arc.c revision 301695
11590Srgrimes/* 250477Speter * CDDL HEADER START 31590Srgrimes * 41590Srgrimes * The contents of this file are subject to the terms of the 5221807Sstas * Common Development and Distribution License (the "License"). 6221807Sstas * You may not use this file except in compliance with the License. 71590Srgrimes * 8221807Sstas * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE 91590Srgrimes * or http://www.opensolaris.org/os/licensing. 10221807Sstas * See the License for the specific language governing permissions 11136744Sphk * and limitations under the License. 121590Srgrimes * 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 <machine/vmparam.h> 143 144#ifdef illumos 145#ifndef _KERNEL 146/* set with ZFS_DEBUG=watch, to enable watchpoints on frozen buffers */ 147boolean_t arc_watch = B_FALSE; 148int arc_procfd; 149#endif 150#endif /* illumos */ 151 152static kmutex_t arc_reclaim_lock; 153static kcondvar_t arc_reclaim_thread_cv; 154static boolean_t arc_reclaim_thread_exit; 155static kcondvar_t arc_reclaim_waiters_cv; 156 157static kmutex_t arc_user_evicts_lock; 158static kcondvar_t arc_user_evicts_cv; 159static boolean_t arc_user_evicts_thread_exit; 160 161uint_t arc_reduce_dnlc_percent = 3; 162 163/* 164 * The number of headers to evict in arc_evict_state_impl() before 165 * dropping the sublist lock and evicting from another sublist. A lower 166 * value means we're more likely to evict the "correct" header (i.e. the 167 * oldest header in the arc state), but comes with higher overhead 168 * (i.e. more invocations of arc_evict_state_impl()). 169 */ 170int zfs_arc_evict_batch_limit = 10; 171 172/* 173 * The number of sublists used for each of the arc state lists. If this 174 * is not set to a suitable value by the user, it will be configured to 175 * the number of CPUs on the system in arc_init(). 176 */ 177int zfs_arc_num_sublists_per_state = 0; 178 179/* number of seconds before growing cache again */ 180static int arc_grow_retry = 60; 181 182/* shift of arc_c for calculating overflow limit in arc_get_data_buf */ 183int zfs_arc_overflow_shift = 8; 184 185/* shift of arc_c for calculating both min and max arc_p */ 186static int arc_p_min_shift = 4; 187 188/* log2(fraction of arc to reclaim) */ 189static int arc_shrink_shift = 7; 190 191/* 192 * log2(fraction of ARC which must be free to allow growing). 193 * I.e. If there is less than arc_c >> arc_no_grow_shift free memory, 194 * when reading a new block into the ARC, we will evict an equal-sized block 195 * from the ARC. 196 * 197 * This must be less than arc_shrink_shift, so that when we shrink the ARC, 198 * we will still not allow it to grow. 199 */ 200int arc_no_grow_shift = 5; 201 202 203/* 204 * minimum lifespan of a prefetch block in clock ticks 205 * (initialized in arc_init()) 206 */ 207static int arc_min_prefetch_lifespan; 208 209/* 210 * If this percent of memory is free, don't throttle. 211 */ 212int arc_lotsfree_percent = 10; 213 214static int arc_dead; 215extern boolean_t zfs_prefetch_disable; 216 217/* 218 * The arc has filled available memory and has now warmed up. 219 */ 220static boolean_t arc_warm; 221 222/* 223 * These tunables are for performance analysis. 224 */ 225uint64_t zfs_arc_max; 226uint64_t zfs_arc_min; 227uint64_t zfs_arc_meta_limit = 0; 228uint64_t zfs_arc_meta_min = 0; 229int zfs_arc_grow_retry = 0; 230int zfs_arc_shrink_shift = 0; 231int zfs_arc_p_min_shift = 0; 232int zfs_disable_dup_eviction = 0; 233uint64_t zfs_arc_average_blocksize = 8 * 1024; /* 8KB */ 234u_int zfs_arc_free_target = 0; 235 236static int sysctl_vfs_zfs_arc_free_target(SYSCTL_HANDLER_ARGS); 237static int sysctl_vfs_zfs_arc_meta_limit(SYSCTL_HANDLER_ARGS); 238 239#ifdef _KERNEL 240static void 241arc_free_target_init(void *unused __unused) 242{ 243 244 zfs_arc_free_target = vm_pageout_wakeup_thresh; 245} 246SYSINIT(arc_free_target_init, SI_SUB_KTHREAD_PAGE, SI_ORDER_ANY, 247 arc_free_target_init, NULL); 248 249TUNABLE_QUAD("vfs.zfs.arc_max", &zfs_arc_max); 250TUNABLE_QUAD("vfs.zfs.arc_min", &zfs_arc_min); 251TUNABLE_QUAD("vfs.zfs.arc_meta_limit", &zfs_arc_meta_limit); 252TUNABLE_QUAD("vfs.zfs.arc_meta_min", &zfs_arc_meta_min); 253TUNABLE_QUAD("vfs.zfs.arc_average_blocksize", &zfs_arc_average_blocksize); 254TUNABLE_INT("vfs.zfs.arc_shrink_shift", &zfs_arc_shrink_shift); 255SYSCTL_DECL(_vfs_zfs); 256SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_max, CTLFLAG_RDTUN, &zfs_arc_max, 0, 257 "Maximum ARC size"); 258SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_min, CTLFLAG_RDTUN, &zfs_arc_min, 0, 259 "Minimum ARC size"); 260SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_average_blocksize, CTLFLAG_RDTUN, 261 &zfs_arc_average_blocksize, 0, 262 "ARC average blocksize"); 263SYSCTL_INT(_vfs_zfs, OID_AUTO, arc_shrink_shift, CTLFLAG_RW, 264 &arc_shrink_shift, 0, 265 "log2(fraction of arc to reclaim)"); 266 267/* 268 * We don't have a tunable for arc_free_target due to the dependency on 269 * pagedaemon initialisation. 270 */ 271SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_free_target, 272 CTLTYPE_UINT | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, sizeof(u_int), 273 sysctl_vfs_zfs_arc_free_target, "IU", 274 "Desired number of free pages below which ARC triggers reclaim"); 275 276static int 277sysctl_vfs_zfs_arc_free_target(SYSCTL_HANDLER_ARGS) 278{ 279 u_int val; 280 int err; 281 282 val = zfs_arc_free_target; 283 err = sysctl_handle_int(oidp, &val, 0, req); 284 if (err != 0 || req->newptr == NULL) 285 return (err); 286 287 if (val < minfree) 288 return (EINVAL); 289 if (val > cnt.v_page_count) 290 return (EINVAL); 291 292 zfs_arc_free_target = val; 293 294 return (0); 295} 296 297/* 298 * Must be declared here, before the definition of corresponding kstat 299 * macro which uses the same names will confuse the compiler. 300 */ 301SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_meta_limit, 302 CTLTYPE_U64 | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, sizeof(uint64_t), 303 sysctl_vfs_zfs_arc_meta_limit, "QU", 304 "ARC metadata limit"); 305#endif 306 307/* 308 * Note that buffers can be in one of 6 states: 309 * ARC_anon - anonymous (discussed below) 310 * ARC_mru - recently used, currently cached 311 * ARC_mru_ghost - recentely used, no longer in cache 312 * ARC_mfu - frequently used, currently cached 313 * ARC_mfu_ghost - frequently used, no longer in cache 314 * ARC_l2c_only - exists in L2ARC but not other states 315 * When there are no active references to the buffer, they are 316 * are linked onto a list in one of these arc states. These are 317 * the only buffers that can be evicted or deleted. Within each 318 * state there are multiple lists, one for meta-data and one for 319 * non-meta-data. Meta-data (indirect blocks, blocks of dnodes, 320 * etc.) is tracked separately so that it can be managed more 321 * explicitly: favored over data, limited explicitly. 322 * 323 * Anonymous buffers are buffers that are not associated with 324 * a DVA. These are buffers that hold dirty block copies 325 * before they are written to stable storage. By definition, 326 * they are "ref'd" and are considered part of arc_mru 327 * that cannot be freed. Generally, they will aquire a DVA 328 * as they are written and migrate onto the arc_mru list. 329 * 330 * The ARC_l2c_only state is for buffers that are in the second 331 * level ARC but no longer in any of the ARC_m* lists. The second 332 * level ARC itself may also contain buffers that are in any of 333 * the ARC_m* states - meaning that a buffer can exist in two 334 * places. The reason for the ARC_l2c_only state is to keep the 335 * buffer header in the hash table, so that reads that hit the 336 * second level ARC benefit from these fast lookups. 337 */ 338 339typedef struct arc_state { 340 /* 341 * list of evictable buffers 342 */ 343 multilist_t arcs_list[ARC_BUFC_NUMTYPES]; 344 /* 345 * total amount of evictable data in this state 346 */ 347 uint64_t arcs_lsize[ARC_BUFC_NUMTYPES]; 348 /* 349 * total amount of data in this state; this includes: evictable, 350 * non-evictable, ARC_BUFC_DATA, and ARC_BUFC_METADATA. 351 */ 352 refcount_t arcs_size; 353} arc_state_t; 354 355/* The 6 states: */ 356static arc_state_t ARC_anon; 357static arc_state_t ARC_mru; 358static arc_state_t ARC_mru_ghost; 359static arc_state_t ARC_mfu; 360static arc_state_t ARC_mfu_ghost; 361static arc_state_t ARC_l2c_only; 362 363typedef struct arc_stats { 364 kstat_named_t arcstat_hits; 365 kstat_named_t arcstat_misses; 366 kstat_named_t arcstat_demand_data_hits; 367 kstat_named_t arcstat_demand_data_misses; 368 kstat_named_t arcstat_demand_metadata_hits; 369 kstat_named_t arcstat_demand_metadata_misses; 370 kstat_named_t arcstat_prefetch_data_hits; 371 kstat_named_t arcstat_prefetch_data_misses; 372 kstat_named_t arcstat_prefetch_metadata_hits; 373 kstat_named_t arcstat_prefetch_metadata_misses; 374 kstat_named_t arcstat_mru_hits; 375 kstat_named_t arcstat_mru_ghost_hits; 376 kstat_named_t arcstat_mfu_hits; 377 kstat_named_t arcstat_mfu_ghost_hits; 378 kstat_named_t arcstat_allocated; 379 kstat_named_t arcstat_deleted; 380 /* 381 * Number of buffers that could not be evicted because the hash lock 382 * was held by another thread. The lock may not necessarily be held 383 * by something using the same buffer, since hash locks are shared 384 * by multiple buffers. 385 */ 386 kstat_named_t arcstat_mutex_miss; 387 /* 388 * Number of buffers skipped because they have I/O in progress, are 389 * indrect prefetch buffers that have not lived long enough, or are 390 * not from the spa we're trying to evict from. 391 */ 392 kstat_named_t arcstat_evict_skip; 393 /* 394 * Number of times arc_evict_state() was unable to evict enough 395 * buffers to reach it's target amount. 396 */ 397 kstat_named_t arcstat_evict_not_enough; 398 kstat_named_t arcstat_evict_l2_cached; 399 kstat_named_t arcstat_evict_l2_eligible; 400 kstat_named_t arcstat_evict_l2_ineligible; 401 kstat_named_t arcstat_evict_l2_skip; 402 kstat_named_t arcstat_hash_elements; 403 kstat_named_t arcstat_hash_elements_max; 404 kstat_named_t arcstat_hash_collisions; 405 kstat_named_t arcstat_hash_chains; 406 kstat_named_t arcstat_hash_chain_max; 407 kstat_named_t arcstat_p; 408 kstat_named_t arcstat_c; 409 kstat_named_t arcstat_c_min; 410 kstat_named_t arcstat_c_max; 411 kstat_named_t arcstat_size; 412 /* 413 * Number of bytes consumed by internal ARC structures necessary 414 * for tracking purposes; these structures are not actually 415 * backed by ARC buffers. This includes arc_buf_hdr_t structures 416 * (allocated via arc_buf_hdr_t_full and arc_buf_hdr_t_l2only 417 * caches), and arc_buf_t structures (allocated via arc_buf_t 418 * cache). 419 */ 420 kstat_named_t arcstat_hdr_size; 421 /* 422 * Number of bytes consumed by ARC buffers of type equal to 423 * ARC_BUFC_DATA. This is generally consumed by buffers backing 424 * on disk user data (e.g. plain file contents). 425 */ 426 kstat_named_t arcstat_data_size; 427 /* 428 * Number of bytes consumed by ARC buffers of type equal to 429 * ARC_BUFC_METADATA. This is generally consumed by buffers 430 * backing on disk data that is used for internal ZFS 431 * structures (e.g. ZAP, dnode, indirect blocks, etc). 432 */ 433 kstat_named_t arcstat_metadata_size; 434 /* 435 * Number of bytes consumed by various buffers and structures 436 * not actually backed with ARC buffers. This includes bonus 437 * buffers (allocated directly via zio_buf_* functions), 438 * dmu_buf_impl_t structures (allocated via dmu_buf_impl_t 439 * cache), and dnode_t structures (allocated via dnode_t cache). 440 */ 441 kstat_named_t arcstat_other_size; 442 /* 443 * Total number of bytes consumed by ARC buffers residing in the 444 * arc_anon state. This includes *all* buffers in the arc_anon 445 * state; e.g. data, metadata, evictable, and unevictable buffers 446 * are all included in this value. 447 */ 448 kstat_named_t arcstat_anon_size; 449 /* 450 * Number of bytes consumed by ARC buffers that meet the 451 * following criteria: backing buffers of type ARC_BUFC_DATA, 452 * residing in the arc_anon state, and are eligible for eviction 453 * (e.g. have no outstanding holds on the buffer). 454 */ 455 kstat_named_t arcstat_anon_evictable_data; 456 /* 457 * Number of bytes consumed by ARC buffers that meet the 458 * following criteria: backing buffers of type ARC_BUFC_METADATA, 459 * residing in the arc_anon state, and are eligible for eviction 460 * (e.g. have no outstanding holds on the buffer). 461 */ 462 kstat_named_t arcstat_anon_evictable_metadata; 463 /* 464 * Total number of bytes consumed by ARC buffers residing in the 465 * arc_mru state. This includes *all* buffers in the arc_mru 466 * state; e.g. data, metadata, evictable, and unevictable buffers 467 * are all included in this value. 468 */ 469 kstat_named_t arcstat_mru_size; 470 /* 471 * Number of bytes consumed by ARC buffers that meet the 472 * following criteria: backing buffers of type ARC_BUFC_DATA, 473 * residing in the arc_mru state, and are eligible for eviction 474 * (e.g. have no outstanding holds on the buffer). 475 */ 476 kstat_named_t arcstat_mru_evictable_data; 477 /* 478 * Number of bytes consumed by ARC buffers that meet the 479 * following criteria: backing buffers of type ARC_BUFC_METADATA, 480 * residing in the arc_mru state, and are eligible for eviction 481 * (e.g. have no outstanding holds on the buffer). 482 */ 483 kstat_named_t arcstat_mru_evictable_metadata; 484 /* 485 * Total number of bytes that *would have been* consumed by ARC 486 * buffers in the arc_mru_ghost state. The key thing to note 487 * here, is the fact that this size doesn't actually indicate 488 * RAM consumption. The ghost lists only consist of headers and 489 * don't actually have ARC buffers linked off of these headers. 490 * Thus, *if* the headers had associated ARC buffers, these 491 * buffers *would have* consumed this number of bytes. 492 */ 493 kstat_named_t arcstat_mru_ghost_size; 494 /* 495 * Number of bytes that *would have been* consumed by ARC 496 * buffers that are eligible for eviction, of type 497 * ARC_BUFC_DATA, and linked off the arc_mru_ghost state. 498 */ 499 kstat_named_t arcstat_mru_ghost_evictable_data; 500 /* 501 * Number of bytes that *would have been* consumed by ARC 502 * buffers that are eligible for eviction, of type 503 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state. 504 */ 505 kstat_named_t arcstat_mru_ghost_evictable_metadata; 506 /* 507 * Total number of bytes consumed by ARC buffers residing in the 508 * arc_mfu state. This includes *all* buffers in the arc_mfu 509 * state; e.g. data, metadata, evictable, and unevictable buffers 510 * are all included in this value. 511 */ 512 kstat_named_t arcstat_mfu_size; 513 /* 514 * Number of bytes consumed by ARC buffers that are eligible for 515 * eviction, of type ARC_BUFC_DATA, and reside in the arc_mfu 516 * state. 517 */ 518 kstat_named_t arcstat_mfu_evictable_data; 519 /* 520 * Number of bytes consumed by ARC buffers that are eligible for 521 * eviction, of type ARC_BUFC_METADATA, and reside in the 522 * arc_mfu state. 523 */ 524 kstat_named_t arcstat_mfu_evictable_metadata; 525 /* 526 * Total number of bytes that *would have been* consumed by ARC 527 * buffers in the arc_mfu_ghost state. See the comment above 528 * arcstat_mru_ghost_size for more details. 529 */ 530 kstat_named_t arcstat_mfu_ghost_size; 531 /* 532 * Number of bytes that *would have been* consumed by ARC 533 * buffers that are eligible for eviction, of type 534 * ARC_BUFC_DATA, and linked off the arc_mfu_ghost state. 535 */ 536 kstat_named_t arcstat_mfu_ghost_evictable_data; 537 /* 538 * Number of bytes that *would have been* consumed by ARC 539 * buffers that are eligible for eviction, of type 540 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state. 541 */ 542 kstat_named_t arcstat_mfu_ghost_evictable_metadata; 543 kstat_named_t arcstat_l2_hits; 544 kstat_named_t arcstat_l2_misses; 545 kstat_named_t arcstat_l2_feeds; 546 kstat_named_t arcstat_l2_rw_clash; 547 kstat_named_t arcstat_l2_read_bytes; 548 kstat_named_t arcstat_l2_write_bytes; 549 kstat_named_t arcstat_l2_writes_sent; 550 kstat_named_t arcstat_l2_writes_done; 551 kstat_named_t arcstat_l2_writes_error; 552 kstat_named_t arcstat_l2_writes_lock_retry; 553 kstat_named_t arcstat_l2_evict_lock_retry; 554 kstat_named_t arcstat_l2_evict_reading; 555 kstat_named_t arcstat_l2_evict_l1cached; 556 kstat_named_t arcstat_l2_free_on_write; 557 kstat_named_t arcstat_l2_cdata_free_on_write; 558 kstat_named_t arcstat_l2_abort_lowmem; 559 kstat_named_t arcstat_l2_cksum_bad; 560 kstat_named_t arcstat_l2_io_error; 561 kstat_named_t arcstat_l2_size; 562 kstat_named_t arcstat_l2_asize; 563 kstat_named_t arcstat_l2_hdr_size; 564 kstat_named_t arcstat_l2_compress_successes; 565 kstat_named_t arcstat_l2_compress_zeros; 566 kstat_named_t arcstat_l2_compress_failures; 567 kstat_named_t arcstat_l2_padding_needed; 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_padding_needed", KSTAT_DATA_UINT64 }, 670 { "l2_write_trylock_fail", KSTAT_DATA_UINT64 }, 671 { "l2_write_passed_headroom", KSTAT_DATA_UINT64 }, 672 { "l2_write_spa_mismatch", KSTAT_DATA_UINT64 }, 673 { "l2_write_in_l2", KSTAT_DATA_UINT64 }, 674 { "l2_write_io_in_progress", KSTAT_DATA_UINT64 }, 675 { "l2_write_not_cacheable", KSTAT_DATA_UINT64 }, 676 { "l2_write_full", KSTAT_DATA_UINT64 }, 677 { "l2_write_buffer_iter", KSTAT_DATA_UINT64 }, 678 { "l2_write_pios", KSTAT_DATA_UINT64 }, 679 { "l2_write_buffer_bytes_scanned", KSTAT_DATA_UINT64 }, 680 { "l2_write_buffer_list_iter", KSTAT_DATA_UINT64 }, 681 { "l2_write_buffer_list_null_iter", KSTAT_DATA_UINT64 }, 682 { "memory_throttle_count", KSTAT_DATA_UINT64 }, 683 { "duplicate_buffers", KSTAT_DATA_UINT64 }, 684 { "duplicate_buffers_size", KSTAT_DATA_UINT64 }, 685 { "duplicate_reads", KSTAT_DATA_UINT64 }, 686 { "arc_meta_used", KSTAT_DATA_UINT64 }, 687 { "arc_meta_limit", KSTAT_DATA_UINT64 }, 688 { "arc_meta_max", KSTAT_DATA_UINT64 }, 689 { "arc_meta_min", KSTAT_DATA_UINT64 }, 690 { "sync_wait_for_async", KSTAT_DATA_UINT64 }, 691 { "demand_hit_predictive_prefetch", KSTAT_DATA_UINT64 }, 692}; 693 694#define ARCSTAT(stat) (arc_stats.stat.value.ui64) 695 696#define ARCSTAT_INCR(stat, val) \ 697 atomic_add_64(&arc_stats.stat.value.ui64, (val)) 698 699#define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1) 700#define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1) 701 702#define ARCSTAT_MAX(stat, val) { \ 703 uint64_t m; \ 704 while ((val) > (m = arc_stats.stat.value.ui64) && \ 705 (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \ 706 continue; \ 707} 708 709#define ARCSTAT_MAXSTAT(stat) \ 710 ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64) 711 712/* 713 * We define a macro to allow ARC hits/misses to be easily broken down by 714 * two separate conditions, giving a total of four different subtypes for 715 * each of hits and misses (so eight statistics total). 716 */ 717#define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \ 718 if (cond1) { \ 719 if (cond2) { \ 720 ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \ 721 } else { \ 722 ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \ 723 } \ 724 } else { \ 725 if (cond2) { \ 726 ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \ 727 } else { \ 728 ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\ 729 } \ 730 } 731 732kstat_t *arc_ksp; 733static arc_state_t *arc_anon; 734static arc_state_t *arc_mru; 735static arc_state_t *arc_mru_ghost; 736static arc_state_t *arc_mfu; 737static arc_state_t *arc_mfu_ghost; 738static arc_state_t *arc_l2c_only; 739 740/* 741 * There are several ARC variables that are critical to export as kstats -- 742 * but we don't want to have to grovel around in the kstat whenever we wish to 743 * manipulate them. For these variables, we therefore define them to be in 744 * terms of the statistic variable. This assures that we are not introducing 745 * the possibility of inconsistency by having shadow copies of the variables, 746 * while still allowing the code to be readable. 747 */ 748#define arc_size ARCSTAT(arcstat_size) /* actual total arc size */ 749#define arc_p ARCSTAT(arcstat_p) /* target size of MRU */ 750#define arc_c ARCSTAT(arcstat_c) /* target size of cache */ 751#define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */ 752#define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */ 753#define arc_meta_limit ARCSTAT(arcstat_meta_limit) /* max size for metadata */ 754#define arc_meta_min ARCSTAT(arcstat_meta_min) /* min size for metadata */ 755#define arc_meta_used ARCSTAT(arcstat_meta_used) /* size of metadata */ 756#define arc_meta_max ARCSTAT(arcstat_meta_max) /* max size of metadata */ 757 758#define L2ARC_IS_VALID_COMPRESS(_c_) \ 759 ((_c_) == ZIO_COMPRESS_LZ4 || (_c_) == ZIO_COMPRESS_EMPTY) 760 761static int arc_no_grow; /* Don't try to grow cache size */ 762static uint64_t arc_tempreserve; 763static uint64_t arc_loaned_bytes; 764 765typedef struct arc_callback arc_callback_t; 766 767struct arc_callback { 768 void *acb_private; 769 arc_done_func_t *acb_done; 770 arc_buf_t *acb_buf; 771 zio_t *acb_zio_dummy; 772 arc_callback_t *acb_next; 773}; 774 775typedef struct arc_write_callback arc_write_callback_t; 776 777struct arc_write_callback { 778 void *awcb_private; 779 arc_done_func_t *awcb_ready; 780 arc_done_func_t *awcb_physdone; 781 arc_done_func_t *awcb_done; 782 arc_buf_t *awcb_buf; 783}; 784 785/* 786 * ARC buffers are separated into multiple structs as a memory saving measure: 787 * - Common fields struct, always defined, and embedded within it: 788 * - L2-only fields, always allocated but undefined when not in L2ARC 789 * - L1-only fields, only allocated when in L1ARC 790 * 791 * Buffer in L1 Buffer only in L2 792 * +------------------------+ +------------------------+ 793 * | arc_buf_hdr_t | | arc_buf_hdr_t | 794 * | | | | 795 * | | | | 796 * | | | | 797 * +------------------------+ +------------------------+ 798 * | l2arc_buf_hdr_t | | l2arc_buf_hdr_t | 799 * | (undefined if L1-only) | | | 800 * +------------------------+ +------------------------+ 801 * | l1arc_buf_hdr_t | 802 * | | 803 * | | 804 * | | 805 * | | 806 * +------------------------+ 807 * 808 * Because it's possible for the L2ARC to become extremely large, we can wind 809 * up eating a lot of memory in L2ARC buffer headers, so the size of a header 810 * is minimized by only allocating the fields necessary for an L1-cached buffer 811 * when a header is actually in the L1 cache. The sub-headers (l1arc_buf_hdr and 812 * l2arc_buf_hdr) are embedded rather than allocated separately to save a couple 813 * words in pointers. arc_hdr_realloc() is used to switch a header between 814 * these two allocation states. 815 */ 816typedef struct l1arc_buf_hdr { 817 kmutex_t b_freeze_lock; 818#ifdef ZFS_DEBUG 819 /* 820 * used for debugging wtih kmem_flags - by allocating and freeing 821 * b_thawed when the buffer is thawed, we get a record of the stack 822 * trace that thawed it. 823 */ 824 void *b_thawed; 825#endif 826 827 arc_buf_t *b_buf; 828 uint32_t b_datacnt; 829 /* for waiting on writes to complete */ 830 kcondvar_t b_cv; 831 832 /* protected by arc state mutex */ 833 arc_state_t *b_state; 834 multilist_node_t b_arc_node; 835 836 /* updated atomically */ 837 clock_t b_arc_access; 838 839 /* self protecting */ 840 refcount_t b_refcnt; 841 842 arc_callback_t *b_acb; 843 /* temporary buffer holder for in-flight compressed or padded data */ 844 void *b_tmp_cdata; 845} l1arc_buf_hdr_t; 846 847typedef struct l2arc_dev l2arc_dev_t; 848 849typedef struct l2arc_buf_hdr { 850 /* protected by arc_buf_hdr mutex */ 851 l2arc_dev_t *b_dev; /* L2ARC device */ 852 uint64_t b_daddr; /* disk address, offset byte */ 853 /* real alloc'd buffer size depending on b_compress applied */ 854 int32_t b_asize; 855 uint8_t b_compress; 856 857 list_node_t b_l2node; 858} l2arc_buf_hdr_t; 859 860struct arc_buf_hdr { 861 /* protected by hash lock */ 862 dva_t b_dva; 863 uint64_t b_birth; 864 /* 865 * Even though this checksum is only set/verified when a buffer is in 866 * the L1 cache, it needs to be in the set of common fields because it 867 * must be preserved from the time before a buffer is written out to 868 * L2ARC until after it is read back in. 869 */ 870 zio_cksum_t *b_freeze_cksum; 871 872 arc_buf_hdr_t *b_hash_next; 873 arc_flags_t b_flags; 874 875 /* immutable */ 876 int32_t b_size; 877 uint64_t b_spa; 878 879 /* L2ARC fields. Undefined when not in L2ARC. */ 880 l2arc_buf_hdr_t b_l2hdr; 881 /* L1ARC fields. Undefined when in l2arc_only state */ 882 l1arc_buf_hdr_t b_l1hdr; 883}; 884 885#ifdef _KERNEL 886static int 887sysctl_vfs_zfs_arc_meta_limit(SYSCTL_HANDLER_ARGS) 888{ 889 uint64_t val; 890 int err; 891 892 val = arc_meta_limit; 893 err = sysctl_handle_64(oidp, &val, 0, req); 894 if (err != 0 || req->newptr == NULL) 895 return (err); 896 897 if (val <= 0 || val > arc_c_max) 898 return (EINVAL); 899 900 arc_meta_limit = val; 901 return (0); 902} 903#endif 904 905static arc_buf_t *arc_eviction_list; 906static arc_buf_hdr_t arc_eviction_hdr; 907 908#define GHOST_STATE(state) \ 909 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \ 910 (state) == arc_l2c_only) 911 912#define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_FLAG_IN_HASH_TABLE) 913#define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) 914#define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_FLAG_IO_ERROR) 915#define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_FLAG_PREFETCH) 916#define HDR_FREED_IN_READ(hdr) ((hdr)->b_flags & ARC_FLAG_FREED_IN_READ) 917#define HDR_BUF_AVAILABLE(hdr) ((hdr)->b_flags & ARC_FLAG_BUF_AVAILABLE) 918 919#define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_FLAG_L2CACHE) 920#define HDR_L2COMPRESS(hdr) ((hdr)->b_flags & ARC_FLAG_L2COMPRESS) 921#define HDR_L2_READING(hdr) \ 922 (((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) && \ 923 ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR)) 924#define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITING) 925#define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_FLAG_L2_EVICTED) 926#define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITE_HEAD) 927 928#define HDR_ISTYPE_METADATA(hdr) \ 929 ((hdr)->b_flags & ARC_FLAG_BUFC_METADATA) 930#define HDR_ISTYPE_DATA(hdr) (!HDR_ISTYPE_METADATA(hdr)) 931 932#define HDR_HAS_L1HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L1HDR) 933#define HDR_HAS_L2HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR) 934 935/* 936 * Other sizes 937 */ 938 939#define HDR_FULL_SIZE ((int64_t)sizeof (arc_buf_hdr_t)) 940#define HDR_L2ONLY_SIZE ((int64_t)offsetof(arc_buf_hdr_t, b_l1hdr)) 941 942/* 943 * Hash table routines 944 */ 945 946#define HT_LOCK_PAD CACHE_LINE_SIZE 947 948struct ht_lock { 949 kmutex_t ht_lock; 950#ifdef _KERNEL 951 unsigned char pad[(HT_LOCK_PAD - sizeof (kmutex_t))]; 952#endif 953}; 954 955#define BUF_LOCKS 256 956typedef struct buf_hash_table { 957 uint64_t ht_mask; 958 arc_buf_hdr_t **ht_table; 959 struct ht_lock ht_locks[BUF_LOCKS] __aligned(CACHE_LINE_SIZE); 960} buf_hash_table_t; 961 962static buf_hash_table_t buf_hash_table; 963 964#define BUF_HASH_INDEX(spa, dva, birth) \ 965 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask) 966#define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)]) 967#define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock)) 968#define HDR_LOCK(hdr) \ 969 (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth))) 970 971uint64_t zfs_crc64_table[256]; 972 973/* 974 * Level 2 ARC 975 */ 976 977#define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */ 978#define L2ARC_HEADROOM 2 /* num of writes */ 979/* 980 * If we discover during ARC scan any buffers to be compressed, we boost 981 * our headroom for the next scanning cycle by this percentage multiple. 982 */ 983#define L2ARC_HEADROOM_BOOST 200 984#define L2ARC_FEED_SECS 1 /* caching interval secs */ 985#define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */ 986 987/* 988 * Used to distinguish headers that are being process by 989 * l2arc_write_buffers(), but have yet to be assigned to a l2arc disk 990 * address. This can happen when the header is added to the l2arc's list 991 * of buffers to write in the first stage of l2arc_write_buffers(), but 992 * has not yet been written out which happens in the second stage of 993 * l2arc_write_buffers(). 994 */ 995#define L2ARC_ADDR_UNSET ((uint64_t)(-1)) 996 997#define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent) 998#define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done) 999 1000/* L2ARC Performance Tunables */ 1001uint64_t l2arc_write_max = L2ARC_WRITE_SIZE; /* default max write size */ 1002uint64_t l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra write during warmup */ 1003uint64_t l2arc_headroom = L2ARC_HEADROOM; /* number of dev writes */ 1004uint64_t l2arc_headroom_boost = L2ARC_HEADROOM_BOOST; 1005uint64_t l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */ 1006uint64_t l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval milliseconds */ 1007boolean_t l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */ 1008boolean_t l2arc_feed_again = B_TRUE; /* turbo warmup */ 1009boolean_t l2arc_norw = B_TRUE; /* no reads during writes */ 1010 1011SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_max, CTLFLAG_RW, 1012 &l2arc_write_max, 0, "max write size"); 1013SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_boost, CTLFLAG_RW, 1014 &l2arc_write_boost, 0, "extra write during warmup"); 1015SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_headroom, CTLFLAG_RW, 1016 &l2arc_headroom, 0, "number of dev writes"); 1017SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_secs, CTLFLAG_RW, 1018 &l2arc_feed_secs, 0, "interval seconds"); 1019SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_min_ms, CTLFLAG_RW, 1020 &l2arc_feed_min_ms, 0, "min interval milliseconds"); 1021 1022SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_noprefetch, CTLFLAG_RW, 1023 &l2arc_noprefetch, 0, "don't cache prefetch bufs"); 1024SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_feed_again, CTLFLAG_RW, 1025 &l2arc_feed_again, 0, "turbo warmup"); 1026SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_norw, CTLFLAG_RW, 1027 &l2arc_norw, 0, "no reads during writes"); 1028 1029SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_size, CTLFLAG_RD, 1030 &ARC_anon.arcs_size.rc_count, 0, "size of anonymous state"); 1031SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_metadata_lsize, CTLFLAG_RD, 1032 &ARC_anon.arcs_lsize[ARC_BUFC_METADATA], 0, "size of anonymous state"); 1033SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_data_lsize, CTLFLAG_RD, 1034 &ARC_anon.arcs_lsize[ARC_BUFC_DATA], 0, "size of anonymous state"); 1035 1036SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_size, CTLFLAG_RD, 1037 &ARC_mru.arcs_size.rc_count, 0, "size of mru state"); 1038SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_metadata_lsize, CTLFLAG_RD, 1039 &ARC_mru.arcs_lsize[ARC_BUFC_METADATA], 0, "size of metadata in mru state"); 1040SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_data_lsize, CTLFLAG_RD, 1041 &ARC_mru.arcs_lsize[ARC_BUFC_DATA], 0, "size of data in mru state"); 1042 1043SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_size, CTLFLAG_RD, 1044 &ARC_mru_ghost.arcs_size.rc_count, 0, "size of mru ghost state"); 1045SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_metadata_lsize, CTLFLAG_RD, 1046 &ARC_mru_ghost.arcs_lsize[ARC_BUFC_METADATA], 0, 1047 "size of metadata in mru ghost state"); 1048SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_data_lsize, CTLFLAG_RD, 1049 &ARC_mru_ghost.arcs_lsize[ARC_BUFC_DATA], 0, 1050 "size of data in mru ghost state"); 1051 1052SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_size, CTLFLAG_RD, 1053 &ARC_mfu.arcs_size.rc_count, 0, "size of mfu state"); 1054SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_metadata_lsize, CTLFLAG_RD, 1055 &ARC_mfu.arcs_lsize[ARC_BUFC_METADATA], 0, "size of metadata in mfu state"); 1056SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_data_lsize, CTLFLAG_RD, 1057 &ARC_mfu.arcs_lsize[ARC_BUFC_DATA], 0, "size of data in mfu state"); 1058 1059SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_size, CTLFLAG_RD, 1060 &ARC_mfu_ghost.arcs_size.rc_count, 0, "size of mfu ghost state"); 1061SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_metadata_lsize, CTLFLAG_RD, 1062 &ARC_mfu_ghost.arcs_lsize[ARC_BUFC_METADATA], 0, 1063 "size of metadata in mfu ghost state"); 1064SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_data_lsize, CTLFLAG_RD, 1065 &ARC_mfu_ghost.arcs_lsize[ARC_BUFC_DATA], 0, 1066 "size of data in mfu ghost state"); 1067 1068SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2c_only_size, CTLFLAG_RD, 1069 &ARC_l2c_only.arcs_size.rc_count, 0, "size of mru state"); 1070 1071/* 1072 * L2ARC Internals 1073 */ 1074struct l2arc_dev { 1075 vdev_t *l2ad_vdev; /* vdev */ 1076 spa_t *l2ad_spa; /* spa */ 1077 uint64_t l2ad_hand; /* next write location */ 1078 uint64_t l2ad_start; /* first addr on device */ 1079 uint64_t l2ad_end; /* last addr on device */ 1080 boolean_t l2ad_first; /* first sweep through */ 1081 boolean_t l2ad_writing; /* currently writing */ 1082 kmutex_t l2ad_mtx; /* lock for buffer list */ 1083 list_t l2ad_buflist; /* buffer list */ 1084 list_node_t l2ad_node; /* device list node */ 1085 refcount_t l2ad_alloc; /* allocated bytes */ 1086}; 1087 1088static list_t L2ARC_dev_list; /* device list */ 1089static list_t *l2arc_dev_list; /* device list pointer */ 1090static kmutex_t l2arc_dev_mtx; /* device list mutex */ 1091static l2arc_dev_t *l2arc_dev_last; /* last device used */ 1092static list_t L2ARC_free_on_write; /* free after write buf list */ 1093static list_t *l2arc_free_on_write; /* free after write list ptr */ 1094static kmutex_t l2arc_free_on_write_mtx; /* mutex for list */ 1095static uint64_t l2arc_ndev; /* number of devices */ 1096 1097typedef struct l2arc_read_callback { 1098 arc_buf_t *l2rcb_buf; /* read buffer */ 1099 spa_t *l2rcb_spa; /* spa */ 1100 blkptr_t l2rcb_bp; /* original blkptr */ 1101 zbookmark_phys_t l2rcb_zb; /* original bookmark */ 1102 int l2rcb_flags; /* original flags */ 1103 enum zio_compress l2rcb_compress; /* applied compress */ 1104 void *l2rcb_data; /* temporary buffer */ 1105} l2arc_read_callback_t; 1106 1107typedef struct l2arc_write_callback { 1108 l2arc_dev_t *l2wcb_dev; /* device info */ 1109 arc_buf_hdr_t *l2wcb_head; /* head of write buflist */ 1110} l2arc_write_callback_t; 1111 1112typedef struct l2arc_data_free { 1113 /* protected by l2arc_free_on_write_mtx */ 1114 void *l2df_data; 1115 size_t l2df_size; 1116 void (*l2df_func)(void *, size_t); 1117 list_node_t l2df_list_node; 1118} l2arc_data_free_t; 1119 1120static kmutex_t l2arc_feed_thr_lock; 1121static kcondvar_t l2arc_feed_thr_cv; 1122static uint8_t l2arc_thread_exit; 1123 1124static void arc_get_data_buf(arc_buf_t *); 1125static void arc_access(arc_buf_hdr_t *, kmutex_t *); 1126static boolean_t arc_is_overflowing(); 1127static void arc_buf_watch(arc_buf_t *); 1128 1129static arc_buf_contents_t arc_buf_type(arc_buf_hdr_t *); 1130static uint32_t arc_bufc_to_flags(arc_buf_contents_t); 1131 1132static boolean_t l2arc_write_eligible(uint64_t, arc_buf_hdr_t *); 1133static void l2arc_read_done(zio_t *); 1134 1135static boolean_t l2arc_transform_buf(arc_buf_hdr_t *, boolean_t); 1136static void l2arc_decompress_zio(zio_t *, arc_buf_hdr_t *, enum zio_compress); 1137static void l2arc_release_cdata_buf(arc_buf_hdr_t *); 1138 1139static void 1140l2arc_trim(const arc_buf_hdr_t *hdr) 1141{ 1142 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev; 1143 1144 ASSERT(HDR_HAS_L2HDR(hdr)); 1145 ASSERT(MUTEX_HELD(&dev->l2ad_mtx)); 1146 1147 if (hdr->b_l2hdr.b_daddr == L2ARC_ADDR_UNSET) 1148 return; 1149 if (hdr->b_l2hdr.b_asize != 0) { 1150 trim_map_free(dev->l2ad_vdev, hdr->b_l2hdr.b_daddr, 1151 hdr->b_l2hdr.b_asize, 0); 1152 } else { 1153 ASSERT3U(hdr->b_l2hdr.b_compress, ==, ZIO_COMPRESS_EMPTY); 1154 } 1155} 1156 1157static uint64_t 1158buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth) 1159{ 1160 uint8_t *vdva = (uint8_t *)dva; 1161 uint64_t crc = -1ULL; 1162 int i; 1163 1164 ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY); 1165 1166 for (i = 0; i < sizeof (dva_t); i++) 1167 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ vdva[i]) & 0xFF]; 1168 1169 crc ^= (spa>>8) ^ birth; 1170 1171 return (crc); 1172} 1173 1174#define BUF_EMPTY(buf) \ 1175 ((buf)->b_dva.dva_word[0] == 0 && \ 1176 (buf)->b_dva.dva_word[1] == 0) 1177 1178#define BUF_EQUAL(spa, dva, birth, buf) \ 1179 ((buf)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \ 1180 ((buf)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \ 1181 ((buf)->b_birth == birth) && ((buf)->b_spa == spa) 1182 1183static void 1184buf_discard_identity(arc_buf_hdr_t *hdr) 1185{ 1186 hdr->b_dva.dva_word[0] = 0; 1187 hdr->b_dva.dva_word[1] = 0; 1188 hdr->b_birth = 0; 1189} 1190 1191static arc_buf_hdr_t * 1192buf_hash_find(uint64_t spa, const blkptr_t *bp, kmutex_t **lockp) 1193{ 1194 const dva_t *dva = BP_IDENTITY(bp); 1195 uint64_t birth = BP_PHYSICAL_BIRTH(bp); 1196 uint64_t idx = BUF_HASH_INDEX(spa, dva, birth); 1197 kmutex_t *hash_lock = BUF_HASH_LOCK(idx); 1198 arc_buf_hdr_t *hdr; 1199 1200 mutex_enter(hash_lock); 1201 for (hdr = buf_hash_table.ht_table[idx]; hdr != NULL; 1202 hdr = hdr->b_hash_next) { 1203 if (BUF_EQUAL(spa, dva, birth, hdr)) { 1204 *lockp = hash_lock; 1205 return (hdr); 1206 } 1207 } 1208 mutex_exit(hash_lock); 1209 *lockp = NULL; 1210 return (NULL); 1211} 1212 1213/* 1214 * Insert an entry into the hash table. If there is already an element 1215 * equal to elem in the hash table, then the already existing element 1216 * will be returned and the new element will not be inserted. 1217 * Otherwise returns NULL. 1218 * If lockp == NULL, the caller is assumed to already hold the hash lock. 1219 */ 1220static arc_buf_hdr_t * 1221buf_hash_insert(arc_buf_hdr_t *hdr, kmutex_t **lockp) 1222{ 1223 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth); 1224 kmutex_t *hash_lock = BUF_HASH_LOCK(idx); 1225 arc_buf_hdr_t *fhdr; 1226 uint32_t i; 1227 1228 ASSERT(!DVA_IS_EMPTY(&hdr->b_dva)); 1229 ASSERT(hdr->b_birth != 0); 1230 ASSERT(!HDR_IN_HASH_TABLE(hdr)); 1231 1232 if (lockp != NULL) { 1233 *lockp = hash_lock; 1234 mutex_enter(hash_lock); 1235 } else { 1236 ASSERT(MUTEX_HELD(hash_lock)); 1237 } 1238 1239 for (fhdr = buf_hash_table.ht_table[idx], i = 0; fhdr != NULL; 1240 fhdr = fhdr->b_hash_next, i++) { 1241 if (BUF_EQUAL(hdr->b_spa, &hdr->b_dva, hdr->b_birth, fhdr)) 1242 return (fhdr); 1243 } 1244 1245 hdr->b_hash_next = buf_hash_table.ht_table[idx]; 1246 buf_hash_table.ht_table[idx] = hdr; 1247 hdr->b_flags |= ARC_FLAG_IN_HASH_TABLE; 1248 1249 /* collect some hash table performance data */ 1250 if (i > 0) { 1251 ARCSTAT_BUMP(arcstat_hash_collisions); 1252 if (i == 1) 1253 ARCSTAT_BUMP(arcstat_hash_chains); 1254 1255 ARCSTAT_MAX(arcstat_hash_chain_max, i); 1256 } 1257 1258 ARCSTAT_BUMP(arcstat_hash_elements); 1259 ARCSTAT_MAXSTAT(arcstat_hash_elements); 1260 1261 return (NULL); 1262} 1263 1264static void 1265buf_hash_remove(arc_buf_hdr_t *hdr) 1266{ 1267 arc_buf_hdr_t *fhdr, **hdrp; 1268 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth); 1269 1270 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx))); 1271 ASSERT(HDR_IN_HASH_TABLE(hdr)); 1272 1273 hdrp = &buf_hash_table.ht_table[idx]; 1274 while ((fhdr = *hdrp) != hdr) { 1275 ASSERT(fhdr != NULL); 1276 hdrp = &fhdr->b_hash_next; 1277 } 1278 *hdrp = hdr->b_hash_next; 1279 hdr->b_hash_next = NULL; 1280 hdr->b_flags &= ~ARC_FLAG_IN_HASH_TABLE; 1281 1282 /* collect some hash table performance data */ 1283 ARCSTAT_BUMPDOWN(arcstat_hash_elements); 1284 1285 if (buf_hash_table.ht_table[idx] && 1286 buf_hash_table.ht_table[idx]->b_hash_next == NULL) 1287 ARCSTAT_BUMPDOWN(arcstat_hash_chains); 1288} 1289 1290/* 1291 * Global data structures and functions for the buf kmem cache. 1292 */ 1293static kmem_cache_t *hdr_full_cache; 1294static kmem_cache_t *hdr_l2only_cache; 1295static kmem_cache_t *buf_cache; 1296 1297static void 1298buf_fini(void) 1299{ 1300 int i; 1301 1302 kmem_free(buf_hash_table.ht_table, 1303 (buf_hash_table.ht_mask + 1) * sizeof (void *)); 1304 for (i = 0; i < BUF_LOCKS; i++) 1305 mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock); 1306 kmem_cache_destroy(hdr_full_cache); 1307 kmem_cache_destroy(hdr_l2only_cache); 1308 kmem_cache_destroy(buf_cache); 1309} 1310 1311/* 1312 * Constructor callback - called when the cache is empty 1313 * and a new buf is requested. 1314 */ 1315/* ARGSUSED */ 1316static int 1317hdr_full_cons(void *vbuf, void *unused, int kmflag) 1318{ 1319 arc_buf_hdr_t *hdr = vbuf; 1320 1321 bzero(hdr, HDR_FULL_SIZE); 1322 cv_init(&hdr->b_l1hdr.b_cv, NULL, CV_DEFAULT, NULL); 1323 refcount_create(&hdr->b_l1hdr.b_refcnt); 1324 mutex_init(&hdr->b_l1hdr.b_freeze_lock, NULL, MUTEX_DEFAULT, NULL); 1325 multilist_link_init(&hdr->b_l1hdr.b_arc_node); 1326 arc_space_consume(HDR_FULL_SIZE, ARC_SPACE_HDRS); 1327 1328 return (0); 1329} 1330 1331/* ARGSUSED */ 1332static int 1333hdr_l2only_cons(void *vbuf, void *unused, int kmflag) 1334{ 1335 arc_buf_hdr_t *hdr = vbuf; 1336 1337 bzero(hdr, HDR_L2ONLY_SIZE); 1338 arc_space_consume(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS); 1339 1340 return (0); 1341} 1342 1343/* ARGSUSED */ 1344static int 1345buf_cons(void *vbuf, void *unused, int kmflag) 1346{ 1347 arc_buf_t *buf = vbuf; 1348 1349 bzero(buf, sizeof (arc_buf_t)); 1350 mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL); 1351 arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS); 1352 1353 return (0); 1354} 1355 1356/* 1357 * Destructor callback - called when a cached buf is 1358 * no longer required. 1359 */ 1360/* ARGSUSED */ 1361static void 1362hdr_full_dest(void *vbuf, void *unused) 1363{ 1364 arc_buf_hdr_t *hdr = vbuf; 1365 1366 ASSERT(BUF_EMPTY(hdr)); 1367 cv_destroy(&hdr->b_l1hdr.b_cv); 1368 refcount_destroy(&hdr->b_l1hdr.b_refcnt); 1369 mutex_destroy(&hdr->b_l1hdr.b_freeze_lock); 1370 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 1371 arc_space_return(HDR_FULL_SIZE, ARC_SPACE_HDRS); 1372} 1373 1374/* ARGSUSED */ 1375static void 1376hdr_l2only_dest(void *vbuf, void *unused) 1377{ 1378 arc_buf_hdr_t *hdr = vbuf; 1379 1380 ASSERT(BUF_EMPTY(hdr)); 1381 arc_space_return(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS); 1382} 1383 1384/* ARGSUSED */ 1385static void 1386buf_dest(void *vbuf, void *unused) 1387{ 1388 arc_buf_t *buf = vbuf; 1389 1390 mutex_destroy(&buf->b_evict_lock); 1391 arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS); 1392} 1393 1394/* 1395 * Reclaim callback -- invoked when memory is low. 1396 */ 1397/* ARGSUSED */ 1398static void 1399hdr_recl(void *unused) 1400{ 1401 dprintf("hdr_recl called\n"); 1402 /* 1403 * umem calls the reclaim func when we destroy the buf cache, 1404 * which is after we do arc_fini(). 1405 */ 1406 if (!arc_dead) 1407 cv_signal(&arc_reclaim_thread_cv); 1408} 1409 1410static void 1411buf_init(void) 1412{ 1413 uint64_t *ct; 1414 uint64_t hsize = 1ULL << 12; 1415 int i, j; 1416 1417 /* 1418 * The hash table is big enough to fill all of physical memory 1419 * with an average block size of zfs_arc_average_blocksize (default 8K). 1420 * By default, the table will take up 1421 * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers). 1422 */ 1423 while (hsize * zfs_arc_average_blocksize < (uint64_t)physmem * PAGESIZE) 1424 hsize <<= 1; 1425retry: 1426 buf_hash_table.ht_mask = hsize - 1; 1427 buf_hash_table.ht_table = 1428 kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP); 1429 if (buf_hash_table.ht_table == NULL) { 1430 ASSERT(hsize > (1ULL << 8)); 1431 hsize >>= 1; 1432 goto retry; 1433 } 1434 1435 hdr_full_cache = kmem_cache_create("arc_buf_hdr_t_full", HDR_FULL_SIZE, 1436 0, hdr_full_cons, hdr_full_dest, hdr_recl, NULL, NULL, 0); 1437 hdr_l2only_cache = kmem_cache_create("arc_buf_hdr_t_l2only", 1438 HDR_L2ONLY_SIZE, 0, hdr_l2only_cons, hdr_l2only_dest, hdr_recl, 1439 NULL, NULL, 0); 1440 buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t), 1441 0, buf_cons, buf_dest, NULL, NULL, NULL, 0); 1442 1443 for (i = 0; i < 256; i++) 1444 for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--) 1445 *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY); 1446 1447 for (i = 0; i < BUF_LOCKS; i++) { 1448 mutex_init(&buf_hash_table.ht_locks[i].ht_lock, 1449 NULL, MUTEX_DEFAULT, NULL); 1450 } 1451} 1452 1453/* 1454 * Transition between the two allocation states for the arc_buf_hdr struct. 1455 * The arc_buf_hdr struct can be allocated with (hdr_full_cache) or without 1456 * (hdr_l2only_cache) the fields necessary for the L1 cache - the smaller 1457 * version is used when a cache buffer is only in the L2ARC in order to reduce 1458 * memory usage. 1459 */ 1460static arc_buf_hdr_t * 1461arc_hdr_realloc(arc_buf_hdr_t *hdr, kmem_cache_t *old, kmem_cache_t *new) 1462{ 1463 ASSERT(HDR_HAS_L2HDR(hdr)); 1464 1465 arc_buf_hdr_t *nhdr; 1466 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev; 1467 1468 ASSERT((old == hdr_full_cache && new == hdr_l2only_cache) || 1469 (old == hdr_l2only_cache && new == hdr_full_cache)); 1470 1471 nhdr = kmem_cache_alloc(new, KM_PUSHPAGE); 1472 1473 ASSERT(MUTEX_HELD(HDR_LOCK(hdr))); 1474 buf_hash_remove(hdr); 1475 1476 bcopy(hdr, nhdr, HDR_L2ONLY_SIZE); 1477 1478 if (new == hdr_full_cache) { 1479 nhdr->b_flags |= ARC_FLAG_HAS_L1HDR; 1480 /* 1481 * arc_access and arc_change_state need to be aware that a 1482 * header has just come out of L2ARC, so we set its state to 1483 * l2c_only even though it's about to change. 1484 */ 1485 nhdr->b_l1hdr.b_state = arc_l2c_only; 1486 1487 /* Verify previous threads set to NULL before freeing */ 1488 ASSERT3P(nhdr->b_l1hdr.b_tmp_cdata, ==, NULL); 1489 } else { 1490 ASSERT(hdr->b_l1hdr.b_buf == NULL); 1491 ASSERT0(hdr->b_l1hdr.b_datacnt); 1492 1493 /* 1494 * If we've reached here, We must have been called from 1495 * arc_evict_hdr(), as such we should have already been 1496 * removed from any ghost list we were previously on 1497 * (which protects us from racing with arc_evict_state), 1498 * thus no locking is needed during this check. 1499 */ 1500 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 1501 1502 /* 1503 * A buffer must not be moved into the arc_l2c_only 1504 * state if it's not finished being written out to the 1505 * l2arc device. Otherwise, the b_l1hdr.b_tmp_cdata field 1506 * might try to be accessed, even though it was removed. 1507 */ 1508 VERIFY(!HDR_L2_WRITING(hdr)); 1509 VERIFY3P(hdr->b_l1hdr.b_tmp_cdata, ==, NULL); 1510 1511#ifdef ZFS_DEBUG 1512 if (hdr->b_l1hdr.b_thawed != NULL) { 1513 kmem_free(hdr->b_l1hdr.b_thawed, 1); 1514 hdr->b_l1hdr.b_thawed = NULL; 1515 } 1516#endif 1517 1518 nhdr->b_flags &= ~ARC_FLAG_HAS_L1HDR; 1519 } 1520 /* 1521 * The header has been reallocated so we need to re-insert it into any 1522 * lists it was on. 1523 */ 1524 (void) buf_hash_insert(nhdr, NULL); 1525 1526 ASSERT(list_link_active(&hdr->b_l2hdr.b_l2node)); 1527 1528 mutex_enter(&dev->l2ad_mtx); 1529 1530 /* 1531 * We must place the realloc'ed header back into the list at 1532 * the same spot. Otherwise, if it's placed earlier in the list, 1533 * l2arc_write_buffers() could find it during the function's 1534 * write phase, and try to write it out to the l2arc. 1535 */ 1536 list_insert_after(&dev->l2ad_buflist, hdr, nhdr); 1537 list_remove(&dev->l2ad_buflist, hdr); 1538 1539 mutex_exit(&dev->l2ad_mtx); 1540 1541 /* 1542 * Since we're using the pointer address as the tag when 1543 * incrementing and decrementing the l2ad_alloc refcount, we 1544 * must remove the old pointer (that we're about to destroy) and 1545 * add the new pointer to the refcount. Otherwise we'd remove 1546 * the wrong pointer address when calling arc_hdr_destroy() later. 1547 */ 1548 1549 (void) refcount_remove_many(&dev->l2ad_alloc, 1550 hdr->b_l2hdr.b_asize, hdr); 1551 1552 (void) refcount_add_many(&dev->l2ad_alloc, 1553 nhdr->b_l2hdr.b_asize, nhdr); 1554 1555 buf_discard_identity(hdr); 1556 hdr->b_freeze_cksum = NULL; 1557 kmem_cache_free(old, hdr); 1558 1559 return (nhdr); 1560} 1561 1562 1563#define ARC_MINTIME (hz>>4) /* 62 ms */ 1564 1565static void 1566arc_cksum_verify(arc_buf_t *buf) 1567{ 1568 zio_cksum_t zc; 1569 1570 if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 1571 return; 1572 1573 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1574 if (buf->b_hdr->b_freeze_cksum == NULL || HDR_IO_ERROR(buf->b_hdr)) { 1575 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1576 return; 1577 } 1578 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, NULL, &zc); 1579 if (!ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc)) 1580 panic("buffer modified while frozen!"); 1581 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1582} 1583 1584static int 1585arc_cksum_equal(arc_buf_t *buf) 1586{ 1587 zio_cksum_t zc; 1588 int equal; 1589 1590 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1591 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, NULL, &zc); 1592 equal = ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc); 1593 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1594 1595 return (equal); 1596} 1597 1598static void 1599arc_cksum_compute(arc_buf_t *buf, boolean_t force) 1600{ 1601 if (!force && !(zfs_flags & ZFS_DEBUG_MODIFY)) 1602 return; 1603 1604 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1605 if (buf->b_hdr->b_freeze_cksum != NULL) { 1606 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1607 return; 1608 } 1609 buf->b_hdr->b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t), KM_SLEEP); 1610 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, 1611 NULL, buf->b_hdr->b_freeze_cksum); 1612 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1613#ifdef illumos 1614 arc_buf_watch(buf); 1615#endif 1616} 1617 1618#ifdef illumos 1619#ifndef _KERNEL 1620typedef struct procctl { 1621 long cmd; 1622 prwatch_t prwatch; 1623} procctl_t; 1624#endif 1625 1626/* ARGSUSED */ 1627static void 1628arc_buf_unwatch(arc_buf_t *buf) 1629{ 1630#ifndef _KERNEL 1631 if (arc_watch) { 1632 int result; 1633 procctl_t ctl; 1634 ctl.cmd = PCWATCH; 1635 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data; 1636 ctl.prwatch.pr_size = 0; 1637 ctl.prwatch.pr_wflags = 0; 1638 result = write(arc_procfd, &ctl, sizeof (ctl)); 1639 ASSERT3U(result, ==, sizeof (ctl)); 1640 } 1641#endif 1642} 1643 1644/* ARGSUSED */ 1645static void 1646arc_buf_watch(arc_buf_t *buf) 1647{ 1648#ifndef _KERNEL 1649 if (arc_watch) { 1650 int result; 1651 procctl_t ctl; 1652 ctl.cmd = PCWATCH; 1653 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data; 1654 ctl.prwatch.pr_size = buf->b_hdr->b_size; 1655 ctl.prwatch.pr_wflags = WA_WRITE; 1656 result = write(arc_procfd, &ctl, sizeof (ctl)); 1657 ASSERT3U(result, ==, sizeof (ctl)); 1658 } 1659#endif 1660} 1661#endif /* illumos */ 1662 1663static arc_buf_contents_t 1664arc_buf_type(arc_buf_hdr_t *hdr) 1665{ 1666 if (HDR_ISTYPE_METADATA(hdr)) { 1667 return (ARC_BUFC_METADATA); 1668 } else { 1669 return (ARC_BUFC_DATA); 1670 } 1671} 1672 1673static uint32_t 1674arc_bufc_to_flags(arc_buf_contents_t type) 1675{ 1676 switch (type) { 1677 case ARC_BUFC_DATA: 1678 /* metadata field is 0 if buffer contains normal data */ 1679 return (0); 1680 case ARC_BUFC_METADATA: 1681 return (ARC_FLAG_BUFC_METADATA); 1682 default: 1683 break; 1684 } 1685 panic("undefined ARC buffer type!"); 1686 return ((uint32_t)-1); 1687} 1688 1689void 1690arc_buf_thaw(arc_buf_t *buf) 1691{ 1692 if (zfs_flags & ZFS_DEBUG_MODIFY) { 1693 if (buf->b_hdr->b_l1hdr.b_state != arc_anon) 1694 panic("modifying non-anon buffer!"); 1695 if (HDR_IO_IN_PROGRESS(buf->b_hdr)) 1696 panic("modifying buffer while i/o in progress!"); 1697 arc_cksum_verify(buf); 1698 } 1699 1700 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1701 if (buf->b_hdr->b_freeze_cksum != NULL) { 1702 kmem_free(buf->b_hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 1703 buf->b_hdr->b_freeze_cksum = NULL; 1704 } 1705 1706#ifdef ZFS_DEBUG 1707 if (zfs_flags & ZFS_DEBUG_MODIFY) { 1708 if (buf->b_hdr->b_l1hdr.b_thawed != NULL) 1709 kmem_free(buf->b_hdr->b_l1hdr.b_thawed, 1); 1710 buf->b_hdr->b_l1hdr.b_thawed = kmem_alloc(1, KM_SLEEP); 1711 } 1712#endif 1713 1714 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1715 1716#ifdef illumos 1717 arc_buf_unwatch(buf); 1718#endif 1719} 1720 1721void 1722arc_buf_freeze(arc_buf_t *buf) 1723{ 1724 kmutex_t *hash_lock; 1725 1726 if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 1727 return; 1728 1729 hash_lock = HDR_LOCK(buf->b_hdr); 1730 mutex_enter(hash_lock); 1731 1732 ASSERT(buf->b_hdr->b_freeze_cksum != NULL || 1733 buf->b_hdr->b_l1hdr.b_state == arc_anon); 1734 arc_cksum_compute(buf, B_FALSE); 1735 mutex_exit(hash_lock); 1736 1737} 1738 1739static void 1740add_reference(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, void *tag) 1741{ 1742 ASSERT(HDR_HAS_L1HDR(hdr)); 1743 ASSERT(MUTEX_HELD(hash_lock)); 1744 arc_state_t *state = hdr->b_l1hdr.b_state; 1745 1746 if ((refcount_add(&hdr->b_l1hdr.b_refcnt, tag) == 1) && 1747 (state != arc_anon)) { 1748 /* We don't use the L2-only state list. */ 1749 if (state != arc_l2c_only) { 1750 arc_buf_contents_t type = arc_buf_type(hdr); 1751 uint64_t delta = hdr->b_size * hdr->b_l1hdr.b_datacnt; 1752 multilist_t *list = &state->arcs_list[type]; 1753 uint64_t *size = &state->arcs_lsize[type]; 1754 1755 multilist_remove(list, hdr); 1756 1757 if (GHOST_STATE(state)) { 1758 ASSERT0(hdr->b_l1hdr.b_datacnt); 1759 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 1760 delta = hdr->b_size; 1761 } 1762 ASSERT(delta > 0); 1763 ASSERT3U(*size, >=, delta); 1764 atomic_add_64(size, -delta); 1765 } 1766 /* remove the prefetch flag if we get a reference */ 1767 hdr->b_flags &= ~ARC_FLAG_PREFETCH; 1768 } 1769} 1770 1771static int 1772remove_reference(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, void *tag) 1773{ 1774 int cnt; 1775 arc_state_t *state = hdr->b_l1hdr.b_state; 1776 1777 ASSERT(HDR_HAS_L1HDR(hdr)); 1778 ASSERT(state == arc_anon || MUTEX_HELD(hash_lock)); 1779 ASSERT(!GHOST_STATE(state)); 1780 1781 /* 1782 * arc_l2c_only counts as a ghost state so we don't need to explicitly 1783 * check to prevent usage of the arc_l2c_only list. 1784 */ 1785 if (((cnt = refcount_remove(&hdr->b_l1hdr.b_refcnt, tag)) == 0) && 1786 (state != arc_anon)) { 1787 arc_buf_contents_t type = arc_buf_type(hdr); 1788 multilist_t *list = &state->arcs_list[type]; 1789 uint64_t *size = &state->arcs_lsize[type]; 1790 1791 multilist_insert(list, hdr); 1792 1793 ASSERT(hdr->b_l1hdr.b_datacnt > 0); 1794 atomic_add_64(size, hdr->b_size * 1795 hdr->b_l1hdr.b_datacnt); 1796 } 1797 return (cnt); 1798} 1799 1800/* 1801 * Move the supplied buffer to the indicated state. The hash lock 1802 * for the buffer must be held by the caller. 1803 */ 1804static void 1805arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *hdr, 1806 kmutex_t *hash_lock) 1807{ 1808 arc_state_t *old_state; 1809 int64_t refcnt; 1810 uint32_t datacnt; 1811 uint64_t from_delta, to_delta; 1812 arc_buf_contents_t buftype = arc_buf_type(hdr); 1813 1814 /* 1815 * We almost always have an L1 hdr here, since we call arc_hdr_realloc() 1816 * in arc_read() when bringing a buffer out of the L2ARC. However, the 1817 * L1 hdr doesn't always exist when we change state to arc_anon before 1818 * destroying a header, in which case reallocating to add the L1 hdr is 1819 * pointless. 1820 */ 1821 if (HDR_HAS_L1HDR(hdr)) { 1822 old_state = hdr->b_l1hdr.b_state; 1823 refcnt = refcount_count(&hdr->b_l1hdr.b_refcnt); 1824 datacnt = hdr->b_l1hdr.b_datacnt; 1825 } else { 1826 old_state = arc_l2c_only; 1827 refcnt = 0; 1828 datacnt = 0; 1829 } 1830 1831 ASSERT(MUTEX_HELD(hash_lock)); 1832 ASSERT3P(new_state, !=, old_state); 1833 ASSERT(refcnt == 0 || datacnt > 0); 1834 ASSERT(!GHOST_STATE(new_state) || datacnt == 0); 1835 ASSERT(old_state != arc_anon || datacnt <= 1); 1836 1837 from_delta = to_delta = datacnt * hdr->b_size; 1838 1839 /* 1840 * If this buffer is evictable, transfer it from the 1841 * old state list to the new state list. 1842 */ 1843 if (refcnt == 0) { 1844 if (old_state != arc_anon && old_state != arc_l2c_only) { 1845 uint64_t *size = &old_state->arcs_lsize[buftype]; 1846 1847 ASSERT(HDR_HAS_L1HDR(hdr)); 1848 multilist_remove(&old_state->arcs_list[buftype], hdr); 1849 1850 /* 1851 * If prefetching out of the ghost cache, 1852 * we will have a non-zero datacnt. 1853 */ 1854 if (GHOST_STATE(old_state) && datacnt == 0) { 1855 /* ghost elements have a ghost size */ 1856 ASSERT(hdr->b_l1hdr.b_buf == NULL); 1857 from_delta = hdr->b_size; 1858 } 1859 ASSERT3U(*size, >=, from_delta); 1860 atomic_add_64(size, -from_delta); 1861 } 1862 if (new_state != arc_anon && new_state != arc_l2c_only) { 1863 uint64_t *size = &new_state->arcs_lsize[buftype]; 1864 1865 /* 1866 * An L1 header always exists here, since if we're 1867 * moving to some L1-cached state (i.e. not l2c_only or 1868 * anonymous), we realloc the header to add an L1hdr 1869 * beforehand. 1870 */ 1871 ASSERT(HDR_HAS_L1HDR(hdr)); 1872 multilist_insert(&new_state->arcs_list[buftype], hdr); 1873 1874 /* ghost elements have a ghost size */ 1875 if (GHOST_STATE(new_state)) { 1876 ASSERT0(datacnt); 1877 ASSERT(hdr->b_l1hdr.b_buf == NULL); 1878 to_delta = hdr->b_size; 1879 } 1880 atomic_add_64(size, to_delta); 1881 } 1882 } 1883 1884 ASSERT(!BUF_EMPTY(hdr)); 1885 if (new_state == arc_anon && HDR_IN_HASH_TABLE(hdr)) 1886 buf_hash_remove(hdr); 1887 1888 /* adjust state sizes (ignore arc_l2c_only) */ 1889 1890 if (to_delta && new_state != arc_l2c_only) { 1891 ASSERT(HDR_HAS_L1HDR(hdr)); 1892 if (GHOST_STATE(new_state)) { 1893 ASSERT0(datacnt); 1894 1895 /* 1896 * We moving a header to a ghost state, we first 1897 * remove all arc buffers. Thus, we'll have a 1898 * datacnt of zero, and no arc buffer to use for 1899 * the reference. As a result, we use the arc 1900 * header pointer for the reference. 1901 */ 1902 (void) refcount_add_many(&new_state->arcs_size, 1903 hdr->b_size, hdr); 1904 } else { 1905 ASSERT3U(datacnt, !=, 0); 1906 1907 /* 1908 * Each individual buffer holds a unique reference, 1909 * thus we must remove each of these references one 1910 * at a time. 1911 */ 1912 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL; 1913 buf = buf->b_next) { 1914 (void) refcount_add_many(&new_state->arcs_size, 1915 hdr->b_size, buf); 1916 } 1917 } 1918 } 1919 1920 if (from_delta && old_state != arc_l2c_only) { 1921 ASSERT(HDR_HAS_L1HDR(hdr)); 1922 if (GHOST_STATE(old_state)) { 1923 /* 1924 * When moving a header off of a ghost state, 1925 * there's the possibility for datacnt to be 1926 * non-zero. This is because we first add the 1927 * arc buffer to the header prior to changing 1928 * the header's state. Since we used the header 1929 * for the reference when putting the header on 1930 * the ghost state, we must balance that and use 1931 * the header when removing off the ghost state 1932 * (even though datacnt is non zero). 1933 */ 1934 1935 IMPLY(datacnt == 0, new_state == arc_anon || 1936 new_state == arc_l2c_only); 1937 1938 (void) refcount_remove_many(&old_state->arcs_size, 1939 hdr->b_size, hdr); 1940 } else { 1941 ASSERT3P(datacnt, !=, 0); 1942 1943 /* 1944 * Each individual buffer holds a unique reference, 1945 * thus we must remove each of these references one 1946 * at a time. 1947 */ 1948 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL; 1949 buf = buf->b_next) { 1950 (void) refcount_remove_many( 1951 &old_state->arcs_size, hdr->b_size, buf); 1952 } 1953 } 1954 } 1955 1956 if (HDR_HAS_L1HDR(hdr)) 1957 hdr->b_l1hdr.b_state = new_state; 1958 1959 /* 1960 * L2 headers should never be on the L2 state list since they don't 1961 * have L1 headers allocated. 1962 */ 1963 ASSERT(multilist_is_empty(&arc_l2c_only->arcs_list[ARC_BUFC_DATA]) && 1964 multilist_is_empty(&arc_l2c_only->arcs_list[ARC_BUFC_METADATA])); 1965} 1966 1967void 1968arc_space_consume(uint64_t space, arc_space_type_t type) 1969{ 1970 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES); 1971 1972 switch (type) { 1973 case ARC_SPACE_DATA: 1974 ARCSTAT_INCR(arcstat_data_size, space); 1975 break; 1976 case ARC_SPACE_META: 1977 ARCSTAT_INCR(arcstat_metadata_size, space); 1978 break; 1979 case ARC_SPACE_OTHER: 1980 ARCSTAT_INCR(arcstat_other_size, space); 1981 break; 1982 case ARC_SPACE_HDRS: 1983 ARCSTAT_INCR(arcstat_hdr_size, space); 1984 break; 1985 case ARC_SPACE_L2HDRS: 1986 ARCSTAT_INCR(arcstat_l2_hdr_size, space); 1987 break; 1988 } 1989 1990 if (type != ARC_SPACE_DATA) 1991 ARCSTAT_INCR(arcstat_meta_used, space); 1992 1993 atomic_add_64(&arc_size, space); 1994} 1995 1996void 1997arc_space_return(uint64_t space, arc_space_type_t type) 1998{ 1999 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES); 2000 2001 switch (type) { 2002 case ARC_SPACE_DATA: 2003 ARCSTAT_INCR(arcstat_data_size, -space); 2004 break; 2005 case ARC_SPACE_META: 2006 ARCSTAT_INCR(arcstat_metadata_size, -space); 2007 break; 2008 case ARC_SPACE_OTHER: 2009 ARCSTAT_INCR(arcstat_other_size, -space); 2010 break; 2011 case ARC_SPACE_HDRS: 2012 ARCSTAT_INCR(arcstat_hdr_size, -space); 2013 break; 2014 case ARC_SPACE_L2HDRS: 2015 ARCSTAT_INCR(arcstat_l2_hdr_size, -space); 2016 break; 2017 } 2018 2019 if (type != ARC_SPACE_DATA) { 2020 ASSERT(arc_meta_used >= space); 2021 if (arc_meta_max < arc_meta_used) 2022 arc_meta_max = arc_meta_used; 2023 ARCSTAT_INCR(arcstat_meta_used, -space); 2024 } 2025 2026 ASSERT(arc_size >= space); 2027 atomic_add_64(&arc_size, -space); 2028} 2029 2030arc_buf_t * 2031arc_buf_alloc(spa_t *spa, int32_t size, void *tag, arc_buf_contents_t type) 2032{ 2033 arc_buf_hdr_t *hdr; 2034 arc_buf_t *buf; 2035 2036 ASSERT3U(size, >, 0); 2037 hdr = kmem_cache_alloc(hdr_full_cache, KM_PUSHPAGE); 2038 ASSERT(BUF_EMPTY(hdr)); 2039 ASSERT3P(hdr->b_freeze_cksum, ==, NULL); 2040 hdr->b_size = size; 2041 hdr->b_spa = spa_load_guid(spa); 2042 2043 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 2044 buf->b_hdr = hdr; 2045 buf->b_data = NULL; 2046 buf->b_efunc = NULL; 2047 buf->b_private = NULL; 2048 buf->b_next = NULL; 2049 2050 hdr->b_flags = arc_bufc_to_flags(type); 2051 hdr->b_flags |= ARC_FLAG_HAS_L1HDR; 2052 2053 hdr->b_l1hdr.b_buf = buf; 2054 hdr->b_l1hdr.b_state = arc_anon; 2055 hdr->b_l1hdr.b_arc_access = 0; 2056 hdr->b_l1hdr.b_datacnt = 1; 2057 hdr->b_l1hdr.b_tmp_cdata = NULL; 2058 2059 arc_get_data_buf(buf); 2060 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 2061 (void) refcount_add(&hdr->b_l1hdr.b_refcnt, tag); 2062 2063 return (buf); 2064} 2065 2066static char *arc_onloan_tag = "onloan"; 2067 2068/* 2069 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in 2070 * flight data by arc_tempreserve_space() until they are "returned". Loaned 2071 * buffers must be returned to the arc before they can be used by the DMU or 2072 * freed. 2073 */ 2074arc_buf_t * 2075arc_loan_buf(spa_t *spa, int size) 2076{ 2077 arc_buf_t *buf; 2078 2079 buf = arc_buf_alloc(spa, size, arc_onloan_tag, ARC_BUFC_DATA); 2080 2081 atomic_add_64(&arc_loaned_bytes, size); 2082 return (buf); 2083} 2084 2085/* 2086 * Return a loaned arc buffer to the arc. 2087 */ 2088void 2089arc_return_buf(arc_buf_t *buf, void *tag) 2090{ 2091 arc_buf_hdr_t *hdr = buf->b_hdr; 2092 2093 ASSERT(buf->b_data != NULL); 2094 ASSERT(HDR_HAS_L1HDR(hdr)); 2095 (void) refcount_add(&hdr->b_l1hdr.b_refcnt, tag); 2096 (void) refcount_remove(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag); 2097 2098 atomic_add_64(&arc_loaned_bytes, -hdr->b_size); 2099} 2100 2101/* Detach an arc_buf from a dbuf (tag) */ 2102void 2103arc_loan_inuse_buf(arc_buf_t *buf, void *tag) 2104{ 2105 arc_buf_hdr_t *hdr = buf->b_hdr; 2106 2107 ASSERT(buf->b_data != NULL); 2108 ASSERT(HDR_HAS_L1HDR(hdr)); 2109 (void) refcount_add(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag); 2110 (void) refcount_remove(&hdr->b_l1hdr.b_refcnt, tag); 2111 buf->b_efunc = NULL; 2112 buf->b_private = NULL; 2113 2114 atomic_add_64(&arc_loaned_bytes, hdr->b_size); 2115} 2116 2117static arc_buf_t * 2118arc_buf_clone(arc_buf_t *from) 2119{ 2120 arc_buf_t *buf; 2121 arc_buf_hdr_t *hdr = from->b_hdr; 2122 uint64_t size = hdr->b_size; 2123 2124 ASSERT(HDR_HAS_L1HDR(hdr)); 2125 ASSERT(hdr->b_l1hdr.b_state != arc_anon); 2126 2127 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 2128 buf->b_hdr = hdr; 2129 buf->b_data = NULL; 2130 buf->b_efunc = NULL; 2131 buf->b_private = NULL; 2132 buf->b_next = hdr->b_l1hdr.b_buf; 2133 hdr->b_l1hdr.b_buf = buf; 2134 arc_get_data_buf(buf); 2135 bcopy(from->b_data, buf->b_data, size); 2136 2137 /* 2138 * This buffer already exists in the arc so create a duplicate 2139 * copy for the caller. If the buffer is associated with user data 2140 * then track the size and number of duplicates. These stats will be 2141 * updated as duplicate buffers are created and destroyed. 2142 */ 2143 if (HDR_ISTYPE_DATA(hdr)) { 2144 ARCSTAT_BUMP(arcstat_duplicate_buffers); 2145 ARCSTAT_INCR(arcstat_duplicate_buffers_size, size); 2146 } 2147 hdr->b_l1hdr.b_datacnt += 1; 2148 return (buf); 2149} 2150 2151void 2152arc_buf_add_ref(arc_buf_t *buf, void* tag) 2153{ 2154 arc_buf_hdr_t *hdr; 2155 kmutex_t *hash_lock; 2156 2157 /* 2158 * Check to see if this buffer is evicted. Callers 2159 * must verify b_data != NULL to know if the add_ref 2160 * was successful. 2161 */ 2162 mutex_enter(&buf->b_evict_lock); 2163 if (buf->b_data == NULL) { 2164 mutex_exit(&buf->b_evict_lock); 2165 return; 2166 } 2167 hash_lock = HDR_LOCK(buf->b_hdr); 2168 mutex_enter(hash_lock); 2169 hdr = buf->b_hdr; 2170 ASSERT(HDR_HAS_L1HDR(hdr)); 2171 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 2172 mutex_exit(&buf->b_evict_lock); 2173 2174 ASSERT(hdr->b_l1hdr.b_state == arc_mru || 2175 hdr->b_l1hdr.b_state == arc_mfu); 2176 2177 add_reference(hdr, hash_lock, tag); 2178 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr); 2179 arc_access(hdr, hash_lock); 2180 mutex_exit(hash_lock); 2181 ARCSTAT_BUMP(arcstat_hits); 2182 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr), 2183 demand, prefetch, !HDR_ISTYPE_METADATA(hdr), 2184 data, metadata, hits); 2185} 2186 2187static void 2188arc_buf_free_on_write(void *data, size_t size, 2189 void (*free_func)(void *, size_t)) 2190{ 2191 l2arc_data_free_t *df; 2192 2193 df = kmem_alloc(sizeof (*df), KM_SLEEP); 2194 df->l2df_data = data; 2195 df->l2df_size = size; 2196 df->l2df_func = free_func; 2197 mutex_enter(&l2arc_free_on_write_mtx); 2198 list_insert_head(l2arc_free_on_write, df); 2199 mutex_exit(&l2arc_free_on_write_mtx); 2200} 2201 2202/* 2203 * Free the arc data buffer. If it is an l2arc write in progress, 2204 * the buffer is placed on l2arc_free_on_write to be freed later. 2205 */ 2206static void 2207arc_buf_data_free(arc_buf_t *buf, void (*free_func)(void *, size_t)) 2208{ 2209 arc_buf_hdr_t *hdr = buf->b_hdr; 2210 2211 if (HDR_L2_WRITING(hdr)) { 2212 arc_buf_free_on_write(buf->b_data, hdr->b_size, free_func); 2213 ARCSTAT_BUMP(arcstat_l2_free_on_write); 2214 } else { 2215 free_func(buf->b_data, hdr->b_size); 2216 } 2217} 2218 2219static void 2220arc_buf_l2_cdata_free(arc_buf_hdr_t *hdr) 2221{ 2222 size_t align, asize, len; 2223 2224 ASSERT(HDR_HAS_L2HDR(hdr)); 2225 ASSERT(MUTEX_HELD(&hdr->b_l2hdr.b_dev->l2ad_mtx)); 2226 2227 /* 2228 * The b_tmp_cdata field is linked off of the b_l1hdr, so if 2229 * that doesn't exist, the header is in the arc_l2c_only state, 2230 * and there isn't anything to free (it's already been freed). 2231 */ 2232 if (!HDR_HAS_L1HDR(hdr)) 2233 return; 2234 2235 /* 2236 * The header isn't being written to the l2arc device, thus it 2237 * shouldn't have a b_tmp_cdata to free. 2238 */ 2239 if (!HDR_L2_WRITING(hdr)) { 2240 ASSERT3P(hdr->b_l1hdr.b_tmp_cdata, ==, NULL); 2241 return; 2242 } 2243 2244 /* 2245 * The bufer has been chosen for writing to L2ARC, but it's 2246 * not being written just yet. In other words, 2247 * b_tmp_cdata points to exactly the same buffer as b_data, 2248 * l2arc_transform_buf hasn't been called. 2249 */ 2250 if (hdr->b_l2hdr.b_daddr == L2ARC_ADDR_UNSET) { 2251 ASSERT3P(hdr->b_l1hdr.b_tmp_cdata, ==, 2252 hdr->b_l1hdr.b_buf->b_data); 2253 ASSERT3U(hdr->b_l2hdr.b_compress, ==, ZIO_COMPRESS_OFF); 2254 return; 2255 } 2256 2257 /* 2258 * There's nothing to free since the buffer was all zero's and 2259 * compressed to a zero length buffer. 2260 */ 2261 if (hdr->b_l2hdr.b_compress == ZIO_COMPRESS_EMPTY) { 2262 ASSERT3P(hdr->b_l1hdr.b_tmp_cdata, ==, NULL); 2263 return; 2264 } 2265 2266 /* 2267 * Nothing to do if the temporary buffer was not required. 2268 */ 2269 if (hdr->b_l1hdr.b_tmp_cdata == NULL) 2270 return; 2271 2272 ARCSTAT_BUMP(arcstat_l2_cdata_free_on_write); 2273 len = hdr->b_size; 2274 align = (size_t)1 << hdr->b_l2hdr.b_dev->l2ad_vdev->vdev_ashift; 2275 asize = P2ROUNDUP(len, align); 2276 arc_buf_free_on_write(hdr->b_l1hdr.b_tmp_cdata, asize, 2277 zio_data_buf_free); 2278 hdr->b_l1hdr.b_tmp_cdata = NULL; 2279} 2280 2281/* 2282 * Free up buf->b_data and if 'remove' is set, then pull the 2283 * arc_buf_t off of the the arc_buf_hdr_t's list and free it. 2284 */ 2285static void 2286arc_buf_destroy(arc_buf_t *buf, boolean_t remove) 2287{ 2288 arc_buf_t **bufp; 2289 2290 /* free up data associated with the buf */ 2291 if (buf->b_data != NULL) { 2292 arc_state_t *state = buf->b_hdr->b_l1hdr.b_state; 2293 uint64_t size = buf->b_hdr->b_size; 2294 arc_buf_contents_t type = arc_buf_type(buf->b_hdr); 2295 2296 arc_cksum_verify(buf); 2297#ifdef illumos 2298 arc_buf_unwatch(buf); 2299#endif 2300 2301 if (type == ARC_BUFC_METADATA) { 2302 arc_buf_data_free(buf, zio_buf_free); 2303 arc_space_return(size, ARC_SPACE_META); 2304 } else { 2305 ASSERT(type == ARC_BUFC_DATA); 2306 arc_buf_data_free(buf, zio_data_buf_free); 2307 arc_space_return(size, ARC_SPACE_DATA); 2308 } 2309 2310 /* protected by hash lock, if in the hash table */ 2311 if (multilist_link_active(&buf->b_hdr->b_l1hdr.b_arc_node)) { 2312 uint64_t *cnt = &state->arcs_lsize[type]; 2313 2314 ASSERT(refcount_is_zero( 2315 &buf->b_hdr->b_l1hdr.b_refcnt)); 2316 ASSERT(state != arc_anon && state != arc_l2c_only); 2317 2318 ASSERT3U(*cnt, >=, size); 2319 atomic_add_64(cnt, -size); 2320 } 2321 2322 (void) refcount_remove_many(&state->arcs_size, size, buf); 2323 buf->b_data = NULL; 2324 2325 /* 2326 * If we're destroying a duplicate buffer make sure 2327 * that the appropriate statistics are updated. 2328 */ 2329 if (buf->b_hdr->b_l1hdr.b_datacnt > 1 && 2330 HDR_ISTYPE_DATA(buf->b_hdr)) { 2331 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers); 2332 ARCSTAT_INCR(arcstat_duplicate_buffers_size, -size); 2333 } 2334 ASSERT(buf->b_hdr->b_l1hdr.b_datacnt > 0); 2335 buf->b_hdr->b_l1hdr.b_datacnt -= 1; 2336 } 2337 2338 /* only remove the buf if requested */ 2339 if (!remove) 2340 return; 2341 2342 /* remove the buf from the hdr list */ 2343 for (bufp = &buf->b_hdr->b_l1hdr.b_buf; *bufp != buf; 2344 bufp = &(*bufp)->b_next) 2345 continue; 2346 *bufp = buf->b_next; 2347 buf->b_next = NULL; 2348 2349 ASSERT(buf->b_efunc == NULL); 2350 2351 /* clean up the buf */ 2352 buf->b_hdr = NULL; 2353 kmem_cache_free(buf_cache, buf); 2354} 2355 2356static void 2357arc_hdr_l2hdr_destroy(arc_buf_hdr_t *hdr) 2358{ 2359 l2arc_buf_hdr_t *l2hdr = &hdr->b_l2hdr; 2360 l2arc_dev_t *dev = l2hdr->b_dev; 2361 2362 ASSERT(MUTEX_HELD(&dev->l2ad_mtx)); 2363 ASSERT(HDR_HAS_L2HDR(hdr)); 2364 2365 list_remove(&dev->l2ad_buflist, hdr); 2366 2367 /* 2368 * We don't want to leak the b_tmp_cdata buffer that was 2369 * allocated in l2arc_write_buffers() 2370 */ 2371 arc_buf_l2_cdata_free(hdr); 2372 2373 /* 2374 * If the l2hdr's b_daddr is equal to L2ARC_ADDR_UNSET, then 2375 * this header is being processed by l2arc_write_buffers() (i.e. 2376 * it's in the first stage of l2arc_write_buffers()). 2377 * Re-affirming that truth here, just to serve as a reminder. If 2378 * b_daddr does not equal L2ARC_ADDR_UNSET, then the header may or 2379 * may not have its HDR_L2_WRITING flag set. (the write may have 2380 * completed, in which case HDR_L2_WRITING will be false and the 2381 * b_daddr field will point to the address of the buffer on disk). 2382 */ 2383 IMPLY(l2hdr->b_daddr == L2ARC_ADDR_UNSET, HDR_L2_WRITING(hdr)); 2384 2385 /* 2386 * If b_daddr is equal to L2ARC_ADDR_UNSET, we're racing with 2387 * l2arc_write_buffers(). Since we've just removed this header 2388 * from the l2arc buffer list, this header will never reach the 2389 * second stage of l2arc_write_buffers(), which increments the 2390 * accounting stats for this header. Thus, we must be careful 2391 * not to decrement them for this header either. 2392 */ 2393 if (l2hdr->b_daddr != L2ARC_ADDR_UNSET) { 2394 ARCSTAT_INCR(arcstat_l2_asize, -l2hdr->b_asize); 2395 ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size); 2396 2397 vdev_space_update(dev->l2ad_vdev, 2398 -l2hdr->b_asize, 0, 0); 2399 2400 (void) refcount_remove_many(&dev->l2ad_alloc, 2401 l2hdr->b_asize, hdr); 2402 } 2403 2404 hdr->b_flags &= ~ARC_FLAG_HAS_L2HDR; 2405} 2406 2407static void 2408arc_hdr_destroy(arc_buf_hdr_t *hdr) 2409{ 2410 if (HDR_HAS_L1HDR(hdr)) { 2411 ASSERT(hdr->b_l1hdr.b_buf == NULL || 2412 hdr->b_l1hdr.b_datacnt > 0); 2413 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 2414 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon); 2415 } 2416 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 2417 ASSERT(!HDR_IN_HASH_TABLE(hdr)); 2418 2419 if (HDR_HAS_L2HDR(hdr)) { 2420 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev; 2421 boolean_t buflist_held = MUTEX_HELD(&dev->l2ad_mtx); 2422 2423 if (!buflist_held) 2424 mutex_enter(&dev->l2ad_mtx); 2425 2426 /* 2427 * Even though we checked this conditional above, we 2428 * need to check this again now that we have the 2429 * l2ad_mtx. This is because we could be racing with 2430 * another thread calling l2arc_evict() which might have 2431 * destroyed this header's L2 portion as we were waiting 2432 * to acquire the l2ad_mtx. If that happens, we don't 2433 * want to re-destroy the header's L2 portion. 2434 */ 2435 if (HDR_HAS_L2HDR(hdr)) { 2436 l2arc_trim(hdr); 2437 arc_hdr_l2hdr_destroy(hdr); 2438 } 2439 2440 if (!buflist_held) 2441 mutex_exit(&dev->l2ad_mtx); 2442 } 2443 2444 if (!BUF_EMPTY(hdr)) 2445 buf_discard_identity(hdr); 2446 2447 if (hdr->b_freeze_cksum != NULL) { 2448 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 2449 hdr->b_freeze_cksum = NULL; 2450 } 2451 2452 if (HDR_HAS_L1HDR(hdr)) { 2453 while (hdr->b_l1hdr.b_buf) { 2454 arc_buf_t *buf = hdr->b_l1hdr.b_buf; 2455 2456 if (buf->b_efunc != NULL) { 2457 mutex_enter(&arc_user_evicts_lock); 2458 mutex_enter(&buf->b_evict_lock); 2459 ASSERT(buf->b_hdr != NULL); 2460 arc_buf_destroy(hdr->b_l1hdr.b_buf, FALSE); 2461 hdr->b_l1hdr.b_buf = buf->b_next; 2462 buf->b_hdr = &arc_eviction_hdr; 2463 buf->b_next = arc_eviction_list; 2464 arc_eviction_list = buf; 2465 mutex_exit(&buf->b_evict_lock); 2466 cv_signal(&arc_user_evicts_cv); 2467 mutex_exit(&arc_user_evicts_lock); 2468 } else { 2469 arc_buf_destroy(hdr->b_l1hdr.b_buf, TRUE); 2470 } 2471 } 2472#ifdef ZFS_DEBUG 2473 if (hdr->b_l1hdr.b_thawed != NULL) { 2474 kmem_free(hdr->b_l1hdr.b_thawed, 1); 2475 hdr->b_l1hdr.b_thawed = NULL; 2476 } 2477#endif 2478 } 2479 2480 ASSERT3P(hdr->b_hash_next, ==, NULL); 2481 if (HDR_HAS_L1HDR(hdr)) { 2482 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 2483 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL); 2484 kmem_cache_free(hdr_full_cache, hdr); 2485 } else { 2486 kmem_cache_free(hdr_l2only_cache, hdr); 2487 } 2488} 2489 2490void 2491arc_buf_free(arc_buf_t *buf, void *tag) 2492{ 2493 arc_buf_hdr_t *hdr = buf->b_hdr; 2494 int hashed = hdr->b_l1hdr.b_state != arc_anon; 2495 2496 ASSERT(buf->b_efunc == NULL); 2497 ASSERT(buf->b_data != NULL); 2498 2499 if (hashed) { 2500 kmutex_t *hash_lock = HDR_LOCK(hdr); 2501 2502 mutex_enter(hash_lock); 2503 hdr = buf->b_hdr; 2504 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 2505 2506 (void) remove_reference(hdr, hash_lock, tag); 2507 if (hdr->b_l1hdr.b_datacnt > 1) { 2508 arc_buf_destroy(buf, TRUE); 2509 } else { 2510 ASSERT(buf == hdr->b_l1hdr.b_buf); 2511 ASSERT(buf->b_efunc == NULL); 2512 hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE; 2513 } 2514 mutex_exit(hash_lock); 2515 } else if (HDR_IO_IN_PROGRESS(hdr)) { 2516 int destroy_hdr; 2517 /* 2518 * We are in the middle of an async write. Don't destroy 2519 * this buffer unless the write completes before we finish 2520 * decrementing the reference count. 2521 */ 2522 mutex_enter(&arc_user_evicts_lock); 2523 (void) remove_reference(hdr, NULL, tag); 2524 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 2525 destroy_hdr = !HDR_IO_IN_PROGRESS(hdr); 2526 mutex_exit(&arc_user_evicts_lock); 2527 if (destroy_hdr) 2528 arc_hdr_destroy(hdr); 2529 } else { 2530 if (remove_reference(hdr, NULL, tag) > 0) 2531 arc_buf_destroy(buf, TRUE); 2532 else 2533 arc_hdr_destroy(hdr); 2534 } 2535} 2536 2537boolean_t 2538arc_buf_remove_ref(arc_buf_t *buf, void* tag) 2539{ 2540 arc_buf_hdr_t *hdr = buf->b_hdr; 2541 kmutex_t *hash_lock = HDR_LOCK(hdr); 2542 boolean_t no_callback = (buf->b_efunc == NULL); 2543 2544 if (hdr->b_l1hdr.b_state == arc_anon) { 2545 ASSERT(hdr->b_l1hdr.b_datacnt == 1); 2546 arc_buf_free(buf, tag); 2547 return (no_callback); 2548 } 2549 2550 mutex_enter(hash_lock); 2551 hdr = buf->b_hdr; 2552 ASSERT(hdr->b_l1hdr.b_datacnt > 0); 2553 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 2554 ASSERT(hdr->b_l1hdr.b_state != arc_anon); 2555 ASSERT(buf->b_data != NULL); 2556 2557 (void) remove_reference(hdr, hash_lock, tag); 2558 if (hdr->b_l1hdr.b_datacnt > 1) { 2559 if (no_callback) 2560 arc_buf_destroy(buf, TRUE); 2561 } else if (no_callback) { 2562 ASSERT(hdr->b_l1hdr.b_buf == buf && buf->b_next == NULL); 2563 ASSERT(buf->b_efunc == NULL); 2564 hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE; 2565 } 2566 ASSERT(no_callback || hdr->b_l1hdr.b_datacnt > 1 || 2567 refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 2568 mutex_exit(hash_lock); 2569 return (no_callback); 2570} 2571 2572int32_t 2573arc_buf_size(arc_buf_t *buf) 2574{ 2575 return (buf->b_hdr->b_size); 2576} 2577 2578/* 2579 * Called from the DMU to determine if the current buffer should be 2580 * evicted. In order to ensure proper locking, the eviction must be initiated 2581 * from the DMU. Return true if the buffer is associated with user data and 2582 * duplicate buffers still exist. 2583 */ 2584boolean_t 2585arc_buf_eviction_needed(arc_buf_t *buf) 2586{ 2587 arc_buf_hdr_t *hdr; 2588 boolean_t evict_needed = B_FALSE; 2589 2590 if (zfs_disable_dup_eviction) 2591 return (B_FALSE); 2592 2593 mutex_enter(&buf->b_evict_lock); 2594 hdr = buf->b_hdr; 2595 if (hdr == NULL) { 2596 /* 2597 * We are in arc_do_user_evicts(); let that function 2598 * perform the eviction. 2599 */ 2600 ASSERT(buf->b_data == NULL); 2601 mutex_exit(&buf->b_evict_lock); 2602 return (B_FALSE); 2603 } else if (buf->b_data == NULL) { 2604 /* 2605 * We have already been added to the arc eviction list; 2606 * recommend eviction. 2607 */ 2608 ASSERT3P(hdr, ==, &arc_eviction_hdr); 2609 mutex_exit(&buf->b_evict_lock); 2610 return (B_TRUE); 2611 } 2612 2613 if (hdr->b_l1hdr.b_datacnt > 1 && HDR_ISTYPE_DATA(hdr)) 2614 evict_needed = B_TRUE; 2615 2616 mutex_exit(&buf->b_evict_lock); 2617 return (evict_needed); 2618} 2619 2620/* 2621 * Evict the arc_buf_hdr that is provided as a parameter. The resultant 2622 * state of the header is dependent on it's state prior to entering this 2623 * function. The following transitions are possible: 2624 * 2625 * - arc_mru -> arc_mru_ghost 2626 * - arc_mfu -> arc_mfu_ghost 2627 * - arc_mru_ghost -> arc_l2c_only 2628 * - arc_mru_ghost -> deleted 2629 * - arc_mfu_ghost -> arc_l2c_only 2630 * - arc_mfu_ghost -> deleted 2631 */ 2632static int64_t 2633arc_evict_hdr(arc_buf_hdr_t *hdr, kmutex_t *hash_lock) 2634{ 2635 arc_state_t *evicted_state, *state; 2636 int64_t bytes_evicted = 0; 2637 2638 ASSERT(MUTEX_HELD(hash_lock)); 2639 ASSERT(HDR_HAS_L1HDR(hdr)); 2640 2641 state = hdr->b_l1hdr.b_state; 2642 if (GHOST_STATE(state)) { 2643 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 2644 ASSERT(hdr->b_l1hdr.b_buf == NULL); 2645 2646 /* 2647 * l2arc_write_buffers() relies on a header's L1 portion 2648 * (i.e. it's b_tmp_cdata field) during it's write phase. 2649 * Thus, we cannot push a header onto the arc_l2c_only 2650 * state (removing it's L1 piece) until the header is 2651 * done being written to the l2arc. 2652 */ 2653 if (HDR_HAS_L2HDR(hdr) && HDR_L2_WRITING(hdr)) { 2654 ARCSTAT_BUMP(arcstat_evict_l2_skip); 2655 return (bytes_evicted); 2656 } 2657 2658 ARCSTAT_BUMP(arcstat_deleted); 2659 bytes_evicted += hdr->b_size; 2660 2661 DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, hdr); 2662 2663 if (HDR_HAS_L2HDR(hdr)) { 2664 /* 2665 * This buffer is cached on the 2nd Level ARC; 2666 * don't destroy the header. 2667 */ 2668 arc_change_state(arc_l2c_only, hdr, hash_lock); 2669 /* 2670 * dropping from L1+L2 cached to L2-only, 2671 * realloc to remove the L1 header. 2672 */ 2673 hdr = arc_hdr_realloc(hdr, hdr_full_cache, 2674 hdr_l2only_cache); 2675 } else { 2676 arc_change_state(arc_anon, hdr, hash_lock); 2677 arc_hdr_destroy(hdr); 2678 } 2679 return (bytes_evicted); 2680 } 2681 2682 ASSERT(state == arc_mru || state == arc_mfu); 2683 evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost; 2684 2685 /* prefetch buffers have a minimum lifespan */ 2686 if (HDR_IO_IN_PROGRESS(hdr) || 2687 ((hdr->b_flags & (ARC_FLAG_PREFETCH | ARC_FLAG_INDIRECT)) && 2688 ddi_get_lbolt() - hdr->b_l1hdr.b_arc_access < 2689 arc_min_prefetch_lifespan)) { 2690 ARCSTAT_BUMP(arcstat_evict_skip); 2691 return (bytes_evicted); 2692 } 2693 2694 ASSERT0(refcount_count(&hdr->b_l1hdr.b_refcnt)); 2695 ASSERT3U(hdr->b_l1hdr.b_datacnt, >, 0); 2696 while (hdr->b_l1hdr.b_buf) { 2697 arc_buf_t *buf = hdr->b_l1hdr.b_buf; 2698 if (!mutex_tryenter(&buf->b_evict_lock)) { 2699 ARCSTAT_BUMP(arcstat_mutex_miss); 2700 break; 2701 } 2702 if (buf->b_data != NULL) 2703 bytes_evicted += hdr->b_size; 2704 if (buf->b_efunc != NULL) { 2705 mutex_enter(&arc_user_evicts_lock); 2706 arc_buf_destroy(buf, FALSE); 2707 hdr->b_l1hdr.b_buf = buf->b_next; 2708 buf->b_hdr = &arc_eviction_hdr; 2709 buf->b_next = arc_eviction_list; 2710 arc_eviction_list = buf; 2711 cv_signal(&arc_user_evicts_cv); 2712 mutex_exit(&arc_user_evicts_lock); 2713 mutex_exit(&buf->b_evict_lock); 2714 } else { 2715 mutex_exit(&buf->b_evict_lock); 2716 arc_buf_destroy(buf, TRUE); 2717 } 2718 } 2719 2720 if (HDR_HAS_L2HDR(hdr)) { 2721 ARCSTAT_INCR(arcstat_evict_l2_cached, hdr->b_size); 2722 } else { 2723 if (l2arc_write_eligible(hdr->b_spa, hdr)) 2724 ARCSTAT_INCR(arcstat_evict_l2_eligible, hdr->b_size); 2725 else 2726 ARCSTAT_INCR(arcstat_evict_l2_ineligible, hdr->b_size); 2727 } 2728 2729 if (hdr->b_l1hdr.b_datacnt == 0) { 2730 arc_change_state(evicted_state, hdr, hash_lock); 2731 ASSERT(HDR_IN_HASH_TABLE(hdr)); 2732 hdr->b_flags |= ARC_FLAG_IN_HASH_TABLE; 2733 hdr->b_flags &= ~ARC_FLAG_BUF_AVAILABLE; 2734 DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, hdr); 2735 } 2736 2737 return (bytes_evicted); 2738} 2739 2740static uint64_t 2741arc_evict_state_impl(multilist_t *ml, int idx, arc_buf_hdr_t *marker, 2742 uint64_t spa, int64_t bytes) 2743{ 2744 multilist_sublist_t *mls; 2745 uint64_t bytes_evicted = 0; 2746 arc_buf_hdr_t *hdr; 2747 kmutex_t *hash_lock; 2748 int evict_count = 0; 2749 2750 ASSERT3P(marker, !=, NULL); 2751 IMPLY(bytes < 0, bytes == ARC_EVICT_ALL); 2752 2753 mls = multilist_sublist_lock(ml, idx); 2754 2755 for (hdr = multilist_sublist_prev(mls, marker); hdr != NULL; 2756 hdr = multilist_sublist_prev(mls, marker)) { 2757 if ((bytes != ARC_EVICT_ALL && bytes_evicted >= bytes) || 2758 (evict_count >= zfs_arc_evict_batch_limit)) 2759 break; 2760 2761 /* 2762 * To keep our iteration location, move the marker 2763 * forward. Since we're not holding hdr's hash lock, we 2764 * must be very careful and not remove 'hdr' from the 2765 * sublist. Otherwise, other consumers might mistake the 2766 * 'hdr' as not being on a sublist when they call the 2767 * multilist_link_active() function (they all rely on 2768 * the hash lock protecting concurrent insertions and 2769 * removals). multilist_sublist_move_forward() was 2770 * specifically implemented to ensure this is the case 2771 * (only 'marker' will be removed and re-inserted). 2772 */ 2773 multilist_sublist_move_forward(mls, marker); 2774 2775 /* 2776 * The only case where the b_spa field should ever be 2777 * zero, is the marker headers inserted by 2778 * arc_evict_state(). It's possible for multiple threads 2779 * to be calling arc_evict_state() concurrently (e.g. 2780 * dsl_pool_close() and zio_inject_fault()), so we must 2781 * skip any markers we see from these other threads. 2782 */ 2783 if (hdr->b_spa == 0) 2784 continue; 2785 2786 /* we're only interested in evicting buffers of a certain spa */ 2787 if (spa != 0 && hdr->b_spa != spa) { 2788 ARCSTAT_BUMP(arcstat_evict_skip); 2789 continue; 2790 } 2791 2792 hash_lock = HDR_LOCK(hdr); 2793 2794 /* 2795 * We aren't calling this function from any code path 2796 * that would already be holding a hash lock, so we're 2797 * asserting on this assumption to be defensive in case 2798 * this ever changes. Without this check, it would be 2799 * possible to incorrectly increment arcstat_mutex_miss 2800 * below (e.g. if the code changed such that we called 2801 * this function with a hash lock held). 2802 */ 2803 ASSERT(!MUTEX_HELD(hash_lock)); 2804 2805 if (mutex_tryenter(hash_lock)) { 2806 uint64_t evicted = arc_evict_hdr(hdr, hash_lock); 2807 mutex_exit(hash_lock); 2808 2809 bytes_evicted += evicted; 2810 2811 /* 2812 * If evicted is zero, arc_evict_hdr() must have 2813 * decided to skip this header, don't increment 2814 * evict_count in this case. 2815 */ 2816 if (evicted != 0) 2817 evict_count++; 2818 2819 /* 2820 * If arc_size isn't overflowing, signal any 2821 * threads that might happen to be waiting. 2822 * 2823 * For each header evicted, we wake up a single 2824 * thread. If we used cv_broadcast, we could 2825 * wake up "too many" threads causing arc_size 2826 * to significantly overflow arc_c; since 2827 * arc_get_data_buf() doesn't check for overflow 2828 * when it's woken up (it doesn't because it's 2829 * possible for the ARC to be overflowing while 2830 * full of un-evictable buffers, and the 2831 * function should proceed in this case). 2832 * 2833 * If threads are left sleeping, due to not 2834 * using cv_broadcast, they will be woken up 2835 * just before arc_reclaim_thread() sleeps. 2836 */ 2837 mutex_enter(&arc_reclaim_lock); 2838 if (!arc_is_overflowing()) 2839 cv_signal(&arc_reclaim_waiters_cv); 2840 mutex_exit(&arc_reclaim_lock); 2841 } else { 2842 ARCSTAT_BUMP(arcstat_mutex_miss); 2843 } 2844 } 2845 2846 multilist_sublist_unlock(mls); 2847 2848 return (bytes_evicted); 2849} 2850 2851/* 2852 * Evict buffers from the given arc state, until we've removed the 2853 * specified number of bytes. Move the removed buffers to the 2854 * appropriate evict state. 2855 * 2856 * This function makes a "best effort". It skips over any buffers 2857 * it can't get a hash_lock on, and so, may not catch all candidates. 2858 * It may also return without evicting as much space as requested. 2859 * 2860 * If bytes is specified using the special value ARC_EVICT_ALL, this 2861 * will evict all available (i.e. unlocked and evictable) buffers from 2862 * the given arc state; which is used by arc_flush(). 2863 */ 2864static uint64_t 2865arc_evict_state(arc_state_t *state, uint64_t spa, int64_t bytes, 2866 arc_buf_contents_t type) 2867{ 2868 uint64_t total_evicted = 0; 2869 multilist_t *ml = &state->arcs_list[type]; 2870 int num_sublists; 2871 arc_buf_hdr_t **markers; 2872 2873 IMPLY(bytes < 0, bytes == ARC_EVICT_ALL); 2874 2875 num_sublists = multilist_get_num_sublists(ml); 2876 2877 /* 2878 * If we've tried to evict from each sublist, made some 2879 * progress, but still have not hit the target number of bytes 2880 * to evict, we want to keep trying. The markers allow us to 2881 * pick up where we left off for each individual sublist, rather 2882 * than starting from the tail each time. 2883 */ 2884 markers = kmem_zalloc(sizeof (*markers) * num_sublists, KM_SLEEP); 2885 for (int i = 0; i < num_sublists; i++) { 2886 markers[i] = kmem_cache_alloc(hdr_full_cache, KM_SLEEP); 2887 2888 /* 2889 * A b_spa of 0 is used to indicate that this header is 2890 * a marker. This fact is used in arc_adjust_type() and 2891 * arc_evict_state_impl(). 2892 */ 2893 markers[i]->b_spa = 0; 2894 2895 multilist_sublist_t *mls = multilist_sublist_lock(ml, i); 2896 multilist_sublist_insert_tail(mls, markers[i]); 2897 multilist_sublist_unlock(mls); 2898 } 2899 2900 /* 2901 * While we haven't hit our target number of bytes to evict, or 2902 * we're evicting all available buffers. 2903 */ 2904 while (total_evicted < bytes || bytes == ARC_EVICT_ALL) { 2905 /* 2906 * Start eviction using a randomly selected sublist, 2907 * this is to try and evenly balance eviction across all 2908 * sublists. Always starting at the same sublist 2909 * (e.g. index 0) would cause evictions to favor certain 2910 * sublists over others. 2911 */ 2912 int sublist_idx = multilist_get_random_index(ml); 2913 uint64_t scan_evicted = 0; 2914 2915 for (int i = 0; i < num_sublists; i++) { 2916 uint64_t bytes_remaining; 2917 uint64_t bytes_evicted; 2918 2919 if (bytes == ARC_EVICT_ALL) 2920 bytes_remaining = ARC_EVICT_ALL; 2921 else if (total_evicted < bytes) 2922 bytes_remaining = bytes - total_evicted; 2923 else 2924 break; 2925 2926 bytes_evicted = arc_evict_state_impl(ml, sublist_idx, 2927 markers[sublist_idx], spa, bytes_remaining); 2928 2929 scan_evicted += bytes_evicted; 2930 total_evicted += bytes_evicted; 2931 2932 /* we've reached the end, wrap to the beginning */ 2933 if (++sublist_idx >= num_sublists) 2934 sublist_idx = 0; 2935 } 2936 2937 /* 2938 * If we didn't evict anything during this scan, we have 2939 * no reason to believe we'll evict more during another 2940 * scan, so break the loop. 2941 */ 2942 if (scan_evicted == 0) { 2943 /* This isn't possible, let's make that obvious */ 2944 ASSERT3S(bytes, !=, 0); 2945 2946 /* 2947 * When bytes is ARC_EVICT_ALL, the only way to 2948 * break the loop is when scan_evicted is zero. 2949 * In that case, we actually have evicted enough, 2950 * so we don't want to increment the kstat. 2951 */ 2952 if (bytes != ARC_EVICT_ALL) { 2953 ASSERT3S(total_evicted, <, bytes); 2954 ARCSTAT_BUMP(arcstat_evict_not_enough); 2955 } 2956 2957 break; 2958 } 2959 } 2960 2961 for (int i = 0; i < num_sublists; i++) { 2962 multilist_sublist_t *mls = multilist_sublist_lock(ml, i); 2963 multilist_sublist_remove(mls, markers[i]); 2964 multilist_sublist_unlock(mls); 2965 2966 kmem_cache_free(hdr_full_cache, markers[i]); 2967 } 2968 kmem_free(markers, sizeof (*markers) * num_sublists); 2969 2970 return (total_evicted); 2971} 2972 2973/* 2974 * Flush all "evictable" data of the given type from the arc state 2975 * specified. This will not evict any "active" buffers (i.e. referenced). 2976 * 2977 * When 'retry' is set to FALSE, the function will make a single pass 2978 * over the state and evict any buffers that it can. Since it doesn't 2979 * continually retry the eviction, it might end up leaving some buffers 2980 * in the ARC due to lock misses. 2981 * 2982 * When 'retry' is set to TRUE, the function will continually retry the 2983 * eviction until *all* evictable buffers have been removed from the 2984 * state. As a result, if concurrent insertions into the state are 2985 * allowed (e.g. if the ARC isn't shutting down), this function might 2986 * wind up in an infinite loop, continually trying to evict buffers. 2987 */ 2988static uint64_t 2989arc_flush_state(arc_state_t *state, uint64_t spa, arc_buf_contents_t type, 2990 boolean_t retry) 2991{ 2992 uint64_t evicted = 0; 2993 2994 while (state->arcs_lsize[type] != 0) { 2995 evicted += arc_evict_state(state, spa, ARC_EVICT_ALL, type); 2996 2997 if (!retry) 2998 break; 2999 } 3000 3001 return (evicted); 3002} 3003 3004/* 3005 * Evict the specified number of bytes from the state specified, 3006 * restricting eviction to the spa and type given. This function 3007 * prevents us from trying to evict more from a state's list than 3008 * is "evictable", and to skip evicting altogether when passed a 3009 * negative value for "bytes". In contrast, arc_evict_state() will 3010 * evict everything it can, when passed a negative value for "bytes". 3011 */ 3012static uint64_t 3013arc_adjust_impl(arc_state_t *state, uint64_t spa, int64_t bytes, 3014 arc_buf_contents_t type) 3015{ 3016 int64_t delta; 3017 3018 if (bytes > 0 && state->arcs_lsize[type] > 0) { 3019 delta = MIN(state->arcs_lsize[type], bytes); 3020 return (arc_evict_state(state, spa, delta, type)); 3021 } 3022 3023 return (0); 3024} 3025 3026/* 3027 * Evict metadata buffers from the cache, such that arc_meta_used is 3028 * capped by the arc_meta_limit tunable. 3029 */ 3030static uint64_t 3031arc_adjust_meta(void) 3032{ 3033 uint64_t total_evicted = 0; 3034 int64_t target; 3035 3036 /* 3037 * If we're over the meta limit, we want to evict enough 3038 * metadata to get back under the meta limit. We don't want to 3039 * evict so much that we drop the MRU below arc_p, though. If 3040 * we're over the meta limit more than we're over arc_p, we 3041 * evict some from the MRU here, and some from the MFU below. 3042 */ 3043 target = MIN((int64_t)(arc_meta_used - arc_meta_limit), 3044 (int64_t)(refcount_count(&arc_anon->arcs_size) + 3045 refcount_count(&arc_mru->arcs_size) - arc_p)); 3046 3047 total_evicted += arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA); 3048 3049 /* 3050 * Similar to the above, we want to evict enough bytes to get us 3051 * below the meta limit, but not so much as to drop us below the 3052 * space alloted to the MFU (which is defined as arc_c - arc_p). 3053 */ 3054 target = MIN((int64_t)(arc_meta_used - arc_meta_limit), 3055 (int64_t)(refcount_count(&arc_mfu->arcs_size) - (arc_c - arc_p))); 3056 3057 total_evicted += arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA); 3058 3059 return (total_evicted); 3060} 3061 3062/* 3063 * Return the type of the oldest buffer in the given arc state 3064 * 3065 * This function will select a random sublist of type ARC_BUFC_DATA and 3066 * a random sublist of type ARC_BUFC_METADATA. The tail of each sublist 3067 * is compared, and the type which contains the "older" buffer will be 3068 * returned. 3069 */ 3070static arc_buf_contents_t 3071arc_adjust_type(arc_state_t *state) 3072{ 3073 multilist_t *data_ml = &state->arcs_list[ARC_BUFC_DATA]; 3074 multilist_t *meta_ml = &state->arcs_list[ARC_BUFC_METADATA]; 3075 int data_idx = multilist_get_random_index(data_ml); 3076 int meta_idx = multilist_get_random_index(meta_ml); 3077 multilist_sublist_t *data_mls; 3078 multilist_sublist_t *meta_mls; 3079 arc_buf_contents_t type; 3080 arc_buf_hdr_t *data_hdr; 3081 arc_buf_hdr_t *meta_hdr; 3082 3083 /* 3084 * We keep the sublist lock until we're finished, to prevent 3085 * the headers from being destroyed via arc_evict_state(). 3086 */ 3087 data_mls = multilist_sublist_lock(data_ml, data_idx); 3088 meta_mls = multilist_sublist_lock(meta_ml, meta_idx); 3089 3090 /* 3091 * These two loops are to ensure we skip any markers that 3092 * might be at the tail of the lists due to arc_evict_state(). 3093 */ 3094 3095 for (data_hdr = multilist_sublist_tail(data_mls); data_hdr != NULL; 3096 data_hdr = multilist_sublist_prev(data_mls, data_hdr)) { 3097 if (data_hdr->b_spa != 0) 3098 break; 3099 } 3100 3101 for (meta_hdr = multilist_sublist_tail(meta_mls); meta_hdr != NULL; 3102 meta_hdr = multilist_sublist_prev(meta_mls, meta_hdr)) { 3103 if (meta_hdr->b_spa != 0) 3104 break; 3105 } 3106 3107 if (data_hdr == NULL && meta_hdr == NULL) { 3108 type = ARC_BUFC_DATA; 3109 } else if (data_hdr == NULL) { 3110 ASSERT3P(meta_hdr, !=, NULL); 3111 type = ARC_BUFC_METADATA; 3112 } else if (meta_hdr == NULL) { 3113 ASSERT3P(data_hdr, !=, NULL); 3114 type = ARC_BUFC_DATA; 3115 } else { 3116 ASSERT3P(data_hdr, !=, NULL); 3117 ASSERT3P(meta_hdr, !=, NULL); 3118 3119 /* The headers can't be on the sublist without an L1 header */ 3120 ASSERT(HDR_HAS_L1HDR(data_hdr)); 3121 ASSERT(HDR_HAS_L1HDR(meta_hdr)); 3122 3123 if (data_hdr->b_l1hdr.b_arc_access < 3124 meta_hdr->b_l1hdr.b_arc_access) { 3125 type = ARC_BUFC_DATA; 3126 } else { 3127 type = ARC_BUFC_METADATA; 3128 } 3129 } 3130 3131 multilist_sublist_unlock(meta_mls); 3132 multilist_sublist_unlock(data_mls); 3133 3134 return (type); 3135} 3136 3137/* 3138 * Evict buffers from the cache, such that arc_size is capped by arc_c. 3139 */ 3140static uint64_t 3141arc_adjust(void) 3142{ 3143 uint64_t total_evicted = 0; 3144 uint64_t bytes; 3145 int64_t target; 3146 3147 /* 3148 * If we're over arc_meta_limit, we want to correct that before 3149 * potentially evicting data buffers below. 3150 */ 3151 total_evicted += arc_adjust_meta(); 3152 3153 /* 3154 * Adjust MRU size 3155 * 3156 * If we're over the target cache size, we want to evict enough 3157 * from the list to get back to our target size. We don't want 3158 * to evict too much from the MRU, such that it drops below 3159 * arc_p. So, if we're over our target cache size more than 3160 * the MRU is over arc_p, we'll evict enough to get back to 3161 * arc_p here, and then evict more from the MFU below. 3162 */ 3163 target = MIN((int64_t)(arc_size - arc_c), 3164 (int64_t)(refcount_count(&arc_anon->arcs_size) + 3165 refcount_count(&arc_mru->arcs_size) + arc_meta_used - arc_p)); 3166 3167 /* 3168 * If we're below arc_meta_min, always prefer to evict data. 3169 * Otherwise, try to satisfy the requested number of bytes to 3170 * evict from the type which contains older buffers; in an 3171 * effort to keep newer buffers in the cache regardless of their 3172 * type. If we cannot satisfy the number of bytes from this 3173 * type, spill over into the next type. 3174 */ 3175 if (arc_adjust_type(arc_mru) == ARC_BUFC_METADATA && 3176 arc_meta_used > arc_meta_min) { 3177 bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA); 3178 total_evicted += bytes; 3179 3180 /* 3181 * If we couldn't evict our target number of bytes from 3182 * metadata, we try to get the rest from data. 3183 */ 3184 target -= bytes; 3185 3186 total_evicted += 3187 arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA); 3188 } else { 3189 bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA); 3190 total_evicted += bytes; 3191 3192 /* 3193 * If we couldn't evict our target number of bytes from 3194 * data, we try to get the rest from metadata. 3195 */ 3196 target -= bytes; 3197 3198 total_evicted += 3199 arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA); 3200 } 3201 3202 /* 3203 * Adjust MFU size 3204 * 3205 * Now that we've tried to evict enough from the MRU to get its 3206 * size back to arc_p, if we're still above the target cache 3207 * size, we evict the rest from the MFU. 3208 */ 3209 target = arc_size - arc_c; 3210 3211 if (arc_adjust_type(arc_mfu) == ARC_BUFC_METADATA && 3212 arc_meta_used > arc_meta_min) { 3213 bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA); 3214 total_evicted += bytes; 3215 3216 /* 3217 * If we couldn't evict our target number of bytes from 3218 * metadata, we try to get the rest from data. 3219 */ 3220 target -= bytes; 3221 3222 total_evicted += 3223 arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA); 3224 } else { 3225 bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA); 3226 total_evicted += bytes; 3227 3228 /* 3229 * If we couldn't evict our target number of bytes from 3230 * data, we try to get the rest from data. 3231 */ 3232 target -= bytes; 3233 3234 total_evicted += 3235 arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA); 3236 } 3237 3238 /* 3239 * Adjust ghost lists 3240 * 3241 * In addition to the above, the ARC also defines target values 3242 * for the ghost lists. The sum of the mru list and mru ghost 3243 * list should never exceed the target size of the cache, and 3244 * the sum of the mru list, mfu list, mru ghost list, and mfu 3245 * ghost list should never exceed twice the target size of the 3246 * cache. The following logic enforces these limits on the ghost 3247 * caches, and evicts from them as needed. 3248 */ 3249 target = refcount_count(&arc_mru->arcs_size) + 3250 refcount_count(&arc_mru_ghost->arcs_size) - arc_c; 3251 3252 bytes = arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_DATA); 3253 total_evicted += bytes; 3254 3255 target -= bytes; 3256 3257 total_evicted += 3258 arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_METADATA); 3259 3260 /* 3261 * We assume the sum of the mru list and mfu list is less than 3262 * or equal to arc_c (we enforced this above), which means we 3263 * can use the simpler of the two equations below: 3264 * 3265 * mru + mfu + mru ghost + mfu ghost <= 2 * arc_c 3266 * mru ghost + mfu ghost <= arc_c 3267 */ 3268 target = refcount_count(&arc_mru_ghost->arcs_size) + 3269 refcount_count(&arc_mfu_ghost->arcs_size) - arc_c; 3270 3271 bytes = arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_DATA); 3272 total_evicted += bytes; 3273 3274 target -= bytes; 3275 3276 total_evicted += 3277 arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_METADATA); 3278 3279 return (total_evicted); 3280} 3281 3282static void 3283arc_do_user_evicts(void) 3284{ 3285 mutex_enter(&arc_user_evicts_lock); 3286 while (arc_eviction_list != NULL) { 3287 arc_buf_t *buf = arc_eviction_list; 3288 arc_eviction_list = buf->b_next; 3289 mutex_enter(&buf->b_evict_lock); 3290 buf->b_hdr = NULL; 3291 mutex_exit(&buf->b_evict_lock); 3292 mutex_exit(&arc_user_evicts_lock); 3293 3294 if (buf->b_efunc != NULL) 3295 VERIFY0(buf->b_efunc(buf->b_private)); 3296 3297 buf->b_efunc = NULL; 3298 buf->b_private = NULL; 3299 kmem_cache_free(buf_cache, buf); 3300 mutex_enter(&arc_user_evicts_lock); 3301 } 3302 mutex_exit(&arc_user_evicts_lock); 3303} 3304 3305void 3306arc_flush(spa_t *spa, boolean_t retry) 3307{ 3308 uint64_t guid = 0; 3309 3310 /* 3311 * If retry is TRUE, a spa must not be specified since we have 3312 * no good way to determine if all of a spa's buffers have been 3313 * evicted from an arc state. 3314 */ 3315 ASSERT(!retry || spa == 0); 3316 3317 if (spa != NULL) 3318 guid = spa_load_guid(spa); 3319 3320 (void) arc_flush_state(arc_mru, guid, ARC_BUFC_DATA, retry); 3321 (void) arc_flush_state(arc_mru, guid, ARC_BUFC_METADATA, retry); 3322 3323 (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_DATA, retry); 3324 (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_METADATA, retry); 3325 3326 (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_DATA, retry); 3327 (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_METADATA, retry); 3328 3329 (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_DATA, retry); 3330 (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_METADATA, retry); 3331 3332 arc_do_user_evicts(); 3333 ASSERT(spa || arc_eviction_list == NULL); 3334} 3335 3336void 3337arc_shrink(int64_t to_free) 3338{ 3339 if (arc_c > arc_c_min) { 3340 DTRACE_PROBE4(arc__shrink, uint64_t, arc_c, uint64_t, 3341 arc_c_min, uint64_t, arc_p, uint64_t, to_free); 3342 if (arc_c > arc_c_min + to_free) 3343 atomic_add_64(&arc_c, -to_free); 3344 else 3345 arc_c = arc_c_min; 3346 3347 atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift)); 3348 if (arc_c > arc_size) 3349 arc_c = MAX(arc_size, arc_c_min); 3350 if (arc_p > arc_c) 3351 arc_p = (arc_c >> 1); 3352 3353 DTRACE_PROBE2(arc__shrunk, uint64_t, arc_c, uint64_t, 3354 arc_p); 3355 3356 ASSERT(arc_c >= arc_c_min); 3357 ASSERT((int64_t)arc_p >= 0); 3358 } 3359 3360 if (arc_size > arc_c) { 3361 DTRACE_PROBE2(arc__shrink_adjust, uint64_t, arc_size, 3362 uint64_t, arc_c); 3363 (void) arc_adjust(); 3364 } 3365} 3366 3367static long needfree = 0; 3368 3369typedef enum free_memory_reason_t { 3370 FMR_UNKNOWN, 3371 FMR_NEEDFREE, 3372 FMR_LOTSFREE, 3373 FMR_SWAPFS_MINFREE, 3374 FMR_PAGES_PP_MAXIMUM, 3375 FMR_HEAP_ARENA, 3376 FMR_ZIO_ARENA, 3377 FMR_ZIO_FRAG, 3378} free_memory_reason_t; 3379 3380int64_t last_free_memory; 3381free_memory_reason_t last_free_reason; 3382 3383/* 3384 * Additional reserve of pages for pp_reserve. 3385 */ 3386int64_t arc_pages_pp_reserve = 64; 3387 3388/* 3389 * Additional reserve of pages for swapfs. 3390 */ 3391int64_t arc_swapfs_reserve = 64; 3392 3393/* 3394 * Return the amount of memory that can be consumed before reclaim will be 3395 * needed. Positive if there is sufficient free memory, negative indicates 3396 * the amount of memory that needs to be freed up. 3397 */ 3398static int64_t 3399arc_available_memory(void) 3400{ 3401 int64_t lowest = INT64_MAX; 3402 int64_t n; 3403 free_memory_reason_t r = FMR_UNKNOWN; 3404 3405#ifdef _KERNEL 3406 if (needfree > 0) { 3407 n = PAGESIZE * (-needfree); 3408 if (n < lowest) { 3409 lowest = n; 3410 r = FMR_NEEDFREE; 3411 } 3412 } 3413 3414 /* 3415 * Cooperate with pagedaemon when it's time for it to scan 3416 * and reclaim some pages. 3417 */ 3418 n = PAGESIZE * ((int64_t)freemem - zfs_arc_free_target); 3419 if (n < lowest) { 3420 lowest = n; 3421 r = FMR_LOTSFREE; 3422 } 3423 3424#ifdef illumos 3425 /* 3426 * check that we're out of range of the pageout scanner. It starts to 3427 * schedule paging if freemem is less than lotsfree and needfree. 3428 * lotsfree is the high-water mark for pageout, and needfree is the 3429 * number of needed free pages. We add extra pages here to make sure 3430 * the scanner doesn't start up while we're freeing memory. 3431 */ 3432 n = PAGESIZE * (freemem - lotsfree - needfree - desfree); 3433 if (n < lowest) { 3434 lowest = n; 3435 r = FMR_LOTSFREE; 3436 } 3437 3438 /* 3439 * check to make sure that swapfs has enough space so that anon 3440 * reservations can still succeed. anon_resvmem() checks that the 3441 * availrmem is greater than swapfs_minfree, and the number of reserved 3442 * swap pages. We also add a bit of extra here just to prevent 3443 * circumstances from getting really dire. 3444 */ 3445 n = PAGESIZE * (availrmem - swapfs_minfree - swapfs_reserve - 3446 desfree - arc_swapfs_reserve); 3447 if (n < lowest) { 3448 lowest = n; 3449 r = FMR_SWAPFS_MINFREE; 3450 } 3451 3452 3453 /* 3454 * Check that we have enough availrmem that memory locking (e.g., via 3455 * mlock(3C) or memcntl(2)) can still succeed. (pages_pp_maximum 3456 * stores the number of pages that cannot be locked; when availrmem 3457 * drops below pages_pp_maximum, page locking mechanisms such as 3458 * page_pp_lock() will fail.) 3459 */ 3460 n = PAGESIZE * (availrmem - pages_pp_maximum - 3461 arc_pages_pp_reserve); 3462 if (n < lowest) { 3463 lowest = n; 3464 r = FMR_PAGES_PP_MAXIMUM; 3465 } 3466 3467#endif /* illumos */ 3468#if defined(__i386) || !defined(UMA_MD_SMALL_ALLOC) 3469 /* 3470 * If we're on an i386 platform, it's possible that we'll exhaust the 3471 * kernel heap space before we ever run out of available physical 3472 * memory. Most checks of the size of the heap_area compare against 3473 * tune.t_minarmem, which is the minimum available real memory that we 3474 * can have in the system. However, this is generally fixed at 25 pages 3475 * which is so low that it's useless. In this comparison, we seek to 3476 * calculate the total heap-size, and reclaim if more than 3/4ths of the 3477 * heap is allocated. (Or, in the calculation, if less than 1/4th is 3478 * free) 3479 */ 3480 n = (int64_t)vmem_size(heap_arena, VMEM_FREE) - 3481 (vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC) >> 2); 3482 if (n < lowest) { 3483 lowest = n; 3484 r = FMR_HEAP_ARENA; 3485 } 3486#define zio_arena NULL 3487#else 3488#define zio_arena heap_arena 3489#endif 3490 3491 /* 3492 * If zio data pages are being allocated out of a separate heap segment, 3493 * then enforce that the size of available vmem for this arena remains 3494 * above about 1/16th free. 3495 * 3496 * Note: The 1/16th arena free requirement was put in place 3497 * to aggressively evict memory from the arc in order to avoid 3498 * memory fragmentation issues. 3499 */ 3500 if (zio_arena != NULL) { 3501 n = (int64_t)vmem_size(zio_arena, VMEM_FREE) - 3502 (vmem_size(zio_arena, VMEM_ALLOC) >> 4); 3503 if (n < lowest) { 3504 lowest = n; 3505 r = FMR_ZIO_ARENA; 3506 } 3507 } 3508 3509 /* 3510 * Above limits know nothing about real level of KVA fragmentation. 3511 * Start aggressive reclamation if too little sequential KVA left. 3512 */ 3513 if (lowest > 0) { 3514 n = (vmem_size(heap_arena, VMEM_MAXFREE) < zfs_max_recordsize) ? 3515 -((int64_t)vmem_size(heap_arena, VMEM_ALLOC) >> 4) : 3516 INT64_MAX; 3517 if (n < lowest) { 3518 lowest = n; 3519 r = FMR_ZIO_FRAG; 3520 } 3521 } 3522 3523#else /* _KERNEL */ 3524 /* Every 100 calls, free a small amount */ 3525 if (spa_get_random(100) == 0) 3526 lowest = -1024; 3527#endif /* _KERNEL */ 3528 3529 last_free_memory = lowest; 3530 last_free_reason = r; 3531 DTRACE_PROBE2(arc__available_memory, int64_t, lowest, int, r); 3532 return (lowest); 3533} 3534 3535 3536/* 3537 * Determine if the system is under memory pressure and is asking 3538 * to reclaim memory. A return value of TRUE indicates that the system 3539 * is under memory pressure and that the arc should adjust accordingly. 3540 */ 3541static boolean_t 3542arc_reclaim_needed(void) 3543{ 3544 return (arc_available_memory() < 0); 3545} 3546 3547extern kmem_cache_t *zio_buf_cache[]; 3548extern kmem_cache_t *zio_data_buf_cache[]; 3549extern kmem_cache_t *range_seg_cache; 3550 3551static __noinline void 3552arc_kmem_reap_now(void) 3553{ 3554 size_t i; 3555 kmem_cache_t *prev_cache = NULL; 3556 kmem_cache_t *prev_data_cache = NULL; 3557 3558 DTRACE_PROBE(arc__kmem_reap_start); 3559#ifdef _KERNEL 3560 if (arc_meta_used >= arc_meta_limit) { 3561 /* 3562 * We are exceeding our meta-data cache limit. 3563 * Purge some DNLC entries to release holds on meta-data. 3564 */ 3565 dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent); 3566 } 3567#if defined(__i386) 3568 /* 3569 * Reclaim unused memory from all kmem caches. 3570 */ 3571 kmem_reap(); 3572#endif 3573#endif 3574 3575 for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) { 3576 if (zio_buf_cache[i] != prev_cache) { 3577 prev_cache = zio_buf_cache[i]; 3578 kmem_cache_reap_now(zio_buf_cache[i]); 3579 } 3580 if (zio_data_buf_cache[i] != prev_data_cache) { 3581 prev_data_cache = zio_data_buf_cache[i]; 3582 kmem_cache_reap_now(zio_data_buf_cache[i]); 3583 } 3584 } 3585 kmem_cache_reap_now(buf_cache); 3586 kmem_cache_reap_now(hdr_full_cache); 3587 kmem_cache_reap_now(hdr_l2only_cache); 3588 kmem_cache_reap_now(range_seg_cache); 3589 3590#ifdef illumos 3591 if (zio_arena != NULL) { 3592 /* 3593 * Ask the vmem arena to reclaim unused memory from its 3594 * quantum caches. 3595 */ 3596 vmem_qcache_reap(zio_arena); 3597 } 3598#endif 3599 DTRACE_PROBE(arc__kmem_reap_end); 3600} 3601 3602/* 3603 * Threads can block in arc_get_data_buf() waiting for this thread to evict 3604 * enough data and signal them to proceed. When this happens, the threads in 3605 * arc_get_data_buf() are sleeping while holding the hash lock for their 3606 * particular arc header. Thus, we must be careful to never sleep on a 3607 * hash lock in this thread. This is to prevent the following deadlock: 3608 * 3609 * - Thread A sleeps on CV in arc_get_data_buf() holding hash lock "L", 3610 * waiting for the reclaim thread to signal it. 3611 * 3612 * - arc_reclaim_thread() tries to acquire hash lock "L" using mutex_enter, 3613 * fails, and goes to sleep forever. 3614 * 3615 * This possible deadlock is avoided by always acquiring a hash lock 3616 * using mutex_tryenter() from arc_reclaim_thread(). 3617 */ 3618static void 3619arc_reclaim_thread(void *dummy __unused) 3620{ 3621 hrtime_t growtime = 0; 3622 callb_cpr_t cpr; 3623 3624 CALLB_CPR_INIT(&cpr, &arc_reclaim_lock, callb_generic_cpr, FTAG); 3625 3626 mutex_enter(&arc_reclaim_lock); 3627 while (!arc_reclaim_thread_exit) { 3628 int64_t free_memory = arc_available_memory(); 3629 uint64_t evicted = 0; 3630 3631 mutex_exit(&arc_reclaim_lock); 3632 3633 if (free_memory < 0) { 3634 3635 arc_no_grow = B_TRUE; 3636 arc_warm = B_TRUE; 3637 3638 /* 3639 * Wait at least zfs_grow_retry (default 60) seconds 3640 * before considering growing. 3641 */ 3642 growtime = gethrtime() + SEC2NSEC(arc_grow_retry); 3643 3644 arc_kmem_reap_now(); 3645 3646 /* 3647 * If we are still low on memory, shrink the ARC 3648 * so that we have arc_shrink_min free space. 3649 */ 3650 free_memory = arc_available_memory(); 3651 3652 int64_t to_free = 3653 (arc_c >> arc_shrink_shift) - free_memory; 3654 if (to_free > 0) { 3655#ifdef _KERNEL 3656 to_free = MAX(to_free, ptob(needfree)); 3657#endif 3658 arc_shrink(to_free); 3659 } 3660 } else if (free_memory < arc_c >> arc_no_grow_shift) { 3661 arc_no_grow = B_TRUE; 3662 } else if (gethrtime() >= growtime) { 3663 arc_no_grow = B_FALSE; 3664 } 3665 3666 evicted = arc_adjust(); 3667 3668 mutex_enter(&arc_reclaim_lock); 3669 3670 /* 3671 * If evicted is zero, we couldn't evict anything via 3672 * arc_adjust(). This could be due to hash lock 3673 * collisions, but more likely due to the majority of 3674 * arc buffers being unevictable. Therefore, even if 3675 * arc_size is above arc_c, another pass is unlikely to 3676 * be helpful and could potentially cause us to enter an 3677 * infinite loop. 3678 */ 3679 if (arc_size <= arc_c || evicted == 0) { 3680#ifdef _KERNEL 3681 needfree = 0; 3682#endif 3683 /* 3684 * We're either no longer overflowing, or we 3685 * can't evict anything more, so we should wake 3686 * up any threads before we go to sleep. 3687 */ 3688 cv_broadcast(&arc_reclaim_waiters_cv); 3689 3690 /* 3691 * Block until signaled, or after one second (we 3692 * might need to perform arc_kmem_reap_now() 3693 * even if we aren't being signalled) 3694 */ 3695 CALLB_CPR_SAFE_BEGIN(&cpr); 3696 (void) cv_timedwait_hires(&arc_reclaim_thread_cv, 3697 &arc_reclaim_lock, SEC2NSEC(1), MSEC2NSEC(1), 0); 3698 CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_lock); 3699 } 3700 } 3701 3702 arc_reclaim_thread_exit = FALSE; 3703 cv_broadcast(&arc_reclaim_thread_cv); 3704 CALLB_CPR_EXIT(&cpr); /* drops arc_reclaim_lock */ 3705 thread_exit(); 3706} 3707 3708static void 3709arc_user_evicts_thread(void *dummy __unused) 3710{ 3711 callb_cpr_t cpr; 3712 3713 CALLB_CPR_INIT(&cpr, &arc_user_evicts_lock, callb_generic_cpr, FTAG); 3714 3715 mutex_enter(&arc_user_evicts_lock); 3716 while (!arc_user_evicts_thread_exit) { 3717 mutex_exit(&arc_user_evicts_lock); 3718 3719 arc_do_user_evicts(); 3720 3721 /* 3722 * This is necessary in order for the mdb ::arc dcmd to 3723 * show up to date information. Since the ::arc command 3724 * does not call the kstat's update function, without 3725 * this call, the command may show stale stats for the 3726 * anon, mru, mru_ghost, mfu, and mfu_ghost lists. Even 3727 * with this change, the data might be up to 1 second 3728 * out of date; but that should suffice. The arc_state_t 3729 * structures can be queried directly if more accurate 3730 * information is needed. 3731 */ 3732 if (arc_ksp != NULL) 3733 arc_ksp->ks_update(arc_ksp, KSTAT_READ); 3734 3735 mutex_enter(&arc_user_evicts_lock); 3736 3737 /* 3738 * Block until signaled, or after one second (we need to 3739 * call the arc's kstat update function regularly). 3740 */ 3741 CALLB_CPR_SAFE_BEGIN(&cpr); 3742 (void) cv_timedwait(&arc_user_evicts_cv, 3743 &arc_user_evicts_lock, hz); 3744 CALLB_CPR_SAFE_END(&cpr, &arc_user_evicts_lock); 3745 } 3746 3747 arc_user_evicts_thread_exit = FALSE; 3748 cv_broadcast(&arc_user_evicts_cv); 3749 CALLB_CPR_EXIT(&cpr); /* drops arc_user_evicts_lock */ 3750 thread_exit(); 3751} 3752 3753/* 3754 * Adapt arc info given the number of bytes we are trying to add and 3755 * the state that we are comming from. This function is only called 3756 * when we are adding new content to the cache. 3757 */ 3758static void 3759arc_adapt(int bytes, arc_state_t *state) 3760{ 3761 int mult; 3762 uint64_t arc_p_min = (arc_c >> arc_p_min_shift); 3763 int64_t mrug_size = refcount_count(&arc_mru_ghost->arcs_size); 3764 int64_t mfug_size = refcount_count(&arc_mfu_ghost->arcs_size); 3765 3766 if (state == arc_l2c_only) 3767 return; 3768 3769 ASSERT(bytes > 0); 3770 /* 3771 * Adapt the target size of the MRU list: 3772 * - if we just hit in the MRU ghost list, then increase 3773 * the target size of the MRU list. 3774 * - if we just hit in the MFU ghost list, then increase 3775 * the target size of the MFU list by decreasing the 3776 * target size of the MRU list. 3777 */ 3778 if (state == arc_mru_ghost) { 3779 mult = (mrug_size >= mfug_size) ? 1 : (mfug_size / mrug_size); 3780 mult = MIN(mult, 10); /* avoid wild arc_p adjustment */ 3781 3782 arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult); 3783 } else if (state == arc_mfu_ghost) { 3784 uint64_t delta; 3785 3786 mult = (mfug_size >= mrug_size) ? 1 : (mrug_size / mfug_size); 3787 mult = MIN(mult, 10); 3788 3789 delta = MIN(bytes * mult, arc_p); 3790 arc_p = MAX(arc_p_min, arc_p - delta); 3791 } 3792 ASSERT((int64_t)arc_p >= 0); 3793 3794 if (arc_reclaim_needed()) { 3795 cv_signal(&arc_reclaim_thread_cv); 3796 return; 3797 } 3798 3799 if (arc_no_grow) 3800 return; 3801 3802 if (arc_c >= arc_c_max) 3803 return; 3804 3805 /* 3806 * If we're within (2 * maxblocksize) bytes of the target 3807 * cache size, increment the target cache size 3808 */ 3809 if (arc_size > arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) { 3810 DTRACE_PROBE1(arc__inc_adapt, int, bytes); 3811 atomic_add_64(&arc_c, (int64_t)bytes); 3812 if (arc_c > arc_c_max) 3813 arc_c = arc_c_max; 3814 else if (state == arc_anon) 3815 atomic_add_64(&arc_p, (int64_t)bytes); 3816 if (arc_p > arc_c) 3817 arc_p = arc_c; 3818 } 3819 ASSERT((int64_t)arc_p >= 0); 3820} 3821 3822/* 3823 * Check if arc_size has grown past our upper threshold, determined by 3824 * zfs_arc_overflow_shift. 3825 */ 3826static boolean_t 3827arc_is_overflowing(void) 3828{ 3829 /* Always allow at least one block of overflow */ 3830 uint64_t overflow = MAX(SPA_MAXBLOCKSIZE, 3831 arc_c >> zfs_arc_overflow_shift); 3832 3833 return (arc_size >= arc_c + overflow); 3834} 3835 3836/* 3837 * The buffer, supplied as the first argument, needs a data block. If we 3838 * are hitting the hard limit for the cache size, we must sleep, waiting 3839 * for the eviction thread to catch up. If we're past the target size 3840 * but below the hard limit, we'll only signal the reclaim thread and 3841 * continue on. 3842 */ 3843static void 3844arc_get_data_buf(arc_buf_t *buf) 3845{ 3846 arc_state_t *state = buf->b_hdr->b_l1hdr.b_state; 3847 uint64_t size = buf->b_hdr->b_size; 3848 arc_buf_contents_t type = arc_buf_type(buf->b_hdr); 3849 3850 arc_adapt(size, state); 3851 3852 /* 3853 * If arc_size is currently overflowing, and has grown past our 3854 * upper limit, we must be adding data faster than the evict 3855 * thread can evict. Thus, to ensure we don't compound the 3856 * problem by adding more data and forcing arc_size to grow even 3857 * further past it's target size, we halt and wait for the 3858 * eviction thread to catch up. 3859 * 3860 * It's also possible that the reclaim thread is unable to evict 3861 * enough buffers to get arc_size below the overflow limit (e.g. 3862 * due to buffers being un-evictable, or hash lock collisions). 3863 * In this case, we want to proceed regardless if we're 3864 * overflowing; thus we don't use a while loop here. 3865 */ 3866 if (arc_is_overflowing()) { 3867 mutex_enter(&arc_reclaim_lock); 3868 3869 /* 3870 * Now that we've acquired the lock, we may no longer be 3871 * over the overflow limit, lets check. 3872 * 3873 * We're ignoring the case of spurious wake ups. If that 3874 * were to happen, it'd let this thread consume an ARC 3875 * buffer before it should have (i.e. before we're under 3876 * the overflow limit and were signalled by the reclaim 3877 * thread). As long as that is a rare occurrence, it 3878 * shouldn't cause any harm. 3879 */ 3880 if (arc_is_overflowing()) { 3881 cv_signal(&arc_reclaim_thread_cv); 3882 cv_wait(&arc_reclaim_waiters_cv, &arc_reclaim_lock); 3883 } 3884 3885 mutex_exit(&arc_reclaim_lock); 3886 } 3887 3888 if (type == ARC_BUFC_METADATA) { 3889 buf->b_data = zio_buf_alloc(size); 3890 arc_space_consume(size, ARC_SPACE_META); 3891 } else { 3892 ASSERT(type == ARC_BUFC_DATA); 3893 buf->b_data = zio_data_buf_alloc(size); 3894 arc_space_consume(size, ARC_SPACE_DATA); 3895 } 3896 3897 /* 3898 * Update the state size. Note that ghost states have a 3899 * "ghost size" and so don't need to be updated. 3900 */ 3901 if (!GHOST_STATE(buf->b_hdr->b_l1hdr.b_state)) { 3902 arc_buf_hdr_t *hdr = buf->b_hdr; 3903 arc_state_t *state = hdr->b_l1hdr.b_state; 3904 3905 (void) refcount_add_many(&state->arcs_size, size, buf); 3906 3907 /* 3908 * If this is reached via arc_read, the link is 3909 * protected by the hash lock. If reached via 3910 * arc_buf_alloc, the header should not be accessed by 3911 * any other thread. And, if reached via arc_read_done, 3912 * the hash lock will protect it if it's found in the 3913 * hash table; otherwise no other thread should be 3914 * trying to [add|remove]_reference it. 3915 */ 3916 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) { 3917 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 3918 atomic_add_64(&hdr->b_l1hdr.b_state->arcs_lsize[type], 3919 size); 3920 } 3921 /* 3922 * If we are growing the cache, and we are adding anonymous 3923 * data, and we have outgrown arc_p, update arc_p 3924 */ 3925 if (arc_size < arc_c && hdr->b_l1hdr.b_state == arc_anon && 3926 (refcount_count(&arc_anon->arcs_size) + 3927 refcount_count(&arc_mru->arcs_size) > arc_p)) 3928 arc_p = MIN(arc_c, arc_p + size); 3929 } 3930 ARCSTAT_BUMP(arcstat_allocated); 3931} 3932 3933/* 3934 * This routine is called whenever a buffer is accessed. 3935 * NOTE: the hash lock is dropped in this function. 3936 */ 3937static void 3938arc_access(arc_buf_hdr_t *hdr, kmutex_t *hash_lock) 3939{ 3940 clock_t now; 3941 3942 ASSERT(MUTEX_HELD(hash_lock)); 3943 ASSERT(HDR_HAS_L1HDR(hdr)); 3944 3945 if (hdr->b_l1hdr.b_state == arc_anon) { 3946 /* 3947 * This buffer is not in the cache, and does not 3948 * appear in our "ghost" list. Add the new buffer 3949 * to the MRU state. 3950 */ 3951 3952 ASSERT0(hdr->b_l1hdr.b_arc_access); 3953 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 3954 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr); 3955 arc_change_state(arc_mru, hdr, hash_lock); 3956 3957 } else if (hdr->b_l1hdr.b_state == arc_mru) { 3958 now = ddi_get_lbolt(); 3959 3960 /* 3961 * If this buffer is here because of a prefetch, then either: 3962 * - clear the flag if this is a "referencing" read 3963 * (any subsequent access will bump this into the MFU state). 3964 * or 3965 * - move the buffer to the head of the list if this is 3966 * another prefetch (to make it less likely to be evicted). 3967 */ 3968 if (HDR_PREFETCH(hdr)) { 3969 if (refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) { 3970 /* link protected by hash lock */ 3971 ASSERT(multilist_link_active( 3972 &hdr->b_l1hdr.b_arc_node)); 3973 } else { 3974 hdr->b_flags &= ~ARC_FLAG_PREFETCH; 3975 ARCSTAT_BUMP(arcstat_mru_hits); 3976 } 3977 hdr->b_l1hdr.b_arc_access = now; 3978 return; 3979 } 3980 3981 /* 3982 * This buffer has been "accessed" only once so far, 3983 * but it is still in the cache. Move it to the MFU 3984 * state. 3985 */ 3986 if (now > hdr->b_l1hdr.b_arc_access + ARC_MINTIME) { 3987 /* 3988 * More than 125ms have passed since we 3989 * instantiated this buffer. Move it to the 3990 * most frequently used state. 3991 */ 3992 hdr->b_l1hdr.b_arc_access = now; 3993 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 3994 arc_change_state(arc_mfu, hdr, hash_lock); 3995 } 3996 ARCSTAT_BUMP(arcstat_mru_hits); 3997 } else if (hdr->b_l1hdr.b_state == arc_mru_ghost) { 3998 arc_state_t *new_state; 3999 /* 4000 * This buffer has been "accessed" recently, but 4001 * was evicted from the cache. Move it to the 4002 * MFU state. 4003 */ 4004 4005 if (HDR_PREFETCH(hdr)) { 4006 new_state = arc_mru; 4007 if (refcount_count(&hdr->b_l1hdr.b_refcnt) > 0) 4008 hdr->b_flags &= ~ARC_FLAG_PREFETCH; 4009 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr); 4010 } else { 4011 new_state = arc_mfu; 4012 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 4013 } 4014 4015 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 4016 arc_change_state(new_state, hdr, hash_lock); 4017 4018 ARCSTAT_BUMP(arcstat_mru_ghost_hits); 4019 } else if (hdr->b_l1hdr.b_state == arc_mfu) { 4020 /* 4021 * This buffer has been accessed more than once and is 4022 * still in the cache. Keep it in the MFU state. 4023 * 4024 * NOTE: an add_reference() that occurred when we did 4025 * the arc_read() will have kicked this off the list. 4026 * If it was a prefetch, we will explicitly move it to 4027 * the head of the list now. 4028 */ 4029 if ((HDR_PREFETCH(hdr)) != 0) { 4030 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 4031 /* link protected by hash_lock */ 4032 ASSERT(multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 4033 } 4034 ARCSTAT_BUMP(arcstat_mfu_hits); 4035 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 4036 } else if (hdr->b_l1hdr.b_state == arc_mfu_ghost) { 4037 arc_state_t *new_state = arc_mfu; 4038 /* 4039 * This buffer has been accessed more than once but has 4040 * been evicted from the cache. Move it back to the 4041 * MFU state. 4042 */ 4043 4044 if (HDR_PREFETCH(hdr)) { 4045 /* 4046 * This is a prefetch access... 4047 * move this block back to the MRU state. 4048 */ 4049 ASSERT0(refcount_count(&hdr->b_l1hdr.b_refcnt)); 4050 new_state = arc_mru; 4051 } 4052 4053 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 4054 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 4055 arc_change_state(new_state, hdr, hash_lock); 4056 4057 ARCSTAT_BUMP(arcstat_mfu_ghost_hits); 4058 } else if (hdr->b_l1hdr.b_state == arc_l2c_only) { 4059 /* 4060 * This buffer is on the 2nd Level ARC. 4061 */ 4062 4063 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 4064 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 4065 arc_change_state(arc_mfu, hdr, hash_lock); 4066 } else { 4067 ASSERT(!"invalid arc state"); 4068 } 4069} 4070 4071/* a generic arc_done_func_t which you can use */ 4072/* ARGSUSED */ 4073void 4074arc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg) 4075{ 4076 if (zio == NULL || zio->io_error == 0) 4077 bcopy(buf->b_data, arg, buf->b_hdr->b_size); 4078 VERIFY(arc_buf_remove_ref(buf, arg)); 4079} 4080 4081/* a generic arc_done_func_t */ 4082void 4083arc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg) 4084{ 4085 arc_buf_t **bufp = arg; 4086 if (zio && zio->io_error) { 4087 VERIFY(arc_buf_remove_ref(buf, arg)); 4088 *bufp = NULL; 4089 } else { 4090 *bufp = buf; 4091 ASSERT(buf->b_data); 4092 } 4093} 4094 4095static void 4096arc_read_done(zio_t *zio) 4097{ 4098 arc_buf_hdr_t *hdr; 4099 arc_buf_t *buf; 4100 arc_buf_t *abuf; /* buffer we're assigning to callback */ 4101 kmutex_t *hash_lock = NULL; 4102 arc_callback_t *callback_list, *acb; 4103 int freeable = FALSE; 4104 4105 buf = zio->io_private; 4106 hdr = buf->b_hdr; 4107 4108 /* 4109 * The hdr was inserted into hash-table and removed from lists 4110 * prior to starting I/O. We should find this header, since 4111 * it's in the hash table, and it should be legit since it's 4112 * not possible to evict it during the I/O. The only possible 4113 * reason for it not to be found is if we were freed during the 4114 * read. 4115 */ 4116 if (HDR_IN_HASH_TABLE(hdr)) { 4117 ASSERT3U(hdr->b_birth, ==, BP_PHYSICAL_BIRTH(zio->io_bp)); 4118 ASSERT3U(hdr->b_dva.dva_word[0], ==, 4119 BP_IDENTITY(zio->io_bp)->dva_word[0]); 4120 ASSERT3U(hdr->b_dva.dva_word[1], ==, 4121 BP_IDENTITY(zio->io_bp)->dva_word[1]); 4122 4123 arc_buf_hdr_t *found = buf_hash_find(hdr->b_spa, zio->io_bp, 4124 &hash_lock); 4125 4126 ASSERT((found == NULL && HDR_FREED_IN_READ(hdr) && 4127 hash_lock == NULL) || 4128 (found == hdr && 4129 DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) || 4130 (found == hdr && HDR_L2_READING(hdr))); 4131 } 4132 4133 hdr->b_flags &= ~ARC_FLAG_L2_EVICTED; 4134 if (l2arc_noprefetch && HDR_PREFETCH(hdr)) 4135 hdr->b_flags &= ~ARC_FLAG_L2CACHE; 4136 4137 /* byteswap if necessary */ 4138 callback_list = hdr->b_l1hdr.b_acb; 4139 ASSERT(callback_list != NULL); 4140 if (BP_SHOULD_BYTESWAP(zio->io_bp) && zio->io_error == 0) { 4141 dmu_object_byteswap_t bswap = 4142 DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp)); 4143 arc_byteswap_func_t *func = BP_GET_LEVEL(zio->io_bp) > 0 ? 4144 byteswap_uint64_array : 4145 dmu_ot_byteswap[bswap].ob_func; 4146 func(buf->b_data, hdr->b_size); 4147 } 4148 4149 arc_cksum_compute(buf, B_FALSE); 4150#ifdef illumos 4151 arc_buf_watch(buf); 4152#endif 4153 4154 if (hash_lock && zio->io_error == 0 && 4155 hdr->b_l1hdr.b_state == arc_anon) { 4156 /* 4157 * Only call arc_access on anonymous buffers. This is because 4158 * if we've issued an I/O for an evicted buffer, we've already 4159 * called arc_access (to prevent any simultaneous readers from 4160 * getting confused). 4161 */ 4162 arc_access(hdr, hash_lock); 4163 } 4164 4165 /* create copies of the data buffer for the callers */ 4166 abuf = buf; 4167 for (acb = callback_list; acb; acb = acb->acb_next) { 4168 if (acb->acb_done) { 4169 if (abuf == NULL) { 4170 ARCSTAT_BUMP(arcstat_duplicate_reads); 4171 abuf = arc_buf_clone(buf); 4172 } 4173 acb->acb_buf = abuf; 4174 abuf = NULL; 4175 } 4176 } 4177 hdr->b_l1hdr.b_acb = NULL; 4178 hdr->b_flags &= ~ARC_FLAG_IO_IN_PROGRESS; 4179 ASSERT(!HDR_BUF_AVAILABLE(hdr)); 4180 if (abuf == buf) { 4181 ASSERT(buf->b_efunc == NULL); 4182 ASSERT(hdr->b_l1hdr.b_datacnt == 1); 4183 hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE; 4184 } 4185 4186 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt) || 4187 callback_list != NULL); 4188 4189 if (zio->io_error != 0) { 4190 hdr->b_flags |= ARC_FLAG_IO_ERROR; 4191 if (hdr->b_l1hdr.b_state != arc_anon) 4192 arc_change_state(arc_anon, hdr, hash_lock); 4193 if (HDR_IN_HASH_TABLE(hdr)) 4194 buf_hash_remove(hdr); 4195 freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt); 4196 } 4197 4198 /* 4199 * Broadcast before we drop the hash_lock to avoid the possibility 4200 * that the hdr (and hence the cv) might be freed before we get to 4201 * the cv_broadcast(). 4202 */ 4203 cv_broadcast(&hdr->b_l1hdr.b_cv); 4204 4205 if (hash_lock != NULL) { 4206 mutex_exit(hash_lock); 4207 } else { 4208 /* 4209 * This block was freed while we waited for the read to 4210 * complete. It has been removed from the hash table and 4211 * moved to the anonymous state (so that it won't show up 4212 * in the cache). 4213 */ 4214 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon); 4215 freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt); 4216 } 4217 4218 /* execute each callback and free its structure */ 4219 while ((acb = callback_list) != NULL) { 4220 if (acb->acb_done) 4221 acb->acb_done(zio, acb->acb_buf, acb->acb_private); 4222 4223 if (acb->acb_zio_dummy != NULL) { 4224 acb->acb_zio_dummy->io_error = zio->io_error; 4225 zio_nowait(acb->acb_zio_dummy); 4226 } 4227 4228 callback_list = acb->acb_next; 4229 kmem_free(acb, sizeof (arc_callback_t)); 4230 } 4231 4232 if (freeable) 4233 arc_hdr_destroy(hdr); 4234} 4235 4236/* 4237 * "Read" the block at the specified DVA (in bp) via the 4238 * cache. If the block is found in the cache, invoke the provided 4239 * callback immediately and return. Note that the `zio' parameter 4240 * in the callback will be NULL in this case, since no IO was 4241 * required. If the block is not in the cache pass the read request 4242 * on to the spa with a substitute callback function, so that the 4243 * requested block will be added to the cache. 4244 * 4245 * If a read request arrives for a block that has a read in-progress, 4246 * either wait for the in-progress read to complete (and return the 4247 * results); or, if this is a read with a "done" func, add a record 4248 * to the read to invoke the "done" func when the read completes, 4249 * and return; or just return. 4250 * 4251 * arc_read_done() will invoke all the requested "done" functions 4252 * for readers of this block. 4253 */ 4254int 4255arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_done_func_t *done, 4256 void *private, zio_priority_t priority, int zio_flags, 4257 arc_flags_t *arc_flags, const zbookmark_phys_t *zb) 4258{ 4259 arc_buf_hdr_t *hdr = NULL; 4260 arc_buf_t *buf = NULL; 4261 kmutex_t *hash_lock = NULL; 4262 zio_t *rzio; 4263 uint64_t guid = spa_load_guid(spa); 4264 4265 ASSERT(!BP_IS_EMBEDDED(bp) || 4266 BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA); 4267 4268top: 4269 if (!BP_IS_EMBEDDED(bp)) { 4270 /* 4271 * Embedded BP's have no DVA and require no I/O to "read". 4272 * Create an anonymous arc buf to back it. 4273 */ 4274 hdr = buf_hash_find(guid, bp, &hash_lock); 4275 } 4276 4277 if (hdr != NULL && HDR_HAS_L1HDR(hdr) && hdr->b_l1hdr.b_datacnt > 0) { 4278 4279 *arc_flags |= ARC_FLAG_CACHED; 4280 4281 if (HDR_IO_IN_PROGRESS(hdr)) { 4282 4283 if ((hdr->b_flags & ARC_FLAG_PRIO_ASYNC_READ) && 4284 priority == ZIO_PRIORITY_SYNC_READ) { 4285 /* 4286 * This sync read must wait for an 4287 * in-progress async read (e.g. a predictive 4288 * prefetch). Async reads are queued 4289 * separately at the vdev_queue layer, so 4290 * this is a form of priority inversion. 4291 * Ideally, we would "inherit" the demand 4292 * i/o's priority by moving the i/o from 4293 * the async queue to the synchronous queue, 4294 * but there is currently no mechanism to do 4295 * so. Track this so that we can evaluate 4296 * the magnitude of this potential performance 4297 * problem. 4298 * 4299 * Note that if the prefetch i/o is already 4300 * active (has been issued to the device), 4301 * the prefetch improved performance, because 4302 * we issued it sooner than we would have 4303 * without the prefetch. 4304 */ 4305 DTRACE_PROBE1(arc__sync__wait__for__async, 4306 arc_buf_hdr_t *, hdr); 4307 ARCSTAT_BUMP(arcstat_sync_wait_for_async); 4308 } 4309 if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) { 4310 hdr->b_flags &= ~ARC_FLAG_PREDICTIVE_PREFETCH; 4311 } 4312 4313 if (*arc_flags & ARC_FLAG_WAIT) { 4314 cv_wait(&hdr->b_l1hdr.b_cv, hash_lock); 4315 mutex_exit(hash_lock); 4316 goto top; 4317 } 4318 ASSERT(*arc_flags & ARC_FLAG_NOWAIT); 4319 4320 if (done) { 4321 arc_callback_t *acb = NULL; 4322 4323 acb = kmem_zalloc(sizeof (arc_callback_t), 4324 KM_SLEEP); 4325 acb->acb_done = done; 4326 acb->acb_private = private; 4327 if (pio != NULL) 4328 acb->acb_zio_dummy = zio_null(pio, 4329 spa, NULL, NULL, NULL, zio_flags); 4330 4331 ASSERT(acb->acb_done != NULL); 4332 acb->acb_next = hdr->b_l1hdr.b_acb; 4333 hdr->b_l1hdr.b_acb = acb; 4334 add_reference(hdr, hash_lock, private); 4335 mutex_exit(hash_lock); 4336 return (0); 4337 } 4338 mutex_exit(hash_lock); 4339 return (0); 4340 } 4341 4342 ASSERT(hdr->b_l1hdr.b_state == arc_mru || 4343 hdr->b_l1hdr.b_state == arc_mfu); 4344 4345 if (done) { 4346 if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) { 4347 /* 4348 * This is a demand read which does not have to 4349 * wait for i/o because we did a predictive 4350 * prefetch i/o for it, which has completed. 4351 */ 4352 DTRACE_PROBE1( 4353 arc__demand__hit__predictive__prefetch, 4354 arc_buf_hdr_t *, hdr); 4355 ARCSTAT_BUMP( 4356 arcstat_demand_hit_predictive_prefetch); 4357 hdr->b_flags &= ~ARC_FLAG_PREDICTIVE_PREFETCH; 4358 } 4359 add_reference(hdr, hash_lock, private); 4360 /* 4361 * If this block is already in use, create a new 4362 * copy of the data so that we will be guaranteed 4363 * that arc_release() will always succeed. 4364 */ 4365 buf = hdr->b_l1hdr.b_buf; 4366 ASSERT(buf); 4367 ASSERT(buf->b_data); 4368 if (HDR_BUF_AVAILABLE(hdr)) { 4369 ASSERT(buf->b_efunc == NULL); 4370 hdr->b_flags &= ~ARC_FLAG_BUF_AVAILABLE; 4371 } else { 4372 buf = arc_buf_clone(buf); 4373 } 4374 4375 } else if (*arc_flags & ARC_FLAG_PREFETCH && 4376 refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) { 4377 hdr->b_flags |= ARC_FLAG_PREFETCH; 4378 } 4379 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr); 4380 arc_access(hdr, hash_lock); 4381 if (*arc_flags & ARC_FLAG_L2CACHE) 4382 hdr->b_flags |= ARC_FLAG_L2CACHE; 4383 if (*arc_flags & ARC_FLAG_L2COMPRESS) 4384 hdr->b_flags |= ARC_FLAG_L2COMPRESS; 4385 mutex_exit(hash_lock); 4386 ARCSTAT_BUMP(arcstat_hits); 4387 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr), 4388 demand, prefetch, !HDR_ISTYPE_METADATA(hdr), 4389 data, metadata, hits); 4390 4391 if (done) 4392 done(NULL, buf, private); 4393 } else { 4394 uint64_t size = BP_GET_LSIZE(bp); 4395 arc_callback_t *acb; 4396 vdev_t *vd = NULL; 4397 uint64_t addr = 0; 4398 boolean_t devw = B_FALSE; 4399 enum zio_compress b_compress = ZIO_COMPRESS_OFF; 4400 int32_t b_asize = 0; 4401 4402 if (hdr == NULL) { 4403 /* this block is not in the cache */ 4404 arc_buf_hdr_t *exists = NULL; 4405 arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp); 4406 buf = arc_buf_alloc(spa, size, private, type); 4407 hdr = buf->b_hdr; 4408 if (!BP_IS_EMBEDDED(bp)) { 4409 hdr->b_dva = *BP_IDENTITY(bp); 4410 hdr->b_birth = BP_PHYSICAL_BIRTH(bp); 4411 exists = buf_hash_insert(hdr, &hash_lock); 4412 } 4413 if (exists != NULL) { 4414 /* somebody beat us to the hash insert */ 4415 mutex_exit(hash_lock); 4416 buf_discard_identity(hdr); 4417 (void) arc_buf_remove_ref(buf, private); 4418 goto top; /* restart the IO request */ 4419 } 4420 4421 /* 4422 * If there is a callback, we pass our reference to 4423 * it; otherwise we remove our reference. 4424 */ 4425 if (done == NULL) { 4426 (void) remove_reference(hdr, hash_lock, 4427 private); 4428 } 4429 if (*arc_flags & ARC_FLAG_PREFETCH) 4430 hdr->b_flags |= ARC_FLAG_PREFETCH; 4431 if (*arc_flags & ARC_FLAG_L2CACHE) 4432 hdr->b_flags |= ARC_FLAG_L2CACHE; 4433 if (*arc_flags & ARC_FLAG_L2COMPRESS) 4434 hdr->b_flags |= ARC_FLAG_L2COMPRESS; 4435 if (BP_GET_LEVEL(bp) > 0) 4436 hdr->b_flags |= ARC_FLAG_INDIRECT; 4437 } else { 4438 /* 4439 * This block is in the ghost cache. If it was L2-only 4440 * (and thus didn't have an L1 hdr), we realloc the 4441 * header to add an L1 hdr. 4442 */ 4443 if (!HDR_HAS_L1HDR(hdr)) { 4444 hdr = arc_hdr_realloc(hdr, hdr_l2only_cache, 4445 hdr_full_cache); 4446 } 4447 4448 ASSERT(GHOST_STATE(hdr->b_l1hdr.b_state)); 4449 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 4450 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 4451 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 4452 4453 /* 4454 * If there is a callback, we pass a reference to it. 4455 */ 4456 if (done != NULL) 4457 add_reference(hdr, hash_lock, private); 4458 if (*arc_flags & ARC_FLAG_PREFETCH) 4459 hdr->b_flags |= ARC_FLAG_PREFETCH; 4460 if (*arc_flags & ARC_FLAG_L2CACHE) 4461 hdr->b_flags |= ARC_FLAG_L2CACHE; 4462 if (*arc_flags & ARC_FLAG_L2COMPRESS) 4463 hdr->b_flags |= ARC_FLAG_L2COMPRESS; 4464 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 4465 buf->b_hdr = hdr; 4466 buf->b_data = NULL; 4467 buf->b_efunc = NULL; 4468 buf->b_private = NULL; 4469 buf->b_next = NULL; 4470 hdr->b_l1hdr.b_buf = buf; 4471 ASSERT0(hdr->b_l1hdr.b_datacnt); 4472 hdr->b_l1hdr.b_datacnt = 1; 4473 arc_get_data_buf(buf); 4474 arc_access(hdr, hash_lock); 4475 } 4476 4477 if (*arc_flags & ARC_FLAG_PREDICTIVE_PREFETCH) 4478 hdr->b_flags |= ARC_FLAG_PREDICTIVE_PREFETCH; 4479 ASSERT(!GHOST_STATE(hdr->b_l1hdr.b_state)); 4480 4481 acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP); 4482 acb->acb_done = done; 4483 acb->acb_private = private; 4484 4485 ASSERT(hdr->b_l1hdr.b_acb == NULL); 4486 hdr->b_l1hdr.b_acb = acb; 4487 hdr->b_flags |= ARC_FLAG_IO_IN_PROGRESS; 4488 4489 if (HDR_HAS_L2HDR(hdr) && 4490 (vd = hdr->b_l2hdr.b_dev->l2ad_vdev) != NULL) { 4491 devw = hdr->b_l2hdr.b_dev->l2ad_writing; 4492 addr = hdr->b_l2hdr.b_daddr; 4493 b_compress = hdr->b_l2hdr.b_compress; 4494 b_asize = hdr->b_l2hdr.b_asize; 4495 /* 4496 * Lock out device removal. 4497 */ 4498 if (vdev_is_dead(vd) || 4499 !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER)) 4500 vd = NULL; 4501 } 4502 4503 if (hash_lock != NULL) 4504 mutex_exit(hash_lock); 4505 4506 /* 4507 * At this point, we have a level 1 cache miss. Try again in 4508 * L2ARC if possible. 4509 */ 4510 ASSERT3U(hdr->b_size, ==, size); 4511 DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp, 4512 uint64_t, size, zbookmark_phys_t *, zb); 4513 ARCSTAT_BUMP(arcstat_misses); 4514 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr), 4515 demand, prefetch, !HDR_ISTYPE_METADATA(hdr), 4516 data, metadata, misses); 4517#ifdef _KERNEL 4518 curthread->td_ru.ru_inblock++; 4519#endif 4520 4521 if (priority == ZIO_PRIORITY_ASYNC_READ) 4522 hdr->b_flags |= ARC_FLAG_PRIO_ASYNC_READ; 4523 else 4524 hdr->b_flags &= ~ARC_FLAG_PRIO_ASYNC_READ; 4525 4526 if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) { 4527 /* 4528 * Read from the L2ARC if the following are true: 4529 * 1. The L2ARC vdev was previously cached. 4530 * 2. This buffer still has L2ARC metadata. 4531 * 3. This buffer isn't currently writing to the L2ARC. 4532 * 4. The L2ARC entry wasn't evicted, which may 4533 * also have invalidated the vdev. 4534 * 5. This isn't prefetch and l2arc_noprefetch is set. 4535 */ 4536 if (HDR_HAS_L2HDR(hdr) && 4537 !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) && 4538 !(l2arc_noprefetch && HDR_PREFETCH(hdr))) { 4539 l2arc_read_callback_t *cb; 4540 void* b_data; 4541 4542 DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr); 4543 ARCSTAT_BUMP(arcstat_l2_hits); 4544 4545 cb = kmem_zalloc(sizeof (l2arc_read_callback_t), 4546 KM_SLEEP); 4547 cb->l2rcb_buf = buf; 4548 cb->l2rcb_spa = spa; 4549 cb->l2rcb_bp = *bp; 4550 cb->l2rcb_zb = *zb; 4551 cb->l2rcb_flags = zio_flags; 4552 cb->l2rcb_compress = b_compress; 4553 if (b_asize > hdr->b_size) { 4554 ASSERT3U(b_compress, ==, 4555 ZIO_COMPRESS_OFF); 4556 b_data = zio_data_buf_alloc(b_asize); 4557 cb->l2rcb_data = b_data; 4558 } else { 4559 b_data = buf->b_data; 4560 } 4561 4562 ASSERT(addr >= VDEV_LABEL_START_SIZE && 4563 addr + size < vd->vdev_psize - 4564 VDEV_LABEL_END_SIZE); 4565 4566 /* 4567 * l2arc read. The SCL_L2ARC lock will be 4568 * released by l2arc_read_done(). 4569 * Issue a null zio if the underlying buffer 4570 * was squashed to zero size by compression. 4571 */ 4572 if (b_compress == ZIO_COMPRESS_EMPTY) { 4573 ASSERT3U(b_asize, ==, 0); 4574 rzio = zio_null(pio, spa, vd, 4575 l2arc_read_done, cb, 4576 zio_flags | ZIO_FLAG_DONT_CACHE | 4577 ZIO_FLAG_CANFAIL | 4578 ZIO_FLAG_DONT_PROPAGATE | 4579 ZIO_FLAG_DONT_RETRY); 4580 } else { 4581 rzio = zio_read_phys(pio, vd, addr, 4582 b_asize, b_data, 4583 ZIO_CHECKSUM_OFF, 4584 l2arc_read_done, cb, priority, 4585 zio_flags | ZIO_FLAG_DONT_CACHE | 4586 ZIO_FLAG_CANFAIL | 4587 ZIO_FLAG_DONT_PROPAGATE | 4588 ZIO_FLAG_DONT_RETRY, B_FALSE); 4589 } 4590 DTRACE_PROBE2(l2arc__read, vdev_t *, vd, 4591 zio_t *, rzio); 4592 ARCSTAT_INCR(arcstat_l2_read_bytes, b_asize); 4593 4594 if (*arc_flags & ARC_FLAG_NOWAIT) { 4595 zio_nowait(rzio); 4596 return (0); 4597 } 4598 4599 ASSERT(*arc_flags & ARC_FLAG_WAIT); 4600 if (zio_wait(rzio) == 0) 4601 return (0); 4602 4603 /* l2arc read error; goto zio_read() */ 4604 } else { 4605 DTRACE_PROBE1(l2arc__miss, 4606 arc_buf_hdr_t *, hdr); 4607 ARCSTAT_BUMP(arcstat_l2_misses); 4608 if (HDR_L2_WRITING(hdr)) 4609 ARCSTAT_BUMP(arcstat_l2_rw_clash); 4610 spa_config_exit(spa, SCL_L2ARC, vd); 4611 } 4612 } else { 4613 if (vd != NULL) 4614 spa_config_exit(spa, SCL_L2ARC, vd); 4615 if (l2arc_ndev != 0) { 4616 DTRACE_PROBE1(l2arc__miss, 4617 arc_buf_hdr_t *, hdr); 4618 ARCSTAT_BUMP(arcstat_l2_misses); 4619 } 4620 } 4621 4622 rzio = zio_read(pio, spa, bp, buf->b_data, size, 4623 arc_read_done, buf, priority, zio_flags, zb); 4624 4625 if (*arc_flags & ARC_FLAG_WAIT) 4626 return (zio_wait(rzio)); 4627 4628 ASSERT(*arc_flags & ARC_FLAG_NOWAIT); 4629 zio_nowait(rzio); 4630 } 4631 return (0); 4632} 4633 4634void 4635arc_set_callback(arc_buf_t *buf, arc_evict_func_t *func, void *private) 4636{ 4637 ASSERT(buf->b_hdr != NULL); 4638 ASSERT(buf->b_hdr->b_l1hdr.b_state != arc_anon); 4639 ASSERT(!refcount_is_zero(&buf->b_hdr->b_l1hdr.b_refcnt) || 4640 func == NULL); 4641 ASSERT(buf->b_efunc == NULL); 4642 ASSERT(!HDR_BUF_AVAILABLE(buf->b_hdr)); 4643 4644 buf->b_efunc = func; 4645 buf->b_private = private; 4646} 4647 4648/* 4649 * Notify the arc that a block was freed, and thus will never be used again. 4650 */ 4651void 4652arc_freed(spa_t *spa, const blkptr_t *bp) 4653{ 4654 arc_buf_hdr_t *hdr; 4655 kmutex_t *hash_lock; 4656 uint64_t guid = spa_load_guid(spa); 4657 4658 ASSERT(!BP_IS_EMBEDDED(bp)); 4659 4660 hdr = buf_hash_find(guid, bp, &hash_lock); 4661 if (hdr == NULL) 4662 return; 4663 if (HDR_BUF_AVAILABLE(hdr)) { 4664 arc_buf_t *buf = hdr->b_l1hdr.b_buf; 4665 add_reference(hdr, hash_lock, FTAG); 4666 hdr->b_flags &= ~ARC_FLAG_BUF_AVAILABLE; 4667 mutex_exit(hash_lock); 4668 4669 arc_release(buf, FTAG); 4670 (void) arc_buf_remove_ref(buf, FTAG); 4671 } else { 4672 mutex_exit(hash_lock); 4673 } 4674 4675} 4676 4677/* 4678 * Clear the user eviction callback set by arc_set_callback(), first calling 4679 * it if it exists. Because the presence of a callback keeps an arc_buf cached 4680 * clearing the callback may result in the arc_buf being destroyed. However, 4681 * it will not result in the *last* arc_buf being destroyed, hence the data 4682 * will remain cached in the ARC. We make a copy of the arc buffer here so 4683 * that we can process the callback without holding any locks. 4684 * 4685 * It's possible that the callback is already in the process of being cleared 4686 * by another thread. In this case we can not clear the callback. 4687 * 4688 * Returns B_TRUE if the callback was successfully called and cleared. 4689 */ 4690boolean_t 4691arc_clear_callback(arc_buf_t *buf) 4692{ 4693 arc_buf_hdr_t *hdr; 4694 kmutex_t *hash_lock; 4695 arc_evict_func_t *efunc = buf->b_efunc; 4696 void *private = buf->b_private; 4697 4698 mutex_enter(&buf->b_evict_lock); 4699 hdr = buf->b_hdr; 4700 if (hdr == NULL) { 4701 /* 4702 * We are in arc_do_user_evicts(). 4703 */ 4704 ASSERT(buf->b_data == NULL); 4705 mutex_exit(&buf->b_evict_lock); 4706 return (B_FALSE); 4707 } else if (buf->b_data == NULL) { 4708 /* 4709 * We are on the eviction list; process this buffer now 4710 * but let arc_do_user_evicts() do the reaping. 4711 */ 4712 buf->b_efunc = NULL; 4713 mutex_exit(&buf->b_evict_lock); 4714 VERIFY0(efunc(private)); 4715 return (B_TRUE); 4716 } 4717 hash_lock = HDR_LOCK(hdr); 4718 mutex_enter(hash_lock); 4719 hdr = buf->b_hdr; 4720 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 4721 4722 ASSERT3U(refcount_count(&hdr->b_l1hdr.b_refcnt), <, 4723 hdr->b_l1hdr.b_datacnt); 4724 ASSERT(hdr->b_l1hdr.b_state == arc_mru || 4725 hdr->b_l1hdr.b_state == arc_mfu); 4726 4727 buf->b_efunc = NULL; 4728 buf->b_private = NULL; 4729 4730 if (hdr->b_l1hdr.b_datacnt > 1) { 4731 mutex_exit(&buf->b_evict_lock); 4732 arc_buf_destroy(buf, TRUE); 4733 } else { 4734 ASSERT(buf == hdr->b_l1hdr.b_buf); 4735 hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE; 4736 mutex_exit(&buf->b_evict_lock); 4737 } 4738 4739 mutex_exit(hash_lock); 4740 VERIFY0(efunc(private)); 4741 return (B_TRUE); 4742} 4743 4744/* 4745 * Release this buffer from the cache, making it an anonymous buffer. This 4746 * must be done after a read and prior to modifying the buffer contents. 4747 * If the buffer has more than one reference, we must make 4748 * a new hdr for the buffer. 4749 */ 4750void 4751arc_release(arc_buf_t *buf, void *tag) 4752{ 4753 arc_buf_hdr_t *hdr = buf->b_hdr; 4754 4755 /* 4756 * It would be nice to assert that if it's DMU metadata (level > 4757 * 0 || it's the dnode file), then it must be syncing context. 4758 * But we don't know that information at this level. 4759 */ 4760 4761 mutex_enter(&buf->b_evict_lock); 4762 4763 ASSERT(HDR_HAS_L1HDR(hdr)); 4764 4765 /* 4766 * We don't grab the hash lock prior to this check, because if 4767 * the buffer's header is in the arc_anon state, it won't be 4768 * linked into the hash table. 4769 */ 4770 if (hdr->b_l1hdr.b_state == arc_anon) { 4771 mutex_exit(&buf->b_evict_lock); 4772 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 4773 ASSERT(!HDR_IN_HASH_TABLE(hdr)); 4774 ASSERT(!HDR_HAS_L2HDR(hdr)); 4775 ASSERT(BUF_EMPTY(hdr)); 4776 ASSERT3U(hdr->b_l1hdr.b_datacnt, ==, 1); 4777 ASSERT3S(refcount_count(&hdr->b_l1hdr.b_refcnt), ==, 1); 4778 ASSERT(!list_link_active(&hdr->b_l1hdr.b_arc_node)); 4779 4780 ASSERT3P(buf->b_efunc, ==, NULL); 4781 ASSERT3P(buf->b_private, ==, NULL); 4782 4783 hdr->b_l1hdr.b_arc_access = 0; 4784 arc_buf_thaw(buf); 4785 4786 return; 4787 } 4788 4789 kmutex_t *hash_lock = HDR_LOCK(hdr); 4790 mutex_enter(hash_lock); 4791 4792 /* 4793 * This assignment is only valid as long as the hash_lock is 4794 * held, we must be careful not to reference state or the 4795 * b_state field after dropping the lock. 4796 */ 4797 arc_state_t *state = hdr->b_l1hdr.b_state; 4798 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 4799 ASSERT3P(state, !=, arc_anon); 4800 4801 /* this buffer is not on any list */ 4802 ASSERT(refcount_count(&hdr->b_l1hdr.b_refcnt) > 0); 4803 4804 if (HDR_HAS_L2HDR(hdr)) { 4805 mutex_enter(&hdr->b_l2hdr.b_dev->l2ad_mtx); 4806 4807 /* 4808 * We have to recheck this conditional again now that 4809 * we're holding the l2ad_mtx to prevent a race with 4810 * another thread which might be concurrently calling 4811 * l2arc_evict(). In that case, l2arc_evict() might have 4812 * destroyed the header's L2 portion as we were waiting 4813 * to acquire the l2ad_mtx. 4814 */ 4815 if (HDR_HAS_L2HDR(hdr)) { 4816 l2arc_trim(hdr); 4817 arc_hdr_l2hdr_destroy(hdr); 4818 } 4819 4820 mutex_exit(&hdr->b_l2hdr.b_dev->l2ad_mtx); 4821 } 4822 4823 /* 4824 * Do we have more than one buf? 4825 */ 4826 if (hdr->b_l1hdr.b_datacnt > 1) { 4827 arc_buf_hdr_t *nhdr; 4828 arc_buf_t **bufp; 4829 uint64_t blksz = hdr->b_size; 4830 uint64_t spa = hdr->b_spa; 4831 arc_buf_contents_t type = arc_buf_type(hdr); 4832 uint32_t flags = hdr->b_flags; 4833 4834 ASSERT(hdr->b_l1hdr.b_buf != buf || buf->b_next != NULL); 4835 /* 4836 * Pull the data off of this hdr and attach it to 4837 * a new anonymous hdr. 4838 */ 4839 (void) remove_reference(hdr, hash_lock, tag); 4840 bufp = &hdr->b_l1hdr.b_buf; 4841 while (*bufp != buf) 4842 bufp = &(*bufp)->b_next; 4843 *bufp = buf->b_next; 4844 buf->b_next = NULL; 4845 4846 ASSERT3P(state, !=, arc_l2c_only); 4847 4848 (void) refcount_remove_many( 4849 &state->arcs_size, hdr->b_size, buf); 4850 4851 if (refcount_is_zero(&hdr->b_l1hdr.b_refcnt)) { 4852 ASSERT3P(state, !=, arc_l2c_only); 4853 uint64_t *size = &state->arcs_lsize[type]; 4854 ASSERT3U(*size, >=, hdr->b_size); 4855 atomic_add_64(size, -hdr->b_size); 4856 } 4857 4858 /* 4859 * We're releasing a duplicate user data buffer, update 4860 * our statistics accordingly. 4861 */ 4862 if (HDR_ISTYPE_DATA(hdr)) { 4863 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers); 4864 ARCSTAT_INCR(arcstat_duplicate_buffers_size, 4865 -hdr->b_size); 4866 } 4867 hdr->b_l1hdr.b_datacnt -= 1; 4868 arc_cksum_verify(buf); 4869#ifdef illumos 4870 arc_buf_unwatch(buf); 4871#endif 4872 4873 mutex_exit(hash_lock); 4874 4875 nhdr = kmem_cache_alloc(hdr_full_cache, KM_PUSHPAGE); 4876 nhdr->b_size = blksz; 4877 nhdr->b_spa = spa; 4878 4879 nhdr->b_flags = flags & ARC_FLAG_L2_WRITING; 4880 nhdr->b_flags |= arc_bufc_to_flags(type); 4881 nhdr->b_flags |= ARC_FLAG_HAS_L1HDR; 4882 4883 nhdr->b_l1hdr.b_buf = buf; 4884 nhdr->b_l1hdr.b_datacnt = 1; 4885 nhdr->b_l1hdr.b_state = arc_anon; 4886 nhdr->b_l1hdr.b_arc_access = 0; 4887 nhdr->b_l1hdr.b_tmp_cdata = NULL; 4888 nhdr->b_freeze_cksum = NULL; 4889 4890 (void) refcount_add(&nhdr->b_l1hdr.b_refcnt, tag); 4891 buf->b_hdr = nhdr; 4892 mutex_exit(&buf->b_evict_lock); 4893 (void) refcount_add_many(&arc_anon->arcs_size, blksz, buf); 4894 } else { 4895 mutex_exit(&buf->b_evict_lock); 4896 ASSERT(refcount_count(&hdr->b_l1hdr.b_refcnt) == 1); 4897 /* protected by hash lock, or hdr is on arc_anon */ 4898 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 4899 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 4900 arc_change_state(arc_anon, hdr, hash_lock); 4901 hdr->b_l1hdr.b_arc_access = 0; 4902 mutex_exit(hash_lock); 4903 4904 buf_discard_identity(hdr); 4905 arc_buf_thaw(buf); 4906 } 4907 buf->b_efunc = NULL; 4908 buf->b_private = NULL; 4909} 4910 4911int 4912arc_released(arc_buf_t *buf) 4913{ 4914 int released; 4915 4916 mutex_enter(&buf->b_evict_lock); 4917 released = (buf->b_data != NULL && 4918 buf->b_hdr->b_l1hdr.b_state == arc_anon); 4919 mutex_exit(&buf->b_evict_lock); 4920 return (released); 4921} 4922 4923#ifdef ZFS_DEBUG 4924int 4925arc_referenced(arc_buf_t *buf) 4926{ 4927 int referenced; 4928 4929 mutex_enter(&buf->b_evict_lock); 4930 referenced = (refcount_count(&buf->b_hdr->b_l1hdr.b_refcnt)); 4931 mutex_exit(&buf->b_evict_lock); 4932 return (referenced); 4933} 4934#endif 4935 4936static void 4937arc_write_ready(zio_t *zio) 4938{ 4939 arc_write_callback_t *callback = zio->io_private; 4940 arc_buf_t *buf = callback->awcb_buf; 4941 arc_buf_hdr_t *hdr = buf->b_hdr; 4942 4943 ASSERT(HDR_HAS_L1HDR(hdr)); 4944 ASSERT(!refcount_is_zero(&buf->b_hdr->b_l1hdr.b_refcnt)); 4945 ASSERT(hdr->b_l1hdr.b_datacnt > 0); 4946 callback->awcb_ready(zio, buf, callback->awcb_private); 4947 4948 /* 4949 * If the IO is already in progress, then this is a re-write 4950 * attempt, so we need to thaw and re-compute the cksum. 4951 * It is the responsibility of the callback to handle the 4952 * accounting for any re-write attempt. 4953 */ 4954 if (HDR_IO_IN_PROGRESS(hdr)) { 4955 mutex_enter(&hdr->b_l1hdr.b_freeze_lock); 4956 if (hdr->b_freeze_cksum != NULL) { 4957 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 4958 hdr->b_freeze_cksum = NULL; 4959 } 4960 mutex_exit(&hdr->b_l1hdr.b_freeze_lock); 4961 } 4962 arc_cksum_compute(buf, B_FALSE); 4963 hdr->b_flags |= ARC_FLAG_IO_IN_PROGRESS; 4964} 4965 4966/* 4967 * The SPA calls this callback for each physical write that happens on behalf 4968 * of a logical write. See the comment in dbuf_write_physdone() for details. 4969 */ 4970static void 4971arc_write_physdone(zio_t *zio) 4972{ 4973 arc_write_callback_t *cb = zio->io_private; 4974 if (cb->awcb_physdone != NULL) 4975 cb->awcb_physdone(zio, cb->awcb_buf, cb->awcb_private); 4976} 4977 4978static void 4979arc_write_done(zio_t *zio) 4980{ 4981 arc_write_callback_t *callback = zio->io_private; 4982 arc_buf_t *buf = callback->awcb_buf; 4983 arc_buf_hdr_t *hdr = buf->b_hdr; 4984 4985 ASSERT(hdr->b_l1hdr.b_acb == NULL); 4986 4987 if (zio->io_error == 0) { 4988 if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) { 4989 buf_discard_identity(hdr); 4990 } else { 4991 hdr->b_dva = *BP_IDENTITY(zio->io_bp); 4992 hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp); 4993 } 4994 } else { 4995 ASSERT(BUF_EMPTY(hdr)); 4996 } 4997 4998 /* 4999 * If the block to be written was all-zero or compressed enough to be 5000 * embedded in the BP, no write was performed so there will be no 5001 * dva/birth/checksum. The buffer must therefore remain anonymous 5002 * (and uncached). 5003 */ 5004 if (!BUF_EMPTY(hdr)) { 5005 arc_buf_hdr_t *exists; 5006 kmutex_t *hash_lock; 5007 5008 ASSERT(zio->io_error == 0); 5009 5010 arc_cksum_verify(buf); 5011 5012 exists = buf_hash_insert(hdr, &hash_lock); 5013 if (exists != NULL) { 5014 /* 5015 * This can only happen if we overwrite for 5016 * sync-to-convergence, because we remove 5017 * buffers from the hash table when we arc_free(). 5018 */ 5019 if (zio->io_flags & ZIO_FLAG_IO_REWRITE) { 5020 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) 5021 panic("bad overwrite, hdr=%p exists=%p", 5022 (void *)hdr, (void *)exists); 5023 ASSERT(refcount_is_zero( 5024 &exists->b_l1hdr.b_refcnt)); 5025 arc_change_state(arc_anon, exists, hash_lock); 5026 mutex_exit(hash_lock); 5027 arc_hdr_destroy(exists); 5028 exists = buf_hash_insert(hdr, &hash_lock); 5029 ASSERT3P(exists, ==, NULL); 5030 } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) { 5031 /* nopwrite */ 5032 ASSERT(zio->io_prop.zp_nopwrite); 5033 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) 5034 panic("bad nopwrite, hdr=%p exists=%p", 5035 (void *)hdr, (void *)exists); 5036 } else { 5037 /* Dedup */ 5038 ASSERT(hdr->b_l1hdr.b_datacnt == 1); 5039 ASSERT(hdr->b_l1hdr.b_state == arc_anon); 5040 ASSERT(BP_GET_DEDUP(zio->io_bp)); 5041 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0); 5042 } 5043 } 5044 hdr->b_flags &= ~ARC_FLAG_IO_IN_PROGRESS; 5045 /* if it's not anon, we are doing a scrub */ 5046 if (exists == NULL && hdr->b_l1hdr.b_state == arc_anon) 5047 arc_access(hdr, hash_lock); 5048 mutex_exit(hash_lock); 5049 } else { 5050 hdr->b_flags &= ~ARC_FLAG_IO_IN_PROGRESS; 5051 } 5052 5053 ASSERT(!refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 5054 callback->awcb_done(zio, buf, callback->awcb_private); 5055 5056 kmem_free(callback, sizeof (arc_write_callback_t)); 5057} 5058 5059zio_t * 5060arc_write(zio_t *pio, spa_t *spa, uint64_t txg, 5061 blkptr_t *bp, arc_buf_t *buf, boolean_t l2arc, boolean_t l2arc_compress, 5062 const zio_prop_t *zp, arc_done_func_t *ready, arc_done_func_t *physdone, 5063 arc_done_func_t *done, void *private, zio_priority_t priority, 5064 int zio_flags, const zbookmark_phys_t *zb) 5065{ 5066 arc_buf_hdr_t *hdr = buf->b_hdr; 5067 arc_write_callback_t *callback; 5068 zio_t *zio; 5069 5070 ASSERT(ready != NULL); 5071 ASSERT(done != NULL); 5072 ASSERT(!HDR_IO_ERROR(hdr)); 5073 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 5074 ASSERT(hdr->b_l1hdr.b_acb == NULL); 5075 ASSERT(hdr->b_l1hdr.b_datacnt > 0); 5076 if (l2arc) 5077 hdr->b_flags |= ARC_FLAG_L2CACHE; 5078 if (l2arc_compress) 5079 hdr->b_flags |= ARC_FLAG_L2COMPRESS; 5080 callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP); 5081 callback->awcb_ready = ready; 5082 callback->awcb_physdone = physdone; 5083 callback->awcb_done = done; 5084 callback->awcb_private = private; 5085 callback->awcb_buf = buf; 5086 5087 zio = zio_write(pio, spa, txg, bp, buf->b_data, hdr->b_size, zp, 5088 arc_write_ready, arc_write_physdone, arc_write_done, callback, 5089 priority, zio_flags, zb); 5090 5091 return (zio); 5092} 5093 5094static int 5095arc_memory_throttle(uint64_t reserve, uint64_t txg) 5096{ 5097#ifdef _KERNEL 5098 uint64_t available_memory = ptob(freemem); 5099 static uint64_t page_load = 0; 5100 static uint64_t last_txg = 0; 5101 5102#if defined(__i386) || !defined(UMA_MD_SMALL_ALLOC) 5103 available_memory = 5104 MIN(available_memory, ptob(vmem_size(heap_arena, VMEM_FREE))); 5105#endif 5106 5107 if (freemem > (uint64_t)physmem * arc_lotsfree_percent / 100) 5108 return (0); 5109 5110 if (txg > last_txg) { 5111 last_txg = txg; 5112 page_load = 0; 5113 } 5114 /* 5115 * If we are in pageout, we know that memory is already tight, 5116 * the arc is already going to be evicting, so we just want to 5117 * continue to let page writes occur as quickly as possible. 5118 */ 5119 if (curproc == pageproc) { 5120 if (page_load > MAX(ptob(minfree), available_memory) / 4) 5121 return (SET_ERROR(ERESTART)); 5122 /* Note: reserve is inflated, so we deflate */ 5123 page_load += reserve / 8; 5124 return (0); 5125 } else if (page_load > 0 && arc_reclaim_needed()) { 5126 /* memory is low, delay before restarting */ 5127 ARCSTAT_INCR(arcstat_memory_throttle_count, 1); 5128 return (SET_ERROR(EAGAIN)); 5129 } 5130 page_load = 0; 5131#endif 5132 return (0); 5133} 5134 5135void 5136arc_tempreserve_clear(uint64_t reserve) 5137{ 5138 atomic_add_64(&arc_tempreserve, -reserve); 5139 ASSERT((int64_t)arc_tempreserve >= 0); 5140} 5141 5142int 5143arc_tempreserve_space(uint64_t reserve, uint64_t txg) 5144{ 5145 int error; 5146 uint64_t anon_size; 5147 5148 if (reserve > arc_c/4 && !arc_no_grow) { 5149 arc_c = MIN(arc_c_max, reserve * 4); 5150 DTRACE_PROBE1(arc__set_reserve, uint64_t, arc_c); 5151 } 5152 if (reserve > arc_c) 5153 return (SET_ERROR(ENOMEM)); 5154 5155 /* 5156 * Don't count loaned bufs as in flight dirty data to prevent long 5157 * network delays from blocking transactions that are ready to be 5158 * assigned to a txg. 5159 */ 5160 anon_size = MAX((int64_t)(refcount_count(&arc_anon->arcs_size) - 5161 arc_loaned_bytes), 0); 5162 5163 /* 5164 * Writes will, almost always, require additional memory allocations 5165 * in order to compress/encrypt/etc the data. We therefore need to 5166 * make sure that there is sufficient available memory for this. 5167 */ 5168 error = arc_memory_throttle(reserve, txg); 5169 if (error != 0) 5170 return (error); 5171 5172 /* 5173 * Throttle writes when the amount of dirty data in the cache 5174 * gets too large. We try to keep the cache less than half full 5175 * of dirty blocks so that our sync times don't grow too large. 5176 * Note: if two requests come in concurrently, we might let them 5177 * both succeed, when one of them should fail. Not a huge deal. 5178 */ 5179 5180 if (reserve + arc_tempreserve + anon_size > arc_c / 2 && 5181 anon_size > arc_c / 4) { 5182 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK " 5183 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n", 5184 arc_tempreserve>>10, 5185 arc_anon->arcs_lsize[ARC_BUFC_METADATA]>>10, 5186 arc_anon->arcs_lsize[ARC_BUFC_DATA]>>10, 5187 reserve>>10, arc_c>>10); 5188 return (SET_ERROR(ERESTART)); 5189 } 5190 atomic_add_64(&arc_tempreserve, reserve); 5191 return (0); 5192} 5193 5194static void 5195arc_kstat_update_state(arc_state_t *state, kstat_named_t *size, 5196 kstat_named_t *evict_data, kstat_named_t *evict_metadata) 5197{ 5198 size->value.ui64 = refcount_count(&state->arcs_size); 5199 evict_data->value.ui64 = state->arcs_lsize[ARC_BUFC_DATA]; 5200 evict_metadata->value.ui64 = state->arcs_lsize[ARC_BUFC_METADATA]; 5201} 5202 5203static int 5204arc_kstat_update(kstat_t *ksp, int rw) 5205{ 5206 arc_stats_t *as = ksp->ks_data; 5207 5208 if (rw == KSTAT_WRITE) { 5209 return (EACCES); 5210 } else { 5211 arc_kstat_update_state(arc_anon, 5212 &as->arcstat_anon_size, 5213 &as->arcstat_anon_evictable_data, 5214 &as->arcstat_anon_evictable_metadata); 5215 arc_kstat_update_state(arc_mru, 5216 &as->arcstat_mru_size, 5217 &as->arcstat_mru_evictable_data, 5218 &as->arcstat_mru_evictable_metadata); 5219 arc_kstat_update_state(arc_mru_ghost, 5220 &as->arcstat_mru_ghost_size, 5221 &as->arcstat_mru_ghost_evictable_data, 5222 &as->arcstat_mru_ghost_evictable_metadata); 5223 arc_kstat_update_state(arc_mfu, 5224 &as->arcstat_mfu_size, 5225 &as->arcstat_mfu_evictable_data, 5226 &as->arcstat_mfu_evictable_metadata); 5227 arc_kstat_update_state(arc_mfu_ghost, 5228 &as->arcstat_mfu_ghost_size, 5229 &as->arcstat_mfu_ghost_evictable_data, 5230 &as->arcstat_mfu_ghost_evictable_metadata); 5231 } 5232 5233 return (0); 5234} 5235 5236/* 5237 * This function *must* return indices evenly distributed between all 5238 * sublists of the multilist. This is needed due to how the ARC eviction 5239 * code is laid out; arc_evict_state() assumes ARC buffers are evenly 5240 * distributed between all sublists and uses this assumption when 5241 * deciding which sublist to evict from and how much to evict from it. 5242 */ 5243unsigned int 5244arc_state_multilist_index_func(multilist_t *ml, void *obj) 5245{ 5246 arc_buf_hdr_t *hdr = obj; 5247 5248 /* 5249 * We rely on b_dva to generate evenly distributed index 5250 * numbers using buf_hash below. So, as an added precaution, 5251 * let's make sure we never add empty buffers to the arc lists. 5252 */ 5253 ASSERT(!BUF_EMPTY(hdr)); 5254 5255 /* 5256 * The assumption here, is the hash value for a given 5257 * arc_buf_hdr_t will remain constant throughout it's lifetime 5258 * (i.e. it's b_spa, b_dva, and b_birth fields don't change). 5259 * Thus, we don't need to store the header's sublist index 5260 * on insertion, as this index can be recalculated on removal. 5261 * 5262 * Also, the low order bits of the hash value are thought to be 5263 * distributed evenly. Otherwise, in the case that the multilist 5264 * has a power of two number of sublists, each sublists' usage 5265 * would not be evenly distributed. 5266 */ 5267 return (buf_hash(hdr->b_spa, &hdr->b_dva, hdr->b_birth) % 5268 multilist_get_num_sublists(ml)); 5269} 5270 5271#ifdef _KERNEL 5272static eventhandler_tag arc_event_lowmem = NULL; 5273 5274static void 5275arc_lowmem(void *arg __unused, int howto __unused) 5276{ 5277 5278 mutex_enter(&arc_reclaim_lock); 5279 /* XXX: Memory deficit should be passed as argument. */ 5280 needfree = btoc(arc_c >> arc_shrink_shift); 5281 DTRACE_PROBE(arc__needfree); 5282 cv_signal(&arc_reclaim_thread_cv); 5283 5284 /* 5285 * It is unsafe to block here in arbitrary threads, because we can come 5286 * here from ARC itself and may hold ARC locks and thus risk a deadlock 5287 * with ARC reclaim thread. 5288 */ 5289 if (curproc == pageproc) 5290 (void) cv_wait(&arc_reclaim_waiters_cv, &arc_reclaim_lock); 5291 mutex_exit(&arc_reclaim_lock); 5292} 5293#endif 5294 5295void 5296arc_init(void) 5297{ 5298 int i, prefetch_tunable_set = 0; 5299 5300 mutex_init(&arc_reclaim_lock, NULL, MUTEX_DEFAULT, NULL); 5301 cv_init(&arc_reclaim_thread_cv, NULL, CV_DEFAULT, NULL); 5302 cv_init(&arc_reclaim_waiters_cv, NULL, CV_DEFAULT, NULL); 5303 5304 mutex_init(&arc_user_evicts_lock, NULL, MUTEX_DEFAULT, NULL); 5305 cv_init(&arc_user_evicts_cv, NULL, CV_DEFAULT, NULL); 5306 5307 /* Convert seconds to clock ticks */ 5308 arc_min_prefetch_lifespan = 1 * hz; 5309 5310 /* Start out with 1/8 of all memory */ 5311 arc_c = kmem_size() / 8; 5312 5313#ifdef illumos 5314#ifdef _KERNEL 5315 /* 5316 * On architectures where the physical memory can be larger 5317 * than the addressable space (intel in 32-bit mode), we may 5318 * need to limit the cache to 1/8 of VM size. 5319 */ 5320 arc_c = MIN(arc_c, vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 8); 5321#endif 5322#endif /* illumos */ 5323 /* set min cache to 1/32 of all memory, or 16MB, whichever is more */ 5324 arc_c_min = MAX(arc_c / 4, 16 << 20); 5325 /* set max to 1/2 of all memory, or all but 1GB, whichever is more */ 5326 if (arc_c * 8 >= 1 << 30) 5327 arc_c_max = (arc_c * 8) - (1 << 30); 5328 else 5329 arc_c_max = arc_c_min; 5330 arc_c_max = MAX(arc_c * 5, arc_c_max); 5331 5332 /* 5333 * In userland, there's only the memory pressure that we artificially 5334 * create (see arc_available_memory()). Don't let arc_c get too 5335 * small, because it can cause transactions to be larger than 5336 * arc_c, causing arc_tempreserve_space() to fail. 5337 */ 5338#ifndef _KERNEL 5339 arc_c_min = arc_c_max / 2; 5340#endif 5341 5342#ifdef _KERNEL 5343 /* 5344 * Allow the tunables to override our calculations if they are 5345 * reasonable (ie. over 16MB) 5346 */ 5347 if (zfs_arc_max > 16 << 20 && zfs_arc_max < kmem_size()) 5348 arc_c_max = zfs_arc_max; 5349 if (zfs_arc_min > 16 << 20 && zfs_arc_min <= arc_c_max) 5350 arc_c_min = zfs_arc_min; 5351#endif 5352 5353 arc_c = arc_c_max; 5354 arc_p = (arc_c >> 1); 5355 5356 /* limit meta-data to 1/4 of the arc capacity */ 5357 arc_meta_limit = arc_c_max / 4; 5358 5359 /* Allow the tunable to override if it is reasonable */ 5360 if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max) 5361 arc_meta_limit = zfs_arc_meta_limit; 5362 5363 if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0) 5364 arc_c_min = arc_meta_limit / 2; 5365 5366 if (zfs_arc_meta_min > 0) { 5367 arc_meta_min = zfs_arc_meta_min; 5368 } else { 5369 arc_meta_min = arc_c_min / 2; 5370 } 5371 5372 if (zfs_arc_grow_retry > 0) 5373 arc_grow_retry = zfs_arc_grow_retry; 5374 5375 if (zfs_arc_shrink_shift > 0) 5376 arc_shrink_shift = zfs_arc_shrink_shift; 5377 5378 /* 5379 * Ensure that arc_no_grow_shift is less than arc_shrink_shift. 5380 */ 5381 if (arc_no_grow_shift >= arc_shrink_shift) 5382 arc_no_grow_shift = arc_shrink_shift - 1; 5383 5384 if (zfs_arc_p_min_shift > 0) 5385 arc_p_min_shift = zfs_arc_p_min_shift; 5386 5387 if (zfs_arc_num_sublists_per_state < 1) 5388 zfs_arc_num_sublists_per_state = MAX(max_ncpus, 1); 5389 5390 /* if kmem_flags are set, lets try to use less memory */ 5391 if (kmem_debugging()) 5392 arc_c = arc_c / 2; 5393 if (arc_c < arc_c_min) 5394 arc_c = arc_c_min; 5395 5396 zfs_arc_min = arc_c_min; 5397 zfs_arc_max = arc_c_max; 5398 5399 arc_anon = &ARC_anon; 5400 arc_mru = &ARC_mru; 5401 arc_mru_ghost = &ARC_mru_ghost; 5402 arc_mfu = &ARC_mfu; 5403 arc_mfu_ghost = &ARC_mfu_ghost; 5404 arc_l2c_only = &ARC_l2c_only; 5405 arc_size = 0; 5406 5407 multilist_create(&arc_mru->arcs_list[ARC_BUFC_METADATA], 5408 sizeof (arc_buf_hdr_t), 5409 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5410 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5411 multilist_create(&arc_mru->arcs_list[ARC_BUFC_DATA], 5412 sizeof (arc_buf_hdr_t), 5413 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5414 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5415 multilist_create(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA], 5416 sizeof (arc_buf_hdr_t), 5417 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5418 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5419 multilist_create(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA], 5420 sizeof (arc_buf_hdr_t), 5421 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5422 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5423 multilist_create(&arc_mfu->arcs_list[ARC_BUFC_METADATA], 5424 sizeof (arc_buf_hdr_t), 5425 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5426 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5427 multilist_create(&arc_mfu->arcs_list[ARC_BUFC_DATA], 5428 sizeof (arc_buf_hdr_t), 5429 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5430 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5431 multilist_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA], 5432 sizeof (arc_buf_hdr_t), 5433 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5434 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5435 multilist_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA], 5436 sizeof (arc_buf_hdr_t), 5437 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5438 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5439 multilist_create(&arc_l2c_only->arcs_list[ARC_BUFC_METADATA], 5440 sizeof (arc_buf_hdr_t), 5441 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5442 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5443 multilist_create(&arc_l2c_only->arcs_list[ARC_BUFC_DATA], 5444 sizeof (arc_buf_hdr_t), 5445 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5446 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5447 5448 refcount_create(&arc_anon->arcs_size); 5449 refcount_create(&arc_mru->arcs_size); 5450 refcount_create(&arc_mru_ghost->arcs_size); 5451 refcount_create(&arc_mfu->arcs_size); 5452 refcount_create(&arc_mfu_ghost->arcs_size); 5453 refcount_create(&arc_l2c_only->arcs_size); 5454 5455 buf_init(); 5456 5457 arc_reclaim_thread_exit = FALSE; 5458 arc_user_evicts_thread_exit = FALSE; 5459 arc_eviction_list = NULL; 5460 bzero(&arc_eviction_hdr, sizeof (arc_buf_hdr_t)); 5461 5462 arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED, 5463 sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL); 5464 5465 if (arc_ksp != NULL) { 5466 arc_ksp->ks_data = &arc_stats; 5467 arc_ksp->ks_update = arc_kstat_update; 5468 kstat_install(arc_ksp); 5469 } 5470 5471 (void) thread_create(NULL, 0, arc_reclaim_thread, NULL, 0, &p0, 5472 TS_RUN, minclsyspri); 5473 5474#ifdef _KERNEL 5475 arc_event_lowmem = EVENTHANDLER_REGISTER(vm_lowmem, arc_lowmem, NULL, 5476 EVENTHANDLER_PRI_FIRST); 5477#endif 5478 5479 (void) thread_create(NULL, 0, arc_user_evicts_thread, NULL, 0, &p0, 5480 TS_RUN, minclsyspri); 5481 5482 arc_dead = FALSE; 5483 arc_warm = B_FALSE; 5484 5485 /* 5486 * Calculate maximum amount of dirty data per pool. 5487 * 5488 * If it has been set by /etc/system, take that. 5489 * Otherwise, use a percentage of physical memory defined by 5490 * zfs_dirty_data_max_percent (default 10%) with a cap at 5491 * zfs_dirty_data_max_max (default 4GB). 5492 */ 5493 if (zfs_dirty_data_max == 0) { 5494 zfs_dirty_data_max = ptob(physmem) * 5495 zfs_dirty_data_max_percent / 100; 5496 zfs_dirty_data_max = MIN(zfs_dirty_data_max, 5497 zfs_dirty_data_max_max); 5498 } 5499 5500#ifdef _KERNEL 5501 if (TUNABLE_INT_FETCH("vfs.zfs.prefetch_disable", &zfs_prefetch_disable)) 5502 prefetch_tunable_set = 1; 5503 5504#ifdef __i386__ 5505 if (prefetch_tunable_set == 0) { 5506 printf("ZFS NOTICE: Prefetch is disabled by default on i386 " 5507 "-- to enable,\n"); 5508 printf(" add \"vfs.zfs.prefetch_disable=0\" " 5509 "to /boot/loader.conf.\n"); 5510 zfs_prefetch_disable = 1; 5511 } 5512#else 5513 if ((((uint64_t)physmem * PAGESIZE) < (1ULL << 32)) && 5514 prefetch_tunable_set == 0) { 5515 printf("ZFS NOTICE: Prefetch is disabled by default if less " 5516 "than 4GB of RAM is present;\n" 5517 " to enable, add \"vfs.zfs.prefetch_disable=0\" " 5518 "to /boot/loader.conf.\n"); 5519 zfs_prefetch_disable = 1; 5520 } 5521#endif 5522 /* Warn about ZFS memory and address space requirements. */ 5523 if (((uint64_t)physmem * PAGESIZE) < (256 + 128 + 64) * (1 << 20)) { 5524 printf("ZFS WARNING: Recommended minimum RAM size is 512MB; " 5525 "expect unstable behavior.\n"); 5526 } 5527 if (kmem_size() < 512 * (1 << 20)) { 5528 printf("ZFS WARNING: Recommended minimum kmem_size is 512MB; " 5529 "expect unstable behavior.\n"); 5530 printf(" Consider tuning vm.kmem_size and " 5531 "vm.kmem_size_max\n"); 5532 printf(" in /boot/loader.conf.\n"); 5533 } 5534#endif 5535} 5536 5537void 5538arc_fini(void) 5539{ 5540 mutex_enter(&arc_reclaim_lock); 5541 arc_reclaim_thread_exit = TRUE; 5542 /* 5543 * The reclaim thread will set arc_reclaim_thread_exit back to 5544 * FALSE when it is finished exiting; we're waiting for that. 5545 */ 5546 while (arc_reclaim_thread_exit) { 5547 cv_signal(&arc_reclaim_thread_cv); 5548 cv_wait(&arc_reclaim_thread_cv, &arc_reclaim_lock); 5549 } 5550 mutex_exit(&arc_reclaim_lock); 5551 5552 mutex_enter(&arc_user_evicts_lock); 5553 arc_user_evicts_thread_exit = TRUE; 5554 /* 5555 * The user evicts thread will set arc_user_evicts_thread_exit 5556 * to FALSE when it is finished exiting; we're waiting for that. 5557 */ 5558 while (arc_user_evicts_thread_exit) { 5559 cv_signal(&arc_user_evicts_cv); 5560 cv_wait(&arc_user_evicts_cv, &arc_user_evicts_lock); 5561 } 5562 mutex_exit(&arc_user_evicts_lock); 5563 5564 /* Use TRUE to ensure *all* buffers are evicted */ 5565 arc_flush(NULL, TRUE); 5566 5567 arc_dead = TRUE; 5568 5569 if (arc_ksp != NULL) { 5570 kstat_delete(arc_ksp); 5571 arc_ksp = NULL; 5572 } 5573 5574 mutex_destroy(&arc_reclaim_lock); 5575 cv_destroy(&arc_reclaim_thread_cv); 5576 cv_destroy(&arc_reclaim_waiters_cv); 5577 5578 mutex_destroy(&arc_user_evicts_lock); 5579 cv_destroy(&arc_user_evicts_cv); 5580 5581 refcount_destroy(&arc_anon->arcs_size); 5582 refcount_destroy(&arc_mru->arcs_size); 5583 refcount_destroy(&arc_mru_ghost->arcs_size); 5584 refcount_destroy(&arc_mfu->arcs_size); 5585 refcount_destroy(&arc_mfu_ghost->arcs_size); 5586 refcount_destroy(&arc_l2c_only->arcs_size); 5587 5588 multilist_destroy(&arc_mru->arcs_list[ARC_BUFC_METADATA]); 5589 multilist_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA]); 5590 multilist_destroy(&arc_mfu->arcs_list[ARC_BUFC_METADATA]); 5591 multilist_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA]); 5592 multilist_destroy(&arc_l2c_only->arcs_list[ARC_BUFC_METADATA]); 5593 multilist_destroy(&arc_mru->arcs_list[ARC_BUFC_DATA]); 5594 multilist_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA]); 5595 multilist_destroy(&arc_mfu->arcs_list[ARC_BUFC_DATA]); 5596 multilist_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA]); 5597 multilist_destroy(&arc_l2c_only->arcs_list[ARC_BUFC_DATA]); 5598 5599 buf_fini(); 5600 5601 ASSERT0(arc_loaned_bytes); 5602 5603#ifdef _KERNEL 5604 if (arc_event_lowmem != NULL) 5605 EVENTHANDLER_DEREGISTER(vm_lowmem, arc_event_lowmem); 5606#endif 5607} 5608 5609/* 5610 * Level 2 ARC 5611 * 5612 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk. 5613 * It uses dedicated storage devices to hold cached data, which are populated 5614 * using large infrequent writes. The main role of this cache is to boost 5615 * the performance of random read workloads. The intended L2ARC devices 5616 * include short-stroked disks, solid state disks, and other media with 5617 * substantially faster read latency than disk. 5618 * 5619 * +-----------------------+ 5620 * | ARC | 5621 * +-----------------------+ 5622 * | ^ ^ 5623 * | | | 5624 * l2arc_feed_thread() arc_read() 5625 * | | | 5626 * | l2arc read | 5627 * V | | 5628 * +---------------+ | 5629 * | L2ARC | | 5630 * +---------------+ | 5631 * | ^ | 5632 * l2arc_write() | | 5633 * | | | 5634 * V | | 5635 * +-------+ +-------+ 5636 * | vdev | | vdev | 5637 * | cache | | cache | 5638 * +-------+ +-------+ 5639 * +=========+ .-----. 5640 * : L2ARC : |-_____-| 5641 * : devices : | Disks | 5642 * +=========+ `-_____-' 5643 * 5644 * Read requests are satisfied from the following sources, in order: 5645 * 5646 * 1) ARC 5647 * 2) vdev cache of L2ARC devices 5648 * 3) L2ARC devices 5649 * 4) vdev cache of disks 5650 * 5) disks 5651 * 5652 * Some L2ARC device types exhibit extremely slow write performance. 5653 * To accommodate for this there are some significant differences between 5654 * the L2ARC and traditional cache design: 5655 * 5656 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from 5657 * the ARC behave as usual, freeing buffers and placing headers on ghost 5658 * lists. The ARC does not send buffers to the L2ARC during eviction as 5659 * this would add inflated write latencies for all ARC memory pressure. 5660 * 5661 * 2. The L2ARC attempts to cache data from the ARC before it is evicted. 5662 * It does this by periodically scanning buffers from the eviction-end of 5663 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are 5664 * not already there. It scans until a headroom of buffers is satisfied, 5665 * which itself is a buffer for ARC eviction. If a compressible buffer is 5666 * found during scanning and selected for writing to an L2ARC device, we 5667 * temporarily boost scanning headroom during the next scan cycle to make 5668 * sure we adapt to compression effects (which might significantly reduce 5669 * the data volume we write to L2ARC). The thread that does this is 5670 * l2arc_feed_thread(), illustrated below; example sizes are included to 5671 * provide a better sense of ratio than this diagram: 5672 * 5673 * head --> tail 5674 * +---------------------+----------+ 5675 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC 5676 * +---------------------+----------+ | o L2ARC eligible 5677 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer 5678 * +---------------------+----------+ | 5679 * 15.9 Gbytes ^ 32 Mbytes | 5680 * headroom | 5681 * l2arc_feed_thread() 5682 * | 5683 * l2arc write hand <--[oooo]--' 5684 * | 8 Mbyte 5685 * | write max 5686 * V 5687 * +==============================+ 5688 * L2ARC dev |####|#|###|###| |####| ... | 5689 * +==============================+ 5690 * 32 Gbytes 5691 * 5692 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of 5693 * evicted, then the L2ARC has cached a buffer much sooner than it probably 5694 * needed to, potentially wasting L2ARC device bandwidth and storage. It is 5695 * safe to say that this is an uncommon case, since buffers at the end of 5696 * the ARC lists have moved there due to inactivity. 5697 * 5698 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom, 5699 * then the L2ARC simply misses copying some buffers. This serves as a 5700 * pressure valve to prevent heavy read workloads from both stalling the ARC 5701 * with waits and clogging the L2ARC with writes. This also helps prevent 5702 * the potential for the L2ARC to churn if it attempts to cache content too 5703 * quickly, such as during backups of the entire pool. 5704 * 5705 * 5. After system boot and before the ARC has filled main memory, there are 5706 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru 5707 * lists can remain mostly static. Instead of searching from tail of these 5708 * lists as pictured, the l2arc_feed_thread() will search from the list heads 5709 * for eligible buffers, greatly increasing its chance of finding them. 5710 * 5711 * The L2ARC device write speed is also boosted during this time so that 5712 * the L2ARC warms up faster. Since there have been no ARC evictions yet, 5713 * there are no L2ARC reads, and no fear of degrading read performance 5714 * through increased writes. 5715 * 5716 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that 5717 * the vdev queue can aggregate them into larger and fewer writes. Each 5718 * device is written to in a rotor fashion, sweeping writes through 5719 * available space then repeating. 5720 * 5721 * 7. The L2ARC does not store dirty content. It never needs to flush 5722 * write buffers back to disk based storage. 5723 * 5724 * 8. If an ARC buffer is written (and dirtied) which also exists in the 5725 * L2ARC, the now stale L2ARC buffer is immediately dropped. 5726 * 5727 * The performance of the L2ARC can be tweaked by a number of tunables, which 5728 * may be necessary for different workloads: 5729 * 5730 * l2arc_write_max max write bytes per interval 5731 * l2arc_write_boost extra write bytes during device warmup 5732 * l2arc_noprefetch skip caching prefetched buffers 5733 * l2arc_headroom number of max device writes to precache 5734 * l2arc_headroom_boost when we find compressed buffers during ARC 5735 * scanning, we multiply headroom by this 5736 * percentage factor for the next scan cycle, 5737 * since more compressed buffers are likely to 5738 * be present 5739 * l2arc_feed_secs seconds between L2ARC writing 5740 * 5741 * Tunables may be removed or added as future performance improvements are 5742 * integrated, and also may become zpool properties. 5743 * 5744 * There are three key functions that control how the L2ARC warms up: 5745 * 5746 * l2arc_write_eligible() check if a buffer is eligible to cache 5747 * l2arc_write_size() calculate how much to write 5748 * l2arc_write_interval() calculate sleep delay between writes 5749 * 5750 * These three functions determine what to write, how much, and how quickly 5751 * to send writes. 5752 */ 5753 5754static boolean_t 5755l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *hdr) 5756{ 5757 /* 5758 * A buffer is *not* eligible for the L2ARC if it: 5759 * 1. belongs to a different spa. 5760 * 2. is already cached on the L2ARC. 5761 * 3. has an I/O in progress (it may be an incomplete read). 5762 * 4. is flagged not eligible (zfs property). 5763 */ 5764 if (hdr->b_spa != spa_guid) { 5765 ARCSTAT_BUMP(arcstat_l2_write_spa_mismatch); 5766 return (B_FALSE); 5767 } 5768 if (HDR_HAS_L2HDR(hdr)) { 5769 ARCSTAT_BUMP(arcstat_l2_write_in_l2); 5770 return (B_FALSE); 5771 } 5772 if (HDR_IO_IN_PROGRESS(hdr)) { 5773 ARCSTAT_BUMP(arcstat_l2_write_hdr_io_in_progress); 5774 return (B_FALSE); 5775 } 5776 if (!HDR_L2CACHE(hdr)) { 5777 ARCSTAT_BUMP(arcstat_l2_write_not_cacheable); 5778 return (B_FALSE); 5779 } 5780 5781 return (B_TRUE); 5782} 5783 5784static uint64_t 5785l2arc_write_size(void) 5786{ 5787 uint64_t size; 5788 5789 /* 5790 * Make sure our globals have meaningful values in case the user 5791 * altered them. 5792 */ 5793 size = l2arc_write_max; 5794 if (size == 0) { 5795 cmn_err(CE_NOTE, "Bad value for l2arc_write_max, value must " 5796 "be greater than zero, resetting it to the default (%d)", 5797 L2ARC_WRITE_SIZE); 5798 size = l2arc_write_max = L2ARC_WRITE_SIZE; 5799 } 5800 5801 if (arc_warm == B_FALSE) 5802 size += l2arc_write_boost; 5803 5804 return (size); 5805 5806} 5807 5808static clock_t 5809l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote) 5810{ 5811 clock_t interval, next, now; 5812 5813 /* 5814 * If the ARC lists are busy, increase our write rate; if the 5815 * lists are stale, idle back. This is achieved by checking 5816 * how much we previously wrote - if it was more than half of 5817 * what we wanted, schedule the next write much sooner. 5818 */ 5819 if (l2arc_feed_again && wrote > (wanted / 2)) 5820 interval = (hz * l2arc_feed_min_ms) / 1000; 5821 else 5822 interval = hz * l2arc_feed_secs; 5823 5824 now = ddi_get_lbolt(); 5825 next = MAX(now, MIN(now + interval, began + interval)); 5826 5827 return (next); 5828} 5829 5830/* 5831 * Cycle through L2ARC devices. This is how L2ARC load balances. 5832 * If a device is returned, this also returns holding the spa config lock. 5833 */ 5834static l2arc_dev_t * 5835l2arc_dev_get_next(void) 5836{ 5837 l2arc_dev_t *first, *next = NULL; 5838 5839 /* 5840 * Lock out the removal of spas (spa_namespace_lock), then removal 5841 * of cache devices (l2arc_dev_mtx). Once a device has been selected, 5842 * both locks will be dropped and a spa config lock held instead. 5843 */ 5844 mutex_enter(&spa_namespace_lock); 5845 mutex_enter(&l2arc_dev_mtx); 5846 5847 /* if there are no vdevs, there is nothing to do */ 5848 if (l2arc_ndev == 0) 5849 goto out; 5850 5851 first = NULL; 5852 next = l2arc_dev_last; 5853 do { 5854 /* loop around the list looking for a non-faulted vdev */ 5855 if (next == NULL) { 5856 next = list_head(l2arc_dev_list); 5857 } else { 5858 next = list_next(l2arc_dev_list, next); 5859 if (next == NULL) 5860 next = list_head(l2arc_dev_list); 5861 } 5862 5863 /* if we have come back to the start, bail out */ 5864 if (first == NULL) 5865 first = next; 5866 else if (next == first) 5867 break; 5868 5869 } while (vdev_is_dead(next->l2ad_vdev)); 5870 5871 /* if we were unable to find any usable vdevs, return NULL */ 5872 if (vdev_is_dead(next->l2ad_vdev)) 5873 next = NULL; 5874 5875 l2arc_dev_last = next; 5876 5877out: 5878 mutex_exit(&l2arc_dev_mtx); 5879 5880 /* 5881 * Grab the config lock to prevent the 'next' device from being 5882 * removed while we are writing to it. 5883 */ 5884 if (next != NULL) 5885 spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER); 5886 mutex_exit(&spa_namespace_lock); 5887 5888 return (next); 5889} 5890 5891/* 5892 * Free buffers that were tagged for destruction. 5893 */ 5894static void 5895l2arc_do_free_on_write() 5896{ 5897 list_t *buflist; 5898 l2arc_data_free_t *df, *df_prev; 5899 5900 mutex_enter(&l2arc_free_on_write_mtx); 5901 buflist = l2arc_free_on_write; 5902 5903 for (df = list_tail(buflist); df; df = df_prev) { 5904 df_prev = list_prev(buflist, df); 5905 ASSERT(df->l2df_data != NULL); 5906 ASSERT(df->l2df_func != NULL); 5907 df->l2df_func(df->l2df_data, df->l2df_size); 5908 list_remove(buflist, df); 5909 kmem_free(df, sizeof (l2arc_data_free_t)); 5910 } 5911 5912 mutex_exit(&l2arc_free_on_write_mtx); 5913} 5914 5915/* 5916 * A write to a cache device has completed. Update all headers to allow 5917 * reads from these buffers to begin. 5918 */ 5919static void 5920l2arc_write_done(zio_t *zio) 5921{ 5922 l2arc_write_callback_t *cb; 5923 l2arc_dev_t *dev; 5924 list_t *buflist; 5925 arc_buf_hdr_t *head, *hdr, *hdr_prev; 5926 kmutex_t *hash_lock; 5927 int64_t bytes_dropped = 0; 5928 5929 cb = zio->io_private; 5930 ASSERT(cb != NULL); 5931 dev = cb->l2wcb_dev; 5932 ASSERT(dev != NULL); 5933 head = cb->l2wcb_head; 5934 ASSERT(head != NULL); 5935 buflist = &dev->l2ad_buflist; 5936 ASSERT(buflist != NULL); 5937 DTRACE_PROBE2(l2arc__iodone, zio_t *, zio, 5938 l2arc_write_callback_t *, cb); 5939 5940 if (zio->io_error != 0) 5941 ARCSTAT_BUMP(arcstat_l2_writes_error); 5942 5943 /* 5944 * All writes completed, or an error was hit. 5945 */ 5946top: 5947 mutex_enter(&dev->l2ad_mtx); 5948 for (hdr = list_prev(buflist, head); hdr; hdr = hdr_prev) { 5949 hdr_prev = list_prev(buflist, hdr); 5950 5951 hash_lock = HDR_LOCK(hdr); 5952 5953 /* 5954 * We cannot use mutex_enter or else we can deadlock 5955 * with l2arc_write_buffers (due to swapping the order 5956 * the hash lock and l2ad_mtx are taken). 5957 */ 5958 if (!mutex_tryenter(hash_lock)) { 5959 /* 5960 * Missed the hash lock. We must retry so we 5961 * don't leave the ARC_FLAG_L2_WRITING bit set. 5962 */ 5963 ARCSTAT_BUMP(arcstat_l2_writes_lock_retry); 5964 5965 /* 5966 * We don't want to rescan the headers we've 5967 * already marked as having been written out, so 5968 * we reinsert the head node so we can pick up 5969 * where we left off. 5970 */ 5971 list_remove(buflist, head); 5972 list_insert_after(buflist, hdr, head); 5973 5974 mutex_exit(&dev->l2ad_mtx); 5975 5976 /* 5977 * We wait for the hash lock to become available 5978 * to try and prevent busy waiting, and increase 5979 * the chance we'll be able to acquire the lock 5980 * the next time around. 5981 */ 5982 mutex_enter(hash_lock); 5983 mutex_exit(hash_lock); 5984 goto top; 5985 } 5986 5987 /* 5988 * We could not have been moved into the arc_l2c_only 5989 * state while in-flight due to our ARC_FLAG_L2_WRITING 5990 * bit being set. Let's just ensure that's being enforced. 5991 */ 5992 ASSERT(HDR_HAS_L1HDR(hdr)); 5993 5994 /* 5995 * We may have allocated a buffer for L2ARC compression, 5996 * we must release it to avoid leaking this data. 5997 */ 5998 l2arc_release_cdata_buf(hdr); 5999 6000 if (zio->io_error != 0) { 6001 /* 6002 * Error - drop L2ARC entry. 6003 */ 6004 list_remove(buflist, hdr); 6005 l2arc_trim(hdr); 6006 hdr->b_flags &= ~ARC_FLAG_HAS_L2HDR; 6007 6008 ARCSTAT_INCR(arcstat_l2_asize, -hdr->b_l2hdr.b_asize); 6009 ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size); 6010 6011 bytes_dropped += hdr->b_l2hdr.b_asize; 6012 (void) refcount_remove_many(&dev->l2ad_alloc, 6013 hdr->b_l2hdr.b_asize, hdr); 6014 } 6015 6016 /* 6017 * Allow ARC to begin reads and ghost list evictions to 6018 * this L2ARC entry. 6019 */ 6020 hdr->b_flags &= ~ARC_FLAG_L2_WRITING; 6021 6022 mutex_exit(hash_lock); 6023 } 6024 6025 atomic_inc_64(&l2arc_writes_done); 6026 list_remove(buflist, head); 6027 ASSERT(!HDR_HAS_L1HDR(head)); 6028 kmem_cache_free(hdr_l2only_cache, head); 6029 mutex_exit(&dev->l2ad_mtx); 6030 6031 vdev_space_update(dev->l2ad_vdev, -bytes_dropped, 0, 0); 6032 6033 l2arc_do_free_on_write(); 6034 6035 kmem_free(cb, sizeof (l2arc_write_callback_t)); 6036} 6037 6038/* 6039 * A read to a cache device completed. Validate buffer contents before 6040 * handing over to the regular ARC routines. 6041 */ 6042static void 6043l2arc_read_done(zio_t *zio) 6044{ 6045 l2arc_read_callback_t *cb; 6046 arc_buf_hdr_t *hdr; 6047 arc_buf_t *buf; 6048 kmutex_t *hash_lock; 6049 int equal; 6050 6051 ASSERT(zio->io_vd != NULL); 6052 ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE); 6053 6054 spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd); 6055 6056 cb = zio->io_private; 6057 ASSERT(cb != NULL); 6058 buf = cb->l2rcb_buf; 6059 ASSERT(buf != NULL); 6060 6061 hash_lock = HDR_LOCK(buf->b_hdr); 6062 mutex_enter(hash_lock); 6063 hdr = buf->b_hdr; 6064 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 6065 6066 /* 6067 * If the data was read into a temporary buffer, 6068 * move it and free the buffer. 6069 */ 6070 if (cb->l2rcb_data != NULL) { 6071 ASSERT3U(hdr->b_size, <, zio->io_size); 6072 ASSERT3U(cb->l2rcb_compress, ==, ZIO_COMPRESS_OFF); 6073 if (zio->io_error == 0) 6074 bcopy(cb->l2rcb_data, buf->b_data, hdr->b_size); 6075 6076 /* 6077 * The following must be done regardless of whether 6078 * there was an error: 6079 * - free the temporary buffer 6080 * - point zio to the real ARC buffer 6081 * - set zio size accordingly 6082 * These are required because zio is either re-used for 6083 * an I/O of the block in the case of the error 6084 * or the zio is passed to arc_read_done() and it 6085 * needs real data. 6086 */ 6087 zio_data_buf_free(cb->l2rcb_data, zio->io_size); 6088 zio->io_size = zio->io_orig_size = hdr->b_size; 6089 zio->io_data = zio->io_orig_data = buf->b_data; 6090 } 6091 6092 /* 6093 * If the buffer was compressed, decompress it first. 6094 */ 6095 if (cb->l2rcb_compress != ZIO_COMPRESS_OFF) 6096 l2arc_decompress_zio(zio, hdr, cb->l2rcb_compress); 6097 ASSERT(zio->io_data != NULL); 6098 ASSERT3U(zio->io_size, ==, hdr->b_size); 6099 ASSERT3U(BP_GET_LSIZE(&cb->l2rcb_bp), ==, hdr->b_size); 6100 6101 /* 6102 * Check this survived the L2ARC journey. 6103 */ 6104 equal = arc_cksum_equal(buf); 6105 if (equal && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) { 6106 mutex_exit(hash_lock); 6107 zio->io_private = buf; 6108 zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */ 6109 zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */ 6110 arc_read_done(zio); 6111 } else { 6112 mutex_exit(hash_lock); 6113 /* 6114 * Buffer didn't survive caching. Increment stats and 6115 * reissue to the original storage device. 6116 */ 6117 if (zio->io_error != 0) { 6118 ARCSTAT_BUMP(arcstat_l2_io_error); 6119 } else { 6120 zio->io_error = SET_ERROR(EIO); 6121 } 6122 if (!equal) 6123 ARCSTAT_BUMP(arcstat_l2_cksum_bad); 6124 6125 /* 6126 * If there's no waiter, issue an async i/o to the primary 6127 * storage now. If there *is* a waiter, the caller must 6128 * issue the i/o in a context where it's OK to block. 6129 */ 6130 if (zio->io_waiter == NULL) { 6131 zio_t *pio = zio_unique_parent(zio); 6132 6133 ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL); 6134 6135 zio_nowait(zio_read(pio, cb->l2rcb_spa, &cb->l2rcb_bp, 6136 buf->b_data, hdr->b_size, arc_read_done, buf, 6137 zio->io_priority, cb->l2rcb_flags, &cb->l2rcb_zb)); 6138 } 6139 } 6140 6141 kmem_free(cb, sizeof (l2arc_read_callback_t)); 6142} 6143 6144/* 6145 * This is the list priority from which the L2ARC will search for pages to 6146 * cache. This is used within loops (0..3) to cycle through lists in the 6147 * desired order. This order can have a significant effect on cache 6148 * performance. 6149 * 6150 * Currently the metadata lists are hit first, MFU then MRU, followed by 6151 * the data lists. This function returns a locked list, and also returns 6152 * the lock pointer. 6153 */ 6154static multilist_sublist_t * 6155l2arc_sublist_lock(int list_num) 6156{ 6157 multilist_t *ml = NULL; 6158 unsigned int idx; 6159 6160 ASSERT(list_num >= 0 && list_num <= 3); 6161 6162 switch (list_num) { 6163 case 0: 6164 ml = &arc_mfu->arcs_list[ARC_BUFC_METADATA]; 6165 break; 6166 case 1: 6167 ml = &arc_mru->arcs_list[ARC_BUFC_METADATA]; 6168 break; 6169 case 2: 6170 ml = &arc_mfu->arcs_list[ARC_BUFC_DATA]; 6171 break; 6172 case 3: 6173 ml = &arc_mru->arcs_list[ARC_BUFC_DATA]; 6174 break; 6175 } 6176 6177 /* 6178 * Return a randomly-selected sublist. This is acceptable 6179 * because the caller feeds only a little bit of data for each 6180 * call (8MB). Subsequent calls will result in different 6181 * sublists being selected. 6182 */ 6183 idx = multilist_get_random_index(ml); 6184 return (multilist_sublist_lock(ml, idx)); 6185} 6186 6187/* 6188 * Evict buffers from the device write hand to the distance specified in 6189 * bytes. This distance may span populated buffers, it may span nothing. 6190 * This is clearing a region on the L2ARC device ready for writing. 6191 * If the 'all' boolean is set, every buffer is evicted. 6192 */ 6193static void 6194l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all) 6195{ 6196 list_t *buflist; 6197 arc_buf_hdr_t *hdr, *hdr_prev; 6198 kmutex_t *hash_lock; 6199 uint64_t taddr; 6200 6201 buflist = &dev->l2ad_buflist; 6202 6203 if (!all && dev->l2ad_first) { 6204 /* 6205 * This is the first sweep through the device. There is 6206 * nothing to evict. 6207 */ 6208 return; 6209 } 6210 6211 if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) { 6212 /* 6213 * When nearing the end of the device, evict to the end 6214 * before the device write hand jumps to the start. 6215 */ 6216 taddr = dev->l2ad_end; 6217 } else { 6218 taddr = dev->l2ad_hand + distance; 6219 } 6220 DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist, 6221 uint64_t, taddr, boolean_t, all); 6222 6223top: 6224 mutex_enter(&dev->l2ad_mtx); 6225 for (hdr = list_tail(buflist); hdr; hdr = hdr_prev) { 6226 hdr_prev = list_prev(buflist, hdr); 6227 6228 hash_lock = HDR_LOCK(hdr); 6229 6230 /* 6231 * We cannot use mutex_enter or else we can deadlock 6232 * with l2arc_write_buffers (due to swapping the order 6233 * the hash lock and l2ad_mtx are taken). 6234 */ 6235 if (!mutex_tryenter(hash_lock)) { 6236 /* 6237 * Missed the hash lock. Retry. 6238 */ 6239 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry); 6240 mutex_exit(&dev->l2ad_mtx); 6241 mutex_enter(hash_lock); 6242 mutex_exit(hash_lock); 6243 goto top; 6244 } 6245 6246 if (HDR_L2_WRITE_HEAD(hdr)) { 6247 /* 6248 * We hit a write head node. Leave it for 6249 * l2arc_write_done(). 6250 */ 6251 list_remove(buflist, hdr); 6252 mutex_exit(hash_lock); 6253 continue; 6254 } 6255 6256 if (!all && HDR_HAS_L2HDR(hdr) && 6257 (hdr->b_l2hdr.b_daddr > taddr || 6258 hdr->b_l2hdr.b_daddr < dev->l2ad_hand)) { 6259 /* 6260 * We've evicted to the target address, 6261 * or the end of the device. 6262 */ 6263 mutex_exit(hash_lock); 6264 break; 6265 } 6266 6267 ASSERT(HDR_HAS_L2HDR(hdr)); 6268 if (!HDR_HAS_L1HDR(hdr)) { 6269 ASSERT(!HDR_L2_READING(hdr)); 6270 /* 6271 * This doesn't exist in the ARC. Destroy. 6272 * arc_hdr_destroy() will call list_remove() 6273 * and decrement arcstat_l2_size. 6274 */ 6275 arc_change_state(arc_anon, hdr, hash_lock); 6276 arc_hdr_destroy(hdr); 6277 } else { 6278 ASSERT(hdr->b_l1hdr.b_state != arc_l2c_only); 6279 ARCSTAT_BUMP(arcstat_l2_evict_l1cached); 6280 /* 6281 * Invalidate issued or about to be issued 6282 * reads, since we may be about to write 6283 * over this location. 6284 */ 6285 if (HDR_L2_READING(hdr)) { 6286 ARCSTAT_BUMP(arcstat_l2_evict_reading); 6287 hdr->b_flags |= ARC_FLAG_L2_EVICTED; 6288 } 6289 6290 /* Ensure this header has finished being written */ 6291 ASSERT(!HDR_L2_WRITING(hdr)); 6292 ASSERT3P(hdr->b_l1hdr.b_tmp_cdata, ==, NULL); 6293 6294 arc_hdr_l2hdr_destroy(hdr); 6295 } 6296 mutex_exit(hash_lock); 6297 } 6298 mutex_exit(&dev->l2ad_mtx); 6299} 6300 6301/* 6302 * Find and write ARC buffers to the L2ARC device. 6303 * 6304 * An ARC_FLAG_L2_WRITING flag is set so that the L2ARC buffers are not valid 6305 * for reading until they have completed writing. 6306 * The headroom_boost is an in-out parameter used to maintain headroom boost 6307 * state between calls to this function. 6308 * 6309 * Returns the number of bytes actually written (which may be smaller than 6310 * the delta by which the device hand has changed due to alignment). 6311 */ 6312static uint64_t 6313l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz, 6314 boolean_t *headroom_boost) 6315{ 6316 arc_buf_hdr_t *hdr, *hdr_prev, *head; 6317 uint64_t write_asize, write_sz, headroom, 6318 buf_compress_minsz; 6319 void *buf_data; 6320 boolean_t full; 6321 l2arc_write_callback_t *cb; 6322 zio_t *pio, *wzio; 6323 uint64_t guid = spa_load_guid(spa); 6324 const boolean_t do_headroom_boost = *headroom_boost; 6325 int try; 6326 6327 ASSERT(dev->l2ad_vdev != NULL); 6328 6329 /* Lower the flag now, we might want to raise it again later. */ 6330 *headroom_boost = B_FALSE; 6331 6332 pio = NULL; 6333 write_sz = write_asize = 0; 6334 full = B_FALSE; 6335 head = kmem_cache_alloc(hdr_l2only_cache, KM_PUSHPAGE); 6336 head->b_flags |= ARC_FLAG_L2_WRITE_HEAD; 6337 head->b_flags |= ARC_FLAG_HAS_L2HDR; 6338 6339 ARCSTAT_BUMP(arcstat_l2_write_buffer_iter); 6340 /* 6341 * We will want to try to compress buffers that are at least 2x the 6342 * device sector size. 6343 */ 6344 buf_compress_minsz = 2 << dev->l2ad_vdev->vdev_ashift; 6345 6346 /* 6347 * Copy buffers for L2ARC writing. 6348 */ 6349 for (try = 0; try <= 3; try++) { 6350 multilist_sublist_t *mls = l2arc_sublist_lock(try); 6351 uint64_t passed_sz = 0; 6352 6353 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_iter); 6354 6355 /* 6356 * L2ARC fast warmup. 6357 * 6358 * Until the ARC is warm and starts to evict, read from the 6359 * head of the ARC lists rather than the tail. 6360 */ 6361 if (arc_warm == B_FALSE) 6362 hdr = multilist_sublist_head(mls); 6363 else 6364 hdr = multilist_sublist_tail(mls); 6365 if (hdr == NULL) 6366 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_null_iter); 6367 6368 headroom = target_sz * l2arc_headroom; 6369 if (do_headroom_boost) 6370 headroom = (headroom * l2arc_headroom_boost) / 100; 6371 6372 for (; hdr; hdr = hdr_prev) { 6373 kmutex_t *hash_lock; 6374 uint64_t buf_sz; 6375 uint64_t buf_a_sz; 6376 size_t align; 6377 6378 if (arc_warm == B_FALSE) 6379 hdr_prev = multilist_sublist_next(mls, hdr); 6380 else 6381 hdr_prev = multilist_sublist_prev(mls, hdr); 6382 ARCSTAT_INCR(arcstat_l2_write_buffer_bytes_scanned, hdr->b_size); 6383 6384 hash_lock = HDR_LOCK(hdr); 6385 if (!mutex_tryenter(hash_lock)) { 6386 ARCSTAT_BUMP(arcstat_l2_write_trylock_fail); 6387 /* 6388 * Skip this buffer rather than waiting. 6389 */ 6390 continue; 6391 } 6392 6393 passed_sz += hdr->b_size; 6394 if (passed_sz > headroom) { 6395 /* 6396 * Searched too far. 6397 */ 6398 mutex_exit(hash_lock); 6399 ARCSTAT_BUMP(arcstat_l2_write_passed_headroom); 6400 break; 6401 } 6402 6403 if (!l2arc_write_eligible(guid, hdr)) { 6404 mutex_exit(hash_lock); 6405 continue; 6406 } 6407 6408 /* 6409 * Assume that the buffer is not going to be compressed 6410 * and could take more space on disk because of a larger 6411 * disk block size. 6412 */ 6413 buf_sz = hdr->b_size; 6414 align = (size_t)1 << dev->l2ad_vdev->vdev_ashift; 6415 buf_a_sz = P2ROUNDUP(buf_sz, align); 6416 6417 if ((write_asize + buf_a_sz) > target_sz) { 6418 full = B_TRUE; 6419 mutex_exit(hash_lock); 6420 ARCSTAT_BUMP(arcstat_l2_write_full); 6421 break; 6422 } 6423 6424 if (pio == NULL) { 6425 /* 6426 * Insert a dummy header on the buflist so 6427 * l2arc_write_done() can find where the 6428 * write buffers begin without searching. 6429 */ 6430 mutex_enter(&dev->l2ad_mtx); 6431 list_insert_head(&dev->l2ad_buflist, head); 6432 mutex_exit(&dev->l2ad_mtx); 6433 6434 cb = kmem_alloc( 6435 sizeof (l2arc_write_callback_t), KM_SLEEP); 6436 cb->l2wcb_dev = dev; 6437 cb->l2wcb_head = head; 6438 pio = zio_root(spa, l2arc_write_done, cb, 6439 ZIO_FLAG_CANFAIL); 6440 ARCSTAT_BUMP(arcstat_l2_write_pios); 6441 } 6442 6443 /* 6444 * Create and add a new L2ARC header. 6445 */ 6446 hdr->b_l2hdr.b_dev = dev; 6447 hdr->b_flags |= ARC_FLAG_L2_WRITING; 6448 /* 6449 * Temporarily stash the data buffer in b_tmp_cdata. 6450 * The subsequent write step will pick it up from 6451 * there. This is because can't access b_l1hdr.b_buf 6452 * without holding the hash_lock, which we in turn 6453 * can't access without holding the ARC list locks 6454 * (which we want to avoid during compression/writing). 6455 */ 6456 hdr->b_l2hdr.b_compress = ZIO_COMPRESS_OFF; 6457 hdr->b_l2hdr.b_asize = hdr->b_size; 6458 hdr->b_l1hdr.b_tmp_cdata = hdr->b_l1hdr.b_buf->b_data; 6459 6460 /* 6461 * Explicitly set the b_daddr field to a known 6462 * value which means "invalid address". This 6463 * enables us to differentiate which stage of 6464 * l2arc_write_buffers() the particular header 6465 * is in (e.g. this loop, or the one below). 6466 * ARC_FLAG_L2_WRITING is not enough to make 6467 * this distinction, and we need to know in 6468 * order to do proper l2arc vdev accounting in 6469 * arc_release() and arc_hdr_destroy(). 6470 * 6471 * Note, we can't use a new flag to distinguish 6472 * the two stages because we don't hold the 6473 * header's hash_lock below, in the second stage 6474 * of this function. Thus, we can't simply 6475 * change the b_flags field to denote that the 6476 * IO has been sent. We can change the b_daddr 6477 * field of the L2 portion, though, since we'll 6478 * be holding the l2ad_mtx; which is why we're 6479 * using it to denote the header's state change. 6480 */ 6481 hdr->b_l2hdr.b_daddr = L2ARC_ADDR_UNSET; 6482 6483 hdr->b_flags |= ARC_FLAG_HAS_L2HDR; 6484 6485 mutex_enter(&dev->l2ad_mtx); 6486 list_insert_head(&dev->l2ad_buflist, hdr); 6487 mutex_exit(&dev->l2ad_mtx); 6488 6489 /* 6490 * Compute and store the buffer cksum before 6491 * writing. On debug the cksum is verified first. 6492 */ 6493 arc_cksum_verify(hdr->b_l1hdr.b_buf); 6494 arc_cksum_compute(hdr->b_l1hdr.b_buf, B_TRUE); 6495 6496 mutex_exit(hash_lock); 6497 6498 write_sz += buf_sz; 6499 write_asize += buf_a_sz; 6500 } 6501 6502 multilist_sublist_unlock(mls); 6503 6504 if (full == B_TRUE) 6505 break; 6506 } 6507 6508 /* No buffers selected for writing? */ 6509 if (pio == NULL) { 6510 ASSERT0(write_sz); 6511 ASSERT(!HDR_HAS_L1HDR(head)); 6512 kmem_cache_free(hdr_l2only_cache, head); 6513 return (0); 6514 } 6515 6516 mutex_enter(&dev->l2ad_mtx); 6517 6518 /* 6519 * Now start writing the buffers. We're starting at the write head 6520 * and work backwards, retracing the course of the buffer selector 6521 * loop above. 6522 */ 6523 write_asize = 0; 6524 for (hdr = list_prev(&dev->l2ad_buflist, head); hdr; 6525 hdr = list_prev(&dev->l2ad_buflist, hdr)) { 6526 uint64_t buf_sz; 6527 boolean_t compress; 6528 6529 /* 6530 * We rely on the L1 portion of the header below, so 6531 * it's invalid for this header to have been evicted out 6532 * of the ghost cache, prior to being written out. The 6533 * ARC_FLAG_L2_WRITING bit ensures this won't happen. 6534 */ 6535 ASSERT(HDR_HAS_L1HDR(hdr)); 6536 6537 /* 6538 * We shouldn't need to lock the buffer here, since we flagged 6539 * it as ARC_FLAG_L2_WRITING in the previous step, but we must 6540 * take care to only access its L2 cache parameters. In 6541 * particular, hdr->l1hdr.b_buf may be invalid by now due to 6542 * ARC eviction. 6543 */ 6544 hdr->b_l2hdr.b_daddr = dev->l2ad_hand; 6545 6546 /* 6547 * Save a pointer to the original buffer data we had previously 6548 * stashed away. 6549 */ 6550 buf_data = hdr->b_l1hdr.b_tmp_cdata; 6551 6552 compress = HDR_L2COMPRESS(hdr) && 6553 hdr->b_l2hdr.b_asize >= buf_compress_minsz; 6554 if (l2arc_transform_buf(hdr, compress)) { 6555 /* 6556 * If compression succeeded, enable headroom 6557 * boost on the next scan cycle. 6558 */ 6559 *headroom_boost = B_TRUE; 6560 } 6561 6562 /* 6563 * Get the new buffer size that accounts for compression 6564 * and padding. 6565 */ 6566 buf_sz = hdr->b_l2hdr.b_asize; 6567 6568 /* 6569 * We need to do this regardless if buf_sz is zero or 6570 * not, otherwise, when this l2hdr is evicted we'll 6571 * remove a reference that was never added. 6572 */ 6573 (void) refcount_add_many(&dev->l2ad_alloc, buf_sz, hdr); 6574 6575 /* Compression may have squashed the buffer to zero length. */ 6576 if (buf_sz != 0) { 6577 /* 6578 * If the data was padded or compressed, then it 6579 * it is in a new buffer. 6580 */ 6581 if (hdr->b_l1hdr.b_tmp_cdata != NULL) 6582 buf_data = hdr->b_l1hdr.b_tmp_cdata; 6583 wzio = zio_write_phys(pio, dev->l2ad_vdev, 6584 dev->l2ad_hand, buf_sz, buf_data, ZIO_CHECKSUM_OFF, 6585 NULL, NULL, ZIO_PRIORITY_ASYNC_WRITE, 6586 ZIO_FLAG_CANFAIL, B_FALSE); 6587 6588 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev, 6589 zio_t *, wzio); 6590 (void) zio_nowait(wzio); 6591 6592 write_asize += buf_sz; 6593 dev->l2ad_hand += buf_sz; 6594 } 6595 } 6596 6597 mutex_exit(&dev->l2ad_mtx); 6598 6599 ASSERT3U(write_asize, <=, target_sz); 6600 ARCSTAT_BUMP(arcstat_l2_writes_sent); 6601 ARCSTAT_INCR(arcstat_l2_write_bytes, write_asize); 6602 ARCSTAT_INCR(arcstat_l2_size, write_sz); 6603 ARCSTAT_INCR(arcstat_l2_asize, write_asize); 6604 vdev_space_update(dev->l2ad_vdev, write_asize, 0, 0); 6605 6606 /* 6607 * Bump device hand to the device start if it is approaching the end. 6608 * l2arc_evict() will already have evicted ahead for this case. 6609 */ 6610 if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) { 6611 dev->l2ad_hand = dev->l2ad_start; 6612 dev->l2ad_first = B_FALSE; 6613 } 6614 6615 dev->l2ad_writing = B_TRUE; 6616 (void) zio_wait(pio); 6617 dev->l2ad_writing = B_FALSE; 6618 6619 return (write_asize); 6620} 6621 6622/* 6623 * Transforms, possibly compresses and pads, an L2ARC buffer. 6624 * The data to be compressed must be prefilled in l1hdr.b_tmp_cdata and its 6625 * size in l2hdr->b_asize. This routine tries to compress the data and 6626 * depending on the compression result there are three possible outcomes: 6627 * *) The buffer was incompressible. The buffer size was already ashift aligned. 6628 * The original hdr contents were left untouched except for b_tmp_cdata, 6629 * which is reset to NULL. The caller must keep a pointer to the original 6630 * data. 6631 * *) The buffer was incompressible. The buffer size was not ashift aligned. 6632 * b_tmp_cdata was replaced with a temporary data buffer which holds a padded 6633 * (aligned) copy of the data. Once writing is done, invoke 6634 * l2arc_release_cdata_buf on this hdr to free the temporary buffer. 6635 * *) The buffer was all-zeros, so there is no need to write it to an L2 6636 * device. To indicate this situation b_tmp_cdata is NULL'ed, b_asize is 6637 * set to zero and b_compress is set to ZIO_COMPRESS_EMPTY. 6638 * *) Compression succeeded and b_tmp_cdata was replaced with a temporary 6639 * data buffer which holds the compressed data to be written, and b_asize 6640 * tells us how much data there is. b_compress is set to the appropriate 6641 * compression algorithm. Once writing is done, invoke 6642 * l2arc_release_cdata_buf on this l2hdr to free this temporary buffer. 6643 * 6644 * Returns B_TRUE if compression succeeded, or B_FALSE if it didn't (the 6645 * buffer was incompressible). 6646 */ 6647static boolean_t 6648l2arc_transform_buf(arc_buf_hdr_t *hdr, boolean_t compress) 6649{ 6650 void *cdata; 6651 size_t align, asize, csize, len, rounded; 6652 6653 ASSERT(HDR_HAS_L2HDR(hdr)); 6654 l2arc_buf_hdr_t *l2hdr = &hdr->b_l2hdr; 6655 6656 ASSERT(HDR_HAS_L1HDR(hdr)); 6657 ASSERT3S(l2hdr->b_compress, ==, ZIO_COMPRESS_OFF); 6658 ASSERT(hdr->b_l1hdr.b_tmp_cdata != NULL); 6659 6660 len = l2hdr->b_asize; 6661 align = (size_t)1 << l2hdr->b_dev->l2ad_vdev->vdev_ashift; 6662 asize = P2ROUNDUP(len, align); 6663 cdata = zio_data_buf_alloc(asize); 6664 ASSERT3P(cdata, !=, NULL); 6665 if (compress) 6666 csize = zio_compress_data(ZIO_COMPRESS_LZ4, 6667 hdr->b_l1hdr.b_tmp_cdata, cdata, len); 6668 else 6669 csize = len; 6670 6671 if (csize == 0) { 6672 /* zero block, indicate that there's nothing to write */ 6673 zio_data_buf_free(cdata, asize); 6674 l2hdr->b_compress = ZIO_COMPRESS_EMPTY; 6675 l2hdr->b_asize = 0; 6676 hdr->b_l1hdr.b_tmp_cdata = NULL; 6677 ARCSTAT_BUMP(arcstat_l2_compress_zeros); 6678 return (B_TRUE); 6679 } 6680 6681 rounded = P2ROUNDUP(csize, align); 6682 ASSERT3U(rounded, <=, asize); 6683 if (rounded < len) { 6684 /* 6685 * Compression succeeded, we'll keep the cdata around for 6686 * writing and release it afterwards. 6687 */ 6688 if (rounded > csize) { 6689 bzero((char *)cdata + csize, rounded - csize); 6690 csize = rounded; 6691 } 6692 l2hdr->b_compress = ZIO_COMPRESS_LZ4; 6693 l2hdr->b_asize = csize; 6694 hdr->b_l1hdr.b_tmp_cdata = cdata; 6695 ARCSTAT_BUMP(arcstat_l2_compress_successes); 6696 return (B_TRUE); 6697 } else { 6698 /* 6699 * Compression did not save space. 6700 */ 6701 if (P2PHASE(len, align) != 0) { 6702 /* 6703 * Use compression buffer for a copy of data padded to 6704 * the proper size. Compression algorithm remains set 6705 * to ZIO_COMPRESS_OFF. 6706 */ 6707 ASSERT3U(len, <, asize); 6708 bcopy(hdr->b_l1hdr.b_tmp_cdata, cdata, len); 6709 bzero((char *)cdata + len, asize - len); 6710 l2hdr->b_asize = asize; 6711 hdr->b_l1hdr.b_tmp_cdata = cdata; 6712 ARCSTAT_BUMP(arcstat_l2_padding_needed); 6713 } else { 6714 ASSERT3U(len, ==, asize); 6715 /* 6716 * The original buffer is good as is, 6717 * release the compressed buffer. 6718 * l2hdr will be left unmodified except for b_tmp_cdata. 6719 */ 6720 zio_data_buf_free(cdata, asize); 6721 hdr->b_l1hdr.b_tmp_cdata = NULL; 6722 } 6723 if (compress) 6724 ARCSTAT_BUMP(arcstat_l2_compress_failures); 6725 return (B_FALSE); 6726 } 6727} 6728 6729/* 6730 * Decompresses a zio read back from an l2arc device. On success, the 6731 * underlying zio's io_data buffer is overwritten by the uncompressed 6732 * version. On decompression error (corrupt compressed stream), the 6733 * zio->io_error value is set to signal an I/O error. 6734 * 6735 * Please note that the compressed data stream is not checksummed, so 6736 * if the underlying device is experiencing data corruption, we may feed 6737 * corrupt data to the decompressor, so the decompressor needs to be 6738 * able to handle this situation (LZ4 does). 6739 */ 6740static void 6741l2arc_decompress_zio(zio_t *zio, arc_buf_hdr_t *hdr, enum zio_compress c) 6742{ 6743 ASSERT(L2ARC_IS_VALID_COMPRESS(c)); 6744 6745 if (zio->io_error != 0) { 6746 /* 6747 * An io error has occured, just restore the original io 6748 * size in preparation for a main pool read. 6749 */ 6750 zio->io_orig_size = zio->io_size = hdr->b_size; 6751 return; 6752 } 6753 6754 if (c == ZIO_COMPRESS_EMPTY) { 6755 /* 6756 * An empty buffer results in a null zio, which means we 6757 * need to fill its io_data after we're done restoring the 6758 * buffer's contents. 6759 */ 6760 ASSERT(hdr->b_l1hdr.b_buf != NULL); 6761 bzero(hdr->b_l1hdr.b_buf->b_data, hdr->b_size); 6762 zio->io_data = zio->io_orig_data = hdr->b_l1hdr.b_buf->b_data; 6763 } else { 6764 ASSERT(zio->io_data != NULL); 6765 /* 6766 * We copy the compressed data from the start of the arc buffer 6767 * (the zio_read will have pulled in only what we need, the 6768 * rest is garbage which we will overwrite at decompression) 6769 * and then decompress back to the ARC data buffer. This way we 6770 * can minimize copying by simply decompressing back over the 6771 * original compressed data (rather than decompressing to an 6772 * aux buffer and then copying back the uncompressed buffer, 6773 * which is likely to be much larger). 6774 */ 6775 uint64_t csize; 6776 void *cdata; 6777 6778 csize = zio->io_size; 6779 cdata = zio_data_buf_alloc(csize); 6780 bcopy(zio->io_data, cdata, csize); 6781 if (zio_decompress_data(c, cdata, zio->io_data, csize, 6782 hdr->b_size) != 0) 6783 zio->io_error = EIO; 6784 zio_data_buf_free(cdata, csize); 6785 } 6786 6787 /* Restore the expected uncompressed IO size. */ 6788 zio->io_orig_size = zio->io_size = hdr->b_size; 6789} 6790 6791/* 6792 * Releases the temporary b_tmp_cdata buffer in an l2arc header structure. 6793 * This buffer serves as a temporary holder of compressed or padded data while 6794 * the buffer entry is being written to an l2arc device. Once that is 6795 * done, we can dispose of it. 6796 */ 6797static void 6798l2arc_release_cdata_buf(arc_buf_hdr_t *hdr) 6799{ 6800 size_t align, asize, len; 6801 enum zio_compress comp = hdr->b_l2hdr.b_compress; 6802 6803 ASSERT(HDR_HAS_L2HDR(hdr)); 6804 ASSERT(HDR_HAS_L1HDR(hdr)); 6805 ASSERT(comp == ZIO_COMPRESS_OFF || L2ARC_IS_VALID_COMPRESS(comp)); 6806 6807 if (hdr->b_l1hdr.b_tmp_cdata != NULL) { 6808 ASSERT(comp != ZIO_COMPRESS_EMPTY); 6809 len = hdr->b_size; 6810 align = (size_t)1 << hdr->b_l2hdr.b_dev->l2ad_vdev->vdev_ashift; 6811 asize = P2ROUNDUP(len, align); 6812 zio_data_buf_free(hdr->b_l1hdr.b_tmp_cdata, asize); 6813 hdr->b_l1hdr.b_tmp_cdata = NULL; 6814 } else { 6815 ASSERT(comp == ZIO_COMPRESS_OFF || comp == ZIO_COMPRESS_EMPTY); 6816 } 6817} 6818 6819/* 6820 * This thread feeds the L2ARC at regular intervals. This is the beating 6821 * heart of the L2ARC. 6822 */ 6823static void 6824l2arc_feed_thread(void *dummy __unused) 6825{ 6826 callb_cpr_t cpr; 6827 l2arc_dev_t *dev; 6828 spa_t *spa; 6829 uint64_t size, wrote; 6830 clock_t begin, next = ddi_get_lbolt(); 6831 boolean_t headroom_boost = B_FALSE; 6832 6833 CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG); 6834 6835 mutex_enter(&l2arc_feed_thr_lock); 6836 6837 while (l2arc_thread_exit == 0) { 6838 CALLB_CPR_SAFE_BEGIN(&cpr); 6839 (void) cv_timedwait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock, 6840 next - ddi_get_lbolt()); 6841 CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock); 6842 next = ddi_get_lbolt() + hz; 6843 6844 /* 6845 * Quick check for L2ARC devices. 6846 */ 6847 mutex_enter(&l2arc_dev_mtx); 6848 if (l2arc_ndev == 0) { 6849 mutex_exit(&l2arc_dev_mtx); 6850 continue; 6851 } 6852 mutex_exit(&l2arc_dev_mtx); 6853 begin = ddi_get_lbolt(); 6854 6855 /* 6856 * This selects the next l2arc device to write to, and in 6857 * doing so the next spa to feed from: dev->l2ad_spa. This 6858 * will return NULL if there are now no l2arc devices or if 6859 * they are all faulted. 6860 * 6861 * If a device is returned, its spa's config lock is also 6862 * held to prevent device removal. l2arc_dev_get_next() 6863 * will grab and release l2arc_dev_mtx. 6864 */ 6865 if ((dev = l2arc_dev_get_next()) == NULL) 6866 continue; 6867 6868 spa = dev->l2ad_spa; 6869 ASSERT(spa != NULL); 6870 6871 /* 6872 * If the pool is read-only then force the feed thread to 6873 * sleep a little longer. 6874 */ 6875 if (!spa_writeable(spa)) { 6876 next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz; 6877 spa_config_exit(spa, SCL_L2ARC, dev); 6878 continue; 6879 } 6880 6881 /* 6882 * Avoid contributing to memory pressure. 6883 */ 6884 if (arc_reclaim_needed()) { 6885 ARCSTAT_BUMP(arcstat_l2_abort_lowmem); 6886 spa_config_exit(spa, SCL_L2ARC, dev); 6887 continue; 6888 } 6889 6890 ARCSTAT_BUMP(arcstat_l2_feeds); 6891 6892 size = l2arc_write_size(); 6893 6894 /* 6895 * Evict L2ARC buffers that will be overwritten. 6896 */ 6897 l2arc_evict(dev, size, B_FALSE); 6898 6899 /* 6900 * Write ARC buffers. 6901 */ 6902 wrote = l2arc_write_buffers(spa, dev, size, &headroom_boost); 6903 6904 /* 6905 * Calculate interval between writes. 6906 */ 6907 next = l2arc_write_interval(begin, size, wrote); 6908 spa_config_exit(spa, SCL_L2ARC, dev); 6909 } 6910 6911 l2arc_thread_exit = 0; 6912 cv_broadcast(&l2arc_feed_thr_cv); 6913 CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */ 6914 thread_exit(); 6915} 6916 6917boolean_t 6918l2arc_vdev_present(vdev_t *vd) 6919{ 6920 l2arc_dev_t *dev; 6921 6922 mutex_enter(&l2arc_dev_mtx); 6923 for (dev = list_head(l2arc_dev_list); dev != NULL; 6924 dev = list_next(l2arc_dev_list, dev)) { 6925 if (dev->l2ad_vdev == vd) 6926 break; 6927 } 6928 mutex_exit(&l2arc_dev_mtx); 6929 6930 return (dev != NULL); 6931} 6932 6933/* 6934 * Add a vdev for use by the L2ARC. By this point the spa has already 6935 * validated the vdev and opened it. 6936 */ 6937void 6938l2arc_add_vdev(spa_t *spa, vdev_t *vd) 6939{ 6940 l2arc_dev_t *adddev; 6941 6942 ASSERT(!l2arc_vdev_present(vd)); 6943 6944 vdev_ashift_optimize(vd); 6945 6946 /* 6947 * Create a new l2arc device entry. 6948 */ 6949 adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP); 6950 adddev->l2ad_spa = spa; 6951 adddev->l2ad_vdev = vd; 6952 adddev->l2ad_start = VDEV_LABEL_START_SIZE; 6953 adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd); 6954 adddev->l2ad_hand = adddev->l2ad_start; 6955 adddev->l2ad_first = B_TRUE; 6956 adddev->l2ad_writing = B_FALSE; 6957 6958 mutex_init(&adddev->l2ad_mtx, NULL, MUTEX_DEFAULT, NULL); 6959 /* 6960 * This is a list of all ARC buffers that are still valid on the 6961 * device. 6962 */ 6963 list_create(&adddev->l2ad_buflist, sizeof (arc_buf_hdr_t), 6964 offsetof(arc_buf_hdr_t, b_l2hdr.b_l2node)); 6965 6966 vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand); 6967 refcount_create(&adddev->l2ad_alloc); 6968 6969 /* 6970 * Add device to global list 6971 */ 6972 mutex_enter(&l2arc_dev_mtx); 6973 list_insert_head(l2arc_dev_list, adddev); 6974 atomic_inc_64(&l2arc_ndev); 6975 mutex_exit(&l2arc_dev_mtx); 6976} 6977 6978/* 6979 * Remove a vdev from the L2ARC. 6980 */ 6981void 6982l2arc_remove_vdev(vdev_t *vd) 6983{ 6984 l2arc_dev_t *dev, *nextdev, *remdev = NULL; 6985 6986 /* 6987 * Find the device by vdev 6988 */ 6989 mutex_enter(&l2arc_dev_mtx); 6990 for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) { 6991 nextdev = list_next(l2arc_dev_list, dev); 6992 if (vd == dev->l2ad_vdev) { 6993 remdev = dev; 6994 break; 6995 } 6996 } 6997 ASSERT(remdev != NULL); 6998 6999 /* 7000 * Remove device from global list 7001 */ 7002 list_remove(l2arc_dev_list, remdev); 7003 l2arc_dev_last = NULL; /* may have been invalidated */ 7004 atomic_dec_64(&l2arc_ndev); 7005 mutex_exit(&l2arc_dev_mtx); 7006 7007 /* 7008 * Clear all buflists and ARC references. L2ARC device flush. 7009 */ 7010 l2arc_evict(remdev, 0, B_TRUE); 7011 list_destroy(&remdev->l2ad_buflist); 7012 mutex_destroy(&remdev->l2ad_mtx); 7013 refcount_destroy(&remdev->l2ad_alloc); 7014 kmem_free(remdev, sizeof (l2arc_dev_t)); 7015} 7016 7017void 7018l2arc_init(void) 7019{ 7020 l2arc_thread_exit = 0; 7021 l2arc_ndev = 0; 7022 l2arc_writes_sent = 0; 7023 l2arc_writes_done = 0; 7024 7025 mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL); 7026 cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL); 7027 mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL); 7028 mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL); 7029 7030 l2arc_dev_list = &L2ARC_dev_list; 7031 l2arc_free_on_write = &L2ARC_free_on_write; 7032 list_create(l2arc_dev_list, sizeof (l2arc_dev_t), 7033 offsetof(l2arc_dev_t, l2ad_node)); 7034 list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t), 7035 offsetof(l2arc_data_free_t, l2df_list_node)); 7036} 7037 7038void 7039l2arc_fini(void) 7040{ 7041 /* 7042 * This is called from dmu_fini(), which is called from spa_fini(); 7043 * Because of this, we can assume that all l2arc devices have 7044 * already been removed when the pools themselves were removed. 7045 */ 7046 7047 l2arc_do_free_on_write(); 7048 7049 mutex_destroy(&l2arc_feed_thr_lock); 7050 cv_destroy(&l2arc_feed_thr_cv); 7051 mutex_destroy(&l2arc_dev_mtx); 7052 mutex_destroy(&l2arc_free_on_write_mtx); 7053 7054 list_destroy(l2arc_dev_list); 7055 list_destroy(l2arc_free_on_write); 7056} 7057 7058void 7059l2arc_start(void) 7060{ 7061 if (!(spa_mode_global & FWRITE)) 7062 return; 7063 7064 (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0, 7065 TS_RUN, minclsyspri); 7066} 7067 7068void 7069l2arc_stop(void) 7070{ 7071 if (!(spa_mode_global & FWRITE)) 7072 return; 7073 7074 mutex_enter(&l2arc_feed_thr_lock); 7075 cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */ 7076 l2arc_thread_exit = 1; 7077 while (l2arc_thread_exit != 0) 7078 cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock); 7079 mutex_exit(&l2arc_feed_thr_lock); 7080} 7081