arc.c revision 286655
1/* 2 * CDDL HEADER START 3 * 4 * The contents of this file are subject to the terms of the 5 * Common Development and Distribution License (the "License"). 6 * You may not use this file except in compliance with the License. 7 * 8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE 9 * or http://www.opensolaris.org/os/licensing. 10 * See the License for the specific language governing permissions 11 * and limitations under the License. 12 * 13 * When distributing Covered Code, include this CDDL HEADER in each 14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE. 15 * If applicable, add the following below this CDDL HEADER, with the 16 * fields enclosed by brackets "[]" replaced with your own identifying 17 * information: Portions Copyright [yyyy] [name of copyright owner] 18 * 19 * CDDL HEADER END 20 */ 21/* 22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved. 23 * Copyright (c) 2012, Joyent, Inc. All rights reserved. 24 * Copyright (c) 2011, 2014 by Delphix. All rights reserved. 25 * Copyright (c) 2014 by Saso Kiselkov. All rights reserved. 26 * Copyright 2014 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 mutexs, rather they rely on the 86 * hash table mutexs for the bulk of their protection (i.e. most 87 * fields in the arc_buf_hdr_t are protected by these mutexs). 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#ifdef _KERNEL 133#include <sys/dnlc.h> 134#endif 135#include <sys/callb.h> 136#include <sys/kstat.h> 137#include <sys/trim_map.h> 138#include <zfs_fletcher.h> 139#include <sys/sdt.h> 140 141#include <vm/vm_pageout.h> 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_thr_lock; 153static kcondvar_t arc_reclaim_thr_cv; /* used to signal reclaim thr */ 154static uint8_t arc_thread_exit; 155 156uint_t arc_reduce_dnlc_percent = 3; 157 158/* 159 * The number of iterations through arc_evict_*() before we 160 * drop & reacquire the lock. 161 */ 162int arc_evict_iterations = 100; 163 164/* number of seconds before growing cache again */ 165static int arc_grow_retry = 60; 166 167/* shift of arc_c for calculating both min and max arc_p */ 168static int arc_p_min_shift = 4; 169 170/* log2(fraction of arc to reclaim) */ 171static int arc_shrink_shift = 7; 172 173/* 174 * log2(fraction of ARC which must be free to allow growing). 175 * I.e. If there is less than arc_c >> arc_no_grow_shift free memory, 176 * when reading a new block into the ARC, we will evict an equal-sized block 177 * from the ARC. 178 * 179 * This must be less than arc_shrink_shift, so that when we shrink the ARC, 180 * we will still not allow it to grow. 181 */ 182int arc_no_grow_shift = 5; 183 184 185/* 186 * minimum lifespan of a prefetch block in clock ticks 187 * (initialized in arc_init()) 188 */ 189static int arc_min_prefetch_lifespan; 190 191/* 192 * If this percent of memory is free, don't throttle. 193 */ 194int arc_lotsfree_percent = 10; 195 196static int arc_dead; 197extern int zfs_prefetch_disable; 198 199/* 200 * The arc has filled available memory and has now warmed up. 201 */ 202static boolean_t arc_warm; 203 204uint64_t zfs_arc_max; 205uint64_t zfs_arc_min; 206uint64_t zfs_arc_meta_limit = 0; 207uint64_t zfs_arc_meta_min = 0; 208int zfs_arc_grow_retry = 0; 209int zfs_arc_shrink_shift = 0; 210int zfs_arc_p_min_shift = 0; 211int zfs_disable_dup_eviction = 0; 212uint64_t zfs_arc_average_blocksize = 8 * 1024; /* 8KB */ 213u_int zfs_arc_free_target = 0; 214 215static int sysctl_vfs_zfs_arc_free_target(SYSCTL_HANDLER_ARGS); 216static int sysctl_vfs_zfs_arc_meta_limit(SYSCTL_HANDLER_ARGS); 217 218#ifdef _KERNEL 219static void 220arc_free_target_init(void *unused __unused) 221{ 222 223 zfs_arc_free_target = vm_pageout_wakeup_thresh; 224} 225SYSINIT(arc_free_target_init, SI_SUB_KTHREAD_PAGE, SI_ORDER_ANY, 226 arc_free_target_init, NULL); 227 228TUNABLE_QUAD("vfs.zfs.arc_meta_limit", &zfs_arc_meta_limit); 229TUNABLE_QUAD("vfs.zfs.arc_meta_min", &zfs_arc_meta_min); 230TUNABLE_INT("vfs.zfs.arc_shrink_shift", &zfs_arc_shrink_shift); 231SYSCTL_DECL(_vfs_zfs); 232SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_max, CTLFLAG_RDTUN, &zfs_arc_max, 0, 233 "Maximum ARC size"); 234SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_min, CTLFLAG_RDTUN, &zfs_arc_min, 0, 235 "Minimum ARC size"); 236SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_average_blocksize, CTLFLAG_RDTUN, 237 &zfs_arc_average_blocksize, 0, 238 "ARC average blocksize"); 239SYSCTL_INT(_vfs_zfs, OID_AUTO, arc_shrink_shift, CTLFLAG_RW, 240 &arc_shrink_shift, 0, 241 "log2(fraction of arc to reclaim)"); 242 243/* 244 * We don't have a tunable for arc_free_target due to the dependency on 245 * pagedaemon initialisation. 246 */ 247SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_free_target, 248 CTLTYPE_UINT | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, sizeof(u_int), 249 sysctl_vfs_zfs_arc_free_target, "IU", 250 "Desired number of free pages below which ARC triggers reclaim"); 251 252static int 253sysctl_vfs_zfs_arc_free_target(SYSCTL_HANDLER_ARGS) 254{ 255 u_int val; 256 int err; 257 258 val = zfs_arc_free_target; 259 err = sysctl_handle_int(oidp, &val, 0, req); 260 if (err != 0 || req->newptr == NULL) 261 return (err); 262 263 if (val < minfree) 264 return (EINVAL); 265 if (val > vm_cnt.v_page_count) 266 return (EINVAL); 267 268 zfs_arc_free_target = val; 269 270 return (0); 271} 272 273/* 274 * Must be declared here, before the definition of corresponding kstat 275 * macro which uses the same names will confuse the compiler. 276 */ 277SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_meta_limit, 278 CTLTYPE_U64 | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, sizeof(uint64_t), 279 sysctl_vfs_zfs_arc_meta_limit, "QU", 280 "ARC metadata limit"); 281#endif 282 283/* 284 * Note that buffers can be in one of 6 states: 285 * ARC_anon - anonymous (discussed below) 286 * ARC_mru - recently used, currently cached 287 * ARC_mru_ghost - recentely used, no longer in cache 288 * ARC_mfu - frequently used, currently cached 289 * ARC_mfu_ghost - frequently used, no longer in cache 290 * ARC_l2c_only - exists in L2ARC but not other states 291 * When there are no active references to the buffer, they are 292 * are linked onto a list in one of these arc states. These are 293 * the only buffers that can be evicted or deleted. Within each 294 * state there are multiple lists, one for meta-data and one for 295 * non-meta-data. Meta-data (indirect blocks, blocks of dnodes, 296 * etc.) is tracked separately so that it can be managed more 297 * explicitly: favored over data, limited explicitly. 298 * 299 * Anonymous buffers are buffers that are not associated with 300 * a DVA. These are buffers that hold dirty block copies 301 * before they are written to stable storage. By definition, 302 * they are "ref'd" and are considered part of arc_mru 303 * that cannot be freed. Generally, they will aquire a DVA 304 * as they are written and migrate onto the arc_mru list. 305 * 306 * The ARC_l2c_only state is for buffers that are in the second 307 * level ARC but no longer in any of the ARC_m* lists. The second 308 * level ARC itself may also contain buffers that are in any of 309 * the ARC_m* states - meaning that a buffer can exist in two 310 * places. The reason for the ARC_l2c_only state is to keep the 311 * buffer header in the hash table, so that reads that hit the 312 * second level ARC benefit from these fast lookups. 313 */ 314 315#define ARCS_LOCK_PAD CACHE_LINE_SIZE 316struct arcs_lock { 317 kmutex_t arcs_lock; 318#ifdef _KERNEL 319 unsigned char pad[(ARCS_LOCK_PAD - sizeof (kmutex_t))]; 320#endif 321}; 322 323/* 324 * must be power of two for mask use to work 325 * 326 */ 327#define ARC_BUFC_NUMDATALISTS 16 328#define ARC_BUFC_NUMMETADATALISTS 16 329#define ARC_BUFC_NUMLISTS (ARC_BUFC_NUMMETADATALISTS + ARC_BUFC_NUMDATALISTS) 330 331typedef struct arc_state { 332 uint64_t arcs_lsize[ARC_BUFC_NUMTYPES]; /* amount of evictable data */ 333 uint64_t arcs_size; /* total amount of data in this state */ 334 list_t arcs_lists[ARC_BUFC_NUMLISTS]; /* list of evictable buffers */ 335 struct arcs_lock arcs_locks[ARC_BUFC_NUMLISTS] __aligned(CACHE_LINE_SIZE); 336} arc_state_t; 337 338#define ARCS_LOCK(s, i) (&((s)->arcs_locks[(i)].arcs_lock)) 339 340/* The 6 states: */ 341static arc_state_t ARC_anon; 342static arc_state_t ARC_mru; 343static arc_state_t ARC_mru_ghost; 344static arc_state_t ARC_mfu; 345static arc_state_t ARC_mfu_ghost; 346static arc_state_t ARC_l2c_only; 347 348typedef struct arc_stats { 349 kstat_named_t arcstat_hits; 350 kstat_named_t arcstat_misses; 351 kstat_named_t arcstat_demand_data_hits; 352 kstat_named_t arcstat_demand_data_misses; 353 kstat_named_t arcstat_demand_metadata_hits; 354 kstat_named_t arcstat_demand_metadata_misses; 355 kstat_named_t arcstat_prefetch_data_hits; 356 kstat_named_t arcstat_prefetch_data_misses; 357 kstat_named_t arcstat_prefetch_metadata_hits; 358 kstat_named_t arcstat_prefetch_metadata_misses; 359 kstat_named_t arcstat_mru_hits; 360 kstat_named_t arcstat_mru_ghost_hits; 361 kstat_named_t arcstat_mfu_hits; 362 kstat_named_t arcstat_mfu_ghost_hits; 363 kstat_named_t arcstat_allocated; 364 kstat_named_t arcstat_deleted; 365 kstat_named_t arcstat_stolen; 366 kstat_named_t arcstat_recycle_miss; 367 /* 368 * Number of buffers that could not be evicted because the hash lock 369 * was held by another thread. The lock may not necessarily be held 370 * by something using the same buffer, since hash locks are shared 371 * by multiple buffers. 372 */ 373 kstat_named_t arcstat_mutex_miss; 374 /* 375 * Number of buffers skipped because they have I/O in progress, are 376 * indrect prefetch buffers that have not lived long enough, or are 377 * not from the spa we're trying to evict from. 378 */ 379 kstat_named_t arcstat_evict_skip; 380 kstat_named_t arcstat_evict_l2_cached; 381 kstat_named_t arcstat_evict_l2_eligible; 382 kstat_named_t arcstat_evict_l2_ineligible; 383 kstat_named_t arcstat_hash_elements; 384 kstat_named_t arcstat_hash_elements_max; 385 kstat_named_t arcstat_hash_collisions; 386 kstat_named_t arcstat_hash_chains; 387 kstat_named_t arcstat_hash_chain_max; 388 kstat_named_t arcstat_p; 389 kstat_named_t arcstat_c; 390 kstat_named_t arcstat_c_min; 391 kstat_named_t arcstat_c_max; 392 kstat_named_t arcstat_size; 393 /* 394 * Number of bytes consumed by internal ARC structures necessary 395 * for tracking purposes; these structures are not actually 396 * backed by ARC buffers. This includes arc_buf_hdr_t structures 397 * (allocated via arc_buf_hdr_t_full and arc_buf_hdr_t_l2only 398 * caches), and arc_buf_t structures (allocated via arc_buf_t 399 * cache). 400 */ 401 kstat_named_t arcstat_hdr_size; 402 /* 403 * Number of bytes consumed by ARC buffers of type equal to 404 * ARC_BUFC_DATA. This is generally consumed by buffers backing 405 * on disk user data (e.g. plain file contents). 406 */ 407 kstat_named_t arcstat_data_size; 408 /* 409 * Number of bytes consumed by ARC buffers of type equal to 410 * ARC_BUFC_METADATA. This is generally consumed by buffers 411 * backing on disk data that is used for internal ZFS 412 * structures (e.g. ZAP, dnode, indirect blocks, etc). 413 */ 414 kstat_named_t arcstat_metadata_size; 415 /* 416 * Number of bytes consumed by various buffers and structures 417 * not actually backed with ARC buffers. This includes bonus 418 * buffers (allocated directly via zio_buf_* functions), 419 * dmu_buf_impl_t structures (allocated via dmu_buf_impl_t 420 * cache), and dnode_t structures (allocated via dnode_t cache). 421 */ 422 kstat_named_t arcstat_other_size; 423 /* 424 * Total number of bytes consumed by ARC buffers residing in the 425 * arc_anon state. This includes *all* buffers in the arc_anon 426 * state; e.g. data, metadata, evictable, and unevictable buffers 427 * are all included in this value. 428 */ 429 kstat_named_t arcstat_anon_size; 430 /* 431 * Number of bytes consumed by ARC buffers that meet the 432 * following criteria: backing buffers of type ARC_BUFC_DATA, 433 * residing in the arc_anon state, and are eligible for eviction 434 * (e.g. have no outstanding holds on the buffer). 435 */ 436 kstat_named_t arcstat_anon_evictable_data; 437 /* 438 * Number of bytes consumed by ARC buffers that meet the 439 * following criteria: backing buffers of type ARC_BUFC_METADATA, 440 * residing in the arc_anon state, and are eligible for eviction 441 * (e.g. have no outstanding holds on the buffer). 442 */ 443 kstat_named_t arcstat_anon_evictable_metadata; 444 /* 445 * Total number of bytes consumed by ARC buffers residing in the 446 * arc_mru state. This includes *all* buffers in the arc_mru 447 * state; e.g. data, metadata, evictable, and unevictable buffers 448 * are all included in this value. 449 */ 450 kstat_named_t arcstat_mru_size; 451 /* 452 * Number of bytes consumed by ARC buffers that meet the 453 * following criteria: backing buffers of type ARC_BUFC_DATA, 454 * residing in the arc_mru state, and are eligible for eviction 455 * (e.g. have no outstanding holds on the buffer). 456 */ 457 kstat_named_t arcstat_mru_evictable_data; 458 /* 459 * Number of bytes consumed by ARC buffers that meet the 460 * following criteria: backing buffers of type ARC_BUFC_METADATA, 461 * residing in the arc_mru state, and are eligible for eviction 462 * (e.g. have no outstanding holds on the buffer). 463 */ 464 kstat_named_t arcstat_mru_evictable_metadata; 465 /* 466 * Total number of bytes that *would have been* consumed by ARC 467 * buffers in the arc_mru_ghost state. The key thing to note 468 * here, is the fact that this size doesn't actually indicate 469 * RAM consumption. The ghost lists only consist of headers and 470 * don't actually have ARC buffers linked off of these headers. 471 * Thus, *if* the headers had associated ARC buffers, these 472 * buffers *would have* consumed this number of bytes. 473 */ 474 kstat_named_t arcstat_mru_ghost_size; 475 /* 476 * Number of bytes that *would have been* consumed by ARC 477 * buffers that are eligible for eviction, of type 478 * ARC_BUFC_DATA, and linked off the arc_mru_ghost state. 479 */ 480 kstat_named_t arcstat_mru_ghost_evictable_data; 481 /* 482 * Number of bytes that *would have been* consumed by ARC 483 * buffers that are eligible for eviction, of type 484 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state. 485 */ 486 kstat_named_t arcstat_mru_ghost_evictable_metadata; 487 /* 488 * Total number of bytes consumed by ARC buffers residing in the 489 * arc_mfu state. This includes *all* buffers in the arc_mfu 490 * state; e.g. data, metadata, evictable, and unevictable buffers 491 * are all included in this value. 492 */ 493 kstat_named_t arcstat_mfu_size; 494 /* 495 * Number of bytes consumed by ARC buffers that are eligible for 496 * eviction, of type ARC_BUFC_DATA, and reside in the arc_mfu 497 * state. 498 */ 499 kstat_named_t arcstat_mfu_evictable_data; 500 /* 501 * Number of bytes consumed by ARC buffers that are eligible for 502 * eviction, of type ARC_BUFC_METADATA, and reside in the 503 * arc_mfu state. 504 */ 505 kstat_named_t arcstat_mfu_evictable_metadata; 506 /* 507 * Total number of bytes that *would have been* consumed by ARC 508 * buffers in the arc_mfu_ghost state. See the comment above 509 * arcstat_mru_ghost_size for more details. 510 */ 511 kstat_named_t arcstat_mfu_ghost_size; 512 /* 513 * Number of bytes that *would have been* consumed by ARC 514 * buffers that are eligible for eviction, of type 515 * ARC_BUFC_DATA, and linked off the arc_mfu_ghost state. 516 */ 517 kstat_named_t arcstat_mfu_ghost_evictable_data; 518 /* 519 * Number of bytes that *would have been* consumed by ARC 520 * buffers that are eligible for eviction, of type 521 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state. 522 */ 523 kstat_named_t arcstat_mfu_ghost_evictable_metadata; 524 kstat_named_t arcstat_l2_hits; 525 kstat_named_t arcstat_l2_misses; 526 kstat_named_t arcstat_l2_feeds; 527 kstat_named_t arcstat_l2_rw_clash; 528 kstat_named_t arcstat_l2_read_bytes; 529 kstat_named_t arcstat_l2_write_bytes; 530 kstat_named_t arcstat_l2_writes_sent; 531 kstat_named_t arcstat_l2_writes_done; 532 kstat_named_t arcstat_l2_writes_error; 533 kstat_named_t arcstat_l2_writes_hdr_miss; 534 kstat_named_t arcstat_l2_evict_lock_retry; 535 kstat_named_t arcstat_l2_evict_reading; 536 kstat_named_t arcstat_l2_evict_l1cached; 537 kstat_named_t arcstat_l2_free_on_write; 538 kstat_named_t arcstat_l2_cdata_free_on_write; 539 kstat_named_t arcstat_l2_abort_lowmem; 540 kstat_named_t arcstat_l2_cksum_bad; 541 kstat_named_t arcstat_l2_io_error; 542 kstat_named_t arcstat_l2_size; 543 kstat_named_t arcstat_l2_asize; 544 kstat_named_t arcstat_l2_hdr_size; 545 kstat_named_t arcstat_l2_compress_successes; 546 kstat_named_t arcstat_l2_compress_zeros; 547 kstat_named_t arcstat_l2_compress_failures; 548 kstat_named_t arcstat_l2_write_trylock_fail; 549 kstat_named_t arcstat_l2_write_passed_headroom; 550 kstat_named_t arcstat_l2_write_spa_mismatch; 551 kstat_named_t arcstat_l2_write_in_l2; 552 kstat_named_t arcstat_l2_write_hdr_io_in_progress; 553 kstat_named_t arcstat_l2_write_not_cacheable; 554 kstat_named_t arcstat_l2_write_full; 555 kstat_named_t arcstat_l2_write_buffer_iter; 556 kstat_named_t arcstat_l2_write_pios; 557 kstat_named_t arcstat_l2_write_buffer_bytes_scanned; 558 kstat_named_t arcstat_l2_write_buffer_list_iter; 559 kstat_named_t arcstat_l2_write_buffer_list_null_iter; 560 kstat_named_t arcstat_memory_throttle_count; 561 kstat_named_t arcstat_duplicate_buffers; 562 kstat_named_t arcstat_duplicate_buffers_size; 563 kstat_named_t arcstat_duplicate_reads; 564 kstat_named_t arcstat_meta_used; 565 kstat_named_t arcstat_meta_limit; 566 kstat_named_t arcstat_meta_max; 567 kstat_named_t arcstat_meta_min; 568} arc_stats_t; 569 570static arc_stats_t arc_stats = { 571 { "hits", KSTAT_DATA_UINT64 }, 572 { "misses", KSTAT_DATA_UINT64 }, 573 { "demand_data_hits", KSTAT_DATA_UINT64 }, 574 { "demand_data_misses", KSTAT_DATA_UINT64 }, 575 { "demand_metadata_hits", KSTAT_DATA_UINT64 }, 576 { "demand_metadata_misses", KSTAT_DATA_UINT64 }, 577 { "prefetch_data_hits", KSTAT_DATA_UINT64 }, 578 { "prefetch_data_misses", KSTAT_DATA_UINT64 }, 579 { "prefetch_metadata_hits", KSTAT_DATA_UINT64 }, 580 { "prefetch_metadata_misses", KSTAT_DATA_UINT64 }, 581 { "mru_hits", KSTAT_DATA_UINT64 }, 582 { "mru_ghost_hits", KSTAT_DATA_UINT64 }, 583 { "mfu_hits", KSTAT_DATA_UINT64 }, 584 { "mfu_ghost_hits", KSTAT_DATA_UINT64 }, 585 { "allocated", KSTAT_DATA_UINT64 }, 586 { "deleted", KSTAT_DATA_UINT64 }, 587 { "stolen", KSTAT_DATA_UINT64 }, 588 { "recycle_miss", KSTAT_DATA_UINT64 }, 589 { "mutex_miss", KSTAT_DATA_UINT64 }, 590 { "evict_skip", KSTAT_DATA_UINT64 }, 591 { "evict_l2_cached", KSTAT_DATA_UINT64 }, 592 { "evict_l2_eligible", KSTAT_DATA_UINT64 }, 593 { "evict_l2_ineligible", KSTAT_DATA_UINT64 }, 594 { "hash_elements", KSTAT_DATA_UINT64 }, 595 { "hash_elements_max", KSTAT_DATA_UINT64 }, 596 { "hash_collisions", KSTAT_DATA_UINT64 }, 597 { "hash_chains", KSTAT_DATA_UINT64 }, 598 { "hash_chain_max", KSTAT_DATA_UINT64 }, 599 { "p", KSTAT_DATA_UINT64 }, 600 { "c", KSTAT_DATA_UINT64 }, 601 { "c_min", KSTAT_DATA_UINT64 }, 602 { "c_max", KSTAT_DATA_UINT64 }, 603 { "size", KSTAT_DATA_UINT64 }, 604 { "hdr_size", KSTAT_DATA_UINT64 }, 605 { "data_size", KSTAT_DATA_UINT64 }, 606 { "metadata_size", KSTAT_DATA_UINT64 }, 607 { "other_size", KSTAT_DATA_UINT64 }, 608 { "anon_size", KSTAT_DATA_UINT64 }, 609 { "anon_evictable_data", KSTAT_DATA_UINT64 }, 610 { "anon_evictable_metadata", KSTAT_DATA_UINT64 }, 611 { "mru_size", KSTAT_DATA_UINT64 }, 612 { "mru_evictable_data", KSTAT_DATA_UINT64 }, 613 { "mru_evictable_metadata", KSTAT_DATA_UINT64 }, 614 { "mru_ghost_size", KSTAT_DATA_UINT64 }, 615 { "mru_ghost_evictable_data", KSTAT_DATA_UINT64 }, 616 { "mru_ghost_evictable_metadata", KSTAT_DATA_UINT64 }, 617 { "mfu_size", KSTAT_DATA_UINT64 }, 618 { "mfu_evictable_data", KSTAT_DATA_UINT64 }, 619 { "mfu_evictable_metadata", KSTAT_DATA_UINT64 }, 620 { "mfu_ghost_size", KSTAT_DATA_UINT64 }, 621 { "mfu_ghost_evictable_data", KSTAT_DATA_UINT64 }, 622 { "mfu_ghost_evictable_metadata", KSTAT_DATA_UINT64 }, 623 { "l2_hits", KSTAT_DATA_UINT64 }, 624 { "l2_misses", KSTAT_DATA_UINT64 }, 625 { "l2_feeds", KSTAT_DATA_UINT64 }, 626 { "l2_rw_clash", KSTAT_DATA_UINT64 }, 627 { "l2_read_bytes", KSTAT_DATA_UINT64 }, 628 { "l2_write_bytes", KSTAT_DATA_UINT64 }, 629 { "l2_writes_sent", KSTAT_DATA_UINT64 }, 630 { "l2_writes_done", KSTAT_DATA_UINT64 }, 631 { "l2_writes_error", KSTAT_DATA_UINT64 }, 632 { "l2_writes_hdr_miss", KSTAT_DATA_UINT64 }, 633 { "l2_evict_lock_retry", KSTAT_DATA_UINT64 }, 634 { "l2_evict_reading", KSTAT_DATA_UINT64 }, 635 { "l2_evict_l1cached", KSTAT_DATA_UINT64 }, 636 { "l2_free_on_write", KSTAT_DATA_UINT64 }, 637 { "l2_cdata_free_on_write", KSTAT_DATA_UINT64 }, 638 { "l2_abort_lowmem", KSTAT_DATA_UINT64 }, 639 { "l2_cksum_bad", KSTAT_DATA_UINT64 }, 640 { "l2_io_error", KSTAT_DATA_UINT64 }, 641 { "l2_size", KSTAT_DATA_UINT64 }, 642 { "l2_asize", KSTAT_DATA_UINT64 }, 643 { "l2_hdr_size", KSTAT_DATA_UINT64 }, 644 { "l2_compress_successes", KSTAT_DATA_UINT64 }, 645 { "l2_compress_zeros", KSTAT_DATA_UINT64 }, 646 { "l2_compress_failures", KSTAT_DATA_UINT64 }, 647 { "l2_write_trylock_fail", KSTAT_DATA_UINT64 }, 648 { "l2_write_passed_headroom", KSTAT_DATA_UINT64 }, 649 { "l2_write_spa_mismatch", KSTAT_DATA_UINT64 }, 650 { "l2_write_in_l2", KSTAT_DATA_UINT64 }, 651 { "l2_write_io_in_progress", KSTAT_DATA_UINT64 }, 652 { "l2_write_not_cacheable", KSTAT_DATA_UINT64 }, 653 { "l2_write_full", KSTAT_DATA_UINT64 }, 654 { "l2_write_buffer_iter", KSTAT_DATA_UINT64 }, 655 { "l2_write_pios", KSTAT_DATA_UINT64 }, 656 { "l2_write_buffer_bytes_scanned", KSTAT_DATA_UINT64 }, 657 { "l2_write_buffer_list_iter", KSTAT_DATA_UINT64 }, 658 { "l2_write_buffer_list_null_iter", KSTAT_DATA_UINT64 }, 659 { "memory_throttle_count", KSTAT_DATA_UINT64 }, 660 { "duplicate_buffers", KSTAT_DATA_UINT64 }, 661 { "duplicate_buffers_size", KSTAT_DATA_UINT64 }, 662 { "duplicate_reads", KSTAT_DATA_UINT64 }, 663 { "arc_meta_used", KSTAT_DATA_UINT64 }, 664 { "arc_meta_limit", KSTAT_DATA_UINT64 }, 665 { "arc_meta_max", KSTAT_DATA_UINT64 }, 666 { "arc_meta_min", KSTAT_DATA_UINT64 } 667}; 668 669#define ARCSTAT(stat) (arc_stats.stat.value.ui64) 670 671#define ARCSTAT_INCR(stat, val) \ 672 atomic_add_64(&arc_stats.stat.value.ui64, (val)) 673 674#define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1) 675#define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1) 676 677#define ARCSTAT_MAX(stat, val) { \ 678 uint64_t m; \ 679 while ((val) > (m = arc_stats.stat.value.ui64) && \ 680 (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \ 681 continue; \ 682} 683 684#define ARCSTAT_MAXSTAT(stat) \ 685 ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64) 686 687/* 688 * We define a macro to allow ARC hits/misses to be easily broken down by 689 * two separate conditions, giving a total of four different subtypes for 690 * each of hits and misses (so eight statistics total). 691 */ 692#define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \ 693 if (cond1) { \ 694 if (cond2) { \ 695 ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \ 696 } else { \ 697 ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \ 698 } \ 699 } else { \ 700 if (cond2) { \ 701 ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \ 702 } else { \ 703 ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\ 704 } \ 705 } 706 707kstat_t *arc_ksp; 708static arc_state_t *arc_anon; 709static arc_state_t *arc_mru; 710static arc_state_t *arc_mru_ghost; 711static arc_state_t *arc_mfu; 712static arc_state_t *arc_mfu_ghost; 713static arc_state_t *arc_l2c_only; 714 715/* 716 * There are several ARC variables that are critical to export as kstats -- 717 * but we don't want to have to grovel around in the kstat whenever we wish to 718 * manipulate them. For these variables, we therefore define them to be in 719 * terms of the statistic variable. This assures that we are not introducing 720 * the possibility of inconsistency by having shadow copies of the variables, 721 * while still allowing the code to be readable. 722 */ 723#define arc_size ARCSTAT(arcstat_size) /* actual total arc size */ 724#define arc_p ARCSTAT(arcstat_p) /* target size of MRU */ 725#define arc_c ARCSTAT(arcstat_c) /* target size of cache */ 726#define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */ 727#define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */ 728#define arc_meta_limit ARCSTAT(arcstat_meta_limit) /* max size for metadata */ 729#define arc_meta_min ARCSTAT(arcstat_meta_min) /* min size for metadata */ 730#define arc_meta_used ARCSTAT(arcstat_meta_used) /* size of metadata */ 731#define arc_meta_max ARCSTAT(arcstat_meta_max) /* max size of metadata */ 732 733#define L2ARC_IS_VALID_COMPRESS(_c_) \ 734 ((_c_) == ZIO_COMPRESS_LZ4 || (_c_) == ZIO_COMPRESS_EMPTY) 735 736static int arc_no_grow; /* Don't try to grow cache size */ 737static uint64_t arc_tempreserve; 738static uint64_t arc_loaned_bytes; 739 740typedef struct arc_callback arc_callback_t; 741 742struct arc_callback { 743 void *acb_private; 744 arc_done_func_t *acb_done; 745 arc_buf_t *acb_buf; 746 zio_t *acb_zio_dummy; 747 arc_callback_t *acb_next; 748}; 749 750typedef struct arc_write_callback arc_write_callback_t; 751 752struct arc_write_callback { 753 void *awcb_private; 754 arc_done_func_t *awcb_ready; 755 arc_done_func_t *awcb_physdone; 756 arc_done_func_t *awcb_done; 757 arc_buf_t *awcb_buf; 758}; 759 760/* 761 * ARC buffers are separated into multiple structs as a memory saving measure: 762 * - Common fields struct, always defined, and embedded within it: 763 * - L2-only fields, always allocated but undefined when not in L2ARC 764 * - L1-only fields, only allocated when in L1ARC 765 * 766 * Buffer in L1 Buffer only in L2 767 * +------------------------+ +------------------------+ 768 * | arc_buf_hdr_t | | arc_buf_hdr_t | 769 * | | | | 770 * | | | | 771 * | | | | 772 * +------------------------+ +------------------------+ 773 * | l2arc_buf_hdr_t | | l2arc_buf_hdr_t | 774 * | (undefined if L1-only) | | | 775 * +------------------------+ +------------------------+ 776 * | l1arc_buf_hdr_t | 777 * | | 778 * | | 779 * | | 780 * | | 781 * +------------------------+ 782 * 783 * Because it's possible for the L2ARC to become extremely large, we can wind 784 * up eating a lot of memory in L2ARC buffer headers, so the size of a header 785 * is minimized by only allocating the fields necessary for an L1-cached buffer 786 * when a header is actually in the L1 cache. The sub-headers (l1arc_buf_hdr and 787 * l2arc_buf_hdr) are embedded rather than allocated separately to save a couple 788 * words in pointers. arc_hdr_realloc() is used to switch a header between 789 * these two allocation states. 790 */ 791typedef struct l1arc_buf_hdr { 792 kmutex_t b_freeze_lock; 793#ifdef ZFS_DEBUG 794 /* 795 * used for debugging wtih kmem_flags - by allocating and freeing 796 * b_thawed when the buffer is thawed, we get a record of the stack 797 * trace that thawed it. 798 */ 799 void *b_thawed; 800#endif 801 802 arc_buf_t *b_buf; 803 uint32_t b_datacnt; 804 /* for waiting on writes to complete */ 805 kcondvar_t b_cv; 806 807 /* protected by arc state mutex */ 808 arc_state_t *b_state; 809 list_node_t b_arc_node; 810 811 /* updated atomically */ 812 clock_t b_arc_access; 813 814 /* self protecting */ 815 refcount_t b_refcnt; 816 817 arc_callback_t *b_acb; 818 /* temporary buffer holder for in-flight compressed data */ 819 void *b_tmp_cdata; 820} l1arc_buf_hdr_t; 821 822typedef struct l2arc_dev l2arc_dev_t; 823 824typedef struct l2arc_buf_hdr { 825 /* protected by arc_buf_hdr mutex */ 826 l2arc_dev_t *b_dev; /* L2ARC device */ 827 uint64_t b_daddr; /* disk address, offset byte */ 828 /* real alloc'd buffer size depending on b_compress applied */ 829 int32_t b_asize; 830 831 list_node_t b_l2node; 832} l2arc_buf_hdr_t; 833 834struct arc_buf_hdr { 835 /* protected by hash lock */ 836 dva_t b_dva; 837 uint64_t b_birth; 838 /* 839 * Even though this checksum is only set/verified when a buffer is in 840 * the L1 cache, it needs to be in the set of common fields because it 841 * must be preserved from the time before a buffer is written out to 842 * L2ARC until after it is read back in. 843 */ 844 zio_cksum_t *b_freeze_cksum; 845 846 arc_buf_hdr_t *b_hash_next; 847 arc_flags_t b_flags; 848 849 /* immutable */ 850 int32_t b_size; 851 uint64_t b_spa; 852 853 /* L2ARC fields. Undefined when not in L2ARC. */ 854 l2arc_buf_hdr_t b_l2hdr; 855 /* L1ARC fields. Undefined when in l2arc_only state */ 856 l1arc_buf_hdr_t b_l1hdr; 857}; 858 859#ifdef _KERNEL 860static int 861sysctl_vfs_zfs_arc_meta_limit(SYSCTL_HANDLER_ARGS) 862{ 863 uint64_t val; 864 int err; 865 866 val = arc_meta_limit; 867 err = sysctl_handle_64(oidp, &val, 0, req); 868 if (err != 0 || req->newptr == NULL) 869 return (err); 870 871 if (val <= 0 || val > arc_c_max) 872 return (EINVAL); 873 874 arc_meta_limit = val; 875 return (0); 876} 877#endif 878 879static arc_buf_t *arc_eviction_list; 880static kmutex_t arc_eviction_mtx; 881static arc_buf_hdr_t arc_eviction_hdr; 882 883#define GHOST_STATE(state) \ 884 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \ 885 (state) == arc_l2c_only) 886 887#define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_FLAG_IN_HASH_TABLE) 888#define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) 889#define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_FLAG_IO_ERROR) 890#define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_FLAG_PREFETCH) 891#define HDR_FREED_IN_READ(hdr) ((hdr)->b_flags & ARC_FLAG_FREED_IN_READ) 892#define HDR_BUF_AVAILABLE(hdr) ((hdr)->b_flags & ARC_FLAG_BUF_AVAILABLE) 893 894#define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_FLAG_L2CACHE) 895#define HDR_L2COMPRESS(hdr) ((hdr)->b_flags & ARC_FLAG_L2COMPRESS) 896#define HDR_L2_READING(hdr) \ 897 (((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) && \ 898 ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR)) 899#define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITING) 900#define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_FLAG_L2_EVICTED) 901#define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITE_HEAD) 902 903#define HDR_ISTYPE_METADATA(hdr) \ 904 ((hdr)->b_flags & ARC_FLAG_BUFC_METADATA) 905#define HDR_ISTYPE_DATA(hdr) (!HDR_ISTYPE_METADATA(hdr)) 906 907#define HDR_HAS_L1HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L1HDR) 908#define HDR_HAS_L2HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR) 909 910/* For storing compression mode in b_flags */ 911#define HDR_COMPRESS_OFFSET 24 912#define HDR_COMPRESS_NBITS 7 913 914#define HDR_GET_COMPRESS(hdr) ((enum zio_compress)BF32_GET(hdr->b_flags, \ 915 HDR_COMPRESS_OFFSET, HDR_COMPRESS_NBITS)) 916#define HDR_SET_COMPRESS(hdr, cmp) BF32_SET(hdr->b_flags, \ 917 HDR_COMPRESS_OFFSET, HDR_COMPRESS_NBITS, (cmp)) 918 919/* 920 * Other sizes 921 */ 922 923#define HDR_FULL_SIZE ((int64_t)sizeof (arc_buf_hdr_t)) 924#define HDR_L2ONLY_SIZE ((int64_t)offsetof(arc_buf_hdr_t, b_l1hdr)) 925 926/* 927 * Hash table routines 928 */ 929 930#define HT_LOCK_PAD CACHE_LINE_SIZE 931 932struct ht_lock { 933 kmutex_t ht_lock; 934#ifdef _KERNEL 935 unsigned char pad[(HT_LOCK_PAD - sizeof (kmutex_t))]; 936#endif 937}; 938 939#define BUF_LOCKS 256 940typedef struct buf_hash_table { 941 uint64_t ht_mask; 942 arc_buf_hdr_t **ht_table; 943 struct ht_lock ht_locks[BUF_LOCKS] __aligned(CACHE_LINE_SIZE); 944} buf_hash_table_t; 945 946static buf_hash_table_t buf_hash_table; 947 948#define BUF_HASH_INDEX(spa, dva, birth) \ 949 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask) 950#define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)]) 951#define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock)) 952#define HDR_LOCK(hdr) \ 953 (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth))) 954 955uint64_t zfs_crc64_table[256]; 956 957/* 958 * Level 2 ARC 959 */ 960 961#define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */ 962#define L2ARC_HEADROOM 2 /* num of writes */ 963/* 964 * If we discover during ARC scan any buffers to be compressed, we boost 965 * our headroom for the next scanning cycle by this percentage multiple. 966 */ 967#define L2ARC_HEADROOM_BOOST 200 968#define L2ARC_FEED_SECS 1 /* caching interval secs */ 969#define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */ 970 971/* 972 * Used to distinguish headers that are being process by 973 * l2arc_write_buffers(), but have yet to be assigned to a l2arc disk 974 * address. This can happen when the header is added to the l2arc's list 975 * of buffers to write in the first stage of l2arc_write_buffers(), but 976 * has not yet been written out which happens in the second stage of 977 * l2arc_write_buffers(). 978 */ 979#define L2ARC_ADDR_UNSET ((uint64_t)(-1)) 980 981#define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent) 982#define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done) 983 984/* L2ARC Performance Tunables */ 985uint64_t l2arc_write_max = L2ARC_WRITE_SIZE; /* default max write size */ 986uint64_t l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra write during warmup */ 987uint64_t l2arc_headroom = L2ARC_HEADROOM; /* number of dev writes */ 988uint64_t l2arc_headroom_boost = L2ARC_HEADROOM_BOOST; 989uint64_t l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */ 990uint64_t l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval milliseconds */ 991boolean_t l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */ 992boolean_t l2arc_feed_again = B_TRUE; /* turbo warmup */ 993boolean_t l2arc_norw = B_TRUE; /* no reads during writes */ 994 995SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_max, CTLFLAG_RW, 996 &l2arc_write_max, 0, "max write size"); 997SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_boost, CTLFLAG_RW, 998 &l2arc_write_boost, 0, "extra write during warmup"); 999SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_headroom, CTLFLAG_RW, 1000 &l2arc_headroom, 0, "number of dev writes"); 1001SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_secs, CTLFLAG_RW, 1002 &l2arc_feed_secs, 0, "interval seconds"); 1003SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_min_ms, CTLFLAG_RW, 1004 &l2arc_feed_min_ms, 0, "min interval milliseconds"); 1005 1006SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_noprefetch, CTLFLAG_RW, 1007 &l2arc_noprefetch, 0, "don't cache prefetch bufs"); 1008SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_feed_again, CTLFLAG_RW, 1009 &l2arc_feed_again, 0, "turbo warmup"); 1010SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_norw, CTLFLAG_RW, 1011 &l2arc_norw, 0, "no reads during writes"); 1012 1013SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_size, CTLFLAG_RD, 1014 &ARC_anon.arcs_size, 0, "size of anonymous state"); 1015SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_metadata_lsize, CTLFLAG_RD, 1016 &ARC_anon.arcs_lsize[ARC_BUFC_METADATA], 0, "size of anonymous state"); 1017SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_data_lsize, CTLFLAG_RD, 1018 &ARC_anon.arcs_lsize[ARC_BUFC_DATA], 0, "size of anonymous state"); 1019 1020SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_size, CTLFLAG_RD, 1021 &ARC_mru.arcs_size, 0, "size of mru state"); 1022SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_metadata_lsize, CTLFLAG_RD, 1023 &ARC_mru.arcs_lsize[ARC_BUFC_METADATA], 0, "size of metadata in mru state"); 1024SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_data_lsize, CTLFLAG_RD, 1025 &ARC_mru.arcs_lsize[ARC_BUFC_DATA], 0, "size of data in mru state"); 1026 1027SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_size, CTLFLAG_RD, 1028 &ARC_mru_ghost.arcs_size, 0, "size of mru ghost state"); 1029SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_metadata_lsize, CTLFLAG_RD, 1030 &ARC_mru_ghost.arcs_lsize[ARC_BUFC_METADATA], 0, 1031 "size of metadata in mru ghost state"); 1032SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_data_lsize, CTLFLAG_RD, 1033 &ARC_mru_ghost.arcs_lsize[ARC_BUFC_DATA], 0, 1034 "size of data in mru ghost state"); 1035 1036SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_size, CTLFLAG_RD, 1037 &ARC_mfu.arcs_size, 0, "size of mfu state"); 1038SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_metadata_lsize, CTLFLAG_RD, 1039 &ARC_mfu.arcs_lsize[ARC_BUFC_METADATA], 0, "size of metadata in mfu state"); 1040SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_data_lsize, CTLFLAG_RD, 1041 &ARC_mfu.arcs_lsize[ARC_BUFC_DATA], 0, "size of data in mfu state"); 1042 1043SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_size, CTLFLAG_RD, 1044 &ARC_mfu_ghost.arcs_size, 0, "size of mfu ghost state"); 1045SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_metadata_lsize, CTLFLAG_RD, 1046 &ARC_mfu_ghost.arcs_lsize[ARC_BUFC_METADATA], 0, 1047 "size of metadata in mfu ghost state"); 1048SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_data_lsize, CTLFLAG_RD, 1049 &ARC_mfu_ghost.arcs_lsize[ARC_BUFC_DATA], 0, 1050 "size of data in mfu ghost state"); 1051 1052SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2c_only_size, CTLFLAG_RD, 1053 &ARC_l2c_only.arcs_size, 0, "size of mru state"); 1054 1055/* 1056 * L2ARC Internals 1057 */ 1058struct l2arc_dev { 1059 vdev_t *l2ad_vdev; /* vdev */ 1060 spa_t *l2ad_spa; /* spa */ 1061 uint64_t l2ad_hand; /* next write location */ 1062 uint64_t l2ad_start; /* first addr on device */ 1063 uint64_t l2ad_end; /* last addr on device */ 1064 boolean_t l2ad_first; /* first sweep through */ 1065 boolean_t l2ad_writing; /* currently writing */ 1066 kmutex_t l2ad_mtx; /* lock for buffer list */ 1067 list_t l2ad_buflist; /* buffer list */ 1068 list_node_t l2ad_node; /* device list node */ 1069 refcount_t l2ad_alloc; /* allocated bytes */ 1070}; 1071 1072static list_t L2ARC_dev_list; /* device list */ 1073static list_t *l2arc_dev_list; /* device list pointer */ 1074static kmutex_t l2arc_dev_mtx; /* device list mutex */ 1075static l2arc_dev_t *l2arc_dev_last; /* last device used */ 1076static list_t L2ARC_free_on_write; /* free after write buf list */ 1077static list_t *l2arc_free_on_write; /* free after write list ptr */ 1078static kmutex_t l2arc_free_on_write_mtx; /* mutex for list */ 1079static uint64_t l2arc_ndev; /* number of devices */ 1080 1081typedef struct l2arc_read_callback { 1082 arc_buf_t *l2rcb_buf; /* read buffer */ 1083 spa_t *l2rcb_spa; /* spa */ 1084 blkptr_t l2rcb_bp; /* original blkptr */ 1085 zbookmark_phys_t l2rcb_zb; /* original bookmark */ 1086 int l2rcb_flags; /* original flags */ 1087 enum zio_compress l2rcb_compress; /* applied compress */ 1088} l2arc_read_callback_t; 1089 1090typedef struct l2arc_write_callback { 1091 l2arc_dev_t *l2wcb_dev; /* device info */ 1092 arc_buf_hdr_t *l2wcb_head; /* head of write buflist */ 1093} l2arc_write_callback_t; 1094 1095typedef struct l2arc_data_free { 1096 /* protected by l2arc_free_on_write_mtx */ 1097 void *l2df_data; 1098 size_t l2df_size; 1099 void (*l2df_func)(void *, size_t); 1100 list_node_t l2df_list_node; 1101} l2arc_data_free_t; 1102 1103static kmutex_t l2arc_feed_thr_lock; 1104static kcondvar_t l2arc_feed_thr_cv; 1105static uint8_t l2arc_thread_exit; 1106 1107static void arc_get_data_buf(arc_buf_t *); 1108static void arc_access(arc_buf_hdr_t *, kmutex_t *); 1109static int arc_evict_needed(arc_buf_contents_t); 1110static void arc_evict_ghost(arc_state_t *, uint64_t, int64_t); 1111static void arc_buf_watch(arc_buf_t *); 1112 1113static arc_buf_contents_t arc_buf_type(arc_buf_hdr_t *); 1114static uint32_t arc_bufc_to_flags(arc_buf_contents_t); 1115 1116static boolean_t l2arc_write_eligible(uint64_t, arc_buf_hdr_t *); 1117static void l2arc_read_done(zio_t *); 1118 1119static boolean_t l2arc_compress_buf(arc_buf_hdr_t *); 1120static void l2arc_decompress_zio(zio_t *, arc_buf_hdr_t *, enum zio_compress); 1121static void l2arc_release_cdata_buf(arc_buf_hdr_t *); 1122 1123static uint64_t 1124buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth) 1125{ 1126 uint8_t *vdva = (uint8_t *)dva; 1127 uint64_t crc = -1ULL; 1128 int i; 1129 1130 ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY); 1131 1132 for (i = 0; i < sizeof (dva_t); i++) 1133 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ vdva[i]) & 0xFF]; 1134 1135 crc ^= (spa>>8) ^ birth; 1136 1137 return (crc); 1138} 1139 1140#define BUF_EMPTY(buf) \ 1141 ((buf)->b_dva.dva_word[0] == 0 && \ 1142 (buf)->b_dva.dva_word[1] == 0) 1143 1144#define BUF_EQUAL(spa, dva, birth, buf) \ 1145 ((buf)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \ 1146 ((buf)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \ 1147 ((buf)->b_birth == birth) && ((buf)->b_spa == spa) 1148 1149static void 1150buf_discard_identity(arc_buf_hdr_t *hdr) 1151{ 1152 hdr->b_dva.dva_word[0] = 0; 1153 hdr->b_dva.dva_word[1] = 0; 1154 hdr->b_birth = 0; 1155} 1156 1157static arc_buf_hdr_t * 1158buf_hash_find(uint64_t spa, const blkptr_t *bp, kmutex_t **lockp) 1159{ 1160 const dva_t *dva = BP_IDENTITY(bp); 1161 uint64_t birth = BP_PHYSICAL_BIRTH(bp); 1162 uint64_t idx = BUF_HASH_INDEX(spa, dva, birth); 1163 kmutex_t *hash_lock = BUF_HASH_LOCK(idx); 1164 arc_buf_hdr_t *hdr; 1165 1166 mutex_enter(hash_lock); 1167 for (hdr = buf_hash_table.ht_table[idx]; hdr != NULL; 1168 hdr = hdr->b_hash_next) { 1169 if (BUF_EQUAL(spa, dva, birth, hdr)) { 1170 *lockp = hash_lock; 1171 return (hdr); 1172 } 1173 } 1174 mutex_exit(hash_lock); 1175 *lockp = NULL; 1176 return (NULL); 1177} 1178 1179/* 1180 * Insert an entry into the hash table. If there is already an element 1181 * equal to elem in the hash table, then the already existing element 1182 * will be returned and the new element will not be inserted. 1183 * Otherwise returns NULL. 1184 * If lockp == NULL, the caller is assumed to already hold the hash lock. 1185 */ 1186static arc_buf_hdr_t * 1187buf_hash_insert(arc_buf_hdr_t *hdr, kmutex_t **lockp) 1188{ 1189 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth); 1190 kmutex_t *hash_lock = BUF_HASH_LOCK(idx); 1191 arc_buf_hdr_t *fhdr; 1192 uint32_t i; 1193 1194 ASSERT(!DVA_IS_EMPTY(&hdr->b_dva)); 1195 ASSERT(hdr->b_birth != 0); 1196 ASSERT(!HDR_IN_HASH_TABLE(hdr)); 1197 1198 if (lockp != NULL) { 1199 *lockp = hash_lock; 1200 mutex_enter(hash_lock); 1201 } else { 1202 ASSERT(MUTEX_HELD(hash_lock)); 1203 } 1204 1205 for (fhdr = buf_hash_table.ht_table[idx], i = 0; fhdr != NULL; 1206 fhdr = fhdr->b_hash_next, i++) { 1207 if (BUF_EQUAL(hdr->b_spa, &hdr->b_dva, hdr->b_birth, fhdr)) 1208 return (fhdr); 1209 } 1210 1211 hdr->b_hash_next = buf_hash_table.ht_table[idx]; 1212 buf_hash_table.ht_table[idx] = hdr; 1213 hdr->b_flags |= ARC_FLAG_IN_HASH_TABLE; 1214 1215 /* collect some hash table performance data */ 1216 if (i > 0) { 1217 ARCSTAT_BUMP(arcstat_hash_collisions); 1218 if (i == 1) 1219 ARCSTAT_BUMP(arcstat_hash_chains); 1220 1221 ARCSTAT_MAX(arcstat_hash_chain_max, i); 1222 } 1223 1224 ARCSTAT_BUMP(arcstat_hash_elements); 1225 ARCSTAT_MAXSTAT(arcstat_hash_elements); 1226 1227 return (NULL); 1228} 1229 1230static void 1231buf_hash_remove(arc_buf_hdr_t *hdr) 1232{ 1233 arc_buf_hdr_t *fhdr, **hdrp; 1234 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth); 1235 1236 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx))); 1237 ASSERT(HDR_IN_HASH_TABLE(hdr)); 1238 1239 hdrp = &buf_hash_table.ht_table[idx]; 1240 while ((fhdr = *hdrp) != hdr) { 1241 ASSERT(fhdr != NULL); 1242 hdrp = &fhdr->b_hash_next; 1243 } 1244 *hdrp = hdr->b_hash_next; 1245 hdr->b_hash_next = NULL; 1246 hdr->b_flags &= ~ARC_FLAG_IN_HASH_TABLE; 1247 1248 /* collect some hash table performance data */ 1249 ARCSTAT_BUMPDOWN(arcstat_hash_elements); 1250 1251 if (buf_hash_table.ht_table[idx] && 1252 buf_hash_table.ht_table[idx]->b_hash_next == NULL) 1253 ARCSTAT_BUMPDOWN(arcstat_hash_chains); 1254} 1255 1256/* 1257 * Global data structures and functions for the buf kmem cache. 1258 */ 1259static kmem_cache_t *hdr_full_cache; 1260static kmem_cache_t *hdr_l2only_cache; 1261static kmem_cache_t *buf_cache; 1262 1263static void 1264buf_fini(void) 1265{ 1266 int i; 1267 1268 kmem_free(buf_hash_table.ht_table, 1269 (buf_hash_table.ht_mask + 1) * sizeof (void *)); 1270 for (i = 0; i < BUF_LOCKS; i++) 1271 mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock); 1272 kmem_cache_destroy(hdr_full_cache); 1273 kmem_cache_destroy(hdr_l2only_cache); 1274 kmem_cache_destroy(buf_cache); 1275} 1276 1277/* 1278 * Constructor callback - called when the cache is empty 1279 * and a new buf is requested. 1280 */ 1281/* ARGSUSED */ 1282static int 1283hdr_full_cons(void *vbuf, void *unused, int kmflag) 1284{ 1285 arc_buf_hdr_t *hdr = vbuf; 1286 1287 bzero(hdr, HDR_FULL_SIZE); 1288 cv_init(&hdr->b_l1hdr.b_cv, NULL, CV_DEFAULT, NULL); 1289 refcount_create(&hdr->b_l1hdr.b_refcnt); 1290 mutex_init(&hdr->b_l1hdr.b_freeze_lock, NULL, MUTEX_DEFAULT, NULL); 1291 arc_space_consume(HDR_FULL_SIZE, ARC_SPACE_HDRS); 1292 1293 return (0); 1294} 1295 1296/* ARGSUSED */ 1297static int 1298hdr_l2only_cons(void *vbuf, void *unused, int kmflag) 1299{ 1300 arc_buf_hdr_t *hdr = vbuf; 1301 1302 bzero(hdr, HDR_L2ONLY_SIZE); 1303 arc_space_consume(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS); 1304 1305 return (0); 1306} 1307 1308/* ARGSUSED */ 1309static int 1310buf_cons(void *vbuf, void *unused, int kmflag) 1311{ 1312 arc_buf_t *buf = vbuf; 1313 1314 bzero(buf, sizeof (arc_buf_t)); 1315 mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL); 1316 arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS); 1317 1318 return (0); 1319} 1320 1321/* 1322 * Destructor callback - called when a cached buf is 1323 * no longer required. 1324 */ 1325/* ARGSUSED */ 1326static void 1327hdr_full_dest(void *vbuf, void *unused) 1328{ 1329 arc_buf_hdr_t *hdr = vbuf; 1330 1331 ASSERT(BUF_EMPTY(hdr)); 1332 cv_destroy(&hdr->b_l1hdr.b_cv); 1333 refcount_destroy(&hdr->b_l1hdr.b_refcnt); 1334 mutex_destroy(&hdr->b_l1hdr.b_freeze_lock); 1335 arc_space_return(HDR_FULL_SIZE, ARC_SPACE_HDRS); 1336} 1337 1338/* ARGSUSED */ 1339static void 1340hdr_l2only_dest(void *vbuf, void *unused) 1341{ 1342 arc_buf_hdr_t *hdr = vbuf; 1343 1344 ASSERT(BUF_EMPTY(hdr)); 1345 arc_space_return(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS); 1346} 1347 1348/* ARGSUSED */ 1349static void 1350buf_dest(void *vbuf, void *unused) 1351{ 1352 arc_buf_t *buf = vbuf; 1353 1354 mutex_destroy(&buf->b_evict_lock); 1355 arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS); 1356} 1357 1358/* 1359 * Reclaim callback -- invoked when memory is low. 1360 */ 1361/* ARGSUSED */ 1362static void 1363hdr_recl(void *unused) 1364{ 1365 dprintf("hdr_recl called\n"); 1366 /* 1367 * umem calls the reclaim func when we destroy the buf cache, 1368 * which is after we do arc_fini(). 1369 */ 1370 if (!arc_dead) 1371 cv_signal(&arc_reclaim_thr_cv); 1372} 1373 1374static void 1375buf_init(void) 1376{ 1377 uint64_t *ct; 1378 uint64_t hsize = 1ULL << 12; 1379 int i, j; 1380 1381 /* 1382 * The hash table is big enough to fill all of physical memory 1383 * with an average block size of zfs_arc_average_blocksize (default 8K). 1384 * By default, the table will take up 1385 * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers). 1386 */ 1387 while (hsize * zfs_arc_average_blocksize < (uint64_t)physmem * PAGESIZE) 1388 hsize <<= 1; 1389retry: 1390 buf_hash_table.ht_mask = hsize - 1; 1391 buf_hash_table.ht_table = 1392 kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP); 1393 if (buf_hash_table.ht_table == NULL) { 1394 ASSERT(hsize > (1ULL << 8)); 1395 hsize >>= 1; 1396 goto retry; 1397 } 1398 1399 hdr_full_cache = kmem_cache_create("arc_buf_hdr_t_full", HDR_FULL_SIZE, 1400 0, hdr_full_cons, hdr_full_dest, hdr_recl, NULL, NULL, 0); 1401 hdr_l2only_cache = kmem_cache_create("arc_buf_hdr_t_l2only", 1402 HDR_L2ONLY_SIZE, 0, hdr_l2only_cons, hdr_l2only_dest, hdr_recl, 1403 NULL, NULL, 0); 1404 buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t), 1405 0, buf_cons, buf_dest, NULL, NULL, NULL, 0); 1406 1407 for (i = 0; i < 256; i++) 1408 for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--) 1409 *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY); 1410 1411 for (i = 0; i < BUF_LOCKS; i++) { 1412 mutex_init(&buf_hash_table.ht_locks[i].ht_lock, 1413 NULL, MUTEX_DEFAULT, NULL); 1414 } 1415} 1416 1417/* 1418 * Transition between the two allocation states for the arc_buf_hdr struct. 1419 * The arc_buf_hdr struct can be allocated with (hdr_full_cache) or without 1420 * (hdr_l2only_cache) the fields necessary for the L1 cache - the smaller 1421 * version is used when a cache buffer is only in the L2ARC in order to reduce 1422 * memory usage. 1423 */ 1424static arc_buf_hdr_t * 1425arc_hdr_realloc(arc_buf_hdr_t *hdr, kmem_cache_t *old, kmem_cache_t *new) 1426{ 1427 ASSERT(HDR_HAS_L2HDR(hdr)); 1428 1429 arc_buf_hdr_t *nhdr; 1430 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev; 1431 1432 ASSERT((old == hdr_full_cache && new == hdr_l2only_cache) || 1433 (old == hdr_l2only_cache && new == hdr_full_cache)); 1434 1435 nhdr = kmem_cache_alloc(new, KM_PUSHPAGE); 1436 1437 ASSERT(MUTEX_HELD(HDR_LOCK(hdr))); 1438 buf_hash_remove(hdr); 1439 1440 bcopy(hdr, nhdr, HDR_L2ONLY_SIZE); 1441 1442 if (new == hdr_full_cache) { 1443 nhdr->b_flags |= ARC_FLAG_HAS_L1HDR; 1444 /* 1445 * arc_access and arc_change_state need to be aware that a 1446 * header has just come out of L2ARC, so we set its state to 1447 * l2c_only even though it's about to change. 1448 */ 1449 nhdr->b_l1hdr.b_state = arc_l2c_only; 1450 } else { 1451 ASSERT(hdr->b_l1hdr.b_buf == NULL); 1452 ASSERT0(hdr->b_l1hdr.b_datacnt); 1453 ASSERT(!list_link_active(&hdr->b_l1hdr.b_arc_node)); 1454 /* 1455 * We might be removing the L1hdr of a buffer which was just 1456 * written out to L2ARC. If such a buffer is compressed then we 1457 * need to free its b_tmp_cdata before destroying the header. 1458 */ 1459 if (hdr->b_l1hdr.b_tmp_cdata != NULL && 1460 HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF) 1461 l2arc_release_cdata_buf(hdr); 1462 nhdr->b_flags &= ~ARC_FLAG_HAS_L1HDR; 1463 } 1464 /* 1465 * The header has been reallocated so we need to re-insert it into any 1466 * lists it was on. 1467 */ 1468 (void) buf_hash_insert(nhdr, NULL); 1469 1470 ASSERT(list_link_active(&hdr->b_l2hdr.b_l2node)); 1471 1472 mutex_enter(&dev->l2ad_mtx); 1473 1474 /* 1475 * We must place the realloc'ed header back into the list at 1476 * the same spot. Otherwise, if it's placed earlier in the list, 1477 * l2arc_write_buffers() could find it during the function's 1478 * write phase, and try to write it out to the l2arc. 1479 */ 1480 list_insert_after(&dev->l2ad_buflist, hdr, nhdr); 1481 list_remove(&dev->l2ad_buflist, hdr); 1482 1483 mutex_exit(&dev->l2ad_mtx); 1484 1485 /* 1486 * Since we're using the pointer address as the tag when 1487 * incrementing and decrementing the l2ad_alloc refcount, we 1488 * must remove the old pointer (that we're about to destroy) and 1489 * add the new pointer to the refcount. Otherwise we'd remove 1490 * the wrong pointer address when calling arc_hdr_destroy() later. 1491 */ 1492 1493 (void) refcount_remove_many(&dev->l2ad_alloc, 1494 hdr->b_l2hdr.b_asize, hdr); 1495 1496 (void) refcount_add_many(&dev->l2ad_alloc, 1497 nhdr->b_l2hdr.b_asize, nhdr); 1498 1499 buf_discard_identity(hdr); 1500 hdr->b_freeze_cksum = NULL; 1501 kmem_cache_free(old, hdr); 1502 1503 return (nhdr); 1504} 1505 1506 1507#define ARC_MINTIME (hz>>4) /* 62 ms */ 1508 1509static void 1510arc_cksum_verify(arc_buf_t *buf) 1511{ 1512 zio_cksum_t zc; 1513 1514 if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 1515 return; 1516 1517 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1518 if (buf->b_hdr->b_freeze_cksum == NULL || HDR_IO_ERROR(buf->b_hdr)) { 1519 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1520 return; 1521 } 1522 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc); 1523 if (!ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc)) 1524 panic("buffer modified while frozen!"); 1525 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1526} 1527 1528static int 1529arc_cksum_equal(arc_buf_t *buf) 1530{ 1531 zio_cksum_t zc; 1532 int equal; 1533 1534 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1535 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc); 1536 equal = ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc); 1537 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1538 1539 return (equal); 1540} 1541 1542static void 1543arc_cksum_compute(arc_buf_t *buf, boolean_t force) 1544{ 1545 if (!force && !(zfs_flags & ZFS_DEBUG_MODIFY)) 1546 return; 1547 1548 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1549 if (buf->b_hdr->b_freeze_cksum != NULL) { 1550 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1551 return; 1552 } 1553 buf->b_hdr->b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t), KM_SLEEP); 1554 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, 1555 buf->b_hdr->b_freeze_cksum); 1556 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1557#ifdef illumos 1558 arc_buf_watch(buf); 1559#endif 1560} 1561 1562#ifdef illumos 1563#ifndef _KERNEL 1564typedef struct procctl { 1565 long cmd; 1566 prwatch_t prwatch; 1567} procctl_t; 1568#endif 1569 1570/* ARGSUSED */ 1571static void 1572arc_buf_unwatch(arc_buf_t *buf) 1573{ 1574#ifndef _KERNEL 1575 if (arc_watch) { 1576 int result; 1577 procctl_t ctl; 1578 ctl.cmd = PCWATCH; 1579 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data; 1580 ctl.prwatch.pr_size = 0; 1581 ctl.prwatch.pr_wflags = 0; 1582 result = write(arc_procfd, &ctl, sizeof (ctl)); 1583 ASSERT3U(result, ==, sizeof (ctl)); 1584 } 1585#endif 1586} 1587 1588/* ARGSUSED */ 1589static void 1590arc_buf_watch(arc_buf_t *buf) 1591{ 1592#ifndef _KERNEL 1593 if (arc_watch) { 1594 int result; 1595 procctl_t ctl; 1596 ctl.cmd = PCWATCH; 1597 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data; 1598 ctl.prwatch.pr_size = buf->b_hdr->b_size; 1599 ctl.prwatch.pr_wflags = WA_WRITE; 1600 result = write(arc_procfd, &ctl, sizeof (ctl)); 1601 ASSERT3U(result, ==, sizeof (ctl)); 1602 } 1603#endif 1604} 1605#endif /* illumos */ 1606 1607static arc_buf_contents_t 1608arc_buf_type(arc_buf_hdr_t *hdr) 1609{ 1610 if (HDR_ISTYPE_METADATA(hdr)) { 1611 return (ARC_BUFC_METADATA); 1612 } else { 1613 return (ARC_BUFC_DATA); 1614 } 1615} 1616 1617static uint32_t 1618arc_bufc_to_flags(arc_buf_contents_t type) 1619{ 1620 switch (type) { 1621 case ARC_BUFC_DATA: 1622 /* metadata field is 0 if buffer contains normal data */ 1623 return (0); 1624 case ARC_BUFC_METADATA: 1625 return (ARC_FLAG_BUFC_METADATA); 1626 default: 1627 break; 1628 } 1629 panic("undefined ARC buffer type!"); 1630 return ((uint32_t)-1); 1631} 1632 1633void 1634arc_buf_thaw(arc_buf_t *buf) 1635{ 1636 if (zfs_flags & ZFS_DEBUG_MODIFY) { 1637 if (buf->b_hdr->b_l1hdr.b_state != arc_anon) 1638 panic("modifying non-anon buffer!"); 1639 if (HDR_IO_IN_PROGRESS(buf->b_hdr)) 1640 panic("modifying buffer while i/o in progress!"); 1641 arc_cksum_verify(buf); 1642 } 1643 1644 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1645 if (buf->b_hdr->b_freeze_cksum != NULL) { 1646 kmem_free(buf->b_hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 1647 buf->b_hdr->b_freeze_cksum = NULL; 1648 } 1649 1650#ifdef ZFS_DEBUG 1651 if (zfs_flags & ZFS_DEBUG_MODIFY) { 1652 if (buf->b_hdr->b_l1hdr.b_thawed != NULL) 1653 kmem_free(buf->b_hdr->b_l1hdr.b_thawed, 1); 1654 buf->b_hdr->b_l1hdr.b_thawed = kmem_alloc(1, KM_SLEEP); 1655 } 1656#endif 1657 1658 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1659 1660#ifdef illumos 1661 arc_buf_unwatch(buf); 1662#endif 1663} 1664 1665void 1666arc_buf_freeze(arc_buf_t *buf) 1667{ 1668 kmutex_t *hash_lock; 1669 1670 if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 1671 return; 1672 1673 hash_lock = HDR_LOCK(buf->b_hdr); 1674 mutex_enter(hash_lock); 1675 1676 ASSERT(buf->b_hdr->b_freeze_cksum != NULL || 1677 buf->b_hdr->b_l1hdr.b_state == arc_anon); 1678 arc_cksum_compute(buf, B_FALSE); 1679 mutex_exit(hash_lock); 1680 1681} 1682 1683static void 1684get_buf_info(arc_buf_hdr_t *hdr, arc_state_t *state, list_t **list, kmutex_t **lock) 1685{ 1686 uint64_t buf_hashid = buf_hash(hdr->b_spa, &hdr->b_dva, hdr->b_birth); 1687 1688 if (arc_buf_type(hdr) == ARC_BUFC_METADATA) 1689 buf_hashid &= (ARC_BUFC_NUMMETADATALISTS - 1); 1690 else { 1691 buf_hashid &= (ARC_BUFC_NUMDATALISTS - 1); 1692 buf_hashid += ARC_BUFC_NUMMETADATALISTS; 1693 } 1694 1695 *list = &state->arcs_lists[buf_hashid]; 1696 *lock = ARCS_LOCK(state, buf_hashid); 1697} 1698 1699 1700static void 1701add_reference(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, void *tag) 1702{ 1703 ASSERT(HDR_HAS_L1HDR(hdr)); 1704 ASSERT(MUTEX_HELD(hash_lock)); 1705 arc_state_t *state = hdr->b_l1hdr.b_state; 1706 1707 if ((refcount_add(&hdr->b_l1hdr.b_refcnt, tag) == 1) && 1708 (state != arc_anon)) { 1709 /* We don't use the L2-only state list. */ 1710 if (state != arc_l2c_only) { 1711 uint64_t delta = hdr->b_size * hdr->b_l1hdr.b_datacnt; 1712 uint64_t *size = &state->arcs_lsize[arc_buf_type(hdr)]; 1713 list_t *list; 1714 kmutex_t *lock; 1715 1716 get_buf_info(hdr, state, &list, &lock); 1717 ASSERT(!MUTEX_HELD(lock)); 1718 mutex_enter(lock); 1719 ASSERT(list_link_active(&hdr->b_l1hdr.b_arc_node)); 1720 list_remove(list, hdr); 1721 if (GHOST_STATE(state)) { 1722 ASSERT0(hdr->b_l1hdr.b_datacnt); 1723 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 1724 delta = hdr->b_size; 1725 } 1726 ASSERT(delta > 0); 1727 ASSERT3U(*size, >=, delta); 1728 atomic_add_64(size, -delta); 1729 mutex_exit(lock); 1730 } 1731 /* remove the prefetch flag if we get a reference */ 1732 hdr->b_flags &= ~ARC_FLAG_PREFETCH; 1733 } 1734} 1735 1736static int 1737remove_reference(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, void *tag) 1738{ 1739 int cnt; 1740 arc_state_t *state = hdr->b_l1hdr.b_state; 1741 1742 ASSERT(HDR_HAS_L1HDR(hdr)); 1743 ASSERT(state == arc_anon || MUTEX_HELD(hash_lock)); 1744 ASSERT(!GHOST_STATE(state)); 1745 1746 /* 1747 * arc_l2c_only counts as a ghost state so we don't need to explicitly 1748 * check to prevent usage of the arc_l2c_only list. 1749 */ 1750 if (((cnt = refcount_remove(&hdr->b_l1hdr.b_refcnt, tag)) == 0) && 1751 (state != arc_anon)) { 1752 uint64_t *size = &state->arcs_lsize[arc_buf_type(hdr)]; 1753 list_t *list; 1754 kmutex_t *lock; 1755 1756 get_buf_info(hdr, state, &list, &lock); 1757 ASSERT(!MUTEX_HELD(lock)); 1758 mutex_enter(lock); 1759 ASSERT(!list_link_active(&hdr->b_l1hdr.b_arc_node)); 1760 list_insert_head(list, hdr); 1761 ASSERT(hdr->b_l1hdr.b_datacnt > 0); 1762 atomic_add_64(size, hdr->b_size * 1763 hdr->b_l1hdr.b_datacnt); 1764 mutex_exit(lock); 1765 } 1766 return (cnt); 1767} 1768 1769/* 1770 * Move the supplied buffer to the indicated state. The mutex 1771 * for the buffer must be held by the caller. 1772 */ 1773static void 1774arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *hdr, 1775 kmutex_t *hash_lock) 1776{ 1777 arc_state_t *old_state; 1778 int64_t refcnt; 1779 uint32_t datacnt; 1780 uint64_t from_delta, to_delta; 1781 arc_buf_contents_t buftype = arc_buf_type(hdr); 1782 list_t *list; 1783 kmutex_t *lock; 1784 1785 /* 1786 * We almost always have an L1 hdr here, since we call arc_hdr_realloc() 1787 * in arc_read() when bringing a buffer out of the L2ARC. However, the 1788 * L1 hdr doesn't always exist when we change state to arc_anon before 1789 * destroying a header, in which case reallocating to add the L1 hdr is 1790 * pointless. 1791 */ 1792 if (HDR_HAS_L1HDR(hdr)) { 1793 old_state = hdr->b_l1hdr.b_state; 1794 refcnt = refcount_count(&hdr->b_l1hdr.b_refcnt); 1795 datacnt = hdr->b_l1hdr.b_datacnt; 1796 } else { 1797 old_state = arc_l2c_only; 1798 refcnt = 0; 1799 datacnt = 0; 1800 } 1801 1802 ASSERT(MUTEX_HELD(hash_lock)); 1803 ASSERT3P(new_state, !=, old_state); 1804 ASSERT(refcnt == 0 || datacnt > 0); 1805 ASSERT(!GHOST_STATE(new_state) || datacnt == 0); 1806 ASSERT(old_state != arc_anon || datacnt <= 1); 1807 1808 from_delta = to_delta = datacnt * hdr->b_size; 1809 1810 /* 1811 * If this buffer is evictable, transfer it from the 1812 * old state list to the new state list. 1813 */ 1814 if (refcnt == 0) { 1815 if (old_state != arc_anon && old_state != arc_l2c_only) { 1816 int use_mutex; 1817 uint64_t *size = &old_state->arcs_lsize[buftype]; 1818 1819 get_buf_info(hdr, old_state, &list, &lock); 1820 use_mutex = !MUTEX_HELD(lock); 1821 if (use_mutex) 1822 mutex_enter(lock); 1823 1824 ASSERT(HDR_HAS_L1HDR(hdr)); 1825 ASSERT(list_link_active(&hdr->b_l1hdr.b_arc_node)); 1826 list_remove(list, hdr); 1827 1828 /* 1829 * If prefetching out of the ghost cache, 1830 * we will have a non-zero datacnt. 1831 */ 1832 if (GHOST_STATE(old_state) && datacnt == 0) { 1833 /* ghost elements have a ghost size */ 1834 ASSERT(hdr->b_l1hdr.b_buf == NULL); 1835 from_delta = hdr->b_size; 1836 } 1837 ASSERT3U(*size, >=, from_delta); 1838 atomic_add_64(size, -from_delta); 1839 1840 if (use_mutex) 1841 mutex_exit(lock); 1842 } 1843 if (new_state != arc_anon && new_state != arc_l2c_only) { 1844 int use_mutex; 1845 uint64_t *size = &new_state->arcs_lsize[buftype]; 1846 1847 /* 1848 * An L1 header always exists here, since if we're 1849 * moving to some L1-cached state (i.e. not l2c_only or 1850 * anonymous), we realloc the header to add an L1hdr 1851 * beforehand. 1852 */ 1853 ASSERT(HDR_HAS_L1HDR(hdr)); 1854 get_buf_info(hdr, new_state, &list, &lock); 1855 use_mutex = !MUTEX_HELD(lock); 1856 if (use_mutex) 1857 mutex_enter(lock); 1858 1859 list_insert_head(list, hdr); 1860 1861 /* ghost elements have a ghost size */ 1862 if (GHOST_STATE(new_state)) { 1863 ASSERT(datacnt == 0); 1864 ASSERT(hdr->b_l1hdr.b_buf == NULL); 1865 to_delta = hdr->b_size; 1866 } 1867 atomic_add_64(size, to_delta); 1868 1869 if (use_mutex) 1870 mutex_exit(lock); 1871 } 1872 } 1873 1874 ASSERT(!BUF_EMPTY(hdr)); 1875 if (new_state == arc_anon && HDR_IN_HASH_TABLE(hdr)) 1876 buf_hash_remove(hdr); 1877 1878 /* adjust state sizes (ignore arc_l2c_only) */ 1879 if (to_delta && new_state != arc_l2c_only) 1880 atomic_add_64(&new_state->arcs_size, to_delta); 1881 if (from_delta && old_state != arc_l2c_only) { 1882 ASSERT3U(old_state->arcs_size, >=, from_delta); 1883 atomic_add_64(&old_state->arcs_size, -from_delta); 1884 } 1885 if (HDR_HAS_L1HDR(hdr)) 1886 hdr->b_l1hdr.b_state = new_state; 1887 1888 /* 1889 * L2 headers should never be on the L2 state list since they don't 1890 * have L1 headers allocated. 1891 */ 1892#ifdef illumos 1893 ASSERT(list_is_empty(&arc_l2c_only->arcs_list[ARC_BUFC_DATA]) && 1894 list_is_empty(&arc_l2c_only->arcs_list[ARC_BUFC_METADATA])); 1895#endif 1896} 1897 1898void 1899arc_space_consume(uint64_t space, arc_space_type_t type) 1900{ 1901 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES); 1902 1903 switch (type) { 1904 case ARC_SPACE_DATA: 1905 ARCSTAT_INCR(arcstat_data_size, space); 1906 break; 1907 case ARC_SPACE_META: 1908 ARCSTAT_INCR(arcstat_metadata_size, space); 1909 break; 1910 case ARC_SPACE_OTHER: 1911 ARCSTAT_INCR(arcstat_other_size, space); 1912 break; 1913 case ARC_SPACE_HDRS: 1914 ARCSTAT_INCR(arcstat_hdr_size, space); 1915 break; 1916 case ARC_SPACE_L2HDRS: 1917 ARCSTAT_INCR(arcstat_l2_hdr_size, space); 1918 break; 1919 } 1920 1921 if (type != ARC_SPACE_DATA) 1922 ARCSTAT_INCR(arcstat_meta_used, space); 1923 1924 atomic_add_64(&arc_size, space); 1925} 1926 1927void 1928arc_space_return(uint64_t space, arc_space_type_t type) 1929{ 1930 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES); 1931 1932 switch (type) { 1933 case ARC_SPACE_DATA: 1934 ARCSTAT_INCR(arcstat_data_size, -space); 1935 break; 1936 case ARC_SPACE_META: 1937 ARCSTAT_INCR(arcstat_metadata_size, -space); 1938 break; 1939 case ARC_SPACE_OTHER: 1940 ARCSTAT_INCR(arcstat_other_size, -space); 1941 break; 1942 case ARC_SPACE_HDRS: 1943 ARCSTAT_INCR(arcstat_hdr_size, -space); 1944 break; 1945 case ARC_SPACE_L2HDRS: 1946 ARCSTAT_INCR(arcstat_l2_hdr_size, -space); 1947 break; 1948 } 1949 1950 if (type != ARC_SPACE_DATA) { 1951 ASSERT(arc_meta_used >= space); 1952 if (arc_meta_max < arc_meta_used) 1953 arc_meta_max = arc_meta_used; 1954 ARCSTAT_INCR(arcstat_meta_used, -space); 1955 } 1956 1957 ASSERT(arc_size >= space); 1958 atomic_add_64(&arc_size, -space); 1959} 1960 1961arc_buf_t * 1962arc_buf_alloc(spa_t *spa, int32_t size, void *tag, arc_buf_contents_t type) 1963{ 1964 arc_buf_hdr_t *hdr; 1965 arc_buf_t *buf; 1966 1967 ASSERT3U(size, >, 0); 1968 hdr = kmem_cache_alloc(hdr_full_cache, KM_PUSHPAGE); 1969 ASSERT(BUF_EMPTY(hdr)); 1970 ASSERT3P(hdr->b_freeze_cksum, ==, NULL); 1971 hdr->b_size = size; 1972 hdr->b_spa = spa_load_guid(spa); 1973 1974 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 1975 buf->b_hdr = hdr; 1976 buf->b_data = NULL; 1977 buf->b_efunc = NULL; 1978 buf->b_private = NULL; 1979 buf->b_next = NULL; 1980 1981 hdr->b_flags = arc_bufc_to_flags(type); 1982 hdr->b_flags |= ARC_FLAG_HAS_L1HDR; 1983 1984 hdr->b_l1hdr.b_buf = buf; 1985 hdr->b_l1hdr.b_state = arc_anon; 1986 hdr->b_l1hdr.b_arc_access = 0; 1987 hdr->b_l1hdr.b_datacnt = 1; 1988 1989 arc_get_data_buf(buf); 1990 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 1991 (void) refcount_add(&hdr->b_l1hdr.b_refcnt, tag); 1992 1993 return (buf); 1994} 1995 1996static char *arc_onloan_tag = "onloan"; 1997 1998/* 1999 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in 2000 * flight data by arc_tempreserve_space() until they are "returned". Loaned 2001 * buffers must be returned to the arc before they can be used by the DMU or 2002 * freed. 2003 */ 2004arc_buf_t * 2005arc_loan_buf(spa_t *spa, int size) 2006{ 2007 arc_buf_t *buf; 2008 2009 buf = arc_buf_alloc(spa, size, arc_onloan_tag, ARC_BUFC_DATA); 2010 2011 atomic_add_64(&arc_loaned_bytes, size); 2012 return (buf); 2013} 2014 2015/* 2016 * Return a loaned arc buffer to the arc. 2017 */ 2018void 2019arc_return_buf(arc_buf_t *buf, void *tag) 2020{ 2021 arc_buf_hdr_t *hdr = buf->b_hdr; 2022 2023 ASSERT(buf->b_data != NULL); 2024 ASSERT(HDR_HAS_L1HDR(hdr)); 2025 (void) refcount_add(&hdr->b_l1hdr.b_refcnt, tag); 2026 (void) refcount_remove(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag); 2027 2028 atomic_add_64(&arc_loaned_bytes, -hdr->b_size); 2029} 2030 2031/* Detach an arc_buf from a dbuf (tag) */ 2032void 2033arc_loan_inuse_buf(arc_buf_t *buf, void *tag) 2034{ 2035 arc_buf_hdr_t *hdr = buf->b_hdr; 2036 2037 ASSERT(buf->b_data != NULL); 2038 ASSERT(HDR_HAS_L1HDR(hdr)); 2039 (void) refcount_add(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag); 2040 (void) refcount_remove(&hdr->b_l1hdr.b_refcnt, tag); 2041 buf->b_efunc = NULL; 2042 buf->b_private = NULL; 2043 2044 atomic_add_64(&arc_loaned_bytes, hdr->b_size); 2045} 2046 2047static arc_buf_t * 2048arc_buf_clone(arc_buf_t *from) 2049{ 2050 arc_buf_t *buf; 2051 arc_buf_hdr_t *hdr = from->b_hdr; 2052 uint64_t size = hdr->b_size; 2053 2054 ASSERT(HDR_HAS_L1HDR(hdr)); 2055 ASSERT(hdr->b_l1hdr.b_state != arc_anon); 2056 2057 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 2058 buf->b_hdr = hdr; 2059 buf->b_data = NULL; 2060 buf->b_efunc = NULL; 2061 buf->b_private = NULL; 2062 buf->b_next = hdr->b_l1hdr.b_buf; 2063 hdr->b_l1hdr.b_buf = buf; 2064 arc_get_data_buf(buf); 2065 bcopy(from->b_data, buf->b_data, size); 2066 2067 /* 2068 * This buffer already exists in the arc so create a duplicate 2069 * copy for the caller. If the buffer is associated with user data 2070 * then track the size and number of duplicates. These stats will be 2071 * updated as duplicate buffers are created and destroyed. 2072 */ 2073 if (HDR_ISTYPE_DATA(hdr)) { 2074 ARCSTAT_BUMP(arcstat_duplicate_buffers); 2075 ARCSTAT_INCR(arcstat_duplicate_buffers_size, size); 2076 } 2077 hdr->b_l1hdr.b_datacnt += 1; 2078 return (buf); 2079} 2080 2081void 2082arc_buf_add_ref(arc_buf_t *buf, void* tag) 2083{ 2084 arc_buf_hdr_t *hdr; 2085 kmutex_t *hash_lock; 2086 2087 /* 2088 * Check to see if this buffer is evicted. Callers 2089 * must verify b_data != NULL to know if the add_ref 2090 * was successful. 2091 */ 2092 mutex_enter(&buf->b_evict_lock); 2093 if (buf->b_data == NULL) { 2094 mutex_exit(&buf->b_evict_lock); 2095 return; 2096 } 2097 hash_lock = HDR_LOCK(buf->b_hdr); 2098 mutex_enter(hash_lock); 2099 hdr = buf->b_hdr; 2100 ASSERT(HDR_HAS_L1HDR(hdr)); 2101 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 2102 mutex_exit(&buf->b_evict_lock); 2103 2104 ASSERT(hdr->b_l1hdr.b_state == arc_mru || 2105 hdr->b_l1hdr.b_state == arc_mfu); 2106 2107 add_reference(hdr, hash_lock, tag); 2108 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr); 2109 arc_access(hdr, hash_lock); 2110 mutex_exit(hash_lock); 2111 ARCSTAT_BUMP(arcstat_hits); 2112 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr), 2113 demand, prefetch, !HDR_ISTYPE_METADATA(hdr), 2114 data, metadata, hits); 2115} 2116 2117static void 2118arc_buf_free_on_write(void *data, size_t size, 2119 void (*free_func)(void *, size_t)) 2120{ 2121 l2arc_data_free_t *df; 2122 2123 df = kmem_alloc(sizeof (l2arc_data_free_t), KM_SLEEP); 2124 df->l2df_data = data; 2125 df->l2df_size = size; 2126 df->l2df_func = free_func; 2127 mutex_enter(&l2arc_free_on_write_mtx); 2128 list_insert_head(l2arc_free_on_write, df); 2129 mutex_exit(&l2arc_free_on_write_mtx); 2130} 2131 2132/* 2133 * Free the arc data buffer. If it is an l2arc write in progress, 2134 * the buffer is placed on l2arc_free_on_write to be freed later. 2135 */ 2136static void 2137arc_buf_data_free(arc_buf_t *buf, void (*free_func)(void *, size_t)) 2138{ 2139 arc_buf_hdr_t *hdr = buf->b_hdr; 2140 2141 if (HDR_L2_WRITING(hdr)) { 2142 arc_buf_free_on_write(buf->b_data, hdr->b_size, free_func); 2143 ARCSTAT_BUMP(arcstat_l2_free_on_write); 2144 } else { 2145 free_func(buf->b_data, hdr->b_size); 2146 } 2147} 2148 2149/* 2150 * Free up buf->b_data and if 'remove' is set, then pull the 2151 * arc_buf_t off of the the arc_buf_hdr_t's list and free it. 2152 */ 2153static void 2154arc_buf_l2_cdata_free(arc_buf_hdr_t *hdr) 2155{ 2156 ASSERT(HDR_HAS_L2HDR(hdr)); 2157 ASSERT(MUTEX_HELD(&hdr->b_l2hdr.b_dev->l2ad_mtx)); 2158 2159 /* 2160 * The b_tmp_cdata field is linked off of the b_l1hdr, so if 2161 * that doesn't exist, the header is in the arc_l2c_only state, 2162 * and there isn't anything to free (it's already been freed). 2163 */ 2164 if (!HDR_HAS_L1HDR(hdr)) 2165 return; 2166 2167 if (hdr->b_l1hdr.b_tmp_cdata == NULL) 2168 return; 2169 2170 ASSERT(HDR_L2_WRITING(hdr)); 2171 arc_buf_free_on_write(hdr->b_l1hdr.b_tmp_cdata, hdr->b_size, 2172 zio_data_buf_free); 2173 2174 ARCSTAT_BUMP(arcstat_l2_cdata_free_on_write); 2175 hdr->b_l1hdr.b_tmp_cdata = NULL; 2176} 2177 2178static void 2179arc_buf_destroy(arc_buf_t *buf, boolean_t recycle, boolean_t remove) 2180{ 2181 arc_buf_t **bufp; 2182 2183 /* free up data associated with the buf */ 2184 if (buf->b_data != NULL) { 2185 arc_state_t *state = buf->b_hdr->b_l1hdr.b_state; 2186 uint64_t size = buf->b_hdr->b_size; 2187 arc_buf_contents_t type = arc_buf_type(buf->b_hdr); 2188 2189 arc_cksum_verify(buf); 2190#ifdef illumos 2191 arc_buf_unwatch(buf); 2192#endif 2193 2194 if (!recycle) { 2195 if (type == ARC_BUFC_METADATA) { 2196 arc_buf_data_free(buf, zio_buf_free); 2197 arc_space_return(size, ARC_SPACE_META); 2198 } else { 2199 ASSERT(type == ARC_BUFC_DATA); 2200 arc_buf_data_free(buf, zio_data_buf_free); 2201 arc_space_return(size, ARC_SPACE_DATA); 2202 } 2203 } 2204 if (list_link_active(&buf->b_hdr->b_l1hdr.b_arc_node)) { 2205 uint64_t *cnt = &state->arcs_lsize[type]; 2206 2207 ASSERT(refcount_is_zero( 2208 &buf->b_hdr->b_l1hdr.b_refcnt)); 2209 ASSERT(state != arc_anon && state != arc_l2c_only); 2210 2211 ASSERT3U(*cnt, >=, size); 2212 atomic_add_64(cnt, -size); 2213 } 2214 ASSERT3U(state->arcs_size, >=, size); 2215 atomic_add_64(&state->arcs_size, -size); 2216 buf->b_data = NULL; 2217 2218 /* 2219 * If we're destroying a duplicate buffer make sure 2220 * that the appropriate statistics are updated. 2221 */ 2222 if (buf->b_hdr->b_l1hdr.b_datacnt > 1 && 2223 HDR_ISTYPE_DATA(buf->b_hdr)) { 2224 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers); 2225 ARCSTAT_INCR(arcstat_duplicate_buffers_size, -size); 2226 } 2227 ASSERT(buf->b_hdr->b_l1hdr.b_datacnt > 0); 2228 buf->b_hdr->b_l1hdr.b_datacnt -= 1; 2229 } 2230 2231 /* only remove the buf if requested */ 2232 if (!remove) 2233 return; 2234 2235 /* remove the buf from the hdr list */ 2236 for (bufp = &buf->b_hdr->b_l1hdr.b_buf; *bufp != buf; 2237 bufp = &(*bufp)->b_next) 2238 continue; 2239 *bufp = buf->b_next; 2240 buf->b_next = NULL; 2241 2242 ASSERT(buf->b_efunc == NULL); 2243 2244 /* clean up the buf */ 2245 buf->b_hdr = NULL; 2246 kmem_cache_free(buf_cache, buf); 2247} 2248 2249static void 2250arc_hdr_l2hdr_destroy(arc_buf_hdr_t *hdr) 2251{ 2252 l2arc_buf_hdr_t *l2hdr = &hdr->b_l2hdr; 2253 l2arc_dev_t *dev = l2hdr->b_dev; 2254 2255 ASSERT(MUTEX_HELD(&dev->l2ad_mtx)); 2256 ASSERT(HDR_HAS_L2HDR(hdr)); 2257 2258 list_remove(&dev->l2ad_buflist, hdr); 2259 2260 /* 2261 * We don't want to leak the b_tmp_cdata buffer that was 2262 * allocated in l2arc_write_buffers() 2263 */ 2264 arc_buf_l2_cdata_free(hdr); 2265 2266 /* 2267 * If the l2hdr's b_daddr is equal to L2ARC_ADDR_UNSET, then 2268 * this header is being processed by l2arc_write_buffers() (i.e. 2269 * it's in the first stage of l2arc_write_buffers()). 2270 * Re-affirming that truth here, just to serve as a reminder. If 2271 * b_daddr does not equal L2ARC_ADDR_UNSET, then the header may or 2272 * may not have its HDR_L2_WRITING flag set. (the write may have 2273 * completed, in which case HDR_L2_WRITING will be false and the 2274 * b_daddr field will point to the address of the buffer on disk). 2275 */ 2276 IMPLY(l2hdr->b_daddr == L2ARC_ADDR_UNSET, HDR_L2_WRITING(hdr)); 2277 2278 /* 2279 * If b_daddr is equal to L2ARC_ADDR_UNSET, we're racing with 2280 * l2arc_write_buffers(). Since we've just removed this header 2281 * from the l2arc buffer list, this header will never reach the 2282 * second stage of l2arc_write_buffers(), which increments the 2283 * accounting stats for this header. Thus, we must be careful 2284 * not to decrement them for this header either. 2285 */ 2286 if (l2hdr->b_daddr != L2ARC_ADDR_UNSET) { 2287 ARCSTAT_INCR(arcstat_l2_asize, -l2hdr->b_asize); 2288 ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size); 2289 2290 vdev_space_update(dev->l2ad_vdev, 2291 -l2hdr->b_asize, 0, 0); 2292 2293 (void) refcount_remove_many(&dev->l2ad_alloc, 2294 l2hdr->b_asize, hdr); 2295 } 2296 2297 hdr->b_flags &= ~ARC_FLAG_HAS_L2HDR; 2298} 2299 2300static void 2301arc_hdr_destroy(arc_buf_hdr_t *hdr) 2302{ 2303 if (HDR_HAS_L1HDR(hdr)) { 2304 ASSERT(hdr->b_l1hdr.b_buf == NULL || 2305 hdr->b_l1hdr.b_datacnt > 0); 2306 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 2307 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon); 2308 } 2309 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 2310 ASSERT(!HDR_IN_HASH_TABLE(hdr)); 2311 2312 if (HDR_HAS_L2HDR(hdr)) { 2313 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev; 2314 boolean_t buflist_held = MUTEX_HELD(&dev->l2ad_mtx); 2315 2316 if (!buflist_held) 2317 mutex_enter(&dev->l2ad_mtx); 2318 2319 /* 2320 * Even though we checked this conditional above, we 2321 * need to check this again now that we have the 2322 * l2ad_mtx. This is because we could be racing with 2323 * another thread calling l2arc_evict() which might have 2324 * destroyed this header's L2 portion as we were waiting 2325 * to acquire the l2ad_mtx. If that happens, we don't 2326 * want to re-destroy the header's L2 portion. 2327 */ 2328 if (HDR_HAS_L2HDR(hdr)) { 2329 if (hdr->b_l2hdr.b_daddr != L2ARC_ADDR_UNSET) 2330 trim_map_free(dev->l2ad_vdev, 2331 hdr->b_l2hdr.b_daddr, 2332 hdr->b_l2hdr.b_asize, 0); 2333 arc_hdr_l2hdr_destroy(hdr); 2334 } 2335 2336 if (!buflist_held) 2337 mutex_exit(&dev->l2ad_mtx); 2338 } 2339 2340 if (!BUF_EMPTY(hdr)) 2341 buf_discard_identity(hdr); 2342 if (hdr->b_freeze_cksum != NULL) { 2343 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 2344 hdr->b_freeze_cksum = NULL; 2345 } 2346 2347 if (HDR_HAS_L1HDR(hdr)) { 2348 while (hdr->b_l1hdr.b_buf) { 2349 arc_buf_t *buf = hdr->b_l1hdr.b_buf; 2350 2351 if (buf->b_efunc != NULL) { 2352 mutex_enter(&arc_eviction_mtx); 2353 mutex_enter(&buf->b_evict_lock); 2354 ASSERT(buf->b_hdr != NULL); 2355 arc_buf_destroy(hdr->b_l1hdr.b_buf, FALSE, 2356 FALSE); 2357 hdr->b_l1hdr.b_buf = buf->b_next; 2358 buf->b_hdr = &arc_eviction_hdr; 2359 buf->b_next = arc_eviction_list; 2360 arc_eviction_list = buf; 2361 mutex_exit(&buf->b_evict_lock); 2362 mutex_exit(&arc_eviction_mtx); 2363 } else { 2364 arc_buf_destroy(hdr->b_l1hdr.b_buf, FALSE, 2365 TRUE); 2366 } 2367 } 2368#ifdef ZFS_DEBUG 2369 if (hdr->b_l1hdr.b_thawed != NULL) { 2370 kmem_free(hdr->b_l1hdr.b_thawed, 1); 2371 hdr->b_l1hdr.b_thawed = NULL; 2372 } 2373#endif 2374 } 2375 2376 ASSERT3P(hdr->b_hash_next, ==, NULL); 2377 if (HDR_HAS_L1HDR(hdr)) { 2378 ASSERT(!list_link_active(&hdr->b_l1hdr.b_arc_node)); 2379 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL); 2380 kmem_cache_free(hdr_full_cache, hdr); 2381 } else { 2382 kmem_cache_free(hdr_l2only_cache, hdr); 2383 } 2384} 2385 2386void 2387arc_buf_free(arc_buf_t *buf, void *tag) 2388{ 2389 arc_buf_hdr_t *hdr = buf->b_hdr; 2390 int hashed = hdr->b_l1hdr.b_state != arc_anon; 2391 2392 ASSERT(buf->b_efunc == NULL); 2393 ASSERT(buf->b_data != NULL); 2394 2395 if (hashed) { 2396 kmutex_t *hash_lock = HDR_LOCK(hdr); 2397 2398 mutex_enter(hash_lock); 2399 hdr = buf->b_hdr; 2400 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 2401 2402 (void) remove_reference(hdr, hash_lock, tag); 2403 if (hdr->b_l1hdr.b_datacnt > 1) { 2404 arc_buf_destroy(buf, FALSE, TRUE); 2405 } else { 2406 ASSERT(buf == hdr->b_l1hdr.b_buf); 2407 ASSERT(buf->b_efunc == NULL); 2408 hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE; 2409 } 2410 mutex_exit(hash_lock); 2411 } else if (HDR_IO_IN_PROGRESS(hdr)) { 2412 int destroy_hdr; 2413 /* 2414 * We are in the middle of an async write. Don't destroy 2415 * this buffer unless the write completes before we finish 2416 * decrementing the reference count. 2417 */ 2418 mutex_enter(&arc_eviction_mtx); 2419 (void) remove_reference(hdr, NULL, tag); 2420 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 2421 destroy_hdr = !HDR_IO_IN_PROGRESS(hdr); 2422 mutex_exit(&arc_eviction_mtx); 2423 if (destroy_hdr) 2424 arc_hdr_destroy(hdr); 2425 } else { 2426 if (remove_reference(hdr, NULL, tag) > 0) 2427 arc_buf_destroy(buf, FALSE, TRUE); 2428 else 2429 arc_hdr_destroy(hdr); 2430 } 2431} 2432 2433boolean_t 2434arc_buf_remove_ref(arc_buf_t *buf, void* tag) 2435{ 2436 arc_buf_hdr_t *hdr = buf->b_hdr; 2437 kmutex_t *hash_lock = HDR_LOCK(hdr); 2438 boolean_t no_callback = (buf->b_efunc == NULL); 2439 2440 if (hdr->b_l1hdr.b_state == arc_anon) { 2441 ASSERT(hdr->b_l1hdr.b_datacnt == 1); 2442 arc_buf_free(buf, tag); 2443 return (no_callback); 2444 } 2445 2446 mutex_enter(hash_lock); 2447 hdr = buf->b_hdr; 2448 ASSERT(hdr->b_l1hdr.b_datacnt > 0); 2449 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 2450 ASSERT(hdr->b_l1hdr.b_state != arc_anon); 2451 ASSERT(buf->b_data != NULL); 2452 2453 (void) remove_reference(hdr, hash_lock, tag); 2454 if (hdr->b_l1hdr.b_datacnt > 1) { 2455 if (no_callback) 2456 arc_buf_destroy(buf, FALSE, TRUE); 2457 } else if (no_callback) { 2458 ASSERT(hdr->b_l1hdr.b_buf == buf && buf->b_next == NULL); 2459 ASSERT(buf->b_efunc == NULL); 2460 hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE; 2461 } 2462 ASSERT(no_callback || hdr->b_l1hdr.b_datacnt > 1 || 2463 refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 2464 mutex_exit(hash_lock); 2465 return (no_callback); 2466} 2467 2468int32_t 2469arc_buf_size(arc_buf_t *buf) 2470{ 2471 return (buf->b_hdr->b_size); 2472} 2473 2474/* 2475 * Called from the DMU to determine if the current buffer should be 2476 * evicted. In order to ensure proper locking, the eviction must be initiated 2477 * from the DMU. Return true if the buffer is associated with user data and 2478 * duplicate buffers still exist. 2479 */ 2480boolean_t 2481arc_buf_eviction_needed(arc_buf_t *buf) 2482{ 2483 arc_buf_hdr_t *hdr; 2484 boolean_t evict_needed = B_FALSE; 2485 2486 if (zfs_disable_dup_eviction) 2487 return (B_FALSE); 2488 2489 mutex_enter(&buf->b_evict_lock); 2490 hdr = buf->b_hdr; 2491 if (hdr == NULL) { 2492 /* 2493 * We are in arc_do_user_evicts(); let that function 2494 * perform the eviction. 2495 */ 2496 ASSERT(buf->b_data == NULL); 2497 mutex_exit(&buf->b_evict_lock); 2498 return (B_FALSE); 2499 } else if (buf->b_data == NULL) { 2500 /* 2501 * We have already been added to the arc eviction list; 2502 * recommend eviction. 2503 */ 2504 ASSERT3P(hdr, ==, &arc_eviction_hdr); 2505 mutex_exit(&buf->b_evict_lock); 2506 return (B_TRUE); 2507 } 2508 2509 if (hdr->b_l1hdr.b_datacnt > 1 && HDR_ISTYPE_DATA(hdr)) 2510 evict_needed = B_TRUE; 2511 2512 mutex_exit(&buf->b_evict_lock); 2513 return (evict_needed); 2514} 2515 2516/* 2517 * Evict buffers from list until we've removed the specified number of 2518 * bytes. Move the removed buffers to the appropriate evict state. 2519 * If the recycle flag is set, then attempt to "recycle" a buffer: 2520 * - look for a buffer to evict that is `bytes' long. 2521 * - return the data block from this buffer rather than freeing it. 2522 * This flag is used by callers that are trying to make space for a 2523 * new buffer in a full arc cache. 2524 * 2525 * This function makes a "best effort". It skips over any buffers 2526 * it can't get a hash_lock on, and so may not catch all candidates. 2527 * It may also return without evicting as much space as requested. 2528 */ 2529static void * 2530arc_evict(arc_state_t *state, uint64_t spa, int64_t bytes, boolean_t recycle, 2531 arc_buf_contents_t type) 2532{ 2533 arc_state_t *evicted_state; 2534 uint64_t bytes_evicted = 0, skipped = 0, missed = 0; 2535 int64_t bytes_remaining; 2536 arc_buf_hdr_t *hdr, *hdr_prev = NULL; 2537 list_t *evicted_list, *list, *evicted_list_start, *list_start; 2538 kmutex_t *lock, *evicted_lock; 2539 kmutex_t *hash_lock; 2540 boolean_t have_lock; 2541 void *stolen = NULL; 2542 arc_buf_hdr_t marker = { 0 }; 2543 int count = 0; 2544 static int evict_metadata_offset, evict_data_offset; 2545 int i, idx, offset, list_count, lists; 2546 2547 ASSERT(state == arc_mru || state == arc_mfu); 2548 2549 evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost; 2550 2551 /* 2552 * Decide which "type" (data vs metadata) to recycle from. 2553 * 2554 * If we are over the metadata limit, recycle from metadata. 2555 * If we are under the metadata minimum, recycle from data. 2556 * Otherwise, recycle from whichever type has the oldest (least 2557 * recently accessed) header. This is not yet implemented. 2558 */ 2559 if (recycle) { 2560 arc_buf_contents_t realtype; 2561 if (state->arcs_lsize[ARC_BUFC_DATA] == 0) { 2562 realtype = ARC_BUFC_METADATA; 2563 } else if (state->arcs_lsize[ARC_BUFC_METADATA] == 0) { 2564 realtype = ARC_BUFC_DATA; 2565 } else if (arc_meta_used >= arc_meta_limit) { 2566 realtype = ARC_BUFC_METADATA; 2567 } else if (arc_meta_used <= arc_meta_min) { 2568 realtype = ARC_BUFC_DATA; 2569#ifdef illumos 2570 } else if (HDR_HAS_L1HDR(data_hdr) && 2571 HDR_HAS_L1HDR(metadata_hdr) && 2572 data_hdr->b_l1hdr.b_arc_access < 2573 metadata_hdr->b_l1hdr.b_arc_access) { 2574 realtype = ARC_BUFC_DATA; 2575 } else { 2576 realtype = ARC_BUFC_METADATA; 2577#else 2578 } else { 2579 /* TODO */ 2580 realtype = type; 2581#endif 2582 } 2583 if (realtype != type) { 2584 /* 2585 * If we want to evict from a different list, 2586 * we can not recycle, because DATA vs METADATA 2587 * buffers are segregated into different kmem 2588 * caches (and vmem arenas). 2589 */ 2590 type = realtype; 2591 recycle = B_FALSE; 2592 } 2593 } 2594 2595 if (type == ARC_BUFC_METADATA) { 2596 offset = 0; 2597 list_count = ARC_BUFC_NUMMETADATALISTS; 2598 list_start = &state->arcs_lists[0]; 2599 evicted_list_start = &evicted_state->arcs_lists[0]; 2600 idx = evict_metadata_offset; 2601 } else { 2602 offset = ARC_BUFC_NUMMETADATALISTS; 2603 list_start = &state->arcs_lists[offset]; 2604 evicted_list_start = &evicted_state->arcs_lists[offset]; 2605 list_count = ARC_BUFC_NUMDATALISTS; 2606 idx = evict_data_offset; 2607 } 2608 bytes_remaining = evicted_state->arcs_lsize[type]; 2609 lists = 0; 2610 2611evict_start: 2612 list = &list_start[idx]; 2613 evicted_list = &evicted_list_start[idx]; 2614 lock = ARCS_LOCK(state, (offset + idx)); 2615 evicted_lock = ARCS_LOCK(evicted_state, (offset + idx)); 2616 2617 /* 2618 * The ghost list lock must be acquired first in order to prevent 2619 * a 3 party deadlock: 2620 * 2621 * - arc_evict_ghost acquires arc_*_ghost->arcs_mtx, followed by 2622 * l2ad_mtx in arc_hdr_realloc 2623 * - l2arc_write_buffers acquires l2ad_mtx, followed by arc_*->arcs_mtx 2624 * - arc_evict acquires arc_*_ghost->arcs_mtx, followed by 2625 * arc_*_ghost->arcs_mtx and forms a deadlock cycle. 2626 * 2627 * This situation is avoided by acquiring the ghost list lock first. 2628 */ 2629 mutex_enter(evicted_lock); 2630 mutex_enter(lock); 2631 2632 for (hdr = list_tail(list); hdr; hdr = hdr_prev) { 2633 hdr_prev = list_prev(list, hdr); 2634 if (HDR_HAS_L1HDR(hdr)) { 2635 bytes_remaining -= 2636 (hdr->b_size * hdr->b_l1hdr.b_datacnt); 2637 } 2638 /* prefetch buffers have a minimum lifespan */ 2639 if (HDR_IO_IN_PROGRESS(hdr) || 2640 (spa && hdr->b_spa != spa) || 2641 ((hdr->b_flags & (ARC_FLAG_PREFETCH | ARC_FLAG_INDIRECT)) && 2642 ddi_get_lbolt() - hdr->b_l1hdr.b_arc_access < 2643 arc_min_prefetch_lifespan)) { 2644 skipped++; 2645 continue; 2646 } 2647 /* "lookahead" for better eviction candidate */ 2648 if (recycle && hdr->b_size != bytes && 2649 hdr_prev && hdr_prev->b_size == bytes) 2650 continue; 2651 2652 /* ignore markers */ 2653 if (hdr->b_spa == 0) 2654 continue; 2655 2656 /* 2657 * It may take a long time to evict all the bufs requested. 2658 * To avoid blocking all arc activity, periodically drop 2659 * the arcs_mtx and give other threads a chance to run 2660 * before reacquiring the lock. 2661 * 2662 * If we are looking for a buffer to recycle, we are in 2663 * the hot code path, so don't sleep. 2664 */ 2665 if (!recycle && count++ > arc_evict_iterations) { 2666 list_insert_after(list, hdr, &marker); 2667 mutex_exit(lock); 2668 mutex_exit(evicted_lock); 2669 kpreempt(KPREEMPT_SYNC); 2670 mutex_enter(evicted_lock); 2671 mutex_enter(lock); 2672 hdr_prev = list_prev(list, &marker); 2673 list_remove(list, &marker); 2674 count = 0; 2675 continue; 2676 } 2677 2678 hash_lock = HDR_LOCK(hdr); 2679 have_lock = MUTEX_HELD(hash_lock); 2680 if (have_lock || mutex_tryenter(hash_lock)) { 2681 ASSERT0(refcount_count(&hdr->b_l1hdr.b_refcnt)); 2682 ASSERT3U(hdr->b_l1hdr.b_datacnt, >, 0); 2683 while (hdr->b_l1hdr.b_buf) { 2684 arc_buf_t *buf = hdr->b_l1hdr.b_buf; 2685 if (!mutex_tryenter(&buf->b_evict_lock)) { 2686 missed += 1; 2687 break; 2688 } 2689 if (buf->b_data != NULL) { 2690 bytes_evicted += hdr->b_size; 2691 if (recycle && 2692 arc_buf_type(hdr) == type && 2693 hdr->b_size == bytes && 2694 !HDR_L2_WRITING(hdr)) { 2695 stolen = buf->b_data; 2696 recycle = FALSE; 2697 } 2698 } 2699 if (buf->b_efunc != NULL) { 2700 mutex_enter(&arc_eviction_mtx); 2701 arc_buf_destroy(buf, 2702 buf->b_data == stolen, FALSE); 2703 hdr->b_l1hdr.b_buf = buf->b_next; 2704 buf->b_hdr = &arc_eviction_hdr; 2705 buf->b_next = arc_eviction_list; 2706 arc_eviction_list = buf; 2707 mutex_exit(&arc_eviction_mtx); 2708 mutex_exit(&buf->b_evict_lock); 2709 } else { 2710 mutex_exit(&buf->b_evict_lock); 2711 arc_buf_destroy(buf, 2712 buf->b_data == stolen, TRUE); 2713 } 2714 } 2715 2716 if (HDR_HAS_L2HDR(hdr)) { 2717 ARCSTAT_INCR(arcstat_evict_l2_cached, 2718 hdr->b_size); 2719 } else { 2720 if (l2arc_write_eligible(hdr->b_spa, hdr)) { 2721 ARCSTAT_INCR(arcstat_evict_l2_eligible, 2722 hdr->b_size); 2723 } else { 2724 ARCSTAT_INCR( 2725 arcstat_evict_l2_ineligible, 2726 hdr->b_size); 2727 } 2728 } 2729 2730 if (hdr->b_l1hdr.b_datacnt == 0) { 2731 arc_change_state(evicted_state, hdr, hash_lock); 2732 ASSERT(HDR_IN_HASH_TABLE(hdr)); 2733 hdr->b_flags |= ARC_FLAG_IN_HASH_TABLE; 2734 hdr->b_flags &= ~ARC_FLAG_BUF_AVAILABLE; 2735 DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, hdr); 2736 } 2737 if (!have_lock) 2738 mutex_exit(hash_lock); 2739 if (bytes >= 0 && bytes_evicted >= bytes) 2740 break; 2741 if (bytes_remaining > 0) { 2742 mutex_exit(evicted_lock); 2743 mutex_exit(lock); 2744 idx = ((idx + 1) & (list_count - 1)); 2745 lists++; 2746 goto evict_start; 2747 } 2748 } else { 2749 missed += 1; 2750 } 2751 } 2752 2753 mutex_exit(lock); 2754 mutex_exit(evicted_lock); 2755 2756 idx = ((idx + 1) & (list_count - 1)); 2757 lists++; 2758 2759 if (bytes_evicted < bytes) { 2760 if (lists < list_count) 2761 goto evict_start; 2762 else 2763 dprintf("only evicted %lld bytes from %x", 2764 (longlong_t)bytes_evicted, state); 2765 } 2766 if (type == ARC_BUFC_METADATA) 2767 evict_metadata_offset = idx; 2768 else 2769 evict_data_offset = idx; 2770 2771 if (skipped) 2772 ARCSTAT_INCR(arcstat_evict_skip, skipped); 2773 2774 if (missed) 2775 ARCSTAT_INCR(arcstat_mutex_miss, missed); 2776 2777 /* 2778 * Note: we have just evicted some data into the ghost state, 2779 * potentially putting the ghost size over the desired size. Rather 2780 * that evicting from the ghost list in this hot code path, leave 2781 * this chore to the arc_reclaim_thread(). 2782 */ 2783 2784 if (stolen) 2785 ARCSTAT_BUMP(arcstat_stolen); 2786 return (stolen); 2787} 2788 2789/* 2790 * Remove buffers from list until we've removed the specified number of 2791 * bytes. Destroy the buffers that are removed. 2792 */ 2793static void 2794arc_evict_ghost(arc_state_t *state, uint64_t spa, int64_t bytes) 2795{ 2796 arc_buf_hdr_t *hdr, *hdr_prev; 2797 arc_buf_hdr_t marker = { 0 }; 2798 list_t *list, *list_start; 2799 kmutex_t *hash_lock, *lock; 2800 uint64_t bytes_deleted = 0; 2801 uint64_t bufs_skipped = 0; 2802 int count = 0; 2803 static int evict_offset; 2804 int list_count, idx = evict_offset; 2805 int offset, lists = 0; 2806 2807 ASSERT(GHOST_STATE(state)); 2808 2809 /* 2810 * data lists come after metadata lists 2811 */ 2812 list_start = &state->arcs_lists[ARC_BUFC_NUMMETADATALISTS]; 2813 list_count = ARC_BUFC_NUMDATALISTS; 2814 offset = ARC_BUFC_NUMMETADATALISTS; 2815 2816evict_start: 2817 list = &list_start[idx]; 2818 lock = ARCS_LOCK(state, idx + offset); 2819 2820 mutex_enter(lock); 2821 for (hdr = list_tail(list); hdr; hdr = hdr_prev) { 2822 hdr_prev = list_prev(list, hdr); 2823 if (arc_buf_type(hdr) >= ARC_BUFC_NUMTYPES) 2824 panic("invalid hdr=%p", (void *)hdr); 2825 if (spa && hdr->b_spa != spa) 2826 continue; 2827 2828 /* ignore markers */ 2829 if (hdr->b_spa == 0) 2830 continue; 2831 2832 hash_lock = HDR_LOCK(hdr); 2833 /* caller may be trying to modify this buffer, skip it */ 2834 if (MUTEX_HELD(hash_lock)) 2835 continue; 2836 2837 /* 2838 * It may take a long time to evict all the bufs requested. 2839 * To avoid blocking all arc activity, periodically drop 2840 * the arcs_mtx and give other threads a chance to run 2841 * before reacquiring the lock. 2842 */ 2843 if (count++ > arc_evict_iterations) { 2844 list_insert_after(list, hdr, &marker); 2845 mutex_exit(lock); 2846 kpreempt(KPREEMPT_SYNC); 2847 mutex_enter(lock); 2848 hdr_prev = list_prev(list, &marker); 2849 list_remove(list, &marker); 2850 count = 0; 2851 continue; 2852 } 2853 if (mutex_tryenter(hash_lock)) { 2854 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 2855 ASSERT(!HDR_HAS_L1HDR(hdr) || 2856 hdr->b_l1hdr.b_buf == NULL); 2857 ARCSTAT_BUMP(arcstat_deleted); 2858 bytes_deleted += hdr->b_size; 2859 2860 if (HDR_HAS_L2HDR(hdr)) { 2861 /* 2862 * This buffer is cached on the 2nd Level ARC; 2863 * don't destroy the header. 2864 */ 2865 arc_change_state(arc_l2c_only, hdr, hash_lock); 2866 /* 2867 * dropping from L1+L2 cached to L2-only, 2868 * realloc to remove the L1 header. 2869 */ 2870 hdr = arc_hdr_realloc(hdr, hdr_full_cache, 2871 hdr_l2only_cache); 2872 mutex_exit(hash_lock); 2873 } else { 2874 arc_change_state(arc_anon, hdr, hash_lock); 2875 mutex_exit(hash_lock); 2876 arc_hdr_destroy(hdr); 2877 } 2878 2879 DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, hdr); 2880 if (bytes >= 0 && bytes_deleted >= bytes) 2881 break; 2882 } else if (bytes < 0) { 2883 /* 2884 * Insert a list marker and then wait for the 2885 * hash lock to become available. Once its 2886 * available, restart from where we left off. 2887 */ 2888 list_insert_after(list, hdr, &marker); 2889 mutex_exit(lock); 2890 mutex_enter(hash_lock); 2891 mutex_exit(hash_lock); 2892 mutex_enter(lock); 2893 hdr_prev = list_prev(list, &marker); 2894 list_remove(list, &marker); 2895 } else { 2896 bufs_skipped += 1; 2897 } 2898 2899 } 2900 mutex_exit(lock); 2901 idx = ((idx + 1) & (ARC_BUFC_NUMDATALISTS - 1)); 2902 lists++; 2903 2904 if (lists < list_count) 2905 goto evict_start; 2906 2907 evict_offset = idx; 2908 if ((uintptr_t)list > (uintptr_t)&state->arcs_lists[ARC_BUFC_NUMMETADATALISTS] && 2909 (bytes < 0 || bytes_deleted < bytes)) { 2910 list_start = &state->arcs_lists[0]; 2911 list_count = ARC_BUFC_NUMMETADATALISTS; 2912 offset = lists = 0; 2913 goto evict_start; 2914 } 2915 2916 if (bufs_skipped) { 2917 ARCSTAT_INCR(arcstat_mutex_miss, bufs_skipped); 2918 ASSERT(bytes >= 0); 2919 } 2920 2921 if (bytes_deleted < bytes) 2922 dprintf("only deleted %lld bytes from %p", 2923 (longlong_t)bytes_deleted, state); 2924} 2925 2926static void 2927arc_adjust(void) 2928{ 2929 int64_t adjustment, delta; 2930 2931 /* 2932 * Adjust MRU size 2933 */ 2934 2935 adjustment = MIN((int64_t)(arc_size - arc_c), 2936 (int64_t)(arc_anon->arcs_size + arc_mru->arcs_size + arc_meta_used - 2937 arc_p)); 2938 2939 if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_DATA] > 0) { 2940 delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_DATA], adjustment); 2941 (void) arc_evict(arc_mru, 0, delta, FALSE, ARC_BUFC_DATA); 2942 adjustment -= delta; 2943 } 2944 2945 if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_METADATA] > 0) { 2946 delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_METADATA], adjustment); 2947 (void) arc_evict(arc_mru, 0, delta, FALSE, 2948 ARC_BUFC_METADATA); 2949 } 2950 2951 /* 2952 * Adjust MFU size 2953 */ 2954 2955 adjustment = arc_size - arc_c; 2956 2957 if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_DATA] > 0) { 2958 delta = MIN(adjustment, arc_mfu->arcs_lsize[ARC_BUFC_DATA]); 2959 (void) arc_evict(arc_mfu, 0, delta, FALSE, ARC_BUFC_DATA); 2960 adjustment -= delta; 2961 } 2962 2963 if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_METADATA] > 0) { 2964 int64_t delta = MIN(adjustment, 2965 arc_mfu->arcs_lsize[ARC_BUFC_METADATA]); 2966 (void) arc_evict(arc_mfu, 0, delta, FALSE, 2967 ARC_BUFC_METADATA); 2968 } 2969 2970 /* 2971 * Adjust ghost lists 2972 */ 2973 2974 adjustment = arc_mru->arcs_size + arc_mru_ghost->arcs_size - arc_c; 2975 2976 if (adjustment > 0 && arc_mru_ghost->arcs_size > 0) { 2977 delta = MIN(arc_mru_ghost->arcs_size, adjustment); 2978 arc_evict_ghost(arc_mru_ghost, 0, delta); 2979 } 2980 2981 adjustment = 2982 arc_mru_ghost->arcs_size + arc_mfu_ghost->arcs_size - arc_c; 2983 2984 if (adjustment > 0 && arc_mfu_ghost->arcs_size > 0) { 2985 delta = MIN(arc_mfu_ghost->arcs_size, adjustment); 2986 arc_evict_ghost(arc_mfu_ghost, 0, delta); 2987 } 2988} 2989 2990static void 2991arc_do_user_evicts(void) 2992{ 2993 static arc_buf_t *tmp_arc_eviction_list; 2994 2995 /* 2996 * Move list over to avoid LOR 2997 */ 2998restart: 2999 mutex_enter(&arc_eviction_mtx); 3000 tmp_arc_eviction_list = arc_eviction_list; 3001 arc_eviction_list = NULL; 3002 mutex_exit(&arc_eviction_mtx); 3003 3004 while (tmp_arc_eviction_list != NULL) { 3005 arc_buf_t *buf = tmp_arc_eviction_list; 3006 tmp_arc_eviction_list = buf->b_next; 3007 mutex_enter(&buf->b_evict_lock); 3008 buf->b_hdr = NULL; 3009 mutex_exit(&buf->b_evict_lock); 3010 3011 if (buf->b_efunc != NULL) 3012 VERIFY0(buf->b_efunc(buf->b_private)); 3013 3014 buf->b_efunc = NULL; 3015 buf->b_private = NULL; 3016 kmem_cache_free(buf_cache, buf); 3017 } 3018 3019 if (arc_eviction_list != NULL) 3020 goto restart; 3021} 3022 3023/* 3024 * Flush all *evictable* data from the cache for the given spa. 3025 * NOTE: this will not touch "active" (i.e. referenced) data. 3026 */ 3027void 3028arc_flush(spa_t *spa) 3029{ 3030 uint64_t guid = 0; 3031 3032 if (spa != NULL) 3033 guid = spa_load_guid(spa); 3034 3035 while (arc_mru->arcs_lsize[ARC_BUFC_DATA]) { 3036 (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_DATA); 3037 if (spa != NULL) 3038 break; 3039 } 3040 while (arc_mru->arcs_lsize[ARC_BUFC_METADATA]) { 3041 (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_METADATA); 3042 if (spa != NULL) 3043 break; 3044 } 3045 while (arc_mfu->arcs_lsize[ARC_BUFC_DATA]) { 3046 (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_DATA); 3047 if (spa != NULL) 3048 break; 3049 } 3050 while (arc_mfu->arcs_lsize[ARC_BUFC_METADATA]) { 3051 (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_METADATA); 3052 if (spa != NULL) 3053 break; 3054 } 3055 3056 arc_evict_ghost(arc_mru_ghost, guid, -1); 3057 arc_evict_ghost(arc_mfu_ghost, guid, -1); 3058 3059 mutex_enter(&arc_reclaim_thr_lock); 3060 arc_do_user_evicts(); 3061 mutex_exit(&arc_reclaim_thr_lock); 3062 ASSERT(spa || arc_eviction_list == NULL); 3063} 3064 3065void 3066arc_shrink(int64_t to_free) 3067{ 3068 3069 if (arc_c > arc_c_min) { 3070 DTRACE_PROBE4(arc__shrink, uint64_t, arc_c, uint64_t, 3071 arc_c_min, uint64_t, arc_p, uint64_t, to_free); 3072 if (arc_c > arc_c_min + to_free) 3073 atomic_add_64(&arc_c, -to_free); 3074 else 3075 arc_c = arc_c_min; 3076 3077 atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift)); 3078 if (arc_c > arc_size) 3079 arc_c = MAX(arc_size, arc_c_min); 3080 if (arc_p > arc_c) 3081 arc_p = (arc_c >> 1); 3082 3083 DTRACE_PROBE2(arc__shrunk, uint64_t, arc_c, uint64_t, 3084 arc_p); 3085 3086 ASSERT(arc_c >= arc_c_min); 3087 ASSERT((int64_t)arc_p >= 0); 3088 } 3089 3090 if (arc_size > arc_c) { 3091 DTRACE_PROBE2(arc__shrink_adjust, uint64_t, arc_size, 3092 uint64_t, arc_c); 3093 arc_adjust(); 3094 } 3095} 3096 3097static long needfree = 0; 3098 3099typedef enum free_memory_reason_t { 3100 FMR_UNKNOWN, 3101 FMR_NEEDFREE, 3102 FMR_LOTSFREE, 3103 FMR_SWAPFS_MINFREE, 3104 FMR_PAGES_PP_MAXIMUM, 3105 FMR_HEAP_ARENA, 3106 FMR_ZIO_ARENA, 3107 FMR_ZIO_FRAG, 3108} free_memory_reason_t; 3109 3110int64_t last_free_memory; 3111free_memory_reason_t last_free_reason; 3112 3113/* 3114 * Additional reserve of pages for pp_reserve. 3115 */ 3116int64_t arc_pages_pp_reserve = 64; 3117 3118/* 3119 * Additional reserve of pages for swapfs. 3120 */ 3121int64_t arc_swapfs_reserve = 64; 3122 3123/* 3124 * Return the amount of memory that can be consumed before reclaim will be 3125 * needed. Positive if there is sufficient free memory, negative indicates 3126 * the amount of memory that needs to be freed up. 3127 */ 3128static int64_t 3129arc_available_memory(void) 3130{ 3131 int64_t lowest = INT64_MAX; 3132 int64_t n; 3133 free_memory_reason_t r = FMR_UNKNOWN; 3134 3135#ifdef _KERNEL 3136 if (needfree > 0) { 3137 n = PAGESIZE * (-needfree); 3138 if (n < lowest) { 3139 lowest = n; 3140 r = FMR_NEEDFREE; 3141 } 3142 } 3143 3144 /* 3145 * Cooperate with pagedaemon when it's time for it to scan 3146 * and reclaim some pages. 3147 */ 3148 n = PAGESIZE * ((int64_t)freemem - zfs_arc_free_target); 3149 if (n < lowest) { 3150 lowest = n; 3151 r = FMR_LOTSFREE; 3152 } 3153 3154#ifdef illumos 3155 /* 3156 * check that we're out of range of the pageout scanner. It starts to 3157 * schedule paging if freemem is less than lotsfree and needfree. 3158 * lotsfree is the high-water mark for pageout, and needfree is the 3159 * number of needed free pages. We add extra pages here to make sure 3160 * the scanner doesn't start up while we're freeing memory. 3161 */ 3162 n = PAGESIZE * (freemem - lotsfree - needfree - desfree); 3163 if (n < lowest) { 3164 lowest = n; 3165 r = FMR_LOTSFREE; 3166 } 3167 3168 /* 3169 * check to make sure that swapfs has enough space so that anon 3170 * reservations can still succeed. anon_resvmem() checks that the 3171 * availrmem is greater than swapfs_minfree, and the number of reserved 3172 * swap pages. We also add a bit of extra here just to prevent 3173 * circumstances from getting really dire. 3174 */ 3175 n = PAGESIZE * (availrmem - swapfs_minfree - swapfs_reserve - 3176 desfree - arc_swapfs_reserve); 3177 if (n < lowest) { 3178 lowest = n; 3179 r = FMR_SWAPFS_MINFREE; 3180 } 3181 3182 3183 /* 3184 * Check that we have enough availrmem that memory locking (e.g., via 3185 * mlock(3C) or memcntl(2)) can still succeed. (pages_pp_maximum 3186 * stores the number of pages that cannot be locked; when availrmem 3187 * drops below pages_pp_maximum, page locking mechanisms such as 3188 * page_pp_lock() will fail.) 3189 */ 3190 n = PAGESIZE * (availrmem - pages_pp_maximum - 3191 arc_pages_pp_reserve); 3192 if (n < lowest) { 3193 lowest = n; 3194 r = FMR_PAGES_PP_MAXIMUM; 3195 } 3196 3197#endif /* illumos */ 3198#if defined(__i386) || !defined(UMA_MD_SMALL_ALLOC) 3199 /* 3200 * If we're on an i386 platform, it's possible that we'll exhaust the 3201 * kernel heap space before we ever run out of available physical 3202 * memory. Most checks of the size of the heap_area compare against 3203 * tune.t_minarmem, which is the minimum available real memory that we 3204 * can have in the system. However, this is generally fixed at 25 pages 3205 * which is so low that it's useless. In this comparison, we seek to 3206 * calculate the total heap-size, and reclaim if more than 3/4ths of the 3207 * heap is allocated. (Or, in the calculation, if less than 1/4th is 3208 * free) 3209 */ 3210 n = (int64_t)vmem_size(heap_arena, VMEM_FREE) - 3211 (vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC) >> 2); 3212 if (n < lowest) { 3213 lowest = n; 3214 r = FMR_HEAP_ARENA; 3215 } 3216#define zio_arena NULL 3217#else 3218#define zio_arena heap_arena 3219#endif 3220 3221 /* 3222 * If zio data pages are being allocated out of a separate heap segment, 3223 * then enforce that the size of available vmem for this arena remains 3224 * above about 1/16th free. 3225 * 3226 * Note: The 1/16th arena free requirement was put in place 3227 * to aggressively evict memory from the arc in order to avoid 3228 * memory fragmentation issues. 3229 */ 3230 if (zio_arena != NULL) { 3231 n = (int64_t)vmem_size(zio_arena, VMEM_FREE) - 3232 (vmem_size(zio_arena, VMEM_ALLOC) >> 4); 3233 if (n < lowest) { 3234 lowest = n; 3235 r = FMR_ZIO_ARENA; 3236 } 3237 } 3238 3239 /* 3240 * Above limits know nothing about real level of KVA fragmentation. 3241 * Start aggressive reclamation if too little sequential KVA left. 3242 */ 3243 if (lowest > 0) { 3244 n = (vmem_size(heap_arena, VMEM_MAXFREE) < zfs_max_recordsize) ? 3245 -((int64_t)vmem_size(heap_arena, VMEM_ALLOC) >> 4) : 3246 INT64_MAX; 3247 if (n < lowest) { 3248 lowest = n; 3249 r = FMR_ZIO_FRAG; 3250 } 3251 } 3252 3253#else /* _KERNEL */ 3254 /* Every 100 calls, free a small amount */ 3255 if (spa_get_random(100) == 0) 3256 lowest = -1024; 3257#endif /* _KERNEL */ 3258 3259 last_free_memory = lowest; 3260 last_free_reason = r; 3261 DTRACE_PROBE2(arc__available_memory, int64_t, lowest, int, r); 3262 return (lowest); 3263} 3264 3265 3266/* 3267 * Determine if the system is under memory pressure and is asking 3268 * to reclaim memory. A return value of TRUE indicates that the system 3269 * is under memory pressure and that the arc should adjust accordingly. 3270 */ 3271static boolean_t 3272arc_reclaim_needed(void) 3273{ 3274 return (arc_available_memory() < 0); 3275} 3276 3277extern kmem_cache_t *zio_buf_cache[]; 3278extern kmem_cache_t *zio_data_buf_cache[]; 3279extern kmem_cache_t *range_seg_cache; 3280 3281static __noinline void 3282arc_kmem_reap_now(void) 3283{ 3284 size_t i; 3285 kmem_cache_t *prev_cache = NULL; 3286 kmem_cache_t *prev_data_cache = NULL; 3287 3288 DTRACE_PROBE(arc__kmem_reap_start); 3289#ifdef _KERNEL 3290 if (arc_meta_used >= arc_meta_limit) { 3291 /* 3292 * We are exceeding our meta-data cache limit. 3293 * Purge some DNLC entries to release holds on meta-data. 3294 */ 3295 dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent); 3296 } 3297#if defined(__i386) 3298 /* 3299 * Reclaim unused memory from all kmem caches. 3300 */ 3301 kmem_reap(); 3302#endif 3303#endif 3304 3305 for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) { 3306 if (zio_buf_cache[i] != prev_cache) { 3307 prev_cache = zio_buf_cache[i]; 3308 kmem_cache_reap_now(zio_buf_cache[i]); 3309 } 3310 if (zio_data_buf_cache[i] != prev_data_cache) { 3311 prev_data_cache = zio_data_buf_cache[i]; 3312 kmem_cache_reap_now(zio_data_buf_cache[i]); 3313 } 3314 } 3315 kmem_cache_reap_now(buf_cache); 3316 kmem_cache_reap_now(hdr_full_cache); 3317 kmem_cache_reap_now(hdr_l2only_cache); 3318 kmem_cache_reap_now(range_seg_cache); 3319 3320#ifdef illumos 3321 if (zio_arena != NULL) { 3322 /* 3323 * Ask the vmem arena to reclaim unused memory from its 3324 * quantum caches. 3325 */ 3326 vmem_qcache_reap(zio_arena); 3327 } 3328#endif 3329 DTRACE_PROBE(arc__kmem_reap_end); 3330} 3331 3332static void 3333arc_reclaim_thread(void *dummy __unused) 3334{ 3335 clock_t growtime = 0; 3336 callb_cpr_t cpr; 3337 3338 CALLB_CPR_INIT(&cpr, &arc_reclaim_thr_lock, callb_generic_cpr, FTAG); 3339 3340 mutex_enter(&arc_reclaim_thr_lock); 3341 while (arc_thread_exit == 0) { 3342 int64_t free_memory = arc_available_memory(); 3343 if (free_memory < 0) { 3344 3345 arc_no_grow = B_TRUE; 3346 arc_warm = B_TRUE; 3347 3348 /* 3349 * Wait at least zfs_grow_retry (default 60) seconds 3350 * before considering growing. 3351 */ 3352 growtime = ddi_get_lbolt() + (arc_grow_retry * hz); 3353 3354 arc_kmem_reap_now(); 3355 3356 /* 3357 * If we are still low on memory, shrink the ARC 3358 * so that we have arc_shrink_min free space. 3359 */ 3360 free_memory = arc_available_memory(); 3361 3362 int64_t to_free = 3363 (arc_c >> arc_shrink_shift) - free_memory; 3364 if (to_free > 0) { 3365#ifdef _KERNEL 3366 to_free = MAX(to_free, ptob(needfree)); 3367#endif 3368 arc_shrink(to_free); 3369 } 3370 } else if (free_memory < arc_c >> arc_no_grow_shift) { 3371 arc_no_grow = B_TRUE; 3372 } else if (ddi_get_lbolt() >= growtime) { 3373 arc_no_grow = B_FALSE; 3374 } 3375 3376 arc_adjust(); 3377 3378 if (arc_eviction_list != NULL) 3379 arc_do_user_evicts(); 3380 3381#ifdef _KERNEL 3382 if (needfree) { 3383 needfree = 0; 3384 wakeup(&needfree); 3385 } 3386#endif 3387 3388 /* 3389 * This is necessary in order for the mdb ::arc dcmd to 3390 * show up to date information. Since the ::arc command 3391 * does not call the kstat's update function, without 3392 * this call, the command may show stale stats for the 3393 * anon, mru, mru_ghost, mfu, and mfu_ghost lists. Even 3394 * with this change, the data might be up to 1 second 3395 * out of date; but that should suffice. The arc_state_t 3396 * structures can be queried directly if more accurate 3397 * information is needed. 3398 */ 3399 if (arc_ksp != NULL) 3400 arc_ksp->ks_update(arc_ksp, KSTAT_READ); 3401 3402 /* block until needed, or one second, whichever is shorter */ 3403 CALLB_CPR_SAFE_BEGIN(&cpr); 3404 (void) cv_timedwait(&arc_reclaim_thr_cv, 3405 &arc_reclaim_thr_lock, hz); 3406 CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_thr_lock); 3407 } 3408 3409 arc_thread_exit = 0; 3410 cv_broadcast(&arc_reclaim_thr_cv); 3411 CALLB_CPR_EXIT(&cpr); /* drops arc_reclaim_thr_lock */ 3412 thread_exit(); 3413} 3414 3415/* 3416 * Adapt arc info given the number of bytes we are trying to add and 3417 * the state that we are comming from. This function is only called 3418 * when we are adding new content to the cache. 3419 */ 3420static void 3421arc_adapt(int bytes, arc_state_t *state) 3422{ 3423 int mult; 3424 uint64_t arc_p_min = (arc_c >> arc_p_min_shift); 3425 3426 if (state == arc_l2c_only) 3427 return; 3428 3429 ASSERT(bytes > 0); 3430 /* 3431 * Adapt the target size of the MRU list: 3432 * - if we just hit in the MRU ghost list, then increase 3433 * the target size of the MRU list. 3434 * - if we just hit in the MFU ghost list, then increase 3435 * the target size of the MFU list by decreasing the 3436 * target size of the MRU list. 3437 */ 3438 if (state == arc_mru_ghost) { 3439 mult = ((arc_mru_ghost->arcs_size >= arc_mfu_ghost->arcs_size) ? 3440 1 : (arc_mfu_ghost->arcs_size/arc_mru_ghost->arcs_size)); 3441 mult = MIN(mult, 10); /* avoid wild arc_p adjustment */ 3442 3443 arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult); 3444 } else if (state == arc_mfu_ghost) { 3445 uint64_t delta; 3446 3447 mult = ((arc_mfu_ghost->arcs_size >= arc_mru_ghost->arcs_size) ? 3448 1 : (arc_mru_ghost->arcs_size/arc_mfu_ghost->arcs_size)); 3449 mult = MIN(mult, 10); 3450 3451 delta = MIN(bytes * mult, arc_p); 3452 arc_p = MAX(arc_p_min, arc_p - delta); 3453 } 3454 ASSERT((int64_t)arc_p >= 0); 3455 3456 if (arc_reclaim_needed()) { 3457 cv_signal(&arc_reclaim_thr_cv); 3458 return; 3459 } 3460 3461 if (arc_no_grow) 3462 return; 3463 3464 if (arc_c >= arc_c_max) 3465 return; 3466 3467 /* 3468 * If we're within (2 * maxblocksize) bytes of the target 3469 * cache size, increment the target cache size 3470 */ 3471 if (arc_size > arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) { 3472 DTRACE_PROBE1(arc__inc_adapt, int, bytes); 3473 atomic_add_64(&arc_c, (int64_t)bytes); 3474 if (arc_c > arc_c_max) 3475 arc_c = arc_c_max; 3476 else if (state == arc_anon) 3477 atomic_add_64(&arc_p, (int64_t)bytes); 3478 if (arc_p > arc_c) 3479 arc_p = arc_c; 3480 } 3481 ASSERT((int64_t)arc_p >= 0); 3482} 3483 3484/* 3485 * Check if the cache has reached its limits and eviction is required 3486 * prior to insert. 3487 */ 3488static int 3489arc_evict_needed(arc_buf_contents_t type) 3490{ 3491 if (type == ARC_BUFC_METADATA && arc_meta_used >= arc_meta_limit) 3492 return (1); 3493 3494 if (arc_reclaim_needed()) 3495 return (1); 3496 3497 return (arc_size > arc_c); 3498} 3499 3500/* 3501 * The buffer, supplied as the first argument, needs a data block. 3502 * So, if we are at cache max, determine which cache should be victimized. 3503 * We have the following cases: 3504 * 3505 * 1. Insert for MRU, p > sizeof(arc_anon + arc_mru) -> 3506 * In this situation if we're out of space, but the resident size of the MFU is 3507 * under the limit, victimize the MFU cache to satisfy this insertion request. 3508 * 3509 * 2. Insert for MRU, p <= sizeof(arc_anon + arc_mru) -> 3510 * Here, we've used up all of the available space for the MRU, so we need to 3511 * evict from our own cache instead. Evict from the set of resident MRU 3512 * entries. 3513 * 3514 * 3. Insert for MFU (c - p) > sizeof(arc_mfu) -> 3515 * c minus p represents the MFU space in the cache, since p is the size of the 3516 * cache that is dedicated to the MRU. In this situation there's still space on 3517 * the MFU side, so the MRU side needs to be victimized. 3518 * 3519 * 4. Insert for MFU (c - p) < sizeof(arc_mfu) -> 3520 * MFU's resident set is consuming more space than it has been allotted. In 3521 * this situation, we must victimize our own cache, the MFU, for this insertion. 3522 */ 3523static void 3524arc_get_data_buf(arc_buf_t *buf) 3525{ 3526 arc_state_t *state = buf->b_hdr->b_l1hdr.b_state; 3527 uint64_t size = buf->b_hdr->b_size; 3528 arc_buf_contents_t type = arc_buf_type(buf->b_hdr); 3529 3530 arc_adapt(size, state); 3531 3532 /* 3533 * We have not yet reached cache maximum size, 3534 * just allocate a new buffer. 3535 */ 3536 if (!arc_evict_needed(type)) { 3537 if (type == ARC_BUFC_METADATA) { 3538 buf->b_data = zio_buf_alloc(size); 3539 arc_space_consume(size, ARC_SPACE_META); 3540 } else { 3541 ASSERT(type == ARC_BUFC_DATA); 3542 buf->b_data = zio_data_buf_alloc(size); 3543 arc_space_consume(size, ARC_SPACE_DATA); 3544 } 3545 goto out; 3546 } 3547 3548 /* 3549 * If we are prefetching from the mfu ghost list, this buffer 3550 * will end up on the mru list; so steal space from there. 3551 */ 3552 if (state == arc_mfu_ghost) 3553 state = HDR_PREFETCH(buf->b_hdr) ? arc_mru : arc_mfu; 3554 else if (state == arc_mru_ghost) 3555 state = arc_mru; 3556 3557 if (state == arc_mru || state == arc_anon) { 3558 uint64_t mru_used = arc_anon->arcs_size + arc_mru->arcs_size; 3559 state = (arc_mfu->arcs_lsize[type] >= size && 3560 arc_p > mru_used) ? arc_mfu : arc_mru; 3561 } else { 3562 /* MFU cases */ 3563 uint64_t mfu_space = arc_c - arc_p; 3564 state = (arc_mru->arcs_lsize[type] >= size && 3565 mfu_space > arc_mfu->arcs_size) ? arc_mru : arc_mfu; 3566 } 3567 if ((buf->b_data = arc_evict(state, 0, size, TRUE, type)) == NULL) { 3568 if (type == ARC_BUFC_METADATA) { 3569 buf->b_data = zio_buf_alloc(size); 3570 arc_space_consume(size, ARC_SPACE_META); 3571 } else { 3572 ASSERT(type == ARC_BUFC_DATA); 3573 buf->b_data = zio_data_buf_alloc(size); 3574 arc_space_consume(size, ARC_SPACE_DATA); 3575 } 3576 ARCSTAT_BUMP(arcstat_recycle_miss); 3577 } 3578 ASSERT(buf->b_data != NULL); 3579out: 3580 /* 3581 * Update the state size. Note that ghost states have a 3582 * "ghost size" and so don't need to be updated. 3583 */ 3584 if (!GHOST_STATE(buf->b_hdr->b_l1hdr.b_state)) { 3585 arc_buf_hdr_t *hdr = buf->b_hdr; 3586 3587 atomic_add_64(&hdr->b_l1hdr.b_state->arcs_size, size); 3588 if (list_link_active(&hdr->b_l1hdr.b_arc_node)) { 3589 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 3590 atomic_add_64(&hdr->b_l1hdr.b_state->arcs_lsize[type], 3591 size); 3592 } 3593 /* 3594 * If we are growing the cache, and we are adding anonymous 3595 * data, and we have outgrown arc_p, update arc_p 3596 */ 3597 if (arc_size < arc_c && hdr->b_l1hdr.b_state == arc_anon && 3598 arc_anon->arcs_size + arc_mru->arcs_size > arc_p) 3599 arc_p = MIN(arc_c, arc_p + size); 3600 } 3601 ARCSTAT_BUMP(arcstat_allocated); 3602} 3603 3604/* 3605 * This routine is called whenever a buffer is accessed. 3606 * NOTE: the hash lock is dropped in this function. 3607 */ 3608static void 3609arc_access(arc_buf_hdr_t *hdr, kmutex_t *hash_lock) 3610{ 3611 clock_t now; 3612 3613 ASSERT(MUTEX_HELD(hash_lock)); 3614 ASSERT(HDR_HAS_L1HDR(hdr)); 3615 3616 if (hdr->b_l1hdr.b_state == arc_anon) { 3617 /* 3618 * This buffer is not in the cache, and does not 3619 * appear in our "ghost" list. Add the new buffer 3620 * to the MRU state. 3621 */ 3622 3623 ASSERT0(hdr->b_l1hdr.b_arc_access); 3624 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 3625 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr); 3626 arc_change_state(arc_mru, hdr, hash_lock); 3627 3628 } else if (hdr->b_l1hdr.b_state == arc_mru) { 3629 now = ddi_get_lbolt(); 3630 3631 /* 3632 * If this buffer is here because of a prefetch, then either: 3633 * - clear the flag if this is a "referencing" read 3634 * (any subsequent access will bump this into the MFU state). 3635 * or 3636 * - move the buffer to the head of the list if this is 3637 * another prefetch (to make it less likely to be evicted). 3638 */ 3639 if (HDR_PREFETCH(hdr)) { 3640 if (refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) { 3641 ASSERT(list_link_active( 3642 &hdr->b_l1hdr.b_arc_node)); 3643 } else { 3644 hdr->b_flags &= ~ARC_FLAG_PREFETCH; 3645 ARCSTAT_BUMP(arcstat_mru_hits); 3646 } 3647 hdr->b_l1hdr.b_arc_access = now; 3648 return; 3649 } 3650 3651 /* 3652 * This buffer has been "accessed" only once so far, 3653 * but it is still in the cache. Move it to the MFU 3654 * state. 3655 */ 3656 if (now > hdr->b_l1hdr.b_arc_access + ARC_MINTIME) { 3657 /* 3658 * More than 125ms have passed since we 3659 * instantiated this buffer. Move it to the 3660 * most frequently used state. 3661 */ 3662 hdr->b_l1hdr.b_arc_access = now; 3663 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 3664 arc_change_state(arc_mfu, hdr, hash_lock); 3665 } 3666 ARCSTAT_BUMP(arcstat_mru_hits); 3667 } else if (hdr->b_l1hdr.b_state == arc_mru_ghost) { 3668 arc_state_t *new_state; 3669 /* 3670 * This buffer has been "accessed" recently, but 3671 * was evicted from the cache. Move it to the 3672 * MFU state. 3673 */ 3674 3675 if (HDR_PREFETCH(hdr)) { 3676 new_state = arc_mru; 3677 if (refcount_count(&hdr->b_l1hdr.b_refcnt) > 0) 3678 hdr->b_flags &= ~ARC_FLAG_PREFETCH; 3679 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr); 3680 } else { 3681 new_state = arc_mfu; 3682 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 3683 } 3684 3685 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 3686 arc_change_state(new_state, hdr, hash_lock); 3687 3688 ARCSTAT_BUMP(arcstat_mru_ghost_hits); 3689 } else if (hdr->b_l1hdr.b_state == arc_mfu) { 3690 /* 3691 * This buffer has been accessed more than once and is 3692 * still in the cache. Keep it in the MFU state. 3693 * 3694 * NOTE: an add_reference() that occurred when we did 3695 * the arc_read() will have kicked this off the list. 3696 * If it was a prefetch, we will explicitly move it to 3697 * the head of the list now. 3698 */ 3699 if ((HDR_PREFETCH(hdr)) != 0) { 3700 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 3701 ASSERT(list_link_active(&hdr->b_l1hdr.b_arc_node)); 3702 } 3703 ARCSTAT_BUMP(arcstat_mfu_hits); 3704 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 3705 } else if (hdr->b_l1hdr.b_state == arc_mfu_ghost) { 3706 arc_state_t *new_state = arc_mfu; 3707 /* 3708 * This buffer has been accessed more than once but has 3709 * been evicted from the cache. Move it back to the 3710 * MFU state. 3711 */ 3712 3713 if (HDR_PREFETCH(hdr)) { 3714 /* 3715 * This is a prefetch access... 3716 * move this block back to the MRU state. 3717 */ 3718 ASSERT0(refcount_count(&hdr->b_l1hdr.b_refcnt)); 3719 new_state = arc_mru; 3720 } 3721 3722 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 3723 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 3724 arc_change_state(new_state, hdr, hash_lock); 3725 3726 ARCSTAT_BUMP(arcstat_mfu_ghost_hits); 3727 } else if (hdr->b_l1hdr.b_state == arc_l2c_only) { 3728 /* 3729 * This buffer is on the 2nd Level ARC. 3730 */ 3731 3732 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 3733 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 3734 arc_change_state(arc_mfu, hdr, hash_lock); 3735 } else { 3736 ASSERT(!"invalid arc state"); 3737 } 3738} 3739 3740/* a generic arc_done_func_t which you can use */ 3741/* ARGSUSED */ 3742void 3743arc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg) 3744{ 3745 if (zio == NULL || zio->io_error == 0) 3746 bcopy(buf->b_data, arg, buf->b_hdr->b_size); 3747 VERIFY(arc_buf_remove_ref(buf, arg)); 3748} 3749 3750/* a generic arc_done_func_t */ 3751void 3752arc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg) 3753{ 3754 arc_buf_t **bufp = arg; 3755 if (zio && zio->io_error) { 3756 VERIFY(arc_buf_remove_ref(buf, arg)); 3757 *bufp = NULL; 3758 } else { 3759 *bufp = buf; 3760 ASSERT(buf->b_data); 3761 } 3762} 3763 3764static void 3765arc_read_done(zio_t *zio) 3766{ 3767 arc_buf_hdr_t *hdr; 3768 arc_buf_t *buf; 3769 arc_buf_t *abuf; /* buffer we're assigning to callback */ 3770 kmutex_t *hash_lock = NULL; 3771 arc_callback_t *callback_list, *acb; 3772 int freeable = FALSE; 3773 3774 buf = zio->io_private; 3775 hdr = buf->b_hdr; 3776 3777 /* 3778 * The hdr was inserted into hash-table and removed from lists 3779 * prior to starting I/O. We should find this header, since 3780 * it's in the hash table, and it should be legit since it's 3781 * not possible to evict it during the I/O. The only possible 3782 * reason for it not to be found is if we were freed during the 3783 * read. 3784 */ 3785 if (HDR_IN_HASH_TABLE(hdr)) { 3786 ASSERT3U(hdr->b_birth, ==, BP_PHYSICAL_BIRTH(zio->io_bp)); 3787 ASSERT3U(hdr->b_dva.dva_word[0], ==, 3788 BP_IDENTITY(zio->io_bp)->dva_word[0]); 3789 ASSERT3U(hdr->b_dva.dva_word[1], ==, 3790 BP_IDENTITY(zio->io_bp)->dva_word[1]); 3791 3792 arc_buf_hdr_t *found = buf_hash_find(hdr->b_spa, zio->io_bp, 3793 &hash_lock); 3794 3795 ASSERT((found == NULL && HDR_FREED_IN_READ(hdr) && 3796 hash_lock == NULL) || 3797 (found == hdr && 3798 DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) || 3799 (found == hdr && HDR_L2_READING(hdr))); 3800 } 3801 3802 hdr->b_flags &= ~ARC_FLAG_L2_EVICTED; 3803 if (l2arc_noprefetch && HDR_PREFETCH(hdr)) 3804 hdr->b_flags &= ~ARC_FLAG_L2CACHE; 3805 3806 /* byteswap if necessary */ 3807 callback_list = hdr->b_l1hdr.b_acb; 3808 ASSERT(callback_list != NULL); 3809 if (BP_SHOULD_BYTESWAP(zio->io_bp) && zio->io_error == 0) { 3810 dmu_object_byteswap_t bswap = 3811 DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp)); 3812 arc_byteswap_func_t *func = BP_GET_LEVEL(zio->io_bp) > 0 ? 3813 byteswap_uint64_array : 3814 dmu_ot_byteswap[bswap].ob_func; 3815 func(buf->b_data, hdr->b_size); 3816 } 3817 3818 arc_cksum_compute(buf, B_FALSE); 3819#ifdef illumos 3820 arc_buf_watch(buf); 3821#endif 3822 3823 if (hash_lock && zio->io_error == 0 && 3824 hdr->b_l1hdr.b_state == arc_anon) { 3825 /* 3826 * Only call arc_access on anonymous buffers. This is because 3827 * if we've issued an I/O for an evicted buffer, we've already 3828 * called arc_access (to prevent any simultaneous readers from 3829 * getting confused). 3830 */ 3831 arc_access(hdr, hash_lock); 3832 } 3833 3834 /* create copies of the data buffer for the callers */ 3835 abuf = buf; 3836 for (acb = callback_list; acb; acb = acb->acb_next) { 3837 if (acb->acb_done) { 3838 if (abuf == NULL) { 3839 ARCSTAT_BUMP(arcstat_duplicate_reads); 3840 abuf = arc_buf_clone(buf); 3841 } 3842 acb->acb_buf = abuf; 3843 abuf = NULL; 3844 } 3845 } 3846 hdr->b_l1hdr.b_acb = NULL; 3847 hdr->b_flags &= ~ARC_FLAG_IO_IN_PROGRESS; 3848 ASSERT(!HDR_BUF_AVAILABLE(hdr)); 3849 if (abuf == buf) { 3850 ASSERT(buf->b_efunc == NULL); 3851 ASSERT(hdr->b_l1hdr.b_datacnt == 1); 3852 hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE; 3853 } 3854 3855 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt) || 3856 callback_list != NULL); 3857 3858 if (zio->io_error != 0) { 3859 hdr->b_flags |= ARC_FLAG_IO_ERROR; 3860 if (hdr->b_l1hdr.b_state != arc_anon) 3861 arc_change_state(arc_anon, hdr, hash_lock); 3862 if (HDR_IN_HASH_TABLE(hdr)) 3863 buf_hash_remove(hdr); 3864 freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt); 3865 } 3866 3867 /* 3868 * Broadcast before we drop the hash_lock to avoid the possibility 3869 * that the hdr (and hence the cv) might be freed before we get to 3870 * the cv_broadcast(). 3871 */ 3872 cv_broadcast(&hdr->b_l1hdr.b_cv); 3873 3874 if (hash_lock != NULL) { 3875 mutex_exit(hash_lock); 3876 } else { 3877 /* 3878 * This block was freed while we waited for the read to 3879 * complete. It has been removed from the hash table and 3880 * moved to the anonymous state (so that it won't show up 3881 * in the cache). 3882 */ 3883 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon); 3884 freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt); 3885 } 3886 3887 /* execute each callback and free its structure */ 3888 while ((acb = callback_list) != NULL) { 3889 if (acb->acb_done) 3890 acb->acb_done(zio, acb->acb_buf, acb->acb_private); 3891 3892 if (acb->acb_zio_dummy != NULL) { 3893 acb->acb_zio_dummy->io_error = zio->io_error; 3894 zio_nowait(acb->acb_zio_dummy); 3895 } 3896 3897 callback_list = acb->acb_next; 3898 kmem_free(acb, sizeof (arc_callback_t)); 3899 } 3900 3901 if (freeable) 3902 arc_hdr_destroy(hdr); 3903} 3904 3905/* 3906 * "Read" the block block at the specified DVA (in bp) via the 3907 * cache. If the block is found in the cache, invoke the provided 3908 * callback immediately and return. Note that the `zio' parameter 3909 * in the callback will be NULL in this case, since no IO was 3910 * required. If the block is not in the cache pass the read request 3911 * on to the spa with a substitute callback function, so that the 3912 * requested block will be added to the cache. 3913 * 3914 * If a read request arrives for a block that has a read in-progress, 3915 * either wait for the in-progress read to complete (and return the 3916 * results); or, if this is a read with a "done" func, add a record 3917 * to the read to invoke the "done" func when the read completes, 3918 * and return; or just return. 3919 * 3920 * arc_read_done() will invoke all the requested "done" functions 3921 * for readers of this block. 3922 */ 3923int 3924arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_done_func_t *done, 3925 void *private, zio_priority_t priority, int zio_flags, 3926 arc_flags_t *arc_flags, const zbookmark_phys_t *zb) 3927{ 3928 arc_buf_hdr_t *hdr = NULL; 3929 arc_buf_t *buf = NULL; 3930 kmutex_t *hash_lock = NULL; 3931 zio_t *rzio; 3932 uint64_t guid = spa_load_guid(spa); 3933 3934 ASSERT(!BP_IS_EMBEDDED(bp) || 3935 BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA); 3936 3937top: 3938 if (!BP_IS_EMBEDDED(bp)) { 3939 /* 3940 * Embedded BP's have no DVA and require no I/O to "read". 3941 * Create an anonymous arc buf to back it. 3942 */ 3943 hdr = buf_hash_find(guid, bp, &hash_lock); 3944 } 3945 3946 if (hdr != NULL && HDR_HAS_L1HDR(hdr) && hdr->b_l1hdr.b_datacnt > 0) { 3947 3948 *arc_flags |= ARC_FLAG_CACHED; 3949 3950 if (HDR_IO_IN_PROGRESS(hdr)) { 3951 3952 if (*arc_flags & ARC_FLAG_WAIT) { 3953 cv_wait(&hdr->b_l1hdr.b_cv, hash_lock); 3954 mutex_exit(hash_lock); 3955 goto top; 3956 } 3957 ASSERT(*arc_flags & ARC_FLAG_NOWAIT); 3958 3959 if (done) { 3960 arc_callback_t *acb = NULL; 3961 3962 acb = kmem_zalloc(sizeof (arc_callback_t), 3963 KM_SLEEP); 3964 acb->acb_done = done; 3965 acb->acb_private = private; 3966 if (pio != NULL) 3967 acb->acb_zio_dummy = zio_null(pio, 3968 spa, NULL, NULL, NULL, zio_flags); 3969 3970 ASSERT(acb->acb_done != NULL); 3971 acb->acb_next = hdr->b_l1hdr.b_acb; 3972 hdr->b_l1hdr.b_acb = acb; 3973 add_reference(hdr, hash_lock, private); 3974 mutex_exit(hash_lock); 3975 return (0); 3976 } 3977 mutex_exit(hash_lock); 3978 return (0); 3979 } 3980 3981 ASSERT(hdr->b_l1hdr.b_state == arc_mru || 3982 hdr->b_l1hdr.b_state == arc_mfu); 3983 3984 if (done) { 3985 add_reference(hdr, hash_lock, private); 3986 /* 3987 * If this block is already in use, create a new 3988 * copy of the data so that we will be guaranteed 3989 * that arc_release() will always succeed. 3990 */ 3991 buf = hdr->b_l1hdr.b_buf; 3992 ASSERT(buf); 3993 ASSERT(buf->b_data); 3994 if (HDR_BUF_AVAILABLE(hdr)) { 3995 ASSERT(buf->b_efunc == NULL); 3996 hdr->b_flags &= ~ARC_FLAG_BUF_AVAILABLE; 3997 } else { 3998 buf = arc_buf_clone(buf); 3999 } 4000 4001 } else if (*arc_flags & ARC_FLAG_PREFETCH && 4002 refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) { 4003 hdr->b_flags |= ARC_FLAG_PREFETCH; 4004 } 4005 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr); 4006 arc_access(hdr, hash_lock); 4007 if (*arc_flags & ARC_FLAG_L2CACHE) 4008 hdr->b_flags |= ARC_FLAG_L2CACHE; 4009 if (*arc_flags & ARC_FLAG_L2COMPRESS) 4010 hdr->b_flags |= ARC_FLAG_L2COMPRESS; 4011 mutex_exit(hash_lock); 4012 ARCSTAT_BUMP(arcstat_hits); 4013 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr), 4014 demand, prefetch, !HDR_ISTYPE_METADATA(hdr), 4015 data, metadata, hits); 4016 4017 if (done) 4018 done(NULL, buf, private); 4019 } else { 4020 uint64_t size = BP_GET_LSIZE(bp); 4021 arc_callback_t *acb; 4022 vdev_t *vd = NULL; 4023 uint64_t addr = 0; 4024 boolean_t devw = B_FALSE; 4025 enum zio_compress b_compress = ZIO_COMPRESS_OFF; 4026 int32_t b_asize = 0; 4027 4028 if (hdr == NULL) { 4029 /* this block is not in the cache */ 4030 arc_buf_hdr_t *exists = NULL; 4031 arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp); 4032 buf = arc_buf_alloc(spa, size, private, type); 4033 hdr = buf->b_hdr; 4034 if (!BP_IS_EMBEDDED(bp)) { 4035 hdr->b_dva = *BP_IDENTITY(bp); 4036 hdr->b_birth = BP_PHYSICAL_BIRTH(bp); 4037 exists = buf_hash_insert(hdr, &hash_lock); 4038 } 4039 if (exists != NULL) { 4040 /* somebody beat us to the hash insert */ 4041 mutex_exit(hash_lock); 4042 buf_discard_identity(hdr); 4043 (void) arc_buf_remove_ref(buf, private); 4044 goto top; /* restart the IO request */ 4045 } 4046 4047 /* if this is a prefetch, we don't have a reference */ 4048 if (*arc_flags & ARC_FLAG_PREFETCH) { 4049 (void) remove_reference(hdr, hash_lock, 4050 private); 4051 hdr->b_flags |= ARC_FLAG_PREFETCH; 4052 } 4053 if (*arc_flags & ARC_FLAG_L2CACHE) 4054 hdr->b_flags |= ARC_FLAG_L2CACHE; 4055 if (*arc_flags & ARC_FLAG_L2COMPRESS) 4056 hdr->b_flags |= ARC_FLAG_L2COMPRESS; 4057 if (BP_GET_LEVEL(bp) > 0) 4058 hdr->b_flags |= ARC_FLAG_INDIRECT; 4059 } else { 4060 /* 4061 * This block is in the ghost cache. If it was L2-only 4062 * (and thus didn't have an L1 hdr), we realloc the 4063 * header to add an L1 hdr. 4064 */ 4065 if (!HDR_HAS_L1HDR(hdr)) { 4066 hdr = arc_hdr_realloc(hdr, hdr_l2only_cache, 4067 hdr_full_cache); 4068 } 4069 4070 ASSERT(GHOST_STATE(hdr->b_l1hdr.b_state)); 4071 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 4072 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 4073 ASSERT(hdr->b_l1hdr.b_buf == NULL); 4074 4075 /* if this is a prefetch, we don't have a reference */ 4076 if (*arc_flags & ARC_FLAG_PREFETCH) 4077 hdr->b_flags |= ARC_FLAG_PREFETCH; 4078 else 4079 add_reference(hdr, hash_lock, private); 4080 if (*arc_flags & ARC_FLAG_L2CACHE) 4081 hdr->b_flags |= ARC_FLAG_L2CACHE; 4082 if (*arc_flags & ARC_FLAG_L2COMPRESS) 4083 hdr->b_flags |= ARC_FLAG_L2COMPRESS; 4084 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 4085 buf->b_hdr = hdr; 4086 buf->b_data = NULL; 4087 buf->b_efunc = NULL; 4088 buf->b_private = NULL; 4089 buf->b_next = NULL; 4090 hdr->b_l1hdr.b_buf = buf; 4091 ASSERT0(hdr->b_l1hdr.b_datacnt); 4092 hdr->b_l1hdr.b_datacnt = 1; 4093 arc_get_data_buf(buf); 4094 arc_access(hdr, hash_lock); 4095 } 4096 4097 ASSERT(!GHOST_STATE(hdr->b_l1hdr.b_state)); 4098 4099 acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP); 4100 acb->acb_done = done; 4101 acb->acb_private = private; 4102 4103 ASSERT(hdr->b_l1hdr.b_acb == NULL); 4104 hdr->b_l1hdr.b_acb = acb; 4105 hdr->b_flags |= ARC_FLAG_IO_IN_PROGRESS; 4106 4107 if (HDR_HAS_L2HDR(hdr) && 4108 (vd = hdr->b_l2hdr.b_dev->l2ad_vdev) != NULL) { 4109 devw = hdr->b_l2hdr.b_dev->l2ad_writing; 4110 addr = hdr->b_l2hdr.b_daddr; 4111 b_compress = HDR_GET_COMPRESS(hdr); 4112 b_asize = hdr->b_l2hdr.b_asize; 4113 /* 4114 * Lock out device removal. 4115 */ 4116 if (vdev_is_dead(vd) || 4117 !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER)) 4118 vd = NULL; 4119 } 4120 4121 if (hash_lock != NULL) 4122 mutex_exit(hash_lock); 4123 4124 /* 4125 * At this point, we have a level 1 cache miss. Try again in 4126 * L2ARC if possible. 4127 */ 4128 ASSERT3U(hdr->b_size, ==, size); 4129 DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp, 4130 uint64_t, size, zbookmark_phys_t *, zb); 4131 ARCSTAT_BUMP(arcstat_misses); 4132 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr), 4133 demand, prefetch, !HDR_ISTYPE_METADATA(hdr), 4134 data, metadata, misses); 4135#ifdef _KERNEL 4136 curthread->td_ru.ru_inblock++; 4137#endif 4138 4139 if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) { 4140 /* 4141 * Read from the L2ARC if the following are true: 4142 * 1. The L2ARC vdev was previously cached. 4143 * 2. This buffer still has L2ARC metadata. 4144 * 3. This buffer isn't currently writing to the L2ARC. 4145 * 4. The L2ARC entry wasn't evicted, which may 4146 * also have invalidated the vdev. 4147 * 5. This isn't prefetch and l2arc_noprefetch is set. 4148 */ 4149 if (HDR_HAS_L2HDR(hdr) && 4150 !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) && 4151 !(l2arc_noprefetch && HDR_PREFETCH(hdr))) { 4152 l2arc_read_callback_t *cb; 4153 4154 DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr); 4155 ARCSTAT_BUMP(arcstat_l2_hits); 4156 4157 cb = kmem_zalloc(sizeof (l2arc_read_callback_t), 4158 KM_SLEEP); 4159 cb->l2rcb_buf = buf; 4160 cb->l2rcb_spa = spa; 4161 cb->l2rcb_bp = *bp; 4162 cb->l2rcb_zb = *zb; 4163 cb->l2rcb_flags = zio_flags; 4164 cb->l2rcb_compress = b_compress; 4165 4166 ASSERT(addr >= VDEV_LABEL_START_SIZE && 4167 addr + size < vd->vdev_psize - 4168 VDEV_LABEL_END_SIZE); 4169 4170 /* 4171 * l2arc read. The SCL_L2ARC lock will be 4172 * released by l2arc_read_done(). 4173 * Issue a null zio if the underlying buffer 4174 * was squashed to zero size by compression. 4175 */ 4176 if (b_compress == ZIO_COMPRESS_EMPTY) { 4177 rzio = zio_null(pio, spa, vd, 4178 l2arc_read_done, cb, 4179 zio_flags | ZIO_FLAG_DONT_CACHE | 4180 ZIO_FLAG_CANFAIL | 4181 ZIO_FLAG_DONT_PROPAGATE | 4182 ZIO_FLAG_DONT_RETRY); 4183 } else { 4184 rzio = zio_read_phys(pio, vd, addr, 4185 b_asize, buf->b_data, 4186 ZIO_CHECKSUM_OFF, 4187 l2arc_read_done, cb, priority, 4188 zio_flags | ZIO_FLAG_DONT_CACHE | 4189 ZIO_FLAG_CANFAIL | 4190 ZIO_FLAG_DONT_PROPAGATE | 4191 ZIO_FLAG_DONT_RETRY, B_FALSE); 4192 } 4193 DTRACE_PROBE2(l2arc__read, vdev_t *, vd, 4194 zio_t *, rzio); 4195 ARCSTAT_INCR(arcstat_l2_read_bytes, b_asize); 4196 4197 if (*arc_flags & ARC_FLAG_NOWAIT) { 4198 zio_nowait(rzio); 4199 return (0); 4200 } 4201 4202 ASSERT(*arc_flags & ARC_FLAG_WAIT); 4203 if (zio_wait(rzio) == 0) 4204 return (0); 4205 4206 /* l2arc read error; goto zio_read() */ 4207 } else { 4208 DTRACE_PROBE1(l2arc__miss, 4209 arc_buf_hdr_t *, hdr); 4210 ARCSTAT_BUMP(arcstat_l2_misses); 4211 if (HDR_L2_WRITING(hdr)) 4212 ARCSTAT_BUMP(arcstat_l2_rw_clash); 4213 spa_config_exit(spa, SCL_L2ARC, vd); 4214 } 4215 } else { 4216 if (vd != NULL) 4217 spa_config_exit(spa, SCL_L2ARC, vd); 4218 if (l2arc_ndev != 0) { 4219 DTRACE_PROBE1(l2arc__miss, 4220 arc_buf_hdr_t *, hdr); 4221 ARCSTAT_BUMP(arcstat_l2_misses); 4222 } 4223 } 4224 4225 rzio = zio_read(pio, spa, bp, buf->b_data, size, 4226 arc_read_done, buf, priority, zio_flags, zb); 4227 4228 if (*arc_flags & ARC_FLAG_WAIT) 4229 return (zio_wait(rzio)); 4230 4231 ASSERT(*arc_flags & ARC_FLAG_NOWAIT); 4232 zio_nowait(rzio); 4233 } 4234 return (0); 4235} 4236 4237void 4238arc_set_callback(arc_buf_t *buf, arc_evict_func_t *func, void *private) 4239{ 4240 ASSERT(buf->b_hdr != NULL); 4241 ASSERT(buf->b_hdr->b_l1hdr.b_state != arc_anon); 4242 ASSERT(!refcount_is_zero(&buf->b_hdr->b_l1hdr.b_refcnt) || 4243 func == NULL); 4244 ASSERT(buf->b_efunc == NULL); 4245 ASSERT(!HDR_BUF_AVAILABLE(buf->b_hdr)); 4246 4247 buf->b_efunc = func; 4248 buf->b_private = private; 4249} 4250 4251/* 4252 * Notify the arc that a block was freed, and thus will never be used again. 4253 */ 4254void 4255arc_freed(spa_t *spa, const blkptr_t *bp) 4256{ 4257 arc_buf_hdr_t *hdr; 4258 kmutex_t *hash_lock; 4259 uint64_t guid = spa_load_guid(spa); 4260 4261 ASSERT(!BP_IS_EMBEDDED(bp)); 4262 4263 hdr = buf_hash_find(guid, bp, &hash_lock); 4264 if (hdr == NULL) 4265 return; 4266 if (HDR_BUF_AVAILABLE(hdr)) { 4267 arc_buf_t *buf = hdr->b_l1hdr.b_buf; 4268 add_reference(hdr, hash_lock, FTAG); 4269 hdr->b_flags &= ~ARC_FLAG_BUF_AVAILABLE; 4270 mutex_exit(hash_lock); 4271 4272 arc_release(buf, FTAG); 4273 (void) arc_buf_remove_ref(buf, FTAG); 4274 } else { 4275 mutex_exit(hash_lock); 4276 } 4277 4278} 4279 4280/* 4281 * Clear the user eviction callback set by arc_set_callback(), first calling 4282 * it if it exists. Because the presence of a callback keeps an arc_buf cached 4283 * clearing the callback may result in the arc_buf being destroyed. However, 4284 * it will not result in the *last* arc_buf being destroyed, hence the data 4285 * will remain cached in the ARC. We make a copy of the arc buffer here so 4286 * that we can process the callback without holding any locks. 4287 * 4288 * It's possible that the callback is already in the process of being cleared 4289 * by another thread. In this case we can not clear the callback. 4290 * 4291 * Returns B_TRUE if the callback was successfully called and cleared. 4292 */ 4293boolean_t 4294arc_clear_callback(arc_buf_t *buf) 4295{ 4296 arc_buf_hdr_t *hdr; 4297 kmutex_t *hash_lock; 4298 arc_evict_func_t *efunc = buf->b_efunc; 4299 void *private = buf->b_private; 4300 list_t *list, *evicted_list; 4301 kmutex_t *lock, *evicted_lock; 4302 4303 mutex_enter(&buf->b_evict_lock); 4304 hdr = buf->b_hdr; 4305 if (hdr == NULL) { 4306 /* 4307 * We are in arc_do_user_evicts(). 4308 */ 4309 ASSERT(buf->b_data == NULL); 4310 mutex_exit(&buf->b_evict_lock); 4311 return (B_FALSE); 4312 } else if (buf->b_data == NULL) { 4313 /* 4314 * We are on the eviction list; process this buffer now 4315 * but let arc_do_user_evicts() do the reaping. 4316 */ 4317 buf->b_efunc = NULL; 4318 mutex_exit(&buf->b_evict_lock); 4319 VERIFY0(efunc(private)); 4320 return (B_TRUE); 4321 } 4322 hash_lock = HDR_LOCK(hdr); 4323 mutex_enter(hash_lock); 4324 hdr = buf->b_hdr; 4325 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 4326 4327 ASSERT3U(refcount_count(&hdr->b_l1hdr.b_refcnt), <, 4328 hdr->b_l1hdr.b_datacnt); 4329 ASSERT(hdr->b_l1hdr.b_state == arc_mru || 4330 hdr->b_l1hdr.b_state == arc_mfu); 4331 4332 buf->b_efunc = NULL; 4333 buf->b_private = NULL; 4334 4335 if (hdr->b_l1hdr.b_datacnt > 1) { 4336 mutex_exit(&buf->b_evict_lock); 4337 arc_buf_destroy(buf, FALSE, TRUE); 4338 } else { 4339 ASSERT(buf == hdr->b_l1hdr.b_buf); 4340 hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE; 4341 mutex_exit(&buf->b_evict_lock); 4342 } 4343 4344 mutex_exit(hash_lock); 4345 VERIFY0(efunc(private)); 4346 return (B_TRUE); 4347} 4348 4349/* 4350 * Release this buffer from the cache, making it an anonymous buffer. This 4351 * must be done after a read and prior to modifying the buffer contents. 4352 * If the buffer has more than one reference, we must make 4353 * a new hdr for the buffer. 4354 */ 4355void 4356arc_release(arc_buf_t *buf, void *tag) 4357{ 4358 arc_buf_hdr_t *hdr = buf->b_hdr; 4359 4360 /* 4361 * It would be nice to assert that if it's DMU metadata (level > 4362 * 0 || it's the dnode file), then it must be syncing context. 4363 * But we don't know that information at this level. 4364 */ 4365 4366 mutex_enter(&buf->b_evict_lock); 4367 /* 4368 * We don't grab the hash lock prior to this check, because if 4369 * the buffer's header is in the arc_anon state, it won't be 4370 * linked into the hash table. 4371 */ 4372 if (hdr->b_l1hdr.b_state == arc_anon) { 4373 mutex_exit(&buf->b_evict_lock); 4374 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 4375 ASSERT(!HDR_IN_HASH_TABLE(hdr)); 4376 ASSERT(!HDR_HAS_L2HDR(hdr)); 4377 ASSERT(BUF_EMPTY(hdr)); 4378 ASSERT3U(hdr->b_l1hdr.b_datacnt, ==, 1); 4379 ASSERT3S(refcount_count(&hdr->b_l1hdr.b_refcnt), ==, 1); 4380 ASSERT(!list_link_active(&hdr->b_l1hdr.b_arc_node)); 4381 4382 ASSERT3P(buf->b_efunc, ==, NULL); 4383 ASSERT3P(buf->b_private, ==, NULL); 4384 4385 hdr->b_l1hdr.b_arc_access = 0; 4386 arc_buf_thaw(buf); 4387 4388 return; 4389 } 4390 4391 kmutex_t *hash_lock = HDR_LOCK(hdr); 4392 mutex_enter(hash_lock); 4393 4394 /* 4395 * This assignment is only valid as long as the hash_lock is 4396 * held, we must be careful not to reference state or the 4397 * b_state field after dropping the lock. 4398 */ 4399 arc_state_t *state = hdr->b_l1hdr.b_state; 4400 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 4401 ASSERT3P(state, !=, arc_anon); 4402 4403 /* this buffer is not on any list */ 4404 ASSERT(refcount_count(&hdr->b_l1hdr.b_refcnt) > 0); 4405 4406 if (HDR_HAS_L2HDR(hdr)) { 4407 mutex_enter(&hdr->b_l2hdr.b_dev->l2ad_mtx); 4408 4409 /* 4410 * We have to recheck this conditional again now that 4411 * we're holding the l2ad_mtx to prevent a race with 4412 * another thread which might be concurrently calling 4413 * l2arc_evict(). In that case, l2arc_evict() might have 4414 * destroyed the header's L2 portion as we were waiting 4415 * to acquire the l2ad_mtx. 4416 */ 4417 if (HDR_HAS_L2HDR(hdr)) { 4418 if (hdr->b_l2hdr.b_daddr != L2ARC_ADDR_UNSET) 4419 trim_map_free(hdr->b_l2hdr.b_dev->l2ad_vdev, 4420 hdr->b_l2hdr.b_daddr, 4421 hdr->b_l2hdr.b_asize, 0); 4422 arc_hdr_l2hdr_destroy(hdr); 4423 } 4424 4425 mutex_exit(&hdr->b_l2hdr.b_dev->l2ad_mtx); 4426 } 4427 4428 /* 4429 * Do we have more than one buf? 4430 */ 4431 if (hdr->b_l1hdr.b_datacnt > 1) { 4432 arc_buf_hdr_t *nhdr; 4433 arc_buf_t **bufp; 4434 uint64_t blksz = hdr->b_size; 4435 uint64_t spa = hdr->b_spa; 4436 arc_buf_contents_t type = arc_buf_type(hdr); 4437 uint32_t flags = hdr->b_flags; 4438 4439 ASSERT(hdr->b_l1hdr.b_buf != buf || buf->b_next != NULL); 4440 /* 4441 * Pull the data off of this hdr and attach it to 4442 * a new anonymous hdr. 4443 */ 4444 (void) remove_reference(hdr, hash_lock, tag); 4445 bufp = &hdr->b_l1hdr.b_buf; 4446 while (*bufp != buf) 4447 bufp = &(*bufp)->b_next; 4448 *bufp = buf->b_next; 4449 buf->b_next = NULL; 4450 4451 ASSERT3P(state, !=, arc_l2c_only); 4452 ASSERT3U(state->arcs_size, >=, hdr->b_size); 4453 atomic_add_64(&state->arcs_size, -hdr->b_size); 4454 if (refcount_is_zero(&hdr->b_l1hdr.b_refcnt)) { 4455 ASSERT3P(state, !=, arc_l2c_only); 4456 uint64_t *size = &state->arcs_lsize[type]; 4457 ASSERT3U(*size, >=, hdr->b_size); 4458 atomic_add_64(size, -hdr->b_size); 4459 } 4460 4461 /* 4462 * We're releasing a duplicate user data buffer, update 4463 * our statistics accordingly. 4464 */ 4465 if (HDR_ISTYPE_DATA(hdr)) { 4466 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers); 4467 ARCSTAT_INCR(arcstat_duplicate_buffers_size, 4468 -hdr->b_size); 4469 } 4470 hdr->b_l1hdr.b_datacnt -= 1; 4471 arc_cksum_verify(buf); 4472#ifdef illumos 4473 arc_buf_unwatch(buf); 4474#endif 4475 4476 mutex_exit(hash_lock); 4477 4478 nhdr = kmem_cache_alloc(hdr_full_cache, KM_PUSHPAGE); 4479 nhdr->b_size = blksz; 4480 nhdr->b_spa = spa; 4481 4482 nhdr->b_flags = flags & ARC_FLAG_L2_WRITING; 4483 nhdr->b_flags |= arc_bufc_to_flags(type); 4484 nhdr->b_flags |= ARC_FLAG_HAS_L1HDR; 4485 4486 nhdr->b_l1hdr.b_buf = buf; 4487 nhdr->b_l1hdr.b_datacnt = 1; 4488 nhdr->b_l1hdr.b_state = arc_anon; 4489 nhdr->b_l1hdr.b_arc_access = 0; 4490 nhdr->b_freeze_cksum = NULL; 4491 4492 (void) refcount_add(&nhdr->b_l1hdr.b_refcnt, tag); 4493 buf->b_hdr = nhdr; 4494 mutex_exit(&buf->b_evict_lock); 4495 atomic_add_64(&arc_anon->arcs_size, blksz); 4496 } else { 4497 mutex_exit(&buf->b_evict_lock); 4498 ASSERT(refcount_count(&hdr->b_l1hdr.b_refcnt) == 1); 4499 /* protected by hash lock */ 4500 ASSERT(!list_link_active(&hdr->b_l1hdr.b_arc_node)); 4501 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 4502 arc_change_state(arc_anon, hdr, hash_lock); 4503 hdr->b_l1hdr.b_arc_access = 0; 4504 mutex_exit(hash_lock); 4505 4506 buf_discard_identity(hdr); 4507 arc_buf_thaw(buf); 4508 } 4509 buf->b_efunc = NULL; 4510 buf->b_private = NULL; 4511} 4512 4513int 4514arc_released(arc_buf_t *buf) 4515{ 4516 int released; 4517 4518 mutex_enter(&buf->b_evict_lock); 4519 released = (buf->b_data != NULL && 4520 buf->b_hdr->b_l1hdr.b_state == arc_anon); 4521 mutex_exit(&buf->b_evict_lock); 4522 return (released); 4523} 4524 4525#ifdef ZFS_DEBUG 4526int 4527arc_referenced(arc_buf_t *buf) 4528{ 4529 int referenced; 4530 4531 mutex_enter(&buf->b_evict_lock); 4532 referenced = (refcount_count(&buf->b_hdr->b_l1hdr.b_refcnt)); 4533 mutex_exit(&buf->b_evict_lock); 4534 return (referenced); 4535} 4536#endif 4537 4538static void 4539arc_write_ready(zio_t *zio) 4540{ 4541 arc_write_callback_t *callback = zio->io_private; 4542 arc_buf_t *buf = callback->awcb_buf; 4543 arc_buf_hdr_t *hdr = buf->b_hdr; 4544 4545 ASSERT(HDR_HAS_L1HDR(hdr)); 4546 ASSERT(!refcount_is_zero(&buf->b_hdr->b_l1hdr.b_refcnt)); 4547 ASSERT(hdr->b_l1hdr.b_datacnt > 0); 4548 callback->awcb_ready(zio, buf, callback->awcb_private); 4549 4550 /* 4551 * If the IO is already in progress, then this is a re-write 4552 * attempt, so we need to thaw and re-compute the cksum. 4553 * It is the responsibility of the callback to handle the 4554 * accounting for any re-write attempt. 4555 */ 4556 if (HDR_IO_IN_PROGRESS(hdr)) { 4557 mutex_enter(&hdr->b_l1hdr.b_freeze_lock); 4558 if (hdr->b_freeze_cksum != NULL) { 4559 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 4560 hdr->b_freeze_cksum = NULL; 4561 } 4562 mutex_exit(&hdr->b_l1hdr.b_freeze_lock); 4563 } 4564 arc_cksum_compute(buf, B_FALSE); 4565 hdr->b_flags |= ARC_FLAG_IO_IN_PROGRESS; 4566} 4567 4568/* 4569 * The SPA calls this callback for each physical write that happens on behalf 4570 * of a logical write. See the comment in dbuf_write_physdone() for details. 4571 */ 4572static void 4573arc_write_physdone(zio_t *zio) 4574{ 4575 arc_write_callback_t *cb = zio->io_private; 4576 if (cb->awcb_physdone != NULL) 4577 cb->awcb_physdone(zio, cb->awcb_buf, cb->awcb_private); 4578} 4579 4580static void 4581arc_write_done(zio_t *zio) 4582{ 4583 arc_write_callback_t *callback = zio->io_private; 4584 arc_buf_t *buf = callback->awcb_buf; 4585 arc_buf_hdr_t *hdr = buf->b_hdr; 4586 4587 ASSERT(hdr->b_l1hdr.b_acb == NULL); 4588 4589 if (zio->io_error == 0) { 4590 if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) { 4591 buf_discard_identity(hdr); 4592 } else { 4593 hdr->b_dva = *BP_IDENTITY(zio->io_bp); 4594 hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp); 4595 } 4596 } else { 4597 ASSERT(BUF_EMPTY(hdr)); 4598 } 4599 4600 /* 4601 * If the block to be written was all-zero or compressed enough to be 4602 * embedded in the BP, no write was performed so there will be no 4603 * dva/birth/checksum. The buffer must therefore remain anonymous 4604 * (and uncached). 4605 */ 4606 if (!BUF_EMPTY(hdr)) { 4607 arc_buf_hdr_t *exists; 4608 kmutex_t *hash_lock; 4609 4610 ASSERT(zio->io_error == 0); 4611 4612 arc_cksum_verify(buf); 4613 4614 exists = buf_hash_insert(hdr, &hash_lock); 4615 if (exists != NULL) { 4616 /* 4617 * This can only happen if we overwrite for 4618 * sync-to-convergence, because we remove 4619 * buffers from the hash table when we arc_free(). 4620 */ 4621 if (zio->io_flags & ZIO_FLAG_IO_REWRITE) { 4622 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) 4623 panic("bad overwrite, hdr=%p exists=%p", 4624 (void *)hdr, (void *)exists); 4625 ASSERT(refcount_is_zero( 4626 &exists->b_l1hdr.b_refcnt)); 4627 arc_change_state(arc_anon, exists, hash_lock); 4628 mutex_exit(hash_lock); 4629 arc_hdr_destroy(exists); 4630 exists = buf_hash_insert(hdr, &hash_lock); 4631 ASSERT3P(exists, ==, NULL); 4632 } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) { 4633 /* nopwrite */ 4634 ASSERT(zio->io_prop.zp_nopwrite); 4635 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) 4636 panic("bad nopwrite, hdr=%p exists=%p", 4637 (void *)hdr, (void *)exists); 4638 } else { 4639 /* Dedup */ 4640 ASSERT(hdr->b_l1hdr.b_datacnt == 1); 4641 ASSERT(hdr->b_l1hdr.b_state == arc_anon); 4642 ASSERT(BP_GET_DEDUP(zio->io_bp)); 4643 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0); 4644 } 4645 } 4646 hdr->b_flags &= ~ARC_FLAG_IO_IN_PROGRESS; 4647 /* if it's not anon, we are doing a scrub */ 4648 if (exists == NULL && hdr->b_l1hdr.b_state == arc_anon) 4649 arc_access(hdr, hash_lock); 4650 mutex_exit(hash_lock); 4651 } else { 4652 hdr->b_flags &= ~ARC_FLAG_IO_IN_PROGRESS; 4653 } 4654 4655 ASSERT(!refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 4656 callback->awcb_done(zio, buf, callback->awcb_private); 4657 4658 kmem_free(callback, sizeof (arc_write_callback_t)); 4659} 4660 4661zio_t * 4662arc_write(zio_t *pio, spa_t *spa, uint64_t txg, 4663 blkptr_t *bp, arc_buf_t *buf, boolean_t l2arc, boolean_t l2arc_compress, 4664 const zio_prop_t *zp, arc_done_func_t *ready, arc_done_func_t *physdone, 4665 arc_done_func_t *done, void *private, zio_priority_t priority, 4666 int zio_flags, const zbookmark_phys_t *zb) 4667{ 4668 arc_buf_hdr_t *hdr = buf->b_hdr; 4669 arc_write_callback_t *callback; 4670 zio_t *zio; 4671 4672 ASSERT(ready != NULL); 4673 ASSERT(done != NULL); 4674 ASSERT(!HDR_IO_ERROR(hdr)); 4675 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 4676 ASSERT(hdr->b_l1hdr.b_acb == NULL); 4677 ASSERT(hdr->b_l1hdr.b_datacnt > 0); 4678 if (l2arc) 4679 hdr->b_flags |= ARC_FLAG_L2CACHE; 4680 if (l2arc_compress) 4681 hdr->b_flags |= ARC_FLAG_L2COMPRESS; 4682 callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP); 4683 callback->awcb_ready = ready; 4684 callback->awcb_physdone = physdone; 4685 callback->awcb_done = done; 4686 callback->awcb_private = private; 4687 callback->awcb_buf = buf; 4688 4689 zio = zio_write(pio, spa, txg, bp, buf->b_data, hdr->b_size, zp, 4690 arc_write_ready, arc_write_physdone, arc_write_done, callback, 4691 priority, zio_flags, zb); 4692 4693 return (zio); 4694} 4695 4696static int 4697arc_memory_throttle(uint64_t reserve, uint64_t txg) 4698{ 4699#ifdef _KERNEL 4700 uint64_t available_memory = ptob(freemem); 4701 static uint64_t page_load = 0; 4702 static uint64_t last_txg = 0; 4703 4704#if defined(__i386) || !defined(UMA_MD_SMALL_ALLOC) 4705 available_memory = 4706 MIN(available_memory, ptob(vmem_size(heap_arena, VMEM_FREE))); 4707#endif 4708 4709 if (freemem > (uint64_t)physmem * arc_lotsfree_percent / 100) 4710 return (0); 4711 4712 if (txg > last_txg) { 4713 last_txg = txg; 4714 page_load = 0; 4715 } 4716 /* 4717 * If we are in pageout, we know that memory is already tight, 4718 * the arc is already going to be evicting, so we just want to 4719 * continue to let page writes occur as quickly as possible. 4720 */ 4721 if (curproc == pageproc) { 4722 if (page_load > MAX(ptob(minfree), available_memory) / 4) 4723 return (SET_ERROR(ERESTART)); 4724 /* Note: reserve is inflated, so we deflate */ 4725 page_load += reserve / 8; 4726 return (0); 4727 } else if (page_load > 0 && arc_reclaim_needed()) { 4728 /* memory is low, delay before restarting */ 4729 ARCSTAT_INCR(arcstat_memory_throttle_count, 1); 4730 return (SET_ERROR(EAGAIN)); 4731 } 4732 page_load = 0; 4733#endif 4734 return (0); 4735} 4736 4737void 4738arc_tempreserve_clear(uint64_t reserve) 4739{ 4740 atomic_add_64(&arc_tempreserve, -reserve); 4741 ASSERT((int64_t)arc_tempreserve >= 0); 4742} 4743 4744int 4745arc_tempreserve_space(uint64_t reserve, uint64_t txg) 4746{ 4747 int error; 4748 uint64_t anon_size; 4749 4750 if (reserve > arc_c/4 && !arc_no_grow) { 4751 arc_c = MIN(arc_c_max, reserve * 4); 4752 DTRACE_PROBE1(arc__set_reserve, uint64_t, arc_c); 4753 } 4754 if (reserve > arc_c) 4755 return (SET_ERROR(ENOMEM)); 4756 4757 /* 4758 * Don't count loaned bufs as in flight dirty data to prevent long 4759 * network delays from blocking transactions that are ready to be 4760 * assigned to a txg. 4761 */ 4762 anon_size = MAX((int64_t)(arc_anon->arcs_size - arc_loaned_bytes), 0); 4763 4764 /* 4765 * Writes will, almost always, require additional memory allocations 4766 * in order to compress/encrypt/etc the data. We therefore need to 4767 * make sure that there is sufficient available memory for this. 4768 */ 4769 error = arc_memory_throttle(reserve, txg); 4770 if (error != 0) 4771 return (error); 4772 4773 /* 4774 * Throttle writes when the amount of dirty data in the cache 4775 * gets too large. We try to keep the cache less than half full 4776 * of dirty blocks so that our sync times don't grow too large. 4777 * Note: if two requests come in concurrently, we might let them 4778 * both succeed, when one of them should fail. Not a huge deal. 4779 */ 4780 4781 if (reserve + arc_tempreserve + anon_size > arc_c / 2 && 4782 anon_size > arc_c / 4) { 4783 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK " 4784 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n", 4785 arc_tempreserve>>10, 4786 arc_anon->arcs_lsize[ARC_BUFC_METADATA]>>10, 4787 arc_anon->arcs_lsize[ARC_BUFC_DATA]>>10, 4788 reserve>>10, arc_c>>10); 4789 return (SET_ERROR(ERESTART)); 4790 } 4791 atomic_add_64(&arc_tempreserve, reserve); 4792 return (0); 4793} 4794 4795static void 4796arc_kstat_update_state(arc_state_t *state, kstat_named_t *size, 4797 kstat_named_t *evict_data, kstat_named_t *evict_metadata) 4798{ 4799 size->value.ui64 = state->arcs_size; 4800 evict_data->value.ui64 = state->arcs_lsize[ARC_BUFC_DATA]; 4801 evict_metadata->value.ui64 = state->arcs_lsize[ARC_BUFC_METADATA]; 4802} 4803 4804static int 4805arc_kstat_update(kstat_t *ksp, int rw) 4806{ 4807 arc_stats_t *as = ksp->ks_data; 4808 4809 if (rw == KSTAT_WRITE) { 4810 return (EACCES); 4811 } else { 4812 arc_kstat_update_state(arc_anon, 4813 &as->arcstat_anon_size, 4814 &as->arcstat_anon_evictable_data, 4815 &as->arcstat_anon_evictable_metadata); 4816 arc_kstat_update_state(arc_mru, 4817 &as->arcstat_mru_size, 4818 &as->arcstat_mru_evictable_data, 4819 &as->arcstat_mru_evictable_metadata); 4820 arc_kstat_update_state(arc_mru_ghost, 4821 &as->arcstat_mru_ghost_size, 4822 &as->arcstat_mru_ghost_evictable_data, 4823 &as->arcstat_mru_ghost_evictable_metadata); 4824 arc_kstat_update_state(arc_mfu, 4825 &as->arcstat_mfu_size, 4826 &as->arcstat_mfu_evictable_data, 4827 &as->arcstat_mfu_evictable_metadata); 4828 arc_kstat_update_state(arc_mfu_ghost, 4829 &as->arcstat_mfu_ghost_size, 4830 &as->arcstat_mfu_ghost_evictable_data, 4831 &as->arcstat_mfu_ghost_evictable_metadata); 4832 } 4833 4834 return (0); 4835} 4836 4837#ifdef _KERNEL 4838static eventhandler_tag arc_event_lowmem = NULL; 4839 4840static void 4841arc_lowmem(void *arg __unused, int howto __unused) 4842{ 4843 4844 mutex_enter(&arc_reclaim_thr_lock); 4845 /* XXX: Memory deficit should be passed as argument. */ 4846 needfree = btoc(arc_c >> arc_shrink_shift); 4847 DTRACE_PROBE(arc__needfree); 4848 cv_signal(&arc_reclaim_thr_cv); 4849 4850 /* 4851 * It is unsafe to block here in arbitrary threads, because we can come 4852 * here from ARC itself and may hold ARC locks and thus risk a deadlock 4853 * with ARC reclaim thread. 4854 */ 4855 if (curproc == pageproc) 4856 msleep(&needfree, &arc_reclaim_thr_lock, 0, "zfs:lowmem", 0); 4857 mutex_exit(&arc_reclaim_thr_lock); 4858} 4859#endif 4860 4861void 4862arc_init(void) 4863{ 4864 int i, prefetch_tunable_set = 0; 4865 4866 mutex_init(&arc_reclaim_thr_lock, NULL, MUTEX_DEFAULT, NULL); 4867 cv_init(&arc_reclaim_thr_cv, NULL, CV_DEFAULT, NULL); 4868 4869 /* Convert seconds to clock ticks */ 4870 arc_min_prefetch_lifespan = 1 * hz; 4871 4872 /* Start out with 1/8 of all memory */ 4873 arc_c = kmem_size() / 8; 4874 4875#ifdef illumos 4876#ifdef _KERNEL 4877 /* 4878 * On architectures where the physical memory can be larger 4879 * than the addressable space (intel in 32-bit mode), we may 4880 * need to limit the cache to 1/8 of VM size. 4881 */ 4882 arc_c = MIN(arc_c, vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 8); 4883#endif 4884#endif /* illumos */ 4885 /* set min cache to 1/32 of all memory, or 16MB, whichever is more */ 4886 arc_c_min = MAX(arc_c / 4, 16 << 20); 4887 /* set max to 1/2 of all memory, or all but 1GB, whichever is more */ 4888 if (arc_c * 8 >= 1 << 30) 4889 arc_c_max = (arc_c * 8) - (1 << 30); 4890 else 4891 arc_c_max = arc_c_min; 4892 arc_c_max = MAX(arc_c * 5, arc_c_max); 4893 4894#ifdef _KERNEL 4895 /* 4896 * Allow the tunables to override our calculations if they are 4897 * reasonable (ie. over 16MB) 4898 */ 4899 if (zfs_arc_max > 16 << 20 && zfs_arc_max < kmem_size()) 4900 arc_c_max = zfs_arc_max; 4901 if (zfs_arc_min > 16 << 20 && zfs_arc_min <= arc_c_max) 4902 arc_c_min = zfs_arc_min; 4903#endif 4904 4905 arc_c = arc_c_max; 4906 arc_p = (arc_c >> 1); 4907 4908 /* limit meta-data to 1/4 of the arc capacity */ 4909 arc_meta_limit = arc_c_max / 4; 4910 4911 /* Allow the tunable to override if it is reasonable */ 4912 if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max) 4913 arc_meta_limit = zfs_arc_meta_limit; 4914 4915 if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0) 4916 arc_c_min = arc_meta_limit / 2; 4917 4918 if (zfs_arc_meta_min > 0) { 4919 arc_meta_min = zfs_arc_meta_min; 4920 } else { 4921 arc_meta_min = arc_c_min / 2; 4922 } 4923 4924 if (zfs_arc_grow_retry > 0) 4925 arc_grow_retry = zfs_arc_grow_retry; 4926 4927 if (zfs_arc_shrink_shift > 0) 4928 arc_shrink_shift = zfs_arc_shrink_shift; 4929 4930 /* 4931 * Ensure that arc_no_grow_shift is less than arc_shrink_shift. 4932 */ 4933 if (arc_no_grow_shift >= arc_shrink_shift) 4934 arc_no_grow_shift = arc_shrink_shift - 1; 4935 4936 if (zfs_arc_p_min_shift > 0) 4937 arc_p_min_shift = zfs_arc_p_min_shift; 4938 4939 /* if kmem_flags are set, lets try to use less memory */ 4940 if (kmem_debugging()) 4941 arc_c = arc_c / 2; 4942 if (arc_c < arc_c_min) 4943 arc_c = arc_c_min; 4944 4945 zfs_arc_min = arc_c_min; 4946 zfs_arc_max = arc_c_max; 4947 4948 arc_anon = &ARC_anon; 4949 arc_mru = &ARC_mru; 4950 arc_mru_ghost = &ARC_mru_ghost; 4951 arc_mfu = &ARC_mfu; 4952 arc_mfu_ghost = &ARC_mfu_ghost; 4953 arc_l2c_only = &ARC_l2c_only; 4954 arc_size = 0; 4955 4956 for (i = 0; i < ARC_BUFC_NUMLISTS; i++) { 4957 mutex_init(&arc_anon->arcs_locks[i].arcs_lock, 4958 NULL, MUTEX_DEFAULT, NULL); 4959 mutex_init(&arc_mru->arcs_locks[i].arcs_lock, 4960 NULL, MUTEX_DEFAULT, NULL); 4961 mutex_init(&arc_mru_ghost->arcs_locks[i].arcs_lock, 4962 NULL, MUTEX_DEFAULT, NULL); 4963 mutex_init(&arc_mfu->arcs_locks[i].arcs_lock, 4964 NULL, MUTEX_DEFAULT, NULL); 4965 mutex_init(&arc_mfu_ghost->arcs_locks[i].arcs_lock, 4966 NULL, MUTEX_DEFAULT, NULL); 4967 mutex_init(&arc_l2c_only->arcs_locks[i].arcs_lock, 4968 NULL, MUTEX_DEFAULT, NULL); 4969 4970 list_create(&arc_mru->arcs_lists[i], 4971 sizeof (arc_buf_hdr_t), 4972 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node)); 4973 list_create(&arc_mru_ghost->arcs_lists[i], 4974 sizeof (arc_buf_hdr_t), 4975 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node)); 4976 list_create(&arc_mfu->arcs_lists[i], 4977 sizeof (arc_buf_hdr_t), 4978 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node)); 4979 list_create(&arc_mfu_ghost->arcs_lists[i], 4980 sizeof (arc_buf_hdr_t), 4981 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node)); 4982 list_create(&arc_mfu_ghost->arcs_lists[i], 4983 sizeof (arc_buf_hdr_t), 4984 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node)); 4985 list_create(&arc_l2c_only->arcs_lists[i], 4986 sizeof (arc_buf_hdr_t), 4987 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node)); 4988 } 4989 4990 buf_init(); 4991 4992 arc_thread_exit = 0; 4993 arc_eviction_list = NULL; 4994 mutex_init(&arc_eviction_mtx, NULL, MUTEX_DEFAULT, NULL); 4995 bzero(&arc_eviction_hdr, sizeof (arc_buf_hdr_t)); 4996 4997 arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED, 4998 sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL); 4999 5000 if (arc_ksp != NULL) { 5001 arc_ksp->ks_data = &arc_stats; 5002 arc_ksp->ks_update = arc_kstat_update; 5003 kstat_install(arc_ksp); 5004 } 5005 5006 (void) thread_create(NULL, 0, arc_reclaim_thread, NULL, 0, &p0, 5007 TS_RUN, minclsyspri); 5008 5009#ifdef _KERNEL 5010 arc_event_lowmem = EVENTHANDLER_REGISTER(vm_lowmem, arc_lowmem, NULL, 5011 EVENTHANDLER_PRI_FIRST); 5012#endif 5013 5014 arc_dead = FALSE; 5015 arc_warm = B_FALSE; 5016 5017 /* 5018 * Calculate maximum amount of dirty data per pool. 5019 * 5020 * If it has been set by /etc/system, take that. 5021 * Otherwise, use a percentage of physical memory defined by 5022 * zfs_dirty_data_max_percent (default 10%) with a cap at 5023 * zfs_dirty_data_max_max (default 4GB). 5024 */ 5025 if (zfs_dirty_data_max == 0) { 5026 zfs_dirty_data_max = ptob(physmem) * 5027 zfs_dirty_data_max_percent / 100; 5028 zfs_dirty_data_max = MIN(zfs_dirty_data_max, 5029 zfs_dirty_data_max_max); 5030 } 5031 5032#ifdef _KERNEL 5033 if (TUNABLE_INT_FETCH("vfs.zfs.prefetch_disable", &zfs_prefetch_disable)) 5034 prefetch_tunable_set = 1; 5035 5036#ifdef __i386__ 5037 if (prefetch_tunable_set == 0) { 5038 printf("ZFS NOTICE: Prefetch is disabled by default on i386 " 5039 "-- to enable,\n"); 5040 printf(" add \"vfs.zfs.prefetch_disable=0\" " 5041 "to /boot/loader.conf.\n"); 5042 zfs_prefetch_disable = 1; 5043 } 5044#else 5045 if ((((uint64_t)physmem * PAGESIZE) < (1ULL << 32)) && 5046 prefetch_tunable_set == 0) { 5047 printf("ZFS NOTICE: Prefetch is disabled by default if less " 5048 "than 4GB of RAM is present;\n" 5049 " to enable, add \"vfs.zfs.prefetch_disable=0\" " 5050 "to /boot/loader.conf.\n"); 5051 zfs_prefetch_disable = 1; 5052 } 5053#endif 5054 /* Warn about ZFS memory and address space requirements. */ 5055 if (((uint64_t)physmem * PAGESIZE) < (256 + 128 + 64) * (1 << 20)) { 5056 printf("ZFS WARNING: Recommended minimum RAM size is 512MB; " 5057 "expect unstable behavior.\n"); 5058 } 5059 if (kmem_size() < 512 * (1 << 20)) { 5060 printf("ZFS WARNING: Recommended minimum kmem_size is 512MB; " 5061 "expect unstable behavior.\n"); 5062 printf(" Consider tuning vm.kmem_size and " 5063 "vm.kmem_size_max\n"); 5064 printf(" in /boot/loader.conf.\n"); 5065 } 5066#endif 5067} 5068 5069void 5070arc_fini(void) 5071{ 5072 int i; 5073 5074 mutex_enter(&arc_reclaim_thr_lock); 5075 arc_thread_exit = 1; 5076 cv_signal(&arc_reclaim_thr_cv); 5077 while (arc_thread_exit != 0) 5078 cv_wait(&arc_reclaim_thr_cv, &arc_reclaim_thr_lock); 5079 mutex_exit(&arc_reclaim_thr_lock); 5080 5081 arc_flush(NULL); 5082 5083 arc_dead = TRUE; 5084 5085 if (arc_ksp != NULL) { 5086 kstat_delete(arc_ksp); 5087 arc_ksp = NULL; 5088 } 5089 5090 mutex_destroy(&arc_eviction_mtx); 5091 mutex_destroy(&arc_reclaim_thr_lock); 5092 cv_destroy(&arc_reclaim_thr_cv); 5093 5094 for (i = 0; i < ARC_BUFC_NUMLISTS; i++) { 5095 list_destroy(&arc_mru->arcs_lists[i]); 5096 list_destroy(&arc_mru_ghost->arcs_lists[i]); 5097 list_destroy(&arc_mfu->arcs_lists[i]); 5098 list_destroy(&arc_mfu_ghost->arcs_lists[i]); 5099 list_destroy(&arc_l2c_only->arcs_lists[i]); 5100 5101 mutex_destroy(&arc_anon->arcs_locks[i].arcs_lock); 5102 mutex_destroy(&arc_mru->arcs_locks[i].arcs_lock); 5103 mutex_destroy(&arc_mru_ghost->arcs_locks[i].arcs_lock); 5104 mutex_destroy(&arc_mfu->arcs_locks[i].arcs_lock); 5105 mutex_destroy(&arc_mfu_ghost->arcs_locks[i].arcs_lock); 5106 mutex_destroy(&arc_l2c_only->arcs_locks[i].arcs_lock); 5107 } 5108 5109 buf_fini(); 5110 5111 ASSERT0(arc_loaned_bytes); 5112 5113#ifdef _KERNEL 5114 if (arc_event_lowmem != NULL) 5115 EVENTHANDLER_DEREGISTER(vm_lowmem, arc_event_lowmem); 5116#endif 5117} 5118 5119/* 5120 * Level 2 ARC 5121 * 5122 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk. 5123 * It uses dedicated storage devices to hold cached data, which are populated 5124 * using large infrequent writes. The main role of this cache is to boost 5125 * the performance of random read workloads. The intended L2ARC devices 5126 * include short-stroked disks, solid state disks, and other media with 5127 * substantially faster read latency than disk. 5128 * 5129 * +-----------------------+ 5130 * | ARC | 5131 * +-----------------------+ 5132 * | ^ ^ 5133 * | | | 5134 * l2arc_feed_thread() arc_read() 5135 * | | | 5136 * | l2arc read | 5137 * V | | 5138 * +---------------+ | 5139 * | L2ARC | | 5140 * +---------------+ | 5141 * | ^ | 5142 * l2arc_write() | | 5143 * | | | 5144 * V | | 5145 * +-------+ +-------+ 5146 * | vdev | | vdev | 5147 * | cache | | cache | 5148 * +-------+ +-------+ 5149 * +=========+ .-----. 5150 * : L2ARC : |-_____-| 5151 * : devices : | Disks | 5152 * +=========+ `-_____-' 5153 * 5154 * Read requests are satisfied from the following sources, in order: 5155 * 5156 * 1) ARC 5157 * 2) vdev cache of L2ARC devices 5158 * 3) L2ARC devices 5159 * 4) vdev cache of disks 5160 * 5) disks 5161 * 5162 * Some L2ARC device types exhibit extremely slow write performance. 5163 * To accommodate for this there are some significant differences between 5164 * the L2ARC and traditional cache design: 5165 * 5166 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from 5167 * the ARC behave as usual, freeing buffers and placing headers on ghost 5168 * lists. The ARC does not send buffers to the L2ARC during eviction as 5169 * this would add inflated write latencies for all ARC memory pressure. 5170 * 5171 * 2. The L2ARC attempts to cache data from the ARC before it is evicted. 5172 * It does this by periodically scanning buffers from the eviction-end of 5173 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are 5174 * not already there. It scans until a headroom of buffers is satisfied, 5175 * which itself is a buffer for ARC eviction. If a compressible buffer is 5176 * found during scanning and selected for writing to an L2ARC device, we 5177 * temporarily boost scanning headroom during the next scan cycle to make 5178 * sure we adapt to compression effects (which might significantly reduce 5179 * the data volume we write to L2ARC). The thread that does this is 5180 * l2arc_feed_thread(), illustrated below; example sizes are included to 5181 * provide a better sense of ratio than this diagram: 5182 * 5183 * head --> tail 5184 * +---------------------+----------+ 5185 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC 5186 * +---------------------+----------+ | o L2ARC eligible 5187 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer 5188 * +---------------------+----------+ | 5189 * 15.9 Gbytes ^ 32 Mbytes | 5190 * headroom | 5191 * l2arc_feed_thread() 5192 * | 5193 * l2arc write hand <--[oooo]--' 5194 * | 8 Mbyte 5195 * | write max 5196 * V 5197 * +==============================+ 5198 * L2ARC dev |####|#|###|###| |####| ... | 5199 * +==============================+ 5200 * 32 Gbytes 5201 * 5202 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of 5203 * evicted, then the L2ARC has cached a buffer much sooner than it probably 5204 * needed to, potentially wasting L2ARC device bandwidth and storage. It is 5205 * safe to say that this is an uncommon case, since buffers at the end of 5206 * the ARC lists have moved there due to inactivity. 5207 * 5208 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom, 5209 * then the L2ARC simply misses copying some buffers. This serves as a 5210 * pressure valve to prevent heavy read workloads from both stalling the ARC 5211 * with waits and clogging the L2ARC with writes. This also helps prevent 5212 * the potential for the L2ARC to churn if it attempts to cache content too 5213 * quickly, such as during backups of the entire pool. 5214 * 5215 * 5. After system boot and before the ARC has filled main memory, there are 5216 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru 5217 * lists can remain mostly static. Instead of searching from tail of these 5218 * lists as pictured, the l2arc_feed_thread() will search from the list heads 5219 * for eligible buffers, greatly increasing its chance of finding them. 5220 * 5221 * The L2ARC device write speed is also boosted during this time so that 5222 * the L2ARC warms up faster. Since there have been no ARC evictions yet, 5223 * there are no L2ARC reads, and no fear of degrading read performance 5224 * through increased writes. 5225 * 5226 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that 5227 * the vdev queue can aggregate them into larger and fewer writes. Each 5228 * device is written to in a rotor fashion, sweeping writes through 5229 * available space then repeating. 5230 * 5231 * 7. The L2ARC does not store dirty content. It never needs to flush 5232 * write buffers back to disk based storage. 5233 * 5234 * 8. If an ARC buffer is written (and dirtied) which also exists in the 5235 * L2ARC, the now stale L2ARC buffer is immediately dropped. 5236 * 5237 * The performance of the L2ARC can be tweaked by a number of tunables, which 5238 * may be necessary for different workloads: 5239 * 5240 * l2arc_write_max max write bytes per interval 5241 * l2arc_write_boost extra write bytes during device warmup 5242 * l2arc_noprefetch skip caching prefetched buffers 5243 * l2arc_headroom number of max device writes to precache 5244 * l2arc_headroom_boost when we find compressed buffers during ARC 5245 * scanning, we multiply headroom by this 5246 * percentage factor for the next scan cycle, 5247 * since more compressed buffers are likely to 5248 * be present 5249 * l2arc_feed_secs seconds between L2ARC writing 5250 * 5251 * Tunables may be removed or added as future performance improvements are 5252 * integrated, and also may become zpool properties. 5253 * 5254 * There are three key functions that control how the L2ARC warms up: 5255 * 5256 * l2arc_write_eligible() check if a buffer is eligible to cache 5257 * l2arc_write_size() calculate how much to write 5258 * l2arc_write_interval() calculate sleep delay between writes 5259 * 5260 * These three functions determine what to write, how much, and how quickly 5261 * to send writes. 5262 */ 5263 5264static boolean_t 5265l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *hdr) 5266{ 5267 /* 5268 * A buffer is *not* eligible for the L2ARC if it: 5269 * 1. belongs to a different spa. 5270 * 2. is already cached on the L2ARC. 5271 * 3. has an I/O in progress (it may be an incomplete read). 5272 * 4. is flagged not eligible (zfs property). 5273 */ 5274 if (hdr->b_spa != spa_guid) { 5275 ARCSTAT_BUMP(arcstat_l2_write_spa_mismatch); 5276 return (B_FALSE); 5277 } 5278 if (HDR_HAS_L2HDR(hdr)) { 5279 ARCSTAT_BUMP(arcstat_l2_write_in_l2); 5280 return (B_FALSE); 5281 } 5282 if (HDR_IO_IN_PROGRESS(hdr)) { 5283 ARCSTAT_BUMP(arcstat_l2_write_hdr_io_in_progress); 5284 return (B_FALSE); 5285 } 5286 if (!HDR_L2CACHE(hdr)) { 5287 ARCSTAT_BUMP(arcstat_l2_write_not_cacheable); 5288 return (B_FALSE); 5289 } 5290 5291 return (B_TRUE); 5292} 5293 5294static uint64_t 5295l2arc_write_size(void) 5296{ 5297 uint64_t size; 5298 5299 /* 5300 * Make sure our globals have meaningful values in case the user 5301 * altered them. 5302 */ 5303 size = l2arc_write_max; 5304 if (size == 0) { 5305 cmn_err(CE_NOTE, "Bad value for l2arc_write_max, value must " 5306 "be greater than zero, resetting it to the default (%d)", 5307 L2ARC_WRITE_SIZE); 5308 size = l2arc_write_max = L2ARC_WRITE_SIZE; 5309 } 5310 5311 if (arc_warm == B_FALSE) 5312 size += l2arc_write_boost; 5313 5314 return (size); 5315 5316} 5317 5318static clock_t 5319l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote) 5320{ 5321 clock_t interval, next, now; 5322 5323 /* 5324 * If the ARC lists are busy, increase our write rate; if the 5325 * lists are stale, idle back. This is achieved by checking 5326 * how much we previously wrote - if it was more than half of 5327 * what we wanted, schedule the next write much sooner. 5328 */ 5329 if (l2arc_feed_again && wrote > (wanted / 2)) 5330 interval = (hz * l2arc_feed_min_ms) / 1000; 5331 else 5332 interval = hz * l2arc_feed_secs; 5333 5334 now = ddi_get_lbolt(); 5335 next = MAX(now, MIN(now + interval, began + interval)); 5336 5337 return (next); 5338} 5339 5340/* 5341 * Cycle through L2ARC devices. This is how L2ARC load balances. 5342 * If a device is returned, this also returns holding the spa config lock. 5343 */ 5344static l2arc_dev_t * 5345l2arc_dev_get_next(void) 5346{ 5347 l2arc_dev_t *first, *next = NULL; 5348 5349 /* 5350 * Lock out the removal of spas (spa_namespace_lock), then removal 5351 * of cache devices (l2arc_dev_mtx). Once a device has been selected, 5352 * both locks will be dropped and a spa config lock held instead. 5353 */ 5354 mutex_enter(&spa_namespace_lock); 5355 mutex_enter(&l2arc_dev_mtx); 5356 5357 /* if there are no vdevs, there is nothing to do */ 5358 if (l2arc_ndev == 0) 5359 goto out; 5360 5361 first = NULL; 5362 next = l2arc_dev_last; 5363 do { 5364 /* loop around the list looking for a non-faulted vdev */ 5365 if (next == NULL) { 5366 next = list_head(l2arc_dev_list); 5367 } else { 5368 next = list_next(l2arc_dev_list, next); 5369 if (next == NULL) 5370 next = list_head(l2arc_dev_list); 5371 } 5372 5373 /* if we have come back to the start, bail out */ 5374 if (first == NULL) 5375 first = next; 5376 else if (next == first) 5377 break; 5378 5379 } while (vdev_is_dead(next->l2ad_vdev)); 5380 5381 /* if we were unable to find any usable vdevs, return NULL */ 5382 if (vdev_is_dead(next->l2ad_vdev)) 5383 next = NULL; 5384 5385 l2arc_dev_last = next; 5386 5387out: 5388 mutex_exit(&l2arc_dev_mtx); 5389 5390 /* 5391 * Grab the config lock to prevent the 'next' device from being 5392 * removed while we are writing to it. 5393 */ 5394 if (next != NULL) 5395 spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER); 5396 mutex_exit(&spa_namespace_lock); 5397 5398 return (next); 5399} 5400 5401/* 5402 * Free buffers that were tagged for destruction. 5403 */ 5404static void 5405l2arc_do_free_on_write() 5406{ 5407 list_t *buflist; 5408 l2arc_data_free_t *df, *df_prev; 5409 5410 mutex_enter(&l2arc_free_on_write_mtx); 5411 buflist = l2arc_free_on_write; 5412 5413 for (df = list_tail(buflist); df; df = df_prev) { 5414 df_prev = list_prev(buflist, df); 5415 ASSERT(df->l2df_data != NULL); 5416 ASSERT(df->l2df_func != NULL); 5417 df->l2df_func(df->l2df_data, df->l2df_size); 5418 list_remove(buflist, df); 5419 kmem_free(df, sizeof (l2arc_data_free_t)); 5420 } 5421 5422 mutex_exit(&l2arc_free_on_write_mtx); 5423} 5424 5425/* 5426 * A write to a cache device has completed. Update all headers to allow 5427 * reads from these buffers to begin. 5428 */ 5429static void 5430l2arc_write_done(zio_t *zio) 5431{ 5432 l2arc_write_callback_t *cb; 5433 l2arc_dev_t *dev; 5434 list_t *buflist; 5435 arc_buf_hdr_t *head, *hdr, *hdr_prev; 5436 kmutex_t *hash_lock; 5437 int64_t bytes_dropped = 0; 5438 5439 cb = zio->io_private; 5440 ASSERT(cb != NULL); 5441 dev = cb->l2wcb_dev; 5442 ASSERT(dev != NULL); 5443 head = cb->l2wcb_head; 5444 ASSERT(head != NULL); 5445 buflist = &dev->l2ad_buflist; 5446 ASSERT(buflist != NULL); 5447 DTRACE_PROBE2(l2arc__iodone, zio_t *, zio, 5448 l2arc_write_callback_t *, cb); 5449 5450 if (zio->io_error != 0) 5451 ARCSTAT_BUMP(arcstat_l2_writes_error); 5452 5453 mutex_enter(&dev->l2ad_mtx); 5454 5455 /* 5456 * All writes completed, or an error was hit. 5457 */ 5458 for (hdr = list_prev(buflist, head); hdr; hdr = hdr_prev) { 5459 hdr_prev = list_prev(buflist, hdr); 5460 5461 hash_lock = HDR_LOCK(hdr); 5462 if (!mutex_tryenter(hash_lock)) { 5463 /* 5464 * This buffer misses out. It may be in a stage 5465 * of eviction. Its ARC_FLAG_L2_WRITING flag will be 5466 * left set, denying reads to this buffer. 5467 */ 5468 ARCSTAT_BUMP(arcstat_l2_writes_hdr_miss); 5469 continue; 5470 } 5471 5472 /* 5473 * It's possible that this buffer got evicted from the L1 cache 5474 * before we grabbed the vdev + hash locks, in which case 5475 * arc_hdr_realloc freed b_tmp_cdata for us if it was allocated. 5476 * Only free the buffer if we still have an L1 hdr. 5477 */ 5478 if (HDR_HAS_L1HDR(hdr) && hdr->b_l1hdr.b_tmp_cdata != NULL && 5479 HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF) 5480 l2arc_release_cdata_buf(hdr); 5481 5482 if (zio->io_error != 0) { 5483 /* 5484 * Error - drop L2ARC entry. 5485 */ 5486 trim_map_free(hdr->b_l2hdr.b_dev->l2ad_vdev, 5487 hdr->b_l2hdr.b_daddr, hdr->b_l2hdr.b_asize, 0); 5488 hdr->b_flags &= ~ARC_FLAG_HAS_L2HDR; 5489 5490 ARCSTAT_INCR(arcstat_l2_asize, -hdr->b_l2hdr.b_asize); 5491 ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size); 5492 5493 bytes_dropped += hdr->b_l2hdr.b_asize; 5494 (void) refcount_remove_many(&dev->l2ad_alloc, 5495 hdr->b_l2hdr.b_asize, hdr); 5496 } 5497 5498 /* 5499 * Allow ARC to begin reads to this L2ARC entry. 5500 */ 5501 hdr->b_flags &= ~ARC_FLAG_L2_WRITING; 5502 5503 mutex_exit(hash_lock); 5504 } 5505 5506 atomic_inc_64(&l2arc_writes_done); 5507 list_remove(buflist, head); 5508 ASSERT(!HDR_HAS_L1HDR(head)); 5509 kmem_cache_free(hdr_l2only_cache, head); 5510 mutex_exit(&dev->l2ad_mtx); 5511 5512 vdev_space_update(dev->l2ad_vdev, -bytes_dropped, 0, 0); 5513 5514 l2arc_do_free_on_write(); 5515 5516 kmem_free(cb, sizeof (l2arc_write_callback_t)); 5517} 5518 5519/* 5520 * A read to a cache device completed. Validate buffer contents before 5521 * handing over to the regular ARC routines. 5522 */ 5523static void 5524l2arc_read_done(zio_t *zio) 5525{ 5526 l2arc_read_callback_t *cb; 5527 arc_buf_hdr_t *hdr; 5528 arc_buf_t *buf; 5529 kmutex_t *hash_lock; 5530 int equal; 5531 5532 ASSERT(zio->io_vd != NULL); 5533 ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE); 5534 5535 spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd); 5536 5537 cb = zio->io_private; 5538 ASSERT(cb != NULL); 5539 buf = cb->l2rcb_buf; 5540 ASSERT(buf != NULL); 5541 5542 hash_lock = HDR_LOCK(buf->b_hdr); 5543 mutex_enter(hash_lock); 5544 hdr = buf->b_hdr; 5545 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 5546 5547 /* 5548 * If the buffer was compressed, decompress it first. 5549 */ 5550 if (cb->l2rcb_compress != ZIO_COMPRESS_OFF) 5551 l2arc_decompress_zio(zio, hdr, cb->l2rcb_compress); 5552 ASSERT(zio->io_data != NULL); 5553 5554 /* 5555 * Check this survived the L2ARC journey. 5556 */ 5557 equal = arc_cksum_equal(buf); 5558 if (equal && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) { 5559 mutex_exit(hash_lock); 5560 zio->io_private = buf; 5561 zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */ 5562 zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */ 5563 arc_read_done(zio); 5564 } else { 5565 mutex_exit(hash_lock); 5566 /* 5567 * Buffer didn't survive caching. Increment stats and 5568 * reissue to the original storage device. 5569 */ 5570 if (zio->io_error != 0) { 5571 ARCSTAT_BUMP(arcstat_l2_io_error); 5572 } else { 5573 zio->io_error = SET_ERROR(EIO); 5574 } 5575 if (!equal) 5576 ARCSTAT_BUMP(arcstat_l2_cksum_bad); 5577 5578 /* 5579 * If there's no waiter, issue an async i/o to the primary 5580 * storage now. If there *is* a waiter, the caller must 5581 * issue the i/o in a context where it's OK to block. 5582 */ 5583 if (zio->io_waiter == NULL) { 5584 zio_t *pio = zio_unique_parent(zio); 5585 5586 ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL); 5587 5588 zio_nowait(zio_read(pio, cb->l2rcb_spa, &cb->l2rcb_bp, 5589 buf->b_data, zio->io_size, arc_read_done, buf, 5590 zio->io_priority, cb->l2rcb_flags, &cb->l2rcb_zb)); 5591 } 5592 } 5593 5594 kmem_free(cb, sizeof (l2arc_read_callback_t)); 5595} 5596 5597/* 5598 * This is the list priority from which the L2ARC will search for pages to 5599 * cache. This is used within loops (0..3) to cycle through lists in the 5600 * desired order. This order can have a significant effect on cache 5601 * performance. 5602 * 5603 * Currently the metadata lists are hit first, MFU then MRU, followed by 5604 * the data lists. This function returns a locked list, and also returns 5605 * the lock pointer. 5606 */ 5607static list_t * 5608l2arc_list_locked(int list_num, kmutex_t **lock) 5609{ 5610 list_t *list = NULL; 5611 int idx; 5612 5613 ASSERT(list_num >= 0 && list_num < 2 * ARC_BUFC_NUMLISTS); 5614 5615 if (list_num < ARC_BUFC_NUMMETADATALISTS) { 5616 idx = list_num; 5617 list = &arc_mfu->arcs_lists[idx]; 5618 *lock = ARCS_LOCK(arc_mfu, idx); 5619 } else if (list_num < ARC_BUFC_NUMMETADATALISTS * 2) { 5620 idx = list_num - ARC_BUFC_NUMMETADATALISTS; 5621 list = &arc_mru->arcs_lists[idx]; 5622 *lock = ARCS_LOCK(arc_mru, idx); 5623 } else if (list_num < (ARC_BUFC_NUMMETADATALISTS * 2 + 5624 ARC_BUFC_NUMDATALISTS)) { 5625 idx = list_num - ARC_BUFC_NUMMETADATALISTS; 5626 list = &arc_mfu->arcs_lists[idx]; 5627 *lock = ARCS_LOCK(arc_mfu, idx); 5628 } else { 5629 idx = list_num - ARC_BUFC_NUMLISTS; 5630 list = &arc_mru->arcs_lists[idx]; 5631 *lock = ARCS_LOCK(arc_mru, idx); 5632 } 5633 5634 ASSERT(!(MUTEX_HELD(*lock))); 5635 mutex_enter(*lock); 5636 return (list); 5637} 5638 5639/* 5640 * Evict buffers from the device write hand to the distance specified in 5641 * bytes. This distance may span populated buffers, it may span nothing. 5642 * This is clearing a region on the L2ARC device ready for writing. 5643 * If the 'all' boolean is set, every buffer is evicted. 5644 */ 5645static void 5646l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all) 5647{ 5648 list_t *buflist; 5649 arc_buf_hdr_t *hdr, *hdr_prev; 5650 kmutex_t *hash_lock; 5651 uint64_t taddr; 5652 5653 buflist = &dev->l2ad_buflist; 5654 5655 if (!all && dev->l2ad_first) { 5656 /* 5657 * This is the first sweep through the device. There is 5658 * nothing to evict. 5659 */ 5660 return; 5661 } 5662 5663 if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) { 5664 /* 5665 * When nearing the end of the device, evict to the end 5666 * before the device write hand jumps to the start. 5667 */ 5668 taddr = dev->l2ad_end; 5669 } else { 5670 taddr = dev->l2ad_hand + distance; 5671 } 5672 DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist, 5673 uint64_t, taddr, boolean_t, all); 5674 5675top: 5676 mutex_enter(&dev->l2ad_mtx); 5677 for (hdr = list_tail(buflist); hdr; hdr = hdr_prev) { 5678 hdr_prev = list_prev(buflist, hdr); 5679 5680 hash_lock = HDR_LOCK(hdr); 5681 if (!mutex_tryenter(hash_lock)) { 5682 /* 5683 * Missed the hash lock. Retry. 5684 */ 5685 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry); 5686 mutex_exit(&dev->l2ad_mtx); 5687 mutex_enter(hash_lock); 5688 mutex_exit(hash_lock); 5689 goto top; 5690 } 5691 5692 if (HDR_L2_WRITE_HEAD(hdr)) { 5693 /* 5694 * We hit a write head node. Leave it for 5695 * l2arc_write_done(). 5696 */ 5697 list_remove(buflist, hdr); 5698 mutex_exit(hash_lock); 5699 continue; 5700 } 5701 5702 if (!all && HDR_HAS_L2HDR(hdr) && 5703 (hdr->b_l2hdr.b_daddr > taddr || 5704 hdr->b_l2hdr.b_daddr < dev->l2ad_hand)) { 5705 /* 5706 * We've evicted to the target address, 5707 * or the end of the device. 5708 */ 5709 mutex_exit(hash_lock); 5710 break; 5711 } 5712 5713 ASSERT(HDR_HAS_L2HDR(hdr)); 5714 if (!HDR_HAS_L1HDR(hdr)) { 5715 ASSERT(!HDR_L2_READING(hdr)); 5716 /* 5717 * This doesn't exist in the ARC. Destroy. 5718 * arc_hdr_destroy() will call list_remove() 5719 * and decrement arcstat_l2_size. 5720 */ 5721 arc_change_state(arc_anon, hdr, hash_lock); 5722 arc_hdr_destroy(hdr); 5723 } else { 5724 ASSERT(hdr->b_l1hdr.b_state != arc_l2c_only); 5725 ARCSTAT_BUMP(arcstat_l2_evict_l1cached); 5726 /* 5727 * Invalidate issued or about to be issued 5728 * reads, since we may be about to write 5729 * over this location. 5730 */ 5731 if (HDR_L2_READING(hdr)) { 5732 ARCSTAT_BUMP(arcstat_l2_evict_reading); 5733 hdr->b_flags |= ARC_FLAG_L2_EVICTED; 5734 } 5735 5736 arc_hdr_l2hdr_destroy(hdr); 5737 } 5738 mutex_exit(hash_lock); 5739 } 5740 mutex_exit(&dev->l2ad_mtx); 5741} 5742 5743/* 5744 * Find and write ARC buffers to the L2ARC device. 5745 * 5746 * An ARC_FLAG_L2_WRITING flag is set so that the L2ARC buffers are not valid 5747 * for reading until they have completed writing. 5748 * The headroom_boost is an in-out parameter used to maintain headroom boost 5749 * state between calls to this function. 5750 * 5751 * Returns the number of bytes actually written (which may be smaller than 5752 * the delta by which the device hand has changed due to alignment). 5753 */ 5754static uint64_t 5755l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz, 5756 boolean_t *headroom_boost) 5757{ 5758 arc_buf_hdr_t *hdr, *hdr_prev, *head; 5759 list_t *list; 5760 uint64_t write_asize, write_psize, write_sz, headroom, 5761 buf_compress_minsz; 5762 void *buf_data; 5763 kmutex_t *list_lock; 5764 boolean_t full; 5765 l2arc_write_callback_t *cb; 5766 zio_t *pio, *wzio; 5767 uint64_t guid = spa_load_guid(spa); 5768 const boolean_t do_headroom_boost = *headroom_boost; 5769 int try; 5770 5771 ASSERT(dev->l2ad_vdev != NULL); 5772 5773 /* Lower the flag now, we might want to raise it again later. */ 5774 *headroom_boost = B_FALSE; 5775 5776 pio = NULL; 5777 write_sz = write_asize = write_psize = 0; 5778 full = B_FALSE; 5779 head = kmem_cache_alloc(hdr_l2only_cache, KM_PUSHPAGE); 5780 head->b_flags |= ARC_FLAG_L2_WRITE_HEAD; 5781 head->b_flags |= ARC_FLAG_HAS_L2HDR; 5782 5783 ARCSTAT_BUMP(arcstat_l2_write_buffer_iter); 5784 /* 5785 * We will want to try to compress buffers that are at least 2x the 5786 * device sector size. 5787 */ 5788 buf_compress_minsz = 2 << dev->l2ad_vdev->vdev_ashift; 5789 5790 /* 5791 * Copy buffers for L2ARC writing. 5792 */ 5793 mutex_enter(&dev->l2ad_mtx); 5794 for (try = 0; try < 2 * ARC_BUFC_NUMLISTS; try++) { 5795 uint64_t passed_sz = 0; 5796 5797 list = l2arc_list_locked(try, &list_lock); 5798 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_iter); 5799 5800 /* 5801 * L2ARC fast warmup. 5802 * 5803 * Until the ARC is warm and starts to evict, read from the 5804 * head of the ARC lists rather than the tail. 5805 */ 5806 if (arc_warm == B_FALSE) 5807 hdr = list_head(list); 5808 else 5809 hdr = list_tail(list); 5810 if (hdr == NULL) 5811 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_null_iter); 5812 5813 headroom = target_sz * l2arc_headroom * 2 / ARC_BUFC_NUMLISTS; 5814 if (do_headroom_boost) 5815 headroom = (headroom * l2arc_headroom_boost) / 100; 5816 5817 for (; hdr; hdr = hdr_prev) { 5818 kmutex_t *hash_lock; 5819 uint64_t buf_sz; 5820 5821 if (arc_warm == B_FALSE) 5822 hdr_prev = list_next(list, hdr); 5823 else 5824 hdr_prev = list_prev(list, hdr); 5825 ARCSTAT_INCR(arcstat_l2_write_buffer_bytes_scanned, hdr->b_size); 5826 5827 hash_lock = HDR_LOCK(hdr); 5828 if (!mutex_tryenter(hash_lock)) { 5829 ARCSTAT_BUMP(arcstat_l2_write_trylock_fail); 5830 /* 5831 * Skip this buffer rather than waiting. 5832 */ 5833 continue; 5834 } 5835 5836 passed_sz += hdr->b_size; 5837 if (passed_sz > headroom) { 5838 /* 5839 * Searched too far. 5840 */ 5841 mutex_exit(hash_lock); 5842 ARCSTAT_BUMP(arcstat_l2_write_passed_headroom); 5843 break; 5844 } 5845 5846 if (!l2arc_write_eligible(guid, hdr)) { 5847 mutex_exit(hash_lock); 5848 continue; 5849 } 5850 5851 if ((write_sz + hdr->b_size) > target_sz) { 5852 full = B_TRUE; 5853 mutex_exit(hash_lock); 5854 ARCSTAT_BUMP(arcstat_l2_write_full); 5855 break; 5856 } 5857 5858 if (pio == NULL) { 5859 /* 5860 * Insert a dummy header on the buflist so 5861 * l2arc_write_done() can find where the 5862 * write buffers begin without searching. 5863 */ 5864 list_insert_head(&dev->l2ad_buflist, head); 5865 5866 cb = kmem_alloc( 5867 sizeof (l2arc_write_callback_t), KM_SLEEP); 5868 cb->l2wcb_dev = dev; 5869 cb->l2wcb_head = head; 5870 pio = zio_root(spa, l2arc_write_done, cb, 5871 ZIO_FLAG_CANFAIL); 5872 ARCSTAT_BUMP(arcstat_l2_write_pios); 5873 } 5874 5875 /* 5876 * Create and add a new L2ARC header. 5877 */ 5878 hdr->b_l2hdr.b_dev = dev; 5879 hdr->b_flags |= ARC_FLAG_L2_WRITING; 5880 /* 5881 * Temporarily stash the data buffer in b_tmp_cdata. 5882 * The subsequent write step will pick it up from 5883 * there. This is because can't access b_l1hdr.b_buf 5884 * without holding the hash_lock, which we in turn 5885 * can't access without holding the ARC list locks 5886 * (which we want to avoid during compression/writing). 5887 */ 5888 HDR_SET_COMPRESS(hdr, ZIO_COMPRESS_OFF); 5889 hdr->b_l2hdr.b_asize = hdr->b_size; 5890 hdr->b_l1hdr.b_tmp_cdata = hdr->b_l1hdr.b_buf->b_data; 5891 5892 /* 5893 * Explicitly set the b_daddr field to a known 5894 * value which means "invalid address". This 5895 * enables us to differentiate which stage of 5896 * l2arc_write_buffers() the particular header 5897 * is in (e.g. this loop, or the one below). 5898 * ARC_FLAG_L2_WRITING is not enough to make 5899 * this distinction, and we need to know in 5900 * order to do proper l2arc vdev accounting in 5901 * arc_release() and arc_hdr_destroy(). 5902 * 5903 * Note, we can't use a new flag to distinguish 5904 * the two stages because we don't hold the 5905 * header's hash_lock below, in the second stage 5906 * of this function. Thus, we can't simply 5907 * change the b_flags field to denote that the 5908 * IO has been sent. We can change the b_daddr 5909 * field of the L2 portion, though, since we'll 5910 * be holding the l2ad_mtx; which is why we're 5911 * using it to denote the header's state change. 5912 */ 5913 hdr->b_l2hdr.b_daddr = L2ARC_ADDR_UNSET; 5914 5915 buf_sz = hdr->b_size; 5916 hdr->b_flags |= ARC_FLAG_HAS_L2HDR; 5917 5918 list_insert_head(&dev->l2ad_buflist, hdr); 5919 5920 /* 5921 * Compute and store the buffer cksum before 5922 * writing. On debug the cksum is verified first. 5923 */ 5924 arc_cksum_verify(hdr->b_l1hdr.b_buf); 5925 arc_cksum_compute(hdr->b_l1hdr.b_buf, B_TRUE); 5926 5927 mutex_exit(hash_lock); 5928 5929 write_sz += buf_sz; 5930 } 5931 5932 mutex_exit(list_lock); 5933 5934 if (full == B_TRUE) 5935 break; 5936 } 5937 5938 /* No buffers selected for writing? */ 5939 if (pio == NULL) { 5940 ASSERT0(write_sz); 5941 mutex_exit(&dev->l2ad_mtx); 5942 ASSERT(!HDR_HAS_L1HDR(head)); 5943 kmem_cache_free(hdr_l2only_cache, head); 5944 return (0); 5945 } 5946 5947 /* 5948 * Now start writing the buffers. We're starting at the write head 5949 * and work backwards, retracing the course of the buffer selector 5950 * loop above. 5951 */ 5952 for (hdr = list_prev(&dev->l2ad_buflist, head); hdr; 5953 hdr = list_prev(&dev->l2ad_buflist, hdr)) { 5954 uint64_t buf_sz; 5955 5956 /* 5957 * We shouldn't need to lock the buffer here, since we flagged 5958 * it as ARC_FLAG_L2_WRITING in the previous step, but we must 5959 * take care to only access its L2 cache parameters. In 5960 * particular, hdr->l1hdr.b_buf may be invalid by now due to 5961 * ARC eviction. 5962 */ 5963 hdr->b_l2hdr.b_daddr = dev->l2ad_hand; 5964 5965 if ((HDR_L2COMPRESS(hdr)) && 5966 hdr->b_l2hdr.b_asize >= buf_compress_minsz) { 5967 if (l2arc_compress_buf(hdr)) { 5968 /* 5969 * If compression succeeded, enable headroom 5970 * boost on the next scan cycle. 5971 */ 5972 *headroom_boost = B_TRUE; 5973 } 5974 } 5975 5976 /* 5977 * Pick up the buffer data we had previously stashed away 5978 * (and now potentially also compressed). 5979 */ 5980 buf_data = hdr->b_l1hdr.b_tmp_cdata; 5981 buf_sz = hdr->b_l2hdr.b_asize; 5982 5983 /* 5984 * If the data has not been compressed, then clear b_tmp_cdata 5985 * to make sure that it points only to a temporary compression 5986 * buffer. 5987 */ 5988 if (!L2ARC_IS_VALID_COMPRESS(HDR_GET_COMPRESS(hdr))) 5989 hdr->b_l1hdr.b_tmp_cdata = NULL; 5990 5991 /* 5992 * We need to do this regardless if buf_sz is zero or 5993 * not, otherwise, when this l2hdr is evicted we'll 5994 * remove a reference that was never added. 5995 */ 5996 (void) refcount_add_many(&dev->l2ad_alloc, buf_sz, hdr); 5997 5998 /* Compression may have squashed the buffer to zero length. */ 5999 if (buf_sz != 0) { 6000 uint64_t buf_p_sz; 6001 6002 wzio = zio_write_phys(pio, dev->l2ad_vdev, 6003 dev->l2ad_hand, buf_sz, buf_data, ZIO_CHECKSUM_OFF, 6004 NULL, NULL, ZIO_PRIORITY_ASYNC_WRITE, 6005 ZIO_FLAG_CANFAIL, B_FALSE); 6006 6007 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev, 6008 zio_t *, wzio); 6009 (void) zio_nowait(wzio); 6010 6011 write_asize += buf_sz; 6012 6013 /* 6014 * Keep the clock hand suitably device-aligned. 6015 */ 6016 buf_p_sz = vdev_psize_to_asize(dev->l2ad_vdev, buf_sz); 6017 write_psize += buf_p_sz; 6018 dev->l2ad_hand += buf_p_sz; 6019 } 6020 } 6021 6022 mutex_exit(&dev->l2ad_mtx); 6023 6024 ASSERT3U(write_asize, <=, target_sz); 6025 ARCSTAT_BUMP(arcstat_l2_writes_sent); 6026 ARCSTAT_INCR(arcstat_l2_write_bytes, write_asize); 6027 ARCSTAT_INCR(arcstat_l2_size, write_sz); 6028 ARCSTAT_INCR(arcstat_l2_asize, write_asize); 6029 vdev_space_update(dev->l2ad_vdev, write_asize, 0, 0); 6030 6031 /* 6032 * Bump device hand to the device start if it is approaching the end. 6033 * l2arc_evict() will already have evicted ahead for this case. 6034 */ 6035 if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) { 6036 dev->l2ad_hand = dev->l2ad_start; 6037 dev->l2ad_first = B_FALSE; 6038 } 6039 6040 dev->l2ad_writing = B_TRUE; 6041 (void) zio_wait(pio); 6042 dev->l2ad_writing = B_FALSE; 6043 6044 return (write_asize); 6045} 6046 6047/* 6048 * Compresses an L2ARC buffer. 6049 * The data to be compressed must be prefilled in l1hdr.b_tmp_cdata and its 6050 * size in l2hdr->b_asize. This routine tries to compress the data and 6051 * depending on the compression result there are three possible outcomes: 6052 * *) The buffer was incompressible. The original l2hdr contents were left 6053 * untouched and are ready for writing to an L2 device. 6054 * *) The buffer was all-zeros, so there is no need to write it to an L2 6055 * device. To indicate this situation b_tmp_cdata is NULL'ed, b_asize is 6056 * set to zero and b_compress is set to ZIO_COMPRESS_EMPTY. 6057 * *) Compression succeeded and b_tmp_cdata was replaced with a temporary 6058 * data buffer which holds the compressed data to be written, and b_asize 6059 * tells us how much data there is. b_compress is set to the appropriate 6060 * compression algorithm. Once writing is done, invoke 6061 * l2arc_release_cdata_buf on this l2hdr to free this temporary buffer. 6062 * 6063 * Returns B_TRUE if compression succeeded, or B_FALSE if it didn't (the 6064 * buffer was incompressible). 6065 */ 6066static boolean_t 6067l2arc_compress_buf(arc_buf_hdr_t *hdr) 6068{ 6069 void *cdata; 6070 size_t csize, len, rounded; 6071 ASSERT(HDR_HAS_L2HDR(hdr)); 6072 l2arc_buf_hdr_t *l2hdr = &hdr->b_l2hdr; 6073 6074 ASSERT(HDR_HAS_L1HDR(hdr)); 6075 ASSERT(HDR_GET_COMPRESS(hdr) == ZIO_COMPRESS_OFF); 6076 ASSERT(hdr->b_l1hdr.b_tmp_cdata != NULL); 6077 6078 len = l2hdr->b_asize; 6079 cdata = zio_data_buf_alloc(len); 6080 ASSERT3P(cdata, !=, NULL); 6081 csize = zio_compress_data(ZIO_COMPRESS_LZ4, hdr->b_l1hdr.b_tmp_cdata, 6082 cdata, l2hdr->b_asize); 6083 6084 if (csize == 0) { 6085 /* zero block, indicate that there's nothing to write */ 6086 zio_data_buf_free(cdata, len); 6087 HDR_SET_COMPRESS(hdr, ZIO_COMPRESS_EMPTY); 6088 l2hdr->b_asize = 0; 6089 hdr->b_l1hdr.b_tmp_cdata = NULL; 6090 ARCSTAT_BUMP(arcstat_l2_compress_zeros); 6091 return (B_TRUE); 6092 } 6093 6094 rounded = P2ROUNDUP(csize, 6095 (size_t)1 << l2hdr->b_dev->l2ad_vdev->vdev_ashift); 6096 if (rounded < len) { 6097 /* 6098 * Compression succeeded, we'll keep the cdata around for 6099 * writing and release it afterwards. 6100 */ 6101 if (rounded > csize) { 6102 bzero((char *)cdata + csize, rounded - csize); 6103 csize = rounded; 6104 } 6105 HDR_SET_COMPRESS(hdr, ZIO_COMPRESS_LZ4); 6106 l2hdr->b_asize = csize; 6107 hdr->b_l1hdr.b_tmp_cdata = cdata; 6108 ARCSTAT_BUMP(arcstat_l2_compress_successes); 6109 return (B_TRUE); 6110 } else { 6111 /* 6112 * Compression failed, release the compressed buffer. 6113 * l2hdr will be left unmodified. 6114 */ 6115 zio_data_buf_free(cdata, len); 6116 ARCSTAT_BUMP(arcstat_l2_compress_failures); 6117 return (B_FALSE); 6118 } 6119} 6120 6121/* 6122 * Decompresses a zio read back from an l2arc device. On success, the 6123 * underlying zio's io_data buffer is overwritten by the uncompressed 6124 * version. On decompression error (corrupt compressed stream), the 6125 * zio->io_error value is set to signal an I/O error. 6126 * 6127 * Please note that the compressed data stream is not checksummed, so 6128 * if the underlying device is experiencing data corruption, we may feed 6129 * corrupt data to the decompressor, so the decompressor needs to be 6130 * able to handle this situation (LZ4 does). 6131 */ 6132static void 6133l2arc_decompress_zio(zio_t *zio, arc_buf_hdr_t *hdr, enum zio_compress c) 6134{ 6135 ASSERT(L2ARC_IS_VALID_COMPRESS(c)); 6136 6137 if (zio->io_error != 0) { 6138 /* 6139 * An io error has occured, just restore the original io 6140 * size in preparation for a main pool read. 6141 */ 6142 zio->io_orig_size = zio->io_size = hdr->b_size; 6143 return; 6144 } 6145 6146 if (c == ZIO_COMPRESS_EMPTY) { 6147 /* 6148 * An empty buffer results in a null zio, which means we 6149 * need to fill its io_data after we're done restoring the 6150 * buffer's contents. 6151 */ 6152 ASSERT(hdr->b_l1hdr.b_buf != NULL); 6153 bzero(hdr->b_l1hdr.b_buf->b_data, hdr->b_size); 6154 zio->io_data = zio->io_orig_data = hdr->b_l1hdr.b_buf->b_data; 6155 } else { 6156 ASSERT(zio->io_data != NULL); 6157 /* 6158 * We copy the compressed data from the start of the arc buffer 6159 * (the zio_read will have pulled in only what we need, the 6160 * rest is garbage which we will overwrite at decompression) 6161 * and then decompress back to the ARC data buffer. This way we 6162 * can minimize copying by simply decompressing back over the 6163 * original compressed data (rather than decompressing to an 6164 * aux buffer and then copying back the uncompressed buffer, 6165 * which is likely to be much larger). 6166 */ 6167 uint64_t csize; 6168 void *cdata; 6169 6170 csize = zio->io_size; 6171 cdata = zio_data_buf_alloc(csize); 6172 bcopy(zio->io_data, cdata, csize); 6173 if (zio_decompress_data(c, cdata, zio->io_data, csize, 6174 hdr->b_size) != 0) 6175 zio->io_error = EIO; 6176 zio_data_buf_free(cdata, csize); 6177 } 6178 6179 /* Restore the expected uncompressed IO size. */ 6180 zio->io_orig_size = zio->io_size = hdr->b_size; 6181} 6182 6183/* 6184 * Releases the temporary b_tmp_cdata buffer in an l2arc header structure. 6185 * This buffer serves as a temporary holder of compressed data while 6186 * the buffer entry is being written to an l2arc device. Once that is 6187 * done, we can dispose of it. 6188 */ 6189static void 6190l2arc_release_cdata_buf(arc_buf_hdr_t *hdr) 6191{ 6192 ASSERT(HDR_HAS_L1HDR(hdr)); 6193 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_EMPTY) { 6194 /* 6195 * If the data was compressed, then we've allocated a 6196 * temporary buffer for it, so now we need to release it. 6197 */ 6198 ASSERT(hdr->b_l1hdr.b_tmp_cdata != NULL); 6199 zio_data_buf_free(hdr->b_l1hdr.b_tmp_cdata, 6200 hdr->b_size); 6201 hdr->b_l1hdr.b_tmp_cdata = NULL; 6202 } else { 6203 ASSERT(hdr->b_l1hdr.b_tmp_cdata == NULL); 6204 } 6205} 6206 6207/* 6208 * This thread feeds the L2ARC at regular intervals. This is the beating 6209 * heart of the L2ARC. 6210 */ 6211static void 6212l2arc_feed_thread(void *dummy __unused) 6213{ 6214 callb_cpr_t cpr; 6215 l2arc_dev_t *dev; 6216 spa_t *spa; 6217 uint64_t size, wrote; 6218 clock_t begin, next = ddi_get_lbolt(); 6219 boolean_t headroom_boost = B_FALSE; 6220 6221 CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG); 6222 6223 mutex_enter(&l2arc_feed_thr_lock); 6224 6225 while (l2arc_thread_exit == 0) { 6226 CALLB_CPR_SAFE_BEGIN(&cpr); 6227 (void) cv_timedwait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock, 6228 next - ddi_get_lbolt()); 6229 CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock); 6230 next = ddi_get_lbolt() + hz; 6231 6232 /* 6233 * Quick check for L2ARC devices. 6234 */ 6235 mutex_enter(&l2arc_dev_mtx); 6236 if (l2arc_ndev == 0) { 6237 mutex_exit(&l2arc_dev_mtx); 6238 continue; 6239 } 6240 mutex_exit(&l2arc_dev_mtx); 6241 begin = ddi_get_lbolt(); 6242 6243 /* 6244 * This selects the next l2arc device to write to, and in 6245 * doing so the next spa to feed from: dev->l2ad_spa. This 6246 * will return NULL if there are now no l2arc devices or if 6247 * they are all faulted. 6248 * 6249 * If a device is returned, its spa's config lock is also 6250 * held to prevent device removal. l2arc_dev_get_next() 6251 * will grab and release l2arc_dev_mtx. 6252 */ 6253 if ((dev = l2arc_dev_get_next()) == NULL) 6254 continue; 6255 6256 spa = dev->l2ad_spa; 6257 ASSERT(spa != NULL); 6258 6259 /* 6260 * If the pool is read-only then force the feed thread to 6261 * sleep a little longer. 6262 */ 6263 if (!spa_writeable(spa)) { 6264 next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz; 6265 spa_config_exit(spa, SCL_L2ARC, dev); 6266 continue; 6267 } 6268 6269 /* 6270 * Avoid contributing to memory pressure. 6271 */ 6272 if (arc_reclaim_needed()) { 6273 ARCSTAT_BUMP(arcstat_l2_abort_lowmem); 6274 spa_config_exit(spa, SCL_L2ARC, dev); 6275 continue; 6276 } 6277 6278 ARCSTAT_BUMP(arcstat_l2_feeds); 6279 6280 size = l2arc_write_size(); 6281 6282 /* 6283 * Evict L2ARC buffers that will be overwritten. 6284 */ 6285 l2arc_evict(dev, size, B_FALSE); 6286 6287 /* 6288 * Write ARC buffers. 6289 */ 6290 wrote = l2arc_write_buffers(spa, dev, size, &headroom_boost); 6291 6292 /* 6293 * Calculate interval between writes. 6294 */ 6295 next = l2arc_write_interval(begin, size, wrote); 6296 spa_config_exit(spa, SCL_L2ARC, dev); 6297 } 6298 6299 l2arc_thread_exit = 0; 6300 cv_broadcast(&l2arc_feed_thr_cv); 6301 CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */ 6302 thread_exit(); 6303} 6304 6305boolean_t 6306l2arc_vdev_present(vdev_t *vd) 6307{ 6308 l2arc_dev_t *dev; 6309 6310 mutex_enter(&l2arc_dev_mtx); 6311 for (dev = list_head(l2arc_dev_list); dev != NULL; 6312 dev = list_next(l2arc_dev_list, dev)) { 6313 if (dev->l2ad_vdev == vd) 6314 break; 6315 } 6316 mutex_exit(&l2arc_dev_mtx); 6317 6318 return (dev != NULL); 6319} 6320 6321/* 6322 * Add a vdev for use by the L2ARC. By this point the spa has already 6323 * validated the vdev and opened it. 6324 */ 6325void 6326l2arc_add_vdev(spa_t *spa, vdev_t *vd) 6327{ 6328 l2arc_dev_t *adddev; 6329 6330 ASSERT(!l2arc_vdev_present(vd)); 6331 6332 vdev_ashift_optimize(vd); 6333 6334 /* 6335 * Create a new l2arc device entry. 6336 */ 6337 adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP); 6338 adddev->l2ad_spa = spa; 6339 adddev->l2ad_vdev = vd; 6340 adddev->l2ad_start = VDEV_LABEL_START_SIZE; 6341 adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd); 6342 adddev->l2ad_hand = adddev->l2ad_start; 6343 adddev->l2ad_first = B_TRUE; 6344 adddev->l2ad_writing = B_FALSE; 6345 6346 mutex_init(&adddev->l2ad_mtx, NULL, MUTEX_DEFAULT, NULL); 6347 /* 6348 * This is a list of all ARC buffers that are still valid on the 6349 * device. 6350 */ 6351 list_create(&adddev->l2ad_buflist, sizeof (arc_buf_hdr_t), 6352 offsetof(arc_buf_hdr_t, b_l2hdr.b_l2node)); 6353 6354 vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand); 6355 refcount_create(&adddev->l2ad_alloc); 6356 6357 /* 6358 * Add device to global list 6359 */ 6360 mutex_enter(&l2arc_dev_mtx); 6361 list_insert_head(l2arc_dev_list, adddev); 6362 atomic_inc_64(&l2arc_ndev); 6363 mutex_exit(&l2arc_dev_mtx); 6364} 6365 6366/* 6367 * Remove a vdev from the L2ARC. 6368 */ 6369void 6370l2arc_remove_vdev(vdev_t *vd) 6371{ 6372 l2arc_dev_t *dev, *nextdev, *remdev = NULL; 6373 6374 /* 6375 * Find the device by vdev 6376 */ 6377 mutex_enter(&l2arc_dev_mtx); 6378 for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) { 6379 nextdev = list_next(l2arc_dev_list, dev); 6380 if (vd == dev->l2ad_vdev) { 6381 remdev = dev; 6382 break; 6383 } 6384 } 6385 ASSERT(remdev != NULL); 6386 6387 /* 6388 * Remove device from global list 6389 */ 6390 list_remove(l2arc_dev_list, remdev); 6391 l2arc_dev_last = NULL; /* may have been invalidated */ 6392 atomic_dec_64(&l2arc_ndev); 6393 mutex_exit(&l2arc_dev_mtx); 6394 6395 /* 6396 * Clear all buflists and ARC references. L2ARC device flush. 6397 */ 6398 l2arc_evict(remdev, 0, B_TRUE); 6399 list_destroy(&remdev->l2ad_buflist); 6400 mutex_destroy(&remdev->l2ad_mtx); 6401 refcount_destroy(&remdev->l2ad_alloc); 6402 kmem_free(remdev, sizeof (l2arc_dev_t)); 6403} 6404 6405void 6406l2arc_init(void) 6407{ 6408 l2arc_thread_exit = 0; 6409 l2arc_ndev = 0; 6410 l2arc_writes_sent = 0; 6411 l2arc_writes_done = 0; 6412 6413 mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL); 6414 cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL); 6415 mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL); 6416 mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL); 6417 6418 l2arc_dev_list = &L2ARC_dev_list; 6419 l2arc_free_on_write = &L2ARC_free_on_write; 6420 list_create(l2arc_dev_list, sizeof (l2arc_dev_t), 6421 offsetof(l2arc_dev_t, l2ad_node)); 6422 list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t), 6423 offsetof(l2arc_data_free_t, l2df_list_node)); 6424} 6425 6426void 6427l2arc_fini(void) 6428{ 6429 /* 6430 * This is called from dmu_fini(), which is called from spa_fini(); 6431 * Because of this, we can assume that all l2arc devices have 6432 * already been removed when the pools themselves were removed. 6433 */ 6434 6435 l2arc_do_free_on_write(); 6436 6437 mutex_destroy(&l2arc_feed_thr_lock); 6438 cv_destroy(&l2arc_feed_thr_cv); 6439 mutex_destroy(&l2arc_dev_mtx); 6440 mutex_destroy(&l2arc_free_on_write_mtx); 6441 6442 list_destroy(l2arc_dev_list); 6443 list_destroy(l2arc_free_on_write); 6444} 6445 6446void 6447l2arc_start(void) 6448{ 6449 if (!(spa_mode_global & FWRITE)) 6450 return; 6451 6452 (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0, 6453 TS_RUN, minclsyspri); 6454} 6455 6456void 6457l2arc_stop(void) 6458{ 6459 if (!(spa_mode_global & FWRITE)) 6460 return; 6461 6462 mutex_enter(&l2arc_feed_thr_lock); 6463 cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */ 6464 l2arc_thread_exit = 1; 6465 while (l2arc_thread_exit != 0) 6466 cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock); 6467 mutex_exit(&l2arc_feed_thr_lock); 6468} 6469