t *hash_lock; 6141 uint64_t taddr; 6142 6143 buflist = &dev->l2ad_buflist; 6144 6145 if (!all && dev->l2ad_first) { 6146 /* 6147 * This is the first sweep through the device. There is 6148 * nothing to evict. 6149 */ 6150 return; 6151 } 6152 6153 if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) { 6154 /* 6155 * When nearing the end of the device, evict to the end 6156 * before the device write hand jumps to the start. 6157 */ 6158 taddr = dev->l2ad_end; 6159 } else { 6160 taddr = dev->l2ad_hand + distance; 6161 } 6162 DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist, 6163 uint64_t, taddr, boolean_t, all); 6164 6165top: 6166 mutex_enter(&dev->l2ad_mtx); 6167 for (hdr = list_tail(buflist); hdr; hdr = hdr_prev) { 6168 hdr_prev = list_prev(buflist, hdr); 6169 6170 hash_lock = HDR_LOCK(hdr); 6171 6172 /* 6173 * We cannot use mutex_enter or else we can deadlock 6174 * with l2arc_write_buffers (due to swapping the order 6175 * the hash lock and l2ad_mtx are taken). 6176 */ 6177 if (!mutex_tryenter(hash_lock)) { 6178 /* 6179 * Missed the hash lock. Retry. 6180 */ 6181 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry); 6182 mutex_exit(&dev->l2ad_mtx); 6183 mutex_enter(hash_lock); 6184 mutex_exit(hash_lock); 6185 goto top; 6186 } 6187 6188 if (HDR_L2_WRITE_HEAD(hdr)) { 6189 /* 6190 * We hit a write head node. Leave it for 6191 * l2arc_write_done(). 6192 */ 6193 list_remove(buflist, hdr); 6194 mutex_exit(hash_lock); 6195 continue; 6196 } 6197 6198 if (!all && HDR_HAS_L2HDR(hdr) && 6199 (hdr->b_l2hdr.b_daddr > taddr || 6200 hdr->b_l2hdr.b_daddr < dev->l2ad_hand)) { 6201 /* 6202 * We've evicted to the target address, 6203 * or the end of the device. 6204 */ 6205 mutex_exit(hash_lock); 6206 break; 6207 } 6208 6209 ASSERT(HDR_HAS_L2HDR(hdr)); 6210 if (!HDR_HAS_L1HDR(hdr)) { 6211 ASSERT(!HDR_L2_READING(hdr)); 6212 /* 6213 * This doesn't exist in the ARC. Destroy. 6214 * arc_hdr_destroy() will call list_remove() 6215 * and decrement arcstat_l2_size. 6216 */ 6217 arc_change_state(arc_anon, hdr, hash_lock); 6218 arc_hdr_destroy(hdr); 6219 } else { 6220 ASSERT(hdr->b_l1hdr.b_state != arc_l2c_only); 6221 ARCSTAT_BUMP(arcstat_l2_evict_l1cached); 6222 /* 6223 * Invalidate issued or about to be issued 6224 * reads, since we may be about to write 6225 * over this location. 6226 */ 6227 if (HDR_L2_READING(hdr)) { 6228 ARCSTAT_BUMP(arcstat_l2_evict_reading); 6229 hdr->b_flags |= ARC_FLAG_L2_EVICTED; 6230 } 6231 6232 /* Ensure this header has finished being written */ 6233 ASSERT(!HDR_L2_WRITING(hdr)); 6234 ASSERT3P(hdr->b_l1hdr.b_tmp_cdata, ==, NULL); 6235 6236 arc_hdr_l2hdr_destroy(hdr); 6237 } 6238 mutex_exit(hash_lock); 6239 } 6240 mutex_exit(&dev->l2ad_mtx); 6241} 6242 6243/* 6244 * Find and write ARC buffers to the L2ARC device. 6245 * 6246 * An ARC_FLAG_L2_WRITING flag is set so that the L2ARC buffers are not valid 6247 * for reading until they have completed writing. 6248 * The headroom_boost is an in-out parameter used to maintain headroom boost 6249 * state between calls to this function. 6250 * 6251 * Returns the number of bytes actually written (which may be smaller than 6252 * the delta by which the device hand has changed due to alignment). 6253 */ 6254static uint64_t 6255l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz, 6256 boolean_t *headroom_boost) 6257{ 6258 arc_buf_hdr_t *hdr, *hdr_prev, *head; 6259 uint64_t write_asize, write_sz, headroom, 6260 buf_compress_minsz; 6261 void *buf_data; 6262 boolean_t full; 6263 l2arc_write_callback_t *cb; 6264 zio_t *pio, *wzio; 6265 uint64_t guid = spa_load_guid(spa); 6266 const boolean_t do_headroom_boost = *headroom_boost; 6267 int try; 6268 6269 ASSERT(dev->l2ad_vdev != NULL); 6270 6271 /* Lower the flag now, we might want to raise it again later. */ 6272 *headroom_boost = B_FALSE; 6273 6274 pio = NULL; 6275 write_sz = write_asize = 0; 6276 full = B_FALSE; 6277 head = kmem_cache_alloc(hdr_l2only_cache, KM_PUSHPAGE); 6278 head->b_flags |= ARC_FLAG_L2_WRITE_HEAD; 6279 head->b_flags |= ARC_FLAG_HAS_L2HDR; 6280 6281 ARCSTAT_BUMP(arcstat_l2_write_buffer_iter); 6282 /* 6283 * We will want to try to compress buffers that are at least 2x the 6284 * device sector size. 6285 */ 6286 buf_compress_minsz = 2 << dev->l2ad_vdev->vdev_ashift; 6287 6288 /* 6289 * Copy buffers for L2ARC writing. 6290 */ 6291 for (try = 0; try <= 3; try++) { 6292 multilist_sublist_t *mls = l2arc_sublist_lock(try); 6293 uint64_t passed_sz = 0; 6294 6295 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_iter); 6296 6297 /* 6298 * L2ARC fast warmup. 6299 * 6300 * Until the ARC is warm and starts to evict, read from the 6301 * head of the ARC lists rather than the tail. 6302 */ 6303 if (arc_warm == B_FALSE) 6304 hdr = multilist_sublist_head(mls); 6305 else 6306 hdr = multilist_sublist_tail(mls); 6307 if (hdr == NULL) 6308 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_null_iter); 6309 6310 headroom = target_sz * l2arc_headroom; 6311 if (do_headroom_boost) 6312 headroom = (headroom * l2arc_headroom_boost) / 100; 6313 6314 for (; hdr; hdr = hdr_prev) { 6315 kmutex_t *hash_lock; 6316 uint64_t buf_sz; 6317 uint64_t buf_a_sz; 6318 6319 if (arc_warm == B_FALSE) 6320 hdr_prev = multilist_sublist_next(mls, hdr); 6321 else 6322 hdr_prev = multilist_sublist_prev(mls, hdr); 6323 ARCSTAT_INCR(arcstat_l2_write_buffer_bytes_scanned, hdr->b_size); 6324 6325 hash_lock = HDR_LOCK(hdr); 6326 if (!mutex_tryenter(hash_lock)) { 6327 ARCSTAT_BUMP(arcstat_l2_write_trylock_fail); 6328 /* 6329 * Skip this buffer rather than waiting. 6330 */ 6331 continue; 6332 } 6333 6334 passed_sz += hdr->b_size; 6335 if (passed_sz > headroom) { 6336 /* 6337 * Searched too far. 6338 */ 6339 mutex_exit(hash_lock); 6340 ARCSTAT_BUMP(arcstat_l2_write_passed_headroom); 6341 break; 6342 } 6343 6344 if (!l2arc_write_eligible(guid, hdr)) { 6345 mutex_exit(hash_lock); 6346 continue; 6347 } 6348 6349 /* 6350 * Assume that the buffer is not going to be compressed 6351 * and could take more space on disk because of a larger 6352 * disk block size. 6353 */ 6354 buf_sz = hdr->b_size; 6355 buf_a_sz = vdev_psize_to_asize(dev->l2ad_vdev, buf_sz); 6356 6357 if ((write_asize + buf_a_sz) > target_sz) { 6358 full = B_TRUE; 6359 mutex_exit(hash_lock); 6360 ARCSTAT_BUMP(arcstat_l2_write_full); 6361 break; 6362 } 6363 6364 if (pio == NULL) { 6365 /* 6366 * Insert a dummy header on the buflist so 6367 * l2arc_write_done() can find where the 6368 * write buffers begin without searching. 6369 */ 6370 mutex_enter(&dev->l2ad_mtx); 6371 list_insert_head(&dev->l2ad_buflist, head); 6372 mutex_exit(&dev->l2ad_mtx); 6373 6374 cb = kmem_alloc( 6375 sizeof (l2arc_write_callback_t), KM_SLEEP); 6376 cb->l2wcb_dev = dev; 6377 cb->l2wcb_head = head; 6378 pio = zio_root(spa, l2arc_write_done, cb, 6379 ZIO_FLAG_CANFAIL); 6380 ARCSTAT_BUMP(arcstat_l2_write_pios); 6381 } 6382 6383 /* 6384 * Create and add a new L2ARC header. 6385 */ 6386 hdr->b_l2hdr.b_dev = dev; 6387 hdr->b_flags |= ARC_FLAG_L2_WRITING; 6388 /* 6389 * Temporarily stash the data buffer in b_tmp_cdata. 6390 * The subsequent write step will pick it up from 6391 * there. This is because can't access b_l1hdr.b_buf 6392 * without holding the hash_lock, which we in turn 6393 * can't access without holding the ARC list locks 6394 * (which we want to avoid during compression/writing). 6395 */ 6396 hdr->b_l2hdr.b_compress = ZIO_COMPRESS_OFF; 6397 hdr->b_l2hdr.b_asize = hdr->b_size; 6398 hdr->b_l1hdr.b_tmp_cdata = hdr->b_l1hdr.b_buf->b_data; 6399 6400 /* 6401 * Explicitly set the b_daddr field to a known 6402 * value which means "invalid address". This 6403 * enables us to differentiate which stage of 6404 * l2arc_write_buffers() the particular header 6405 * is in (e.g. this loop, or the one below). 6406 * ARC_FLAG_L2_WRITING is not enough to make 6407 * this distinction, and we need to know in 6408 * order to do proper l2arc vdev accounting in 6409 * arc_release() and arc_hdr_destroy(). 6410 * 6411 * Note, we can't use a new flag to distinguish 6412 * the two stages because we don't hold the 6413 * header's hash_lock below, in the second stage 6414 * of this function. Thus, we can't simply 6415 * change the b_flags field to denote that the 6416 * IO has been sent. We can change the b_daddr 6417 * field of the L2 portion, though, since we'll 6418 * be holding the l2ad_mtx; which is why we're 6419 * using it to denote the header's state change. 6420 */ 6421 hdr->b_l2hdr.b_daddr = L2ARC_ADDR_UNSET; 6422 6423 hdr->b_flags |= ARC_FLAG_HAS_L2HDR; 6424 6425 mutex_enter(&dev->l2ad_mtx); 6426 list_insert_head(&dev->l2ad_buflist, hdr); 6427 mutex_exit(&dev->l2ad_mtx); 6428 6429 /* 6430 * Compute and store the buffer cksum before 6431 * writing. On debug the cksum is verified first. 6432 */ 6433 arc_cksum_verify(hdr->b_l1hdr.b_buf); 6434 arc_cksum_compute(hdr->b_l1hdr.b_buf, B_TRUE); 6435 6436 mutex_exit(hash_lock); 6437 6438 write_sz += buf_sz; 6439 write_asize += buf_a_sz; 6440 } 6441 6442 multilist_sublist_unlock(mls); 6443 6444 if (full == B_TRUE) 6445 break; 6446 } 6447 6448 /* No buffers selected for writing? */ 6449 if (pio == NULL) { 6450 ASSERT0(write_sz); 6451 ASSERT(!HDR_HAS_L1HDR(head)); 6452 kmem_cache_free(hdr_l2only_cache, head); 6453 return (0); 6454 } 6455 6456 mutex_enter(&dev->l2ad_mtx); 6457 6458 /* 6459 * Note that elsewhere in this file arcstat_l2_asize 6460 * and the used space on l2ad_vdev are updated using b_asize, 6461 * which is not necessarily rounded up to the device block size. 6462 * Too keep accounting consistent we do the same here as well: 6463 * stats_size accumulates the sum of b_asize of the written buffers, 6464 * while write_asize accumulates the sum of b_asize rounded up 6465 * to the device block size. 6466 * The latter sum is used only to validate the corectness of the code. 6467 */ 6468 uint64_t stats_size = 0; 6469 write_asize = 0; 6470 6471 /* 6472 * Now start writing the buffers. We're starting at the write head 6473 * and work backwards, retracing the course of the buffer selector 6474 * loop above. 6475 */ 6476 for (hdr = list_prev(&dev->l2ad_buflist, head); hdr; 6477 hdr = list_prev(&dev->l2ad_buflist, hdr)) { 6478 uint64_t buf_sz; 6479 6480 /* 6481 * We rely on the L1 portion of the header below, so 6482 * it's invalid for this header to have been evicted out 6483 * of the ghost cache, prior to being written out. The 6484 * ARC_FLAG_L2_WRITING bit ensures this won't happen. 6485 */ 6486 ASSERT(HDR_HAS_L1HDR(hdr)); 6487 6488 /* 6489 * We shouldn't need to lock the buffer here, since we flagged 6490 * it as ARC_FLAG_L2_WRITING in the previous step, but we must 6491 * take care to only access its L2 cache parameters. In 6492 * particular, hdr->l1hdr.b_buf may be invalid by now due to 6493 * ARC eviction. 6494 */ 6495 hdr->b_l2hdr.b_daddr = dev->l2ad_hand; 6496 6497 if ((HDR_L2COMPRESS(hdr)) && 6498 hdr->b_l2hdr.b_asize >= buf_compress_minsz) { 6499 if (l2arc_compress_buf(hdr)) { 6500 /* 6501 * If compression succeeded, enable headroom 6502 * boost on the next scan cycle. 6503 */ 6504 *headroom_boost = B_TRUE; 6505 } 6506 } 6507 6508 /* 6509 * Pick up the buffer data we had previously stashed away 6510 * (and now potentially also compressed). 6511 */ 6512 buf_data = hdr->b_l1hdr.b_tmp_cdata; 6513 buf_sz = hdr->b_l2hdr.b_asize; 6514 6515 /* 6516 * We need to do this regardless if buf_sz is zero or 6517 * not, otherwise, when this l2hdr is evicted we'll 6518 * remove a reference that was never added. 6519 */ 6520 (void) refcount_add_many(&dev->l2ad_alloc, buf_sz, hdr); 6521 6522 /* Compression may have squashed the buffer to zero length. */ 6523 if (buf_sz != 0) { 6524 uint64_t buf_a_sz; 6525 6526 wzio = zio_write_phys(pio, dev->l2ad_vdev, 6527 dev->l2ad_hand, buf_sz, buf_data, ZIO_CHECKSUM_OFF, 6528 NULL, NULL, ZIO_PRIORITY_ASYNC_WRITE, 6529 ZIO_FLAG_CANFAIL, B_FALSE); 6530 6531 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev, 6532 zio_t *, wzio); 6533 (void) zio_nowait(wzio); 6534 6535 stats_size += buf_sz; 6536 6537 /* 6538 * Keep the clock hand suitably device-aligned. 6539 */ 6540 buf_a_sz = vdev_psize_to_asize(dev->l2ad_vdev, buf_sz); 6541 write_asize += buf_a_sz; 6542 dev->l2ad_hand += buf_a_sz; 6543 } 6544 } 6545 6546 mutex_exit(&dev->l2ad_mtx); 6547 6548 ASSERT3U(write_asize, <=, target_sz); 6549 ARCSTAT_BUMP(arcstat_l2_writes_sent); 6550 ARCSTAT_INCR(arcstat_l2_write_bytes, write_asize); 6551 ARCSTAT_INCR(arcstat_l2_size, write_sz); 6552 ARCSTAT_INCR(arcstat_l2_asize, stats_size); 6553 vdev_space_update(dev->l2ad_vdev, stats_size, 0, 0); 6554 6555 /* 6556 * Bump device hand to the device start if it is approaching the end. 6557 * l2arc_evict() will already have evicted ahead for this case. 6558 */ 6559 if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) { 6560 dev->l2ad_hand = dev->l2ad_start; 6561 dev->l2ad_first = B_FALSE; 6562 } 6563 6564 dev->l2ad_writing = B_TRUE; 6565 (void) zio_wait(pio); 6566 dev->l2ad_writing = B_FALSE; 6567 6568 return (write_asize); 6569} 6570 6571/* 6572 * Compresses an L2ARC buffer. 6573 * The data to be compressed must be prefilled in l1hdr.b_tmp_cdata and its 6574 * size in l2hdr->b_asize. This routine tries to compress the data and 6575 * depending on the compression result there are three possible outcomes: 6576 * *) The buffer was incompressible. The original l2hdr contents were left 6577 * untouched and are ready for writing to an L2 device. 6578 * *) The buffer was all-zeros, so there is no need to write it to an L2 6579 * device. To indicate this situation b_tmp_cdata is NULL'ed, b_asize is 6580 * set to zero and b_compress is set to ZIO_COMPRESS_EMPTY. 6581 * *) Compression succeeded and b_tmp_cdata was replaced with a temporary 6582 * data buffer which holds the compressed data to be written, and b_asize 6583 * tells us how much data there is. b_compress is set to the appropriate 6584 * compression algorithm. Once writing is done, invoke 6585 * l2arc_release_cdata_buf on this l2hdr to free this temporary buffer. 6586 * 6587 * Returns B_TRUE if compression succeeded, or B_FALSE if it didn't (the 6588 * buffer was incompressible). 6589 */ 6590static boolean_t 6591l2arc_compress_buf(arc_buf_hdr_t *hdr) 6592{ 6593 void *cdata; 6594 size_t csize, len, rounded; 6595 ASSERT(HDR_HAS_L2HDR(hdr)); 6596 l2arc_buf_hdr_t *l2hdr = &hdr->b_l2hdr; 6597 6598 ASSERT(HDR_HAS_L1HDR(hdr)); 6599 ASSERT3S(l2hdr->b_compress, ==, ZIO_COMPRESS_OFF); 6600 ASSERT(hdr->b_l1hdr.b_tmp_cdata != NULL); 6601 6602 len = l2hdr->b_asize; 6603 cdata = zio_data_buf_alloc(len); 6604 ASSERT3P(cdata, !=, NULL); 6605 csize = zio_compress_data(ZIO_COMPRESS_LZ4, hdr->b_l1hdr.b_tmp_cdata, 6606 cdata, l2hdr->b_asize); 6607 6608 if (csize == 0) { 6609 /* zero block, indicate that there's nothing to write */ 6610 zio_data_buf_free(cdata, len); 6611 l2hdr->b_compress = ZIO_COMPRESS_EMPTY; 6612 l2hdr->b_asize = 0; 6613 hdr->b_l1hdr.b_tmp_cdata = NULL; 6614 ARCSTAT_BUMP(arcstat_l2_compress_zeros); 6615 return (B_TRUE); 6616 } 6617 6618 rounded = P2ROUNDUP(csize, 6619 (size_t)1 << l2hdr->b_dev->l2ad_vdev->vdev_ashift); 6620 if (rounded < len) { 6621 /* 6622 * Compression succeeded, we'll keep the cdata around for 6623 * writing and release it afterwards. 6624 */ 6625 if (rounded > csize) { 6626 bzero((char *)cdata + csize, rounded - csize); 6627 csize = rounded; 6628 } 6629 l2hdr->b_compress = ZIO_COMPRESS_LZ4; 6630 l2hdr->b_asize = csize; 6631 hdr->b_l1hdr.b_tmp_cdata = cdata; 6632 ARCSTAT_BUMP(arcstat_l2_compress_successes); 6633 return (B_TRUE); 6634 } else { 6635 /* 6636 * Compression failed, release the compressed buffer. 6637 * l2hdr will be left unmodified. 6638 */ 6639 zio_data_buf_free(cdata, len); 6640 ARCSTAT_BUMP(arcstat_l2_compress_failures); 6641 return (B_FALSE); 6642 } 6643} 6644 6645/* 6646 * Decompresses a zio read back from an l2arc device. On success, the 6647 * underlying zio's io_data buffer is overwritten by the uncompressed 6648 * version. On decompression error (corrupt compressed stream), the 6649 * zio->io_error value is set to signal an I/O error. 6650 * 6651 * Please note that the compressed data stream is not checksummed, so 6652 * if the underlying device is experiencing data corruption, we may feed 6653 * corrupt data to the decompressor, so the decompressor needs to be 6654 * able to handle this situation (LZ4 does). 6655 */ 6656static void 6657l2arc_decompress_zio(zio_t *zio, arc_buf_hdr_t *hdr, enum zio_compress c) 6658{ 6659 ASSERT(L2ARC_IS_VALID_COMPRESS(c)); 6660 6661 if (zio->io_error != 0) { 6662 /* 6663 * An io error has occured, just restore the original io 6664 * size in preparation for a main pool read. 6665 */ 6666 zio->io_orig_size = zio->io_size = hdr->b_size; 6667 return; 6668 } 6669 6670 if (c == ZIO_COMPRESS_EMPTY) { 6671 /* 6672 * An empty buffer results in a null zio, which means we 6673 * need to fill its io_data after we're done restoring the 6674 * buffer's contents. 6675 */ 6676 ASSERT(hdr->b_l1hdr.b_buf != NULL); 6677 bzero(hdr->b_l1hdr.b_buf->b_data, hdr->b_size); 6678 zio->io_data = zio->io_orig_data = hdr->b_l1hdr.b_buf->b_data; 6679 } else { 6680 ASSERT(zio->io_data != NULL); 6681 /* 6682 * We copy the compressed data from the start of the arc buffer 6683 * (the zio_read will have pulled in only what we need, the 6684 * rest is garbage which we will overwrite at decompression) 6685 * and then decompress back to the ARC data buffer. This way we 6686 * can minimize copying by simply decompressing back over the 6687 * original compressed data (rather than decompressing to an 6688 * aux buffer and then copying back the uncompressed buffer, 6689 * which is likely to be much larger). 6690 */ 6691 uint64_t csize; 6692 void *cdata; 6693 6694 csize = zio->io_size; 6695 cdata = zio_data_buf_alloc(csize); 6696 bcopy(zio->io_data, cdata, csize); 6697 if (zio_decompress_data(c, cdata, zio->io_data, csize, 6698 hdr->b_size) != 0) 6699 zio->io_error = EIO; 6700 zio_data_buf_free(cdata, csize); 6701 } 6702 6703 /* Restore the expected uncompressed IO size. */ 6704 zio->io_orig_size = zio->io_size = hdr->b_size; 6705} 6706 6707/* 6708 * Releases the temporary b_tmp_cdata buffer in an l2arc header structure. 6709 * This buffer serves as a temporary holder of compressed data while 6710 * the buffer entry is being written to an l2arc device. Once that is 6711 * done, we can dispose of it. 6712 */ 6713static void 6714l2arc_release_cdata_buf(arc_buf_hdr_t *hdr) 6715{ 6716 ASSERT(HDR_HAS_L2HDR(hdr)); 6717 enum zio_compress comp = hdr->b_l2hdr.b_compress; 6718 6719 ASSERT(HDR_HAS_L1HDR(hdr)); 6720 ASSERT(comp == ZIO_COMPRESS_OFF || L2ARC_IS_VALID_COMPRESS(comp)); 6721 6722 if (comp == ZIO_COMPRESS_OFF) { 6723 /* 6724 * In this case, b_tmp_cdata points to the same buffer 6725 * as the arc_buf_t's b_data field. We don't want to 6726 * free it, since the arc_buf_t will handle that. 6727 */ 6728 hdr->b_l1hdr.b_tmp_cdata = NULL; 6729 } else if (comp == ZIO_COMPRESS_EMPTY) { 6730 /* 6731 * In this case, b_tmp_cdata was compressed to an empty 6732 * buffer, thus there's nothing to free and b_tmp_cdata 6733 * should have been set to NULL in l2arc_write_buffers(). 6734 */ 6735 ASSERT3P(hdr->b_l1hdr.b_tmp_cdata, ==, NULL); 6736 } else { 6737 /* 6738 * If the data was compressed, then we've allocated a 6739 * temporary buffer for it, so now we need to release it. 6740 */ 6741 ASSERT(hdr->b_l1hdr.b_tmp_cdata != NULL); 6742 zio_data_buf_free(hdr->b_l1hdr.b_tmp_cdata, 6743 hdr->b_size); 6744 hdr->b_l1hdr.b_tmp_cdata = NULL; 6745 } 6746 6747} 6748 6749/* 6750 * This thread feeds the L2ARC at regular intervals. This is the beating 6751 * heart of the L2ARC. 6752 */ 6753static void 6754l2arc_feed_thread(void *dummy __unused) 6755{ 6756 callb_cpr_t cpr; 6757 l2arc_dev_t *dev; 6758 spa_t *spa; 6759 uint64_t size, wrote; 6760 clock_t begin, next = ddi_get_lbolt(); 6761 boolean_t headroom_boost = B_FALSE; 6762 6763 CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG); 6764 6765 mutex_enter(&l2arc_feed_thr_lock); 6766 6767 while (l2arc_thread_exit == 0) { 6768 CALLB_CPR_SAFE_BEGIN(&cpr); 6769 (void) cv_timedwait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock, 6770 next - ddi_get_lbolt()); 6771 CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock); 6772 next = ddi_get_lbolt() + hz; 6773 6774 /* 6775 * Quick check for L2ARC devices. 6776 */ 6777 mutex_enter(&l2arc_dev_mtx); 6778 if (l2arc_ndev == 0) { 6779 mutex_exit(&l2arc_dev_mtx); 6780 continue; 6781 } 6782 mutex_exit(&l2arc_dev_mtx); 6783 begin = ddi_get_lbolt(); 6784 6785 /* 6786 * This selects the next l2arc device to write to, and in 6787 * doing so the next spa to feed from: dev->l2ad_spa. This 6788 * will return NULL if there are now no l2arc devices or if 6789 * they are all faulted. 6790 * 6791 * If a device is returned, its spa's config lock is also 6792 * held to prevent device removal. l2arc_dev_get_next() 6793 * will grab and release l2arc_dev_mtx. 6794 */ 6795 if ((dev = l2arc_dev_get_next()) == NULL) 6796 continue; 6797 6798 spa = dev->l2ad_spa; 6799 ASSERT(spa != NULL); 6800 6801 /* 6802 * If the pool is read-only then force the feed thread to 6803 * sleep a little longer. 6804 */ 6805 if (!spa_writeable(spa)) { 6806 next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz; 6807 spa_config_exit(spa, SCL_L2ARC, dev); 6808 continue; 6809 } 6810 6811 /* 6812 * Avoid contributing to memory pressure. 6813 */ 6814 if (arc_reclaim_needed()) { 6815 ARCSTAT_BUMP(arcstat_l2_abort_lowmem); 6816 spa_config_exit(spa, SCL_L2ARC, dev); 6817 continue; 6818 } 6819 6820 ARCSTAT_BUMP(arcstat_l2_feeds); 6821 6822 size = l2arc_write_size(); 6823 6824 /* 6825 * Evict L2ARC buffers that will be overwritten. 6826 */ 6827 l2arc_evict(dev, size, B_FALSE); 6828 6829 /* 6830 * Write ARC buffers. 6831 */ 6832 wrote = l2arc_write_buffers(spa, dev, size, &headroom_boost); 6833 6834 /* 6835 * Calculate interval between writes. 6836 */ 6837 next = l2arc_write_interval(begin, size, wrote); 6838 spa_config_exit(spa, SCL_L2ARC, dev); 6839 } 6840 6841 l2arc_thread_exit = 0; 6842 cv_broadcast(&l2arc_feed_thr_cv); 6843 CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */ 6844 thread_exit(); 6845} 6846 6847boolean_t 6848l2arc_vdev_present(vdev_t *vd) 6849{ 6850 l2arc_dev_t *dev; 6851 6852 mutex_enter(&l2arc_dev_mtx); 6853 for (dev = list_head(l2arc_dev_list); dev != NULL; 6854 dev = list_next(l2arc_dev_list, dev)) { 6855 if (dev->l2ad_vdev == vd) 6856 break; 6857 } 6858 mutex_exit(&l2arc_dev_mtx); 6859 6860 return (dev != NULL); 6861} 6862 6863/* 6864 * Add a vdev for use by the L2ARC. By this point the spa has already 6865 * validated the vdev and opened it. 6866 */ 6867void 6868l2arc_add_vdev(spa_t *spa, vdev_t *vd) 6869{ 6870 l2arc_dev_t *adddev; 6871 6872 ASSERT(!l2arc_vdev_present(vd)); 6873 6874 vdev_ashift_optimize(vd); 6875 6876 /* 6877 * Create a new l2arc device entry. 6878 */ 6879 adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP); 6880 adddev->l2ad_spa = spa; 6881 adddev->l2ad_vdev = vd; 6882 adddev->l2ad_start = VDEV_LABEL_START_SIZE; 6883 adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd); 6884 adddev->l2ad_hand = adddev->l2ad_start; 6885 adddev->l2ad_first = B_TRUE; 6886 adddev->l2ad_writing = B_FALSE; 6887 6888 mutex_init(&adddev->l2ad_mtx, NULL, MUTEX_DEFAULT, NULL); 6889 /* 6890 * This is a list of all ARC buffers that are still valid on the 6891 * device. 6892 */ 6893 list_create(&adddev->l2ad_buflist, sizeof (arc_buf_hdr_t), 6894 offsetof(arc_buf_hdr_t, b_l2hdr.b_l2node)); 6895 6896 vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand); 6897 refcount_create(&adddev->l2ad_alloc); 6898 6899 /* 6900 * Add device to global list 6901 */ 6902 mutex_enter(&l2arc_dev_mtx); 6903 list_insert_head(l2arc_dev_list, adddev); 6904 atomic_inc_64(&l2arc_ndev); 6905 mutex_exit(&l2arc_dev_mtx); 6906} 6907 6908/* 6909 * Remove a vdev from the L2ARC. 6910 */ 6911void 6912l2arc_remove_vdev(vdev_t *vd) 6913{ 6914 l2arc_dev_t *dev, *nextdev, *remdev = NULL; 6915 6916 /* 6917 * Find the device by vdev 6918 */ 6919 mutex_enter(&l2arc_dev_mtx); 6920 for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) { 6921 nextdev = list_next(l2arc_dev_list, dev); 6922 if (vd == dev->l2ad_vdev) { 6923 remdev = dev; 6924 break; 6925 } 6926 } 6927 ASSERT(remdev != NULL); 6928 6929 /* 6930 * Remove device from global list 6931 */ 6932 list_remove(l2arc_dev_list, remdev); 6933 l2arc_dev_last = NULL; /* may have been invalidated */ 6934 atomic_dec_64(&l2arc_ndev); 6935 mutex_exit(&l2arc_dev_mtx); 6936 6937 /* 6938 * Clear all buflists and ARC references. L2ARC device flush. 6939 */ 6940 l2arc_evict(remdev, 0, B_TRUE); 6941 list_destroy(&remdev->l2ad_buflist); 6942 mutex_destroy(&remdev->l2ad_mtx); 6943 refcount_destroy(&remdev->l2ad_alloc); 6944 kmem_free(remdev, sizeof (l2arc_dev_t)); 6945} 6946 6947void 6948l2arc_init(void) 6949{ 6950 l2arc_thread_exit = 0; 6951 l2arc_ndev = 0; 6952 l2arc_writes_sent = 0; 6953 l2arc_writes_done = 0; 6954 6955 mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL); 6956 cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL); 6957 mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL); 6958 mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL); 6959 6960 l2arc_dev_list = &L2ARC_dev_list; 6961 l2arc_free_on_write = &L2ARC_free_on_write; 6962 list_create(l2arc_dev_list, sizeof (l2arc_dev_t), 6963 offsetof(l2arc_dev_t, l2ad_node)); 6964 list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t), 6965 offsetof(l2arc_data_free_t, l2df_list_node)); 6966} 6967 6968void 6969l2arc_fini(void) 6970{ 6971 /* 6972 * This is called from dmu_fini(), which is called from spa_fini(); 6973 * Because of this, we can assume that all l2arc devices have 6974 * already been removed when the pools themselves were removed. 6975 */ 6976 6977 l2arc_do_free_on_write(); 6978 6979 mutex_destroy(&l2arc_feed_thr_lock); 6980 cv_destroy(&l2arc_feed_thr_cv); 6981 mutex_destroy(&l2arc_dev_mtx); 6982 mutex_destroy(&l2arc_free_on_write_mtx); 6983 6984 list_destroy(l2arc_dev_list); 6985 list_destroy(l2arc_free_on_write); 6986} 6987 6988void 6989l2arc_start(void) 6990{ 6991 if (!(spa_mode_global & FWRITE)) 6992 return; 6993 6994 (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0, 6995 TS_RUN, minclsyspri); 6996} 6997 6998void 6999l2arc_stop(void) 7000{ 7001 if (!(spa_mode_global & FWRITE)) 7002 return; 7003 7004 mutex_enter(&l2arc_feed_thr_lock); 7005 cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */ 7006 l2arc_thread_exit = 1; 7007 while (l2arc_thread_exit != 0) 7008 cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock); 7009 mutex_exit(&l2arc_feed_thr_lock); 7010} 7011