arc.c revision 288565
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_max", &zfs_arc_max); 229TUNABLE_QUAD("vfs.zfs.arc_min", &zfs_arc_min); 230TUNABLE_QUAD("vfs.zfs.arc_meta_limit", &zfs_arc_meta_limit); 231TUNABLE_QUAD("vfs.zfs.arc_meta_min", &zfs_arc_meta_min); 232TUNABLE_QUAD("vfs.zfs.arc_average_blocksize", &zfs_arc_average_blocksize); 233TUNABLE_INT("vfs.zfs.arc_shrink_shift", &zfs_arc_shrink_shift); 234SYSCTL_DECL(_vfs_zfs); 235SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_max, CTLFLAG_RDTUN, &zfs_arc_max, 0, 236 "Maximum ARC size"); 237SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_min, CTLFLAG_RDTUN, &zfs_arc_min, 0, 238 "Minimum ARC size"); 239SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_average_blocksize, CTLFLAG_RDTUN, 240 &zfs_arc_average_blocksize, 0, 241 "ARC average blocksize"); 242SYSCTL_INT(_vfs_zfs, OID_AUTO, arc_shrink_shift, CTLFLAG_RW, 243 &arc_shrink_shift, 0, 244 "log2(fraction of arc to reclaim)"); 245 246/* 247 * We don't have a tunable for arc_free_target due to the dependency on 248 * pagedaemon initialisation. 249 */ 250SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_free_target, 251 CTLTYPE_UINT | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, sizeof(u_int), 252 sysctl_vfs_zfs_arc_free_target, "IU", 253 "Desired number of free pages below which ARC triggers reclaim"); 254 255static int 256sysctl_vfs_zfs_arc_free_target(SYSCTL_HANDLER_ARGS) 257{ 258 u_int val; 259 int err; 260 261 val = zfs_arc_free_target; 262 err = sysctl_handle_int(oidp, &val, 0, req); 263 if (err != 0 || req->newptr == NULL) 264 return (err); 265 266 if (val < minfree) 267 return (EINVAL); 268 if (val > cnt.v_page_count) 269 return (EINVAL); 270 271 zfs_arc_free_target = val; 272 273 return (0); 274} 275 276/* 277 * Must be declared here, before the definition of corresponding kstat 278 * macro which uses the same names will confuse the compiler. 279 */ 280SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_meta_limit, 281 CTLTYPE_U64 | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, sizeof(uint64_t), 282 sysctl_vfs_zfs_arc_meta_limit, "QU", 283 "ARC metadata limit"); 284#endif 285 286/* 287 * Note that buffers can be in one of 6 states: 288 * ARC_anon - anonymous (discussed below) 289 * ARC_mru - recently used, currently cached 290 * ARC_mru_ghost - recentely used, no longer in cache 291 * ARC_mfu - frequently used, currently cached 292 * ARC_mfu_ghost - frequently used, no longer in cache 293 * ARC_l2c_only - exists in L2ARC but not other states 294 * When there are no active references to the buffer, they are 295 * are linked onto a list in one of these arc states. These are 296 * the only buffers that can be evicted or deleted. Within each 297 * state there are multiple lists, one for meta-data and one for 298 * non-meta-data. Meta-data (indirect blocks, blocks of dnodes, 299 * etc.) is tracked separately so that it can be managed more 300 * explicitly: favored over data, limited explicitly. 301 * 302 * Anonymous buffers are buffers that are not associated with 303 * a DVA. These are buffers that hold dirty block copies 304 * before they are written to stable storage. By definition, 305 * they are "ref'd" and are considered part of arc_mru 306 * that cannot be freed. Generally, they will aquire a DVA 307 * as they are written and migrate onto the arc_mru list. 308 * 309 * The ARC_l2c_only state is for buffers that are in the second 310 * level ARC but no longer in any of the ARC_m* lists. The second 311 * level ARC itself may also contain buffers that are in any of 312 * the ARC_m* states - meaning that a buffer can exist in two 313 * places. The reason for the ARC_l2c_only state is to keep the 314 * buffer header in the hash table, so that reads that hit the 315 * second level ARC benefit from these fast lookups. 316 */ 317 318#define ARCS_LOCK_PAD CACHE_LINE_SIZE 319struct arcs_lock { 320 kmutex_t arcs_lock; 321#ifdef _KERNEL 322 unsigned char pad[(ARCS_LOCK_PAD - sizeof (kmutex_t))]; 323#endif 324}; 325 326/* 327 * must be power of two for mask use to work 328 * 329 */ 330#define ARC_BUFC_NUMDATALISTS 16 331#define ARC_BUFC_NUMMETADATALISTS 16 332#define ARC_BUFC_NUMLISTS (ARC_BUFC_NUMMETADATALISTS + ARC_BUFC_NUMDATALISTS) 333 334typedef struct arc_state { 335 uint64_t arcs_lsize[ARC_BUFC_NUMTYPES]; /* amount of evictable data */ 336 uint64_t arcs_size; /* total amount of data in this state */ 337 list_t arcs_lists[ARC_BUFC_NUMLISTS]; /* list of evictable buffers */ 338 struct arcs_lock arcs_locks[ARC_BUFC_NUMLISTS] __aligned(CACHE_LINE_SIZE); 339} arc_state_t; 340 341#define ARCS_LOCK(s, i) (&((s)->arcs_locks[(i)].arcs_lock)) 342 343/* The 6 states: */ 344static arc_state_t ARC_anon; 345static arc_state_t ARC_mru; 346static arc_state_t ARC_mru_ghost; 347static arc_state_t ARC_mfu; 348static arc_state_t ARC_mfu_ghost; 349static arc_state_t ARC_l2c_only; 350 351typedef struct arc_stats { 352 kstat_named_t arcstat_hits; 353 kstat_named_t arcstat_misses; 354 kstat_named_t arcstat_demand_data_hits; 355 kstat_named_t arcstat_demand_data_misses; 356 kstat_named_t arcstat_demand_metadata_hits; 357 kstat_named_t arcstat_demand_metadata_misses; 358 kstat_named_t arcstat_prefetch_data_hits; 359 kstat_named_t arcstat_prefetch_data_misses; 360 kstat_named_t arcstat_prefetch_metadata_hits; 361 kstat_named_t arcstat_prefetch_metadata_misses; 362 kstat_named_t arcstat_mru_hits; 363 kstat_named_t arcstat_mru_ghost_hits; 364 kstat_named_t arcstat_mfu_hits; 365 kstat_named_t arcstat_mfu_ghost_hits; 366 kstat_named_t arcstat_allocated; 367 kstat_named_t arcstat_deleted; 368 kstat_named_t arcstat_stolen; 369 kstat_named_t arcstat_recycle_miss; 370 /* 371 * Number of buffers that could not be evicted because the hash lock 372 * was held by another thread. The lock may not necessarily be held 373 * by something using the same buffer, since hash locks are shared 374 * by multiple buffers. 375 */ 376 kstat_named_t arcstat_mutex_miss; 377 /* 378 * Number of buffers skipped because they have I/O in progress, are 379 * indrect prefetch buffers that have not lived long enough, or are 380 * not from the spa we're trying to evict from. 381 */ 382 kstat_named_t arcstat_evict_skip; 383 kstat_named_t arcstat_evict_l2_cached; 384 kstat_named_t arcstat_evict_l2_eligible; 385 kstat_named_t arcstat_evict_l2_ineligible; 386 kstat_named_t arcstat_hash_elements; 387 kstat_named_t arcstat_hash_elements_max; 388 kstat_named_t arcstat_hash_collisions; 389 kstat_named_t arcstat_hash_chains; 390 kstat_named_t arcstat_hash_chain_max; 391 kstat_named_t arcstat_p; 392 kstat_named_t arcstat_c; 393 kstat_named_t arcstat_c_min; 394 kstat_named_t arcstat_c_max; 395 kstat_named_t arcstat_size; 396 /* 397 * Number of bytes consumed by internal ARC structures necessary 398 * for tracking purposes; these structures are not actually 399 * backed by ARC buffers. This includes arc_buf_hdr_t structures 400 * (allocated via arc_buf_hdr_t_full and arc_buf_hdr_t_l2only 401 * caches), and arc_buf_t structures (allocated via arc_buf_t 402 * cache). 403 */ 404 kstat_named_t arcstat_hdr_size; 405 /* 406 * Number of bytes consumed by ARC buffers of type equal to 407 * ARC_BUFC_DATA. This is generally consumed by buffers backing 408 * on disk user data (e.g. plain file contents). 409 */ 410 kstat_named_t arcstat_data_size; 411 /* 412 * Number of bytes consumed by ARC buffers of type equal to 413 * ARC_BUFC_METADATA. This is generally consumed by buffers 414 * backing on disk data that is used for internal ZFS 415 * structures (e.g. ZAP, dnode, indirect blocks, etc). 416 */ 417 kstat_named_t arcstat_metadata_size; 418 /* 419 * Number of bytes consumed by various buffers and structures 420 * not actually backed with ARC buffers. This includes bonus 421 * buffers (allocated directly via zio_buf_* functions), 422 * dmu_buf_impl_t structures (allocated via dmu_buf_impl_t 423 * cache), and dnode_t structures (allocated via dnode_t cache). 424 */ 425 kstat_named_t arcstat_other_size; 426 /* 427 * Total number of bytes consumed by ARC buffers residing in the 428 * arc_anon state. This includes *all* buffers in the arc_anon 429 * state; e.g. data, metadata, evictable, and unevictable buffers 430 * are all included in this value. 431 */ 432 kstat_named_t arcstat_anon_size; 433 /* 434 * Number of bytes consumed by ARC buffers that meet the 435 * following criteria: backing buffers of type ARC_BUFC_DATA, 436 * residing in the arc_anon state, and are eligible for eviction 437 * (e.g. have no outstanding holds on the buffer). 438 */ 439 kstat_named_t arcstat_anon_evictable_data; 440 /* 441 * Number of bytes consumed by ARC buffers that meet the 442 * following criteria: backing buffers of type ARC_BUFC_METADATA, 443 * residing in the arc_anon state, and are eligible for eviction 444 * (e.g. have no outstanding holds on the buffer). 445 */ 446 kstat_named_t arcstat_anon_evictable_metadata; 447 /* 448 * Total number of bytes consumed by ARC buffers residing in the 449 * arc_mru state. This includes *all* buffers in the arc_mru 450 * state; e.g. data, metadata, evictable, and unevictable buffers 451 * are all included in this value. 452 */ 453 kstat_named_t arcstat_mru_size; 454 /* 455 * Number of bytes consumed by ARC buffers that meet the 456 * following criteria: backing buffers of type ARC_BUFC_DATA, 457 * residing in the arc_mru state, and are eligible for eviction 458 * (e.g. have no outstanding holds on the buffer). 459 */ 460 kstat_named_t arcstat_mru_evictable_data; 461 /* 462 * Number of bytes consumed by ARC buffers that meet the 463 * following criteria: backing buffers of type ARC_BUFC_METADATA, 464 * residing in the arc_mru state, and are eligible for eviction 465 * (e.g. have no outstanding holds on the buffer). 466 */ 467 kstat_named_t arcstat_mru_evictable_metadata; 468 /* 469 * Total number of bytes that *would have been* consumed by ARC 470 * buffers in the arc_mru_ghost state. The key thing to note 471 * here, is the fact that this size doesn't actually indicate 472 * RAM consumption. The ghost lists only consist of headers and 473 * don't actually have ARC buffers linked off of these headers. 474 * Thus, *if* the headers had associated ARC buffers, these 475 * buffers *would have* consumed this number of bytes. 476 */ 477 kstat_named_t arcstat_mru_ghost_size; 478 /* 479 * Number of bytes that *would have been* consumed by ARC 480 * buffers that are eligible for eviction, of type 481 * ARC_BUFC_DATA, and linked off the arc_mru_ghost state. 482 */ 483 kstat_named_t arcstat_mru_ghost_evictable_data; 484 /* 485 * Number of bytes that *would have been* consumed by ARC 486 * buffers that are eligible for eviction, of type 487 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state. 488 */ 489 kstat_named_t arcstat_mru_ghost_evictable_metadata; 490 /* 491 * Total number of bytes consumed by ARC buffers residing in the 492 * arc_mfu state. This includes *all* buffers in the arc_mfu 493 * state; e.g. data, metadata, evictable, and unevictable buffers 494 * are all included in this value. 495 */ 496 kstat_named_t arcstat_mfu_size; 497 /* 498 * Number of bytes consumed by ARC buffers that are eligible for 499 * eviction, of type ARC_BUFC_DATA, and reside in the arc_mfu 500 * state. 501 */ 502 kstat_named_t arcstat_mfu_evictable_data; 503 /* 504 * Number of bytes consumed by ARC buffers that are eligible for 505 * eviction, of type ARC_BUFC_METADATA, and reside in the 506 * arc_mfu state. 507 */ 508 kstat_named_t arcstat_mfu_evictable_metadata; 509 /* 510 * Total number of bytes that *would have been* consumed by ARC 511 * buffers in the arc_mfu_ghost state. See the comment above 512 * arcstat_mru_ghost_size for more details. 513 */ 514 kstat_named_t arcstat_mfu_ghost_size; 515 /* 516 * Number of bytes that *would have been* consumed by ARC 517 * buffers that are eligible for eviction, of type 518 * ARC_BUFC_DATA, and linked off the arc_mfu_ghost state. 519 */ 520 kstat_named_t arcstat_mfu_ghost_evictable_data; 521 /* 522 * Number of bytes that *would have been* consumed by ARC 523 * buffers that are eligible for eviction, of type 524 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state. 525 */ 526 kstat_named_t arcstat_mfu_ghost_evictable_metadata; 527 kstat_named_t arcstat_l2_hits; 528 kstat_named_t arcstat_l2_misses; 529 kstat_named_t arcstat_l2_feeds; 530 kstat_named_t arcstat_l2_rw_clash; 531 kstat_named_t arcstat_l2_read_bytes; 532 kstat_named_t arcstat_l2_write_bytes; 533 kstat_named_t arcstat_l2_writes_sent; 534 kstat_named_t arcstat_l2_writes_done; 535 kstat_named_t arcstat_l2_writes_error; 536 kstat_named_t arcstat_l2_writes_hdr_miss; 537 kstat_named_t arcstat_l2_evict_lock_retry; 538 kstat_named_t arcstat_l2_evict_reading; 539 kstat_named_t arcstat_l2_evict_l1cached; 540 kstat_named_t arcstat_l2_free_on_write; 541 kstat_named_t arcstat_l2_cdata_free_on_write; 542 kstat_named_t arcstat_l2_abort_lowmem; 543 kstat_named_t arcstat_l2_cksum_bad; 544 kstat_named_t arcstat_l2_io_error; 545 kstat_named_t arcstat_l2_size; 546 kstat_named_t arcstat_l2_asize; 547 kstat_named_t arcstat_l2_hdr_size; 548 kstat_named_t arcstat_l2_compress_successes; 549 kstat_named_t arcstat_l2_compress_zeros; 550 kstat_named_t arcstat_l2_compress_failures; 551 kstat_named_t arcstat_l2_write_trylock_fail; 552 kstat_named_t arcstat_l2_write_passed_headroom; 553 kstat_named_t arcstat_l2_write_spa_mismatch; 554 kstat_named_t arcstat_l2_write_in_l2; 555 kstat_named_t arcstat_l2_write_hdr_io_in_progress; 556 kstat_named_t arcstat_l2_write_not_cacheable; 557 kstat_named_t arcstat_l2_write_full; 558 kstat_named_t arcstat_l2_write_buffer_iter; 559 kstat_named_t arcstat_l2_write_pios; 560 kstat_named_t arcstat_l2_write_buffer_bytes_scanned; 561 kstat_named_t arcstat_l2_write_buffer_list_iter; 562 kstat_named_t arcstat_l2_write_buffer_list_null_iter; 563 kstat_named_t arcstat_memory_throttle_count; 564 kstat_named_t arcstat_duplicate_buffers; 565 kstat_named_t arcstat_duplicate_buffers_size; 566 kstat_named_t arcstat_duplicate_reads; 567 kstat_named_t arcstat_meta_used; 568 kstat_named_t arcstat_meta_limit; 569 kstat_named_t arcstat_meta_max; 570 kstat_named_t arcstat_meta_min; 571} arc_stats_t; 572 573static arc_stats_t arc_stats = { 574 { "hits", KSTAT_DATA_UINT64 }, 575 { "misses", KSTAT_DATA_UINT64 }, 576 { "demand_data_hits", KSTAT_DATA_UINT64 }, 577 { "demand_data_misses", KSTAT_DATA_UINT64 }, 578 { "demand_metadata_hits", KSTAT_DATA_UINT64 }, 579 { "demand_metadata_misses", KSTAT_DATA_UINT64 }, 580 { "prefetch_data_hits", KSTAT_DATA_UINT64 }, 581 { "prefetch_data_misses", KSTAT_DATA_UINT64 }, 582 { "prefetch_metadata_hits", KSTAT_DATA_UINT64 }, 583 { "prefetch_metadata_misses", KSTAT_DATA_UINT64 }, 584 { "mru_hits", KSTAT_DATA_UINT64 }, 585 { "mru_ghost_hits", KSTAT_DATA_UINT64 }, 586 { "mfu_hits", KSTAT_DATA_UINT64 }, 587 { "mfu_ghost_hits", KSTAT_DATA_UINT64 }, 588 { "allocated", KSTAT_DATA_UINT64 }, 589 { "deleted", KSTAT_DATA_UINT64 }, 590 { "stolen", KSTAT_DATA_UINT64 }, 591 { "recycle_miss", KSTAT_DATA_UINT64 }, 592 { "mutex_miss", KSTAT_DATA_UINT64 }, 593 { "evict_skip", KSTAT_DATA_UINT64 }, 594 { "evict_l2_cached", KSTAT_DATA_UINT64 }, 595 { "evict_l2_eligible", KSTAT_DATA_UINT64 }, 596 { "evict_l2_ineligible", KSTAT_DATA_UINT64 }, 597 { "hash_elements", KSTAT_DATA_UINT64 }, 598 { "hash_elements_max", KSTAT_DATA_UINT64 }, 599 { "hash_collisions", KSTAT_DATA_UINT64 }, 600 { "hash_chains", KSTAT_DATA_UINT64 }, 601 { "hash_chain_max", KSTAT_DATA_UINT64 }, 602 { "p", KSTAT_DATA_UINT64 }, 603 { "c", KSTAT_DATA_UINT64 }, 604 { "c_min", KSTAT_DATA_UINT64 }, 605 { "c_max", KSTAT_DATA_UINT64 }, 606 { "size", KSTAT_DATA_UINT64 }, 607 { "hdr_size", KSTAT_DATA_UINT64 }, 608 { "data_size", KSTAT_DATA_UINT64 }, 609 { "metadata_size", KSTAT_DATA_UINT64 }, 610 { "other_size", KSTAT_DATA_UINT64 }, 611 { "anon_size", KSTAT_DATA_UINT64 }, 612 { "anon_evictable_data", KSTAT_DATA_UINT64 }, 613 { "anon_evictable_metadata", KSTAT_DATA_UINT64 }, 614 { "mru_size", KSTAT_DATA_UINT64 }, 615 { "mru_evictable_data", KSTAT_DATA_UINT64 }, 616 { "mru_evictable_metadata", KSTAT_DATA_UINT64 }, 617 { "mru_ghost_size", KSTAT_DATA_UINT64 }, 618 { "mru_ghost_evictable_data", KSTAT_DATA_UINT64 }, 619 { "mru_ghost_evictable_metadata", KSTAT_DATA_UINT64 }, 620 { "mfu_size", KSTAT_DATA_UINT64 }, 621 { "mfu_evictable_data", KSTAT_DATA_UINT64 }, 622 { "mfu_evictable_metadata", KSTAT_DATA_UINT64 }, 623 { "mfu_ghost_size", KSTAT_DATA_UINT64 }, 624 { "mfu_ghost_evictable_data", KSTAT_DATA_UINT64 }, 625 { "mfu_ghost_evictable_metadata", KSTAT_DATA_UINT64 }, 626 { "l2_hits", KSTAT_DATA_UINT64 }, 627 { "l2_misses", KSTAT_DATA_UINT64 }, 628 { "l2_feeds", KSTAT_DATA_UINT64 }, 629 { "l2_rw_clash", KSTAT_DATA_UINT64 }, 630 { "l2_read_bytes", KSTAT_DATA_UINT64 }, 631 { "l2_write_bytes", KSTAT_DATA_UINT64 }, 632 { "l2_writes_sent", KSTAT_DATA_UINT64 }, 633 { "l2_writes_done", KSTAT_DATA_UINT64 }, 634 { "l2_writes_error", KSTAT_DATA_UINT64 }, 635 { "l2_writes_hdr_miss", KSTAT_DATA_UINT64 }, 636 { "l2_evict_lock_retry", KSTAT_DATA_UINT64 }, 637 { "l2_evict_reading", KSTAT_DATA_UINT64 }, 638 { "l2_evict_l1cached", KSTAT_DATA_UINT64 }, 639 { "l2_free_on_write", KSTAT_DATA_UINT64 }, 640 { "l2_cdata_free_on_write", KSTAT_DATA_UINT64 }, 641 { "l2_abort_lowmem", KSTAT_DATA_UINT64 }, 642 { "l2_cksum_bad", KSTAT_DATA_UINT64 }, 643 { "l2_io_error", KSTAT_DATA_UINT64 }, 644 { "l2_size", KSTAT_DATA_UINT64 }, 645 { "l2_asize", KSTAT_DATA_UINT64 }, 646 { "l2_hdr_size", KSTAT_DATA_UINT64 }, 647 { "l2_compress_successes", KSTAT_DATA_UINT64 }, 648 { "l2_compress_zeros", KSTAT_DATA_UINT64 }, 649 { "l2_compress_failures", KSTAT_DATA_UINT64 }, 650 { "l2_write_trylock_fail", KSTAT_DATA_UINT64 }, 651 { "l2_write_passed_headroom", KSTAT_DATA_UINT64 }, 652 { "l2_write_spa_mismatch", KSTAT_DATA_UINT64 }, 653 { "l2_write_in_l2", KSTAT_DATA_UINT64 }, 654 { "l2_write_io_in_progress", KSTAT_DATA_UINT64 }, 655 { "l2_write_not_cacheable", KSTAT_DATA_UINT64 }, 656 { "l2_write_full", KSTAT_DATA_UINT64 }, 657 { "l2_write_buffer_iter", KSTAT_DATA_UINT64 }, 658 { "l2_write_pios", KSTAT_DATA_UINT64 }, 659 { "l2_write_buffer_bytes_scanned", KSTAT_DATA_UINT64 }, 660 { "l2_write_buffer_list_iter", KSTAT_DATA_UINT64 }, 661 { "l2_write_buffer_list_null_iter", KSTAT_DATA_UINT64 }, 662 { "memory_throttle_count", KSTAT_DATA_UINT64 }, 663 { "duplicate_buffers", KSTAT_DATA_UINT64 }, 664 { "duplicate_buffers_size", KSTAT_DATA_UINT64 }, 665 { "duplicate_reads", KSTAT_DATA_UINT64 }, 666 { "arc_meta_used", KSTAT_DATA_UINT64 }, 667 { "arc_meta_limit", KSTAT_DATA_UINT64 }, 668 { "arc_meta_max", KSTAT_DATA_UINT64 }, 669 { "arc_meta_min", KSTAT_DATA_UINT64 } 670}; 671 672#define ARCSTAT(stat) (arc_stats.stat.value.ui64) 673 674#define ARCSTAT_INCR(stat, val) \ 675 atomic_add_64(&arc_stats.stat.value.ui64, (val)) 676 677#define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1) 678#define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1) 679 680#define ARCSTAT_MAX(stat, val) { \ 681 uint64_t m; \ 682 while ((val) > (m = arc_stats.stat.value.ui64) && \ 683 (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \ 684 continue; \ 685} 686 687#define ARCSTAT_MAXSTAT(stat) \ 688 ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64) 689 690/* 691 * We define a macro to allow ARC hits/misses to be easily broken down by 692 * two separate conditions, giving a total of four different subtypes for 693 * each of hits and misses (so eight statistics total). 694 */ 695#define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \ 696 if (cond1) { \ 697 if (cond2) { \ 698 ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \ 699 } else { \ 700 ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \ 701 } \ 702 } else { \ 703 if (cond2) { \ 704 ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \ 705 } else { \ 706 ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\ 707 } \ 708 } 709 710kstat_t *arc_ksp; 711static arc_state_t *arc_anon; 712static arc_state_t *arc_mru; 713static arc_state_t *arc_mru_ghost; 714static arc_state_t *arc_mfu; 715static arc_state_t *arc_mfu_ghost; 716static arc_state_t *arc_l2c_only; 717 718/* 719 * There are several ARC variables that are critical to export as kstats -- 720 * but we don't want to have to grovel around in the kstat whenever we wish to 721 * manipulate them. For these variables, we therefore define them to be in 722 * terms of the statistic variable. This assures that we are not introducing 723 * the possibility of inconsistency by having shadow copies of the variables, 724 * while still allowing the code to be readable. 725 */ 726#define arc_size ARCSTAT(arcstat_size) /* actual total arc size */ 727#define arc_p ARCSTAT(arcstat_p) /* target size of MRU */ 728#define arc_c ARCSTAT(arcstat_c) /* target size of cache */ 729#define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */ 730#define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */ 731#define arc_meta_limit ARCSTAT(arcstat_meta_limit) /* max size for metadata */ 732#define arc_meta_min ARCSTAT(arcstat_meta_min) /* min size for metadata */ 733#define arc_meta_used ARCSTAT(arcstat_meta_used) /* size of metadata */ 734#define arc_meta_max ARCSTAT(arcstat_meta_max) /* max size of metadata */ 735 736#define L2ARC_IS_VALID_COMPRESS(_c_) \ 737 ((_c_) == ZIO_COMPRESS_LZ4 || (_c_) == ZIO_COMPRESS_EMPTY) 738 739static int arc_no_grow; /* Don't try to grow cache size */ 740static uint64_t arc_tempreserve; 741static uint64_t arc_loaned_bytes; 742 743typedef struct arc_callback arc_callback_t; 744 745struct arc_callback { 746 void *acb_private; 747 arc_done_func_t *acb_done; 748 arc_buf_t *acb_buf; 749 zio_t *acb_zio_dummy; 750 arc_callback_t *acb_next; 751}; 752 753typedef struct arc_write_callback arc_write_callback_t; 754 755struct arc_write_callback { 756 void *awcb_private; 757 arc_done_func_t *awcb_ready; 758 arc_done_func_t *awcb_physdone; 759 arc_done_func_t *awcb_done; 760 arc_buf_t *awcb_buf; 761}; 762 763/* 764 * ARC buffers are separated into multiple structs as a memory saving measure: 765 * - Common fields struct, always defined, and embedded within it: 766 * - L2-only fields, always allocated but undefined when not in L2ARC 767 * - L1-only fields, only allocated when in L1ARC 768 * 769 * Buffer in L1 Buffer only in L2 770 * +------------------------+ +------------------------+ 771 * | arc_buf_hdr_t | | arc_buf_hdr_t | 772 * | | | | 773 * | | | | 774 * | | | | 775 * +------------------------+ +------------------------+ 776 * | l2arc_buf_hdr_t | | l2arc_buf_hdr_t | 777 * | (undefined if L1-only) | | | 778 * +------------------------+ +------------------------+ 779 * | l1arc_buf_hdr_t | 780 * | | 781 * | | 782 * | | 783 * | | 784 * +------------------------+ 785 * 786 * Because it's possible for the L2ARC to become extremely large, we can wind 787 * up eating a lot of memory in L2ARC buffer headers, so the size of a header 788 * is minimized by only allocating the fields necessary for an L1-cached buffer 789 * when a header is actually in the L1 cache. The sub-headers (l1arc_buf_hdr and 790 * l2arc_buf_hdr) are embedded rather than allocated separately to save a couple 791 * words in pointers. arc_hdr_realloc() is used to switch a header between 792 * these two allocation states. 793 */ 794typedef struct l1arc_buf_hdr { 795 kmutex_t b_freeze_lock; 796#ifdef ZFS_DEBUG 797 /* 798 * used for debugging wtih kmem_flags - by allocating and freeing 799 * b_thawed when the buffer is thawed, we get a record of the stack 800 * trace that thawed it. 801 */ 802 void *b_thawed; 803#endif 804 805 arc_buf_t *b_buf; 806 uint32_t b_datacnt; 807 /* for waiting on writes to complete */ 808 kcondvar_t b_cv; 809 810 /* protected by arc state mutex */ 811 arc_state_t *b_state; 812 list_node_t b_arc_node; 813 814 /* updated atomically */ 815 clock_t b_arc_access; 816 817 /* self protecting */ 818 refcount_t b_refcnt; 819 820 arc_callback_t *b_acb; 821 /* temporary buffer holder for in-flight compressed data */ 822 void *b_tmp_cdata; 823} l1arc_buf_hdr_t; 824 825typedef struct l2arc_dev l2arc_dev_t; 826 827typedef struct l2arc_buf_hdr { 828 /* protected by arc_buf_hdr mutex */ 829 l2arc_dev_t *b_dev; /* L2ARC device */ 830 uint64_t b_daddr; /* disk address, offset byte */ 831 /* real alloc'd buffer size depending on b_compress applied */ 832 int32_t b_asize; 833 834 list_node_t b_l2node; 835} l2arc_buf_hdr_t; 836 837struct arc_buf_hdr { 838 /* protected by hash lock */ 839 dva_t b_dva; 840 uint64_t b_birth; 841 /* 842 * Even though this checksum is only set/verified when a buffer is in 843 * the L1 cache, it needs to be in the set of common fields because it 844 * must be preserved from the time before a buffer is written out to 845 * L2ARC until after it is read back in. 846 */ 847 zio_cksum_t *b_freeze_cksum; 848 849 arc_buf_hdr_t *b_hash_next; 850 arc_flags_t b_flags; 851 852 /* immutable */ 853 int32_t b_size; 854 uint64_t b_spa; 855 856 /* L2ARC fields. Undefined when not in L2ARC. */ 857 l2arc_buf_hdr_t b_l2hdr; 858 /* L1ARC fields. Undefined when in l2arc_only state */ 859 l1arc_buf_hdr_t b_l1hdr; 860}; 861 862#ifdef _KERNEL 863static int 864sysctl_vfs_zfs_arc_meta_limit(SYSCTL_HANDLER_ARGS) 865{ 866 uint64_t val; 867 int err; 868 869 val = arc_meta_limit; 870 err = sysctl_handle_64(oidp, &val, 0, req); 871 if (err != 0 || req->newptr == NULL) 872 return (err); 873 874 if (val <= 0 || val > arc_c_max) 875 return (EINVAL); 876 877 arc_meta_limit = val; 878 return (0); 879} 880#endif 881 882static arc_buf_t *arc_eviction_list; 883static kmutex_t arc_eviction_mtx; 884static arc_buf_hdr_t arc_eviction_hdr; 885 886#define GHOST_STATE(state) \ 887 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \ 888 (state) == arc_l2c_only) 889 890#define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_FLAG_IN_HASH_TABLE) 891#define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) 892#define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_FLAG_IO_ERROR) 893#define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_FLAG_PREFETCH) 894#define HDR_FREED_IN_READ(hdr) ((hdr)->b_flags & ARC_FLAG_FREED_IN_READ) 895#define HDR_BUF_AVAILABLE(hdr) ((hdr)->b_flags & ARC_FLAG_BUF_AVAILABLE) 896 897#define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_FLAG_L2CACHE) 898#define HDR_L2COMPRESS(hdr) ((hdr)->b_flags & ARC_FLAG_L2COMPRESS) 899#define HDR_L2_READING(hdr) \ 900 (((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) && \ 901 ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR)) 902#define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITING) 903#define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_FLAG_L2_EVICTED) 904#define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITE_HEAD) 905 906#define HDR_ISTYPE_METADATA(hdr) \ 907 ((hdr)->b_flags & ARC_FLAG_BUFC_METADATA) 908#define HDR_ISTYPE_DATA(hdr) (!HDR_ISTYPE_METADATA(hdr)) 909 910#define HDR_HAS_L1HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L1HDR) 911#define HDR_HAS_L2HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR) 912 913/* For storing compression mode in b_flags */ 914#define HDR_COMPRESS_OFFSET 24 915#define HDR_COMPRESS_NBITS 7 916 917#define HDR_GET_COMPRESS(hdr) ((enum zio_compress)BF32_GET(hdr->b_flags, \ 918 HDR_COMPRESS_OFFSET, HDR_COMPRESS_NBITS)) 919#define HDR_SET_COMPRESS(hdr, cmp) BF32_SET(hdr->b_flags, \ 920 HDR_COMPRESS_OFFSET, HDR_COMPRESS_NBITS, (cmp)) 921 922/* 923 * Other sizes 924 */ 925 926#define HDR_FULL_SIZE ((int64_t)sizeof (arc_buf_hdr_t)) 927#define HDR_L2ONLY_SIZE ((int64_t)offsetof(arc_buf_hdr_t, b_l1hdr)) 928 929/* 930 * Hash table routines 931 */ 932 933#define HT_LOCK_PAD CACHE_LINE_SIZE 934 935struct ht_lock { 936 kmutex_t ht_lock; 937#ifdef _KERNEL 938 unsigned char pad[(HT_LOCK_PAD - sizeof (kmutex_t))]; 939#endif 940}; 941 942#define BUF_LOCKS 256 943typedef struct buf_hash_table { 944 uint64_t ht_mask; 945 arc_buf_hdr_t **ht_table; 946 struct ht_lock ht_locks[BUF_LOCKS] __aligned(CACHE_LINE_SIZE); 947} buf_hash_table_t; 948 949static buf_hash_table_t buf_hash_table; 950 951#define BUF_HASH_INDEX(spa, dva, birth) \ 952 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask) 953#define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)]) 954#define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock)) 955#define HDR_LOCK(hdr) \ 956 (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth))) 957 958uint64_t zfs_crc64_table[256]; 959 960/* 961 * Level 2 ARC 962 */ 963 964#define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */ 965#define L2ARC_HEADROOM 2 /* num of writes */ 966/* 967 * If we discover during ARC scan any buffers to be compressed, we boost 968 * our headroom for the next scanning cycle by this percentage multiple. 969 */ 970#define L2ARC_HEADROOM_BOOST 200 971#define L2ARC_FEED_SECS 1 /* caching interval secs */ 972#define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */ 973 974/* 975 * Used to distinguish headers that are being process by 976 * l2arc_write_buffers(), but have yet to be assigned to a l2arc disk 977 * address. This can happen when the header is added to the l2arc's list 978 * of buffers to write in the first stage of l2arc_write_buffers(), but 979 * has not yet been written out which happens in the second stage of 980 * l2arc_write_buffers(). 981 */ 982#define L2ARC_ADDR_UNSET ((uint64_t)(-1)) 983 984#define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent) 985#define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done) 986 987/* L2ARC Performance Tunables */ 988uint64_t l2arc_write_max = L2ARC_WRITE_SIZE; /* default max write size */ 989uint64_t l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra write during warmup */ 990uint64_t l2arc_headroom = L2ARC_HEADROOM; /* number of dev writes */ 991uint64_t l2arc_headroom_boost = L2ARC_HEADROOM_BOOST; 992uint64_t l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */ 993uint64_t l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval milliseconds */ 994boolean_t l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */ 995boolean_t l2arc_feed_again = B_TRUE; /* turbo warmup */ 996boolean_t l2arc_norw = B_TRUE; /* no reads during writes */ 997 998SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_max, CTLFLAG_RW, 999 &l2arc_write_max, 0, "max write size"); 1000SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_boost, CTLFLAG_RW, 1001 &l2arc_write_boost, 0, "extra write during warmup"); 1002SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_headroom, CTLFLAG_RW, 1003 &l2arc_headroom, 0, "number of dev writes"); 1004SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_secs, CTLFLAG_RW, 1005 &l2arc_feed_secs, 0, "interval seconds"); 1006SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_min_ms, CTLFLAG_RW, 1007 &l2arc_feed_min_ms, 0, "min interval milliseconds"); 1008 1009SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_noprefetch, CTLFLAG_RW, 1010 &l2arc_noprefetch, 0, "don't cache prefetch bufs"); 1011SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_feed_again, CTLFLAG_RW, 1012 &l2arc_feed_again, 0, "turbo warmup"); 1013SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_norw, CTLFLAG_RW, 1014 &l2arc_norw, 0, "no reads during writes"); 1015 1016SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_size, CTLFLAG_RD, 1017 &ARC_anon.arcs_size, 0, "size of anonymous state"); 1018SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_metadata_lsize, CTLFLAG_RD, 1019 &ARC_anon.arcs_lsize[ARC_BUFC_METADATA], 0, "size of anonymous state"); 1020SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_data_lsize, CTLFLAG_RD, 1021 &ARC_anon.arcs_lsize[ARC_BUFC_DATA], 0, "size of anonymous state"); 1022 1023SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_size, CTLFLAG_RD, 1024 &ARC_mru.arcs_size, 0, "size of mru state"); 1025SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_metadata_lsize, CTLFLAG_RD, 1026 &ARC_mru.arcs_lsize[ARC_BUFC_METADATA], 0, "size of metadata in mru state"); 1027SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_data_lsize, CTLFLAG_RD, 1028 &ARC_mru.arcs_lsize[ARC_BUFC_DATA], 0, "size of data in mru state"); 1029 1030SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_size, CTLFLAG_RD, 1031 &ARC_mru_ghost.arcs_size, 0, "size of mru ghost state"); 1032SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_metadata_lsize, CTLFLAG_RD, 1033 &ARC_mru_ghost.arcs_lsize[ARC_BUFC_METADATA], 0, 1034 "size of metadata in mru ghost state"); 1035SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_data_lsize, CTLFLAG_RD, 1036 &ARC_mru_ghost.arcs_lsize[ARC_BUFC_DATA], 0, 1037 "size of data in mru ghost state"); 1038 1039SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_size, CTLFLAG_RD, 1040 &ARC_mfu.arcs_size, 0, "size of mfu state"); 1041SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_metadata_lsize, CTLFLAG_RD, 1042 &ARC_mfu.arcs_lsize[ARC_BUFC_METADATA], 0, "size of metadata in mfu state"); 1043SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_data_lsize, CTLFLAG_RD, 1044 &ARC_mfu.arcs_lsize[ARC_BUFC_DATA], 0, "size of data in mfu state"); 1045 1046SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_size, CTLFLAG_RD, 1047 &ARC_mfu_ghost.arcs_size, 0, "size of mfu ghost state"); 1048SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_metadata_lsize, CTLFLAG_RD, 1049 &ARC_mfu_ghost.arcs_lsize[ARC_BUFC_METADATA], 0, 1050 "size of metadata in mfu ghost state"); 1051SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_data_lsize, CTLFLAG_RD, 1052 &ARC_mfu_ghost.arcs_lsize[ARC_BUFC_DATA], 0, 1053 "size of data in mfu ghost state"); 1054 1055SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2c_only_size, CTLFLAG_RD, 1056 &ARC_l2c_only.arcs_size, 0, "size of mru state"); 1057 1058/* 1059 * L2ARC Internals 1060 */ 1061struct l2arc_dev { 1062 vdev_t *l2ad_vdev; /* vdev */ 1063 spa_t *l2ad_spa; /* spa */ 1064 uint64_t l2ad_hand; /* next write location */ 1065 uint64_t l2ad_start; /* first addr on device */ 1066 uint64_t l2ad_end; /* last addr on device */ 1067 boolean_t l2ad_first; /* first sweep through */ 1068 boolean_t l2ad_writing; /* currently writing */ 1069 kmutex_t l2ad_mtx; /* lock for buffer list */ 1070 list_t l2ad_buflist; /* buffer list */ 1071 list_node_t l2ad_node; /* device list node */ 1072 refcount_t l2ad_alloc; /* allocated bytes */ 1073}; 1074 1075static list_t L2ARC_dev_list; /* device list */ 1076static list_t *l2arc_dev_list; /* device list pointer */ 1077static kmutex_t l2arc_dev_mtx; /* device list mutex */ 1078static l2arc_dev_t *l2arc_dev_last; /* last device used */ 1079static list_t L2ARC_free_on_write; /* free after write buf list */ 1080static list_t *l2arc_free_on_write; /* free after write list ptr */ 1081static kmutex_t l2arc_free_on_write_mtx; /* mutex for list */ 1082static uint64_t l2arc_ndev; /* number of devices */ 1083 1084typedef struct l2arc_read_callback { 1085 arc_buf_t *l2rcb_buf; /* read buffer */ 1086 spa_t *l2rcb_spa; /* spa */ 1087 blkptr_t l2rcb_bp; /* original blkptr */ 1088 zbookmark_phys_t l2rcb_zb; /* original bookmark */ 1089 int l2rcb_flags; /* original flags */ 1090 enum zio_compress l2rcb_compress; /* applied compress */ 1091} l2arc_read_callback_t; 1092 1093typedef struct l2arc_write_callback { 1094 l2arc_dev_t *l2wcb_dev; /* device info */ 1095 arc_buf_hdr_t *l2wcb_head; /* head of write buflist */ 1096} l2arc_write_callback_t; 1097 1098typedef struct l2arc_data_free { 1099 /* protected by l2arc_free_on_write_mtx */ 1100 void *l2df_data; 1101 size_t l2df_size; 1102 void (*l2df_func)(void *, size_t); 1103 list_node_t l2df_list_node; 1104} l2arc_data_free_t; 1105 1106static kmutex_t l2arc_feed_thr_lock; 1107static kcondvar_t l2arc_feed_thr_cv; 1108static uint8_t l2arc_thread_exit; 1109 1110static void arc_get_data_buf(arc_buf_t *); 1111static void arc_access(arc_buf_hdr_t *, kmutex_t *); 1112static int arc_evict_needed(arc_buf_contents_t); 1113static void arc_evict_ghost(arc_state_t *, uint64_t, int64_t); 1114static void arc_buf_watch(arc_buf_t *); 1115 1116static arc_buf_contents_t arc_buf_type(arc_buf_hdr_t *); 1117static uint32_t arc_bufc_to_flags(arc_buf_contents_t); 1118 1119static boolean_t l2arc_write_eligible(uint64_t, arc_buf_hdr_t *); 1120static void l2arc_read_done(zio_t *); 1121 1122static boolean_t l2arc_compress_buf(arc_buf_hdr_t *); 1123static void l2arc_decompress_zio(zio_t *, arc_buf_hdr_t *, enum zio_compress); 1124static void l2arc_release_cdata_buf(arc_buf_hdr_t *); 1125 1126static uint64_t 1127buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth) 1128{ 1129 uint8_t *vdva = (uint8_t *)dva; 1130 uint64_t crc = -1ULL; 1131 int i; 1132 1133 ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY); 1134 1135 for (i = 0; i < sizeof (dva_t); i++) 1136 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ vdva[i]) & 0xFF]; 1137 1138 crc ^= (spa>>8) ^ birth; 1139 1140 return (crc); 1141} 1142 1143#define BUF_EMPTY(buf) \ 1144 ((buf)->b_dva.dva_word[0] == 0 && \ 1145 (buf)->b_dva.dva_word[1] == 0) 1146 1147#define BUF_EQUAL(spa, dva, birth, buf) \ 1148 ((buf)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \ 1149 ((buf)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \ 1150 ((buf)->b_birth == birth) && ((buf)->b_spa == spa) 1151 1152static void 1153buf_discard_identity(arc_buf_hdr_t *hdr) 1154{ 1155 hdr->b_dva.dva_word[0] = 0; 1156 hdr->b_dva.dva_word[1] = 0; 1157 hdr->b_birth = 0; 1158} 1159 1160static arc_buf_hdr_t * 1161buf_hash_find(uint64_t spa, const blkptr_t *bp, kmutex_t **lockp) 1162{ 1163 const dva_t *dva = BP_IDENTITY(bp); 1164 uint64_t birth = BP_PHYSICAL_BIRTH(bp); 1165 uint64_t idx = BUF_HASH_INDEX(spa, dva, birth); 1166 kmutex_t *hash_lock = BUF_HASH_LOCK(idx); 1167 arc_buf_hdr_t *hdr; 1168 1169 mutex_enter(hash_lock); 1170 for (hdr = buf_hash_table.ht_table[idx]; hdr != NULL; 1171 hdr = hdr->b_hash_next) { 1172 if (BUF_EQUAL(spa, dva, birth, hdr)) { 1173 *lockp = hash_lock; 1174 return (hdr); 1175 } 1176 } 1177 mutex_exit(hash_lock); 1178 *lockp = NULL; 1179 return (NULL); 1180} 1181 1182/* 1183 * Insert an entry into the hash table. If there is already an element 1184 * equal to elem in the hash table, then the already existing element 1185 * will be returned and the new element will not be inserted. 1186 * Otherwise returns NULL. 1187 * If lockp == NULL, the caller is assumed to already hold the hash lock. 1188 */ 1189static arc_buf_hdr_t * 1190buf_hash_insert(arc_buf_hdr_t *hdr, kmutex_t **lockp) 1191{ 1192 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth); 1193 kmutex_t *hash_lock = BUF_HASH_LOCK(idx); 1194 arc_buf_hdr_t *fhdr; 1195 uint32_t i; 1196 1197 ASSERT(!DVA_IS_EMPTY(&hdr->b_dva)); 1198 ASSERT(hdr->b_birth != 0); 1199 ASSERT(!HDR_IN_HASH_TABLE(hdr)); 1200 1201 if (lockp != NULL) { 1202 *lockp = hash_lock; 1203 mutex_enter(hash_lock); 1204 } else { 1205 ASSERT(MUTEX_HELD(hash_lock)); 1206 } 1207 1208 for (fhdr = buf_hash_table.ht_table[idx], i = 0; fhdr != NULL; 1209 fhdr = fhdr->b_hash_next, i++) { 1210 if (BUF_EQUAL(hdr->b_spa, &hdr->b_dva, hdr->b_birth, fhdr)) 1211 return (fhdr); 1212 } 1213 1214 hdr->b_hash_next = buf_hash_table.ht_table[idx]; 1215 buf_hash_table.ht_table[idx] = hdr; 1216 hdr->b_flags |= ARC_FLAG_IN_HASH_TABLE; 1217 1218 /* collect some hash table performance data */ 1219 if (i > 0) { 1220 ARCSTAT_BUMP(arcstat_hash_collisions); 1221 if (i == 1) 1222 ARCSTAT_BUMP(arcstat_hash_chains); 1223 1224 ARCSTAT_MAX(arcstat_hash_chain_max, i); 1225 } 1226 1227 ARCSTAT_BUMP(arcstat_hash_elements); 1228 ARCSTAT_MAXSTAT(arcstat_hash_elements); 1229 1230 return (NULL); 1231} 1232 1233static void 1234buf_hash_remove(arc_buf_hdr_t *hdr) 1235{ 1236 arc_buf_hdr_t *fhdr, **hdrp; 1237 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth); 1238 1239 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx))); 1240 ASSERT(HDR_IN_HASH_TABLE(hdr)); 1241 1242 hdrp = &buf_hash_table.ht_table[idx]; 1243 while ((fhdr = *hdrp) != hdr) { 1244 ASSERT(fhdr != NULL); 1245 hdrp = &fhdr->b_hash_next; 1246 } 1247 *hdrp = hdr->b_hash_next; 1248 hdr->b_hash_next = NULL; 1249 hdr->b_flags &= ~ARC_FLAG_IN_HASH_TABLE; 1250 1251 /* collect some hash table performance data */ 1252 ARCSTAT_BUMPDOWN(arcstat_hash_elements); 1253 1254 if (buf_hash_table.ht_table[idx] && 1255 buf_hash_table.ht_table[idx]->b_hash_next == NULL) 1256 ARCSTAT_BUMPDOWN(arcstat_hash_chains); 1257} 1258 1259/* 1260 * Global data structures and functions for the buf kmem cache. 1261 */ 1262static kmem_cache_t *hdr_full_cache; 1263static kmem_cache_t *hdr_l2only_cache; 1264static kmem_cache_t *buf_cache; 1265 1266static void 1267buf_fini(void) 1268{ 1269 int i; 1270 1271 kmem_free(buf_hash_table.ht_table, 1272 (buf_hash_table.ht_mask + 1) * sizeof (void *)); 1273 for (i = 0; i < BUF_LOCKS; i++) 1274 mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock); 1275 kmem_cache_destroy(hdr_full_cache); 1276 kmem_cache_destroy(hdr_l2only_cache); 1277 kmem_cache_destroy(buf_cache); 1278} 1279 1280/* 1281 * Constructor callback - called when the cache is empty 1282 * and a new buf is requested. 1283 */ 1284/* ARGSUSED */ 1285static int 1286hdr_full_cons(void *vbuf, void *unused, int kmflag) 1287{ 1288 arc_buf_hdr_t *hdr = vbuf; 1289 1290 bzero(hdr, HDR_FULL_SIZE); 1291 cv_init(&hdr->b_l1hdr.b_cv, NULL, CV_DEFAULT, NULL); 1292 refcount_create(&hdr->b_l1hdr.b_refcnt); 1293 mutex_init(&hdr->b_l1hdr.b_freeze_lock, NULL, MUTEX_DEFAULT, NULL); 1294 arc_space_consume(HDR_FULL_SIZE, ARC_SPACE_HDRS); 1295 1296 return (0); 1297} 1298 1299/* ARGSUSED */ 1300static int 1301hdr_l2only_cons(void *vbuf, void *unused, int kmflag) 1302{ 1303 arc_buf_hdr_t *hdr = vbuf; 1304 1305 bzero(hdr, HDR_L2ONLY_SIZE); 1306 arc_space_consume(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS); 1307 1308 return (0); 1309} 1310 1311/* ARGSUSED */ 1312static int 1313buf_cons(void *vbuf, void *unused, int kmflag) 1314{ 1315 arc_buf_t *buf = vbuf; 1316 1317 bzero(buf, sizeof (arc_buf_t)); 1318 mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL); 1319 arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS); 1320 1321 return (0); 1322} 1323 1324/* 1325 * Destructor callback - called when a cached buf is 1326 * no longer required. 1327 */ 1328/* ARGSUSED */ 1329static void 1330hdr_full_dest(void *vbuf, void *unused) 1331{ 1332 arc_buf_hdr_t *hdr = vbuf; 1333 1334 ASSERT(BUF_EMPTY(hdr)); 1335 cv_destroy(&hdr->b_l1hdr.b_cv); 1336 refcount_destroy(&hdr->b_l1hdr.b_refcnt); 1337 mutex_destroy(&hdr->b_l1hdr.b_freeze_lock); 1338 arc_space_return(HDR_FULL_SIZE, ARC_SPACE_HDRS); 1339} 1340 1341/* ARGSUSED */ 1342static void 1343hdr_l2only_dest(void *vbuf, void *unused) 1344{ 1345 arc_buf_hdr_t *hdr = vbuf; 1346 1347 ASSERT(BUF_EMPTY(hdr)); 1348 arc_space_return(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS); 1349} 1350 1351/* ARGSUSED */ 1352static void 1353buf_dest(void *vbuf, void *unused) 1354{ 1355 arc_buf_t *buf = vbuf; 1356 1357 mutex_destroy(&buf->b_evict_lock); 1358 arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS); 1359} 1360 1361/* 1362 * Reclaim callback -- invoked when memory is low. 1363 */ 1364/* ARGSUSED */ 1365static void 1366hdr_recl(void *unused) 1367{ 1368 dprintf("hdr_recl called\n"); 1369 /* 1370 * umem calls the reclaim func when we destroy the buf cache, 1371 * which is after we do arc_fini(). 1372 */ 1373 if (!arc_dead) 1374 cv_signal(&arc_reclaim_thr_cv); 1375} 1376 1377static void 1378buf_init(void) 1379{ 1380 uint64_t *ct; 1381 uint64_t hsize = 1ULL << 12; 1382 int i, j; 1383 1384 /* 1385 * The hash table is big enough to fill all of physical memory 1386 * with an average block size of zfs_arc_average_blocksize (default 8K). 1387 * By default, the table will take up 1388 * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers). 1389 */ 1390 while (hsize * zfs_arc_average_blocksize < (uint64_t)physmem * PAGESIZE) 1391 hsize <<= 1; 1392retry: 1393 buf_hash_table.ht_mask = hsize - 1; 1394 buf_hash_table.ht_table = 1395 kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP); 1396 if (buf_hash_table.ht_table == NULL) { 1397 ASSERT(hsize > (1ULL << 8)); 1398 hsize >>= 1; 1399 goto retry; 1400 } 1401 1402 hdr_full_cache = kmem_cache_create("arc_buf_hdr_t_full", HDR_FULL_SIZE, 1403 0, hdr_full_cons, hdr_full_dest, hdr_recl, NULL, NULL, 0); 1404 hdr_l2only_cache = kmem_cache_create("arc_buf_hdr_t_l2only", 1405 HDR_L2ONLY_SIZE, 0, hdr_l2only_cons, hdr_l2only_dest, hdr_recl, 1406 NULL, NULL, 0); 1407 buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t), 1408 0, buf_cons, buf_dest, NULL, NULL, NULL, 0); 1409 1410 for (i = 0; i < 256; i++) 1411 for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--) 1412 *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY); 1413 1414 for (i = 0; i < BUF_LOCKS; i++) { 1415 mutex_init(&buf_hash_table.ht_locks[i].ht_lock, 1416 NULL, MUTEX_DEFAULT, NULL); 1417 } 1418} 1419 1420/* 1421 * Transition between the two allocation states for the arc_buf_hdr struct. 1422 * The arc_buf_hdr struct can be allocated with (hdr_full_cache) or without 1423 * (hdr_l2only_cache) the fields necessary for the L1 cache - the smaller 1424 * version is used when a cache buffer is only in the L2ARC in order to reduce 1425 * memory usage. 1426 */ 1427static arc_buf_hdr_t * 1428arc_hdr_realloc(arc_buf_hdr_t *hdr, kmem_cache_t *old, kmem_cache_t *new) 1429{ 1430 ASSERT(HDR_HAS_L2HDR(hdr)); 1431 1432 arc_buf_hdr_t *nhdr; 1433 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev; 1434 1435 ASSERT((old == hdr_full_cache && new == hdr_l2only_cache) || 1436 (old == hdr_l2only_cache && new == hdr_full_cache)); 1437 1438 nhdr = kmem_cache_alloc(new, KM_PUSHPAGE); 1439 1440 ASSERT(MUTEX_HELD(HDR_LOCK(hdr))); 1441 buf_hash_remove(hdr); 1442 1443 bcopy(hdr, nhdr, HDR_L2ONLY_SIZE); 1444 1445 if (new == hdr_full_cache) { 1446 nhdr->b_flags |= ARC_FLAG_HAS_L1HDR; 1447 /* 1448 * arc_access and arc_change_state need to be aware that a 1449 * header has just come out of L2ARC, so we set its state to 1450 * l2c_only even though it's about to change. 1451 */ 1452 nhdr->b_l1hdr.b_state = arc_l2c_only; 1453 } else { 1454 ASSERT(hdr->b_l1hdr.b_buf == NULL); 1455 ASSERT0(hdr->b_l1hdr.b_datacnt); 1456 ASSERT(!list_link_active(&hdr->b_l1hdr.b_arc_node)); 1457 /* 1458 * We might be removing the L1hdr of a buffer which was just 1459 * written out to L2ARC. If such a buffer is compressed then we 1460 * need to free its b_tmp_cdata before destroying the header. 1461 */ 1462 if (hdr->b_l1hdr.b_tmp_cdata != NULL && 1463 HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF) 1464 l2arc_release_cdata_buf(hdr); 1465 nhdr->b_flags &= ~ARC_FLAG_HAS_L1HDR; 1466 } 1467 /* 1468 * The header has been reallocated so we need to re-insert it into any 1469 * lists it was on. 1470 */ 1471 (void) buf_hash_insert(nhdr, NULL); 1472 1473 ASSERT(list_link_active(&hdr->b_l2hdr.b_l2node)); 1474 1475 mutex_enter(&dev->l2ad_mtx); 1476 1477 /* 1478 * We must place the realloc'ed header back into the list at 1479 * the same spot. Otherwise, if it's placed earlier in the list, 1480 * l2arc_write_buffers() could find it during the function's 1481 * write phase, and try to write it out to the l2arc. 1482 */ 1483 list_insert_after(&dev->l2ad_buflist, hdr, nhdr); 1484 list_remove(&dev->l2ad_buflist, hdr); 1485 1486 mutex_exit(&dev->l2ad_mtx); 1487 1488 /* 1489 * Since we're using the pointer address as the tag when 1490 * incrementing and decrementing the l2ad_alloc refcount, we 1491 * must remove the old pointer (that we're about to destroy) and 1492 * add the new pointer to the refcount. Otherwise we'd remove 1493 * the wrong pointer address when calling arc_hdr_destroy() later. 1494 */ 1495 1496 (void) refcount_remove_many(&dev->l2ad_alloc, 1497 hdr->b_l2hdr.b_asize, hdr); 1498 1499 (void) refcount_add_many(&dev->l2ad_alloc, 1500 nhdr->b_l2hdr.b_asize, nhdr); 1501 1502 buf_discard_identity(hdr); 1503 hdr->b_freeze_cksum = NULL; 1504 kmem_cache_free(old, hdr); 1505 1506 return (nhdr); 1507} 1508 1509 1510#define ARC_MINTIME (hz>>4) /* 62 ms */ 1511 1512static void 1513arc_cksum_verify(arc_buf_t *buf) 1514{ 1515 zio_cksum_t zc; 1516 1517 if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 1518 return; 1519 1520 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1521 if (buf->b_hdr->b_freeze_cksum == NULL || HDR_IO_ERROR(buf->b_hdr)) { 1522 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1523 return; 1524 } 1525 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc); 1526 if (!ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc)) 1527 panic("buffer modified while frozen!"); 1528 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1529} 1530 1531static int 1532arc_cksum_equal(arc_buf_t *buf) 1533{ 1534 zio_cksum_t zc; 1535 int equal; 1536 1537 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1538 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, &zc); 1539 equal = ZIO_CHECKSUM_EQUAL(*buf->b_hdr->b_freeze_cksum, zc); 1540 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1541 1542 return (equal); 1543} 1544 1545static void 1546arc_cksum_compute(arc_buf_t *buf, boolean_t force) 1547{ 1548 if (!force && !(zfs_flags & ZFS_DEBUG_MODIFY)) 1549 return; 1550 1551 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1552 if (buf->b_hdr->b_freeze_cksum != NULL) { 1553 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1554 return; 1555 } 1556 buf->b_hdr->b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t), KM_SLEEP); 1557 fletcher_2_native(buf->b_data, buf->b_hdr->b_size, 1558 buf->b_hdr->b_freeze_cksum); 1559 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1560#ifdef illumos 1561 arc_buf_watch(buf); 1562#endif /* illumos */ 1563} 1564 1565#ifdef illumos 1566#ifndef _KERNEL 1567typedef struct procctl { 1568 long cmd; 1569 prwatch_t prwatch; 1570} procctl_t; 1571#endif 1572 1573/* ARGSUSED */ 1574static void 1575arc_buf_unwatch(arc_buf_t *buf) 1576{ 1577#ifndef _KERNEL 1578 if (arc_watch) { 1579 int result; 1580 procctl_t ctl; 1581 ctl.cmd = PCWATCH; 1582 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data; 1583 ctl.prwatch.pr_size = 0; 1584 ctl.prwatch.pr_wflags = 0; 1585 result = write(arc_procfd, &ctl, sizeof (ctl)); 1586 ASSERT3U(result, ==, sizeof (ctl)); 1587 } 1588#endif 1589} 1590 1591/* ARGSUSED */ 1592static void 1593arc_buf_watch(arc_buf_t *buf) 1594{ 1595#ifndef _KERNEL 1596 if (arc_watch) { 1597 int result; 1598 procctl_t ctl; 1599 ctl.cmd = PCWATCH; 1600 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data; 1601 ctl.prwatch.pr_size = buf->b_hdr->b_size; 1602 ctl.prwatch.pr_wflags = WA_WRITE; 1603 result = write(arc_procfd, &ctl, sizeof (ctl)); 1604 ASSERT3U(result, ==, sizeof (ctl)); 1605 } 1606#endif 1607} 1608#endif /* illumos */ 1609 1610static arc_buf_contents_t 1611arc_buf_type(arc_buf_hdr_t *hdr) 1612{ 1613 if (HDR_ISTYPE_METADATA(hdr)) { 1614 return (ARC_BUFC_METADATA); 1615 } else { 1616 return (ARC_BUFC_DATA); 1617 } 1618} 1619 1620static uint32_t 1621arc_bufc_to_flags(arc_buf_contents_t type) 1622{ 1623 switch (type) { 1624 case ARC_BUFC_DATA: 1625 /* metadata field is 0 if buffer contains normal data */ 1626 return (0); 1627 case ARC_BUFC_METADATA: 1628 return (ARC_FLAG_BUFC_METADATA); 1629 default: 1630 break; 1631 } 1632 panic("undefined ARC buffer type!"); 1633 return ((uint32_t)-1); 1634} 1635 1636void 1637arc_buf_thaw(arc_buf_t *buf) 1638{ 1639 if (zfs_flags & ZFS_DEBUG_MODIFY) { 1640 if (buf->b_hdr->b_l1hdr.b_state != arc_anon) 1641 panic("modifying non-anon buffer!"); 1642 if (HDR_IO_IN_PROGRESS(buf->b_hdr)) 1643 panic("modifying buffer while i/o in progress!"); 1644 arc_cksum_verify(buf); 1645 } 1646 1647 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1648 if (buf->b_hdr->b_freeze_cksum != NULL) { 1649 kmem_free(buf->b_hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 1650 buf->b_hdr->b_freeze_cksum = NULL; 1651 } 1652 1653#ifdef ZFS_DEBUG 1654 if (zfs_flags & ZFS_DEBUG_MODIFY) { 1655 if (buf->b_hdr->b_l1hdr.b_thawed != NULL) 1656 kmem_free(buf->b_hdr->b_l1hdr.b_thawed, 1); 1657 buf->b_hdr->b_l1hdr.b_thawed = kmem_alloc(1, KM_SLEEP); 1658 } 1659#endif 1660 1661 mutex_exit(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1662 1663#ifdef illumos 1664 arc_buf_unwatch(buf); 1665#endif /* illumos */ 1666} 1667 1668void 1669arc_buf_freeze(arc_buf_t *buf) 1670{ 1671 kmutex_t *hash_lock; 1672 1673 if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 1674 return; 1675 1676 hash_lock = HDR_LOCK(buf->b_hdr); 1677 mutex_enter(hash_lock); 1678 1679 ASSERT(buf->b_hdr->b_freeze_cksum != NULL || 1680 buf->b_hdr->b_l1hdr.b_state == arc_anon); 1681 arc_cksum_compute(buf, B_FALSE); 1682 mutex_exit(hash_lock); 1683 1684} 1685 1686static void 1687get_buf_info(arc_buf_hdr_t *hdr, arc_state_t *state, list_t **list, kmutex_t **lock) 1688{ 1689 uint64_t buf_hashid = buf_hash(hdr->b_spa, &hdr->b_dva, hdr->b_birth); 1690 1691 if (arc_buf_type(hdr) == ARC_BUFC_METADATA) 1692 buf_hashid &= (ARC_BUFC_NUMMETADATALISTS - 1); 1693 else { 1694 buf_hashid &= (ARC_BUFC_NUMDATALISTS - 1); 1695 buf_hashid += ARC_BUFC_NUMMETADATALISTS; 1696 } 1697 1698 *list = &state->arcs_lists[buf_hashid]; 1699 *lock = ARCS_LOCK(state, buf_hashid); 1700} 1701 1702 1703static void 1704add_reference(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, void *tag) 1705{ 1706 ASSERT(HDR_HAS_L1HDR(hdr)); 1707 ASSERT(MUTEX_HELD(hash_lock)); 1708 arc_state_t *state = hdr->b_l1hdr.b_state; 1709 1710 if ((refcount_add(&hdr->b_l1hdr.b_refcnt, tag) == 1) && 1711 (state != arc_anon)) { 1712 /* We don't use the L2-only state list. */ 1713 if (state != arc_l2c_only) { 1714 uint64_t delta = hdr->b_size * hdr->b_l1hdr.b_datacnt; 1715 uint64_t *size = &state->arcs_lsize[arc_buf_type(hdr)]; 1716 list_t *list; 1717 kmutex_t *lock; 1718 1719 get_buf_info(hdr, state, &list, &lock); 1720 ASSERT(!MUTEX_HELD(lock)); 1721 mutex_enter(lock); 1722 ASSERT(list_link_active(&hdr->b_l1hdr.b_arc_node)); 1723 list_remove(list, hdr); 1724 if (GHOST_STATE(state)) { 1725 ASSERT0(hdr->b_l1hdr.b_datacnt); 1726 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 1727 delta = hdr->b_size; 1728 } 1729 ASSERT(delta > 0); 1730 ASSERT3U(*size, >=, delta); 1731 atomic_add_64(size, -delta); 1732 mutex_exit(lock); 1733 } 1734 /* remove the prefetch flag if we get a reference */ 1735 hdr->b_flags &= ~ARC_FLAG_PREFETCH; 1736 } 1737} 1738 1739static int 1740remove_reference(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, void *tag) 1741{ 1742 int cnt; 1743 arc_state_t *state = hdr->b_l1hdr.b_state; 1744 1745 ASSERT(HDR_HAS_L1HDR(hdr)); 1746 ASSERT(state == arc_anon || MUTEX_HELD(hash_lock)); 1747 ASSERT(!GHOST_STATE(state)); 1748 1749 /* 1750 * arc_l2c_only counts as a ghost state so we don't need to explicitly 1751 * check to prevent usage of the arc_l2c_only list. 1752 */ 1753 if (((cnt = refcount_remove(&hdr->b_l1hdr.b_refcnt, tag)) == 0) && 1754 (state != arc_anon)) { 1755 uint64_t *size = &state->arcs_lsize[arc_buf_type(hdr)]; 1756 list_t *list; 1757 kmutex_t *lock; 1758 1759 get_buf_info(hdr, state, &list, &lock); 1760 ASSERT(!MUTEX_HELD(lock)); 1761 mutex_enter(lock); 1762 ASSERT(!list_link_active(&hdr->b_l1hdr.b_arc_node)); 1763 list_insert_head(list, hdr); 1764 ASSERT(hdr->b_l1hdr.b_datacnt > 0); 1765 atomic_add_64(size, hdr->b_size * 1766 hdr->b_l1hdr.b_datacnt); 1767 mutex_exit(lock); 1768 } 1769 return (cnt); 1770} 1771 1772/* 1773 * Move the supplied buffer to the indicated state. The mutex 1774 * for the buffer must be held by the caller. 1775 */ 1776static void 1777arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *hdr, 1778 kmutex_t *hash_lock) 1779{ 1780 arc_state_t *old_state; 1781 int64_t refcnt; 1782 uint32_t datacnt; 1783 uint64_t from_delta, to_delta; 1784 arc_buf_contents_t buftype = arc_buf_type(hdr); 1785 list_t *list; 1786 kmutex_t *lock; 1787 1788 /* 1789 * We almost always have an L1 hdr here, since we call arc_hdr_realloc() 1790 * in arc_read() when bringing a buffer out of the L2ARC. However, the 1791 * L1 hdr doesn't always exist when we change state to arc_anon before 1792 * destroying a header, in which case reallocating to add the L1 hdr is 1793 * pointless. 1794 */ 1795 if (HDR_HAS_L1HDR(hdr)) { 1796 old_state = hdr->b_l1hdr.b_state; 1797 refcnt = refcount_count(&hdr->b_l1hdr.b_refcnt); 1798 datacnt = hdr->b_l1hdr.b_datacnt; 1799 } else { 1800 old_state = arc_l2c_only; 1801 refcnt = 0; 1802 datacnt = 0; 1803 } 1804 1805 ASSERT(MUTEX_HELD(hash_lock)); 1806 ASSERT3P(new_state, !=, old_state); 1807 ASSERT(refcnt == 0 || datacnt > 0); 1808 ASSERT(!GHOST_STATE(new_state) || datacnt == 0); 1809 ASSERT(old_state != arc_anon || datacnt <= 1); 1810 1811 from_delta = to_delta = datacnt * hdr->b_size; 1812 1813 /* 1814 * If this buffer is evictable, transfer it from the 1815 * old state list to the new state list. 1816 */ 1817 if (refcnt == 0) { 1818 if (old_state != arc_anon && old_state != arc_l2c_only) { 1819 int use_mutex; 1820 uint64_t *size = &old_state->arcs_lsize[buftype]; 1821 1822 get_buf_info(hdr, old_state, &list, &lock); 1823 use_mutex = !MUTEX_HELD(lock); 1824 if (use_mutex) 1825 mutex_enter(lock); 1826 1827 ASSERT(HDR_HAS_L1HDR(hdr)); 1828 ASSERT(list_link_active(&hdr->b_l1hdr.b_arc_node)); 1829 list_remove(list, hdr); 1830 1831 /* 1832 * If prefetching out of the ghost cache, 1833 * we will have a non-zero datacnt. 1834 */ 1835 if (GHOST_STATE(old_state) && datacnt == 0) { 1836 /* ghost elements have a ghost size */ 1837 ASSERT(hdr->b_l1hdr.b_buf == NULL); 1838 from_delta = hdr->b_size; 1839 } 1840 ASSERT3U(*size, >=, from_delta); 1841 atomic_add_64(size, -from_delta); 1842 1843 if (use_mutex) 1844 mutex_exit(lock); 1845 } 1846 if (new_state != arc_anon && new_state != arc_l2c_only) { 1847 int use_mutex; 1848 uint64_t *size = &new_state->arcs_lsize[buftype]; 1849 1850 /* 1851 * An L1 header always exists here, since if we're 1852 * moving to some L1-cached state (i.e. not l2c_only or 1853 * anonymous), we realloc the header to add an L1hdr 1854 * beforehand. 1855 */ 1856 ASSERT(HDR_HAS_L1HDR(hdr)); 1857 get_buf_info(hdr, new_state, &list, &lock); 1858 use_mutex = !MUTEX_HELD(lock); 1859 if (use_mutex) 1860 mutex_enter(lock); 1861 1862 list_insert_head(list, hdr); 1863 1864 /* ghost elements have a ghost size */ 1865 if (GHOST_STATE(new_state)) { 1866 ASSERT(datacnt == 0); 1867 ASSERT(hdr->b_l1hdr.b_buf == NULL); 1868 to_delta = hdr->b_size; 1869 } 1870 atomic_add_64(size, to_delta); 1871 1872 if (use_mutex) 1873 mutex_exit(lock); 1874 } 1875 } 1876 1877 ASSERT(!BUF_EMPTY(hdr)); 1878 if (new_state == arc_anon && HDR_IN_HASH_TABLE(hdr)) 1879 buf_hash_remove(hdr); 1880 1881 /* adjust state sizes (ignore arc_l2c_only) */ 1882 if (to_delta && new_state != arc_l2c_only) 1883 atomic_add_64(&new_state->arcs_size, to_delta); 1884 if (from_delta && old_state != arc_l2c_only) { 1885 ASSERT3U(old_state->arcs_size, >=, from_delta); 1886 atomic_add_64(&old_state->arcs_size, -from_delta); 1887 } 1888 if (HDR_HAS_L1HDR(hdr)) 1889 hdr->b_l1hdr.b_state = new_state; 1890 1891 /* 1892 * L2 headers should never be on the L2 state list since they don't 1893 * have L1 headers allocated. 1894 */ 1895#ifdef illumos 1896 ASSERT(list_is_empty(&arc_l2c_only->arcs_list[ARC_BUFC_DATA]) && 1897 list_is_empty(&arc_l2c_only->arcs_list[ARC_BUFC_METADATA])); 1898#endif 1899} 1900 1901void 1902arc_space_consume(uint64_t space, arc_space_type_t type) 1903{ 1904 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES); 1905 1906 switch (type) { 1907 case ARC_SPACE_DATA: 1908 ARCSTAT_INCR(arcstat_data_size, space); 1909 break; 1910 case ARC_SPACE_META: 1911 ARCSTAT_INCR(arcstat_metadata_size, space); 1912 break; 1913 case ARC_SPACE_OTHER: 1914 ARCSTAT_INCR(arcstat_other_size, space); 1915 break; 1916 case ARC_SPACE_HDRS: 1917 ARCSTAT_INCR(arcstat_hdr_size, space); 1918 break; 1919 case ARC_SPACE_L2HDRS: 1920 ARCSTAT_INCR(arcstat_l2_hdr_size, space); 1921 break; 1922 } 1923 1924 if (type != ARC_SPACE_DATA) 1925 ARCSTAT_INCR(arcstat_meta_used, space); 1926 1927 atomic_add_64(&arc_size, space); 1928} 1929 1930void 1931arc_space_return(uint64_t space, arc_space_type_t type) 1932{ 1933 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES); 1934 1935 switch (type) { 1936 case ARC_SPACE_DATA: 1937 ARCSTAT_INCR(arcstat_data_size, -space); 1938 break; 1939 case ARC_SPACE_META: 1940 ARCSTAT_INCR(arcstat_metadata_size, -space); 1941 break; 1942 case ARC_SPACE_OTHER: 1943 ARCSTAT_INCR(arcstat_other_size, -space); 1944 break; 1945 case ARC_SPACE_HDRS: 1946 ARCSTAT_INCR(arcstat_hdr_size, -space); 1947 break; 1948 case ARC_SPACE_L2HDRS: 1949 ARCSTAT_INCR(arcstat_l2_hdr_size, -space); 1950 break; 1951 } 1952 1953 if (type != ARC_SPACE_DATA) { 1954 ASSERT(arc_meta_used >= space); 1955 if (arc_meta_max < arc_meta_used) 1956 arc_meta_max = arc_meta_used; 1957 ARCSTAT_INCR(arcstat_meta_used, -space); 1958 } 1959 1960 ASSERT(arc_size >= space); 1961 atomic_add_64(&arc_size, -space); 1962} 1963 1964arc_buf_t * 1965arc_buf_alloc(spa_t *spa, int32_t size, void *tag, arc_buf_contents_t type) 1966{ 1967 arc_buf_hdr_t *hdr; 1968 arc_buf_t *buf; 1969 1970 ASSERT3U(size, >, 0); 1971 hdr = kmem_cache_alloc(hdr_full_cache, KM_PUSHPAGE); 1972 ASSERT(BUF_EMPTY(hdr)); 1973 ASSERT3P(hdr->b_freeze_cksum, ==, NULL); 1974 hdr->b_size = size; 1975 hdr->b_spa = spa_load_guid(spa); 1976 1977 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 1978 buf->b_hdr = hdr; 1979 buf->b_data = NULL; 1980 buf->b_efunc = NULL; 1981 buf->b_private = NULL; 1982 buf->b_next = NULL; 1983 1984 hdr->b_flags = arc_bufc_to_flags(type); 1985 hdr->b_flags |= ARC_FLAG_HAS_L1HDR; 1986 1987 hdr->b_l1hdr.b_buf = buf; 1988 hdr->b_l1hdr.b_state = arc_anon; 1989 hdr->b_l1hdr.b_arc_access = 0; 1990 hdr->b_l1hdr.b_datacnt = 1; 1991 1992 arc_get_data_buf(buf); 1993 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 1994 (void) refcount_add(&hdr->b_l1hdr.b_refcnt, tag); 1995 1996 return (buf); 1997} 1998 1999static char *arc_onloan_tag = "onloan"; 2000 2001/* 2002 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in 2003 * flight data by arc_tempreserve_space() until they are "returned". Loaned 2004 * buffers must be returned to the arc before they can be used by the DMU or 2005 * freed. 2006 */ 2007arc_buf_t * 2008arc_loan_buf(spa_t *spa, int size) 2009{ 2010 arc_buf_t *buf; 2011 2012 buf = arc_buf_alloc(spa, size, arc_onloan_tag, ARC_BUFC_DATA); 2013 2014 atomic_add_64(&arc_loaned_bytes, size); 2015 return (buf); 2016} 2017 2018/* 2019 * Return a loaned arc buffer to the arc. 2020 */ 2021void 2022arc_return_buf(arc_buf_t *buf, void *tag) 2023{ 2024 arc_buf_hdr_t *hdr = buf->b_hdr; 2025 2026 ASSERT(buf->b_data != NULL); 2027 ASSERT(HDR_HAS_L1HDR(hdr)); 2028 (void) refcount_add(&hdr->b_l1hdr.b_refcnt, tag); 2029 (void) refcount_remove(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag); 2030 2031 atomic_add_64(&arc_loaned_bytes, -hdr->b_size); 2032} 2033 2034/* Detach an arc_buf from a dbuf (tag) */ 2035void 2036arc_loan_inuse_buf(arc_buf_t *buf, void *tag) 2037{ 2038 arc_buf_hdr_t *hdr = buf->b_hdr; 2039 2040 ASSERT(buf->b_data != NULL); 2041 ASSERT(HDR_HAS_L1HDR(hdr)); 2042 (void) refcount_add(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag); 2043 (void) refcount_remove(&hdr->b_l1hdr.b_refcnt, tag); 2044 buf->b_efunc = NULL; 2045 buf->b_private = NULL; 2046 2047 atomic_add_64(&arc_loaned_bytes, hdr->b_size); 2048} 2049 2050static arc_buf_t * 2051arc_buf_clone(arc_buf_t *from) 2052{ 2053 arc_buf_t *buf; 2054 arc_buf_hdr_t *hdr = from->b_hdr; 2055 uint64_t size = hdr->b_size; 2056 2057 ASSERT(HDR_HAS_L1HDR(hdr)); 2058 ASSERT(hdr->b_l1hdr.b_state != arc_anon); 2059 2060 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 2061 buf->b_hdr = hdr; 2062 buf->b_data = NULL; 2063 buf->b_efunc = NULL; 2064 buf->b_private = NULL; 2065 buf->b_next = hdr->b_l1hdr.b_buf; 2066 hdr->b_l1hdr.b_buf = buf; 2067 arc_get_data_buf(buf); 2068 bcopy(from->b_data, buf->b_data, size); 2069 2070 /* 2071 * This buffer already exists in the arc so create a duplicate 2072 * copy for the caller. If the buffer is associated with user data 2073 * then track the size and number of duplicates. These stats will be 2074 * updated as duplicate buffers are created and destroyed. 2075 */ 2076 if (HDR_ISTYPE_DATA(hdr)) { 2077 ARCSTAT_BUMP(arcstat_duplicate_buffers); 2078 ARCSTAT_INCR(arcstat_duplicate_buffers_size, size); 2079 } 2080 hdr->b_l1hdr.b_datacnt += 1; 2081 return (buf); 2082} 2083 2084void 2085arc_buf_add_ref(arc_buf_t *buf, void* tag) 2086{ 2087 arc_buf_hdr_t *hdr; 2088 kmutex_t *hash_lock; 2089 2090 /* 2091 * Check to see if this buffer is evicted. Callers 2092 * must verify b_data != NULL to know if the add_ref 2093 * was successful. 2094 */ 2095 mutex_enter(&buf->b_evict_lock); 2096 if (buf->b_data == NULL) { 2097 mutex_exit(&buf->b_evict_lock); 2098 return; 2099 } 2100 hash_lock = HDR_LOCK(buf->b_hdr); 2101 mutex_enter(hash_lock); 2102 hdr = buf->b_hdr; 2103 ASSERT(HDR_HAS_L1HDR(hdr)); 2104 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 2105 mutex_exit(&buf->b_evict_lock); 2106 2107 ASSERT(hdr->b_l1hdr.b_state == arc_mru || 2108 hdr->b_l1hdr.b_state == arc_mfu); 2109 2110 add_reference(hdr, hash_lock, tag); 2111 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr); 2112 arc_access(hdr, hash_lock); 2113 mutex_exit(hash_lock); 2114 ARCSTAT_BUMP(arcstat_hits); 2115 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr), 2116 demand, prefetch, !HDR_ISTYPE_METADATA(hdr), 2117 data, metadata, hits); 2118} 2119 2120static void 2121arc_buf_free_on_write(void *data, size_t size, 2122 void (*free_func)(void *, size_t)) 2123{ 2124 l2arc_data_free_t *df; 2125 2126 df = kmem_alloc(sizeof (l2arc_data_free_t), KM_SLEEP); 2127 df->l2df_data = data; 2128 df->l2df_size = size; 2129 df->l2df_func = free_func; 2130 mutex_enter(&l2arc_free_on_write_mtx); 2131 list_insert_head(l2arc_free_on_write, df); 2132 mutex_exit(&l2arc_free_on_write_mtx); 2133} 2134 2135/* 2136 * Free the arc data buffer. If it is an l2arc write in progress, 2137 * the buffer is placed on l2arc_free_on_write to be freed later. 2138 */ 2139static void 2140arc_buf_data_free(arc_buf_t *buf, void (*free_func)(void *, size_t)) 2141{ 2142 arc_buf_hdr_t *hdr = buf->b_hdr; 2143 2144 if (HDR_L2_WRITING(hdr)) { 2145 arc_buf_free_on_write(buf->b_data, hdr->b_size, free_func); 2146 ARCSTAT_BUMP(arcstat_l2_free_on_write); 2147 } else { 2148 free_func(buf->b_data, hdr->b_size); 2149 } 2150} 2151 2152/* 2153 * Free up buf->b_data and if 'remove' is set, then pull the 2154 * arc_buf_t off of the the arc_buf_hdr_t's list and free it. 2155 */ 2156static void 2157arc_buf_l2_cdata_free(arc_buf_hdr_t *hdr) 2158{ 2159 ASSERT(HDR_HAS_L2HDR(hdr)); 2160 ASSERT(MUTEX_HELD(&hdr->b_l2hdr.b_dev->l2ad_mtx)); 2161 2162 /* 2163 * The b_tmp_cdata field is linked off of the b_l1hdr, so if 2164 * that doesn't exist, the header is in the arc_l2c_only state, 2165 * and there isn't anything to free (it's already been freed). 2166 */ 2167 if (!HDR_HAS_L1HDR(hdr)) 2168 return; 2169 2170 if (hdr->b_l1hdr.b_tmp_cdata == NULL) 2171 return; 2172 2173 ASSERT(HDR_L2_WRITING(hdr)); 2174 arc_buf_free_on_write(hdr->b_l1hdr.b_tmp_cdata, hdr->b_size, 2175 zio_data_buf_free); 2176 2177 ARCSTAT_BUMP(arcstat_l2_cdata_free_on_write); 2178 hdr->b_l1hdr.b_tmp_cdata = NULL; 2179} 2180 2181static void 2182arc_buf_destroy(arc_buf_t *buf, boolean_t recycle, boolean_t remove) 2183{ 2184 arc_buf_t **bufp; 2185 2186 /* free up data associated with the buf */ 2187 if (buf->b_data != NULL) { 2188 arc_state_t *state = buf->b_hdr->b_l1hdr.b_state; 2189 uint64_t size = buf->b_hdr->b_size; 2190 arc_buf_contents_t type = arc_buf_type(buf->b_hdr); 2191 2192 arc_cksum_verify(buf); 2193#ifdef illumos 2194 arc_buf_unwatch(buf); 2195#endif /* illumos */ 2196 2197 if (!recycle) { 2198 if (type == ARC_BUFC_METADATA) { 2199 arc_buf_data_free(buf, zio_buf_free); 2200 arc_space_return(size, ARC_SPACE_META); 2201 } else { 2202 ASSERT(type == ARC_BUFC_DATA); 2203 arc_buf_data_free(buf, zio_data_buf_free); 2204 arc_space_return(size, ARC_SPACE_DATA); 2205 } 2206 } 2207 if (list_link_active(&buf->b_hdr->b_l1hdr.b_arc_node)) { 2208 uint64_t *cnt = &state->arcs_lsize[type]; 2209 2210 ASSERT(refcount_is_zero( 2211 &buf->b_hdr->b_l1hdr.b_refcnt)); 2212 ASSERT(state != arc_anon && state != arc_l2c_only); 2213 2214 ASSERT3U(*cnt, >=, size); 2215 atomic_add_64(cnt, -size); 2216 } 2217 ASSERT3U(state->arcs_size, >=, size); 2218 atomic_add_64(&state->arcs_size, -size); 2219 buf->b_data = NULL; 2220 2221 /* 2222 * If we're destroying a duplicate buffer make sure 2223 * that the appropriate statistics are updated. 2224 */ 2225 if (buf->b_hdr->b_l1hdr.b_datacnt > 1 && 2226 HDR_ISTYPE_DATA(buf->b_hdr)) { 2227 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers); 2228 ARCSTAT_INCR(arcstat_duplicate_buffers_size, -size); 2229 } 2230 ASSERT(buf->b_hdr->b_l1hdr.b_datacnt > 0); 2231 buf->b_hdr->b_l1hdr.b_datacnt -= 1; 2232 } 2233 2234 /* only remove the buf if requested */ 2235 if (!remove) 2236 return; 2237 2238 /* remove the buf from the hdr list */ 2239 for (bufp = &buf->b_hdr->b_l1hdr.b_buf; *bufp != buf; 2240 bufp = &(*bufp)->b_next) 2241 continue; 2242 *bufp = buf->b_next; 2243 buf->b_next = NULL; 2244 2245 ASSERT(buf->b_efunc == NULL); 2246 2247 /* clean up the buf */ 2248 buf->b_hdr = NULL; 2249 kmem_cache_free(buf_cache, buf); 2250} 2251 2252static void 2253arc_hdr_l2hdr_destroy(arc_buf_hdr_t *hdr) 2254{ 2255 l2arc_buf_hdr_t *l2hdr = &hdr->b_l2hdr; 2256 l2arc_dev_t *dev = l2hdr->b_dev; 2257 2258 ASSERT(MUTEX_HELD(&dev->l2ad_mtx)); 2259 ASSERT(HDR_HAS_L2HDR(hdr)); 2260 2261 list_remove(&dev->l2ad_buflist, hdr); 2262 2263 /* 2264 * We don't want to leak the b_tmp_cdata buffer that was 2265 * allocated in l2arc_write_buffers() 2266 */ 2267 arc_buf_l2_cdata_free(hdr); 2268 2269 /* 2270 * If the l2hdr's b_daddr is equal to L2ARC_ADDR_UNSET, then 2271 * this header is being processed by l2arc_write_buffers() (i.e. 2272 * it's in the first stage of l2arc_write_buffers()). 2273 * Re-affirming that truth here, just to serve as a reminder. If 2274 * b_daddr does not equal L2ARC_ADDR_UNSET, then the header may or 2275 * may not have its HDR_L2_WRITING flag set. (the write may have 2276 * completed, in which case HDR_L2_WRITING will be false and the 2277 * b_daddr field will point to the address of the buffer on disk). 2278 */ 2279 IMPLY(l2hdr->b_daddr == L2ARC_ADDR_UNSET, HDR_L2_WRITING(hdr)); 2280 2281 /* 2282 * If b_daddr is equal to L2ARC_ADDR_UNSET, we're racing with 2283 * l2arc_write_buffers(). Since we've just removed this header 2284 * from the l2arc buffer list, this header will never reach the 2285 * second stage of l2arc_write_buffers(), which increments the 2286 * accounting stats for this header. Thus, we must be careful 2287 * not to decrement them for this header either. 2288 */ 2289 if (l2hdr->b_daddr != L2ARC_ADDR_UNSET) { 2290 ARCSTAT_INCR(arcstat_l2_asize, -l2hdr->b_asize); 2291 ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size); 2292 2293 vdev_space_update(dev->l2ad_vdev, 2294 -l2hdr->b_asize, 0, 0); 2295 2296 (void) refcount_remove_many(&dev->l2ad_alloc, 2297 l2hdr->b_asize, hdr); 2298 } 2299 2300 hdr->b_flags &= ~ARC_FLAG_HAS_L2HDR; 2301} 2302 2303static void 2304arc_hdr_destroy(arc_buf_hdr_t *hdr) 2305{ 2306 if (HDR_HAS_L1HDR(hdr)) { 2307 ASSERT(hdr->b_l1hdr.b_buf == NULL || 2308 hdr->b_l1hdr.b_datacnt > 0); 2309 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 2310 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon); 2311 } 2312 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 2313 ASSERT(!HDR_IN_HASH_TABLE(hdr)); 2314 2315 if (HDR_HAS_L2HDR(hdr)) { 2316 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev; 2317 boolean_t buflist_held = MUTEX_HELD(&dev->l2ad_mtx); 2318 2319 if (!buflist_held) 2320 mutex_enter(&dev->l2ad_mtx); 2321 2322 /* 2323 * Even though we checked this conditional above, we 2324 * need to check this again now that we have the 2325 * l2ad_mtx. This is because we could be racing with 2326 * another thread calling l2arc_evict() which might have 2327 * destroyed this header's L2 portion as we were waiting 2328 * to acquire the l2ad_mtx. If that happens, we don't 2329 * want to re-destroy the header's L2 portion. 2330 */ 2331 if (HDR_HAS_L2HDR(hdr)) { 2332 if (hdr->b_l2hdr.b_daddr != L2ARC_ADDR_UNSET) 2333 trim_map_free(dev->l2ad_vdev, 2334 hdr->b_l2hdr.b_daddr, 2335 hdr->b_l2hdr.b_asize, 0); 2336 arc_hdr_l2hdr_destroy(hdr); 2337 } 2338 2339 if (!buflist_held) 2340 mutex_exit(&dev->l2ad_mtx); 2341 } 2342 2343 if (!BUF_EMPTY(hdr)) 2344 buf_discard_identity(hdr); 2345 if (hdr->b_freeze_cksum != NULL) { 2346 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 2347 hdr->b_freeze_cksum = NULL; 2348 } 2349 2350 if (HDR_HAS_L1HDR(hdr)) { 2351 while (hdr->b_l1hdr.b_buf) { 2352 arc_buf_t *buf = hdr->b_l1hdr.b_buf; 2353 2354 if (buf->b_efunc != NULL) { 2355 mutex_enter(&arc_eviction_mtx); 2356 mutex_enter(&buf->b_evict_lock); 2357 ASSERT(buf->b_hdr != NULL); 2358 arc_buf_destroy(hdr->b_l1hdr.b_buf, FALSE, 2359 FALSE); 2360 hdr->b_l1hdr.b_buf = buf->b_next; 2361 buf->b_hdr = &arc_eviction_hdr; 2362 buf->b_next = arc_eviction_list; 2363 arc_eviction_list = buf; 2364 mutex_exit(&buf->b_evict_lock); 2365 mutex_exit(&arc_eviction_mtx); 2366 } else { 2367 arc_buf_destroy(hdr->b_l1hdr.b_buf, FALSE, 2368 TRUE); 2369 } 2370 } 2371#ifdef ZFS_DEBUG 2372 if (hdr->b_l1hdr.b_thawed != NULL) { 2373 kmem_free(hdr->b_l1hdr.b_thawed, 1); 2374 hdr->b_l1hdr.b_thawed = NULL; 2375 } 2376#endif 2377 } 2378 2379 ASSERT3P(hdr->b_hash_next, ==, NULL); 2380 if (HDR_HAS_L1HDR(hdr)) { 2381 ASSERT(!list_link_active(&hdr->b_l1hdr.b_arc_node)); 2382 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL); 2383 kmem_cache_free(hdr_full_cache, hdr); 2384 } else { 2385 kmem_cache_free(hdr_l2only_cache, hdr); 2386 } 2387} 2388 2389void 2390arc_buf_free(arc_buf_t *buf, void *tag) 2391{ 2392 arc_buf_hdr_t *hdr = buf->b_hdr; 2393 int hashed = hdr->b_l1hdr.b_state != arc_anon; 2394 2395 ASSERT(buf->b_efunc == NULL); 2396 ASSERT(buf->b_data != NULL); 2397 2398 if (hashed) { 2399 kmutex_t *hash_lock = HDR_LOCK(hdr); 2400 2401 mutex_enter(hash_lock); 2402 hdr = buf->b_hdr; 2403 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 2404 2405 (void) remove_reference(hdr, hash_lock, tag); 2406 if (hdr->b_l1hdr.b_datacnt > 1) { 2407 arc_buf_destroy(buf, FALSE, TRUE); 2408 } else { 2409 ASSERT(buf == hdr->b_l1hdr.b_buf); 2410 ASSERT(buf->b_efunc == NULL); 2411 hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE; 2412 } 2413 mutex_exit(hash_lock); 2414 } else if (HDR_IO_IN_PROGRESS(hdr)) { 2415 int destroy_hdr; 2416 /* 2417 * We are in the middle of an async write. Don't destroy 2418 * this buffer unless the write completes before we finish 2419 * decrementing the reference count. 2420 */ 2421 mutex_enter(&arc_eviction_mtx); 2422 (void) remove_reference(hdr, NULL, tag); 2423 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 2424 destroy_hdr = !HDR_IO_IN_PROGRESS(hdr); 2425 mutex_exit(&arc_eviction_mtx); 2426 if (destroy_hdr) 2427 arc_hdr_destroy(hdr); 2428 } else { 2429 if (remove_reference(hdr, NULL, tag) > 0) 2430 arc_buf_destroy(buf, FALSE, TRUE); 2431 else 2432 arc_hdr_destroy(hdr); 2433 } 2434} 2435 2436boolean_t 2437arc_buf_remove_ref(arc_buf_t *buf, void* tag) 2438{ 2439 arc_buf_hdr_t *hdr = buf->b_hdr; 2440 kmutex_t *hash_lock = HDR_LOCK(hdr); 2441 boolean_t no_callback = (buf->b_efunc == NULL); 2442 2443 if (hdr->b_l1hdr.b_state == arc_anon) { 2444 ASSERT(hdr->b_l1hdr.b_datacnt == 1); 2445 arc_buf_free(buf, tag); 2446 return (no_callback); 2447 } 2448 2449 mutex_enter(hash_lock); 2450 hdr = buf->b_hdr; 2451 ASSERT(hdr->b_l1hdr.b_datacnt > 0); 2452 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 2453 ASSERT(hdr->b_l1hdr.b_state != arc_anon); 2454 ASSERT(buf->b_data != NULL); 2455 2456 (void) remove_reference(hdr, hash_lock, tag); 2457 if (hdr->b_l1hdr.b_datacnt > 1) { 2458 if (no_callback) 2459 arc_buf_destroy(buf, FALSE, TRUE); 2460 } else if (no_callback) { 2461 ASSERT(hdr->b_l1hdr.b_buf == buf && buf->b_next == NULL); 2462 ASSERT(buf->b_efunc == NULL); 2463 hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE; 2464 } 2465 ASSERT(no_callback || hdr->b_l1hdr.b_datacnt > 1 || 2466 refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 2467 mutex_exit(hash_lock); 2468 return (no_callback); 2469} 2470 2471int32_t 2472arc_buf_size(arc_buf_t *buf) 2473{ 2474 return (buf->b_hdr->b_size); 2475} 2476 2477/* 2478 * Called from the DMU to determine if the current buffer should be 2479 * evicted. In order to ensure proper locking, the eviction must be initiated 2480 * from the DMU. Return true if the buffer is associated with user data and 2481 * duplicate buffers still exist. 2482 */ 2483boolean_t 2484arc_buf_eviction_needed(arc_buf_t *buf) 2485{ 2486 arc_buf_hdr_t *hdr; 2487 boolean_t evict_needed = B_FALSE; 2488 2489 if (zfs_disable_dup_eviction) 2490 return (B_FALSE); 2491 2492 mutex_enter(&buf->b_evict_lock); 2493 hdr = buf->b_hdr; 2494 if (hdr == NULL) { 2495 /* 2496 * We are in arc_do_user_evicts(); let that function 2497 * perform the eviction. 2498 */ 2499 ASSERT(buf->b_data == NULL); 2500 mutex_exit(&buf->b_evict_lock); 2501 return (B_FALSE); 2502 } else if (buf->b_data == NULL) { 2503 /* 2504 * We have already been added to the arc eviction list; 2505 * recommend eviction. 2506 */ 2507 ASSERT3P(hdr, ==, &arc_eviction_hdr); 2508 mutex_exit(&buf->b_evict_lock); 2509 return (B_TRUE); 2510 } 2511 2512 if (hdr->b_l1hdr.b_datacnt > 1 && HDR_ISTYPE_DATA(hdr)) 2513 evict_needed = B_TRUE; 2514 2515 mutex_exit(&buf->b_evict_lock); 2516 return (evict_needed); 2517} 2518 2519/* 2520 * Evict buffers from list until we've removed the specified number of 2521 * bytes. Move the removed buffers to the appropriate evict state. 2522 * If the recycle flag is set, then attempt to "recycle" a buffer: 2523 * - look for a buffer to evict that is `bytes' long. 2524 * - return the data block from this buffer rather than freeing it. 2525 * This flag is used by callers that are trying to make space for a 2526 * new buffer in a full arc cache. 2527 * 2528 * This function makes a "best effort". It skips over any buffers 2529 * it can't get a hash_lock on, and so may not catch all candidates. 2530 * It may also return without evicting as much space as requested. 2531 */ 2532static void * 2533arc_evict(arc_state_t *state, uint64_t spa, int64_t bytes, boolean_t recycle, 2534 arc_buf_contents_t type) 2535{ 2536 arc_state_t *evicted_state; 2537 uint64_t bytes_evicted = 0, skipped = 0, missed = 0; 2538 int64_t bytes_remaining; 2539 arc_buf_hdr_t *hdr, *hdr_prev = NULL; 2540 list_t *evicted_list, *list, *evicted_list_start, *list_start; 2541 kmutex_t *lock, *evicted_lock; 2542 kmutex_t *hash_lock; 2543 boolean_t have_lock; 2544 void *stolen = NULL; 2545 arc_buf_hdr_t marker = { 0 }; 2546 int count = 0; 2547 static int evict_metadata_offset, evict_data_offset; 2548 int i, idx, offset, list_count, lists; 2549 2550 ASSERT(state == arc_mru || state == arc_mfu); 2551 2552 evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost; 2553 2554 /* 2555 * Decide which "type" (data vs metadata) to recycle from. 2556 * 2557 * If we are over the metadata limit, recycle from metadata. 2558 * If we are under the metadata minimum, recycle from data. 2559 * Otherwise, recycle from whichever type has the oldest (least 2560 * recently accessed) header. This is not yet implemented. 2561 */ 2562 if (recycle) { 2563 arc_buf_contents_t realtype; 2564 if (state->arcs_lsize[ARC_BUFC_DATA] == 0) { 2565 realtype = ARC_BUFC_METADATA; 2566 } else if (state->arcs_lsize[ARC_BUFC_METADATA] == 0) { 2567 realtype = ARC_BUFC_DATA; 2568 } else if (arc_meta_used >= arc_meta_limit) { 2569 realtype = ARC_BUFC_METADATA; 2570 } else if (arc_meta_used <= arc_meta_min) { 2571 realtype = ARC_BUFC_DATA; 2572#ifdef illumos 2573 } else if (HDR_HAS_L1HDR(data_hdr) && 2574 HDR_HAS_L1HDR(metadata_hdr) && 2575 data_hdr->b_l1hdr.b_arc_access < 2576 metadata_hdr->b_l1hdr.b_arc_access) { 2577 realtype = ARC_BUFC_DATA; 2578 } else { 2579 realtype = ARC_BUFC_METADATA; 2580#else 2581 } else { 2582 /* TODO */ 2583 realtype = type; 2584#endif 2585 } 2586 if (realtype != type) { 2587 /* 2588 * If we want to evict from a different list, 2589 * we can not recycle, because DATA vs METADATA 2590 * buffers are segregated into different kmem 2591 * caches (and vmem arenas). 2592 */ 2593 type = realtype; 2594 recycle = B_FALSE; 2595 } 2596 } 2597 2598 if (type == ARC_BUFC_METADATA) { 2599 offset = 0; 2600 list_count = ARC_BUFC_NUMMETADATALISTS; 2601 list_start = &state->arcs_lists[0]; 2602 evicted_list_start = &evicted_state->arcs_lists[0]; 2603 idx = evict_metadata_offset; 2604 } else { 2605 offset = ARC_BUFC_NUMMETADATALISTS; 2606 list_start = &state->arcs_lists[offset]; 2607 evicted_list_start = &evicted_state->arcs_lists[offset]; 2608 list_count = ARC_BUFC_NUMDATALISTS; 2609 idx = evict_data_offset; 2610 } 2611 bytes_remaining = evicted_state->arcs_lsize[type]; 2612 lists = 0; 2613 2614evict_start: 2615 list = &list_start[idx]; 2616 evicted_list = &evicted_list_start[idx]; 2617 lock = ARCS_LOCK(state, (offset + idx)); 2618 evicted_lock = ARCS_LOCK(evicted_state, (offset + idx)); 2619 2620 /* 2621 * The ghost list lock must be acquired first in order to prevent 2622 * a 3 party deadlock: 2623 * 2624 * - arc_evict_ghost acquires arc_*_ghost->arcs_mtx, followed by 2625 * l2ad_mtx in arc_hdr_realloc 2626 * - l2arc_write_buffers acquires l2ad_mtx, followed by arc_*->arcs_mtx 2627 * - arc_evict acquires arc_*_ghost->arcs_mtx, followed by 2628 * arc_*_ghost->arcs_mtx and forms a deadlock cycle. 2629 * 2630 * This situation is avoided by acquiring the ghost list lock first. 2631 */ 2632 mutex_enter(evicted_lock); 2633 mutex_enter(lock); 2634 2635 for (hdr = list_tail(list); hdr; hdr = hdr_prev) { 2636 hdr_prev = list_prev(list, hdr); 2637 if (HDR_HAS_L1HDR(hdr)) { 2638 bytes_remaining -= 2639 (hdr->b_size * hdr->b_l1hdr.b_datacnt); 2640 } 2641 /* prefetch buffers have a minimum lifespan */ 2642 if (HDR_IO_IN_PROGRESS(hdr) || 2643 (spa && hdr->b_spa != spa) || 2644 ((hdr->b_flags & (ARC_FLAG_PREFETCH | ARC_FLAG_INDIRECT)) && 2645 ddi_get_lbolt() - hdr->b_l1hdr.b_arc_access < 2646 arc_min_prefetch_lifespan)) { 2647 skipped++; 2648 continue; 2649 } 2650 /* "lookahead" for better eviction candidate */ 2651 if (recycle && hdr->b_size != bytes && 2652 hdr_prev && hdr_prev->b_size == bytes) 2653 continue; 2654 2655 /* ignore markers */ 2656 if (hdr->b_spa == 0) 2657 continue; 2658 2659 /* 2660 * It may take a long time to evict all the bufs requested. 2661 * To avoid blocking all arc activity, periodically drop 2662 * the arcs_mtx and give other threads a chance to run 2663 * before reacquiring the lock. 2664 * 2665 * If we are looking for a buffer to recycle, we are in 2666 * the hot code path, so don't sleep. 2667 */ 2668 if (!recycle && count++ > arc_evict_iterations) { 2669 list_insert_after(list, hdr, &marker); 2670 mutex_exit(lock); 2671 mutex_exit(evicted_lock); 2672 kpreempt(KPREEMPT_SYNC); 2673 mutex_enter(evicted_lock); 2674 mutex_enter(lock); 2675 hdr_prev = list_prev(list, &marker); 2676 list_remove(list, &marker); 2677 count = 0; 2678 continue; 2679 } 2680 2681 hash_lock = HDR_LOCK(hdr); 2682 have_lock = MUTEX_HELD(hash_lock); 2683 if (have_lock || mutex_tryenter(hash_lock)) { 2684 ASSERT0(refcount_count(&hdr->b_l1hdr.b_refcnt)); 2685 ASSERT3U(hdr->b_l1hdr.b_datacnt, >, 0); 2686 while (hdr->b_l1hdr.b_buf) { 2687 arc_buf_t *buf = hdr->b_l1hdr.b_buf; 2688 if (!mutex_tryenter(&buf->b_evict_lock)) { 2689 missed += 1; 2690 break; 2691 } 2692 if (buf->b_data != NULL) { 2693 bytes_evicted += hdr->b_size; 2694 if (recycle && 2695 arc_buf_type(hdr) == type && 2696 hdr->b_size == bytes && 2697 !HDR_L2_WRITING(hdr)) { 2698 stolen = buf->b_data; 2699 recycle = FALSE; 2700 } 2701 } 2702 if (buf->b_efunc != NULL) { 2703 mutex_enter(&arc_eviction_mtx); 2704 arc_buf_destroy(buf, 2705 buf->b_data == stolen, FALSE); 2706 hdr->b_l1hdr.b_buf = buf->b_next; 2707 buf->b_hdr = &arc_eviction_hdr; 2708 buf->b_next = arc_eviction_list; 2709 arc_eviction_list = buf; 2710 mutex_exit(&arc_eviction_mtx); 2711 mutex_exit(&buf->b_evict_lock); 2712 } else { 2713 mutex_exit(&buf->b_evict_lock); 2714 arc_buf_destroy(buf, 2715 buf->b_data == stolen, TRUE); 2716 } 2717 } 2718 2719 if (HDR_HAS_L2HDR(hdr)) { 2720 ARCSTAT_INCR(arcstat_evict_l2_cached, 2721 hdr->b_size); 2722 } else { 2723 if (l2arc_write_eligible(hdr->b_spa, hdr)) { 2724 ARCSTAT_INCR(arcstat_evict_l2_eligible, 2725 hdr->b_size); 2726 } else { 2727 ARCSTAT_INCR( 2728 arcstat_evict_l2_ineligible, 2729 hdr->b_size); 2730 } 2731 } 2732 2733 if (hdr->b_l1hdr.b_datacnt == 0) { 2734 arc_change_state(evicted_state, hdr, hash_lock); 2735 ASSERT(HDR_IN_HASH_TABLE(hdr)); 2736 hdr->b_flags |= ARC_FLAG_IN_HASH_TABLE; 2737 hdr->b_flags &= ~ARC_FLAG_BUF_AVAILABLE; 2738 DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, hdr); 2739 } 2740 if (!have_lock) 2741 mutex_exit(hash_lock); 2742 if (bytes >= 0 && bytes_evicted >= bytes) 2743 break; 2744 if (bytes_remaining > 0) { 2745 mutex_exit(evicted_lock); 2746 mutex_exit(lock); 2747 idx = ((idx + 1) & (list_count - 1)); 2748 lists++; 2749 goto evict_start; 2750 } 2751 } else { 2752 missed += 1; 2753 } 2754 } 2755 2756 mutex_exit(lock); 2757 mutex_exit(evicted_lock); 2758 2759 idx = ((idx + 1) & (list_count - 1)); 2760 lists++; 2761 2762 if (bytes_evicted < bytes) { 2763 if (lists < list_count) 2764 goto evict_start; 2765 else 2766 dprintf("only evicted %lld bytes from %x", 2767 (longlong_t)bytes_evicted, state); 2768 } 2769 if (type == ARC_BUFC_METADATA) 2770 evict_metadata_offset = idx; 2771 else 2772 evict_data_offset = idx; 2773 2774 if (skipped) 2775 ARCSTAT_INCR(arcstat_evict_skip, skipped); 2776 2777 if (missed) 2778 ARCSTAT_INCR(arcstat_mutex_miss, missed); 2779 2780 /* 2781 * Note: we have just evicted some data into the ghost state, 2782 * potentially putting the ghost size over the desired size. Rather 2783 * that evicting from the ghost list in this hot code path, leave 2784 * this chore to the arc_reclaim_thread(). 2785 */ 2786 2787 if (stolen) 2788 ARCSTAT_BUMP(arcstat_stolen); 2789 return (stolen); 2790} 2791 2792/* 2793 * Remove buffers from list until we've removed the specified number of 2794 * bytes. Destroy the buffers that are removed. 2795 */ 2796static void 2797arc_evict_ghost(arc_state_t *state, uint64_t spa, int64_t bytes) 2798{ 2799 arc_buf_hdr_t *hdr, *hdr_prev; 2800 arc_buf_hdr_t marker = { 0 }; 2801 list_t *list, *list_start; 2802 kmutex_t *hash_lock, *lock; 2803 uint64_t bytes_deleted = 0; 2804 uint64_t bufs_skipped = 0; 2805 int count = 0; 2806 static int evict_offset; 2807 int list_count, idx = evict_offset; 2808 int offset, lists = 0; 2809 2810 ASSERT(GHOST_STATE(state)); 2811 2812 /* 2813 * data lists come after metadata lists 2814 */ 2815 list_start = &state->arcs_lists[ARC_BUFC_NUMMETADATALISTS]; 2816 list_count = ARC_BUFC_NUMDATALISTS; 2817 offset = ARC_BUFC_NUMMETADATALISTS; 2818 2819evict_start: 2820 list = &list_start[idx]; 2821 lock = ARCS_LOCK(state, idx + offset); 2822 2823 mutex_enter(lock); 2824 for (hdr = list_tail(list); hdr; hdr = hdr_prev) { 2825 hdr_prev = list_prev(list, hdr); 2826 if (arc_buf_type(hdr) >= ARC_BUFC_NUMTYPES) 2827 panic("invalid hdr=%p", (void *)hdr); 2828 if (spa && hdr->b_spa != spa) 2829 continue; 2830 2831 /* ignore markers */ 2832 if (hdr->b_spa == 0) 2833 continue; 2834 2835 hash_lock = HDR_LOCK(hdr); 2836 /* caller may be trying to modify this buffer, skip it */ 2837 if (MUTEX_HELD(hash_lock)) 2838 continue; 2839 2840 /* 2841 * It may take a long time to evict all the bufs requested. 2842 * To avoid blocking all arc activity, periodically drop 2843 * the arcs_mtx and give other threads a chance to run 2844 * before reacquiring the lock. 2845 */ 2846 if (count++ > arc_evict_iterations) { 2847 list_insert_after(list, hdr, &marker); 2848 mutex_exit(lock); 2849 kpreempt(KPREEMPT_SYNC); 2850 mutex_enter(lock); 2851 hdr_prev = list_prev(list, &marker); 2852 list_remove(list, &marker); 2853 count = 0; 2854 continue; 2855 } 2856 if (mutex_tryenter(hash_lock)) { 2857 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 2858 ASSERT(!HDR_HAS_L1HDR(hdr) || 2859 hdr->b_l1hdr.b_buf == NULL); 2860 ARCSTAT_BUMP(arcstat_deleted); 2861 bytes_deleted += hdr->b_size; 2862 2863 if (HDR_HAS_L2HDR(hdr)) { 2864 /* 2865 * This buffer is cached on the 2nd Level ARC; 2866 * don't destroy the header. 2867 */ 2868 arc_change_state(arc_l2c_only, hdr, hash_lock); 2869 /* 2870 * dropping from L1+L2 cached to L2-only, 2871 * realloc to remove the L1 header. 2872 */ 2873 hdr = arc_hdr_realloc(hdr, hdr_full_cache, 2874 hdr_l2only_cache); 2875 mutex_exit(hash_lock); 2876 } else { 2877 arc_change_state(arc_anon, hdr, hash_lock); 2878 mutex_exit(hash_lock); 2879 arc_hdr_destroy(hdr); 2880 } 2881 2882 DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, hdr); 2883 if (bytes >= 0 && bytes_deleted >= bytes) 2884 break; 2885 } else if (bytes < 0) { 2886 /* 2887 * Insert a list marker and then wait for the 2888 * hash lock to become available. Once its 2889 * available, restart from where we left off. 2890 */ 2891 list_insert_after(list, hdr, &marker); 2892 mutex_exit(lock); 2893 mutex_enter(hash_lock); 2894 mutex_exit(hash_lock); 2895 mutex_enter(lock); 2896 hdr_prev = list_prev(list, &marker); 2897 list_remove(list, &marker); 2898 } else { 2899 bufs_skipped += 1; 2900 } 2901 2902 } 2903 mutex_exit(lock); 2904 idx = ((idx + 1) & (ARC_BUFC_NUMDATALISTS - 1)); 2905 lists++; 2906 2907 if (lists < list_count) 2908 goto evict_start; 2909 2910 evict_offset = idx; 2911 if ((uintptr_t)list > (uintptr_t)&state->arcs_lists[ARC_BUFC_NUMMETADATALISTS] && 2912 (bytes < 0 || bytes_deleted < bytes)) { 2913 list_start = &state->arcs_lists[0]; 2914 list_count = ARC_BUFC_NUMMETADATALISTS; 2915 offset = lists = 0; 2916 goto evict_start; 2917 } 2918 2919 if (bufs_skipped) { 2920 ARCSTAT_INCR(arcstat_mutex_miss, bufs_skipped); 2921 ASSERT(bytes >= 0); 2922 } 2923 2924 if (bytes_deleted < bytes) 2925 dprintf("only deleted %lld bytes from %p", 2926 (longlong_t)bytes_deleted, state); 2927} 2928 2929static void 2930arc_adjust(void) 2931{ 2932 int64_t adjustment, delta; 2933 2934 /* 2935 * Adjust MRU size 2936 */ 2937 2938 adjustment = MIN((int64_t)(arc_size - arc_c), 2939 (int64_t)(arc_anon->arcs_size + arc_mru->arcs_size + arc_meta_used - 2940 arc_p)); 2941 2942 if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_DATA] > 0) { 2943 delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_DATA], adjustment); 2944 (void) arc_evict(arc_mru, 0, delta, FALSE, ARC_BUFC_DATA); 2945 adjustment -= delta; 2946 } 2947 2948 if (adjustment > 0 && arc_mru->arcs_lsize[ARC_BUFC_METADATA] > 0) { 2949 delta = MIN(arc_mru->arcs_lsize[ARC_BUFC_METADATA], adjustment); 2950 (void) arc_evict(arc_mru, 0, delta, FALSE, 2951 ARC_BUFC_METADATA); 2952 } 2953 2954 /* 2955 * Adjust MFU size 2956 */ 2957 2958 adjustment = arc_size - arc_c; 2959 2960 if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_DATA] > 0) { 2961 delta = MIN(adjustment, arc_mfu->arcs_lsize[ARC_BUFC_DATA]); 2962 (void) arc_evict(arc_mfu, 0, delta, FALSE, ARC_BUFC_DATA); 2963 adjustment -= delta; 2964 } 2965 2966 if (adjustment > 0 && arc_mfu->arcs_lsize[ARC_BUFC_METADATA] > 0) { 2967 int64_t delta = MIN(adjustment, 2968 arc_mfu->arcs_lsize[ARC_BUFC_METADATA]); 2969 (void) arc_evict(arc_mfu, 0, delta, FALSE, 2970 ARC_BUFC_METADATA); 2971 } 2972 2973 /* 2974 * Adjust ghost lists 2975 */ 2976 2977 adjustment = arc_mru->arcs_size + arc_mru_ghost->arcs_size - arc_c; 2978 2979 if (adjustment > 0 && arc_mru_ghost->arcs_size > 0) { 2980 delta = MIN(arc_mru_ghost->arcs_size, adjustment); 2981 arc_evict_ghost(arc_mru_ghost, 0, delta); 2982 } 2983 2984 adjustment = 2985 arc_mru_ghost->arcs_size + arc_mfu_ghost->arcs_size - arc_c; 2986 2987 if (adjustment > 0 && arc_mfu_ghost->arcs_size > 0) { 2988 delta = MIN(arc_mfu_ghost->arcs_size, adjustment); 2989 arc_evict_ghost(arc_mfu_ghost, 0, delta); 2990 } 2991} 2992 2993static void 2994arc_do_user_evicts(void) 2995{ 2996 static arc_buf_t *tmp_arc_eviction_list; 2997 2998 /* 2999 * Move list over to avoid LOR 3000 */ 3001restart: 3002 mutex_enter(&arc_eviction_mtx); 3003 tmp_arc_eviction_list = arc_eviction_list; 3004 arc_eviction_list = NULL; 3005 mutex_exit(&arc_eviction_mtx); 3006 3007 while (tmp_arc_eviction_list != NULL) { 3008 arc_buf_t *buf = tmp_arc_eviction_list; 3009 tmp_arc_eviction_list = buf->b_next; 3010 mutex_enter(&buf->b_evict_lock); 3011 buf->b_hdr = NULL; 3012 mutex_exit(&buf->b_evict_lock); 3013 3014 if (buf->b_efunc != NULL) 3015 VERIFY0(buf->b_efunc(buf->b_private)); 3016 3017 buf->b_efunc = NULL; 3018 buf->b_private = NULL; 3019 kmem_cache_free(buf_cache, buf); 3020 } 3021 3022 if (arc_eviction_list != NULL) 3023 goto restart; 3024} 3025 3026/* 3027 * Flush all *evictable* data from the cache for the given spa. 3028 * NOTE: this will not touch "active" (i.e. referenced) data. 3029 */ 3030void 3031arc_flush(spa_t *spa) 3032{ 3033 uint64_t guid = 0; 3034 3035 if (spa != NULL) 3036 guid = spa_load_guid(spa); 3037 3038 while (arc_mru->arcs_lsize[ARC_BUFC_DATA]) { 3039 (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_DATA); 3040 if (spa != NULL) 3041 break; 3042 } 3043 while (arc_mru->arcs_lsize[ARC_BUFC_METADATA]) { 3044 (void) arc_evict(arc_mru, guid, -1, FALSE, ARC_BUFC_METADATA); 3045 if (spa != NULL) 3046 break; 3047 } 3048 while (arc_mfu->arcs_lsize[ARC_BUFC_DATA]) { 3049 (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_DATA); 3050 if (spa != NULL) 3051 break; 3052 } 3053 while (arc_mfu->arcs_lsize[ARC_BUFC_METADATA]) { 3054 (void) arc_evict(arc_mfu, guid, -1, FALSE, ARC_BUFC_METADATA); 3055 if (spa != NULL) 3056 break; 3057 } 3058 3059 arc_evict_ghost(arc_mru_ghost, guid, -1); 3060 arc_evict_ghost(arc_mfu_ghost, guid, -1); 3061 3062 mutex_enter(&arc_reclaim_thr_lock); 3063 arc_do_user_evicts(); 3064 mutex_exit(&arc_reclaim_thr_lock); 3065 ASSERT(spa || arc_eviction_list == NULL); 3066} 3067 3068void 3069arc_shrink(int64_t to_free) 3070{ 3071 3072 if (arc_c > arc_c_min) { 3073 DTRACE_PROBE4(arc__shrink, uint64_t, arc_c, uint64_t, 3074 arc_c_min, uint64_t, arc_p, uint64_t, to_free); 3075 if (arc_c > arc_c_min + to_free) 3076 atomic_add_64(&arc_c, -to_free); 3077 else 3078 arc_c = arc_c_min; 3079 3080 atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift)); 3081 if (arc_c > arc_size) 3082 arc_c = MAX(arc_size, arc_c_min); 3083 if (arc_p > arc_c) 3084 arc_p = (arc_c >> 1); 3085 3086 DTRACE_PROBE2(arc__shrunk, uint64_t, arc_c, uint64_t, 3087 arc_p); 3088 3089 ASSERT(arc_c >= arc_c_min); 3090 ASSERT((int64_t)arc_p >= 0); 3091 } 3092 3093 if (arc_size > arc_c) { 3094 DTRACE_PROBE2(arc__shrink_adjust, uint64_t, arc_size, 3095 uint64_t, arc_c); 3096 arc_adjust(); 3097 } 3098} 3099 3100static long needfree = 0; 3101 3102typedef enum free_memory_reason_t { 3103 FMR_UNKNOWN, 3104 FMR_NEEDFREE, 3105 FMR_LOTSFREE, 3106 FMR_SWAPFS_MINFREE, 3107 FMR_PAGES_PP_MAXIMUM, 3108 FMR_HEAP_ARENA, 3109 FMR_ZIO_ARENA, 3110 FMR_ZIO_FRAG, 3111} free_memory_reason_t; 3112 3113int64_t last_free_memory; 3114free_memory_reason_t last_free_reason; 3115 3116/* 3117 * Additional reserve of pages for pp_reserve. 3118 */ 3119int64_t arc_pages_pp_reserve = 64; 3120 3121/* 3122 * Additional reserve of pages for swapfs. 3123 */ 3124int64_t arc_swapfs_reserve = 64; 3125 3126/* 3127 * Return the amount of memory that can be consumed before reclaim will be 3128 * needed. Positive if there is sufficient free memory, negative indicates 3129 * the amount of memory that needs to be freed up. 3130 */ 3131static int64_t 3132arc_available_memory(void) 3133{ 3134 int64_t lowest = INT64_MAX; 3135 int64_t n; 3136 free_memory_reason_t r = FMR_UNKNOWN; 3137 3138#ifdef _KERNEL 3139 if (needfree > 0) { 3140 n = PAGESIZE * (-needfree); 3141 if (n < lowest) { 3142 lowest = n; 3143 r = FMR_NEEDFREE; 3144 } 3145 } 3146 3147 /* 3148 * Cooperate with pagedaemon when it's time for it to scan 3149 * and reclaim some pages. 3150 */ 3151 n = PAGESIZE * (int64_t)(freemem - zfs_arc_free_target); 3152 if (n < lowest) { 3153 lowest = n; 3154 r = FMR_LOTSFREE; 3155 } 3156 3157#ifdef sun 3158 /* 3159 * check that we're out of range of the pageout scanner. It starts to 3160 * schedule paging if freemem is less than lotsfree and needfree. 3161 * lotsfree is the high-water mark for pageout, and needfree is the 3162 * number of needed free pages. We add extra pages here to make sure 3163 * the scanner doesn't start up while we're freeing memory. 3164 */ 3165 n = PAGESIZE * (freemem - lotsfree - needfree - desfree); 3166 if (n < lowest) { 3167 lowest = n; 3168 r = FMR_LOTSFREE; 3169 } 3170 3171 /* 3172 * check to make sure that swapfs has enough space so that anon 3173 * reservations can still succeed. anon_resvmem() checks that the 3174 * availrmem is greater than swapfs_minfree, and the number of reserved 3175 * swap pages. We also add a bit of extra here just to prevent 3176 * circumstances from getting really dire. 3177 */ 3178 n = PAGESIZE * (availrmem - swapfs_minfree - swapfs_reserve - 3179 desfree - arc_swapfs_reserve); 3180 if (n < lowest) { 3181 lowest = n; 3182 r = FMR_SWAPFS_MINFREE; 3183 } 3184 3185 3186 /* 3187 * Check that we have enough availrmem that memory locking (e.g., via 3188 * mlock(3C) or memcntl(2)) can still succeed. (pages_pp_maximum 3189 * stores the number of pages that cannot be locked; when availrmem 3190 * drops below pages_pp_maximum, page locking mechanisms such as 3191 * page_pp_lock() will fail.) 3192 */ 3193 n = PAGESIZE * (availrmem - pages_pp_maximum - 3194 arc_pages_pp_reserve); 3195 if (n < lowest) { 3196 lowest = n; 3197 r = FMR_PAGES_PP_MAXIMUM; 3198 } 3199 3200#endif /* sun */ 3201#if defined(__i386) || !defined(UMA_MD_SMALL_ALLOC) 3202 /* 3203 * If we're on an i386 platform, it's possible that we'll exhaust the 3204 * kernel heap space before we ever run out of available physical 3205 * memory. Most checks of the size of the heap_area compare against 3206 * tune.t_minarmem, which is the minimum available real memory that we 3207 * can have in the system. However, this is generally fixed at 25 pages 3208 * which is so low that it's useless. In this comparison, we seek to 3209 * calculate the total heap-size, and reclaim if more than 3/4ths of the 3210 * heap is allocated. (Or, in the calculation, if less than 1/4th is 3211 * free) 3212 */ 3213 n = vmem_size(heap_arena, VMEM_FREE) - 3214 (vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC) >> 2); 3215 if (n < lowest) { 3216 lowest = n; 3217 r = FMR_HEAP_ARENA; 3218 } 3219#define zio_arena NULL 3220#else 3221#define zio_arena heap_arena 3222#endif 3223 3224 /* 3225 * If zio data pages are being allocated out of a separate heap segment, 3226 * then enforce that the size of available vmem for this arena remains 3227 * above about 1/16th free. 3228 * 3229 * Note: The 1/16th arena free requirement was put in place 3230 * to aggressively evict memory from the arc in order to avoid 3231 * memory fragmentation issues. 3232 */ 3233 if (zio_arena != NULL) { 3234 n = vmem_size(zio_arena, VMEM_FREE) - 3235 (vmem_size(zio_arena, VMEM_ALLOC) >> 4); 3236 if (n < lowest) { 3237 lowest = n; 3238 r = FMR_ZIO_ARENA; 3239 } 3240 } 3241 3242 /* 3243 * Above limits know nothing about real level of KVA fragmentation. 3244 * Start aggressive reclamation if too little sequential KVA left. 3245 */ 3246 if (lowest > 0) { 3247 n = (vmem_size(heap_arena, VMEM_MAXFREE) < zfs_max_recordsize) ? 3248 -(vmem_size(heap_arena, VMEM_ALLOC) >> 4) : INT64_MAX; 3249 if (n < lowest) { 3250 lowest = n; 3251 r = FMR_ZIO_FRAG; 3252 } 3253 } 3254 3255#else /* _KERNEL */ 3256 /* Every 100 calls, free a small amount */ 3257 if (spa_get_random(100) == 0) 3258 lowest = -1024; 3259#endif /* _KERNEL */ 3260 3261 last_free_memory = lowest; 3262 last_free_reason = r; 3263 DTRACE_PROBE2(arc__available_memory, int64_t, lowest, int, r); 3264 return (lowest); 3265} 3266 3267 3268/* 3269 * Determine if the system is under memory pressure and is asking 3270 * to reclaim memory. A return value of TRUE indicates that the system 3271 * is under memory pressure and that the arc should adjust accordingly. 3272 */ 3273static boolean_t 3274arc_reclaim_needed(void) 3275{ 3276 return (arc_available_memory() < 0); 3277} 3278 3279extern kmem_cache_t *zio_buf_cache[]; 3280extern kmem_cache_t *zio_data_buf_cache[]; 3281extern kmem_cache_t *range_seg_cache; 3282 3283static __noinline void 3284arc_kmem_reap_now(void) 3285{ 3286 size_t i; 3287 kmem_cache_t *prev_cache = NULL; 3288 kmem_cache_t *prev_data_cache = NULL; 3289 3290 DTRACE_PROBE(arc__kmem_reap_start); 3291#ifdef _KERNEL 3292 if (arc_meta_used >= arc_meta_limit) { 3293 /* 3294 * We are exceeding our meta-data cache limit. 3295 * Purge some DNLC entries to release holds on meta-data. 3296 */ 3297 dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent); 3298 } 3299#if defined(__i386) 3300 /* 3301 * Reclaim unused memory from all kmem caches. 3302 */ 3303 kmem_reap(); 3304#endif 3305#endif 3306 3307 for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) { 3308 if (zio_buf_cache[i] != prev_cache) { 3309 prev_cache = zio_buf_cache[i]; 3310 kmem_cache_reap_now(zio_buf_cache[i]); 3311 } 3312 if (zio_data_buf_cache[i] != prev_data_cache) { 3313 prev_data_cache = zio_data_buf_cache[i]; 3314 kmem_cache_reap_now(zio_data_buf_cache[i]); 3315 } 3316 } 3317 kmem_cache_reap_now(buf_cache); 3318 kmem_cache_reap_now(hdr_full_cache); 3319 kmem_cache_reap_now(hdr_l2only_cache); 3320 kmem_cache_reap_now(range_seg_cache); 3321 3322#ifdef sun 3323 if (zio_arena != NULL) { 3324 /* 3325 * Ask the vmem arena to reclaim unused memory from its 3326 * quantum caches. 3327 */ 3328 vmem_qcache_reap(zio_arena); 3329 } 3330#endif 3331 DTRACE_PROBE(arc__kmem_reap_end); 3332} 3333 3334static void 3335arc_reclaim_thread(void *dummy __unused) 3336{ 3337 clock_t growtime = 0; 3338 callb_cpr_t cpr; 3339 3340 CALLB_CPR_INIT(&cpr, &arc_reclaim_thr_lock, callb_generic_cpr, FTAG); 3341 3342 mutex_enter(&arc_reclaim_thr_lock); 3343 while (arc_thread_exit == 0) { 3344 int64_t free_memory = arc_available_memory(); 3345 if (free_memory < 0) { 3346 3347 arc_no_grow = B_TRUE; 3348 arc_warm = B_TRUE; 3349 3350 /* 3351 * Wait at least zfs_grow_retry (default 60) seconds 3352 * before considering growing. 3353 */ 3354 growtime = ddi_get_lbolt() + (arc_grow_retry * hz); 3355 3356 arc_kmem_reap_now(); 3357 3358 /* 3359 * If we are still low on memory, shrink the ARC 3360 * so that we have arc_shrink_min free space. 3361 */ 3362 free_memory = arc_available_memory(); 3363 3364 int64_t to_free = 3365 (arc_c >> arc_shrink_shift) - free_memory; 3366 if (to_free > 0) { 3367#ifdef _KERNEL 3368 to_free = MAX(to_free, ptob(needfree)); 3369#endif 3370 arc_shrink(to_free); 3371 } 3372 } else if (free_memory < arc_c >> arc_no_grow_shift) { 3373 arc_no_grow = B_TRUE; 3374 } else if (ddi_get_lbolt() >= growtime) { 3375 arc_no_grow = B_FALSE; 3376 } 3377 3378 arc_adjust(); 3379 3380 if (arc_eviction_list != NULL) 3381 arc_do_user_evicts(); 3382 3383#ifdef _KERNEL 3384 if (needfree) { 3385 needfree = 0; 3386 wakeup(&needfree); 3387 } 3388#endif 3389 3390 /* 3391 * This is necessary in order for the mdb ::arc dcmd to 3392 * show up to date information. Since the ::arc command 3393 * does not call the kstat's update function, without 3394 * this call, the command may show stale stats for the 3395 * anon, mru, mru_ghost, mfu, and mfu_ghost lists. Even 3396 * with this change, the data might be up to 1 second 3397 * out of date; but that should suffice. The arc_state_t 3398 * structures can be queried directly if more accurate 3399 * information is needed. 3400 */ 3401 if (arc_ksp != NULL) 3402 arc_ksp->ks_update(arc_ksp, KSTAT_READ); 3403 3404 /* block until needed, or one second, whichever is shorter */ 3405 CALLB_CPR_SAFE_BEGIN(&cpr); 3406 (void) cv_timedwait(&arc_reclaim_thr_cv, 3407 &arc_reclaim_thr_lock, hz); 3408 CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_thr_lock); 3409 } 3410 3411 arc_thread_exit = 0; 3412 cv_broadcast(&arc_reclaim_thr_cv); 3413 CALLB_CPR_EXIT(&cpr); /* drops arc_reclaim_thr_lock */ 3414 thread_exit(); 3415} 3416 3417/* 3418 * Adapt arc info given the number of bytes we are trying to add and 3419 * the state that we are comming from. This function is only called 3420 * when we are adding new content to the cache. 3421 */ 3422static void 3423arc_adapt(int bytes, arc_state_t *state) 3424{ 3425 int mult; 3426 uint64_t arc_p_min = (arc_c >> arc_p_min_shift); 3427 3428 if (state == arc_l2c_only) 3429 return; 3430 3431 ASSERT(bytes > 0); 3432 /* 3433 * Adapt the target size of the MRU list: 3434 * - if we just hit in the MRU ghost list, then increase 3435 * the target size of the MRU list. 3436 * - if we just hit in the MFU ghost list, then increase 3437 * the target size of the MFU list by decreasing the 3438 * target size of the MRU list. 3439 */ 3440 if (state == arc_mru_ghost) { 3441 mult = ((arc_mru_ghost->arcs_size >= arc_mfu_ghost->arcs_size) ? 3442 1 : (arc_mfu_ghost->arcs_size/arc_mru_ghost->arcs_size)); 3443 mult = MIN(mult, 10); /* avoid wild arc_p adjustment */ 3444 3445 arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult); 3446 } else if (state == arc_mfu_ghost) { 3447 uint64_t delta; 3448 3449 mult = ((arc_mfu_ghost->arcs_size >= arc_mru_ghost->arcs_size) ? 3450 1 : (arc_mru_ghost->arcs_size/arc_mfu_ghost->arcs_size)); 3451 mult = MIN(mult, 10); 3452 3453 delta = MIN(bytes * mult, arc_p); 3454 arc_p = MAX(arc_p_min, arc_p - delta); 3455 } 3456 ASSERT((int64_t)arc_p >= 0); 3457 3458 if (arc_reclaim_needed()) { 3459 cv_signal(&arc_reclaim_thr_cv); 3460 return; 3461 } 3462 3463 if (arc_no_grow) 3464 return; 3465 3466 if (arc_c >= arc_c_max) 3467 return; 3468 3469 /* 3470 * If we're within (2 * maxblocksize) bytes of the target 3471 * cache size, increment the target cache size 3472 */ 3473 if (arc_size > arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) { 3474 DTRACE_PROBE1(arc__inc_adapt, int, bytes); 3475 atomic_add_64(&arc_c, (int64_t)bytes); 3476 if (arc_c > arc_c_max) 3477 arc_c = arc_c_max; 3478 else if (state == arc_anon) 3479 atomic_add_64(&arc_p, (int64_t)bytes); 3480 if (arc_p > arc_c) 3481 arc_p = arc_c; 3482 } 3483 ASSERT((int64_t)arc_p >= 0); 3484} 3485 3486/* 3487 * Check if the cache has reached its limits and eviction is required 3488 * prior to insert. 3489 */ 3490static int 3491arc_evict_needed(arc_buf_contents_t type) 3492{ 3493 if (type == ARC_BUFC_METADATA && arc_meta_used >= arc_meta_limit) 3494 return (1); 3495 3496 if (arc_reclaim_needed()) 3497 return (1); 3498 3499 return (arc_size > arc_c); 3500} 3501 3502/* 3503 * The buffer, supplied as the first argument, needs a data block. 3504 * So, if we are at cache max, determine which cache should be victimized. 3505 * We have the following cases: 3506 * 3507 * 1. Insert for MRU, p > sizeof(arc_anon + arc_mru) -> 3508 * In this situation if we're out of space, but the resident size of the MFU is 3509 * under the limit, victimize the MFU cache to satisfy this insertion request. 3510 * 3511 * 2. Insert for MRU, p <= sizeof(arc_anon + arc_mru) -> 3512 * Here, we've used up all of the available space for the MRU, so we need to 3513 * evict from our own cache instead. Evict from the set of resident MRU 3514 * entries. 3515 * 3516 * 3. Insert for MFU (c - p) > sizeof(arc_mfu) -> 3517 * c minus p represents the MFU space in the cache, since p is the size of the 3518 * cache that is dedicated to the MRU. In this situation there's still space on 3519 * the MFU side, so the MRU side needs to be victimized. 3520 * 3521 * 4. Insert for MFU (c - p) < sizeof(arc_mfu) -> 3522 * MFU's resident set is consuming more space than it has been allotted. In 3523 * this situation, we must victimize our own cache, the MFU, for this insertion. 3524 */ 3525static void 3526arc_get_data_buf(arc_buf_t *buf) 3527{ 3528 arc_state_t *state = buf->b_hdr->b_l1hdr.b_state; 3529 uint64_t size = buf->b_hdr->b_size; 3530 arc_buf_contents_t type = arc_buf_type(buf->b_hdr); 3531 3532 arc_adapt(size, state); 3533 3534 /* 3535 * We have not yet reached cache maximum size, 3536 * just allocate a new buffer. 3537 */ 3538 if (!arc_evict_needed(type)) { 3539 if (type == ARC_BUFC_METADATA) { 3540 buf->b_data = zio_buf_alloc(size); 3541 arc_space_consume(size, ARC_SPACE_META); 3542 } else { 3543 ASSERT(type == ARC_BUFC_DATA); 3544 buf->b_data = zio_data_buf_alloc(size); 3545 arc_space_consume(size, ARC_SPACE_DATA); 3546 } 3547 goto out; 3548 } 3549 3550 /* 3551 * If we are prefetching from the mfu ghost list, this buffer 3552 * will end up on the mru list; so steal space from there. 3553 */ 3554 if (state == arc_mfu_ghost) 3555 state = HDR_PREFETCH(buf->b_hdr) ? arc_mru : arc_mfu; 3556 else if (state == arc_mru_ghost) 3557 state = arc_mru; 3558 3559 if (state == arc_mru || state == arc_anon) { 3560 uint64_t mru_used = arc_anon->arcs_size + arc_mru->arcs_size; 3561 state = (arc_mfu->arcs_lsize[type] >= size && 3562 arc_p > mru_used) ? arc_mfu : arc_mru; 3563 } else { 3564 /* MFU cases */ 3565 uint64_t mfu_space = arc_c - arc_p; 3566 state = (arc_mru->arcs_lsize[type] >= size && 3567 mfu_space > arc_mfu->arcs_size) ? arc_mru : arc_mfu; 3568 } 3569 if ((buf->b_data = arc_evict(state, 0, size, TRUE, type)) == NULL) { 3570 if (type == ARC_BUFC_METADATA) { 3571 buf->b_data = zio_buf_alloc(size); 3572 arc_space_consume(size, ARC_SPACE_META); 3573 } else { 3574 ASSERT(type == ARC_BUFC_DATA); 3575 buf->b_data = zio_data_buf_alloc(size); 3576 arc_space_consume(size, ARC_SPACE_DATA); 3577 } 3578 ARCSTAT_BUMP(arcstat_recycle_miss); 3579 } 3580 ASSERT(buf->b_data != NULL); 3581out: 3582 /* 3583 * Update the state size. Note that ghost states have a 3584 * "ghost size" and so don't need to be updated. 3585 */ 3586 if (!GHOST_STATE(buf->b_hdr->b_l1hdr.b_state)) { 3587 arc_buf_hdr_t *hdr = buf->b_hdr; 3588 3589 atomic_add_64(&hdr->b_l1hdr.b_state->arcs_size, size); 3590 if (list_link_active(&hdr->b_l1hdr.b_arc_node)) { 3591 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 3592 atomic_add_64(&hdr->b_l1hdr.b_state->arcs_lsize[type], 3593 size); 3594 } 3595 /* 3596 * If we are growing the cache, and we are adding anonymous 3597 * data, and we have outgrown arc_p, update arc_p 3598 */ 3599 if (arc_size < arc_c && hdr->b_l1hdr.b_state == arc_anon && 3600 arc_anon->arcs_size + arc_mru->arcs_size > arc_p) 3601 arc_p = MIN(arc_c, arc_p + size); 3602 } 3603 ARCSTAT_BUMP(arcstat_allocated); 3604} 3605 3606/* 3607 * This routine is called whenever a buffer is accessed. 3608 * NOTE: the hash lock is dropped in this function. 3609 */ 3610static void 3611arc_access(arc_buf_hdr_t *hdr, kmutex_t *hash_lock) 3612{ 3613 clock_t now; 3614 3615 ASSERT(MUTEX_HELD(hash_lock)); 3616 ASSERT(HDR_HAS_L1HDR(hdr)); 3617 3618 if (hdr->b_l1hdr.b_state == arc_anon) { 3619 /* 3620 * This buffer is not in the cache, and does not 3621 * appear in our "ghost" list. Add the new buffer 3622 * to the MRU state. 3623 */ 3624 3625 ASSERT0(hdr->b_l1hdr.b_arc_access); 3626 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 3627 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr); 3628 arc_change_state(arc_mru, hdr, hash_lock); 3629 3630 } else if (hdr->b_l1hdr.b_state == arc_mru) { 3631 now = ddi_get_lbolt(); 3632 3633 /* 3634 * If this buffer is here because of a prefetch, then either: 3635 * - clear the flag if this is a "referencing" read 3636 * (any subsequent access will bump this into the MFU state). 3637 * or 3638 * - move the buffer to the head of the list if this is 3639 * another prefetch (to make it less likely to be evicted). 3640 */ 3641 if (HDR_PREFETCH(hdr)) { 3642 if (refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) { 3643 ASSERT(list_link_active( 3644 &hdr->b_l1hdr.b_arc_node)); 3645 } else { 3646 hdr->b_flags &= ~ARC_FLAG_PREFETCH; 3647 ARCSTAT_BUMP(arcstat_mru_hits); 3648 } 3649 hdr->b_l1hdr.b_arc_access = now; 3650 return; 3651 } 3652 3653 /* 3654 * This buffer has been "accessed" only once so far, 3655 * but it is still in the cache. Move it to the MFU 3656 * state. 3657 */ 3658 if (now > hdr->b_l1hdr.b_arc_access + ARC_MINTIME) { 3659 /* 3660 * More than 125ms have passed since we 3661 * instantiated this buffer. Move it to the 3662 * most frequently used state. 3663 */ 3664 hdr->b_l1hdr.b_arc_access = now; 3665 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 3666 arc_change_state(arc_mfu, hdr, hash_lock); 3667 } 3668 ARCSTAT_BUMP(arcstat_mru_hits); 3669 } else if (hdr->b_l1hdr.b_state == arc_mru_ghost) { 3670 arc_state_t *new_state; 3671 /* 3672 * This buffer has been "accessed" recently, but 3673 * was evicted from the cache. Move it to the 3674 * MFU state. 3675 */ 3676 3677 if (HDR_PREFETCH(hdr)) { 3678 new_state = arc_mru; 3679 if (refcount_count(&hdr->b_l1hdr.b_refcnt) > 0) 3680 hdr->b_flags &= ~ARC_FLAG_PREFETCH; 3681 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr); 3682 } else { 3683 new_state = arc_mfu; 3684 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 3685 } 3686 3687 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 3688 arc_change_state(new_state, hdr, hash_lock); 3689 3690 ARCSTAT_BUMP(arcstat_mru_ghost_hits); 3691 } else if (hdr->b_l1hdr.b_state == arc_mfu) { 3692 /* 3693 * This buffer has been accessed more than once and is 3694 * still in the cache. Keep it in the MFU state. 3695 * 3696 * NOTE: an add_reference() that occurred when we did 3697 * the arc_read() will have kicked this off the list. 3698 * If it was a prefetch, we will explicitly move it to 3699 * the head of the list now. 3700 */ 3701 if ((HDR_PREFETCH(hdr)) != 0) { 3702 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 3703 ASSERT(list_link_active(&hdr->b_l1hdr.b_arc_node)); 3704 } 3705 ARCSTAT_BUMP(arcstat_mfu_hits); 3706 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 3707 } else if (hdr->b_l1hdr.b_state == arc_mfu_ghost) { 3708 arc_state_t *new_state = arc_mfu; 3709 /* 3710 * This buffer has been accessed more than once but has 3711 * been evicted from the cache. Move it back to the 3712 * MFU state. 3713 */ 3714 3715 if (HDR_PREFETCH(hdr)) { 3716 /* 3717 * This is a prefetch access... 3718 * move this block back to the MRU state. 3719 */ 3720 ASSERT0(refcount_count(&hdr->b_l1hdr.b_refcnt)); 3721 new_state = arc_mru; 3722 } 3723 3724 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 3725 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 3726 arc_change_state(new_state, hdr, hash_lock); 3727 3728 ARCSTAT_BUMP(arcstat_mfu_ghost_hits); 3729 } else if (hdr->b_l1hdr.b_state == arc_l2c_only) { 3730 /* 3731 * This buffer is on the 2nd Level ARC. 3732 */ 3733 3734 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 3735 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 3736 arc_change_state(arc_mfu, hdr, hash_lock); 3737 } else { 3738 ASSERT(!"invalid arc state"); 3739 } 3740} 3741 3742/* a generic arc_done_func_t which you can use */ 3743/* ARGSUSED */ 3744void 3745arc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg) 3746{ 3747 if (zio == NULL || zio->io_error == 0) 3748 bcopy(buf->b_data, arg, buf->b_hdr->b_size); 3749 VERIFY(arc_buf_remove_ref(buf, arg)); 3750} 3751 3752/* a generic arc_done_func_t */ 3753void 3754arc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg) 3755{ 3756 arc_buf_t **bufp = arg; 3757 if (zio && zio->io_error) { 3758 VERIFY(arc_buf_remove_ref(buf, arg)); 3759 *bufp = NULL; 3760 } else { 3761 *bufp = buf; 3762 ASSERT(buf->b_data); 3763 } 3764} 3765 3766static void 3767arc_read_done(zio_t *zio) 3768{ 3769 arc_buf_hdr_t *hdr; 3770 arc_buf_t *buf; 3771 arc_buf_t *abuf; /* buffer we're assigning to callback */ 3772 kmutex_t *hash_lock = NULL; 3773 arc_callback_t *callback_list, *acb; 3774 int freeable = FALSE; 3775 3776 buf = zio->io_private; 3777 hdr = buf->b_hdr; 3778 3779 /* 3780 * The hdr was inserted into hash-table and removed from lists 3781 * prior to starting I/O. We should find this header, since 3782 * it's in the hash table, and it should be legit since it's 3783 * not possible to evict it during the I/O. The only possible 3784 * reason for it not to be found is if we were freed during the 3785 * read. 3786 */ 3787 if (HDR_IN_HASH_TABLE(hdr)) { 3788 ASSERT3U(hdr->b_birth, ==, BP_PHYSICAL_BIRTH(zio->io_bp)); 3789 ASSERT3U(hdr->b_dva.dva_word[0], ==, 3790 BP_IDENTITY(zio->io_bp)->dva_word[0]); 3791 ASSERT3U(hdr->b_dva.dva_word[1], ==, 3792 BP_IDENTITY(zio->io_bp)->dva_word[1]); 3793 3794 arc_buf_hdr_t *found = buf_hash_find(hdr->b_spa, zio->io_bp, 3795 &hash_lock); 3796 3797 ASSERT((found == NULL && HDR_FREED_IN_READ(hdr) && 3798 hash_lock == NULL) || 3799 (found == hdr && 3800 DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) || 3801 (found == hdr && HDR_L2_READING(hdr))); 3802 } 3803 3804 hdr->b_flags &= ~ARC_FLAG_L2_EVICTED; 3805 if (l2arc_noprefetch && HDR_PREFETCH(hdr)) 3806 hdr->b_flags &= ~ARC_FLAG_L2CACHE; 3807 3808 /* byteswap if necessary */ 3809 callback_list = hdr->b_l1hdr.b_acb; 3810 ASSERT(callback_list != NULL); 3811 if (BP_SHOULD_BYTESWAP(zio->io_bp) && zio->io_error == 0) { 3812 dmu_object_byteswap_t bswap = 3813 DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp)); 3814 arc_byteswap_func_t *func = BP_GET_LEVEL(zio->io_bp) > 0 ? 3815 byteswap_uint64_array : 3816 dmu_ot_byteswap[bswap].ob_func; 3817 func(buf->b_data, hdr->b_size); 3818 } 3819 3820 arc_cksum_compute(buf, B_FALSE); 3821#ifdef illumos 3822 arc_buf_watch(buf); 3823#endif /* illumos */ 3824 3825 if (hash_lock && zio->io_error == 0 && 3826 hdr->b_l1hdr.b_state == arc_anon) { 3827 /* 3828 * Only call arc_access on anonymous buffers. This is because 3829 * if we've issued an I/O for an evicted buffer, we've already 3830 * called arc_access (to prevent any simultaneous readers from 3831 * getting confused). 3832 */ 3833 arc_access(hdr, hash_lock); 3834 } 3835 3836 /* create copies of the data buffer for the callers */ 3837 abuf = buf; 3838 for (acb = callback_list; acb; acb = acb->acb_next) { 3839 if (acb->acb_done) { 3840 if (abuf == NULL) { 3841 ARCSTAT_BUMP(arcstat_duplicate_reads); 3842 abuf = arc_buf_clone(buf); 3843 } 3844 acb->acb_buf = abuf; 3845 abuf = NULL; 3846 } 3847 } 3848 hdr->b_l1hdr.b_acb = NULL; 3849 hdr->b_flags &= ~ARC_FLAG_IO_IN_PROGRESS; 3850 ASSERT(!HDR_BUF_AVAILABLE(hdr)); 3851 if (abuf == buf) { 3852 ASSERT(buf->b_efunc == NULL); 3853 ASSERT(hdr->b_l1hdr.b_datacnt == 1); 3854 hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE; 3855 } 3856 3857 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt) || 3858 callback_list != NULL); 3859 3860 if (zio->io_error != 0) { 3861 hdr->b_flags |= ARC_FLAG_IO_ERROR; 3862 if (hdr->b_l1hdr.b_state != arc_anon) 3863 arc_change_state(arc_anon, hdr, hash_lock); 3864 if (HDR_IN_HASH_TABLE(hdr)) 3865 buf_hash_remove(hdr); 3866 freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt); 3867 } 3868 3869 /* 3870 * Broadcast before we drop the hash_lock to avoid the possibility 3871 * that the hdr (and hence the cv) might be freed before we get to 3872 * the cv_broadcast(). 3873 */ 3874 cv_broadcast(&hdr->b_l1hdr.b_cv); 3875 3876 if (hash_lock != NULL) { 3877 mutex_exit(hash_lock); 3878 } else { 3879 /* 3880 * This block was freed while we waited for the read to 3881 * complete. It has been removed from the hash table and 3882 * moved to the anonymous state (so that it won't show up 3883 * in the cache). 3884 */ 3885 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon); 3886 freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt); 3887 } 3888 3889 /* execute each callback and free its structure */ 3890 while ((acb = callback_list) != NULL) { 3891 if (acb->acb_done) 3892 acb->acb_done(zio, acb->acb_buf, acb->acb_private); 3893 3894 if (acb->acb_zio_dummy != NULL) { 3895 acb->acb_zio_dummy->io_error = zio->io_error; 3896 zio_nowait(acb->acb_zio_dummy); 3897 } 3898 3899 callback_list = acb->acb_next; 3900 kmem_free(acb, sizeof (arc_callback_t)); 3901 } 3902 3903 if (freeable) 3904 arc_hdr_destroy(hdr); 3905} 3906 3907/* 3908 * "Read" the block block at the specified DVA (in bp) via the 3909 * cache. If the block is found in the cache, invoke the provided 3910 * callback immediately and return. Note that the `zio' parameter 3911 * in the callback will be NULL in this case, since no IO was 3912 * required. If the block is not in the cache pass the read request 3913 * on to the spa with a substitute callback function, so that the 3914 * requested block will be added to the cache. 3915 * 3916 * If a read request arrives for a block that has a read in-progress, 3917 * either wait for the in-progress read to complete (and return the 3918 * results); or, if this is a read with a "done" func, add a record 3919 * to the read to invoke the "done" func when the read completes, 3920 * and return; or just return. 3921 * 3922 * arc_read_done() will invoke all the requested "done" functions 3923 * for readers of this block. 3924 */ 3925int 3926arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_done_func_t *done, 3927 void *private, zio_priority_t priority, int zio_flags, 3928 arc_flags_t *arc_flags, const zbookmark_phys_t *zb) 3929{ 3930 arc_buf_hdr_t *hdr = NULL; 3931 arc_buf_t *buf = NULL; 3932 kmutex_t *hash_lock = NULL; 3933 zio_t *rzio; 3934 uint64_t guid = spa_load_guid(spa); 3935 3936 ASSERT(!BP_IS_EMBEDDED(bp) || 3937 BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA); 3938 3939top: 3940 if (!BP_IS_EMBEDDED(bp)) { 3941 /* 3942 * Embedded BP's have no DVA and require no I/O to "read". 3943 * Create an anonymous arc buf to back it. 3944 */ 3945 hdr = buf_hash_find(guid, bp, &hash_lock); 3946 } 3947 3948 if (hdr != NULL && HDR_HAS_L1HDR(hdr) && hdr->b_l1hdr.b_datacnt > 0) { 3949 3950 *arc_flags |= ARC_FLAG_CACHED; 3951 3952 if (HDR_IO_IN_PROGRESS(hdr)) { 3953 3954 if (*arc_flags & ARC_FLAG_WAIT) { 3955 cv_wait(&hdr->b_l1hdr.b_cv, hash_lock); 3956 mutex_exit(hash_lock); 3957 goto top; 3958 } 3959 ASSERT(*arc_flags & ARC_FLAG_NOWAIT); 3960 3961 if (done) { 3962 arc_callback_t *acb = NULL; 3963 3964 acb = kmem_zalloc(sizeof (arc_callback_t), 3965 KM_SLEEP); 3966 acb->acb_done = done; 3967 acb->acb_private = private; 3968 if (pio != NULL) 3969 acb->acb_zio_dummy = zio_null(pio, 3970 spa, NULL, NULL, NULL, zio_flags); 3971 3972 ASSERT(acb->acb_done != NULL); 3973 acb->acb_next = hdr->b_l1hdr.b_acb; 3974 hdr->b_l1hdr.b_acb = acb; 3975 add_reference(hdr, hash_lock, private); 3976 mutex_exit(hash_lock); 3977 return (0); 3978 } 3979 mutex_exit(hash_lock); 3980 return (0); 3981 } 3982 3983 ASSERT(hdr->b_l1hdr.b_state == arc_mru || 3984 hdr->b_l1hdr.b_state == arc_mfu); 3985 3986 if (done) { 3987 add_reference(hdr, hash_lock, private); 3988 /* 3989 * If this block is already in use, create a new 3990 * copy of the data so that we will be guaranteed 3991 * that arc_release() will always succeed. 3992 */ 3993 buf = hdr->b_l1hdr.b_buf; 3994 ASSERT(buf); 3995 ASSERT(buf->b_data); 3996 if (HDR_BUF_AVAILABLE(hdr)) { 3997 ASSERT(buf->b_efunc == NULL); 3998 hdr->b_flags &= ~ARC_FLAG_BUF_AVAILABLE; 3999 } else { 4000 buf = arc_buf_clone(buf); 4001 } 4002 4003 } else if (*arc_flags & ARC_FLAG_PREFETCH && 4004 refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) { 4005 hdr->b_flags |= ARC_FLAG_PREFETCH; 4006 } 4007 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr); 4008 arc_access(hdr, hash_lock); 4009 if (*arc_flags & ARC_FLAG_L2CACHE) 4010 hdr->b_flags |= ARC_FLAG_L2CACHE; 4011 if (*arc_flags & ARC_FLAG_L2COMPRESS) 4012 hdr->b_flags |= ARC_FLAG_L2COMPRESS; 4013 mutex_exit(hash_lock); 4014 ARCSTAT_BUMP(arcstat_hits); 4015 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr), 4016 demand, prefetch, !HDR_ISTYPE_METADATA(hdr), 4017 data, metadata, hits); 4018 4019 if (done) 4020 done(NULL, buf, private); 4021 } else { 4022 uint64_t size = BP_GET_LSIZE(bp); 4023 arc_callback_t *acb; 4024 vdev_t *vd = NULL; 4025 uint64_t addr = 0; 4026 boolean_t devw = B_FALSE; 4027 enum zio_compress b_compress = ZIO_COMPRESS_OFF; 4028 int32_t b_asize = 0; 4029 4030 if (hdr == NULL) { 4031 /* this block is not in the cache */ 4032 arc_buf_hdr_t *exists = NULL; 4033 arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp); 4034 buf = arc_buf_alloc(spa, size, private, type); 4035 hdr = buf->b_hdr; 4036 if (!BP_IS_EMBEDDED(bp)) { 4037 hdr->b_dva = *BP_IDENTITY(bp); 4038 hdr->b_birth = BP_PHYSICAL_BIRTH(bp); 4039 exists = buf_hash_insert(hdr, &hash_lock); 4040 } 4041 if (exists != NULL) { 4042 /* somebody beat us to the hash insert */ 4043 mutex_exit(hash_lock); 4044 buf_discard_identity(hdr); 4045 (void) arc_buf_remove_ref(buf, private); 4046 goto top; /* restart the IO request */ 4047 } 4048 4049 /* if this is a prefetch, we don't have a reference */ 4050 if (*arc_flags & ARC_FLAG_PREFETCH) { 4051 (void) remove_reference(hdr, hash_lock, 4052 private); 4053 hdr->b_flags |= ARC_FLAG_PREFETCH; 4054 } 4055 if (*arc_flags & ARC_FLAG_L2CACHE) 4056 hdr->b_flags |= ARC_FLAG_L2CACHE; 4057 if (*arc_flags & ARC_FLAG_L2COMPRESS) 4058 hdr->b_flags |= ARC_FLAG_L2COMPRESS; 4059 if (BP_GET_LEVEL(bp) > 0) 4060 hdr->b_flags |= ARC_FLAG_INDIRECT; 4061 } else { 4062 /* 4063 * This block is in the ghost cache. If it was L2-only 4064 * (and thus didn't have an L1 hdr), we realloc the 4065 * header to add an L1 hdr. 4066 */ 4067 if (!HDR_HAS_L1HDR(hdr)) { 4068 hdr = arc_hdr_realloc(hdr, hdr_l2only_cache, 4069 hdr_full_cache); 4070 } 4071 4072 ASSERT(GHOST_STATE(hdr->b_l1hdr.b_state)); 4073 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 4074 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 4075 ASSERT(hdr->b_l1hdr.b_buf == NULL); 4076 4077 /* if this is a prefetch, we don't have a reference */ 4078 if (*arc_flags & ARC_FLAG_PREFETCH) 4079 hdr->b_flags |= ARC_FLAG_PREFETCH; 4080 else 4081 add_reference(hdr, hash_lock, private); 4082 if (*arc_flags & ARC_FLAG_L2CACHE) 4083 hdr->b_flags |= ARC_FLAG_L2CACHE; 4084 if (*arc_flags & ARC_FLAG_L2COMPRESS) 4085 hdr->b_flags |= ARC_FLAG_L2COMPRESS; 4086 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 4087 buf->b_hdr = hdr; 4088 buf->b_data = NULL; 4089 buf->b_efunc = NULL; 4090 buf->b_private = NULL; 4091 buf->b_next = NULL; 4092 hdr->b_l1hdr.b_buf = buf; 4093 ASSERT0(hdr->b_l1hdr.b_datacnt); 4094 hdr->b_l1hdr.b_datacnt = 1; 4095 arc_get_data_buf(buf); 4096 arc_access(hdr, hash_lock); 4097 } 4098 4099 ASSERT(!GHOST_STATE(hdr->b_l1hdr.b_state)); 4100 4101 acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP); 4102 acb->acb_done = done; 4103 acb->acb_private = private; 4104 4105 ASSERT(hdr->b_l1hdr.b_acb == NULL); 4106 hdr->b_l1hdr.b_acb = acb; 4107 hdr->b_flags |= ARC_FLAG_IO_IN_PROGRESS; 4108 4109 if (HDR_HAS_L2HDR(hdr) && 4110 (vd = hdr->b_l2hdr.b_dev->l2ad_vdev) != NULL) { 4111 devw = hdr->b_l2hdr.b_dev->l2ad_writing; 4112 addr = hdr->b_l2hdr.b_daddr; 4113 b_compress = HDR_GET_COMPRESS(hdr); 4114 b_asize = hdr->b_l2hdr.b_asize; 4115 /* 4116 * Lock out device removal. 4117 */ 4118 if (vdev_is_dead(vd) || 4119 !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER)) 4120 vd = NULL; 4121 } 4122 4123 if (hash_lock != NULL) 4124 mutex_exit(hash_lock); 4125 4126 /* 4127 * At this point, we have a level 1 cache miss. Try again in 4128 * L2ARC if possible. 4129 */ 4130 ASSERT3U(hdr->b_size, ==, size); 4131 DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp, 4132 uint64_t, size, zbookmark_phys_t *, zb); 4133 ARCSTAT_BUMP(arcstat_misses); 4134 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr), 4135 demand, prefetch, !HDR_ISTYPE_METADATA(hdr), 4136 data, metadata, misses); 4137#ifdef _KERNEL 4138 curthread->td_ru.ru_inblock++; 4139#endif 4140 4141 if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) { 4142 /* 4143 * Read from the L2ARC if the following are true: 4144 * 1. The L2ARC vdev was previously cached. 4145 * 2. This buffer still has L2ARC metadata. 4146 * 3. This buffer isn't currently writing to the L2ARC. 4147 * 4. The L2ARC entry wasn't evicted, which may 4148 * also have invalidated the vdev. 4149 * 5. This isn't prefetch and l2arc_noprefetch is set. 4150 */ 4151 if (HDR_HAS_L2HDR(hdr) && 4152 !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) && 4153 !(l2arc_noprefetch && HDR_PREFETCH(hdr))) { 4154 l2arc_read_callback_t *cb; 4155 4156 DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr); 4157 ARCSTAT_BUMP(arcstat_l2_hits); 4158 4159 cb = kmem_zalloc(sizeof (l2arc_read_callback_t), 4160 KM_SLEEP); 4161 cb->l2rcb_buf = buf; 4162 cb->l2rcb_spa = spa; 4163 cb->l2rcb_bp = *bp; 4164 cb->l2rcb_zb = *zb; 4165 cb->l2rcb_flags = zio_flags; 4166 cb->l2rcb_compress = b_compress; 4167 4168 ASSERT(addr >= VDEV_LABEL_START_SIZE && 4169 addr + size < vd->vdev_psize - 4170 VDEV_LABEL_END_SIZE); 4171 4172 /* 4173 * l2arc read. The SCL_L2ARC lock will be 4174 * released by l2arc_read_done(). 4175 * Issue a null zio if the underlying buffer 4176 * was squashed to zero size by compression. 4177 */ 4178 if (b_compress == ZIO_COMPRESS_EMPTY) { 4179 rzio = zio_null(pio, spa, vd, 4180 l2arc_read_done, cb, 4181 zio_flags | ZIO_FLAG_DONT_CACHE | 4182 ZIO_FLAG_CANFAIL | 4183 ZIO_FLAG_DONT_PROPAGATE | 4184 ZIO_FLAG_DONT_RETRY); 4185 } else { 4186 rzio = zio_read_phys(pio, vd, addr, 4187 b_asize, buf->b_data, 4188 ZIO_CHECKSUM_OFF, 4189 l2arc_read_done, cb, priority, 4190 zio_flags | ZIO_FLAG_DONT_CACHE | 4191 ZIO_FLAG_CANFAIL | 4192 ZIO_FLAG_DONT_PROPAGATE | 4193 ZIO_FLAG_DONT_RETRY, B_FALSE); 4194 } 4195 DTRACE_PROBE2(l2arc__read, vdev_t *, vd, 4196 zio_t *, rzio); 4197 ARCSTAT_INCR(arcstat_l2_read_bytes, b_asize); 4198 4199 if (*arc_flags & ARC_FLAG_NOWAIT) { 4200 zio_nowait(rzio); 4201 return (0); 4202 } 4203 4204 ASSERT(*arc_flags & ARC_FLAG_WAIT); 4205 if (zio_wait(rzio) == 0) 4206 return (0); 4207 4208 /* l2arc read error; goto zio_read() */ 4209 } else { 4210 DTRACE_PROBE1(l2arc__miss, 4211 arc_buf_hdr_t *, hdr); 4212 ARCSTAT_BUMP(arcstat_l2_misses); 4213 if (HDR_L2_WRITING(hdr)) 4214 ARCSTAT_BUMP(arcstat_l2_rw_clash); 4215 spa_config_exit(spa, SCL_L2ARC, vd); 4216 } 4217 } else { 4218 if (vd != NULL) 4219 spa_config_exit(spa, SCL_L2ARC, vd); 4220 if (l2arc_ndev != 0) { 4221 DTRACE_PROBE1(l2arc__miss, 4222 arc_buf_hdr_t *, hdr); 4223 ARCSTAT_BUMP(arcstat_l2_misses); 4224 } 4225 } 4226 4227 rzio = zio_read(pio, spa, bp, buf->b_data, size, 4228 arc_read_done, buf, priority, zio_flags, zb); 4229 4230 if (*arc_flags & ARC_FLAG_WAIT) 4231 return (zio_wait(rzio)); 4232 4233 ASSERT(*arc_flags & ARC_FLAG_NOWAIT); 4234 zio_nowait(rzio); 4235 } 4236 return (0); 4237} 4238 4239void 4240arc_set_callback(arc_buf_t *buf, arc_evict_func_t *func, void *private) 4241{ 4242 ASSERT(buf->b_hdr != NULL); 4243 ASSERT(buf->b_hdr->b_l1hdr.b_state != arc_anon); 4244 ASSERT(!refcount_is_zero(&buf->b_hdr->b_l1hdr.b_refcnt) || 4245 func == NULL); 4246 ASSERT(buf->b_efunc == NULL); 4247 ASSERT(!HDR_BUF_AVAILABLE(buf->b_hdr)); 4248 4249 buf->b_efunc = func; 4250 buf->b_private = private; 4251} 4252 4253/* 4254 * Notify the arc that a block was freed, and thus will never be used again. 4255 */ 4256void 4257arc_freed(spa_t *spa, const blkptr_t *bp) 4258{ 4259 arc_buf_hdr_t *hdr; 4260 kmutex_t *hash_lock; 4261 uint64_t guid = spa_load_guid(spa); 4262 4263 ASSERT(!BP_IS_EMBEDDED(bp)); 4264 4265 hdr = buf_hash_find(guid, bp, &hash_lock); 4266 if (hdr == NULL) 4267 return; 4268 if (HDR_BUF_AVAILABLE(hdr)) { 4269 arc_buf_t *buf = hdr->b_l1hdr.b_buf; 4270 add_reference(hdr, hash_lock, FTAG); 4271 hdr->b_flags &= ~ARC_FLAG_BUF_AVAILABLE; 4272 mutex_exit(hash_lock); 4273 4274 arc_release(buf, FTAG); 4275 (void) arc_buf_remove_ref(buf, FTAG); 4276 } else { 4277 mutex_exit(hash_lock); 4278 } 4279 4280} 4281 4282/* 4283 * Clear the user eviction callback set by arc_set_callback(), first calling 4284 * it if it exists. Because the presence of a callback keeps an arc_buf cached 4285 * clearing the callback may result in the arc_buf being destroyed. However, 4286 * it will not result in the *last* arc_buf being destroyed, hence the data 4287 * will remain cached in the ARC. We make a copy of the arc buffer here so 4288 * that we can process the callback without holding any locks. 4289 * 4290 * It's possible that the callback is already in the process of being cleared 4291 * by another thread. In this case we can not clear the callback. 4292 * 4293 * Returns B_TRUE if the callback was successfully called and cleared. 4294 */ 4295boolean_t 4296arc_clear_callback(arc_buf_t *buf) 4297{ 4298 arc_buf_hdr_t *hdr; 4299 kmutex_t *hash_lock; 4300 arc_evict_func_t *efunc = buf->b_efunc; 4301 void *private = buf->b_private; 4302 list_t *list, *evicted_list; 4303 kmutex_t *lock, *evicted_lock; 4304 4305 mutex_enter(&buf->b_evict_lock); 4306 hdr = buf->b_hdr; 4307 if (hdr == NULL) { 4308 /* 4309 * We are in arc_do_user_evicts(). 4310 */ 4311 ASSERT(buf->b_data == NULL); 4312 mutex_exit(&buf->b_evict_lock); 4313 return (B_FALSE); 4314 } else if (buf->b_data == NULL) { 4315 /* 4316 * We are on the eviction list; process this buffer now 4317 * but let arc_do_user_evicts() do the reaping. 4318 */ 4319 buf->b_efunc = NULL; 4320 mutex_exit(&buf->b_evict_lock); 4321 VERIFY0(efunc(private)); 4322 return (B_TRUE); 4323 } 4324 hash_lock = HDR_LOCK(hdr); 4325 mutex_enter(hash_lock); 4326 hdr = buf->b_hdr; 4327 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 4328 4329 ASSERT3U(refcount_count(&hdr->b_l1hdr.b_refcnt), <, 4330 hdr->b_l1hdr.b_datacnt); 4331 ASSERT(hdr->b_l1hdr.b_state == arc_mru || 4332 hdr->b_l1hdr.b_state == arc_mfu); 4333 4334 buf->b_efunc = NULL; 4335 buf->b_private = NULL; 4336 4337 if (hdr->b_l1hdr.b_datacnt > 1) { 4338 mutex_exit(&buf->b_evict_lock); 4339 arc_buf_destroy(buf, FALSE, TRUE); 4340 } else { 4341 ASSERT(buf == hdr->b_l1hdr.b_buf); 4342 hdr->b_flags |= ARC_FLAG_BUF_AVAILABLE; 4343 mutex_exit(&buf->b_evict_lock); 4344 } 4345 4346 mutex_exit(hash_lock); 4347 VERIFY0(efunc(private)); 4348 return (B_TRUE); 4349} 4350 4351/* 4352 * Release this buffer from the cache, making it an anonymous buffer. This 4353 * must be done after a read and prior to modifying the buffer contents. 4354 * If the buffer has more than one reference, we must make 4355 * a new hdr for the buffer. 4356 */ 4357void 4358arc_release(arc_buf_t *buf, void *tag) 4359{ 4360 arc_buf_hdr_t *hdr = buf->b_hdr; 4361 4362 /* 4363 * It would be nice to assert that if it's DMU metadata (level > 4364 * 0 || it's the dnode file), then it must be syncing context. 4365 * But we don't know that information at this level. 4366 */ 4367 4368 mutex_enter(&buf->b_evict_lock); 4369 /* 4370 * We don't grab the hash lock prior to this check, because if 4371 * the buffer's header is in the arc_anon state, it won't be 4372 * linked into the hash table. 4373 */ 4374 if (hdr->b_l1hdr.b_state == arc_anon) { 4375 mutex_exit(&buf->b_evict_lock); 4376 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 4377 ASSERT(!HDR_IN_HASH_TABLE(hdr)); 4378 ASSERT(!HDR_HAS_L2HDR(hdr)); 4379 ASSERT(BUF_EMPTY(hdr)); 4380 ASSERT3U(hdr->b_l1hdr.b_datacnt, ==, 1); 4381 ASSERT3S(refcount_count(&hdr->b_l1hdr.b_refcnt), ==, 1); 4382 ASSERT(!list_link_active(&hdr->b_l1hdr.b_arc_node)); 4383 4384 ASSERT3P(buf->b_efunc, ==, NULL); 4385 ASSERT3P(buf->b_private, ==, NULL); 4386 4387 hdr->b_l1hdr.b_arc_access = 0; 4388 arc_buf_thaw(buf); 4389 4390 return; 4391 } 4392 4393 kmutex_t *hash_lock = HDR_LOCK(hdr); 4394 mutex_enter(hash_lock); 4395 4396 /* 4397 * This assignment is only valid as long as the hash_lock is 4398 * held, we must be careful not to reference state or the 4399 * b_state field after dropping the lock. 4400 */ 4401 arc_state_t *state = hdr->b_l1hdr.b_state; 4402 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 4403 ASSERT3P(state, !=, arc_anon); 4404 4405 /* this buffer is not on any list */ 4406 ASSERT(refcount_count(&hdr->b_l1hdr.b_refcnt) > 0); 4407 4408 if (HDR_HAS_L2HDR(hdr)) { 4409 mutex_enter(&hdr->b_l2hdr.b_dev->l2ad_mtx); 4410 4411 /* 4412 * We have to recheck this conditional again now that 4413 * we're holding the l2ad_mtx to prevent a race with 4414 * another thread which might be concurrently calling 4415 * l2arc_evict(). In that case, l2arc_evict() might have 4416 * destroyed the header's L2 portion as we were waiting 4417 * to acquire the l2ad_mtx. 4418 */ 4419 if (HDR_HAS_L2HDR(hdr)) { 4420 if (hdr->b_l2hdr.b_daddr != L2ARC_ADDR_UNSET) 4421 trim_map_free(hdr->b_l2hdr.b_dev->l2ad_vdev, 4422 hdr->b_l2hdr.b_daddr, 4423 hdr->b_l2hdr.b_asize, 0); 4424 arc_hdr_l2hdr_destroy(hdr); 4425 } 4426 4427 mutex_exit(&hdr->b_l2hdr.b_dev->l2ad_mtx); 4428 } 4429 4430 /* 4431 * Do we have more than one buf? 4432 */ 4433 if (hdr->b_l1hdr.b_datacnt > 1) { 4434 arc_buf_hdr_t *nhdr; 4435 arc_buf_t **bufp; 4436 uint64_t blksz = hdr->b_size; 4437 uint64_t spa = hdr->b_spa; 4438 arc_buf_contents_t type = arc_buf_type(hdr); 4439 uint32_t flags = hdr->b_flags; 4440 4441 ASSERT(hdr->b_l1hdr.b_buf != buf || buf->b_next != NULL); 4442 /* 4443 * Pull the data off of this hdr and attach it to 4444 * a new anonymous hdr. 4445 */ 4446 (void) remove_reference(hdr, hash_lock, tag); 4447 bufp = &hdr->b_l1hdr.b_buf; 4448 while (*bufp != buf) 4449 bufp = &(*bufp)->b_next; 4450 *bufp = buf->b_next; 4451 buf->b_next = NULL; 4452 4453 ASSERT3P(state, !=, arc_l2c_only); 4454 ASSERT3U(state->arcs_size, >=, hdr->b_size); 4455 atomic_add_64(&state->arcs_size, -hdr->b_size); 4456 if (refcount_is_zero(&hdr->b_l1hdr.b_refcnt)) { 4457 ASSERT3P(state, !=, arc_l2c_only); 4458 uint64_t *size = &state->arcs_lsize[type]; 4459 ASSERT3U(*size, >=, hdr->b_size); 4460 atomic_add_64(size, -hdr->b_size); 4461 } 4462 4463 /* 4464 * We're releasing a duplicate user data buffer, update 4465 * our statistics accordingly. 4466 */ 4467 if (HDR_ISTYPE_DATA(hdr)) { 4468 ARCSTAT_BUMPDOWN(arcstat_duplicate_buffers); 4469 ARCSTAT_INCR(arcstat_duplicate_buffers_size, 4470 -hdr->b_size); 4471 } 4472 hdr->b_l1hdr.b_datacnt -= 1; 4473 arc_cksum_verify(buf); 4474#ifdef illumos 4475 arc_buf_unwatch(buf); 4476#endif /* illumos */ 4477 4478 mutex_exit(hash_lock); 4479 4480 nhdr = kmem_cache_alloc(hdr_full_cache, KM_PUSHPAGE); 4481 nhdr->b_size = blksz; 4482 nhdr->b_spa = spa; 4483 4484 nhdr->b_flags = flags & ARC_FLAG_L2_WRITING; 4485 nhdr->b_flags |= arc_bufc_to_flags(type); 4486 nhdr->b_flags |= ARC_FLAG_HAS_L1HDR; 4487 4488 nhdr->b_l1hdr.b_buf = buf; 4489 nhdr->b_l1hdr.b_datacnt = 1; 4490 nhdr->b_l1hdr.b_state = arc_anon; 4491 nhdr->b_l1hdr.b_arc_access = 0; 4492 nhdr->b_freeze_cksum = NULL; 4493 4494 (void) refcount_add(&nhdr->b_l1hdr.b_refcnt, tag); 4495 buf->b_hdr = nhdr; 4496 mutex_exit(&buf->b_evict_lock); 4497 atomic_add_64(&arc_anon->arcs_size, blksz); 4498 } else { 4499 mutex_exit(&buf->b_evict_lock); 4500 ASSERT(refcount_count(&hdr->b_l1hdr.b_refcnt) == 1); 4501 /* protected by hash lock */ 4502 ASSERT(!list_link_active(&hdr->b_l1hdr.b_arc_node)); 4503 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 4504 arc_change_state(arc_anon, hdr, hash_lock); 4505 hdr->b_l1hdr.b_arc_access = 0; 4506 mutex_exit(hash_lock); 4507 4508 buf_discard_identity(hdr); 4509 arc_buf_thaw(buf); 4510 } 4511 buf->b_efunc = NULL; 4512 buf->b_private = NULL; 4513} 4514 4515int 4516arc_released(arc_buf_t *buf) 4517{ 4518 int released; 4519 4520 mutex_enter(&buf->b_evict_lock); 4521 released = (buf->b_data != NULL && 4522 buf->b_hdr->b_l1hdr.b_state == arc_anon); 4523 mutex_exit(&buf->b_evict_lock); 4524 return (released); 4525} 4526 4527#ifdef ZFS_DEBUG 4528int 4529arc_referenced(arc_buf_t *buf) 4530{ 4531 int referenced; 4532 4533 mutex_enter(&buf->b_evict_lock); 4534 referenced = (refcount_count(&buf->b_hdr->b_l1hdr.b_refcnt)); 4535 mutex_exit(&buf->b_evict_lock); 4536 return (referenced); 4537} 4538#endif 4539 4540static void 4541arc_write_ready(zio_t *zio) 4542{ 4543 arc_write_callback_t *callback = zio->io_private; 4544 arc_buf_t *buf = callback->awcb_buf; 4545 arc_buf_hdr_t *hdr = buf->b_hdr; 4546 4547 ASSERT(HDR_HAS_L1HDR(hdr)); 4548 ASSERT(!refcount_is_zero(&buf->b_hdr->b_l1hdr.b_refcnt)); 4549 ASSERT(hdr->b_l1hdr.b_datacnt > 0); 4550 callback->awcb_ready(zio, buf, callback->awcb_private); 4551 4552 /* 4553 * If the IO is already in progress, then this is a re-write 4554 * attempt, so we need to thaw and re-compute the cksum. 4555 * It is the responsibility of the callback to handle the 4556 * accounting for any re-write attempt. 4557 */ 4558 if (HDR_IO_IN_PROGRESS(hdr)) { 4559 mutex_enter(&hdr->b_l1hdr.b_freeze_lock); 4560 if (hdr->b_freeze_cksum != NULL) { 4561 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t)); 4562 hdr->b_freeze_cksum = NULL; 4563 } 4564 mutex_exit(&hdr->b_l1hdr.b_freeze_lock); 4565 } 4566 arc_cksum_compute(buf, B_FALSE); 4567 hdr->b_flags |= ARC_FLAG_IO_IN_PROGRESS; 4568} 4569 4570/* 4571 * The SPA calls this callback for each physical write that happens on behalf 4572 * of a logical write. See the comment in dbuf_write_physdone() for details. 4573 */ 4574static void 4575arc_write_physdone(zio_t *zio) 4576{ 4577 arc_write_callback_t *cb = zio->io_private; 4578 if (cb->awcb_physdone != NULL) 4579 cb->awcb_physdone(zio, cb->awcb_buf, cb->awcb_private); 4580} 4581 4582static void 4583arc_write_done(zio_t *zio) 4584{ 4585 arc_write_callback_t *callback = zio->io_private; 4586 arc_buf_t *buf = callback->awcb_buf; 4587 arc_buf_hdr_t *hdr = buf->b_hdr; 4588 4589 ASSERT(hdr->b_l1hdr.b_acb == NULL); 4590 4591 if (zio->io_error == 0) { 4592 if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) { 4593 buf_discard_identity(hdr); 4594 } else { 4595 hdr->b_dva = *BP_IDENTITY(zio->io_bp); 4596 hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp); 4597 } 4598 } else { 4599 ASSERT(BUF_EMPTY(hdr)); 4600 } 4601 4602 /* 4603 * If the block to be written was all-zero or compressed enough to be 4604 * embedded in the BP, no write was performed so there will be no 4605 * dva/birth/checksum. The buffer must therefore remain anonymous 4606 * (and uncached). 4607 */ 4608 if (!BUF_EMPTY(hdr)) { 4609 arc_buf_hdr_t *exists; 4610 kmutex_t *hash_lock; 4611 4612 ASSERT(zio->io_error == 0); 4613 4614 arc_cksum_verify(buf); 4615 4616 exists = buf_hash_insert(hdr, &hash_lock); 4617 if (exists != NULL) { 4618 /* 4619 * This can only happen if we overwrite for 4620 * sync-to-convergence, because we remove 4621 * buffers from the hash table when we arc_free(). 4622 */ 4623 if (zio->io_flags & ZIO_FLAG_IO_REWRITE) { 4624 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) 4625 panic("bad overwrite, hdr=%p exists=%p", 4626 (void *)hdr, (void *)exists); 4627 ASSERT(refcount_is_zero( 4628 &exists->b_l1hdr.b_refcnt)); 4629 arc_change_state(arc_anon, exists, hash_lock); 4630 mutex_exit(hash_lock); 4631 arc_hdr_destroy(exists); 4632 exists = buf_hash_insert(hdr, &hash_lock); 4633 ASSERT3P(exists, ==, NULL); 4634 } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) { 4635 /* nopwrite */ 4636 ASSERT(zio->io_prop.zp_nopwrite); 4637 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) 4638 panic("bad nopwrite, hdr=%p exists=%p", 4639 (void *)hdr, (void *)exists); 4640 } else { 4641 /* Dedup */ 4642 ASSERT(hdr->b_l1hdr.b_datacnt == 1); 4643 ASSERT(hdr->b_l1hdr.b_state == arc_anon); 4644 ASSERT(BP_GET_DEDUP(zio->io_bp)); 4645 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0); 4646 } 4647 } 4648 hdr->b_flags &= ~ARC_FLAG_IO_IN_PROGRESS; 4649 /* if it's not anon, we are doing a scrub */ 4650 if (exists == NULL && hdr->b_l1hdr.b_state == arc_anon) 4651 arc_access(hdr, hash_lock); 4652 mutex_exit(hash_lock); 4653 } else { 4654 hdr->b_flags &= ~ARC_FLAG_IO_IN_PROGRESS; 4655 } 4656 4657 ASSERT(!refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 4658 callback->awcb_done(zio, buf, callback->awcb_private); 4659 4660 kmem_free(callback, sizeof (arc_write_callback_t)); 4661} 4662 4663zio_t * 4664arc_write(zio_t *pio, spa_t *spa, uint64_t txg, 4665 blkptr_t *bp, arc_buf_t *buf, boolean_t l2arc, boolean_t l2arc_compress, 4666 const zio_prop_t *zp, arc_done_func_t *ready, arc_done_func_t *physdone, 4667 arc_done_func_t *done, void *private, zio_priority_t priority, 4668 int zio_flags, const zbookmark_phys_t *zb) 4669{ 4670 arc_buf_hdr_t *hdr = buf->b_hdr; 4671 arc_write_callback_t *callback; 4672 zio_t *zio; 4673 4674 ASSERT(ready != NULL); 4675 ASSERT(done != NULL); 4676 ASSERT(!HDR_IO_ERROR(hdr)); 4677 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 4678 ASSERT(hdr->b_l1hdr.b_acb == NULL); 4679 ASSERT(hdr->b_l1hdr.b_datacnt > 0); 4680 if (l2arc) 4681 hdr->b_flags |= ARC_FLAG_L2CACHE; 4682 if (l2arc_compress) 4683 hdr->b_flags |= ARC_FLAG_L2COMPRESS; 4684 callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP); 4685 callback->awcb_ready = ready; 4686 callback->awcb_physdone = physdone; 4687 callback->awcb_done = done; 4688 callback->awcb_private = private; 4689 callback->awcb_buf = buf; 4690 4691 zio = zio_write(pio, spa, txg, bp, buf->b_data, hdr->b_size, zp, 4692 arc_write_ready, arc_write_physdone, arc_write_done, callback, 4693 priority, zio_flags, zb); 4694 4695 return (zio); 4696} 4697 4698static int 4699arc_memory_throttle(uint64_t reserve, uint64_t txg) 4700{ 4701#ifdef _KERNEL 4702 uint64_t available_memory = ptob(freemem); 4703 static uint64_t page_load = 0; 4704 static uint64_t last_txg = 0; 4705 4706#if defined(__i386) || !defined(UMA_MD_SMALL_ALLOC) 4707 available_memory = 4708 MIN(available_memory, ptob(vmem_size(heap_arena, VMEM_FREE))); 4709#endif 4710 4711 if (freemem > (uint64_t)physmem * arc_lotsfree_percent / 100) 4712 return (0); 4713 4714 if (txg > last_txg) { 4715 last_txg = txg; 4716 page_load = 0; 4717 } 4718 /* 4719 * If we are in pageout, we know that memory is already tight, 4720 * the arc is already going to be evicting, so we just want to 4721 * continue to let page writes occur as quickly as possible. 4722 */ 4723 if (curproc == pageproc) { 4724 if (page_load > MAX(ptob(minfree), available_memory) / 4) 4725 return (SET_ERROR(ERESTART)); 4726 /* Note: reserve is inflated, so we deflate */ 4727 page_load += reserve / 8; 4728 return (0); 4729 } else if (page_load > 0 && arc_reclaim_needed()) { 4730 /* memory is low, delay before restarting */ 4731 ARCSTAT_INCR(arcstat_memory_throttle_count, 1); 4732 return (SET_ERROR(EAGAIN)); 4733 } 4734 page_load = 0; 4735#endif 4736 return (0); 4737} 4738 4739void 4740arc_tempreserve_clear(uint64_t reserve) 4741{ 4742 atomic_add_64(&arc_tempreserve, -reserve); 4743 ASSERT((int64_t)arc_tempreserve >= 0); 4744} 4745 4746int 4747arc_tempreserve_space(uint64_t reserve, uint64_t txg) 4748{ 4749 int error; 4750 uint64_t anon_size; 4751 4752 if (reserve > arc_c/4 && !arc_no_grow) { 4753 arc_c = MIN(arc_c_max, reserve * 4); 4754 DTRACE_PROBE1(arc__set_reserve, uint64_t, arc_c); 4755 } 4756 if (reserve > arc_c) 4757 return (SET_ERROR(ENOMEM)); 4758 4759 /* 4760 * Don't count loaned bufs as in flight dirty data to prevent long 4761 * network delays from blocking transactions that are ready to be 4762 * assigned to a txg. 4763 */ 4764 anon_size = MAX((int64_t)(arc_anon->arcs_size - arc_loaned_bytes), 0); 4765 4766 /* 4767 * Writes will, almost always, require additional memory allocations 4768 * in order to compress/encrypt/etc the data. We therefore need to 4769 * make sure that there is sufficient available memory for this. 4770 */ 4771 error = arc_memory_throttle(reserve, txg); 4772 if (error != 0) 4773 return (error); 4774 4775 /* 4776 * Throttle writes when the amount of dirty data in the cache 4777 * gets too large. We try to keep the cache less than half full 4778 * of dirty blocks so that our sync times don't grow too large. 4779 * Note: if two requests come in concurrently, we might let them 4780 * both succeed, when one of them should fail. Not a huge deal. 4781 */ 4782 4783 if (reserve + arc_tempreserve + anon_size > arc_c / 2 && 4784 anon_size > arc_c / 4) { 4785 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK " 4786 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n", 4787 arc_tempreserve>>10, 4788 arc_anon->arcs_lsize[ARC_BUFC_METADATA]>>10, 4789 arc_anon->arcs_lsize[ARC_BUFC_DATA]>>10, 4790 reserve>>10, arc_c>>10); 4791 return (SET_ERROR(ERESTART)); 4792 } 4793 atomic_add_64(&arc_tempreserve, reserve); 4794 return (0); 4795} 4796 4797static void 4798arc_kstat_update_state(arc_state_t *state, kstat_named_t *size, 4799 kstat_named_t *evict_data, kstat_named_t *evict_metadata) 4800{ 4801 size->value.ui64 = state->arcs_size; 4802 evict_data->value.ui64 = state->arcs_lsize[ARC_BUFC_DATA]; 4803 evict_metadata->value.ui64 = state->arcs_lsize[ARC_BUFC_METADATA]; 4804} 4805 4806static int 4807arc_kstat_update(kstat_t *ksp, int rw) 4808{ 4809 arc_stats_t *as = ksp->ks_data; 4810 4811 if (rw == KSTAT_WRITE) { 4812 return (EACCES); 4813 } else { 4814 arc_kstat_update_state(arc_anon, 4815 &as->arcstat_anon_size, 4816 &as->arcstat_anon_evictable_data, 4817 &as->arcstat_anon_evictable_metadata); 4818 arc_kstat_update_state(arc_mru, 4819 &as->arcstat_mru_size, 4820 &as->arcstat_mru_evictable_data, 4821 &as->arcstat_mru_evictable_metadata); 4822 arc_kstat_update_state(arc_mru_ghost, 4823 &as->arcstat_mru_ghost_size, 4824 &as->arcstat_mru_ghost_evictable_data, 4825 &as->arcstat_mru_ghost_evictable_metadata); 4826 arc_kstat_update_state(arc_mfu, 4827 &as->arcstat_mfu_size, 4828 &as->arcstat_mfu_evictable_data, 4829 &as->arcstat_mfu_evictable_metadata); 4830 arc_kstat_update_state(arc_mfu_ghost, 4831 &as->arcstat_mfu_ghost_size, 4832 &as->arcstat_mfu_ghost_evictable_data, 4833 &as->arcstat_mfu_ghost_evictable_metadata); 4834 } 4835 4836 return (0); 4837} 4838 4839#ifdef _KERNEL 4840static eventhandler_tag arc_event_lowmem = NULL; 4841 4842static void 4843arc_lowmem(void *arg __unused, int howto __unused) 4844{ 4845 4846 mutex_enter(&arc_reclaim_thr_lock); 4847 /* XXX: Memory deficit should be passed as argument. */ 4848 needfree = btoc(arc_c >> arc_shrink_shift); 4849 DTRACE_PROBE(arc__needfree); 4850 cv_signal(&arc_reclaim_thr_cv); 4851 4852 /* 4853 * It is unsafe to block here in arbitrary threads, because we can come 4854 * here from ARC itself and may hold ARC locks and thus risk a deadlock 4855 * with ARC reclaim thread. 4856 */ 4857 if (curproc == pageproc) 4858 msleep(&needfree, &arc_reclaim_thr_lock, 0, "zfs:lowmem", 0); 4859 mutex_exit(&arc_reclaim_thr_lock); 4860} 4861#endif 4862 4863void 4864arc_init(void) 4865{ 4866 int i, prefetch_tunable_set = 0; 4867 4868 mutex_init(&arc_reclaim_thr_lock, NULL, MUTEX_DEFAULT, NULL); 4869 cv_init(&arc_reclaim_thr_cv, NULL, CV_DEFAULT, NULL); 4870 4871 /* Convert seconds to clock ticks */ 4872 arc_min_prefetch_lifespan = 1 * hz; 4873 4874 /* Start out with 1/8 of all memory */ 4875 arc_c = kmem_size() / 8; 4876 4877#ifdef sun 4878#ifdef _KERNEL 4879 /* 4880 * On architectures where the physical memory can be larger 4881 * than the addressable space (intel in 32-bit mode), we may 4882 * need to limit the cache to 1/8 of VM size. 4883 */ 4884 arc_c = MIN(arc_c, vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 8); 4885#endif 4886#endif /* sun */ 4887 /* set min cache to 1/32 of all memory, or 16MB, whichever is more */ 4888 arc_c_min = MAX(arc_c / 4, 16 << 20); 4889 /* set max to 1/2 of all memory, or all but 1GB, whichever is more */ 4890 if (arc_c * 8 >= 1 << 30) 4891 arc_c_max = (arc_c * 8) - (1 << 30); 4892 else 4893 arc_c_max = arc_c_min; 4894 arc_c_max = MAX(arc_c * 5, arc_c_max); 4895 4896#ifdef _KERNEL 4897 /* 4898 * Allow the tunables to override our calculations if they are 4899 * reasonable (ie. over 16MB) 4900 */ 4901 if (zfs_arc_max > 16 << 20 && zfs_arc_max < kmem_size()) 4902 arc_c_max = zfs_arc_max; 4903 if (zfs_arc_min > 16 << 20 && zfs_arc_min <= arc_c_max) 4904 arc_c_min = zfs_arc_min; 4905#endif 4906 4907 arc_c = arc_c_max; 4908 arc_p = (arc_c >> 1); 4909 4910 /* limit meta-data to 1/4 of the arc capacity */ 4911 arc_meta_limit = arc_c_max / 4; 4912 4913 /* Allow the tunable to override if it is reasonable */ 4914 if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max) 4915 arc_meta_limit = zfs_arc_meta_limit; 4916 4917 if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0) 4918 arc_c_min = arc_meta_limit / 2; 4919 4920 if (zfs_arc_meta_min > 0) { 4921 arc_meta_min = zfs_arc_meta_min; 4922 } else { 4923 arc_meta_min = arc_c_min / 2; 4924 } 4925 4926 if (zfs_arc_grow_retry > 0) 4927 arc_grow_retry = zfs_arc_grow_retry; 4928 4929 if (zfs_arc_shrink_shift > 0) 4930 arc_shrink_shift = zfs_arc_shrink_shift; 4931 4932 /* 4933 * Ensure that arc_no_grow_shift is less than arc_shrink_shift. 4934 */ 4935 if (arc_no_grow_shift >= arc_shrink_shift) 4936 arc_no_grow_shift = arc_shrink_shift - 1; 4937 4938 if (zfs_arc_p_min_shift > 0) 4939 arc_p_min_shift = zfs_arc_p_min_shift; 4940 4941 /* if kmem_flags are set, lets try to use less memory */ 4942 if (kmem_debugging()) 4943 arc_c = arc_c / 2; 4944 if (arc_c < arc_c_min) 4945 arc_c = arc_c_min; 4946 4947 zfs_arc_min = arc_c_min; 4948 zfs_arc_max = arc_c_max; 4949 4950 arc_anon = &ARC_anon; 4951 arc_mru = &ARC_mru; 4952 arc_mru_ghost = &ARC_mru_ghost; 4953 arc_mfu = &ARC_mfu; 4954 arc_mfu_ghost = &ARC_mfu_ghost; 4955 arc_l2c_only = &ARC_l2c_only; 4956 arc_size = 0; 4957 4958 for (i = 0; i < ARC_BUFC_NUMLISTS; i++) { 4959 mutex_init(&arc_anon->arcs_locks[i].arcs_lock, 4960 NULL, MUTEX_DEFAULT, NULL); 4961 mutex_init(&arc_mru->arcs_locks[i].arcs_lock, 4962 NULL, MUTEX_DEFAULT, NULL); 4963 mutex_init(&arc_mru_ghost->arcs_locks[i].arcs_lock, 4964 NULL, MUTEX_DEFAULT, NULL); 4965 mutex_init(&arc_mfu->arcs_locks[i].arcs_lock, 4966 NULL, MUTEX_DEFAULT, NULL); 4967 mutex_init(&arc_mfu_ghost->arcs_locks[i].arcs_lock, 4968 NULL, MUTEX_DEFAULT, NULL); 4969 mutex_init(&arc_l2c_only->arcs_locks[i].arcs_lock, 4970 NULL, MUTEX_DEFAULT, NULL); 4971 4972 list_create(&arc_mru->arcs_lists[i], 4973 sizeof (arc_buf_hdr_t), 4974 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node)); 4975 list_create(&arc_mru_ghost->arcs_lists[i], 4976 sizeof (arc_buf_hdr_t), 4977 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node)); 4978 list_create(&arc_mfu->arcs_lists[i], 4979 sizeof (arc_buf_hdr_t), 4980 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node)); 4981 list_create(&arc_mfu_ghost->arcs_lists[i], 4982 sizeof (arc_buf_hdr_t), 4983 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node)); 4984 list_create(&arc_mfu_ghost->arcs_lists[i], 4985 sizeof (arc_buf_hdr_t), 4986 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node)); 4987 list_create(&arc_l2c_only->arcs_lists[i], 4988 sizeof (arc_buf_hdr_t), 4989 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node)); 4990 } 4991 4992 buf_init(); 4993 4994 arc_thread_exit = 0; 4995 arc_eviction_list = NULL; 4996 mutex_init(&arc_eviction_mtx, NULL, MUTEX_DEFAULT, NULL); 4997 bzero(&arc_eviction_hdr, sizeof (arc_buf_hdr_t)); 4998 4999 arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED, 5000 sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL); 5001 5002 if (arc_ksp != NULL) { 5003 arc_ksp->ks_data = &arc_stats; 5004 arc_ksp->ks_update = arc_kstat_update; 5005 kstat_install(arc_ksp); 5006 } 5007 5008 (void) thread_create(NULL, 0, arc_reclaim_thread, NULL, 0, &p0, 5009 TS_RUN, minclsyspri); 5010 5011#ifdef _KERNEL 5012 arc_event_lowmem = EVENTHANDLER_REGISTER(vm_lowmem, arc_lowmem, NULL, 5013 EVENTHANDLER_PRI_FIRST); 5014#endif 5015 5016 arc_dead = FALSE; 5017 arc_warm = B_FALSE; 5018 5019 /* 5020 * Calculate maximum amount of dirty data per pool. 5021 * 5022 * If it has been set by /etc/system, take that. 5023 * Otherwise, use a percentage of physical memory defined by 5024 * zfs_dirty_data_max_percent (default 10%) with a cap at 5025 * zfs_dirty_data_max_max (default 4GB). 5026 */ 5027 if (zfs_dirty_data_max == 0) { 5028 zfs_dirty_data_max = ptob(physmem) * 5029 zfs_dirty_data_max_percent / 100; 5030 zfs_dirty_data_max = MIN(zfs_dirty_data_max, 5031 zfs_dirty_data_max_max); 5032 } 5033 5034#ifdef _KERNEL 5035 if (TUNABLE_INT_FETCH("vfs.zfs.prefetch_disable", &zfs_prefetch_disable)) 5036 prefetch_tunable_set = 1; 5037 5038#ifdef __i386__ 5039 if (prefetch_tunable_set == 0) { 5040 printf("ZFS NOTICE: Prefetch is disabled by default on i386 " 5041 "-- to enable,\n"); 5042 printf(" add \"vfs.zfs.prefetch_disable=0\" " 5043 "to /boot/loader.conf.\n"); 5044 zfs_prefetch_disable = 1; 5045 } 5046#else 5047 if ((((uint64_t)physmem * PAGESIZE) < (1ULL << 32)) && 5048 prefetch_tunable_set == 0) { 5049 printf("ZFS NOTICE: Prefetch is disabled by default if less " 5050 "than 4GB of RAM is present;\n" 5051 " to enable, add \"vfs.zfs.prefetch_disable=0\" " 5052 "to /boot/loader.conf.\n"); 5053 zfs_prefetch_disable = 1; 5054 } 5055#endif 5056 /* Warn about ZFS memory and address space requirements. */ 5057 if (((uint64_t)physmem * PAGESIZE) < (256 + 128 + 64) * (1 << 20)) { 5058 printf("ZFS WARNING: Recommended minimum RAM size is 512MB; " 5059 "expect unstable behavior.\n"); 5060 } 5061 if (kmem_size() < 512 * (1 << 20)) { 5062 printf("ZFS WARNING: Recommended minimum kmem_size is 512MB; " 5063 "expect unstable behavior.\n"); 5064 printf(" Consider tuning vm.kmem_size and " 5065 "vm.kmem_size_max\n"); 5066 printf(" in /boot/loader.conf.\n"); 5067 } 5068#endif 5069} 5070 5071void 5072arc_fini(void) 5073{ 5074 int i; 5075 5076 mutex_enter(&arc_reclaim_thr_lock); 5077 arc_thread_exit = 1; 5078 cv_signal(&arc_reclaim_thr_cv); 5079 while (arc_thread_exit != 0) 5080 cv_wait(&arc_reclaim_thr_cv, &arc_reclaim_thr_lock); 5081 mutex_exit(&arc_reclaim_thr_lock); 5082 5083 arc_flush(NULL); 5084 5085 arc_dead = TRUE; 5086 5087 if (arc_ksp != NULL) { 5088 kstat_delete(arc_ksp); 5089 arc_ksp = NULL; 5090 } 5091 5092 mutex_destroy(&arc_eviction_mtx); 5093 mutex_destroy(&arc_reclaim_thr_lock); 5094 cv_destroy(&arc_reclaim_thr_cv); 5095 5096 for (i = 0; i < ARC_BUFC_NUMLISTS; i++) { 5097 list_destroy(&arc_mru->arcs_lists[i]); 5098 list_destroy(&arc_mru_ghost->arcs_lists[i]); 5099 list_destroy(&arc_mfu->arcs_lists[i]); 5100 list_destroy(&arc_mfu_ghost->arcs_lists[i]); 5101 list_destroy(&arc_l2c_only->arcs_lists[i]); 5102 5103 mutex_destroy(&arc_anon->arcs_locks[i].arcs_lock); 5104 mutex_destroy(&arc_mru->arcs_locks[i].arcs_lock); 5105 mutex_destroy(&arc_mru_ghost->arcs_locks[i].arcs_lock); 5106 mutex_destroy(&arc_mfu->arcs_locks[i].arcs_lock); 5107 mutex_destroy(&arc_mfu_ghost->arcs_locks[i].arcs_lock); 5108 mutex_destroy(&arc_l2c_only->arcs_locks[i].arcs_lock); 5109 } 5110 5111 buf_fini(); 5112 5113 ASSERT0(arc_loaned_bytes); 5114 5115#ifdef _KERNEL 5116 if (arc_event_lowmem != NULL) 5117 EVENTHANDLER_DEREGISTER(vm_lowmem, arc_event_lowmem); 5118#endif 5119} 5120 5121/* 5122 * Level 2 ARC 5123 * 5124 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk. 5125 * It uses dedicated storage devices to hold cached data, which are populated 5126 * using large infrequent writes. The main role of this cache is to boost 5127 * the performance of random read workloads. The intended L2ARC devices 5128 * include short-stroked disks, solid state disks, and other media with 5129 * substantially faster read latency than disk. 5130 * 5131 * +-----------------------+ 5132 * | ARC | 5133 * +-----------------------+ 5134 * | ^ ^ 5135 * | | | 5136 * l2arc_feed_thread() arc_read() 5137 * | | | 5138 * | l2arc read | 5139 * V | | 5140 * +---------------+ | 5141 * | L2ARC | | 5142 * +---------------+ | 5143 * | ^ | 5144 * l2arc_write() | | 5145 * | | | 5146 * V | | 5147 * +-------+ +-------+ 5148 * | vdev | | vdev | 5149 * | cache | | cache | 5150 * +-------+ +-------+ 5151 * +=========+ .-----. 5152 * : L2ARC : |-_____-| 5153 * : devices : | Disks | 5154 * +=========+ `-_____-' 5155 * 5156 * Read requests are satisfied from the following sources, in order: 5157 * 5158 * 1) ARC 5159 * 2) vdev cache of L2ARC devices 5160 * 3) L2ARC devices 5161 * 4) vdev cache of disks 5162 * 5) disks 5163 * 5164 * Some L2ARC device types exhibit extremely slow write performance. 5165 * To accommodate for this there are some significant differences between 5166 * the L2ARC and traditional cache design: 5167 * 5168 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from 5169 * the ARC behave as usual, freeing buffers and placing headers on ghost 5170 * lists. The ARC does not send buffers to the L2ARC during eviction as 5171 * this would add inflated write latencies for all ARC memory pressure. 5172 * 5173 * 2. The L2ARC attempts to cache data from the ARC before it is evicted. 5174 * It does this by periodically scanning buffers from the eviction-end of 5175 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are 5176 * not already there. It scans until a headroom of buffers is satisfied, 5177 * which itself is a buffer for ARC eviction. If a compressible buffer is 5178 * found during scanning and selected for writing to an L2ARC device, we 5179 * temporarily boost scanning headroom during the next scan cycle to make 5180 * sure we adapt to compression effects (which might significantly reduce 5181 * the data volume we write to L2ARC). The thread that does this is 5182 * l2arc_feed_thread(), illustrated below; example sizes are included to 5183 * provide a better sense of ratio than this diagram: 5184 * 5185 * head --> tail 5186 * +---------------------+----------+ 5187 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC 5188 * +---------------------+----------+ | o L2ARC eligible 5189 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer 5190 * +---------------------+----------+ | 5191 * 15.9 Gbytes ^ 32 Mbytes | 5192 * headroom | 5193 * l2arc_feed_thread() 5194 * | 5195 * l2arc write hand <--[oooo]--' 5196 * | 8 Mbyte 5197 * | write max 5198 * V 5199 * +==============================+ 5200 * L2ARC dev |####|#|###|###| |####| ... | 5201 * +==============================+ 5202 * 32 Gbytes 5203 * 5204 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of 5205 * evicted, then the L2ARC has cached a buffer much sooner than it probably 5206 * needed to, potentially wasting L2ARC device bandwidth and storage. It is 5207 * safe to say that this is an uncommon case, since buffers at the end of 5208 * the ARC lists have moved there due to inactivity. 5209 * 5210 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom, 5211 * then the L2ARC simply misses copying some buffers. This serves as a 5212 * pressure valve to prevent heavy read workloads from both stalling the ARC 5213 * with waits and clogging the L2ARC with writes. This also helps prevent 5214 * the potential for the L2ARC to churn if it attempts to cache content too 5215 * quickly, such as during backups of the entire pool. 5216 * 5217 * 5. After system boot and before the ARC has filled main memory, there are 5218 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru 5219 * lists can remain mostly static. Instead of searching from tail of these 5220 * lists as pictured, the l2arc_feed_thread() will search from the list heads 5221 * for eligible buffers, greatly increasing its chance of finding them. 5222 * 5223 * The L2ARC device write speed is also boosted during this time so that 5224 * the L2ARC warms up faster. Since there have been no ARC evictions yet, 5225 * there are no L2ARC reads, and no fear of degrading read performance 5226 * through increased writes. 5227 * 5228 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that 5229 * the vdev queue can aggregate them into larger and fewer writes. Each 5230 * device is written to in a rotor fashion, sweeping writes through 5231 * available space then repeating. 5232 * 5233 * 7. The L2ARC does not store dirty content. It never needs to flush 5234 * write buffers back to disk based storage. 5235 * 5236 * 8. If an ARC buffer is written (and dirtied) which also exists in the 5237 * L2ARC, the now stale L2ARC buffer is immediately dropped. 5238 * 5239 * The performance of the L2ARC can be tweaked by a number of tunables, which 5240 * may be necessary for different workloads: 5241 * 5242 * l2arc_write_max max write bytes per interval 5243 * l2arc_write_boost extra write bytes during device warmup 5244 * l2arc_noprefetch skip caching prefetched buffers 5245 * l2arc_headroom number of max device writes to precache 5246 * l2arc_headroom_boost when we find compressed buffers during ARC 5247 * scanning, we multiply headroom by this 5248 * percentage factor for the next scan cycle, 5249 * since more compressed buffers are likely to 5250 * be present 5251 * l2arc_feed_secs seconds between L2ARC writing 5252 * 5253 * Tunables may be removed or added as future performance improvements are 5254 * integrated, and also may become zpool properties. 5255 * 5256 * There are three key functions that control how the L2ARC warms up: 5257 * 5258 * l2arc_write_eligible() check if a buffer is eligible to cache 5259 * l2arc_write_size() calculate how much to write 5260 * l2arc_write_interval() calculate sleep delay between writes 5261 * 5262 * These three functions determine what to write, how much, and how quickly 5263 * to send writes. 5264 */ 5265 5266static boolean_t 5267l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *hdr) 5268{ 5269 /* 5270 * A buffer is *not* eligible for the L2ARC if it: 5271 * 1. belongs to a different spa. 5272 * 2. is already cached on the L2ARC. 5273 * 3. has an I/O in progress (it may be an incomplete read). 5274 * 4. is flagged not eligible (zfs property). 5275 */ 5276 if (hdr->b_spa != spa_guid) { 5277 ARCSTAT_BUMP(arcstat_l2_write_spa_mismatch); 5278 return (B_FALSE); 5279 } 5280 if (HDR_HAS_L2HDR(hdr)) { 5281 ARCSTAT_BUMP(arcstat_l2_write_in_l2); 5282 return (B_FALSE); 5283 } 5284 if (HDR_IO_IN_PROGRESS(hdr)) { 5285 ARCSTAT_BUMP(arcstat_l2_write_hdr_io_in_progress); 5286 return (B_FALSE); 5287 } 5288 if (!HDR_L2CACHE(hdr)) { 5289 ARCSTAT_BUMP(arcstat_l2_write_not_cacheable); 5290 return (B_FALSE); 5291 } 5292 5293 return (B_TRUE); 5294} 5295 5296static uint64_t 5297l2arc_write_size(void) 5298{ 5299 uint64_t size; 5300 5301 /* 5302 * Make sure our globals have meaningful values in case the user 5303 * altered them. 5304 */ 5305 size = l2arc_write_max; 5306 if (size == 0) { 5307 cmn_err(CE_NOTE, "Bad value for l2arc_write_max, value must " 5308 "be greater than zero, resetting it to the default (%d)", 5309 L2ARC_WRITE_SIZE); 5310 size = l2arc_write_max = L2ARC_WRITE_SIZE; 5311 } 5312 5313 if (arc_warm == B_FALSE) 5314 size += l2arc_write_boost; 5315 5316 return (size); 5317 5318} 5319 5320static clock_t 5321l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote) 5322{ 5323 clock_t interval, next, now; 5324 5325 /* 5326 * If the ARC lists are busy, increase our write rate; if the 5327 * lists are stale, idle back. This is achieved by checking 5328 * how much we previously wrote - if it was more than half of 5329 * what we wanted, schedule the next write much sooner. 5330 */ 5331 if (l2arc_feed_again && wrote > (wanted / 2)) 5332 interval = (hz * l2arc_feed_min_ms) / 1000; 5333 else 5334 interval = hz * l2arc_feed_secs; 5335 5336 now = ddi_get_lbolt(); 5337 next = MAX(now, MIN(now + interval, began + interval)); 5338 5339 return (next); 5340} 5341 5342/* 5343 * Cycle through L2ARC devices. This is how L2ARC load balances. 5344 * If a device is returned, this also returns holding the spa config lock. 5345 */ 5346static l2arc_dev_t * 5347l2arc_dev_get_next(void) 5348{ 5349 l2arc_dev_t *first, *next = NULL; 5350 5351 /* 5352 * Lock out the removal of spas (spa_namespace_lock), then removal 5353 * of cache devices (l2arc_dev_mtx). Once a device has been selected, 5354 * both locks will be dropped and a spa config lock held instead. 5355 */ 5356 mutex_enter(&spa_namespace_lock); 5357 mutex_enter(&l2arc_dev_mtx); 5358 5359 /* if there are no vdevs, there is nothing to do */ 5360 if (l2arc_ndev == 0) 5361 goto out; 5362 5363 first = NULL; 5364 next = l2arc_dev_last; 5365 do { 5366 /* loop around the list looking for a non-faulted vdev */ 5367 if (next == NULL) { 5368 next = list_head(l2arc_dev_list); 5369 } else { 5370 next = list_next(l2arc_dev_list, next); 5371 if (next == NULL) 5372 next = list_head(l2arc_dev_list); 5373 } 5374 5375 /* if we have come back to the start, bail out */ 5376 if (first == NULL) 5377 first = next; 5378 else if (next == first) 5379 break; 5380 5381 } while (vdev_is_dead(next->l2ad_vdev)); 5382 5383 /* if we were unable to find any usable vdevs, return NULL */ 5384 if (vdev_is_dead(next->l2ad_vdev)) 5385 next = NULL; 5386 5387 l2arc_dev_last = next; 5388 5389out: 5390 mutex_exit(&l2arc_dev_mtx); 5391 5392 /* 5393 * Grab the config lock to prevent the 'next' device from being 5394 * removed while we are writing to it. 5395 */ 5396 if (next != NULL) 5397 spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER); 5398 mutex_exit(&spa_namespace_lock); 5399 5400 return (next); 5401} 5402 5403/* 5404 * Free buffers that were tagged for destruction. 5405 */ 5406static void 5407l2arc_do_free_on_write() 5408{ 5409 list_t *buflist; 5410 l2arc_data_free_t *df, *df_prev; 5411 5412 mutex_enter(&l2arc_free_on_write_mtx); 5413 buflist = l2arc_free_on_write; 5414 5415 for (df = list_tail(buflist); df; df = df_prev) { 5416 df_prev = list_prev(buflist, df); 5417 ASSERT(df->l2df_data != NULL); 5418 ASSERT(df->l2df_func != NULL); 5419 df->l2df_func(df->l2df_data, df->l2df_size); 5420 list_remove(buflist, df); 5421 kmem_free(df, sizeof (l2arc_data_free_t)); 5422 } 5423 5424 mutex_exit(&l2arc_free_on_write_mtx); 5425} 5426 5427/* 5428 * A write to a cache device has completed. Update all headers to allow 5429 * reads from these buffers to begin. 5430 */ 5431static void 5432l2arc_write_done(zio_t *zio) 5433{ 5434 l2arc_write_callback_t *cb; 5435 l2arc_dev_t *dev; 5436 list_t *buflist; 5437 arc_buf_hdr_t *head, *hdr, *hdr_prev; 5438 kmutex_t *hash_lock; 5439 int64_t bytes_dropped = 0; 5440 5441 cb = zio->io_private; 5442 ASSERT(cb != NULL); 5443 dev = cb->l2wcb_dev; 5444 ASSERT(dev != NULL); 5445 head = cb->l2wcb_head; 5446 ASSERT(head != NULL); 5447 buflist = &dev->l2ad_buflist; 5448 ASSERT(buflist != NULL); 5449 DTRACE_PROBE2(l2arc__iodone, zio_t *, zio, 5450 l2arc_write_callback_t *, cb); 5451 5452 if (zio->io_error != 0) 5453 ARCSTAT_BUMP(arcstat_l2_writes_error); 5454 5455 mutex_enter(&dev->l2ad_mtx); 5456 5457 /* 5458 * All writes completed, or an error was hit. 5459 */ 5460 for (hdr = list_prev(buflist, head); hdr; hdr = hdr_prev) { 5461 hdr_prev = list_prev(buflist, hdr); 5462 5463 hash_lock = HDR_LOCK(hdr); 5464 if (!mutex_tryenter(hash_lock)) { 5465 /* 5466 * This buffer misses out. It may be in a stage 5467 * of eviction. Its ARC_FLAG_L2_WRITING flag will be 5468 * left set, denying reads to this buffer. 5469 */ 5470 ARCSTAT_BUMP(arcstat_l2_writes_hdr_miss); 5471 continue; 5472 } 5473 5474 /* 5475 * It's possible that this buffer got evicted from the L1 cache 5476 * before we grabbed the vdev + hash locks, in which case 5477 * arc_hdr_realloc freed b_tmp_cdata for us if it was allocated. 5478 * Only free the buffer if we still have an L1 hdr. 5479 */ 5480 if (HDR_HAS_L1HDR(hdr) && hdr->b_l1hdr.b_tmp_cdata != NULL && 5481 HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF) 5482 l2arc_release_cdata_buf(hdr); 5483 5484 if (zio->io_error != 0) { 5485 /* 5486 * Error - drop L2ARC entry. 5487 */ 5488 trim_map_free(hdr->b_l2hdr.b_dev->l2ad_vdev, 5489 hdr->b_l2hdr.b_daddr, hdr->b_l2hdr.b_asize, 0); 5490 hdr->b_flags &= ~ARC_FLAG_HAS_L2HDR; 5491 5492 ARCSTAT_INCR(arcstat_l2_asize, -hdr->b_l2hdr.b_asize); 5493 ARCSTAT_INCR(arcstat_l2_size, -hdr->b_size); 5494 5495 bytes_dropped += hdr->b_l2hdr.b_asize; 5496 (void) refcount_remove_many(&dev->l2ad_alloc, 5497 hdr->b_l2hdr.b_asize, hdr); 5498 } 5499 5500 /* 5501 * Allow ARC to begin reads to this L2ARC entry. 5502 */ 5503 hdr->b_flags &= ~ARC_FLAG_L2_WRITING; 5504 5505 mutex_exit(hash_lock); 5506 } 5507 5508 atomic_inc_64(&l2arc_writes_done); 5509 list_remove(buflist, head); 5510 ASSERT(!HDR_HAS_L1HDR(head)); 5511 kmem_cache_free(hdr_l2only_cache, head); 5512 mutex_exit(&dev->l2ad_mtx); 5513 5514 vdev_space_update(dev->l2ad_vdev, -bytes_dropped, 0, 0); 5515 5516 l2arc_do_free_on_write(); 5517 5518 kmem_free(cb, sizeof (l2arc_write_callback_t)); 5519} 5520 5521/* 5522 * A read to a cache device completed. Validate buffer contents before 5523 * handing over to the regular ARC routines. 5524 */ 5525static void 5526l2arc_read_done(zio_t *zio) 5527{ 5528 l2arc_read_callback_t *cb; 5529 arc_buf_hdr_t *hdr; 5530 arc_buf_t *buf; 5531 kmutex_t *hash_lock; 5532 int equal; 5533 5534 ASSERT(zio->io_vd != NULL); 5535 ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE); 5536 5537 spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd); 5538 5539 cb = zio->io_private; 5540 ASSERT(cb != NULL); 5541 buf = cb->l2rcb_buf; 5542 ASSERT(buf != NULL); 5543 5544 hash_lock = HDR_LOCK(buf->b_hdr); 5545 mutex_enter(hash_lock); 5546 hdr = buf->b_hdr; 5547 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 5548 5549 /* 5550 * If the buffer was compressed, decompress it first. 5551 */ 5552 if (cb->l2rcb_compress != ZIO_COMPRESS_OFF) 5553 l2arc_decompress_zio(zio, hdr, cb->l2rcb_compress); 5554 ASSERT(zio->io_data != NULL); 5555 5556 /* 5557 * Check this survived the L2ARC journey. 5558 */ 5559 equal = arc_cksum_equal(buf); 5560 if (equal && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) { 5561 mutex_exit(hash_lock); 5562 zio->io_private = buf; 5563 zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */ 5564 zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */ 5565 arc_read_done(zio); 5566 } else { 5567 mutex_exit(hash_lock); 5568 /* 5569 * Buffer didn't survive caching. Increment stats and 5570 * reissue to the original storage device. 5571 */ 5572 if (zio->io_error != 0) { 5573 ARCSTAT_BUMP(arcstat_l2_io_error); 5574 } else { 5575 zio->io_error = SET_ERROR(EIO); 5576 } 5577 if (!equal) 5578 ARCSTAT_BUMP(arcstat_l2_cksum_bad); 5579 5580 /* 5581 * If there's no waiter, issue an async i/o to the primary 5582 * storage now. If there *is* a waiter, the caller must 5583 * issue the i/o in a context where it's OK to block. 5584 */ 5585 if (zio->io_waiter == NULL) { 5586 zio_t *pio = zio_unique_parent(zio); 5587 5588 ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL); 5589 5590 zio_nowait(zio_read(pio, cb->l2rcb_spa, &cb->l2rcb_bp, 5591 buf->b_data, zio->io_size, arc_read_done, buf, 5592 zio->io_priority, cb->l2rcb_flags, &cb->l2rcb_zb)); 5593 } 5594 } 5595 5596 kmem_free(cb, sizeof (l2arc_read_callback_t)); 5597} 5598 5599/* 5600 * This is the list priority from which the L2ARC will search for pages to 5601 * cache. This is used within loops (0..3) to cycle through lists in the 5602 * desired order. This order can have a significant effect on cache 5603 * performance. 5604 * 5605 * Currently the metadata lists are hit first, MFU then MRU, followed by 5606 * the data lists. This function returns a locked list, and also returns 5607 * the lock pointer. 5608 */ 5609static list_t * 5610l2arc_list_locked(int list_num, kmutex_t **lock) 5611{ 5612 list_t *list = NULL; 5613 int idx; 5614 5615 ASSERT(list_num >= 0 && list_num < 2 * ARC_BUFC_NUMLISTS); 5616 5617 if (list_num < ARC_BUFC_NUMMETADATALISTS) { 5618 idx = list_num; 5619 list = &arc_mfu->arcs_lists[idx]; 5620 *lock = ARCS_LOCK(arc_mfu, idx); 5621 } else if (list_num < ARC_BUFC_NUMMETADATALISTS * 2) { 5622 idx = list_num - ARC_BUFC_NUMMETADATALISTS; 5623 list = &arc_mru->arcs_lists[idx]; 5624 *lock = ARCS_LOCK(arc_mru, idx); 5625 } else if (list_num < (ARC_BUFC_NUMMETADATALISTS * 2 + 5626 ARC_BUFC_NUMDATALISTS)) { 5627 idx = list_num - ARC_BUFC_NUMMETADATALISTS; 5628 list = &arc_mfu->arcs_lists[idx]; 5629 *lock = ARCS_LOCK(arc_mfu, idx); 5630 } else { 5631 idx = list_num - ARC_BUFC_NUMLISTS; 5632 list = &arc_mru->arcs_lists[idx]; 5633 *lock = ARCS_LOCK(arc_mru, idx); 5634 } 5635 5636 ASSERT(!(MUTEX_HELD(*lock))); 5637 mutex_enter(*lock); 5638 return (list); 5639} 5640 5641/* 5642 * Evict buffers from the device write hand to the distance specified in 5643 * bytes. This distance may span populated buffers, it may span nothing. 5644 * This is clearing a region on the L2ARC device ready for writing. 5645 * If the 'all' boolean is set, every buffer is evicted. 5646 */ 5647static void 5648l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all) 5649{ 5650 list_t *buflist; 5651 arc_buf_hdr_t *hdr, *hdr_prev; 5652 kmutex_t *hash_lock; 5653 uint64_t taddr; 5654 5655 buflist = &dev->l2ad_buflist; 5656 5657 if (!all && dev->l2ad_first) { 5658 /* 5659 * This is the first sweep through the device. There is 5660 * nothing to evict. 5661 */ 5662 return; 5663 } 5664 5665 if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) { 5666 /* 5667 * When nearing the end of the device, evict to the end 5668 * before the device write hand jumps to the start. 5669 */ 5670 taddr = dev->l2ad_end; 5671 } else { 5672 taddr = dev->l2ad_hand + distance; 5673 } 5674 DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist, 5675 uint64_t, taddr, boolean_t, all); 5676 5677top: 5678 mutex_enter(&dev->l2ad_mtx); 5679 for (hdr = list_tail(buflist); hdr; hdr = hdr_prev) { 5680 hdr_prev = list_prev(buflist, hdr); 5681 5682 hash_lock = HDR_LOCK(hdr); 5683 if (!mutex_tryenter(hash_lock)) { 5684 /* 5685 * Missed the hash lock. Retry. 5686 */ 5687 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry); 5688 mutex_exit(&dev->l2ad_mtx); 5689 mutex_enter(hash_lock); 5690 mutex_exit(hash_lock); 5691 goto top; 5692 } 5693 5694 if (HDR_L2_WRITE_HEAD(hdr)) { 5695 /* 5696 * We hit a write head node. Leave it for 5697 * l2arc_write_done(). 5698 */ 5699 list_remove(buflist, hdr); 5700 mutex_exit(hash_lock); 5701 continue; 5702 } 5703 5704 if (!all && HDR_HAS_L2HDR(hdr) && 5705 (hdr->b_l2hdr.b_daddr > taddr || 5706 hdr->b_l2hdr.b_daddr < dev->l2ad_hand)) { 5707 /* 5708 * We've evicted to the target address, 5709 * or the end of the device. 5710 */ 5711 mutex_exit(hash_lock); 5712 break; 5713 } 5714 5715 ASSERT(HDR_HAS_L2HDR(hdr)); 5716 if (!HDR_HAS_L1HDR(hdr)) { 5717 ASSERT(!HDR_L2_READING(hdr)); 5718 /* 5719 * This doesn't exist in the ARC. Destroy. 5720 * arc_hdr_destroy() will call list_remove() 5721 * and decrement arcstat_l2_size. 5722 */ 5723 arc_change_state(arc_anon, hdr, hash_lock); 5724 arc_hdr_destroy(hdr); 5725 } else { 5726 ASSERT(hdr->b_l1hdr.b_state != arc_l2c_only); 5727 ARCSTAT_BUMP(arcstat_l2_evict_l1cached); 5728 /* 5729 * Invalidate issued or about to be issued 5730 * reads, since we may be about to write 5731 * over this location. 5732 */ 5733 if (HDR_L2_READING(hdr)) { 5734 ARCSTAT_BUMP(arcstat_l2_evict_reading); 5735 hdr->b_flags |= ARC_FLAG_L2_EVICTED; 5736 } 5737 5738 arc_hdr_l2hdr_destroy(hdr); 5739 } 5740 mutex_exit(hash_lock); 5741 } 5742 mutex_exit(&dev->l2ad_mtx); 5743} 5744 5745/* 5746 * Find and write ARC buffers to the L2ARC device. 5747 * 5748 * An ARC_FLAG_L2_WRITING flag is set so that the L2ARC buffers are not valid 5749 * for reading until they have completed writing. 5750 * The headroom_boost is an in-out parameter used to maintain headroom boost 5751 * state between calls to this function. 5752 * 5753 * Returns the number of bytes actually written (which may be smaller than 5754 * the delta by which the device hand has changed due to alignment). 5755 */ 5756static uint64_t 5757l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz, 5758 boolean_t *headroom_boost) 5759{ 5760 arc_buf_hdr_t *hdr, *hdr_prev, *head; 5761 list_t *list; 5762 uint64_t write_asize, write_sz, headroom, buf_compress_minsz; 5763 void *buf_data; 5764 kmutex_t *list_lock; 5765 boolean_t full; 5766 l2arc_write_callback_t *cb; 5767 zio_t *pio, *wzio; 5768 uint64_t guid = spa_load_guid(spa); 5769 const boolean_t do_headroom_boost = *headroom_boost; 5770 int try; 5771 5772 ASSERT(dev->l2ad_vdev != NULL); 5773 5774 /* Lower the flag now, we might want to raise it again later. */ 5775 *headroom_boost = B_FALSE; 5776 5777 pio = NULL; 5778 write_sz = write_asize = 0; 5779 full = B_FALSE; 5780 head = kmem_cache_alloc(hdr_l2only_cache, KM_PUSHPAGE); 5781 head->b_flags |= ARC_FLAG_L2_WRITE_HEAD; 5782 head->b_flags |= ARC_FLAG_HAS_L2HDR; 5783 5784 ARCSTAT_BUMP(arcstat_l2_write_buffer_iter); 5785 /* 5786 * We will want to try to compress buffers that are at least 2x the 5787 * device sector size. 5788 */ 5789 buf_compress_minsz = 2 << dev->l2ad_vdev->vdev_ashift; 5790 5791 /* 5792 * Copy buffers for L2ARC writing. 5793 */ 5794 mutex_enter(&dev->l2ad_mtx); 5795 for (try = 0; try < 2 * ARC_BUFC_NUMLISTS; try++) { 5796 uint64_t passed_sz = 0; 5797 5798 list = l2arc_list_locked(try, &list_lock); 5799 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_iter); 5800 5801 /* 5802 * L2ARC fast warmup. 5803 * 5804 * Until the ARC is warm and starts to evict, read from the 5805 * head of the ARC lists rather than the tail. 5806 */ 5807 if (arc_warm == B_FALSE) 5808 hdr = list_head(list); 5809 else 5810 hdr = list_tail(list); 5811 if (hdr == NULL) 5812 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_null_iter); 5813 5814 headroom = target_sz * l2arc_headroom * 2 / ARC_BUFC_NUMLISTS; 5815 if (do_headroom_boost) 5816 headroom = (headroom * l2arc_headroom_boost) / 100; 5817 5818 for (; hdr; hdr = hdr_prev) { 5819 kmutex_t *hash_lock; 5820 uint64_t buf_sz; 5821 uint64_t buf_a_sz; 5822 5823 if (arc_warm == B_FALSE) 5824 hdr_prev = list_next(list, hdr); 5825 else 5826 hdr_prev = list_prev(list, hdr); 5827 ARCSTAT_INCR(arcstat_l2_write_buffer_bytes_scanned, hdr->b_size); 5828 5829 hash_lock = HDR_LOCK(hdr); 5830 if (!mutex_tryenter(hash_lock)) { 5831 ARCSTAT_BUMP(arcstat_l2_write_trylock_fail); 5832 /* 5833 * Skip this buffer rather than waiting. 5834 */ 5835 continue; 5836 } 5837 5838 passed_sz += hdr->b_size; 5839 if (passed_sz > headroom) { 5840 /* 5841 * Searched too far. 5842 */ 5843 mutex_exit(hash_lock); 5844 ARCSTAT_BUMP(arcstat_l2_write_passed_headroom); 5845 break; 5846 } 5847 5848 if (!l2arc_write_eligible(guid, hdr)) { 5849 mutex_exit(hash_lock); 5850 continue; 5851 } 5852 5853 /* 5854 * Assume that the buffer is not going to be compressed 5855 * and could take more space on disk because of a larger 5856 * disk block size. 5857 */ 5858 buf_sz = hdr->b_size; 5859 buf_a_sz = vdev_psize_to_asize(dev->l2ad_vdev, buf_sz); 5860 5861 if ((write_asize + buf_a_sz) > target_sz) { 5862 full = B_TRUE; 5863 mutex_exit(hash_lock); 5864 ARCSTAT_BUMP(arcstat_l2_write_full); 5865 break; 5866 } 5867 5868 if (pio == NULL) { 5869 /* 5870 * Insert a dummy header on the buflist so 5871 * l2arc_write_done() can find where the 5872 * write buffers begin without searching. 5873 */ 5874 list_insert_head(&dev->l2ad_buflist, head); 5875 5876 cb = kmem_alloc( 5877 sizeof (l2arc_write_callback_t), KM_SLEEP); 5878 cb->l2wcb_dev = dev; 5879 cb->l2wcb_head = head; 5880 pio = zio_root(spa, l2arc_write_done, cb, 5881 ZIO_FLAG_CANFAIL); 5882 ARCSTAT_BUMP(arcstat_l2_write_pios); 5883 } 5884 5885 /* 5886 * Create and add a new L2ARC header. 5887 */ 5888 hdr->b_l2hdr.b_dev = dev; 5889 hdr->b_flags |= ARC_FLAG_L2_WRITING; 5890 /* 5891 * Temporarily stash the data buffer in b_tmp_cdata. 5892 * The subsequent write step will pick it up from 5893 * there. This is because can't access b_l1hdr.b_buf 5894 * without holding the hash_lock, which we in turn 5895 * can't access without holding the ARC list locks 5896 * (which we want to avoid during compression/writing). 5897 */ 5898 HDR_SET_COMPRESS(hdr, ZIO_COMPRESS_OFF); 5899 hdr->b_l2hdr.b_asize = hdr->b_size; 5900 hdr->b_l1hdr.b_tmp_cdata = hdr->b_l1hdr.b_buf->b_data; 5901 5902 /* 5903 * Explicitly set the b_daddr field to a known 5904 * value which means "invalid address". This 5905 * enables us to differentiate which stage of 5906 * l2arc_write_buffers() the particular header 5907 * is in (e.g. this loop, or the one below). 5908 * ARC_FLAG_L2_WRITING is not enough to make 5909 * this distinction, and we need to know in 5910 * order to do proper l2arc vdev accounting in 5911 * arc_release() and arc_hdr_destroy(). 5912 * 5913 * Note, we can't use a new flag to distinguish 5914 * the two stages because we don't hold the 5915 * header's hash_lock below, in the second stage 5916 * of this function. Thus, we can't simply 5917 * change the b_flags field to denote that the 5918 * IO has been sent. We can change the b_daddr 5919 * field of the L2 portion, though, since we'll 5920 * be holding the l2ad_mtx; which is why we're 5921 * using it to denote the header's state change. 5922 */ 5923 hdr->b_l2hdr.b_daddr = L2ARC_ADDR_UNSET; 5924 hdr->b_flags |= ARC_FLAG_HAS_L2HDR; 5925 5926 list_insert_head(&dev->l2ad_buflist, hdr); 5927 5928 /* 5929 * Compute and store the buffer cksum before 5930 * writing. On debug the cksum is verified first. 5931 */ 5932 arc_cksum_verify(hdr->b_l1hdr.b_buf); 5933 arc_cksum_compute(hdr->b_l1hdr.b_buf, B_TRUE); 5934 5935 mutex_exit(hash_lock); 5936 5937 write_sz += buf_sz; 5938 write_asize += buf_a_sz; 5939 } 5940 5941 mutex_exit(list_lock); 5942 5943 if (full == B_TRUE) 5944 break; 5945 } 5946 5947 /* No buffers selected for writing? */ 5948 if (pio == NULL) { 5949 ASSERT0(write_sz); 5950 mutex_exit(&dev->l2ad_mtx); 5951 ASSERT(!HDR_HAS_L1HDR(head)); 5952 kmem_cache_free(hdr_l2only_cache, head); 5953 return (0); 5954 } 5955 5956 /* 5957 * Note that elsewhere in this file arcstat_l2_asize 5958 * and the used space on l2ad_vdev are updated using b_asize, 5959 * which is not necessarily rounded up to the device block size. 5960 * Too keep accounting consistent we do the same here as well: 5961 * stats_size accumulates the sum of b_asize of the written buffers, 5962 * while write_asize accumulates the sum of b_asize rounded up 5963 * to the device block size. 5964 * The latter sum is used only to validate the corectness of the code. 5965 */ 5966 uint64_t stats_size = 0; 5967 write_asize = 0; 5968 5969 /* 5970 * Now start writing the buffers. We're starting at the write head 5971 * and work backwards, retracing the course of the buffer selector 5972 * loop above. 5973 */ 5974 for (hdr = list_prev(&dev->l2ad_buflist, head); hdr; 5975 hdr = list_prev(&dev->l2ad_buflist, hdr)) { 5976 uint64_t buf_sz; 5977 5978 /* 5979 * We shouldn't need to lock the buffer here, since we flagged 5980 * it as ARC_FLAG_L2_WRITING in the previous step, but we must 5981 * take care to only access its L2 cache parameters. In 5982 * particular, hdr->l1hdr.b_buf may be invalid by now due to 5983 * ARC eviction. 5984 */ 5985 hdr->b_l2hdr.b_daddr = dev->l2ad_hand; 5986 5987 if ((HDR_L2COMPRESS(hdr)) && 5988 hdr->b_l2hdr.b_asize >= buf_compress_minsz) { 5989 if (l2arc_compress_buf(hdr)) { 5990 /* 5991 * If compression succeeded, enable headroom 5992 * boost on the next scan cycle. 5993 */ 5994 *headroom_boost = B_TRUE; 5995 } 5996 } 5997 5998 /* 5999 * Pick up the buffer data we had previously stashed away 6000 * (and now potentially also compressed). 6001 */ 6002 buf_data = hdr->b_l1hdr.b_tmp_cdata; 6003 buf_sz = hdr->b_l2hdr.b_asize; 6004 6005 /* 6006 * If the data has not been compressed, then clear b_tmp_cdata 6007 * to make sure that it points only to a temporary compression 6008 * buffer. 6009 */ 6010 if (!L2ARC_IS_VALID_COMPRESS(HDR_GET_COMPRESS(hdr))) 6011 hdr->b_l1hdr.b_tmp_cdata = NULL; 6012 6013 /* 6014 * We need to do this regardless if buf_sz is zero or 6015 * not, otherwise, when this l2hdr is evicted we'll 6016 * remove a reference that was never added. 6017 */ 6018 (void) refcount_add_many(&dev->l2ad_alloc, buf_sz, hdr); 6019 6020 /* Compression may have squashed the buffer to zero length. */ 6021 if (buf_sz != 0) { 6022 uint64_t buf_a_sz; 6023 6024 wzio = zio_write_phys(pio, dev->l2ad_vdev, 6025 dev->l2ad_hand, buf_sz, buf_data, ZIO_CHECKSUM_OFF, 6026 NULL, NULL, ZIO_PRIORITY_ASYNC_WRITE, 6027 ZIO_FLAG_CANFAIL, B_FALSE); 6028 6029 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev, 6030 zio_t *, wzio); 6031 (void) zio_nowait(wzio); 6032 6033 stats_size += buf_sz; 6034 6035 /* 6036 * Keep the clock hand suitably device-aligned. 6037 */ 6038 buf_a_sz = vdev_psize_to_asize(dev->l2ad_vdev, buf_sz); 6039 write_asize += buf_a_sz; 6040 dev->l2ad_hand += buf_a_sz; 6041 } 6042 } 6043 6044 mutex_exit(&dev->l2ad_mtx); 6045 6046 ASSERT3U(write_asize, <=, target_sz); 6047 ARCSTAT_BUMP(arcstat_l2_writes_sent); 6048 ARCSTAT_INCR(arcstat_l2_write_bytes, write_asize); 6049 ARCSTAT_INCR(arcstat_l2_size, write_sz); 6050 ARCSTAT_INCR(arcstat_l2_asize, stats_size); 6051 vdev_space_update(dev->l2ad_vdev, stats_size, 0, 0); 6052 6053 /* 6054 * Bump device hand to the device start if it is approaching the end. 6055 * l2arc_evict() will already have evicted ahead for this case. 6056 */ 6057 if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) { 6058 dev->l2ad_hand = dev->l2ad_start; 6059 dev->l2ad_first = B_FALSE; 6060 } 6061 6062 dev->l2ad_writing = B_TRUE; 6063 (void) zio_wait(pio); 6064 dev->l2ad_writing = B_FALSE; 6065 6066 return (write_asize); 6067} 6068 6069/* 6070 * Compresses an L2ARC buffer. 6071 * The data to be compressed must be prefilled in l1hdr.b_tmp_cdata and its 6072 * size in l2hdr->b_asize. This routine tries to compress the data and 6073 * depending on the compression result there are three possible outcomes: 6074 * *) The buffer was incompressible. The original l2hdr contents were left 6075 * untouched and are ready for writing to an L2 device. 6076 * *) The buffer was all-zeros, so there is no need to write it to an L2 6077 * device. To indicate this situation b_tmp_cdata is NULL'ed, b_asize is 6078 * set to zero and b_compress is set to ZIO_COMPRESS_EMPTY. 6079 * *) Compression succeeded and b_tmp_cdata was replaced with a temporary 6080 * data buffer which holds the compressed data to be written, and b_asize 6081 * tells us how much data there is. b_compress is set to the appropriate 6082 * compression algorithm. Once writing is done, invoke 6083 * l2arc_release_cdata_buf on this l2hdr to free this temporary buffer. 6084 * 6085 * Returns B_TRUE if compression succeeded, or B_FALSE if it didn't (the 6086 * buffer was incompressible). 6087 */ 6088static boolean_t 6089l2arc_compress_buf(arc_buf_hdr_t *hdr) 6090{ 6091 void *cdata; 6092 size_t csize, len, rounded; 6093 ASSERT(HDR_HAS_L2HDR(hdr)); 6094 l2arc_buf_hdr_t *l2hdr = &hdr->b_l2hdr; 6095 6096 ASSERT(HDR_HAS_L1HDR(hdr)); 6097 ASSERT(HDR_GET_COMPRESS(hdr) == ZIO_COMPRESS_OFF); 6098 ASSERT(hdr->b_l1hdr.b_tmp_cdata != NULL); 6099 6100 len = l2hdr->b_asize; 6101 cdata = zio_data_buf_alloc(len); 6102 ASSERT3P(cdata, !=, NULL); 6103 csize = zio_compress_data(ZIO_COMPRESS_LZ4, hdr->b_l1hdr.b_tmp_cdata, 6104 cdata, l2hdr->b_asize); 6105 6106 if (csize == 0) { 6107 /* zero block, indicate that there's nothing to write */ 6108 zio_data_buf_free(cdata, len); 6109 HDR_SET_COMPRESS(hdr, ZIO_COMPRESS_EMPTY); 6110 l2hdr->b_asize = 0; 6111 hdr->b_l1hdr.b_tmp_cdata = NULL; 6112 ARCSTAT_BUMP(arcstat_l2_compress_zeros); 6113 return (B_TRUE); 6114 } 6115 6116 rounded = P2ROUNDUP(csize, 6117 (size_t)1 << l2hdr->b_dev->l2ad_vdev->vdev_ashift); 6118 if (rounded < len) { 6119 /* 6120 * Compression succeeded, we'll keep the cdata around for 6121 * writing and release it afterwards. 6122 */ 6123 if (rounded > csize) { 6124 bzero((char *)cdata + csize, rounded - csize); 6125 csize = rounded; 6126 } 6127 HDR_SET_COMPRESS(hdr, ZIO_COMPRESS_LZ4); 6128 l2hdr->b_asize = csize; 6129 hdr->b_l1hdr.b_tmp_cdata = cdata; 6130 ARCSTAT_BUMP(arcstat_l2_compress_successes); 6131 return (B_TRUE); 6132 } else { 6133 /* 6134 * Compression failed, release the compressed buffer. 6135 * l2hdr will be left unmodified. 6136 */ 6137 zio_data_buf_free(cdata, len); 6138 ARCSTAT_BUMP(arcstat_l2_compress_failures); 6139 return (B_FALSE); 6140 } 6141} 6142 6143/* 6144 * Decompresses a zio read back from an l2arc device. On success, the 6145 * underlying zio's io_data buffer is overwritten by the uncompressed 6146 * version. On decompression error (corrupt compressed stream), the 6147 * zio->io_error value is set to signal an I/O error. 6148 * 6149 * Please note that the compressed data stream is not checksummed, so 6150 * if the underlying device is experiencing data corruption, we may feed 6151 * corrupt data to the decompressor, so the decompressor needs to be 6152 * able to handle this situation (LZ4 does). 6153 */ 6154static void 6155l2arc_decompress_zio(zio_t *zio, arc_buf_hdr_t *hdr, enum zio_compress c) 6156{ 6157 ASSERT(L2ARC_IS_VALID_COMPRESS(c)); 6158 6159 if (zio->io_error != 0) { 6160 /* 6161 * An io error has occured, just restore the original io 6162 * size in preparation for a main pool read. 6163 */ 6164 zio->io_orig_size = zio->io_size = hdr->b_size; 6165 return; 6166 } 6167 6168 if (c == ZIO_COMPRESS_EMPTY) { 6169 /* 6170 * An empty buffer results in a null zio, which means we 6171 * need to fill its io_data after we're done restoring the 6172 * buffer's contents. 6173 */ 6174 ASSERT(hdr->b_l1hdr.b_buf != NULL); 6175 bzero(hdr->b_l1hdr.b_buf->b_data, hdr->b_size); 6176 zio->io_data = zio->io_orig_data = hdr->b_l1hdr.b_buf->b_data; 6177 } else { 6178 ASSERT(zio->io_data != NULL); 6179 /* 6180 * We copy the compressed data from the start of the arc buffer 6181 * (the zio_read will have pulled in only what we need, the 6182 * rest is garbage which we will overwrite at decompression) 6183 * and then decompress back to the ARC data buffer. This way we 6184 * can minimize copying by simply decompressing back over the 6185 * original compressed data (rather than decompressing to an 6186 * aux buffer and then copying back the uncompressed buffer, 6187 * which is likely to be much larger). 6188 */ 6189 uint64_t csize; 6190 void *cdata; 6191 6192 csize = zio->io_size; 6193 cdata = zio_data_buf_alloc(csize); 6194 bcopy(zio->io_data, cdata, csize); 6195 if (zio_decompress_data(c, cdata, zio->io_data, csize, 6196 hdr->b_size) != 0) 6197 zio->io_error = EIO; 6198 zio_data_buf_free(cdata, csize); 6199 } 6200 6201 /* Restore the expected uncompressed IO size. */ 6202 zio->io_orig_size = zio->io_size = hdr->b_size; 6203} 6204 6205/* 6206 * Releases the temporary b_tmp_cdata buffer in an l2arc header structure. 6207 * This buffer serves as a temporary holder of compressed data while 6208 * the buffer entry is being written to an l2arc device. Once that is 6209 * done, we can dispose of it. 6210 */ 6211static void 6212l2arc_release_cdata_buf(arc_buf_hdr_t *hdr) 6213{ 6214 ASSERT(HDR_HAS_L1HDR(hdr)); 6215 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_EMPTY) { 6216 /* 6217 * If the data was compressed, then we've allocated a 6218 * temporary buffer for it, so now we need to release it. 6219 */ 6220 ASSERT(hdr->b_l1hdr.b_tmp_cdata != NULL); 6221 zio_data_buf_free(hdr->b_l1hdr.b_tmp_cdata, 6222 hdr->b_size); 6223 hdr->b_l1hdr.b_tmp_cdata = NULL; 6224 } else { 6225 ASSERT(hdr->b_l1hdr.b_tmp_cdata == NULL); 6226 } 6227} 6228 6229/* 6230 * This thread feeds the L2ARC at regular intervals. This is the beating 6231 * heart of the L2ARC. 6232 */ 6233static void 6234l2arc_feed_thread(void *dummy __unused) 6235{ 6236 callb_cpr_t cpr; 6237 l2arc_dev_t *dev; 6238 spa_t *spa; 6239 uint64_t size, wrote; 6240 clock_t begin, next = ddi_get_lbolt(); 6241 boolean_t headroom_boost = B_FALSE; 6242 6243 CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG); 6244 6245 mutex_enter(&l2arc_feed_thr_lock); 6246 6247 while (l2arc_thread_exit == 0) { 6248 CALLB_CPR_SAFE_BEGIN(&cpr); 6249 (void) cv_timedwait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock, 6250 next - ddi_get_lbolt()); 6251 CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock); 6252 next = ddi_get_lbolt() + hz; 6253 6254 /* 6255 * Quick check for L2ARC devices. 6256 */ 6257 mutex_enter(&l2arc_dev_mtx); 6258 if (l2arc_ndev == 0) { 6259 mutex_exit(&l2arc_dev_mtx); 6260 continue; 6261 } 6262 mutex_exit(&l2arc_dev_mtx); 6263 begin = ddi_get_lbolt(); 6264 6265 /* 6266 * This selects the next l2arc device to write to, and in 6267 * doing so the next spa to feed from: dev->l2ad_spa. This 6268 * will return NULL if there are now no l2arc devices or if 6269 * they are all faulted. 6270 * 6271 * If a device is returned, its spa's config lock is also 6272 * held to prevent device removal. l2arc_dev_get_next() 6273 * will grab and release l2arc_dev_mtx. 6274 */ 6275 if ((dev = l2arc_dev_get_next()) == NULL) 6276 continue; 6277 6278 spa = dev->l2ad_spa; 6279 ASSERT(spa != NULL); 6280 6281 /* 6282 * If the pool is read-only then force the feed thread to 6283 * sleep a little longer. 6284 */ 6285 if (!spa_writeable(spa)) { 6286 next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz; 6287 spa_config_exit(spa, SCL_L2ARC, dev); 6288 continue; 6289 } 6290 6291 /* 6292 * Avoid contributing to memory pressure. 6293 */ 6294 if (arc_reclaim_needed()) { 6295 ARCSTAT_BUMP(arcstat_l2_abort_lowmem); 6296 spa_config_exit(spa, SCL_L2ARC, dev); 6297 continue; 6298 } 6299 6300 ARCSTAT_BUMP(arcstat_l2_feeds); 6301 6302 size = l2arc_write_size(); 6303 6304 /* 6305 * Evict L2ARC buffers that will be overwritten. 6306 */ 6307 l2arc_evict(dev, size, B_FALSE); 6308 6309 /* 6310 * Write ARC buffers. 6311 */ 6312 wrote = l2arc_write_buffers(spa, dev, size, &headroom_boost); 6313 6314 /* 6315 * Calculate interval between writes. 6316 */ 6317 next = l2arc_write_interval(begin, size, wrote); 6318 spa_config_exit(spa, SCL_L2ARC, dev); 6319 } 6320 6321 l2arc_thread_exit = 0; 6322 cv_broadcast(&l2arc_feed_thr_cv); 6323 CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */ 6324 thread_exit(); 6325} 6326 6327boolean_t 6328l2arc_vdev_present(vdev_t *vd) 6329{ 6330 l2arc_dev_t *dev; 6331 6332 mutex_enter(&l2arc_dev_mtx); 6333 for (dev = list_head(l2arc_dev_list); dev != NULL; 6334 dev = list_next(l2arc_dev_list, dev)) { 6335 if (dev->l2ad_vdev == vd) 6336 break; 6337 } 6338 mutex_exit(&l2arc_dev_mtx); 6339 6340 return (dev != NULL); 6341} 6342 6343/* 6344 * Add a vdev for use by the L2ARC. By this point the spa has already 6345 * validated the vdev and opened it. 6346 */ 6347void 6348l2arc_add_vdev(spa_t *spa, vdev_t *vd) 6349{ 6350 l2arc_dev_t *adddev; 6351 6352 ASSERT(!l2arc_vdev_present(vd)); 6353 6354 vdev_ashift_optimize(vd); 6355 6356 /* 6357 * Create a new l2arc device entry. 6358 */ 6359 adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP); 6360 adddev->l2ad_spa = spa; 6361 adddev->l2ad_vdev = vd; 6362 adddev->l2ad_start = VDEV_LABEL_START_SIZE; 6363 adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd); 6364 adddev->l2ad_hand = adddev->l2ad_start; 6365 adddev->l2ad_first = B_TRUE; 6366 adddev->l2ad_writing = B_FALSE; 6367 6368 mutex_init(&adddev->l2ad_mtx, NULL, MUTEX_DEFAULT, NULL); 6369 /* 6370 * This is a list of all ARC buffers that are still valid on the 6371 * device. 6372 */ 6373 list_create(&adddev->l2ad_buflist, sizeof (arc_buf_hdr_t), 6374 offsetof(arc_buf_hdr_t, b_l2hdr.b_l2node)); 6375 6376 vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand); 6377 refcount_create(&adddev->l2ad_alloc); 6378 6379 /* 6380 * Add device to global list 6381 */ 6382 mutex_enter(&l2arc_dev_mtx); 6383 list_insert_head(l2arc_dev_list, adddev); 6384 atomic_inc_64(&l2arc_ndev); 6385 mutex_exit(&l2arc_dev_mtx); 6386} 6387 6388/* 6389 * Remove a vdev from the L2ARC. 6390 */ 6391void 6392l2arc_remove_vdev(vdev_t *vd) 6393{ 6394 l2arc_dev_t *dev, *nextdev, *remdev = NULL; 6395 6396 /* 6397 * Find the device by vdev 6398 */ 6399 mutex_enter(&l2arc_dev_mtx); 6400 for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) { 6401 nextdev = list_next(l2arc_dev_list, dev); 6402 if (vd == dev->l2ad_vdev) { 6403 remdev = dev; 6404 break; 6405 } 6406 } 6407 ASSERT(remdev != NULL); 6408 6409 /* 6410 * Remove device from global list 6411 */ 6412 list_remove(l2arc_dev_list, remdev); 6413 l2arc_dev_last = NULL; /* may have been invalidated */ 6414 atomic_dec_64(&l2arc_ndev); 6415 mutex_exit(&l2arc_dev_mtx); 6416 6417 /* 6418 * Clear all buflists and ARC references. L2ARC device flush. 6419 */ 6420 l2arc_evict(remdev, 0, B_TRUE); 6421 list_destroy(&remdev->l2ad_buflist); 6422 mutex_destroy(&remdev->l2ad_mtx); 6423 refcount_destroy(&remdev->l2ad_alloc); 6424 kmem_free(remdev, sizeof (l2arc_dev_t)); 6425} 6426 6427void 6428l2arc_init(void) 6429{ 6430 l2arc_thread_exit = 0; 6431 l2arc_ndev = 0; 6432 l2arc_writes_sent = 0; 6433 l2arc_writes_done = 0; 6434 6435 mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL); 6436 cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL); 6437 mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL); 6438 mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL); 6439 6440 l2arc_dev_list = &L2ARC_dev_list; 6441 l2arc_free_on_write = &L2ARC_free_on_write; 6442 list_create(l2arc_dev_list, sizeof (l2arc_dev_t), 6443 offsetof(l2arc_dev_t, l2ad_node)); 6444 list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t), 6445 offsetof(l2arc_data_free_t, l2df_list_node)); 6446} 6447 6448void 6449l2arc_fini(void) 6450{ 6451 /* 6452 * This is called from dmu_fini(), which is called from spa_fini(); 6453 * Because of this, we can assume that all l2arc devices have 6454 * already been removed when the pools themselves were removed. 6455 */ 6456 6457 l2arc_do_free_on_write(); 6458 6459 mutex_destroy(&l2arc_feed_thr_lock); 6460 cv_destroy(&l2arc_feed_thr_cv); 6461 mutex_destroy(&l2arc_dev_mtx); 6462 mutex_destroy(&l2arc_free_on_write_mtx); 6463 6464 list_destroy(l2arc_dev_list); 6465 list_destroy(l2arc_free_on_write); 6466} 6467 6468void 6469l2arc_start(void) 6470{ 6471 if (!(spa_mode_global & FWRITE)) 6472 return; 6473 6474 (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0, 6475 TS_RUN, minclsyspri); 6476} 6477 6478void 6479l2arc_stop(void) 6480{ 6481 if (!(spa_mode_global & FWRITE)) 6482 return; 6483 6484 mutex_enter(&l2arc_feed_thr_lock); 6485 cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */ 6486 l2arc_thread_exit = 1; 6487 while (l2arc_thread_exit != 0) 6488 cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock); 6489 mutex_exit(&l2arc_feed_thr_lock); 6490} 6491