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, 2016 by Delphix. All rights reserved. 25 * Copyright (c) 2014 by Saso Kiselkov. All rights reserved. 26 * Copyright 2015 Nexenta Systems, Inc. All rights reserved. 27 */ 28 29/* 30 * DVA-based Adjustable Replacement Cache 31 * 32 * While much of the theory of operation used here is 33 * based on the self-tuning, low overhead replacement cache 34 * presented by Megiddo and Modha at FAST 2003, there are some 35 * significant differences: 36 * 37 * 1. The Megiddo and Modha model assumes any page is evictable. 38 * Pages in its cache cannot be "locked" into memory. This makes 39 * the eviction algorithm simple: evict the last page in the list. 40 * This also make the performance characteristics easy to reason 41 * about. Our cache is not so simple. At any given moment, some 42 * subset of the blocks in the cache are un-evictable because we 43 * have handed out a reference to them. Blocks are only evictable 44 * when there are no external references active. This makes 45 * eviction far more problematic: we choose to evict the evictable 46 * blocks that are the "lowest" in the list. 47 * 48 * There are times when it is not possible to evict the requested 49 * space. In these circumstances we are unable to adjust the cache 50 * size. To prevent the cache growing unbounded at these times we 51 * implement a "cache throttle" that slows the flow of new data 52 * into the cache until we can make space available. 53 * 54 * 2. The Megiddo and Modha model assumes a fixed cache size. 55 * Pages are evicted when the cache is full and there is a cache 56 * miss. Our model has a variable sized cache. It grows with 57 * high use, but also tries to react to memory pressure from the 58 * operating system: decreasing its size when system memory is 59 * tight. 60 * 61 * 3. The Megiddo and Modha model assumes a fixed page size. All 62 * elements of the cache are therefore exactly the same size. So 63 * when adjusting the cache size following a cache miss, its simply 64 * a matter of choosing a single page to evict. In our model, we 65 * have variable sized cache blocks (rangeing from 512 bytes to 66 * 128K bytes). We therefore choose a set of blocks to evict to make 67 * space for a cache miss that approximates as closely as possible 68 * the space used by the new block. 69 * 70 * See also: "ARC: A Self-Tuning, Low Overhead Replacement Cache" 71 * by N. Megiddo & D. Modha, FAST 2003 72 */ 73 74/* 75 * The locking model: 76 * 77 * A new reference to a cache buffer can be obtained in two 78 * ways: 1) via a hash table lookup using the DVA as a key, 79 * or 2) via one of the ARC lists. The arc_read() interface 80 * uses method 1, while the internal arc algorithms for 81 * adjusting the cache use method 2. We therefore provide two 82 * types of locks: 1) the hash table lock array, and 2) the 83 * arc list locks. 84 * 85 * Buffers do not have their own mutexes, rather they rely on the 86 * hash table mutexes for the bulk of their protection (i.e. most 87 * fields in the arc_buf_hdr_t are protected by these mutexes). 88 * 89 * buf_hash_find() returns the appropriate mutex (held) when it 90 * locates the requested buffer in the hash table. It returns 91 * NULL for the mutex if the buffer was not in the table. 92 * 93 * buf_hash_remove() expects the appropriate hash mutex to be 94 * already held before it is invoked. 95 * 96 * Each arc state also has a mutex which is used to protect the 97 * buffer list associated with the state. When attempting to 98 * obtain a hash table lock while holding an arc list lock you 99 * must use: mutex_tryenter() to avoid deadlock. Also note that 100 * the active state mutex must be held before the ghost state mutex. 101 * 102 * Arc buffers may have an associated eviction callback function. 103 * This function will be invoked prior to removing the buffer (e.g. 104 * in arc_do_user_evicts()). Note however that the data associated 105 * with the buffer may be evicted prior to the callback. The callback 106 * must be made with *no locks held* (to prevent deadlock). Additionally, 107 * the users of callbacks must ensure that their private data is 108 * protected from simultaneous callbacks from arc_clear_callback() 109 * and arc_do_user_evicts(). 110 * 111 * Note that the majority of the performance stats are manipulated 112 * with atomic operations. 113 * 114 * The L2ARC uses the l2ad_mtx on each vdev for the following: 115 * 116 * - L2ARC buflist creation 117 * - L2ARC buflist eviction 118 * - L2ARC write completion, which walks L2ARC buflists 119 * - ARC header destruction, as it removes from L2ARC buflists 120 * - ARC header release, as it removes from L2ARC buflists 121 */ 122 123/* 124 * ARC operation: 125 * 126 * Every block that is in the ARC is tracked by an arc_buf_hdr_t structure. 127 * This structure can point either to a block that is still in the cache or to 128 * one that is only accessible in an L2 ARC device, or it can provide 129 * information about a block that was recently evicted. If a block is 130 * only accessible in the L2ARC, then the arc_buf_hdr_t only has enough 131 * information to retrieve it from the L2ARC device. This information is 132 * stored in the l2arc_buf_hdr_t sub-structure of the arc_buf_hdr_t. A block 133 * that is in this state cannot access the data directly. 134 * 135 * Blocks that are actively being referenced or have not been evicted 136 * are cached in the L1ARC. The L1ARC (l1arc_buf_hdr_t) is a structure within 137 * the arc_buf_hdr_t that will point to the data block in memory. A block can 138 * only be read by a consumer if it has an l1arc_buf_hdr_t. The L1ARC 139 * caches data in two ways -- in a list of arc buffers (arc_buf_t) and 140 * also in the arc_buf_hdr_t's private physical data block pointer (b_pdata). 141 * Each arc buffer (arc_buf_t) is being actively accessed by a specific ARC 142 * consumer, and always contains uncompressed data. The ARC will provide 143 * references to this data and will keep it cached until it is no longer in 144 * use. Typically, the arc will try to cache only the L1ARC's physical data 145 * block and will aggressively evict any arc_buf_t that is no longer referenced. 146 * The amount of memory consumed by the arc_buf_t's can be seen via the 147 * "overhead_size" kstat. 148 * 149 * 150 * arc_buf_hdr_t 151 * +-----------+ 152 * | | 153 * | | 154 * | | 155 * +-----------+ 156 * l2arc_buf_hdr_t| | 157 * | | 158 * +-----------+ 159 * l1arc_buf_hdr_t| | 160 * | | arc_buf_t 161 * | b_buf +------------>+---------+ arc_buf_t 162 * | | |b_next +---->+---------+ 163 * | b_pdata +-+ |---------| |b_next +-->NULL 164 * +-----------+ | | | +---------+ 165 * | |b_data +-+ | | 166 * | +---------+ | |b_data +-+ 167 * +->+------+ | +---------+ | 168 * (potentially) | | | | 169 * compressed | | | | 170 * data +------+ | v 171 * +->+------+ +------+ 172 * uncompressed | | | | 173 * data | | | | 174 * +------+ +------+ 175 * 176 * The L1ARC's data pointer, however, may or may not be uncompressed. The 177 * ARC has the ability to store the physical data (b_pdata) associated with 178 * the DVA of the arc_buf_hdr_t. Since the b_pdata is a copy of the on-disk 179 * physical block, it will match its on-disk compression characteristics. 180 * If the block on-disk is compressed, then the physical data block 181 * in the cache will also be compressed and vice-versa. This behavior 182 * can be disabled by setting 'zfs_compressed_arc_enabled' to B_FALSE. When the 183 * compressed ARC functionality is disabled, the b_pdata will point to an 184 * uncompressed version of the on-disk data. 185 * 186 * When a consumer reads a block, the ARC must first look to see if the 187 * arc_buf_hdr_t is cached. If the hdr is cached and already has an arc_buf_t, 188 * then an additional arc_buf_t is allocated and the uncompressed data is 189 * bcopied from the existing arc_buf_t. If the hdr is cached but does not 190 * have an arc_buf_t, then the ARC allocates a new arc_buf_t and decompresses 191 * the b_pdata contents into the arc_buf_t's b_data. If the arc_buf_hdr_t's 192 * b_pdata is not compressed, then the block is shared with the newly 193 * allocated arc_buf_t. This block sharing only occurs with one arc_buf_t 194 * in the arc buffer chain. Sharing the block reduces the memory overhead 195 * required when the hdr is caching uncompressed blocks or the compressed 196 * arc functionality has been disabled via 'zfs_compressed_arc_enabled'. 197 * 198 * The diagram below shows an example of an uncompressed ARC hdr that is 199 * sharing its data with an arc_buf_t: 200 * 201 * arc_buf_hdr_t 202 * +-----------+ 203 * | | 204 * | | 205 * | | 206 * +-----------+ 207 * l2arc_buf_hdr_t| | 208 * | | 209 * +-----------+ 210 * l1arc_buf_hdr_t| | 211 * | | arc_buf_t (shared) 212 * | b_buf +------------>+---------+ arc_buf_t 213 * | | |b_next +---->+---------+ 214 * | b_pdata +-+ |---------| |b_next +-->NULL 215 * +-----------+ | | | +---------+ 216 * | |b_data +-+ | | 217 * | +---------+ | |b_data +-+ 218 * +->+------+ | +---------+ | 219 * | | | | 220 * uncompressed | | | | 221 * data +------+ | | 222 * ^ +->+------+ | 223 * | uncompressed | | | 224 * | data | | | 225 * | +------+ | 226 * +---------------------------------+ 227 * 228 * Writing to the arc requires that the ARC first discard the b_pdata 229 * since the physical block is about to be rewritten. The new data contents 230 * will be contained in the arc_buf_t (uncompressed). As the I/O pipeline 231 * performs the write, it may compress the data before writing it to disk. 232 * The ARC will be called with the transformed data and will bcopy the 233 * transformed on-disk block into a newly allocated b_pdata. 234 * 235 * When the L2ARC is in use, it will also take advantage of the b_pdata. The 236 * L2ARC will always write the contents of b_pdata to the L2ARC. This means 237 * that when compressed arc is enabled that the L2ARC blocks are identical 238 * to the on-disk block in the main data pool. This provides a significant 239 * advantage since the ARC can leverage the bp's checksum when reading from the 240 * L2ARC to determine if the contents are valid. However, if the compressed 241 * arc is disabled, then the L2ARC's block must be transformed to look 242 * like the physical block in the main data pool before comparing the 243 * checksum and determining its validity. 244 */ 245 246#include <sys/spa.h> 247#include <sys/zio.h> 248#include <sys/spa_impl.h> 249#include <sys/zio_compress.h> 250#include <sys/zio_checksum.h> 251#include <sys/zfs_context.h> 252#include <sys/arc.h> 253#include <sys/refcount.h> 254#include <sys/vdev.h> 255#include <sys/vdev_impl.h> 256#include <sys/dsl_pool.h> 257#include <sys/multilist.h> 258#ifdef _KERNEL 259#include <sys/dnlc.h> 260#endif 261#include <sys/callb.h> 262#include <sys/kstat.h> 263#include <sys/trim_map.h> 264#include <zfs_fletcher.h> 265#include <sys/sdt.h> 266 267#include <machine/vmparam.h> 268 269#ifdef illumos 270#ifndef _KERNEL 271/* set with ZFS_DEBUG=watch, to enable watchpoints on frozen buffers */ 272boolean_t arc_watch = B_FALSE; 273int arc_procfd; 274#endif 275#endif /* illumos */ 276 277static kmutex_t arc_reclaim_lock; 278static kcondvar_t arc_reclaim_thread_cv; 279static boolean_t arc_reclaim_thread_exit; 280static kcondvar_t arc_reclaim_waiters_cv; 281 282uint_t arc_reduce_dnlc_percent = 3; 283 284/* 285 * The number of headers to evict in arc_evict_state_impl() before 286 * dropping the sublist lock and evicting from another sublist. A lower 287 * value means we're more likely to evict the "correct" header (i.e. the 288 * oldest header in the arc state), but comes with higher overhead 289 * (i.e. more invocations of arc_evict_state_impl()). 290 */ 291int zfs_arc_evict_batch_limit = 10; 292 293/* 294 * The number of sublists used for each of the arc state lists. If this 295 * is not set to a suitable value by the user, it will be configured to 296 * the number of CPUs on the system in arc_init(). 297 */ 298int zfs_arc_num_sublists_per_state = 0; 299 300/* number of seconds before growing cache again */ 301static int arc_grow_retry = 60; 302 303/* shift of arc_c for calculating overflow limit in arc_get_data_buf */ 304int zfs_arc_overflow_shift = 8; 305 306/* shift of arc_c for calculating both min and max arc_p */ 307static int arc_p_min_shift = 4; 308 309/* log2(fraction of arc to reclaim) */ 310static int arc_shrink_shift = 7; 311 312/* 313 * log2(fraction of ARC which must be free to allow growing). 314 * I.e. If there is less than arc_c >> arc_no_grow_shift free memory, 315 * when reading a new block into the ARC, we will evict an equal-sized block 316 * from the ARC. 317 * 318 * This must be less than arc_shrink_shift, so that when we shrink the ARC, 319 * we will still not allow it to grow. 320 */ 321int arc_no_grow_shift = 5; 322 323 324/* 325 * minimum lifespan of a prefetch block in clock ticks 326 * (initialized in arc_init()) 327 */ 328static int arc_min_prefetch_lifespan; 329 330/* 331 * If this percent of memory is free, don't throttle. 332 */ 333int arc_lotsfree_percent = 10; 334 335static int arc_dead; 336extern boolean_t zfs_prefetch_disable; 337 338/* 339 * The arc has filled available memory and has now warmed up. 340 */ 341static boolean_t arc_warm; 342 343/* 344 * These tunables are for performance analysis. 345 */ 346uint64_t zfs_arc_max; 347uint64_t zfs_arc_min; 348uint64_t zfs_arc_meta_limit = 0; 349uint64_t zfs_arc_meta_min = 0; 350int zfs_arc_grow_retry = 0; 351int zfs_arc_shrink_shift = 0; 352int zfs_arc_p_min_shift = 0; 353uint64_t zfs_arc_average_blocksize = 8 * 1024; /* 8KB */ 354u_int zfs_arc_free_target = 0; 355 356/* Absolute min for arc min / max is 16MB. */ 357static uint64_t arc_abs_min = 16 << 20; 358 359boolean_t zfs_compressed_arc_enabled = B_TRUE; 360 361static int sysctl_vfs_zfs_arc_free_target(SYSCTL_HANDLER_ARGS); 362static int sysctl_vfs_zfs_arc_meta_limit(SYSCTL_HANDLER_ARGS); 363static int sysctl_vfs_zfs_arc_max(SYSCTL_HANDLER_ARGS); 364static int sysctl_vfs_zfs_arc_min(SYSCTL_HANDLER_ARGS); 365 366#if defined(__FreeBSD__) && defined(_KERNEL) 367static void 368arc_free_target_init(void *unused __unused) 369{ 370 371 zfs_arc_free_target = vm_pageout_wakeup_thresh; 372} 373SYSINIT(arc_free_target_init, SI_SUB_KTHREAD_PAGE, SI_ORDER_ANY, 374 arc_free_target_init, NULL); 375 376TUNABLE_QUAD("vfs.zfs.arc_max", &zfs_arc_max); 377TUNABLE_QUAD("vfs.zfs.arc_min", &zfs_arc_min); 378TUNABLE_QUAD("vfs.zfs.arc_meta_limit", &zfs_arc_meta_limit); 379TUNABLE_QUAD("vfs.zfs.arc_meta_min", &zfs_arc_meta_min); 380TUNABLE_QUAD("vfs.zfs.arc_average_blocksize", &zfs_arc_average_blocksize); 381TUNABLE_INT("vfs.zfs.arc_shrink_shift", &zfs_arc_shrink_shift); 382SYSCTL_DECL(_vfs_zfs); 383SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_max, CTLTYPE_U64 | CTLFLAG_RWTUN, 384 0, sizeof(uint64_t), sysctl_vfs_zfs_arc_max, "QU", "Maximum ARC size"); 385SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_min, CTLTYPE_U64 | CTLFLAG_RWTUN, 386 0, sizeof(uint64_t), sysctl_vfs_zfs_arc_min, "QU", "Minimum ARC size"); 387SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_average_blocksize, CTLFLAG_RDTUN, 388 &zfs_arc_average_blocksize, 0, 389 "ARC average blocksize"); 390SYSCTL_INT(_vfs_zfs, OID_AUTO, arc_shrink_shift, CTLFLAG_RW, 391 &arc_shrink_shift, 0, 392 "log2(fraction of arc to reclaim)"); 393SYSCTL_INT(_vfs_zfs, OID_AUTO, compressed_arc_enabled, CTLFLAG_RDTUN, 394 &zfs_compressed_arc_enabled, 0, "Enable compressed ARC"); 395 396/* 397 * We don't have a tunable for arc_free_target due to the dependency on 398 * pagedaemon initialisation. 399 */ 400SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_free_target, 401 CTLTYPE_UINT | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, sizeof(u_int), 402 sysctl_vfs_zfs_arc_free_target, "IU", 403 "Desired number of free pages below which ARC triggers reclaim"); 404 405static int 406sysctl_vfs_zfs_arc_free_target(SYSCTL_HANDLER_ARGS) 407{ 408 u_int val; 409 int err; 410 411 val = zfs_arc_free_target; 412 err = sysctl_handle_int(oidp, &val, 0, req); 413 if (err != 0 || req->newptr == NULL) 414 return (err); 415 416 if (val < minfree) 417 return (EINVAL); 418 if (val > cnt.v_page_count) 419 return (EINVAL); 420 421 zfs_arc_free_target = val; 422 423 return (0); 424} 425 426/* 427 * Must be declared here, before the definition of corresponding kstat 428 * macro which uses the same names will confuse the compiler. 429 */ 430SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_meta_limit, 431 CTLTYPE_U64 | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, sizeof(uint64_t), 432 sysctl_vfs_zfs_arc_meta_limit, "QU", 433 "ARC metadata limit"); 434#endif 435 436/* 437 * Note that buffers can be in one of 6 states: 438 * ARC_anon - anonymous (discussed below) 439 * ARC_mru - recently used, currently cached 440 * ARC_mru_ghost - recentely used, no longer in cache 441 * ARC_mfu - frequently used, currently cached 442 * ARC_mfu_ghost - frequently used, no longer in cache 443 * ARC_l2c_only - exists in L2ARC but not other states 444 * When there are no active references to the buffer, they are 445 * are linked onto a list in one of these arc states. These are 446 * the only buffers that can be evicted or deleted. Within each 447 * state there are multiple lists, one for meta-data and one for 448 * non-meta-data. Meta-data (indirect blocks, blocks of dnodes, 449 * etc.) is tracked separately so that it can be managed more 450 * explicitly: favored over data, limited explicitly. 451 * 452 * Anonymous buffers are buffers that are not associated with 453 * a DVA. These are buffers that hold dirty block copies 454 * before they are written to stable storage. By definition, 455 * they are "ref'd" and are considered part of arc_mru 456 * that cannot be freed. Generally, they will aquire a DVA 457 * as they are written and migrate onto the arc_mru list. 458 * 459 * The ARC_l2c_only state is for buffers that are in the second 460 * level ARC but no longer in any of the ARC_m* lists. The second 461 * level ARC itself may also contain buffers that are in any of 462 * the ARC_m* states - meaning that a buffer can exist in two 463 * places. The reason for the ARC_l2c_only state is to keep the 464 * buffer header in the hash table, so that reads that hit the 465 * second level ARC benefit from these fast lookups. 466 */ 467 468typedef struct arc_state { 469 /* 470 * list of evictable buffers 471 */ 472 multilist_t arcs_list[ARC_BUFC_NUMTYPES]; 473 /* 474 * total amount of evictable data in this state 475 */ 476 refcount_t arcs_esize[ARC_BUFC_NUMTYPES]; 477 /* 478 * total amount of data in this state; this includes: evictable, 479 * non-evictable, ARC_BUFC_DATA, and ARC_BUFC_METADATA. 480 */ 481 refcount_t arcs_size; 482} arc_state_t; 483 484/* The 6 states: */ 485static arc_state_t ARC_anon; 486static arc_state_t ARC_mru; 487static arc_state_t ARC_mru_ghost; 488static arc_state_t ARC_mfu; 489static arc_state_t ARC_mfu_ghost; 490static arc_state_t ARC_l2c_only; 491 492typedef struct arc_stats { 493 kstat_named_t arcstat_hits; 494 kstat_named_t arcstat_misses; 495 kstat_named_t arcstat_demand_data_hits; 496 kstat_named_t arcstat_demand_data_misses; 497 kstat_named_t arcstat_demand_metadata_hits; 498 kstat_named_t arcstat_demand_metadata_misses; 499 kstat_named_t arcstat_prefetch_data_hits; 500 kstat_named_t arcstat_prefetch_data_misses; 501 kstat_named_t arcstat_prefetch_metadata_hits; 502 kstat_named_t arcstat_prefetch_metadata_misses; 503 kstat_named_t arcstat_mru_hits; 504 kstat_named_t arcstat_mru_ghost_hits; 505 kstat_named_t arcstat_mfu_hits; 506 kstat_named_t arcstat_mfu_ghost_hits; 507 kstat_named_t arcstat_allocated; 508 kstat_named_t arcstat_deleted; 509 /* 510 * Number of buffers that could not be evicted because the hash lock 511 * was held by another thread. The lock may not necessarily be held 512 * by something using the same buffer, since hash locks are shared 513 * by multiple buffers. 514 */ 515 kstat_named_t arcstat_mutex_miss; 516 /* 517 * Number of buffers skipped because they have I/O in progress, are 518 * indrect prefetch buffers that have not lived long enough, or are 519 * not from the spa we're trying to evict from. 520 */ 521 kstat_named_t arcstat_evict_skip; 522 /* 523 * Number of times arc_evict_state() was unable to evict enough 524 * buffers to reach it's target amount. 525 */ 526 kstat_named_t arcstat_evict_not_enough; 527 kstat_named_t arcstat_evict_l2_cached; 528 kstat_named_t arcstat_evict_l2_eligible; 529 kstat_named_t arcstat_evict_l2_ineligible; 530 kstat_named_t arcstat_evict_l2_skip; 531 kstat_named_t arcstat_hash_elements; 532 kstat_named_t arcstat_hash_elements_max; 533 kstat_named_t arcstat_hash_collisions; 534 kstat_named_t arcstat_hash_chains; 535 kstat_named_t arcstat_hash_chain_max; 536 kstat_named_t arcstat_p; 537 kstat_named_t arcstat_c; 538 kstat_named_t arcstat_c_min; 539 kstat_named_t arcstat_c_max; 540 kstat_named_t arcstat_size; 541 /* 542 * Number of compressed bytes stored in the arc_buf_hdr_t's b_pdata. 543 * Note that the compressed bytes may match the uncompressed bytes 544 * if the block is either not compressed or compressed arc is disabled. 545 */ 546 kstat_named_t arcstat_compressed_size; 547 /* 548 * Uncompressed size of the data stored in b_pdata. If compressed 549 * arc is disabled then this value will be identical to the stat 550 * above. 551 */ 552 kstat_named_t arcstat_uncompressed_size; 553 /* 554 * Number of bytes stored in all the arc_buf_t's. This is classified 555 * as "overhead" since this data is typically short-lived and will 556 * be evicted from the arc when it becomes unreferenced unless the 557 * zfs_keep_uncompressed_metadata or zfs_keep_uncompressed_level 558 * values have been set (see comment in dbuf.c for more information). 559 */ 560 kstat_named_t arcstat_overhead_size; 561 /* 562 * Number of bytes consumed by internal ARC structures necessary 563 * for tracking purposes; these structures are not actually 564 * backed by ARC buffers. This includes arc_buf_hdr_t structures 565 * (allocated via arc_buf_hdr_t_full and arc_buf_hdr_t_l2only 566 * caches), and arc_buf_t structures (allocated via arc_buf_t 567 * cache). 568 */ 569 kstat_named_t arcstat_hdr_size; 570 /* 571 * Number of bytes consumed by ARC buffers of type equal to 572 * ARC_BUFC_DATA. This is generally consumed by buffers backing 573 * on disk user data (e.g. plain file contents). 574 */ 575 kstat_named_t arcstat_data_size; 576 /* 577 * Number of bytes consumed by ARC buffers of type equal to 578 * ARC_BUFC_METADATA. This is generally consumed by buffers 579 * backing on disk data that is used for internal ZFS 580 * structures (e.g. ZAP, dnode, indirect blocks, etc). 581 */ 582 kstat_named_t arcstat_metadata_size; 583 /* 584 * Number of bytes consumed by various buffers and structures 585 * not actually backed with ARC buffers. This includes bonus 586 * buffers (allocated directly via zio_buf_* functions), 587 * dmu_buf_impl_t structures (allocated via dmu_buf_impl_t 588 * cache), and dnode_t structures (allocated via dnode_t cache). 589 */ 590 kstat_named_t arcstat_other_size; 591 /* 592 * Total number of bytes consumed by ARC buffers residing in the 593 * arc_anon state. This includes *all* buffers in the arc_anon 594 * state; e.g. data, metadata, evictable, and unevictable buffers 595 * are all included in this value. 596 */ 597 kstat_named_t arcstat_anon_size; 598 /* 599 * Number of bytes consumed by ARC buffers that meet the 600 * following criteria: backing buffers of type ARC_BUFC_DATA, 601 * residing in the arc_anon state, and are eligible for eviction 602 * (e.g. have no outstanding holds on the buffer). 603 */ 604 kstat_named_t arcstat_anon_evictable_data; 605 /* 606 * Number of bytes consumed by ARC buffers that meet the 607 * following criteria: backing buffers of type ARC_BUFC_METADATA, 608 * residing in the arc_anon state, and are eligible for eviction 609 * (e.g. have no outstanding holds on the buffer). 610 */ 611 kstat_named_t arcstat_anon_evictable_metadata; 612 /* 613 * Total number of bytes consumed by ARC buffers residing in the 614 * arc_mru state. This includes *all* buffers in the arc_mru 615 * state; e.g. data, metadata, evictable, and unevictable buffers 616 * are all included in this value. 617 */ 618 kstat_named_t arcstat_mru_size; 619 /* 620 * Number of bytes consumed by ARC buffers that meet the 621 * following criteria: backing buffers of type ARC_BUFC_DATA, 622 * residing in the arc_mru state, and are eligible for eviction 623 * (e.g. have no outstanding holds on the buffer). 624 */ 625 kstat_named_t arcstat_mru_evictable_data; 626 /* 627 * Number of bytes consumed by ARC buffers that meet the 628 * following criteria: backing buffers of type ARC_BUFC_METADATA, 629 * residing in the arc_mru state, and are eligible for eviction 630 * (e.g. have no outstanding holds on the buffer). 631 */ 632 kstat_named_t arcstat_mru_evictable_metadata; 633 /* 634 * Total number of bytes that *would have been* consumed by ARC 635 * buffers in the arc_mru_ghost state. The key thing to note 636 * here, is the fact that this size doesn't actually indicate 637 * RAM consumption. The ghost lists only consist of headers and 638 * don't actually have ARC buffers linked off of these headers. 639 * Thus, *if* the headers had associated ARC buffers, these 640 * buffers *would have* consumed this number of bytes. 641 */ 642 kstat_named_t arcstat_mru_ghost_size; 643 /* 644 * Number of bytes that *would have been* consumed by ARC 645 * buffers that are eligible for eviction, of type 646 * ARC_BUFC_DATA, and linked off the arc_mru_ghost state. 647 */ 648 kstat_named_t arcstat_mru_ghost_evictable_data; 649 /* 650 * Number of bytes that *would have been* consumed by ARC 651 * buffers that are eligible for eviction, of type 652 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state. 653 */ 654 kstat_named_t arcstat_mru_ghost_evictable_metadata; 655 /* 656 * Total number of bytes consumed by ARC buffers residing in the 657 * arc_mfu state. This includes *all* buffers in the arc_mfu 658 * state; e.g. data, metadata, evictable, and unevictable buffers 659 * are all included in this value. 660 */ 661 kstat_named_t arcstat_mfu_size; 662 /* 663 * Number of bytes consumed by ARC buffers that are eligible for 664 * eviction, of type ARC_BUFC_DATA, and reside in the arc_mfu 665 * state. 666 */ 667 kstat_named_t arcstat_mfu_evictable_data; 668 /* 669 * Number of bytes consumed by ARC buffers that are eligible for 670 * eviction, of type ARC_BUFC_METADATA, and reside in the 671 * arc_mfu state. 672 */ 673 kstat_named_t arcstat_mfu_evictable_metadata; 674 /* 675 * Total number of bytes that *would have been* consumed by ARC 676 * buffers in the arc_mfu_ghost state. See the comment above 677 * arcstat_mru_ghost_size for more details. 678 */ 679 kstat_named_t arcstat_mfu_ghost_size; 680 /* 681 * Number of bytes that *would have been* consumed by ARC 682 * buffers that are eligible for eviction, of type 683 * ARC_BUFC_DATA, and linked off the arc_mfu_ghost state. 684 */ 685 kstat_named_t arcstat_mfu_ghost_evictable_data; 686 /* 687 * Number of bytes that *would have been* consumed by ARC 688 * buffers that are eligible for eviction, of type 689 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state. 690 */ 691 kstat_named_t arcstat_mfu_ghost_evictable_metadata; 692 kstat_named_t arcstat_l2_hits; 693 kstat_named_t arcstat_l2_misses; 694 kstat_named_t arcstat_l2_feeds; 695 kstat_named_t arcstat_l2_rw_clash; 696 kstat_named_t arcstat_l2_read_bytes; 697 kstat_named_t arcstat_l2_write_bytes; 698 kstat_named_t arcstat_l2_writes_sent; 699 kstat_named_t arcstat_l2_writes_done; 700 kstat_named_t arcstat_l2_writes_error; 701 kstat_named_t arcstat_l2_writes_lock_retry; 702 kstat_named_t arcstat_l2_evict_lock_retry; 703 kstat_named_t arcstat_l2_evict_reading; 704 kstat_named_t arcstat_l2_evict_l1cached; 705 kstat_named_t arcstat_l2_free_on_write; 706 kstat_named_t arcstat_l2_abort_lowmem; 707 kstat_named_t arcstat_l2_cksum_bad; 708 kstat_named_t arcstat_l2_io_error; 709 kstat_named_t arcstat_l2_size; 710 kstat_named_t arcstat_l2_asize; 711 kstat_named_t arcstat_l2_hdr_size; 712 kstat_named_t arcstat_l2_write_trylock_fail; 713 kstat_named_t arcstat_l2_write_passed_headroom; 714 kstat_named_t arcstat_l2_write_spa_mismatch; 715 kstat_named_t arcstat_l2_write_in_l2; 716 kstat_named_t arcstat_l2_write_hdr_io_in_progress; 717 kstat_named_t arcstat_l2_write_not_cacheable; 718 kstat_named_t arcstat_l2_write_full; 719 kstat_named_t arcstat_l2_write_buffer_iter; 720 kstat_named_t arcstat_l2_write_pios; 721 kstat_named_t arcstat_l2_write_buffer_bytes_scanned; 722 kstat_named_t arcstat_l2_write_buffer_list_iter; 723 kstat_named_t arcstat_l2_write_buffer_list_null_iter; 724 kstat_named_t arcstat_memory_throttle_count; 725 kstat_named_t arcstat_meta_used; 726 kstat_named_t arcstat_meta_limit; 727 kstat_named_t arcstat_meta_max; 728 kstat_named_t arcstat_meta_min; 729 kstat_named_t arcstat_sync_wait_for_async; 730 kstat_named_t arcstat_demand_hit_predictive_prefetch; 731} arc_stats_t; 732 733static arc_stats_t arc_stats = { 734 { "hits", KSTAT_DATA_UINT64 }, 735 { "misses", KSTAT_DATA_UINT64 }, 736 { "demand_data_hits", KSTAT_DATA_UINT64 }, 737 { "demand_data_misses", KSTAT_DATA_UINT64 }, 738 { "demand_metadata_hits", KSTAT_DATA_UINT64 }, 739 { "demand_metadata_misses", KSTAT_DATA_UINT64 }, 740 { "prefetch_data_hits", KSTAT_DATA_UINT64 }, 741 { "prefetch_data_misses", KSTAT_DATA_UINT64 }, 742 { "prefetch_metadata_hits", KSTAT_DATA_UINT64 }, 743 { "prefetch_metadata_misses", KSTAT_DATA_UINT64 }, 744 { "mru_hits", KSTAT_DATA_UINT64 }, 745 { "mru_ghost_hits", KSTAT_DATA_UINT64 }, 746 { "mfu_hits", KSTAT_DATA_UINT64 }, 747 { "mfu_ghost_hits", KSTAT_DATA_UINT64 }, 748 { "allocated", KSTAT_DATA_UINT64 }, 749 { "deleted", KSTAT_DATA_UINT64 }, 750 { "mutex_miss", KSTAT_DATA_UINT64 }, 751 { "evict_skip", KSTAT_DATA_UINT64 }, 752 { "evict_not_enough", KSTAT_DATA_UINT64 }, 753 { "evict_l2_cached", KSTAT_DATA_UINT64 }, 754 { "evict_l2_eligible", KSTAT_DATA_UINT64 }, 755 { "evict_l2_ineligible", KSTAT_DATA_UINT64 }, 756 { "evict_l2_skip", KSTAT_DATA_UINT64 }, 757 { "hash_elements", KSTAT_DATA_UINT64 }, 758 { "hash_elements_max", KSTAT_DATA_UINT64 }, 759 { "hash_collisions", KSTAT_DATA_UINT64 }, 760 { "hash_chains", KSTAT_DATA_UINT64 }, 761 { "hash_chain_max", KSTAT_DATA_UINT64 }, 762 { "p", KSTAT_DATA_UINT64 }, 763 { "c", KSTAT_DATA_UINT64 }, 764 { "c_min", KSTAT_DATA_UINT64 }, 765 { "c_max", KSTAT_DATA_UINT64 }, 766 { "size", KSTAT_DATA_UINT64 }, 767 { "compressed_size", KSTAT_DATA_UINT64 }, 768 { "uncompressed_size", KSTAT_DATA_UINT64 }, 769 { "overhead_size", KSTAT_DATA_UINT64 }, 770 { "hdr_size", KSTAT_DATA_UINT64 }, 771 { "data_size", KSTAT_DATA_UINT64 }, 772 { "metadata_size", KSTAT_DATA_UINT64 }, 773 { "other_size", KSTAT_DATA_UINT64 }, 774 { "anon_size", KSTAT_DATA_UINT64 }, 775 { "anon_evictable_data", KSTAT_DATA_UINT64 }, 776 { "anon_evictable_metadata", KSTAT_DATA_UINT64 }, 777 { "mru_size", KSTAT_DATA_UINT64 }, 778 { "mru_evictable_data", KSTAT_DATA_UINT64 }, 779 { "mru_evictable_metadata", KSTAT_DATA_UINT64 }, 780 { "mru_ghost_size", KSTAT_DATA_UINT64 }, 781 { "mru_ghost_evictable_data", KSTAT_DATA_UINT64 }, 782 { "mru_ghost_evictable_metadata", KSTAT_DATA_UINT64 }, 783 { "mfu_size", KSTAT_DATA_UINT64 }, 784 { "mfu_evictable_data", KSTAT_DATA_UINT64 }, 785 { "mfu_evictable_metadata", KSTAT_DATA_UINT64 }, 786 { "mfu_ghost_size", KSTAT_DATA_UINT64 }, 787 { "mfu_ghost_evictable_data", KSTAT_DATA_UINT64 }, 788 { "mfu_ghost_evictable_metadata", KSTAT_DATA_UINT64 }, 789 { "l2_hits", KSTAT_DATA_UINT64 }, 790 { "l2_misses", KSTAT_DATA_UINT64 }, 791 { "l2_feeds", KSTAT_DATA_UINT64 }, 792 { "l2_rw_clash", KSTAT_DATA_UINT64 }, 793 { "l2_read_bytes", KSTAT_DATA_UINT64 }, 794 { "l2_write_bytes", KSTAT_DATA_UINT64 }, 795 { "l2_writes_sent", KSTAT_DATA_UINT64 }, 796 { "l2_writes_done", KSTAT_DATA_UINT64 }, 797 { "l2_writes_error", KSTAT_DATA_UINT64 }, 798 { "l2_writes_lock_retry", KSTAT_DATA_UINT64 }, 799 { "l2_evict_lock_retry", KSTAT_DATA_UINT64 }, 800 { "l2_evict_reading", KSTAT_DATA_UINT64 }, 801 { "l2_evict_l1cached", KSTAT_DATA_UINT64 }, 802 { "l2_free_on_write", KSTAT_DATA_UINT64 }, 803 { "l2_abort_lowmem", KSTAT_DATA_UINT64 }, 804 { "l2_cksum_bad", KSTAT_DATA_UINT64 }, 805 { "l2_io_error", KSTAT_DATA_UINT64 }, 806 { "l2_size", KSTAT_DATA_UINT64 }, 807 { "l2_asize", KSTAT_DATA_UINT64 }, 808 { "l2_hdr_size", KSTAT_DATA_UINT64 }, 809 { "l2_write_trylock_fail", KSTAT_DATA_UINT64 }, 810 { "l2_write_passed_headroom", KSTAT_DATA_UINT64 }, 811 { "l2_write_spa_mismatch", KSTAT_DATA_UINT64 }, 812 { "l2_write_in_l2", KSTAT_DATA_UINT64 }, 813 { "l2_write_io_in_progress", KSTAT_DATA_UINT64 }, 814 { "l2_write_not_cacheable", KSTAT_DATA_UINT64 }, 815 { "l2_write_full", KSTAT_DATA_UINT64 }, 816 { "l2_write_buffer_iter", KSTAT_DATA_UINT64 }, 817 { "l2_write_pios", KSTAT_DATA_UINT64 }, 818 { "l2_write_buffer_bytes_scanned", KSTAT_DATA_UINT64 }, 819 { "l2_write_buffer_list_iter", KSTAT_DATA_UINT64 }, 820 { "l2_write_buffer_list_null_iter", KSTAT_DATA_UINT64 }, 821 { "memory_throttle_count", KSTAT_DATA_UINT64 }, 822 { "arc_meta_used", KSTAT_DATA_UINT64 }, 823 { "arc_meta_limit", KSTAT_DATA_UINT64 }, 824 { "arc_meta_max", KSTAT_DATA_UINT64 }, 825 { "arc_meta_min", KSTAT_DATA_UINT64 }, 826 { "sync_wait_for_async", KSTAT_DATA_UINT64 }, 827 { "demand_hit_predictive_prefetch", KSTAT_DATA_UINT64 }, 828}; 829 830#define ARCSTAT(stat) (arc_stats.stat.value.ui64) 831 832#define ARCSTAT_INCR(stat, val) \ 833 atomic_add_64(&arc_stats.stat.value.ui64, (val)) 834 835#define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1) 836#define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1) 837 838#define ARCSTAT_MAX(stat, val) { \ 839 uint64_t m; \ 840 while ((val) > (m = arc_stats.stat.value.ui64) && \ 841 (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \ 842 continue; \ 843} 844 845#define ARCSTAT_MAXSTAT(stat) \ 846 ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64) 847 848/* 849 * We define a macro to allow ARC hits/misses to be easily broken down by 850 * two separate conditions, giving a total of four different subtypes for 851 * each of hits and misses (so eight statistics total). 852 */ 853#define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \ 854 if (cond1) { \ 855 if (cond2) { \ 856 ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \ 857 } else { \ 858 ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \ 859 } \ 860 } else { \ 861 if (cond2) { \ 862 ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \ 863 } else { \ 864 ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\ 865 } \ 866 } 867 868kstat_t *arc_ksp; 869static arc_state_t *arc_anon; 870static arc_state_t *arc_mru; 871static arc_state_t *arc_mru_ghost; 872static arc_state_t *arc_mfu; 873static arc_state_t *arc_mfu_ghost; 874static arc_state_t *arc_l2c_only; 875 876/* 877 * There are several ARC variables that are critical to export as kstats -- 878 * but we don't want to have to grovel around in the kstat whenever we wish to 879 * manipulate them. For these variables, we therefore define them to be in 880 * terms of the statistic variable. This assures that we are not introducing 881 * the possibility of inconsistency by having shadow copies of the variables, 882 * while still allowing the code to be readable. 883 */ 884#define arc_size ARCSTAT(arcstat_size) /* actual total arc size */ 885#define arc_p ARCSTAT(arcstat_p) /* target size of MRU */ 886#define arc_c ARCSTAT(arcstat_c) /* target size of cache */ 887#define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */ 888#define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */ 889#define arc_meta_limit ARCSTAT(arcstat_meta_limit) /* max size for metadata */ 890#define arc_meta_min ARCSTAT(arcstat_meta_min) /* min size for metadata */ 891#define arc_meta_used ARCSTAT(arcstat_meta_used) /* size of metadata */ 892#define arc_meta_max ARCSTAT(arcstat_meta_max) /* max size of metadata */ 893 894/* compressed size of entire arc */ 895#define arc_compressed_size ARCSTAT(arcstat_compressed_size) 896/* uncompressed size of entire arc */ 897#define arc_uncompressed_size ARCSTAT(arcstat_uncompressed_size) 898/* number of bytes in the arc from arc_buf_t's */ 899#define arc_overhead_size ARCSTAT(arcstat_overhead_size) 900 901static int arc_no_grow; /* Don't try to grow cache size */ 902static uint64_t arc_tempreserve; 903static uint64_t arc_loaned_bytes; 904 905typedef struct arc_callback arc_callback_t; 906 907struct arc_callback { 908 void *acb_private; 909 arc_done_func_t *acb_done; 910 arc_buf_t *acb_buf; 911 zio_t *acb_zio_dummy; 912 arc_callback_t *acb_next; 913}; 914 915typedef struct arc_write_callback arc_write_callback_t; 916 917struct arc_write_callback { 918 void *awcb_private; 919 arc_done_func_t *awcb_ready; 920 arc_done_func_t *awcb_children_ready; 921 arc_done_func_t *awcb_physdone; 922 arc_done_func_t *awcb_done; 923 arc_buf_t *awcb_buf; 924}; 925 926/* 927 * ARC buffers are separated into multiple structs as a memory saving measure: 928 * - Common fields struct, always defined, and embedded within it: 929 * - L2-only fields, always allocated but undefined when not in L2ARC 930 * - L1-only fields, only allocated when in L1ARC 931 * 932 * Buffer in L1 Buffer only in L2 933 * +------------------------+ +------------------------+ 934 * | arc_buf_hdr_t | | arc_buf_hdr_t | 935 * | | | | 936 * | | | | 937 * | | | | 938 * +------------------------+ +------------------------+ 939 * | l2arc_buf_hdr_t | | l2arc_buf_hdr_t | 940 * | (undefined if L1-only) | | | 941 * +------------------------+ +------------------------+ 942 * | l1arc_buf_hdr_t | 943 * | | 944 * | | 945 * | | 946 * | | 947 * +------------------------+ 948 * 949 * Because it's possible for the L2ARC to become extremely large, we can wind 950 * up eating a lot of memory in L2ARC buffer headers, so the size of a header 951 * is minimized by only allocating the fields necessary for an L1-cached buffer 952 * when a header is actually in the L1 cache. The sub-headers (l1arc_buf_hdr and 953 * l2arc_buf_hdr) are embedded rather than allocated separately to save a couple 954 * words in pointers. arc_hdr_realloc() is used to switch a header between 955 * these two allocation states. 956 */ 957typedef struct l1arc_buf_hdr { 958 kmutex_t b_freeze_lock; 959 zio_cksum_t *b_freeze_cksum; 960#ifdef ZFS_DEBUG 961 /* 962 * used for debugging wtih kmem_flags - by allocating and freeing 963 * b_thawed when the buffer is thawed, we get a record of the stack 964 * trace that thawed it. 965 */ 966 void *b_thawed; 967#endif 968 969 arc_buf_t *b_buf; 970 uint32_t b_bufcnt; 971 /* for waiting on writes to complete */ 972 kcondvar_t b_cv; 973 uint8_t b_byteswap; 974 975 /* protected by arc state mutex */ 976 arc_state_t *b_state; 977 multilist_node_t b_arc_node; 978 979 /* updated atomically */ 980 clock_t b_arc_access; 981 982 /* self protecting */ 983 refcount_t b_refcnt; 984 985 arc_callback_t *b_acb; 986 void *b_pdata; 987} l1arc_buf_hdr_t; 988 989typedef struct l2arc_dev l2arc_dev_t; 990 991typedef struct l2arc_buf_hdr { 992 /* protected by arc_buf_hdr mutex */ 993 l2arc_dev_t *b_dev; /* L2ARC device */ 994 uint64_t b_daddr; /* disk address, offset byte */ 995 996 list_node_t b_l2node; 997} l2arc_buf_hdr_t; 998 999struct arc_buf_hdr { 1000 /* protected by hash lock */ 1001 dva_t b_dva; 1002 uint64_t b_birth; 1003 1004 arc_buf_contents_t b_type; 1005 arc_buf_hdr_t *b_hash_next; 1006 arc_flags_t b_flags; 1007 1008 /* 1009 * This field stores the size of the data buffer after 1010 * compression, and is set in the arc's zio completion handlers. 1011 * It is in units of SPA_MINBLOCKSIZE (e.g. 1 == 512 bytes). 1012 * 1013 * While the block pointers can store up to 32MB in their psize 1014 * field, we can only store up to 32MB minus 512B. This is due 1015 * to the bp using a bias of 1, whereas we use a bias of 0 (i.e. 1016 * a field of zeros represents 512B in the bp). We can't use a 1017 * bias of 1 since we need to reserve a psize of zero, here, to 1018 * represent holes and embedded blocks. 1019 * 1020 * This isn't a problem in practice, since the maximum size of a 1021 * buffer is limited to 16MB, so we never need to store 32MB in 1022 * this field. Even in the upstream illumos code base, the 1023 * maximum size of a buffer is limited to 16MB. 1024 */ 1025 uint16_t b_psize; 1026 1027 /* 1028 * This field stores the size of the data buffer before 1029 * compression, and cannot change once set. It is in units 1030 * of SPA_MINBLOCKSIZE (e.g. 2 == 1024 bytes) 1031 */ 1032 uint16_t b_lsize; /* immutable */ 1033 uint64_t b_spa; /* immutable */ 1034 1035 /* L2ARC fields. Undefined when not in L2ARC. */ 1036 l2arc_buf_hdr_t b_l2hdr; 1037 /* L1ARC fields. Undefined when in l2arc_only state */ 1038 l1arc_buf_hdr_t b_l1hdr; 1039}; 1040 1041#if defined(__FreeBSD__) && defined(_KERNEL) 1042static int 1043sysctl_vfs_zfs_arc_meta_limit(SYSCTL_HANDLER_ARGS) 1044{ 1045 uint64_t val; 1046 int err; 1047 1048 val = arc_meta_limit; 1049 err = sysctl_handle_64(oidp, &val, 0, req); 1050 if (err != 0 || req->newptr == NULL) 1051 return (err); 1052 1053 if (val <= 0 || val > arc_c_max) 1054 return (EINVAL); 1055 1056 arc_meta_limit = val; 1057 return (0); 1058} 1059 1060static int 1061sysctl_vfs_zfs_arc_max(SYSCTL_HANDLER_ARGS) 1062{ 1063 uint64_t val; 1064 int err; 1065 1066 val = zfs_arc_max; 1067 err = sysctl_handle_64(oidp, &val, 0, req); 1068 if (err != 0 || req->newptr == NULL) 1069 return (err); 1070 1071 if (zfs_arc_max == 0) { 1072 /* Loader tunable so blindly set */ 1073 zfs_arc_max = val; 1074 return (0); 1075 } 1076 1077 if (val < arc_abs_min || val > kmem_size()) 1078 return (EINVAL); 1079 if (val < arc_c_min) 1080 return (EINVAL); 1081 if (zfs_arc_meta_limit > 0 && val < zfs_arc_meta_limit) 1082 return (EINVAL); 1083 1084 arc_c_max = val; 1085 1086 arc_c = arc_c_max; 1087 arc_p = (arc_c >> 1); 1088 1089 if (zfs_arc_meta_limit == 0) { 1090 /* limit meta-data to 1/4 of the arc capacity */ 1091 arc_meta_limit = arc_c_max / 4; 1092 } 1093 1094 /* if kmem_flags are set, lets try to use less memory */ 1095 if (kmem_debugging()) 1096 arc_c = arc_c / 2; 1097 1098 zfs_arc_max = arc_c; 1099 1100 return (0); 1101} 1102 1103static int 1104sysctl_vfs_zfs_arc_min(SYSCTL_HANDLER_ARGS) 1105{ 1106 uint64_t val; 1107 int err; 1108 1109 val = zfs_arc_min; 1110 err = sysctl_handle_64(oidp, &val, 0, req); 1111 if (err != 0 || req->newptr == NULL) 1112 return (err); 1113 1114 if (zfs_arc_min == 0) { 1115 /* Loader tunable so blindly set */ 1116 zfs_arc_min = val; 1117 return (0); 1118 } 1119 1120 if (val < arc_abs_min || val > arc_c_max) 1121 return (EINVAL); 1122 1123 arc_c_min = val; 1124 1125 if (zfs_arc_meta_min == 0) 1126 arc_meta_min = arc_c_min / 2; 1127 1128 if (arc_c < arc_c_min) 1129 arc_c = arc_c_min; 1130 1131 zfs_arc_min = arc_c_min; 1132 1133 return (0); 1134} 1135#endif 1136 1137#define GHOST_STATE(state) \ 1138 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \ 1139 (state) == arc_l2c_only) 1140 1141#define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_FLAG_IN_HASH_TABLE) 1142#define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) 1143#define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_FLAG_IO_ERROR) 1144#define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_FLAG_PREFETCH) 1145#define HDR_COMPRESSION_ENABLED(hdr) \ 1146 ((hdr)->b_flags & ARC_FLAG_COMPRESSED_ARC) 1147 1148#define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_FLAG_L2CACHE) 1149#define HDR_L2_READING(hdr) \ 1150 (((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) && \ 1151 ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR)) 1152#define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITING) 1153#define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_FLAG_L2_EVICTED) 1154#define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITE_HEAD) 1155#define HDR_SHARED_DATA(hdr) ((hdr)->b_flags & ARC_FLAG_SHARED_DATA) 1156 1157#define HDR_ISTYPE_METADATA(hdr) \ 1158 ((hdr)->b_flags & ARC_FLAG_BUFC_METADATA) 1159#define HDR_ISTYPE_DATA(hdr) (!HDR_ISTYPE_METADATA(hdr)) 1160 1161#define HDR_HAS_L1HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L1HDR) 1162#define HDR_HAS_L2HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR) 1163 1164/* For storing compression mode in b_flags */ 1165#define HDR_COMPRESS_OFFSET (highbit64(ARC_FLAG_COMPRESS_0) - 1) 1166 1167#define HDR_GET_COMPRESS(hdr) ((enum zio_compress)BF32_GET((hdr)->b_flags, \ 1168 HDR_COMPRESS_OFFSET, SPA_COMPRESSBITS)) 1169#define HDR_SET_COMPRESS(hdr, cmp) BF32_SET((hdr)->b_flags, \ 1170 HDR_COMPRESS_OFFSET, SPA_COMPRESSBITS, (cmp)); 1171 1172#define ARC_BUF_LAST(buf) ((buf)->b_next == NULL) 1173 1174/* 1175 * Other sizes 1176 */ 1177 1178#define HDR_FULL_SIZE ((int64_t)sizeof (arc_buf_hdr_t)) 1179#define HDR_L2ONLY_SIZE ((int64_t)offsetof(arc_buf_hdr_t, b_l1hdr)) 1180 1181/* 1182 * Hash table routines 1183 */ 1184 1185#define HT_LOCK_PAD CACHE_LINE_SIZE 1186 1187struct ht_lock { 1188 kmutex_t ht_lock; 1189#ifdef _KERNEL 1190 unsigned char pad[(HT_LOCK_PAD - sizeof (kmutex_t))]; 1191#endif 1192}; 1193 1194#define BUF_LOCKS 256 1195typedef struct buf_hash_table { 1196 uint64_t ht_mask; 1197 arc_buf_hdr_t **ht_table; 1198 struct ht_lock ht_locks[BUF_LOCKS] __aligned(CACHE_LINE_SIZE); 1199} buf_hash_table_t; 1200 1201static buf_hash_table_t buf_hash_table; 1202 1203#define BUF_HASH_INDEX(spa, dva, birth) \ 1204 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask) 1205#define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)]) 1206#define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock)) 1207#define HDR_LOCK(hdr) \ 1208 (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth))) 1209 1210uint64_t zfs_crc64_table[256]; 1211 1212/* 1213 * Level 2 ARC 1214 */ 1215 1216#define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */ 1217#define L2ARC_HEADROOM 2 /* num of writes */ 1218/* 1219 * If we discover during ARC scan any buffers to be compressed, we boost 1220 * our headroom for the next scanning cycle by this percentage multiple. 1221 */ 1222#define L2ARC_HEADROOM_BOOST 200 1223#define L2ARC_FEED_SECS 1 /* caching interval secs */ 1224#define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */ 1225 1226#define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent) 1227#define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done) 1228 1229/* L2ARC Performance Tunables */ 1230uint64_t l2arc_write_max = L2ARC_WRITE_SIZE; /* default max write size */ 1231uint64_t l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra write during warmup */ 1232uint64_t l2arc_headroom = L2ARC_HEADROOM; /* number of dev writes */ 1233uint64_t l2arc_headroom_boost = L2ARC_HEADROOM_BOOST; 1234uint64_t l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */ 1235uint64_t l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval milliseconds */ 1236boolean_t l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */ 1237boolean_t l2arc_feed_again = B_TRUE; /* turbo warmup */ 1238boolean_t l2arc_norw = B_TRUE; /* no reads during writes */ 1239 1240SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_max, CTLFLAG_RW, 1241 &l2arc_write_max, 0, "max write size"); 1242SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_boost, CTLFLAG_RW, 1243 &l2arc_write_boost, 0, "extra write during warmup"); 1244SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_headroom, CTLFLAG_RW, 1245 &l2arc_headroom, 0, "number of dev writes"); 1246SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_secs, CTLFLAG_RW, 1247 &l2arc_feed_secs, 0, "interval seconds"); 1248SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_min_ms, CTLFLAG_RW, 1249 &l2arc_feed_min_ms, 0, "min interval milliseconds"); 1250 1251SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_noprefetch, CTLFLAG_RW, 1252 &l2arc_noprefetch, 0, "don't cache prefetch bufs"); 1253SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_feed_again, CTLFLAG_RW, 1254 &l2arc_feed_again, 0, "turbo warmup"); 1255SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_norw, CTLFLAG_RW, 1256 &l2arc_norw, 0, "no reads during writes"); 1257 1258SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_size, CTLFLAG_RD, 1259 &ARC_anon.arcs_size.rc_count, 0, "size of anonymous state"); 1260SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_metadata_esize, CTLFLAG_RD, 1261 &ARC_anon.arcs_esize[ARC_BUFC_METADATA].rc_count, 0, 1262 "size of anonymous state"); 1263SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_data_esize, CTLFLAG_RD, 1264 &ARC_anon.arcs_esize[ARC_BUFC_DATA].rc_count, 0, 1265 "size of anonymous state"); 1266 1267SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_size, CTLFLAG_RD, 1268 &ARC_mru.arcs_size.rc_count, 0, "size of mru state"); 1269SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_metadata_esize, CTLFLAG_RD, 1270 &ARC_mru.arcs_esize[ARC_BUFC_METADATA].rc_count, 0, 1271 "size of metadata in mru state"); 1272SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_data_esize, CTLFLAG_RD, 1273 &ARC_mru.arcs_esize[ARC_BUFC_DATA].rc_count, 0, 1274 "size of data in mru state"); 1275 1276SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_size, CTLFLAG_RD, 1277 &ARC_mru_ghost.arcs_size.rc_count, 0, "size of mru ghost state"); 1278SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_metadata_esize, CTLFLAG_RD, 1279 &ARC_mru_ghost.arcs_esize[ARC_BUFC_METADATA].rc_count, 0, 1280 "size of metadata in mru ghost state"); 1281SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_data_esize, CTLFLAG_RD, 1282 &ARC_mru_ghost.arcs_esize[ARC_BUFC_DATA].rc_count, 0, 1283 "size of data in mru ghost state"); 1284 1285SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_size, CTLFLAG_RD, 1286 &ARC_mfu.arcs_size.rc_count, 0, "size of mfu state"); 1287SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_metadata_esize, CTLFLAG_RD, 1288 &ARC_mfu.arcs_esize[ARC_BUFC_METADATA].rc_count, 0, 1289 "size of metadata in mfu state"); 1290SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_data_esize, CTLFLAG_RD, 1291 &ARC_mfu.arcs_esize[ARC_BUFC_DATA].rc_count, 0, 1292 "size of data in mfu state"); 1293 1294SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_size, CTLFLAG_RD, 1295 &ARC_mfu_ghost.arcs_size.rc_count, 0, "size of mfu ghost state"); 1296SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_metadata_esize, CTLFLAG_RD, 1297 &ARC_mfu_ghost.arcs_esize[ARC_BUFC_METADATA].rc_count, 0, 1298 "size of metadata in mfu ghost state"); 1299SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_data_esize, CTLFLAG_RD, 1300 &ARC_mfu_ghost.arcs_esize[ARC_BUFC_DATA].rc_count, 0, 1301 "size of data in mfu ghost state"); 1302 1303SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2c_only_size, CTLFLAG_RD, 1304 &ARC_l2c_only.arcs_size.rc_count, 0, "size of mru state"); 1305 1306/* 1307 * L2ARC Internals 1308 */ 1309struct l2arc_dev { 1310 vdev_t *l2ad_vdev; /* vdev */ 1311 spa_t *l2ad_spa; /* spa */ 1312 uint64_t l2ad_hand; /* next write location */ 1313 uint64_t l2ad_start; /* first addr on device */ 1314 uint64_t l2ad_end; /* last addr on device */ 1315 boolean_t l2ad_first; /* first sweep through */ 1316 boolean_t l2ad_writing; /* currently writing */ 1317 kmutex_t l2ad_mtx; /* lock for buffer list */ 1318 list_t l2ad_buflist; /* buffer list */ 1319 list_node_t l2ad_node; /* device list node */ 1320 refcount_t l2ad_alloc; /* allocated bytes */ 1321}; 1322 1323static list_t L2ARC_dev_list; /* device list */ 1324static list_t *l2arc_dev_list; /* device list pointer */ 1325static kmutex_t l2arc_dev_mtx; /* device list mutex */ 1326static l2arc_dev_t *l2arc_dev_last; /* last device used */ 1327static list_t L2ARC_free_on_write; /* free after write buf list */ 1328static list_t *l2arc_free_on_write; /* free after write list ptr */ 1329static kmutex_t l2arc_free_on_write_mtx; /* mutex for list */ 1330static uint64_t l2arc_ndev; /* number of devices */ 1331 1332typedef struct l2arc_read_callback { 1333 arc_buf_hdr_t *l2rcb_hdr; /* read buffer */ 1334 blkptr_t l2rcb_bp; /* original blkptr */ 1335 zbookmark_phys_t l2rcb_zb; /* original bookmark */ 1336 int l2rcb_flags; /* original flags */ 1337 void *l2rcb_data; /* temporary buffer */ 1338} l2arc_read_callback_t; 1339 1340typedef struct l2arc_write_callback { 1341 l2arc_dev_t *l2wcb_dev; /* device info */ 1342 arc_buf_hdr_t *l2wcb_head; /* head of write buflist */ 1343} l2arc_write_callback_t; 1344 1345typedef struct l2arc_data_free { 1346 /* protected by l2arc_free_on_write_mtx */ 1347 void *l2df_data; 1348 size_t l2df_size; 1349 arc_buf_contents_t l2df_type; 1350 list_node_t l2df_list_node; 1351} l2arc_data_free_t; 1352 1353static kmutex_t l2arc_feed_thr_lock; 1354static kcondvar_t l2arc_feed_thr_cv; 1355static uint8_t l2arc_thread_exit; 1356 1357static void *arc_get_data_buf(arc_buf_hdr_t *, uint64_t, void *); 1358static void arc_free_data_buf(arc_buf_hdr_t *, void *, uint64_t, void *); 1359static void arc_hdr_free_pdata(arc_buf_hdr_t *hdr); 1360static void arc_hdr_alloc_pdata(arc_buf_hdr_t *); 1361static void arc_access(arc_buf_hdr_t *, kmutex_t *); 1362static boolean_t arc_is_overflowing(); 1363static void arc_buf_watch(arc_buf_t *); 1364 1365static arc_buf_contents_t arc_buf_type(arc_buf_hdr_t *); 1366static uint32_t arc_bufc_to_flags(arc_buf_contents_t); 1367static inline void arc_hdr_set_flags(arc_buf_hdr_t *hdr, arc_flags_t flags); 1368static inline void arc_hdr_clear_flags(arc_buf_hdr_t *hdr, arc_flags_t flags); 1369 1370static boolean_t l2arc_write_eligible(uint64_t, arc_buf_hdr_t *); 1371static void l2arc_read_done(zio_t *); 1372 1373static void 1374l2arc_trim(const arc_buf_hdr_t *hdr) 1375{ 1376 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev; 1377 1378 ASSERT(HDR_HAS_L2HDR(hdr)); 1379 ASSERT(MUTEX_HELD(&dev->l2ad_mtx)); 1380 1381 if (HDR_GET_PSIZE(hdr) != 0) { 1382 trim_map_free(dev->l2ad_vdev, hdr->b_l2hdr.b_daddr, 1383 HDR_GET_PSIZE(hdr), 0); 1384 } 1385} 1386 1387static uint64_t 1388buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth) 1389{ 1390 uint8_t *vdva = (uint8_t *)dva; 1391 uint64_t crc = -1ULL; 1392 int i; 1393 1394 ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY); 1395 1396 for (i = 0; i < sizeof (dva_t); i++) 1397 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ vdva[i]) & 0xFF]; 1398 1399 crc ^= (spa>>8) ^ birth; 1400 1401 return (crc); 1402} 1403 1404#define HDR_EMPTY(hdr) \ 1405 ((hdr)->b_dva.dva_word[0] == 0 && \ 1406 (hdr)->b_dva.dva_word[1] == 0) 1407 1408#define HDR_EQUAL(spa, dva, birth, hdr) \ 1409 ((hdr)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \ 1410 ((hdr)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \ 1411 ((hdr)->b_birth == birth) && ((hdr)->b_spa == spa) 1412 1413static void 1414buf_discard_identity(arc_buf_hdr_t *hdr) 1415{ 1416 hdr->b_dva.dva_word[0] = 0; 1417 hdr->b_dva.dva_word[1] = 0; 1418 hdr->b_birth = 0; 1419} 1420 1421static arc_buf_hdr_t * 1422buf_hash_find(uint64_t spa, const blkptr_t *bp, kmutex_t **lockp) 1423{ 1424 const dva_t *dva = BP_IDENTITY(bp); 1425 uint64_t birth = BP_PHYSICAL_BIRTH(bp); 1426 uint64_t idx = BUF_HASH_INDEX(spa, dva, birth); 1427 kmutex_t *hash_lock = BUF_HASH_LOCK(idx); 1428 arc_buf_hdr_t *hdr; 1429 1430 mutex_enter(hash_lock); 1431 for (hdr = buf_hash_table.ht_table[idx]; hdr != NULL; 1432 hdr = hdr->b_hash_next) { 1433 if (HDR_EQUAL(spa, dva, birth, hdr)) { 1434 *lockp = hash_lock; 1435 return (hdr); 1436 } 1437 } 1438 mutex_exit(hash_lock); 1439 *lockp = NULL; 1440 return (NULL); 1441} 1442 1443/* 1444 * Insert an entry into the hash table. If there is already an element 1445 * equal to elem in the hash table, then the already existing element 1446 * will be returned and the new element will not be inserted. 1447 * Otherwise returns NULL. 1448 * If lockp == NULL, the caller is assumed to already hold the hash lock. 1449 */ 1450static arc_buf_hdr_t * 1451buf_hash_insert(arc_buf_hdr_t *hdr, kmutex_t **lockp) 1452{ 1453 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth); 1454 kmutex_t *hash_lock = BUF_HASH_LOCK(idx); 1455 arc_buf_hdr_t *fhdr; 1456 uint32_t i; 1457 1458 ASSERT(!DVA_IS_EMPTY(&hdr->b_dva)); 1459 ASSERT(hdr->b_birth != 0); 1460 ASSERT(!HDR_IN_HASH_TABLE(hdr)); 1461 1462 if (lockp != NULL) { 1463 *lockp = hash_lock; 1464 mutex_enter(hash_lock); 1465 } else { 1466 ASSERT(MUTEX_HELD(hash_lock)); 1467 } 1468 1469 for (fhdr = buf_hash_table.ht_table[idx], i = 0; fhdr != NULL; 1470 fhdr = fhdr->b_hash_next, i++) { 1471 if (HDR_EQUAL(hdr->b_spa, &hdr->b_dva, hdr->b_birth, fhdr)) 1472 return (fhdr); 1473 } 1474 1475 hdr->b_hash_next = buf_hash_table.ht_table[idx]; 1476 buf_hash_table.ht_table[idx] = hdr; 1477 arc_hdr_set_flags(hdr, ARC_FLAG_IN_HASH_TABLE); 1478 1479 /* collect some hash table performance data */ 1480 if (i > 0) { 1481 ARCSTAT_BUMP(arcstat_hash_collisions); 1482 if (i == 1) 1483 ARCSTAT_BUMP(arcstat_hash_chains); 1484 1485 ARCSTAT_MAX(arcstat_hash_chain_max, i); 1486 } 1487 1488 ARCSTAT_BUMP(arcstat_hash_elements); 1489 ARCSTAT_MAXSTAT(arcstat_hash_elements); 1490 1491 return (NULL); 1492} 1493 1494static void 1495buf_hash_remove(arc_buf_hdr_t *hdr) 1496{ 1497 arc_buf_hdr_t *fhdr, **hdrp; 1498 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth); 1499 1500 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx))); 1501 ASSERT(HDR_IN_HASH_TABLE(hdr)); 1502 1503 hdrp = &buf_hash_table.ht_table[idx]; 1504 while ((fhdr = *hdrp) != hdr) { 1505 ASSERT3P(fhdr, !=, NULL); 1506 hdrp = &fhdr->b_hash_next; 1507 } 1508 *hdrp = hdr->b_hash_next; 1509 hdr->b_hash_next = NULL; 1510 arc_hdr_clear_flags(hdr, ARC_FLAG_IN_HASH_TABLE); 1511 1512 /* collect some hash table performance data */ 1513 ARCSTAT_BUMPDOWN(arcstat_hash_elements); 1514 1515 if (buf_hash_table.ht_table[idx] && 1516 buf_hash_table.ht_table[idx]->b_hash_next == NULL) 1517 ARCSTAT_BUMPDOWN(arcstat_hash_chains); 1518} 1519 1520/* 1521 * Global data structures and functions for the buf kmem cache. 1522 */ 1523static kmem_cache_t *hdr_full_cache; 1524static kmem_cache_t *hdr_l2only_cache; 1525static kmem_cache_t *buf_cache; 1526 1527static void 1528buf_fini(void) 1529{ 1530 int i; 1531 1532 kmem_free(buf_hash_table.ht_table, 1533 (buf_hash_table.ht_mask + 1) * sizeof (void *)); 1534 for (i = 0; i < BUF_LOCKS; i++) 1535 mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock); 1536 kmem_cache_destroy(hdr_full_cache); 1537 kmem_cache_destroy(hdr_l2only_cache); 1538 kmem_cache_destroy(buf_cache); 1539} 1540 1541/* 1542 * Constructor callback - called when the cache is empty 1543 * and a new buf is requested. 1544 */ 1545/* ARGSUSED */ 1546static int 1547hdr_full_cons(void *vbuf, void *unused, int kmflag) 1548{ 1549 arc_buf_hdr_t *hdr = vbuf; 1550 1551 bzero(hdr, HDR_FULL_SIZE); 1552 cv_init(&hdr->b_l1hdr.b_cv, NULL, CV_DEFAULT, NULL); 1553 refcount_create(&hdr->b_l1hdr.b_refcnt); 1554 mutex_init(&hdr->b_l1hdr.b_freeze_lock, NULL, MUTEX_DEFAULT, NULL); 1555 multilist_link_init(&hdr->b_l1hdr.b_arc_node); 1556 arc_space_consume(HDR_FULL_SIZE, ARC_SPACE_HDRS); 1557 1558 return (0); 1559} 1560 1561/* ARGSUSED */ 1562static int 1563hdr_l2only_cons(void *vbuf, void *unused, int kmflag) 1564{ 1565 arc_buf_hdr_t *hdr = vbuf; 1566 1567 bzero(hdr, HDR_L2ONLY_SIZE); 1568 arc_space_consume(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS); 1569 1570 return (0); 1571} 1572 1573/* ARGSUSED */ 1574static int 1575buf_cons(void *vbuf, void *unused, int kmflag) 1576{ 1577 arc_buf_t *buf = vbuf; 1578 1579 bzero(buf, sizeof (arc_buf_t)); 1580 mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL); 1581 arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS); 1582 1583 return (0); 1584} 1585 1586/* 1587 * Destructor callback - called when a cached buf is 1588 * no longer required. 1589 */ 1590/* ARGSUSED */ 1591static void 1592hdr_full_dest(void *vbuf, void *unused) 1593{ 1594 arc_buf_hdr_t *hdr = vbuf; 1595 1596 ASSERT(HDR_EMPTY(hdr)); 1597 cv_destroy(&hdr->b_l1hdr.b_cv); 1598 refcount_destroy(&hdr->b_l1hdr.b_refcnt); 1599 mutex_destroy(&hdr->b_l1hdr.b_freeze_lock); 1600 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 1601 arc_space_return(HDR_FULL_SIZE, ARC_SPACE_HDRS); 1602} 1603 1604/* ARGSUSED */ 1605static void 1606hdr_l2only_dest(void *vbuf, void *unused) 1607{ 1608 arc_buf_hdr_t *hdr = vbuf; 1609 1610 ASSERT(HDR_EMPTY(hdr)); 1611 arc_space_return(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS); 1612} 1613 1614/* ARGSUSED */ 1615static void 1616buf_dest(void *vbuf, void *unused) 1617{ 1618 arc_buf_t *buf = vbuf; 1619 1620 mutex_destroy(&buf->b_evict_lock); 1621 arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS); 1622} 1623 1624/* 1625 * Reclaim callback -- invoked when memory is low. 1626 */ 1627/* ARGSUSED */ 1628static void 1629hdr_recl(void *unused) 1630{ 1631 dprintf("hdr_recl called\n"); 1632 /* 1633 * umem calls the reclaim func when we destroy the buf cache, 1634 * which is after we do arc_fini(). 1635 */ 1636 if (!arc_dead) 1637 cv_signal(&arc_reclaim_thread_cv); 1638} 1639 1640static void 1641buf_init(void) 1642{ 1643 uint64_t *ct; 1644 uint64_t hsize = 1ULL << 12; 1645 int i, j; 1646 1647 /* 1648 * The hash table is big enough to fill all of physical memory 1649 * with an average block size of zfs_arc_average_blocksize (default 8K). 1650 * By default, the table will take up 1651 * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers). 1652 */ 1653 while (hsize * zfs_arc_average_blocksize < (uint64_t)physmem * PAGESIZE) 1654 hsize <<= 1; 1655retry: 1656 buf_hash_table.ht_mask = hsize - 1; 1657 buf_hash_table.ht_table = 1658 kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP); 1659 if (buf_hash_table.ht_table == NULL) { 1660 ASSERT(hsize > (1ULL << 8)); 1661 hsize >>= 1; 1662 goto retry; 1663 } 1664 1665 hdr_full_cache = kmem_cache_create("arc_buf_hdr_t_full", HDR_FULL_SIZE, 1666 0, hdr_full_cons, hdr_full_dest, hdr_recl, NULL, NULL, 0); 1667 hdr_l2only_cache = kmem_cache_create("arc_buf_hdr_t_l2only", 1668 HDR_L2ONLY_SIZE, 0, hdr_l2only_cons, hdr_l2only_dest, hdr_recl, 1669 NULL, NULL, 0); 1670 buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t), 1671 0, buf_cons, buf_dest, NULL, NULL, NULL, 0); 1672 1673 for (i = 0; i < 256; i++) 1674 for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--) 1675 *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY); 1676 1677 for (i = 0; i < BUF_LOCKS; i++) { 1678 mutex_init(&buf_hash_table.ht_locks[i].ht_lock, 1679 NULL, MUTEX_DEFAULT, NULL); 1680 } 1681} 1682 1683#define ARC_MINTIME (hz>>4) /* 62 ms */ 1684 1685static inline boolean_t 1686arc_buf_is_shared(arc_buf_t *buf) 1687{ 1688 boolean_t shared = (buf->b_data != NULL && 1689 buf->b_data == buf->b_hdr->b_l1hdr.b_pdata); 1690 IMPLY(shared, HDR_SHARED_DATA(buf->b_hdr)); 1691 return (shared); 1692} 1693 1694static inline void 1695arc_cksum_free(arc_buf_hdr_t *hdr) 1696{ 1697 ASSERT(HDR_HAS_L1HDR(hdr)); 1698 mutex_enter(&hdr->b_l1hdr.b_freeze_lock); 1699 if (hdr->b_l1hdr.b_freeze_cksum != NULL) { 1700 kmem_free(hdr->b_l1hdr.b_freeze_cksum, sizeof (zio_cksum_t)); 1701 hdr->b_l1hdr.b_freeze_cksum = NULL; 1702 } 1703 mutex_exit(&hdr->b_l1hdr.b_freeze_lock); 1704} 1705 1706static void 1707arc_cksum_verify(arc_buf_t *buf) 1708{ 1709 arc_buf_hdr_t *hdr = buf->b_hdr; 1710 zio_cksum_t zc; 1711 1712 if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 1713 return; 1714 1715 ASSERT(HDR_HAS_L1HDR(hdr)); 1716 1717 mutex_enter(&hdr->b_l1hdr.b_freeze_lock); 1718 if (hdr->b_l1hdr.b_freeze_cksum == NULL || HDR_IO_ERROR(hdr)) { 1719 mutex_exit(&hdr->b_l1hdr.b_freeze_lock); 1720 return; 1721 } 1722 fletcher_2_native(buf->b_data, HDR_GET_LSIZE(hdr), NULL, &zc); 1723 if (!ZIO_CHECKSUM_EQUAL(*hdr->b_l1hdr.b_freeze_cksum, zc)) 1724 panic("buffer modified while frozen!"); 1725 mutex_exit(&hdr->b_l1hdr.b_freeze_lock); 1726} 1727 1728static boolean_t 1729arc_cksum_is_equal(arc_buf_hdr_t *hdr, zio_t *zio) 1730{ 1731 enum zio_compress compress = BP_GET_COMPRESS(zio->io_bp); 1732 boolean_t valid_cksum; 1733 1734 ASSERT(!BP_IS_EMBEDDED(zio->io_bp)); 1735 VERIFY3U(BP_GET_PSIZE(zio->io_bp), ==, HDR_GET_PSIZE(hdr)); 1736 1737 /* 1738 * We rely on the blkptr's checksum to determine if the block 1739 * is valid or not. When compressed arc is enabled, the l2arc 1740 * writes the block to the l2arc just as it appears in the pool. 1741 * This allows us to use the blkptr's checksum to validate the 1742 * data that we just read off of the l2arc without having to store 1743 * a separate checksum in the arc_buf_hdr_t. However, if compressed 1744 * arc is disabled, then the data written to the l2arc is always 1745 * uncompressed and won't match the block as it exists in the main 1746 * pool. When this is the case, we must first compress it if it is 1747 * compressed on the main pool before we can validate the checksum. 1748 */ 1749 if (!HDR_COMPRESSION_ENABLED(hdr) && compress != ZIO_COMPRESS_OFF) { 1750 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF); 1751 uint64_t lsize = HDR_GET_LSIZE(hdr); 1752 uint64_t csize; 1753 1754 void *cbuf = zio_buf_alloc(HDR_GET_PSIZE(hdr)); 1755 csize = zio_compress_data(compress, zio->io_data, cbuf, lsize); 1756 ASSERT3U(csize, <=, HDR_GET_PSIZE(hdr)); 1757 if (csize < HDR_GET_PSIZE(hdr)) { 1758 /* 1759 * Compressed blocks are always a multiple of the 1760 * smallest ashift in the pool. Ideally, we would 1761 * like to round up the csize to the next 1762 * spa_min_ashift but that value may have changed 1763 * since the block was last written. Instead, 1764 * we rely on the fact that the hdr's psize 1765 * was set to the psize of the block when it was 1766 * last written. We set the csize to that value 1767 * and zero out any part that should not contain 1768 * data. 1769 */ 1770 bzero((char *)cbuf + csize, HDR_GET_PSIZE(hdr) - csize); 1771 csize = HDR_GET_PSIZE(hdr); 1772 } 1773 zio_push_transform(zio, cbuf, csize, HDR_GET_PSIZE(hdr), NULL); 1774 } 1775 1776 /* 1777 * Block pointers always store the checksum for the logical data. 1778 * If the block pointer has the gang bit set, then the checksum 1779 * it represents is for the reconstituted data and not for an 1780 * individual gang member. The zio pipeline, however, must be able to 1781 * determine the checksum of each of the gang constituents so it 1782 * treats the checksum comparison differently than what we need 1783 * for l2arc blocks. This prevents us from using the 1784 * zio_checksum_error() interface directly. Instead we must call the 1785 * zio_checksum_error_impl() so that we can ensure the checksum is 1786 * generated using the correct checksum algorithm and accounts for the 1787 * logical I/O size and not just a gang fragment. 1788 */ 1789 valid_cksum = (zio_checksum_error_impl(zio->io_spa, zio->io_bp, 1790 BP_GET_CHECKSUM(zio->io_bp), zio->io_data, zio->io_size, 1791 zio->io_offset, NULL) == 0); 1792 zio_pop_transforms(zio); 1793 return (valid_cksum); 1794} 1795 1796static void 1797arc_cksum_compute(arc_buf_t *buf) 1798{ 1799 arc_buf_hdr_t *hdr = buf->b_hdr; 1800 1801 if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 1802 return; 1803 1804 ASSERT(HDR_HAS_L1HDR(hdr)); 1805 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1806 if (hdr->b_l1hdr.b_freeze_cksum != NULL) { 1807 mutex_exit(&hdr->b_l1hdr.b_freeze_lock); 1808 return; 1809 } 1810 hdr->b_l1hdr.b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t), 1811 KM_SLEEP); 1812 fletcher_2_native(buf->b_data, HDR_GET_LSIZE(hdr), NULL, 1813 hdr->b_l1hdr.b_freeze_cksum); 1814 mutex_exit(&hdr->b_l1hdr.b_freeze_lock); 1815#ifdef illumos 1816 arc_buf_watch(buf); 1817#endif 1818} 1819 1820#ifdef illumos 1821#ifndef _KERNEL 1822typedef struct procctl { 1823 long cmd; 1824 prwatch_t prwatch; 1825} procctl_t; 1826#endif 1827 1828/* ARGSUSED */ 1829static void 1830arc_buf_unwatch(arc_buf_t *buf) 1831{ 1832#ifndef _KERNEL 1833 if (arc_watch) { 1834 int result; 1835 procctl_t ctl; 1836 ctl.cmd = PCWATCH; 1837 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data; 1838 ctl.prwatch.pr_size = 0; 1839 ctl.prwatch.pr_wflags = 0; 1840 result = write(arc_procfd, &ctl, sizeof (ctl)); 1841 ASSERT3U(result, ==, sizeof (ctl)); 1842 } 1843#endif 1844} 1845 1846/* ARGSUSED */ 1847static void 1848arc_buf_watch(arc_buf_t *buf) 1849{ 1850#ifndef _KERNEL 1851 if (arc_watch) { 1852 int result; 1853 procctl_t ctl; 1854 ctl.cmd = PCWATCH; 1855 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data; 1856 ctl.prwatch.pr_size = HDR_GET_LSIZE(buf->b_hdr); 1857 ctl.prwatch.pr_wflags = WA_WRITE; 1858 result = write(arc_procfd, &ctl, sizeof (ctl)); 1859 ASSERT3U(result, ==, sizeof (ctl)); 1860 } 1861#endif 1862} 1863#endif /* illumos */ 1864 1865static arc_buf_contents_t 1866arc_buf_type(arc_buf_hdr_t *hdr) 1867{ 1868 arc_buf_contents_t type; 1869 if (HDR_ISTYPE_METADATA(hdr)) { 1870 type = ARC_BUFC_METADATA; 1871 } else { 1872 type = ARC_BUFC_DATA; 1873 } 1874 VERIFY3U(hdr->b_type, ==, type); 1875 return (type); 1876} 1877 1878static uint32_t 1879arc_bufc_to_flags(arc_buf_contents_t type) 1880{ 1881 switch (type) { 1882 case ARC_BUFC_DATA: 1883 /* metadata field is 0 if buffer contains normal data */ 1884 return (0); 1885 case ARC_BUFC_METADATA: 1886 return (ARC_FLAG_BUFC_METADATA); 1887 default: 1888 break; 1889 } 1890 panic("undefined ARC buffer type!"); 1891 return ((uint32_t)-1); 1892} 1893 1894void 1895arc_buf_thaw(arc_buf_t *buf) 1896{ 1897 arc_buf_hdr_t *hdr = buf->b_hdr; 1898 1899 if (zfs_flags & ZFS_DEBUG_MODIFY) { 1900 if (hdr->b_l1hdr.b_state != arc_anon) 1901 panic("modifying non-anon buffer!"); 1902 if (HDR_IO_IN_PROGRESS(hdr)) 1903 panic("modifying buffer while i/o in progress!"); 1904 arc_cksum_verify(buf); 1905 } 1906 1907 ASSERT(HDR_HAS_L1HDR(hdr)); 1908 arc_cksum_free(hdr); 1909 1910 mutex_enter(&hdr->b_l1hdr.b_freeze_lock); 1911#ifdef ZFS_DEBUG 1912 if (zfs_flags & ZFS_DEBUG_MODIFY) { 1913 if (hdr->b_l1hdr.b_thawed != NULL) 1914 kmem_free(hdr->b_l1hdr.b_thawed, 1); 1915 hdr->b_l1hdr.b_thawed = kmem_alloc(1, KM_SLEEP); 1916 } 1917#endif 1918 1919 mutex_exit(&hdr->b_l1hdr.b_freeze_lock); 1920 1921#ifdef illumos 1922 arc_buf_unwatch(buf); 1923#endif 1924} 1925 1926void 1927arc_buf_freeze(arc_buf_t *buf) 1928{ 1929 arc_buf_hdr_t *hdr = buf->b_hdr; 1930 kmutex_t *hash_lock; 1931 1932 if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 1933 return; 1934 1935 hash_lock = HDR_LOCK(hdr); 1936 mutex_enter(hash_lock); 1937 1938 ASSERT(HDR_HAS_L1HDR(hdr)); 1939 ASSERT(hdr->b_l1hdr.b_freeze_cksum != NULL || 1940 hdr->b_l1hdr.b_state == arc_anon); 1941 arc_cksum_compute(buf); 1942 mutex_exit(hash_lock); 1943 1944} 1945 1946/* 1947 * The arc_buf_hdr_t's b_flags should never be modified directly. Instead, 1948 * the following functions should be used to ensure that the flags are 1949 * updated in a thread-safe way. When manipulating the flags either 1950 * the hash_lock must be held or the hdr must be undiscoverable. This 1951 * ensures that we're not racing with any other threads when updating 1952 * the flags. 1953 */ 1954static inline void 1955arc_hdr_set_flags(arc_buf_hdr_t *hdr, arc_flags_t flags) 1956{ 1957 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); 1958 hdr->b_flags |= flags; 1959} 1960 1961static inline void 1962arc_hdr_clear_flags(arc_buf_hdr_t *hdr, arc_flags_t flags) 1963{ 1964 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); 1965 hdr->b_flags &= ~flags; 1966} 1967 1968/* 1969 * Setting the compression bits in the arc_buf_hdr_t's b_flags is 1970 * done in a special way since we have to clear and set bits 1971 * at the same time. Consumers that wish to set the compression bits 1972 * must use this function to ensure that the flags are updated in 1973 * thread-safe manner. 1974 */ 1975static void 1976arc_hdr_set_compress(arc_buf_hdr_t *hdr, enum zio_compress cmp) 1977{ 1978 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); 1979 1980 /* 1981 * Holes and embedded blocks will always have a psize = 0 so 1982 * we ignore the compression of the blkptr and set the 1983 * arc_buf_hdr_t's compression to ZIO_COMPRESS_OFF. 1984 * Holes and embedded blocks remain anonymous so we don't 1985 * want to uncompress them. Mark them as uncompressed. 1986 */ 1987 if (!zfs_compressed_arc_enabled || HDR_GET_PSIZE(hdr) == 0) { 1988 arc_hdr_clear_flags(hdr, ARC_FLAG_COMPRESSED_ARC); 1989 HDR_SET_COMPRESS(hdr, ZIO_COMPRESS_OFF); 1990 ASSERT(!HDR_COMPRESSION_ENABLED(hdr)); 1991 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF); 1992 } else { 1993 arc_hdr_set_flags(hdr, ARC_FLAG_COMPRESSED_ARC); 1994 HDR_SET_COMPRESS(hdr, cmp); 1995 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, cmp); 1996 ASSERT(HDR_COMPRESSION_ENABLED(hdr)); 1997 } 1998} 1999 2000static int 2001arc_decompress(arc_buf_t *buf) 2002{ 2003 arc_buf_hdr_t *hdr = buf->b_hdr; 2004 dmu_object_byteswap_t bswap = hdr->b_l1hdr.b_byteswap; 2005 int error; 2006 2007 if (arc_buf_is_shared(buf)) { 2008 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF); 2009 } else if (HDR_GET_COMPRESS(hdr) == ZIO_COMPRESS_OFF) { 2010 /* 2011 * The arc_buf_hdr_t is either not compressed or is 2012 * associated with an embedded block or a hole in which 2013 * case they remain anonymous. 2014 */ 2015 IMPLY(HDR_COMPRESSION_ENABLED(hdr), HDR_GET_PSIZE(hdr) == 0 || 2016 HDR_GET_PSIZE(hdr) == HDR_GET_LSIZE(hdr)); 2017 ASSERT(!HDR_SHARED_DATA(hdr)); 2018 bcopy(hdr->b_l1hdr.b_pdata, buf->b_data, HDR_GET_LSIZE(hdr)); 2019 } else { 2020 ASSERT(!HDR_SHARED_DATA(hdr)); 2021 ASSERT3U(HDR_GET_LSIZE(hdr), !=, HDR_GET_PSIZE(hdr)); 2022 error = zio_decompress_data(HDR_GET_COMPRESS(hdr), 2023 hdr->b_l1hdr.b_pdata, buf->b_data, HDR_GET_PSIZE(hdr), 2024 HDR_GET_LSIZE(hdr)); 2025 if (error != 0) { 2026 zfs_dbgmsg("hdr %p, compress %d, psize %d, lsize %d", 2027 hdr, HDR_GET_COMPRESS(hdr), HDR_GET_PSIZE(hdr), 2028 HDR_GET_LSIZE(hdr)); 2029 return (SET_ERROR(EIO)); 2030 } 2031 } 2032 if (bswap != DMU_BSWAP_NUMFUNCS) { 2033 ASSERT(!HDR_SHARED_DATA(hdr)); 2034 ASSERT3U(bswap, <, DMU_BSWAP_NUMFUNCS); 2035 dmu_ot_byteswap[bswap].ob_func(buf->b_data, HDR_GET_LSIZE(hdr)); 2036 } 2037 arc_cksum_compute(buf); 2038 return (0); 2039} 2040 2041/* 2042 * Return the size of the block, b_pdata, that is stored in the arc_buf_hdr_t. 2043 */ 2044static uint64_t 2045arc_hdr_size(arc_buf_hdr_t *hdr) 2046{ 2047 uint64_t size; 2048 2049 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF && 2050 HDR_GET_PSIZE(hdr) > 0) { 2051 size = HDR_GET_PSIZE(hdr); 2052 } else { 2053 ASSERT3U(HDR_GET_LSIZE(hdr), !=, 0); 2054 size = HDR_GET_LSIZE(hdr); 2055 } 2056 return (size); 2057} 2058 2059/* 2060 * Increment the amount of evictable space in the arc_state_t's refcount. 2061 * We account for the space used by the hdr and the arc buf individually 2062 * so that we can add and remove them from the refcount individually. 2063 */ 2064static void 2065arc_evictable_space_increment(arc_buf_hdr_t *hdr, arc_state_t *state) 2066{ 2067 arc_buf_contents_t type = arc_buf_type(hdr); 2068 uint64_t lsize = HDR_GET_LSIZE(hdr); 2069 2070 ASSERT(HDR_HAS_L1HDR(hdr)); 2071 2072 if (GHOST_STATE(state)) { 2073 ASSERT0(hdr->b_l1hdr.b_bufcnt); 2074 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 2075 ASSERT3P(hdr->b_l1hdr.b_pdata, ==, NULL); 2076 (void) refcount_add_many(&state->arcs_esize[type], lsize, hdr); 2077 return; 2078 } 2079 2080 ASSERT(!GHOST_STATE(state)); 2081 if (hdr->b_l1hdr.b_pdata != NULL) { 2082 (void) refcount_add_many(&state->arcs_esize[type], 2083 arc_hdr_size(hdr), hdr); 2084 } 2085 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL; 2086 buf = buf->b_next) { 2087 if (arc_buf_is_shared(buf)) { 2088 ASSERT(ARC_BUF_LAST(buf)); 2089 continue; 2090 } 2091 (void) refcount_add_many(&state->arcs_esize[type], lsize, buf); 2092 } 2093} 2094 2095/* 2096 * Decrement the amount of evictable space in the arc_state_t's refcount. 2097 * We account for the space used by the hdr and the arc buf individually 2098 * so that we can add and remove them from the refcount individually. 2099 */ 2100static void 2101arc_evitable_space_decrement(arc_buf_hdr_t *hdr, arc_state_t *state) 2102{ 2103 arc_buf_contents_t type = arc_buf_type(hdr); 2104 uint64_t lsize = HDR_GET_LSIZE(hdr); 2105 2106 ASSERT(HDR_HAS_L1HDR(hdr)); 2107 2108 if (GHOST_STATE(state)) { 2109 ASSERT0(hdr->b_l1hdr.b_bufcnt); 2110 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 2111 ASSERT3P(hdr->b_l1hdr.b_pdata, ==, NULL); 2112 (void) refcount_remove_many(&state->arcs_esize[type], 2113 lsize, hdr); 2114 return; 2115 } 2116 2117 ASSERT(!GHOST_STATE(state)); 2118 if (hdr->b_l1hdr.b_pdata != NULL) { 2119 (void) refcount_remove_many(&state->arcs_esize[type], 2120 arc_hdr_size(hdr), hdr); 2121 } 2122 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL; 2123 buf = buf->b_next) { 2124 if (arc_buf_is_shared(buf)) { 2125 ASSERT(ARC_BUF_LAST(buf)); 2126 continue; 2127 } 2128 (void) refcount_remove_many(&state->arcs_esize[type], 2129 lsize, buf); 2130 } 2131} 2132 2133/* 2134 * Add a reference to this hdr indicating that someone is actively 2135 * referencing that memory. When the refcount transitions from 0 to 1, 2136 * we remove it from the respective arc_state_t list to indicate that 2137 * it is not evictable. 2138 */ 2139static void 2140add_reference(arc_buf_hdr_t *hdr, void *tag) 2141{ 2142 ASSERT(HDR_HAS_L1HDR(hdr)); 2143 if (!MUTEX_HELD(HDR_LOCK(hdr))) { 2144 ASSERT(hdr->b_l1hdr.b_state == arc_anon); 2145 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 2146 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 2147 } 2148 2149 arc_state_t *state = hdr->b_l1hdr.b_state; 2150 2151 if ((refcount_add(&hdr->b_l1hdr.b_refcnt, tag) == 1) && 2152 (state != arc_anon)) { 2153 /* We don't use the L2-only state list. */ 2154 if (state != arc_l2c_only) { 2155 multilist_remove(&state->arcs_list[arc_buf_type(hdr)], 2156 hdr); 2157 arc_evitable_space_decrement(hdr, state); 2158 } 2159 /* remove the prefetch flag if we get a reference */ 2160 arc_hdr_clear_flags(hdr, ARC_FLAG_PREFETCH); 2161 } 2162} 2163 2164/* 2165 * Remove a reference from this hdr. When the reference transitions from 2166 * 1 to 0 and we're not anonymous, then we add this hdr to the arc_state_t's 2167 * list making it eligible for eviction. 2168 */ 2169static int 2170remove_reference(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, void *tag) 2171{ 2172 int cnt; 2173 arc_state_t *state = hdr->b_l1hdr.b_state; 2174 2175 ASSERT(HDR_HAS_L1HDR(hdr)); 2176 ASSERT(state == arc_anon || MUTEX_HELD(hash_lock)); 2177 ASSERT(!GHOST_STATE(state)); 2178 2179 /* 2180 * arc_l2c_only counts as a ghost state so we don't need to explicitly 2181 * check to prevent usage of the arc_l2c_only list. 2182 */ 2183 if (((cnt = refcount_remove(&hdr->b_l1hdr.b_refcnt, tag)) == 0) && 2184 (state != arc_anon)) { 2185 multilist_insert(&state->arcs_list[arc_buf_type(hdr)], hdr); 2186 ASSERT3U(hdr->b_l1hdr.b_bufcnt, >, 0); 2187 arc_evictable_space_increment(hdr, state); 2188 } 2189 return (cnt); 2190} 2191 2192/* 2193 * Move the supplied buffer to the indicated state. The hash lock 2194 * for the buffer must be held by the caller. 2195 */ 2196static void 2197arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *hdr, 2198 kmutex_t *hash_lock) 2199{ 2200 arc_state_t *old_state; 2201 int64_t refcnt; 2202 uint32_t bufcnt; 2203 boolean_t update_old, update_new; 2204 arc_buf_contents_t buftype = arc_buf_type(hdr); 2205 2206 /* 2207 * We almost always have an L1 hdr here, since we call arc_hdr_realloc() 2208 * in arc_read() when bringing a buffer out of the L2ARC. However, the 2209 * L1 hdr doesn't always exist when we change state to arc_anon before 2210 * destroying a header, in which case reallocating to add the L1 hdr is 2211 * pointless. 2212 */ 2213 if (HDR_HAS_L1HDR(hdr)) { 2214 old_state = hdr->b_l1hdr.b_state; 2215 refcnt = refcount_count(&hdr->b_l1hdr.b_refcnt); 2216 bufcnt = hdr->b_l1hdr.b_bufcnt; 2217 update_old = (bufcnt > 0 || hdr->b_l1hdr.b_pdata != NULL); 2218 } else { 2219 old_state = arc_l2c_only; 2220 refcnt = 0; 2221 bufcnt = 0; 2222 update_old = B_FALSE; 2223 } 2224 update_new = update_old; 2225 2226 ASSERT(MUTEX_HELD(hash_lock)); 2227 ASSERT3P(new_state, !=, old_state); 2228 ASSERT(!GHOST_STATE(new_state) || bufcnt == 0); 2229 ASSERT(old_state != arc_anon || bufcnt <= 1); 2230 2231 /* 2232 * If this buffer is evictable, transfer it from the 2233 * old state list to the new state list. 2234 */ 2235 if (refcnt == 0) { 2236 if (old_state != arc_anon && old_state != arc_l2c_only) { 2237 ASSERT(HDR_HAS_L1HDR(hdr)); 2238 multilist_remove(&old_state->arcs_list[buftype], hdr); 2239 2240 if (GHOST_STATE(old_state)) { 2241 ASSERT0(bufcnt); 2242 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 2243 update_old = B_TRUE; 2244 } 2245 arc_evitable_space_decrement(hdr, old_state); 2246 } 2247 if (new_state != arc_anon && new_state != arc_l2c_only) { 2248 2249 /* 2250 * An L1 header always exists here, since if we're 2251 * moving to some L1-cached state (i.e. not l2c_only or 2252 * anonymous), we realloc the header to add an L1hdr 2253 * beforehand. 2254 */ 2255 ASSERT(HDR_HAS_L1HDR(hdr)); 2256 multilist_insert(&new_state->arcs_list[buftype], hdr); 2257 2258 if (GHOST_STATE(new_state)) { 2259 ASSERT0(bufcnt); 2260 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 2261 update_new = B_TRUE; 2262 } 2263 arc_evictable_space_increment(hdr, new_state); 2264 } 2265 } 2266 2267 ASSERT(!HDR_EMPTY(hdr)); 2268 if (new_state == arc_anon && HDR_IN_HASH_TABLE(hdr)) 2269 buf_hash_remove(hdr); 2270 2271 /* adjust state sizes (ignore arc_l2c_only) */ 2272 2273 if (update_new && new_state != arc_l2c_only) { 2274 ASSERT(HDR_HAS_L1HDR(hdr)); 2275 if (GHOST_STATE(new_state)) { 2276 ASSERT0(bufcnt); 2277 2278 /* 2279 * When moving a header to a ghost state, we first 2280 * remove all arc buffers. Thus, we'll have a 2281 * bufcnt of zero, and no arc buffer to use for 2282 * the reference. As a result, we use the arc 2283 * header pointer for the reference. 2284 */ 2285 (void) refcount_add_many(&new_state->arcs_size, 2286 HDR_GET_LSIZE(hdr), hdr); 2287 ASSERT3P(hdr->b_l1hdr.b_pdata, ==, NULL); 2288 } else { 2289 uint32_t buffers = 0; 2290 2291 /* 2292 * Each individual buffer holds a unique reference, 2293 * thus we must remove each of these references one 2294 * at a time. 2295 */ 2296 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL; 2297 buf = buf->b_next) { 2298 ASSERT3U(bufcnt, !=, 0); 2299 buffers++; 2300 2301 /* 2302 * When the arc_buf_t is sharing the data 2303 * block with the hdr, the owner of the 2304 * reference belongs to the hdr. Only 2305 * add to the refcount if the arc_buf_t is 2306 * not shared. 2307 */ 2308 if (arc_buf_is_shared(buf)) { 2309 ASSERT(ARC_BUF_LAST(buf)); 2310 continue; 2311 } 2312 2313 (void) refcount_add_many(&new_state->arcs_size, 2314 HDR_GET_LSIZE(hdr), buf); 2315 } 2316 ASSERT3U(bufcnt, ==, buffers); 2317 2318 if (hdr->b_l1hdr.b_pdata != NULL) { 2319 (void) refcount_add_many(&new_state->arcs_size, 2320 arc_hdr_size(hdr), hdr); 2321 } else { 2322 ASSERT(GHOST_STATE(old_state)); 2323 } 2324 } 2325 } 2326 2327 if (update_old && old_state != arc_l2c_only) { 2328 ASSERT(HDR_HAS_L1HDR(hdr)); 2329 if (GHOST_STATE(old_state)) { 2330 ASSERT0(bufcnt); 2331 2332 /* 2333 * When moving a header off of a ghost state, 2334 * the header will not contain any arc buffers. 2335 * We use the arc header pointer for the reference 2336 * which is exactly what we did when we put the 2337 * header on the ghost state. 2338 */ 2339 2340 (void) refcount_remove_many(&old_state->arcs_size, 2341 HDR_GET_LSIZE(hdr), hdr); 2342 ASSERT3P(hdr->b_l1hdr.b_pdata, ==, NULL); 2343 } else { 2344 uint32_t buffers = 0; 2345 2346 /* 2347 * Each individual buffer holds a unique reference, 2348 * thus we must remove each of these references one 2349 * at a time. 2350 */ 2351 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL; 2352 buf = buf->b_next) { 2353 ASSERT3P(bufcnt, !=, 0); 2354 buffers++; 2355 2356 /* 2357 * When the arc_buf_t is sharing the data 2358 * block with the hdr, the owner of the 2359 * reference belongs to the hdr. Only 2360 * add to the refcount if the arc_buf_t is 2361 * not shared. 2362 */ 2363 if (arc_buf_is_shared(buf)) { 2364 ASSERT(ARC_BUF_LAST(buf)); 2365 continue; 2366 } 2367 2368 (void) refcount_remove_many( 2369 &old_state->arcs_size, HDR_GET_LSIZE(hdr), 2370 buf); 2371 } 2372 ASSERT3U(bufcnt, ==, buffers); 2373 ASSERT3P(hdr->b_l1hdr.b_pdata, !=, NULL); 2374 (void) refcount_remove_many( 2375 &old_state->arcs_size, arc_hdr_size(hdr), hdr); 2376 } 2377 } 2378 2379 if (HDR_HAS_L1HDR(hdr)) 2380 hdr->b_l1hdr.b_state = new_state; 2381 2382 /* 2383 * L2 headers should never be on the L2 state list since they don't 2384 * have L1 headers allocated. 2385 */ 2386 ASSERT(multilist_is_empty(&arc_l2c_only->arcs_list[ARC_BUFC_DATA]) && 2387 multilist_is_empty(&arc_l2c_only->arcs_list[ARC_BUFC_METADATA])); 2388} 2389 2390void 2391arc_space_consume(uint64_t space, arc_space_type_t type) 2392{ 2393 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES); 2394 2395 switch (type) { 2396 case ARC_SPACE_DATA: 2397 ARCSTAT_INCR(arcstat_data_size, space); 2398 break; 2399 case ARC_SPACE_META: 2400 ARCSTAT_INCR(arcstat_metadata_size, space); 2401 break; 2402 case ARC_SPACE_OTHER: 2403 ARCSTAT_INCR(arcstat_other_size, space); 2404 break; 2405 case ARC_SPACE_HDRS: 2406 ARCSTAT_INCR(arcstat_hdr_size, space); 2407 break; 2408 case ARC_SPACE_L2HDRS: 2409 ARCSTAT_INCR(arcstat_l2_hdr_size, space); 2410 break; 2411 } 2412 2413 if (type != ARC_SPACE_DATA) 2414 ARCSTAT_INCR(arcstat_meta_used, space); 2415 2416 atomic_add_64(&arc_size, space); 2417} 2418 2419void 2420arc_space_return(uint64_t space, arc_space_type_t type) 2421{ 2422 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES); 2423 2424 switch (type) { 2425 case ARC_SPACE_DATA: 2426 ARCSTAT_INCR(arcstat_data_size, -space); 2427 break; 2428 case ARC_SPACE_META: 2429 ARCSTAT_INCR(arcstat_metadata_size, -space); 2430 break; 2431 case ARC_SPACE_OTHER: 2432 ARCSTAT_INCR(arcstat_other_size, -space); 2433 break; 2434 case ARC_SPACE_HDRS: 2435 ARCSTAT_INCR(arcstat_hdr_size, -space); 2436 break; 2437 case ARC_SPACE_L2HDRS: 2438 ARCSTAT_INCR(arcstat_l2_hdr_size, -space); 2439 break; 2440 } 2441 2442 if (type != ARC_SPACE_DATA) { 2443 ASSERT(arc_meta_used >= space); 2444 if (arc_meta_max < arc_meta_used) 2445 arc_meta_max = arc_meta_used; 2446 ARCSTAT_INCR(arcstat_meta_used, -space); 2447 } 2448 2449 ASSERT(arc_size >= space); 2450 atomic_add_64(&arc_size, -space); 2451} 2452 2453/* 2454 * Allocate an initial buffer for this hdr, subsequent buffers will 2455 * use arc_buf_clone(). 2456 */ 2457static arc_buf_t * 2458arc_buf_alloc_impl(arc_buf_hdr_t *hdr, void *tag) 2459{ 2460 arc_buf_t *buf; 2461 2462 ASSERT(HDR_HAS_L1HDR(hdr)); 2463 ASSERT3U(HDR_GET_LSIZE(hdr), >, 0); 2464 VERIFY(hdr->b_type == ARC_BUFC_DATA || 2465 hdr->b_type == ARC_BUFC_METADATA); 2466 2467 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 2468 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 2469 ASSERT0(hdr->b_l1hdr.b_bufcnt); 2470 2471 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 2472 buf->b_hdr = hdr; 2473 buf->b_data = NULL; 2474 buf->b_next = NULL; 2475 2476 add_reference(hdr, tag); 2477 2478 /* 2479 * We're about to change the hdr's b_flags. We must either 2480 * hold the hash_lock or be undiscoverable. 2481 */ 2482 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); 2483 2484 /* 2485 * If the hdr's data can be shared (no byteswapping, hdr is 2486 * uncompressed, hdr's data is not currently being written to the 2487 * L2ARC write) then we share the data buffer and set the appropriate 2488 * bit in the hdr's b_flags to indicate the hdr is sharing it's 2489 * b_pdata with the arc_buf_t. Otherwise, we allocate a new buffer to 2490 * store the buf's data. 2491 */ 2492 if (hdr->b_l1hdr.b_byteswap == DMU_BSWAP_NUMFUNCS && 2493 HDR_GET_COMPRESS(hdr) == ZIO_COMPRESS_OFF && !HDR_L2_WRITING(hdr)) { 2494 buf->b_data = hdr->b_l1hdr.b_pdata; 2495 arc_hdr_set_flags(hdr, ARC_FLAG_SHARED_DATA); 2496 } else { 2497 buf->b_data = arc_get_data_buf(hdr, HDR_GET_LSIZE(hdr), buf); 2498 ARCSTAT_INCR(arcstat_overhead_size, HDR_GET_LSIZE(hdr)); 2499 arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA); 2500 } 2501 VERIFY3P(buf->b_data, !=, NULL); 2502 2503 hdr->b_l1hdr.b_buf = buf; 2504 hdr->b_l1hdr.b_bufcnt += 1; 2505 2506 return (buf); 2507} 2508 2509/* 2510 * Used when allocating additional buffers. 2511 */ 2512static arc_buf_t * 2513arc_buf_clone(arc_buf_t *from) 2514{ 2515 arc_buf_t *buf; 2516 arc_buf_hdr_t *hdr = from->b_hdr; 2517 uint64_t size = HDR_GET_LSIZE(hdr); 2518 2519 ASSERT(HDR_HAS_L1HDR(hdr)); 2520 ASSERT(hdr->b_l1hdr.b_state != arc_anon); 2521 2522 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 2523 buf->b_hdr = hdr; 2524 buf->b_data = NULL; 2525 buf->b_next = hdr->b_l1hdr.b_buf; 2526 hdr->b_l1hdr.b_buf = buf; 2527 buf->b_data = arc_get_data_buf(hdr, HDR_GET_LSIZE(hdr), buf); 2528 bcopy(from->b_data, buf->b_data, size); 2529 hdr->b_l1hdr.b_bufcnt += 1; 2530 2531 ARCSTAT_INCR(arcstat_overhead_size, HDR_GET_LSIZE(hdr)); 2532 return (buf); 2533} 2534 2535static char *arc_onloan_tag = "onloan"; 2536 2537/* 2538 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in 2539 * flight data by arc_tempreserve_space() until they are "returned". Loaned 2540 * buffers must be returned to the arc before they can be used by the DMU or 2541 * freed. 2542 */ 2543arc_buf_t * 2544arc_loan_buf(spa_t *spa, int size) 2545{ 2546 arc_buf_t *buf; 2547 2548 buf = arc_alloc_buf(spa, size, arc_onloan_tag, ARC_BUFC_DATA); 2549 2550 atomic_add_64(&arc_loaned_bytes, size); 2551 return (buf); 2552} 2553 2554/* 2555 * Return a loaned arc buffer to the arc. 2556 */ 2557void 2558arc_return_buf(arc_buf_t *buf, void *tag) 2559{ 2560 arc_buf_hdr_t *hdr = buf->b_hdr; 2561 2562 ASSERT3P(buf->b_data, !=, NULL); 2563 ASSERT(HDR_HAS_L1HDR(hdr)); 2564 (void) refcount_add(&hdr->b_l1hdr.b_refcnt, tag); 2565 (void) refcount_remove(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag); 2566 2567 atomic_add_64(&arc_loaned_bytes, -HDR_GET_LSIZE(hdr)); 2568} 2569 2570/* Detach an arc_buf from a dbuf (tag) */ 2571void 2572arc_loan_inuse_buf(arc_buf_t *buf, void *tag) 2573{ 2574 arc_buf_hdr_t *hdr = buf->b_hdr; 2575 2576 ASSERT3P(buf->b_data, !=, NULL); 2577 ASSERT(HDR_HAS_L1HDR(hdr)); 2578 (void) refcount_add(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag); 2579 (void) refcount_remove(&hdr->b_l1hdr.b_refcnt, tag); 2580 2581 atomic_add_64(&arc_loaned_bytes, HDR_GET_LSIZE(hdr)); 2582} 2583 2584static void 2585l2arc_free_data_on_write(void *data, size_t size, arc_buf_contents_t type) 2586{ 2587 l2arc_data_free_t *df = kmem_alloc(sizeof (*df), KM_SLEEP); 2588 2589 df->l2df_data = data; 2590 df->l2df_size = size; 2591 df->l2df_type = type; 2592 mutex_enter(&l2arc_free_on_write_mtx); 2593 list_insert_head(l2arc_free_on_write, df); 2594 mutex_exit(&l2arc_free_on_write_mtx); 2595} 2596 2597static void 2598arc_hdr_free_on_write(arc_buf_hdr_t *hdr) 2599{ 2600 arc_state_t *state = hdr->b_l1hdr.b_state; 2601 arc_buf_contents_t type = arc_buf_type(hdr); 2602 uint64_t size = arc_hdr_size(hdr); 2603 2604 /* protected by hash lock, if in the hash table */ 2605 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) { 2606 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 2607 ASSERT(state != arc_anon && state != arc_l2c_only); 2608 2609 (void) refcount_remove_many(&state->arcs_esize[type], 2610 size, hdr); 2611 } 2612 (void) refcount_remove_many(&state->arcs_size, size, hdr); 2613 if (type == ARC_BUFC_METADATA) { 2614 arc_space_return(size, ARC_SPACE_META); 2615 } else { 2616 ASSERT(type == ARC_BUFC_DATA); 2617 arc_space_return(size, ARC_SPACE_DATA); 2618 } 2619 2620 l2arc_free_data_on_write(hdr->b_l1hdr.b_pdata, size, type); 2621} 2622 2623/* 2624 * Share the arc_buf_t's data with the hdr. Whenever we are sharing the 2625 * data buffer, we transfer the refcount ownership to the hdr and update 2626 * the appropriate kstats. 2627 */ 2628static void 2629arc_share_buf(arc_buf_hdr_t *hdr, arc_buf_t *buf) 2630{ 2631 arc_state_t *state = hdr->b_l1hdr.b_state; 2632 2633 ASSERT(!HDR_SHARED_DATA(hdr)); 2634 ASSERT(!arc_buf_is_shared(buf)); 2635 ASSERT3P(hdr->b_l1hdr.b_pdata, ==, NULL); 2636 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); 2637 2638 /* 2639 * Start sharing the data buffer. We transfer the 2640 * refcount ownership to the hdr since it always owns 2641 * the refcount whenever an arc_buf_t is shared. 2642 */ 2643 refcount_transfer_ownership(&state->arcs_size, buf, hdr); 2644 hdr->b_l1hdr.b_pdata = buf->b_data; 2645 arc_hdr_set_flags(hdr, ARC_FLAG_SHARED_DATA); 2646 2647 /* 2648 * Since we've transferred ownership to the hdr we need 2649 * to increment its compressed and uncompressed kstats and 2650 * decrement the overhead size. 2651 */ 2652 ARCSTAT_INCR(arcstat_compressed_size, arc_hdr_size(hdr)); 2653 ARCSTAT_INCR(arcstat_uncompressed_size, HDR_GET_LSIZE(hdr)); 2654 ARCSTAT_INCR(arcstat_overhead_size, -HDR_GET_LSIZE(hdr)); 2655} 2656 2657static void 2658arc_unshare_buf(arc_buf_hdr_t *hdr, arc_buf_t *buf) 2659{ 2660 arc_state_t *state = hdr->b_l1hdr.b_state; 2661 2662 ASSERT(HDR_SHARED_DATA(hdr)); 2663 ASSERT(arc_buf_is_shared(buf)); 2664 ASSERT3P(hdr->b_l1hdr.b_pdata, !=, NULL); 2665 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); 2666 2667 /* 2668 * We are no longer sharing this buffer so we need 2669 * to transfer its ownership to the rightful owner. 2670 */ 2671 refcount_transfer_ownership(&state->arcs_size, hdr, buf); 2672 arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA); 2673 hdr->b_l1hdr.b_pdata = NULL; 2674 2675 /* 2676 * Since the buffer is no longer shared between 2677 * the arc buf and the hdr, count it as overhead. 2678 */ 2679 ARCSTAT_INCR(arcstat_compressed_size, -arc_hdr_size(hdr)); 2680 ARCSTAT_INCR(arcstat_uncompressed_size, -HDR_GET_LSIZE(hdr)); 2681 ARCSTAT_INCR(arcstat_overhead_size, HDR_GET_LSIZE(hdr)); 2682} 2683 2684/* 2685 * Free up buf->b_data and if 'remove' is set, then pull the 2686 * arc_buf_t off of the the arc_buf_hdr_t's list and free it. 2687 */ 2688static void 2689arc_buf_destroy_impl(arc_buf_t *buf, boolean_t remove) 2690{ 2691 arc_buf_t **bufp; 2692 arc_buf_hdr_t *hdr = buf->b_hdr; 2693 uint64_t size = HDR_GET_LSIZE(hdr); 2694 boolean_t destroyed_buf_is_shared = arc_buf_is_shared(buf); 2695 2696 /* 2697 * Free up the data associated with the buf but only 2698 * if we're not sharing this with the hdr. If we are sharing 2699 * it with the hdr, then hdr will have performed the allocation 2700 * so allow it to do the free. 2701 */ 2702 if (buf->b_data != NULL) { 2703 /* 2704 * We're about to change the hdr's b_flags. We must either 2705 * hold the hash_lock or be undiscoverable. 2706 */ 2707 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); 2708 2709 arc_cksum_verify(buf); 2710#ifdef illumos 2711 arc_buf_unwatch(buf); 2712#endif 2713 2714 if (destroyed_buf_is_shared) { 2715 ASSERT(ARC_BUF_LAST(buf)); 2716 ASSERT(HDR_SHARED_DATA(hdr)); 2717 arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA); 2718 } else { 2719 arc_free_data_buf(hdr, buf->b_data, size, buf); 2720 ARCSTAT_INCR(arcstat_overhead_size, -size); 2721 } 2722 buf->b_data = NULL; 2723 2724 ASSERT(hdr->b_l1hdr.b_bufcnt > 0); 2725 hdr->b_l1hdr.b_bufcnt -= 1; 2726 } 2727 2728 /* only remove the buf if requested */ 2729 if (!remove) 2730 return; 2731 2732 /* remove the buf from the hdr list */ 2733 arc_buf_t *lastbuf = NULL; 2734 bufp = &hdr->b_l1hdr.b_buf; 2735 while (*bufp != NULL) { 2736 if (*bufp == buf) 2737 *bufp = buf->b_next; 2738 2739 /* 2740 * If we've removed a buffer in the middle of 2741 * the list then update the lastbuf and update 2742 * bufp. 2743 */ 2744 if (*bufp != NULL) { 2745 lastbuf = *bufp; 2746 bufp = &(*bufp)->b_next; 2747 } 2748 } 2749 buf->b_next = NULL; 2750 ASSERT3P(lastbuf, !=, buf); 2751 2752 /* 2753 * If the current arc_buf_t is sharing its data 2754 * buffer with the hdr, then reassign the hdr's 2755 * b_pdata to share it with the new buffer at the end 2756 * of the list. The shared buffer is always the last one 2757 * on the hdr's buffer list. 2758 */ 2759 if (destroyed_buf_is_shared && lastbuf != NULL) { 2760 ASSERT(ARC_BUF_LAST(buf)); 2761 ASSERT(ARC_BUF_LAST(lastbuf)); 2762 VERIFY(!arc_buf_is_shared(lastbuf)); 2763 2764 ASSERT3P(hdr->b_l1hdr.b_pdata, !=, NULL); 2765 arc_hdr_free_pdata(hdr); 2766 2767 /* 2768 * We must setup a new shared block between the 2769 * last buffer and the hdr. The data would have 2770 * been allocated by the arc buf so we need to transfer 2771 * ownership to the hdr since it's now being shared. 2772 */ 2773 arc_share_buf(hdr, lastbuf); 2774 } else if (HDR_SHARED_DATA(hdr)) { 2775 ASSERT(arc_buf_is_shared(lastbuf)); 2776 } 2777 2778 if (hdr->b_l1hdr.b_bufcnt == 0) 2779 arc_cksum_free(hdr); 2780 2781 /* clean up the buf */ 2782 buf->b_hdr = NULL; 2783 kmem_cache_free(buf_cache, buf); 2784} 2785 2786static void 2787arc_hdr_alloc_pdata(arc_buf_hdr_t *hdr) 2788{ 2789 ASSERT3U(HDR_GET_LSIZE(hdr), >, 0); 2790 ASSERT(HDR_HAS_L1HDR(hdr)); 2791 ASSERT(!HDR_SHARED_DATA(hdr)); 2792 2793 ASSERT3P(hdr->b_l1hdr.b_pdata, ==, NULL); 2794 hdr->b_l1hdr.b_pdata = arc_get_data_buf(hdr, arc_hdr_size(hdr), hdr); 2795 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS; 2796 ASSERT3P(hdr->b_l1hdr.b_pdata, !=, NULL); 2797 2798 ARCSTAT_INCR(arcstat_compressed_size, arc_hdr_size(hdr)); 2799 ARCSTAT_INCR(arcstat_uncompressed_size, HDR_GET_LSIZE(hdr)); 2800} 2801 2802static void 2803arc_hdr_free_pdata(arc_buf_hdr_t *hdr) 2804{ 2805 ASSERT(HDR_HAS_L1HDR(hdr)); 2806 ASSERT3P(hdr->b_l1hdr.b_pdata, !=, NULL); 2807 2808 /* 2809 * If the hdr is currently being written to the l2arc then 2810 * we defer freeing the data by adding it to the l2arc_free_on_write 2811 * list. The l2arc will free the data once it's finished 2812 * writing it to the l2arc device. 2813 */ 2814 if (HDR_L2_WRITING(hdr)) { 2815 arc_hdr_free_on_write(hdr); 2816 ARCSTAT_BUMP(arcstat_l2_free_on_write); 2817 } else { 2818 arc_free_data_buf(hdr, hdr->b_l1hdr.b_pdata, 2819 arc_hdr_size(hdr), hdr); 2820 } 2821 hdr->b_l1hdr.b_pdata = NULL; 2822 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS; 2823 2824 ARCSTAT_INCR(arcstat_compressed_size, -arc_hdr_size(hdr)); 2825 ARCSTAT_INCR(arcstat_uncompressed_size, -HDR_GET_LSIZE(hdr)); 2826} 2827 2828static arc_buf_hdr_t * 2829arc_hdr_alloc(uint64_t spa, int32_t psize, int32_t lsize, 2830 enum zio_compress compress, arc_buf_contents_t type) 2831{ 2832 arc_buf_hdr_t *hdr; 2833 2834 ASSERT3U(lsize, >, 0); 2835 VERIFY(type == ARC_BUFC_DATA || type == ARC_BUFC_METADATA); 2836 2837 hdr = kmem_cache_alloc(hdr_full_cache, KM_PUSHPAGE); 2838 ASSERT(HDR_EMPTY(hdr)); 2839 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL); 2840 ASSERT3P(hdr->b_l1hdr.b_thawed, ==, NULL); 2841 HDR_SET_PSIZE(hdr, psize); 2842 HDR_SET_LSIZE(hdr, lsize); 2843 hdr->b_spa = spa; 2844 hdr->b_type = type; 2845 hdr->b_flags = 0; 2846 arc_hdr_set_flags(hdr, arc_bufc_to_flags(type) | ARC_FLAG_HAS_L1HDR); 2847 arc_hdr_set_compress(hdr, compress); 2848 2849 hdr->b_l1hdr.b_state = arc_anon; 2850 hdr->b_l1hdr.b_arc_access = 0; 2851 hdr->b_l1hdr.b_bufcnt = 0; 2852 hdr->b_l1hdr.b_buf = NULL; 2853 2854 /* 2855 * Allocate the hdr's buffer. This will contain either 2856 * the compressed or uncompressed data depending on the block 2857 * it references and compressed arc enablement. 2858 */ 2859 arc_hdr_alloc_pdata(hdr); 2860 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 2861 2862 return (hdr); 2863} 2864 2865/* 2866 * Transition between the two allocation states for the arc_buf_hdr struct. 2867 * The arc_buf_hdr struct can be allocated with (hdr_full_cache) or without 2868 * (hdr_l2only_cache) the fields necessary for the L1 cache - the smaller 2869 * version is used when a cache buffer is only in the L2ARC in order to reduce 2870 * memory usage. 2871 */ 2872static arc_buf_hdr_t * 2873arc_hdr_realloc(arc_buf_hdr_t *hdr, kmem_cache_t *old, kmem_cache_t *new) 2874{ 2875 ASSERT(HDR_HAS_L2HDR(hdr)); 2876 2877 arc_buf_hdr_t *nhdr; 2878 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev; 2879 2880 ASSERT((old == hdr_full_cache && new == hdr_l2only_cache) || 2881 (old == hdr_l2only_cache && new == hdr_full_cache)); 2882 2883 nhdr = kmem_cache_alloc(new, KM_PUSHPAGE); 2884 2885 ASSERT(MUTEX_HELD(HDR_LOCK(hdr))); 2886 buf_hash_remove(hdr); 2887 2888 bcopy(hdr, nhdr, HDR_L2ONLY_SIZE); 2889 2890 if (new == hdr_full_cache) { 2891 arc_hdr_set_flags(nhdr, ARC_FLAG_HAS_L1HDR); 2892 /* 2893 * arc_access and arc_change_state need to be aware that a 2894 * header has just come out of L2ARC, so we set its state to 2895 * l2c_only even though it's about to change. 2896 */ 2897 nhdr->b_l1hdr.b_state = arc_l2c_only; 2898 2899 /* Verify previous threads set to NULL before freeing */ 2900 ASSERT3P(nhdr->b_l1hdr.b_pdata, ==, NULL); 2901 } else { 2902 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 2903 ASSERT0(hdr->b_l1hdr.b_bufcnt); 2904 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL); 2905 2906 /* 2907 * If we've reached here, We must have been called from 2908 * arc_evict_hdr(), as such we should have already been 2909 * removed from any ghost list we were previously on 2910 * (which protects us from racing with arc_evict_state), 2911 * thus no locking is needed during this check. 2912 */ 2913 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 2914 2915 /* 2916 * A buffer must not be moved into the arc_l2c_only 2917 * state if it's not finished being written out to the 2918 * l2arc device. Otherwise, the b_l1hdr.b_pdata field 2919 * might try to be accessed, even though it was removed. 2920 */ 2921 VERIFY(!HDR_L2_WRITING(hdr)); 2922 VERIFY3P(hdr->b_l1hdr.b_pdata, ==, NULL); 2923 2924#ifdef ZFS_DEBUG 2925 if (hdr->b_l1hdr.b_thawed != NULL) { 2926 kmem_free(hdr->b_l1hdr.b_thawed, 1); 2927 hdr->b_l1hdr.b_thawed = NULL; 2928 } 2929#endif 2930 2931 arc_hdr_clear_flags(nhdr, ARC_FLAG_HAS_L1HDR); 2932 } 2933 /* 2934 * The header has been reallocated so we need to re-insert it into any 2935 * lists it was on. 2936 */ 2937 (void) buf_hash_insert(nhdr, NULL); 2938 2939 ASSERT(list_link_active(&hdr->b_l2hdr.b_l2node)); 2940 2941 mutex_enter(&dev->l2ad_mtx); 2942 2943 /* 2944 * We must place the realloc'ed header back into the list at 2945 * the same spot. Otherwise, if it's placed earlier in the list, 2946 * l2arc_write_buffers() could find it during the function's 2947 * write phase, and try to write it out to the l2arc. 2948 */ 2949 list_insert_after(&dev->l2ad_buflist, hdr, nhdr); 2950 list_remove(&dev->l2ad_buflist, hdr); 2951 2952 mutex_exit(&dev->l2ad_mtx); 2953 2954 /* 2955 * Since we're using the pointer address as the tag when 2956 * incrementing and decrementing the l2ad_alloc refcount, we 2957 * must remove the old pointer (that we're about to destroy) and 2958 * add the new pointer to the refcount. Otherwise we'd remove 2959 * the wrong pointer address when calling arc_hdr_destroy() later. 2960 */ 2961 2962 (void) refcount_remove_many(&dev->l2ad_alloc, arc_hdr_size(hdr), hdr); 2963 (void) refcount_add_many(&dev->l2ad_alloc, arc_hdr_size(nhdr), nhdr); 2964 2965 buf_discard_identity(hdr); 2966 kmem_cache_free(old, hdr); 2967 2968 return (nhdr); 2969} 2970 2971/* 2972 * Allocate a new arc_buf_hdr_t and arc_buf_t and return the buf to the caller. 2973 * The buf is returned thawed since we expect the consumer to modify it. 2974 */ 2975arc_buf_t * 2976arc_alloc_buf(spa_t *spa, int32_t size, void *tag, arc_buf_contents_t type) 2977{ 2978 arc_buf_hdr_t *hdr = arc_hdr_alloc(spa_load_guid(spa), size, size, 2979 ZIO_COMPRESS_OFF, type); 2980 ASSERT(!MUTEX_HELD(HDR_LOCK(hdr))); 2981 arc_buf_t *buf = arc_buf_alloc_impl(hdr, tag); 2982 arc_buf_thaw(buf); 2983 return (buf); 2984} 2985 2986static void 2987arc_hdr_l2hdr_destroy(arc_buf_hdr_t *hdr) 2988{ 2989 l2arc_buf_hdr_t *l2hdr = &hdr->b_l2hdr; 2990 l2arc_dev_t *dev = l2hdr->b_dev; 2991 uint64_t asize = arc_hdr_size(hdr); 2992 2993 ASSERT(MUTEX_HELD(&dev->l2ad_mtx)); 2994 ASSERT(HDR_HAS_L2HDR(hdr)); 2995 2996 list_remove(&dev->l2ad_buflist, hdr); 2997 2998 ARCSTAT_INCR(arcstat_l2_asize, -asize); 2999 ARCSTAT_INCR(arcstat_l2_size, -HDR_GET_LSIZE(hdr)); 3000 3001 vdev_space_update(dev->l2ad_vdev, -asize, 0, 0); 3002 3003 (void) refcount_remove_many(&dev->l2ad_alloc, asize, hdr); 3004 arc_hdr_clear_flags(hdr, ARC_FLAG_HAS_L2HDR); 3005} 3006 3007static void 3008arc_hdr_destroy(arc_buf_hdr_t *hdr) 3009{ 3010 if (HDR_HAS_L1HDR(hdr)) { 3011 ASSERT(hdr->b_l1hdr.b_buf == NULL || 3012 hdr->b_l1hdr.b_bufcnt > 0); 3013 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 3014 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon); 3015 } 3016 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 3017 ASSERT(!HDR_IN_HASH_TABLE(hdr)); 3018 3019 if (!HDR_EMPTY(hdr)) 3020 buf_discard_identity(hdr); 3021 3022 if (HDR_HAS_L2HDR(hdr)) { 3023 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev; 3024 boolean_t buflist_held = MUTEX_HELD(&dev->l2ad_mtx); 3025 3026 if (!buflist_held) 3027 mutex_enter(&dev->l2ad_mtx); 3028 3029 /* 3030 * Even though we checked this conditional above, we 3031 * need to check this again now that we have the 3032 * l2ad_mtx. This is because we could be racing with 3033 * another thread calling l2arc_evict() which might have 3034 * destroyed this header's L2 portion as we were waiting 3035 * to acquire the l2ad_mtx. If that happens, we don't 3036 * want to re-destroy the header's L2 portion. 3037 */ 3038 if (HDR_HAS_L2HDR(hdr)) { 3039 l2arc_trim(hdr); 3040 arc_hdr_l2hdr_destroy(hdr); 3041 } 3042 3043 if (!buflist_held) 3044 mutex_exit(&dev->l2ad_mtx); 3045 } 3046 3047 if (HDR_HAS_L1HDR(hdr)) { 3048 arc_cksum_free(hdr); 3049 3050 while (hdr->b_l1hdr.b_buf != NULL) 3051 arc_buf_destroy_impl(hdr->b_l1hdr.b_buf, B_TRUE); 3052 3053#ifdef ZFS_DEBUG 3054 if (hdr->b_l1hdr.b_thawed != NULL) { 3055 kmem_free(hdr->b_l1hdr.b_thawed, 1); 3056 hdr->b_l1hdr.b_thawed = NULL; 3057 } 3058#endif 3059 3060 if (hdr->b_l1hdr.b_pdata != NULL) { 3061 arc_hdr_free_pdata(hdr); 3062 } 3063 } 3064 3065 ASSERT3P(hdr->b_hash_next, ==, NULL); 3066 if (HDR_HAS_L1HDR(hdr)) { 3067 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 3068 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL); 3069 kmem_cache_free(hdr_full_cache, hdr); 3070 } else { 3071 kmem_cache_free(hdr_l2only_cache, hdr); 3072 } 3073} 3074 3075void 3076arc_buf_destroy(arc_buf_t *buf, void* tag) 3077{ 3078 arc_buf_hdr_t *hdr = buf->b_hdr; 3079 kmutex_t *hash_lock = HDR_LOCK(hdr); 3080 3081 if (hdr->b_l1hdr.b_state == arc_anon) { 3082 ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1); 3083 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 3084 VERIFY0(remove_reference(hdr, NULL, tag)); 3085 arc_hdr_destroy(hdr); 3086 return; 3087 } 3088 3089 mutex_enter(hash_lock); 3090 ASSERT3P(hdr, ==, buf->b_hdr); 3091 ASSERT(hdr->b_l1hdr.b_bufcnt > 0); 3092 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 3093 ASSERT3P(hdr->b_l1hdr.b_state, !=, arc_anon); 3094 ASSERT3P(buf->b_data, !=, NULL); 3095 3096 (void) remove_reference(hdr, hash_lock, tag); 3097 arc_buf_destroy_impl(buf, B_TRUE); 3098 mutex_exit(hash_lock); 3099} 3100 3101int32_t 3102arc_buf_size(arc_buf_t *buf) 3103{ 3104 return (HDR_GET_LSIZE(buf->b_hdr)); 3105} 3106 3107/* 3108 * Evict the arc_buf_hdr that is provided as a parameter. The resultant 3109 * state of the header is dependent on it's state prior to entering this 3110 * function. The following transitions are possible: 3111 * 3112 * - arc_mru -> arc_mru_ghost 3113 * - arc_mfu -> arc_mfu_ghost 3114 * - arc_mru_ghost -> arc_l2c_only 3115 * - arc_mru_ghost -> deleted 3116 * - arc_mfu_ghost -> arc_l2c_only 3117 * - arc_mfu_ghost -> deleted 3118 */ 3119static int64_t 3120arc_evict_hdr(arc_buf_hdr_t *hdr, kmutex_t *hash_lock) 3121{ 3122 arc_state_t *evicted_state, *state; 3123 int64_t bytes_evicted = 0; 3124 3125 ASSERT(MUTEX_HELD(hash_lock)); 3126 ASSERT(HDR_HAS_L1HDR(hdr)); 3127 3128 state = hdr->b_l1hdr.b_state; 3129 if (GHOST_STATE(state)) { 3130 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 3131 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 3132 3133 /* 3134 * l2arc_write_buffers() relies on a header's L1 portion 3135 * (i.e. its b_pdata field) during its write phase. 3136 * Thus, we cannot push a header onto the arc_l2c_only 3137 * state (removing it's L1 piece) until the header is 3138 * done being written to the l2arc. 3139 */ 3140 if (HDR_HAS_L2HDR(hdr) && HDR_L2_WRITING(hdr)) { 3141 ARCSTAT_BUMP(arcstat_evict_l2_skip); 3142 return (bytes_evicted); 3143 } 3144 3145 ARCSTAT_BUMP(arcstat_deleted); 3146 bytes_evicted += HDR_GET_LSIZE(hdr); 3147 3148 DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, hdr); 3149 3150 ASSERT3P(hdr->b_l1hdr.b_pdata, ==, NULL); 3151 if (HDR_HAS_L2HDR(hdr)) { 3152 ASSERT(hdr->b_l1hdr.b_pdata == NULL); 3153 /* 3154 * This buffer is cached on the 2nd Level ARC; 3155 * don't destroy the header. 3156 */ 3157 arc_change_state(arc_l2c_only, hdr, hash_lock); 3158 /* 3159 * dropping from L1+L2 cached to L2-only, 3160 * realloc to remove the L1 header. 3161 */ 3162 hdr = arc_hdr_realloc(hdr, hdr_full_cache, 3163 hdr_l2only_cache); 3164 } else { 3165 ASSERT(hdr->b_l1hdr.b_pdata == NULL); 3166 arc_change_state(arc_anon, hdr, hash_lock); 3167 arc_hdr_destroy(hdr); 3168 } 3169 return (bytes_evicted); 3170 } 3171 3172 ASSERT(state == arc_mru || state == arc_mfu); 3173 evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost; 3174 3175 /* prefetch buffers have a minimum lifespan */ 3176 if (HDR_IO_IN_PROGRESS(hdr) || 3177 ((hdr->b_flags & (ARC_FLAG_PREFETCH | ARC_FLAG_INDIRECT)) && 3178 ddi_get_lbolt() - hdr->b_l1hdr.b_arc_access < 3179 arc_min_prefetch_lifespan)) { 3180 ARCSTAT_BUMP(arcstat_evict_skip); 3181 return (bytes_evicted); 3182 } 3183 3184 ASSERT0(refcount_count(&hdr->b_l1hdr.b_refcnt)); 3185 while (hdr->b_l1hdr.b_buf) { 3186 arc_buf_t *buf = hdr->b_l1hdr.b_buf; 3187 if (!mutex_tryenter(&buf->b_evict_lock)) { 3188 ARCSTAT_BUMP(arcstat_mutex_miss); 3189 break; 3190 } 3191 if (buf->b_data != NULL) 3192 bytes_evicted += HDR_GET_LSIZE(hdr); 3193 mutex_exit(&buf->b_evict_lock); 3194 arc_buf_destroy_impl(buf, B_TRUE); 3195 } 3196 3197 if (HDR_HAS_L2HDR(hdr)) { 3198 ARCSTAT_INCR(arcstat_evict_l2_cached, HDR_GET_LSIZE(hdr)); 3199 } else { 3200 if (l2arc_write_eligible(hdr->b_spa, hdr)) { 3201 ARCSTAT_INCR(arcstat_evict_l2_eligible, 3202 HDR_GET_LSIZE(hdr)); 3203 } else { 3204 ARCSTAT_INCR(arcstat_evict_l2_ineligible, 3205 HDR_GET_LSIZE(hdr)); 3206 } 3207 } 3208 3209 if (hdr->b_l1hdr.b_bufcnt == 0) { 3210 arc_cksum_free(hdr); 3211 3212 bytes_evicted += arc_hdr_size(hdr); 3213 3214 /* 3215 * If this hdr is being evicted and has a compressed 3216 * buffer then we discard it here before we change states. 3217 * This ensures that the accounting is updated correctly 3218 * in arc_free_data_buf(). 3219 */ 3220 arc_hdr_free_pdata(hdr); 3221 3222 arc_change_state(evicted_state, hdr, hash_lock); 3223 ASSERT(HDR_IN_HASH_TABLE(hdr)); 3224 arc_hdr_set_flags(hdr, ARC_FLAG_IN_HASH_TABLE); 3225 DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, hdr); 3226 } 3227 3228 return (bytes_evicted); 3229} 3230 3231static uint64_t 3232arc_evict_state_impl(multilist_t *ml, int idx, arc_buf_hdr_t *marker, 3233 uint64_t spa, int64_t bytes) 3234{ 3235 multilist_sublist_t *mls; 3236 uint64_t bytes_evicted = 0; 3237 arc_buf_hdr_t *hdr; 3238 kmutex_t *hash_lock; 3239 int evict_count = 0; 3240 3241 ASSERT3P(marker, !=, NULL); 3242 IMPLY(bytes < 0, bytes == ARC_EVICT_ALL); 3243 3244 mls = multilist_sublist_lock(ml, idx); 3245 3246 for (hdr = multilist_sublist_prev(mls, marker); hdr != NULL; 3247 hdr = multilist_sublist_prev(mls, marker)) { 3248 if ((bytes != ARC_EVICT_ALL && bytes_evicted >= bytes) || 3249 (evict_count >= zfs_arc_evict_batch_limit)) 3250 break; 3251 3252 /* 3253 * To keep our iteration location, move the marker 3254 * forward. Since we're not holding hdr's hash lock, we 3255 * must be very careful and not remove 'hdr' from the 3256 * sublist. Otherwise, other consumers might mistake the 3257 * 'hdr' as not being on a sublist when they call the 3258 * multilist_link_active() function (they all rely on 3259 * the hash lock protecting concurrent insertions and 3260 * removals). multilist_sublist_move_forward() was 3261 * specifically implemented to ensure this is the case 3262 * (only 'marker' will be removed and re-inserted). 3263 */ 3264 multilist_sublist_move_forward(mls, marker); 3265 3266 /* 3267 * The only case where the b_spa field should ever be 3268 * zero, is the marker headers inserted by 3269 * arc_evict_state(). It's possible for multiple threads 3270 * to be calling arc_evict_state() concurrently (e.g. 3271 * dsl_pool_close() and zio_inject_fault()), so we must 3272 * skip any markers we see from these other threads. 3273 */ 3274 if (hdr->b_spa == 0) 3275 continue; 3276 3277 /* we're only interested in evicting buffers of a certain spa */ 3278 if (spa != 0 && hdr->b_spa != spa) { 3279 ARCSTAT_BUMP(arcstat_evict_skip); 3280 continue; 3281 } 3282 3283 hash_lock = HDR_LOCK(hdr); 3284 3285 /* 3286 * We aren't calling this function from any code path 3287 * that would already be holding a hash lock, so we're 3288 * asserting on this assumption to be defensive in case 3289 * this ever changes. Without this check, it would be 3290 * possible to incorrectly increment arcstat_mutex_miss 3291 * below (e.g. if the code changed such that we called 3292 * this function with a hash lock held). 3293 */ 3294 ASSERT(!MUTEX_HELD(hash_lock)); 3295 3296 if (mutex_tryenter(hash_lock)) { 3297 uint64_t evicted = arc_evict_hdr(hdr, hash_lock); 3298 mutex_exit(hash_lock); 3299 3300 bytes_evicted += evicted; 3301 3302 /* 3303 * If evicted is zero, arc_evict_hdr() must have 3304 * decided to skip this header, don't increment 3305 * evict_count in this case. 3306 */ 3307 if (evicted != 0) 3308 evict_count++; 3309 3310 /* 3311 * If arc_size isn't overflowing, signal any 3312 * threads that might happen to be waiting. 3313 * 3314 * For each header evicted, we wake up a single 3315 * thread. If we used cv_broadcast, we could 3316 * wake up "too many" threads causing arc_size 3317 * to significantly overflow arc_c; since 3318 * arc_get_data_buf() doesn't check for overflow 3319 * when it's woken up (it doesn't because it's 3320 * possible for the ARC to be overflowing while 3321 * full of un-evictable buffers, and the 3322 * function should proceed in this case). 3323 * 3324 * If threads are left sleeping, due to not 3325 * using cv_broadcast, they will be woken up 3326 * just before arc_reclaim_thread() sleeps. 3327 */ 3328 mutex_enter(&arc_reclaim_lock); 3329 if (!arc_is_overflowing()) 3330 cv_signal(&arc_reclaim_waiters_cv); 3331 mutex_exit(&arc_reclaim_lock); 3332 } else { 3333 ARCSTAT_BUMP(arcstat_mutex_miss); 3334 } 3335 } 3336 3337 multilist_sublist_unlock(mls); 3338 3339 return (bytes_evicted); 3340} 3341 3342/* 3343 * Evict buffers from the given arc state, until we've removed the 3344 * specified number of bytes. Move the removed buffers to the 3345 * appropriate evict state. 3346 * 3347 * This function makes a "best effort". It skips over any buffers 3348 * it can't get a hash_lock on, and so, may not catch all candidates. 3349 * It may also return without evicting as much space as requested. 3350 * 3351 * If bytes is specified using the special value ARC_EVICT_ALL, this 3352 * will evict all available (i.e. unlocked and evictable) buffers from 3353 * the given arc state; which is used by arc_flush(). 3354 */ 3355static uint64_t 3356arc_evict_state(arc_state_t *state, uint64_t spa, int64_t bytes, 3357 arc_buf_contents_t type) 3358{ 3359 uint64_t total_evicted = 0; 3360 multilist_t *ml = &state->arcs_list[type]; 3361 int num_sublists; 3362 arc_buf_hdr_t **markers; 3363 3364 IMPLY(bytes < 0, bytes == ARC_EVICT_ALL); 3365 3366 num_sublists = multilist_get_num_sublists(ml); 3367 3368 /* 3369 * If we've tried to evict from each sublist, made some 3370 * progress, but still have not hit the target number of bytes 3371 * to evict, we want to keep trying. The markers allow us to 3372 * pick up where we left off for each individual sublist, rather 3373 * than starting from the tail each time. 3374 */ 3375 markers = kmem_zalloc(sizeof (*markers) * num_sublists, KM_SLEEP); 3376 for (int i = 0; i < num_sublists; i++) { 3377 markers[i] = kmem_cache_alloc(hdr_full_cache, KM_SLEEP); 3378 3379 /* 3380 * A b_spa of 0 is used to indicate that this header is 3381 * a marker. This fact is used in arc_adjust_type() and 3382 * arc_evict_state_impl(). 3383 */ 3384 markers[i]->b_spa = 0; 3385 3386 multilist_sublist_t *mls = multilist_sublist_lock(ml, i); 3387 multilist_sublist_insert_tail(mls, markers[i]); 3388 multilist_sublist_unlock(mls); 3389 } 3390 3391 /* 3392 * While we haven't hit our target number of bytes to evict, or 3393 * we're evicting all available buffers. 3394 */ 3395 while (total_evicted < bytes || bytes == ARC_EVICT_ALL) { 3396 /* 3397 * Start eviction using a randomly selected sublist, 3398 * this is to try and evenly balance eviction across all 3399 * sublists. Always starting at the same sublist 3400 * (e.g. index 0) would cause evictions to favor certain 3401 * sublists over others. 3402 */ 3403 int sublist_idx = multilist_get_random_index(ml); 3404 uint64_t scan_evicted = 0; 3405 3406 for (int i = 0; i < num_sublists; i++) { 3407 uint64_t bytes_remaining; 3408 uint64_t bytes_evicted; 3409 3410 if (bytes == ARC_EVICT_ALL) 3411 bytes_remaining = ARC_EVICT_ALL; 3412 else if (total_evicted < bytes) 3413 bytes_remaining = bytes - total_evicted; 3414 else 3415 break; 3416 3417 bytes_evicted = arc_evict_state_impl(ml, sublist_idx, 3418 markers[sublist_idx], spa, bytes_remaining); 3419 3420 scan_evicted += bytes_evicted; 3421 total_evicted += bytes_evicted; 3422 3423 /* we've reached the end, wrap to the beginning */ 3424 if (++sublist_idx >= num_sublists) 3425 sublist_idx = 0; 3426 } 3427 3428 /* 3429 * If we didn't evict anything during this scan, we have 3430 * no reason to believe we'll evict more during another 3431 * scan, so break the loop. 3432 */ 3433 if (scan_evicted == 0) { 3434 /* This isn't possible, let's make that obvious */ 3435 ASSERT3S(bytes, !=, 0); 3436 3437 /* 3438 * When bytes is ARC_EVICT_ALL, the only way to 3439 * break the loop is when scan_evicted is zero. 3440 * In that case, we actually have evicted enough, 3441 * so we don't want to increment the kstat. 3442 */ 3443 if (bytes != ARC_EVICT_ALL) { 3444 ASSERT3S(total_evicted, <, bytes); 3445 ARCSTAT_BUMP(arcstat_evict_not_enough); 3446 } 3447 3448 break; 3449 } 3450 } 3451 3452 for (int i = 0; i < num_sublists; i++) { 3453 multilist_sublist_t *mls = multilist_sublist_lock(ml, i); 3454 multilist_sublist_remove(mls, markers[i]); 3455 multilist_sublist_unlock(mls); 3456 3457 kmem_cache_free(hdr_full_cache, markers[i]); 3458 } 3459 kmem_free(markers, sizeof (*markers) * num_sublists); 3460 3461 return (total_evicted); 3462} 3463 3464/* 3465 * Flush all "evictable" data of the given type from the arc state 3466 * specified. This will not evict any "active" buffers (i.e. referenced). 3467 * 3468 * When 'retry' is set to B_FALSE, the function will make a single pass 3469 * over the state and evict any buffers that it can. Since it doesn't 3470 * continually retry the eviction, it might end up leaving some buffers 3471 * in the ARC due to lock misses. 3472 * 3473 * When 'retry' is set to B_TRUE, the function will continually retry the 3474 * eviction until *all* evictable buffers have been removed from the 3475 * state. As a result, if concurrent insertions into the state are 3476 * allowed (e.g. if the ARC isn't shutting down), this function might 3477 * wind up in an infinite loop, continually trying to evict buffers. 3478 */ 3479static uint64_t 3480arc_flush_state(arc_state_t *state, uint64_t spa, arc_buf_contents_t type, 3481 boolean_t retry) 3482{ 3483 uint64_t evicted = 0; 3484 3485 while (refcount_count(&state->arcs_esize[type]) != 0) { 3486 evicted += arc_evict_state(state, spa, ARC_EVICT_ALL, type); 3487 3488 if (!retry) 3489 break; 3490 } 3491 3492 return (evicted); 3493} 3494 3495/* 3496 * Evict the specified number of bytes from the state specified, 3497 * restricting eviction to the spa and type given. This function 3498 * prevents us from trying to evict more from a state's list than 3499 * is "evictable", and to skip evicting altogether when passed a 3500 * negative value for "bytes". In contrast, arc_evict_state() will 3501 * evict everything it can, when passed a negative value for "bytes". 3502 */ 3503static uint64_t 3504arc_adjust_impl(arc_state_t *state, uint64_t spa, int64_t bytes, 3505 arc_buf_contents_t type) 3506{ 3507 int64_t delta; 3508 3509 if (bytes > 0 && refcount_count(&state->arcs_esize[type]) > 0) { 3510 delta = MIN(refcount_count(&state->arcs_esize[type]), bytes); 3511 return (arc_evict_state(state, spa, delta, type)); 3512 } 3513 3514 return (0); 3515} 3516 3517/* 3518 * Evict metadata buffers from the cache, such that arc_meta_used is 3519 * capped by the arc_meta_limit tunable. 3520 */ 3521static uint64_t 3522arc_adjust_meta(void) 3523{ 3524 uint64_t total_evicted = 0; 3525 int64_t target; 3526 3527 /* 3528 * If we're over the meta limit, we want to evict enough 3529 * metadata to get back under the meta limit. We don't want to 3530 * evict so much that we drop the MRU below arc_p, though. If 3531 * we're over the meta limit more than we're over arc_p, we 3532 * evict some from the MRU here, and some from the MFU below. 3533 */ 3534 target = MIN((int64_t)(arc_meta_used - arc_meta_limit), 3535 (int64_t)(refcount_count(&arc_anon->arcs_size) + 3536 refcount_count(&arc_mru->arcs_size) - arc_p)); 3537 3538 total_evicted += arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA); 3539 3540 /* 3541 * Similar to the above, we want to evict enough bytes to get us 3542 * below the meta limit, but not so much as to drop us below the 3543 * space alloted to the MFU (which is defined as arc_c - arc_p). 3544 */ 3545 target = MIN((int64_t)(arc_meta_used - arc_meta_limit), 3546 (int64_t)(refcount_count(&arc_mfu->arcs_size) - (arc_c - arc_p))); 3547 3548 total_evicted += arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA); 3549 3550 return (total_evicted); 3551} 3552 3553/* 3554 * Return the type of the oldest buffer in the given arc state 3555 * 3556 * This function will select a random sublist of type ARC_BUFC_DATA and 3557 * a random sublist of type ARC_BUFC_METADATA. The tail of each sublist 3558 * is compared, and the type which contains the "older" buffer will be 3559 * returned. 3560 */ 3561static arc_buf_contents_t 3562arc_adjust_type(arc_state_t *state) 3563{ 3564 multilist_t *data_ml = &state->arcs_list[ARC_BUFC_DATA]; 3565 multilist_t *meta_ml = &state->arcs_list[ARC_BUFC_METADATA]; 3566 int data_idx = multilist_get_random_index(data_ml); 3567 int meta_idx = multilist_get_random_index(meta_ml); 3568 multilist_sublist_t *data_mls; 3569 multilist_sublist_t *meta_mls; 3570 arc_buf_contents_t type; 3571 arc_buf_hdr_t *data_hdr; 3572 arc_buf_hdr_t *meta_hdr; 3573 3574 /* 3575 * We keep the sublist lock until we're finished, to prevent 3576 * the headers from being destroyed via arc_evict_state(). 3577 */ 3578 data_mls = multilist_sublist_lock(data_ml, data_idx); 3579 meta_mls = multilist_sublist_lock(meta_ml, meta_idx); 3580 3581 /* 3582 * These two loops are to ensure we skip any markers that 3583 * might be at the tail of the lists due to arc_evict_state(). 3584 */ 3585 3586 for (data_hdr = multilist_sublist_tail(data_mls); data_hdr != NULL; 3587 data_hdr = multilist_sublist_prev(data_mls, data_hdr)) { 3588 if (data_hdr->b_spa != 0) 3589 break; 3590 } 3591 3592 for (meta_hdr = multilist_sublist_tail(meta_mls); meta_hdr != NULL; 3593 meta_hdr = multilist_sublist_prev(meta_mls, meta_hdr)) { 3594 if (meta_hdr->b_spa != 0) 3595 break; 3596 } 3597 3598 if (data_hdr == NULL && meta_hdr == NULL) { 3599 type = ARC_BUFC_DATA; 3600 } else if (data_hdr == NULL) { 3601 ASSERT3P(meta_hdr, !=, NULL); 3602 type = ARC_BUFC_METADATA; 3603 } else if (meta_hdr == NULL) { 3604 ASSERT3P(data_hdr, !=, NULL); 3605 type = ARC_BUFC_DATA; 3606 } else { 3607 ASSERT3P(data_hdr, !=, NULL); 3608 ASSERT3P(meta_hdr, !=, NULL); 3609 3610 /* The headers can't be on the sublist without an L1 header */ 3611 ASSERT(HDR_HAS_L1HDR(data_hdr)); 3612 ASSERT(HDR_HAS_L1HDR(meta_hdr)); 3613 3614 if (data_hdr->b_l1hdr.b_arc_access < 3615 meta_hdr->b_l1hdr.b_arc_access) { 3616 type = ARC_BUFC_DATA; 3617 } else { 3618 type = ARC_BUFC_METADATA; 3619 } 3620 } 3621 3622 multilist_sublist_unlock(meta_mls); 3623 multilist_sublist_unlock(data_mls); 3624 3625 return (type); 3626} 3627 3628/* 3629 * Evict buffers from the cache, such that arc_size is capped by arc_c. 3630 */ 3631static uint64_t 3632arc_adjust(void) 3633{ 3634 uint64_t total_evicted = 0; 3635 uint64_t bytes; 3636 int64_t target; 3637 3638 /* 3639 * If we're over arc_meta_limit, we want to correct that before 3640 * potentially evicting data buffers below. 3641 */ 3642 total_evicted += arc_adjust_meta(); 3643 3644 /* 3645 * Adjust MRU size 3646 * 3647 * If we're over the target cache size, we want to evict enough 3648 * from the list to get back to our target size. We don't want 3649 * to evict too much from the MRU, such that it drops below 3650 * arc_p. So, if we're over our target cache size more than 3651 * the MRU is over arc_p, we'll evict enough to get back to 3652 * arc_p here, and then evict more from the MFU below. 3653 */ 3654 target = MIN((int64_t)(arc_size - arc_c), 3655 (int64_t)(refcount_count(&arc_anon->arcs_size) + 3656 refcount_count(&arc_mru->arcs_size) + arc_meta_used - arc_p)); 3657 3658 /* 3659 * If we're below arc_meta_min, always prefer to evict data. 3660 * Otherwise, try to satisfy the requested number of bytes to 3661 * evict from the type which contains older buffers; in an 3662 * effort to keep newer buffers in the cache regardless of their 3663 * type. If we cannot satisfy the number of bytes from this 3664 * type, spill over into the next type. 3665 */ 3666 if (arc_adjust_type(arc_mru) == ARC_BUFC_METADATA && 3667 arc_meta_used > arc_meta_min) { 3668 bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA); 3669 total_evicted += bytes; 3670 3671 /* 3672 * If we couldn't evict our target number of bytes from 3673 * metadata, we try to get the rest from data. 3674 */ 3675 target -= bytes; 3676 3677 total_evicted += 3678 arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA); 3679 } else { 3680 bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA); 3681 total_evicted += bytes; 3682 3683 /* 3684 * If we couldn't evict our target number of bytes from 3685 * data, we try to get the rest from metadata. 3686 */ 3687 target -= bytes; 3688 3689 total_evicted += 3690 arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA); 3691 } 3692 3693 /* 3694 * Adjust MFU size 3695 * 3696 * Now that we've tried to evict enough from the MRU to get its 3697 * size back to arc_p, if we're still above the target cache 3698 * size, we evict the rest from the MFU. 3699 */ 3700 target = arc_size - arc_c; 3701 3702 if (arc_adjust_type(arc_mfu) == ARC_BUFC_METADATA && 3703 arc_meta_used > arc_meta_min) { 3704 bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA); 3705 total_evicted += bytes; 3706 3707 /* 3708 * If we couldn't evict our target number of bytes from 3709 * metadata, we try to get the rest from data. 3710 */ 3711 target -= bytes; 3712 3713 total_evicted += 3714 arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA); 3715 } else { 3716 bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA); 3717 total_evicted += bytes; 3718 3719 /* 3720 * If we couldn't evict our target number of bytes from 3721 * data, we try to get the rest from data. 3722 */ 3723 target -= bytes; 3724 3725 total_evicted += 3726 arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA); 3727 } 3728 3729 /* 3730 * Adjust ghost lists 3731 * 3732 * In addition to the above, the ARC also defines target values 3733 * for the ghost lists. The sum of the mru list and mru ghost 3734 * list should never exceed the target size of the cache, and 3735 * the sum of the mru list, mfu list, mru ghost list, and mfu 3736 * ghost list should never exceed twice the target size of the 3737 * cache. The following logic enforces these limits on the ghost 3738 * caches, and evicts from them as needed. 3739 */ 3740 target = refcount_count(&arc_mru->arcs_size) + 3741 refcount_count(&arc_mru_ghost->arcs_size) - arc_c; 3742 3743 bytes = arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_DATA); 3744 total_evicted += bytes; 3745 3746 target -= bytes; 3747 3748 total_evicted += 3749 arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_METADATA); 3750 3751 /* 3752 * We assume the sum of the mru list and mfu list is less than 3753 * or equal to arc_c (we enforced this above), which means we 3754 * can use the simpler of the two equations below: 3755 * 3756 * mru + mfu + mru ghost + mfu ghost <= 2 * arc_c 3757 * mru ghost + mfu ghost <= arc_c 3758 */ 3759 target = refcount_count(&arc_mru_ghost->arcs_size) + 3760 refcount_count(&arc_mfu_ghost->arcs_size) - arc_c; 3761 3762 bytes = arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_DATA); 3763 total_evicted += bytes; 3764 3765 target -= bytes; 3766 3767 total_evicted += 3768 arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_METADATA); 3769 3770 return (total_evicted); 3771} 3772 3773void 3774arc_flush(spa_t *spa, boolean_t retry) 3775{ 3776 uint64_t guid = 0; 3777 3778 /* 3779 * If retry is B_TRUE, a spa must not be specified since we have 3780 * no good way to determine if all of a spa's buffers have been 3781 * evicted from an arc state. 3782 */ 3783 ASSERT(!retry || spa == 0); 3784 3785 if (spa != NULL) 3786 guid = spa_load_guid(spa); 3787 3788 (void) arc_flush_state(arc_mru, guid, ARC_BUFC_DATA, retry); 3789 (void) arc_flush_state(arc_mru, guid, ARC_BUFC_METADATA, retry); 3790 3791 (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_DATA, retry); 3792 (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_METADATA, retry); 3793 3794 (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_DATA, retry); 3795 (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_METADATA, retry); 3796 3797 (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_DATA, retry); 3798 (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_METADATA, retry); 3799} 3800 3801void 3802arc_shrink(int64_t to_free) 3803{ 3804 if (arc_c > arc_c_min) { 3805 DTRACE_PROBE4(arc__shrink, uint64_t, arc_c, uint64_t, 3806 arc_c_min, uint64_t, arc_p, uint64_t, to_free); 3807 if (arc_c > arc_c_min + to_free) 3808 atomic_add_64(&arc_c, -to_free); 3809 else 3810 arc_c = arc_c_min; 3811 3812 atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift)); 3813 if (arc_c > arc_size) 3814 arc_c = MAX(arc_size, arc_c_min); 3815 if (arc_p > arc_c) 3816 arc_p = (arc_c >> 1); 3817 3818 DTRACE_PROBE2(arc__shrunk, uint64_t, arc_c, uint64_t, 3819 arc_p); 3820 3821 ASSERT(arc_c >= arc_c_min); 3822 ASSERT((int64_t)arc_p >= 0); 3823 } 3824 3825 if (arc_size > arc_c) { 3826 DTRACE_PROBE2(arc__shrink_adjust, uint64_t, arc_size, 3827 uint64_t, arc_c); 3828 (void) arc_adjust(); 3829 } 3830} 3831 3832static long needfree = 0; 3833 3834typedef enum free_memory_reason_t { 3835 FMR_UNKNOWN, 3836 FMR_NEEDFREE, 3837 FMR_LOTSFREE, 3838 FMR_SWAPFS_MINFREE, 3839 FMR_PAGES_PP_MAXIMUM, 3840 FMR_HEAP_ARENA, 3841 FMR_ZIO_ARENA, 3842 FMR_ZIO_FRAG, 3843} free_memory_reason_t; 3844 3845int64_t last_free_memory; 3846free_memory_reason_t last_free_reason; 3847 3848/* 3849 * Additional reserve of pages for pp_reserve. 3850 */ 3851int64_t arc_pages_pp_reserve = 64; 3852 3853/* 3854 * Additional reserve of pages for swapfs. 3855 */ 3856int64_t arc_swapfs_reserve = 64; 3857 3858/* 3859 * Return the amount of memory that can be consumed before reclaim will be 3860 * needed. Positive if there is sufficient free memory, negative indicates 3861 * the amount of memory that needs to be freed up. 3862 */ 3863static int64_t 3864arc_available_memory(void) 3865{ 3866 int64_t lowest = INT64_MAX; 3867 int64_t n; 3868 free_memory_reason_t r = FMR_UNKNOWN; 3869 3870#ifdef _KERNEL 3871 if (needfree > 0) { 3872 n = PAGESIZE * (-needfree); 3873 if (n < lowest) { 3874 lowest = n; 3875 r = FMR_NEEDFREE; 3876 } 3877 } 3878 3879 /* 3880 * Cooperate with pagedaemon when it's time for it to scan 3881 * and reclaim some pages. 3882 */ 3883 n = PAGESIZE * ((int64_t)freemem - zfs_arc_free_target); 3884 if (n < lowest) { 3885 lowest = n; 3886 r = FMR_LOTSFREE; 3887 } 3888 3889#ifdef illumos 3890 /* 3891 * check that we're out of range of the pageout scanner. It starts to 3892 * schedule paging if freemem is less than lotsfree and needfree. 3893 * lotsfree is the high-water mark for pageout, and needfree is the 3894 * number of needed free pages. We add extra pages here to make sure 3895 * the scanner doesn't start up while we're freeing memory. 3896 */ 3897 n = PAGESIZE * (freemem - lotsfree - needfree - desfree); 3898 if (n < lowest) { 3899 lowest = n; 3900 r = FMR_LOTSFREE; 3901 } 3902 3903 /* 3904 * check to make sure that swapfs has enough space so that anon 3905 * reservations can still succeed. anon_resvmem() checks that the 3906 * availrmem is greater than swapfs_minfree, and the number of reserved 3907 * swap pages. We also add a bit of extra here just to prevent 3908 * circumstances from getting really dire. 3909 */ 3910 n = PAGESIZE * (availrmem - swapfs_minfree - swapfs_reserve - 3911 desfree - arc_swapfs_reserve); 3912 if (n < lowest) { 3913 lowest = n; 3914 r = FMR_SWAPFS_MINFREE; 3915 } 3916 3917 3918 /* 3919 * Check that we have enough availrmem that memory locking (e.g., via 3920 * mlock(3C) or memcntl(2)) can still succeed. (pages_pp_maximum 3921 * stores the number of pages that cannot be locked; when availrmem 3922 * drops below pages_pp_maximum, page locking mechanisms such as 3923 * page_pp_lock() will fail.) 3924 */ 3925 n = PAGESIZE * (availrmem - pages_pp_maximum - 3926 arc_pages_pp_reserve); 3927 if (n < lowest) { 3928 lowest = n; 3929 r = FMR_PAGES_PP_MAXIMUM; 3930 } 3931 3932#endif /* illumos */ 3933#if defined(__i386) || !defined(UMA_MD_SMALL_ALLOC) 3934 /* 3935 * If we're on an i386 platform, it's possible that we'll exhaust the 3936 * kernel heap space before we ever run out of available physical 3937 * memory. Most checks of the size of the heap_area compare against 3938 * tune.t_minarmem, which is the minimum available real memory that we 3939 * can have in the system. However, this is generally fixed at 25 pages 3940 * which is so low that it's useless. In this comparison, we seek to 3941 * calculate the total heap-size, and reclaim if more than 3/4ths of the 3942 * heap is allocated. (Or, in the calculation, if less than 1/4th is 3943 * free) 3944 */ 3945 n = (int64_t)vmem_size(heap_arena, VMEM_FREE) - 3946 (vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC) >> 2); 3947 if (n < lowest) { 3948 lowest = n; 3949 r = FMR_HEAP_ARENA; 3950 } 3951#define zio_arena NULL 3952#else 3953#define zio_arena heap_arena 3954#endif 3955 3956 /* 3957 * If zio data pages are being allocated out of a separate heap segment, 3958 * then enforce that the size of available vmem for this arena remains 3959 * above about 1/16th free. 3960 * 3961 * Note: The 1/16th arena free requirement was put in place 3962 * to aggressively evict memory from the arc in order to avoid 3963 * memory fragmentation issues. 3964 */ 3965 if (zio_arena != NULL) { 3966 n = (int64_t)vmem_size(zio_arena, VMEM_FREE) - 3967 (vmem_size(zio_arena, VMEM_ALLOC) >> 4); 3968 if (n < lowest) { 3969 lowest = n; 3970 r = FMR_ZIO_ARENA; 3971 } 3972 } 3973 3974 /* 3975 * Above limits know nothing about real level of KVA fragmentation. 3976 * Start aggressive reclamation if too little sequential KVA left. 3977 */ 3978 if (lowest > 0) { 3979 n = (vmem_size(heap_arena, VMEM_MAXFREE) < zfs_max_recordsize) ? 3980 -((int64_t)vmem_size(heap_arena, VMEM_ALLOC) >> 4) : 3981 INT64_MAX; 3982 if (n < lowest) { 3983 lowest = n; 3984 r = FMR_ZIO_FRAG; 3985 } 3986 } 3987 3988#else /* _KERNEL */ 3989 /* Every 100 calls, free a small amount */ 3990 if (spa_get_random(100) == 0) 3991 lowest = -1024; 3992#endif /* _KERNEL */ 3993 3994 last_free_memory = lowest; 3995 last_free_reason = r; 3996 DTRACE_PROBE2(arc__available_memory, int64_t, lowest, int, r); 3997 return (lowest); 3998} 3999 4000 4001/* 4002 * Determine if the system is under memory pressure and is asking 4003 * to reclaim memory. A return value of B_TRUE indicates that the system 4004 * is under memory pressure and that the arc should adjust accordingly. 4005 */ 4006static boolean_t 4007arc_reclaim_needed(void) 4008{ 4009 return (arc_available_memory() < 0); 4010} 4011 4012extern kmem_cache_t *zio_buf_cache[]; 4013extern kmem_cache_t *zio_data_buf_cache[]; 4014extern kmem_cache_t *range_seg_cache; 4015 4016static __noinline void 4017arc_kmem_reap_now(void) 4018{ 4019 size_t i; 4020 kmem_cache_t *prev_cache = NULL; 4021 kmem_cache_t *prev_data_cache = NULL; 4022 4023 DTRACE_PROBE(arc__kmem_reap_start); 4024#ifdef _KERNEL 4025 if (arc_meta_used >= arc_meta_limit) { 4026 /* 4027 * We are exceeding our meta-data cache limit. 4028 * Purge some DNLC entries to release holds on meta-data. 4029 */ 4030 dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent); 4031 } 4032#if defined(__i386) 4033 /* 4034 * Reclaim unused memory from all kmem caches. 4035 */ 4036 kmem_reap(); 4037#endif 4038#endif 4039 4040 for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) { 4041 if (zio_buf_cache[i] != prev_cache) { 4042 prev_cache = zio_buf_cache[i]; 4043 kmem_cache_reap_now(zio_buf_cache[i]); 4044 } 4045 if (zio_data_buf_cache[i] != prev_data_cache) { 4046 prev_data_cache = zio_data_buf_cache[i]; 4047 kmem_cache_reap_now(zio_data_buf_cache[i]); 4048 } 4049 } 4050 kmem_cache_reap_now(buf_cache); 4051 kmem_cache_reap_now(hdr_full_cache); 4052 kmem_cache_reap_now(hdr_l2only_cache); 4053 kmem_cache_reap_now(range_seg_cache); 4054 4055#ifdef illumos 4056 if (zio_arena != NULL) { 4057 /* 4058 * Ask the vmem arena to reclaim unused memory from its 4059 * quantum caches. 4060 */ 4061 vmem_qcache_reap(zio_arena); 4062 } 4063#endif 4064 DTRACE_PROBE(arc__kmem_reap_end); 4065} 4066 4067/* 4068 * Threads can block in arc_get_data_buf() waiting for this thread to evict 4069 * enough data and signal them to proceed. When this happens, the threads in 4070 * arc_get_data_buf() are sleeping while holding the hash lock for their 4071 * particular arc header. Thus, we must be careful to never sleep on a 4072 * hash lock in this thread. This is to prevent the following deadlock: 4073 * 4074 * - Thread A sleeps on CV in arc_get_data_buf() holding hash lock "L", 4075 * waiting for the reclaim thread to signal it. 4076 * 4077 * - arc_reclaim_thread() tries to acquire hash lock "L" using mutex_enter, 4078 * fails, and goes to sleep forever. 4079 * 4080 * This possible deadlock is avoided by always acquiring a hash lock 4081 * using mutex_tryenter() from arc_reclaim_thread(). 4082 */ 4083static void 4084arc_reclaim_thread(void *dummy __unused) 4085{ 4086 hrtime_t growtime = 0; 4087 callb_cpr_t cpr; 4088 4089 CALLB_CPR_INIT(&cpr, &arc_reclaim_lock, callb_generic_cpr, FTAG); 4090 4091 mutex_enter(&arc_reclaim_lock); 4092 while (!arc_reclaim_thread_exit) { 4093 uint64_t evicted = 0; 4094 4095 /* 4096 * This is necessary in order for the mdb ::arc dcmd to 4097 * show up to date information. Since the ::arc command 4098 * does not call the kstat's update function, without 4099 * this call, the command may show stale stats for the 4100 * anon, mru, mru_ghost, mfu, and mfu_ghost lists. Even 4101 * with this change, the data might be up to 1 second 4102 * out of date; but that should suffice. The arc_state_t 4103 * structures can be queried directly if more accurate 4104 * information is needed. 4105 */ 4106 if (arc_ksp != NULL) 4107 arc_ksp->ks_update(arc_ksp, KSTAT_READ); 4108 4109 mutex_exit(&arc_reclaim_lock); 4110 4111 /* 4112 * We call arc_adjust() before (possibly) calling 4113 * arc_kmem_reap_now(), so that we can wake up 4114 * arc_get_data_buf() sooner. 4115 */ 4116 evicted = arc_adjust(); 4117 4118 int64_t free_memory = arc_available_memory(); 4119 if (free_memory < 0) { 4120 4121 arc_no_grow = B_TRUE; 4122 arc_warm = B_TRUE; 4123 4124 /* 4125 * Wait at least zfs_grow_retry (default 60) seconds 4126 * before considering growing. 4127 */ 4128 growtime = gethrtime() + SEC2NSEC(arc_grow_retry); 4129 4130 arc_kmem_reap_now(); 4131 4132 /* 4133 * If we are still low on memory, shrink the ARC 4134 * so that we have arc_shrink_min free space. 4135 */ 4136 free_memory = arc_available_memory(); 4137 4138 int64_t to_free = 4139 (arc_c >> arc_shrink_shift) - free_memory; 4140 if (to_free > 0) { 4141#ifdef _KERNEL 4142 to_free = MAX(to_free, ptob(needfree)); 4143#endif 4144 arc_shrink(to_free); 4145 } 4146 } else if (free_memory < arc_c >> arc_no_grow_shift) { 4147 arc_no_grow = B_TRUE; 4148 } else if (gethrtime() >= growtime) { 4149 arc_no_grow = B_FALSE; 4150 } 4151 4152 mutex_enter(&arc_reclaim_lock); 4153 4154 /* 4155 * If evicted is zero, we couldn't evict anything via 4156 * arc_adjust(). This could be due to hash lock 4157 * collisions, but more likely due to the majority of 4158 * arc buffers being unevictable. Therefore, even if 4159 * arc_size is above arc_c, another pass is unlikely to 4160 * be helpful and could potentially cause us to enter an 4161 * infinite loop. 4162 */ 4163 if (arc_size <= arc_c || evicted == 0) { 4164#ifdef _KERNEL 4165 needfree = 0; 4166#endif 4167 /* 4168 * We're either no longer overflowing, or we 4169 * can't evict anything more, so we should wake 4170 * up any threads before we go to sleep. 4171 */ 4172 cv_broadcast(&arc_reclaim_waiters_cv); 4173 4174 /* 4175 * Block until signaled, or after one second (we 4176 * might need to perform arc_kmem_reap_now() 4177 * even if we aren't being signalled) 4178 */ 4179 CALLB_CPR_SAFE_BEGIN(&cpr); 4180 (void) cv_timedwait_hires(&arc_reclaim_thread_cv, 4181 &arc_reclaim_lock, SEC2NSEC(1), MSEC2NSEC(1), 0); 4182 CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_lock); 4183 } 4184 } 4185 4186 arc_reclaim_thread_exit = B_FALSE; 4187 cv_broadcast(&arc_reclaim_thread_cv); 4188 CALLB_CPR_EXIT(&cpr); /* drops arc_reclaim_lock */ 4189 thread_exit(); 4190} 4191 4192/* 4193 * Adapt arc info given the number of bytes we are trying to add and 4194 * the state that we are comming from. This function is only called 4195 * when we are adding new content to the cache. 4196 */ 4197static void 4198arc_adapt(int bytes, arc_state_t *state) 4199{ 4200 int mult; 4201 uint64_t arc_p_min = (arc_c >> arc_p_min_shift); 4202 int64_t mrug_size = refcount_count(&arc_mru_ghost->arcs_size); 4203 int64_t mfug_size = refcount_count(&arc_mfu_ghost->arcs_size); 4204 4205 if (state == arc_l2c_only) 4206 return; 4207 4208 ASSERT(bytes > 0); 4209 /* 4210 * Adapt the target size of the MRU list: 4211 * - if we just hit in the MRU ghost list, then increase 4212 * the target size of the MRU list. 4213 * - if we just hit in the MFU ghost list, then increase 4214 * the target size of the MFU list by decreasing the 4215 * target size of the MRU list. 4216 */ 4217 if (state == arc_mru_ghost) { 4218 mult = (mrug_size >= mfug_size) ? 1 : (mfug_size / mrug_size); 4219 mult = MIN(mult, 10); /* avoid wild arc_p adjustment */ 4220 4221 arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult); 4222 } else if (state == arc_mfu_ghost) { 4223 uint64_t delta; 4224 4225 mult = (mfug_size >= mrug_size) ? 1 : (mrug_size / mfug_size); 4226 mult = MIN(mult, 10); 4227 4228 delta = MIN(bytes * mult, arc_p); 4229 arc_p = MAX(arc_p_min, arc_p - delta); 4230 } 4231 ASSERT((int64_t)arc_p >= 0); 4232 4233 if (arc_reclaim_needed()) { 4234 cv_signal(&arc_reclaim_thread_cv); 4235 return; 4236 } 4237 4238 if (arc_no_grow) 4239 return; 4240 4241 if (arc_c >= arc_c_max) 4242 return; 4243 4244 /* 4245 * If we're within (2 * maxblocksize) bytes of the target 4246 * cache size, increment the target cache size 4247 */ 4248 if (arc_size > arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) { 4249 DTRACE_PROBE1(arc__inc_adapt, int, bytes); 4250 atomic_add_64(&arc_c, (int64_t)bytes); 4251 if (arc_c > arc_c_max) 4252 arc_c = arc_c_max; 4253 else if (state == arc_anon) 4254 atomic_add_64(&arc_p, (int64_t)bytes); 4255 if (arc_p > arc_c) 4256 arc_p = arc_c; 4257 } 4258 ASSERT((int64_t)arc_p >= 0); 4259} 4260 4261/* 4262 * Check if arc_size has grown past our upper threshold, determined by 4263 * zfs_arc_overflow_shift. 4264 */ 4265static boolean_t 4266arc_is_overflowing(void) 4267{ 4268 /* Always allow at least one block of overflow */ 4269 uint64_t overflow = MAX(SPA_MAXBLOCKSIZE, 4270 arc_c >> zfs_arc_overflow_shift); 4271 4272 return (arc_size >= arc_c + overflow); 4273} 4274 4275/* 4276 * Allocate a block and return it to the caller. If we are hitting the 4277 * hard limit for the cache size, we must sleep, waiting for the eviction 4278 * thread to catch up. If we're past the target size but below the hard 4279 * limit, we'll only signal the reclaim thread and continue on. 4280 */ 4281static void * 4282arc_get_data_buf(arc_buf_hdr_t *hdr, uint64_t size, void *tag) 4283{ 4284 void *datap = NULL; 4285 arc_state_t *state = hdr->b_l1hdr.b_state; 4286 arc_buf_contents_t type = arc_buf_type(hdr); 4287 4288 arc_adapt(size, state); 4289 4290 /* 4291 * If arc_size is currently overflowing, and has grown past our 4292 * upper limit, we must be adding data faster than the evict 4293 * thread can evict. Thus, to ensure we don't compound the 4294 * problem by adding more data and forcing arc_size to grow even 4295 * further past it's target size, we halt and wait for the 4296 * eviction thread to catch up. 4297 * 4298 * It's also possible that the reclaim thread is unable to evict 4299 * enough buffers to get arc_size below the overflow limit (e.g. 4300 * due to buffers being un-evictable, or hash lock collisions). 4301 * In this case, we want to proceed regardless if we're 4302 * overflowing; thus we don't use a while loop here. 4303 */ 4304 if (arc_is_overflowing()) { 4305 mutex_enter(&arc_reclaim_lock); 4306 4307 /* 4308 * Now that we've acquired the lock, we may no longer be 4309 * over the overflow limit, lets check. 4310 * 4311 * We're ignoring the case of spurious wake ups. If that 4312 * were to happen, it'd let this thread consume an ARC 4313 * buffer before it should have (i.e. before we're under 4314 * the overflow limit and were signalled by the reclaim 4315 * thread). As long as that is a rare occurrence, it 4316 * shouldn't cause any harm. 4317 */ 4318 if (arc_is_overflowing()) { 4319 cv_signal(&arc_reclaim_thread_cv); 4320 cv_wait(&arc_reclaim_waiters_cv, &arc_reclaim_lock); 4321 } 4322 4323 mutex_exit(&arc_reclaim_lock); 4324 } 4325 4326 VERIFY3U(hdr->b_type, ==, type); 4327 if (type == ARC_BUFC_METADATA) { 4328 datap = zio_buf_alloc(size); 4329 arc_space_consume(size, ARC_SPACE_META); 4330 } else { 4331 ASSERT(type == ARC_BUFC_DATA); 4332 datap = zio_data_buf_alloc(size); 4333 arc_space_consume(size, ARC_SPACE_DATA); 4334 } 4335 4336 /* 4337 * Update the state size. Note that ghost states have a 4338 * "ghost size" and so don't need to be updated. 4339 */ 4340 if (!GHOST_STATE(state)) { 4341 4342 (void) refcount_add_many(&state->arcs_size, size, tag); 4343 4344 /* 4345 * If this is reached via arc_read, the link is 4346 * protected by the hash lock. If reached via 4347 * arc_buf_alloc, the header should not be accessed by 4348 * any other thread. And, if reached via arc_read_done, 4349 * the hash lock will protect it if it's found in the 4350 * hash table; otherwise no other thread should be 4351 * trying to [add|remove]_reference it. 4352 */ 4353 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) { 4354 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 4355 (void) refcount_add_many(&state->arcs_esize[type], 4356 size, tag); 4357 } 4358 4359 /* 4360 * If we are growing the cache, and we are adding anonymous 4361 * data, and we have outgrown arc_p, update arc_p 4362 */ 4363 if (arc_size < arc_c && hdr->b_l1hdr.b_state == arc_anon && 4364 (refcount_count(&arc_anon->arcs_size) + 4365 refcount_count(&arc_mru->arcs_size) > arc_p)) 4366 arc_p = MIN(arc_c, arc_p + size); 4367 } 4368 ARCSTAT_BUMP(arcstat_allocated); 4369 return (datap); 4370} 4371 4372/* 4373 * Free the arc data buffer. 4374 */ 4375static void 4376arc_free_data_buf(arc_buf_hdr_t *hdr, void *data, uint64_t size, void *tag) 4377{ 4378 arc_state_t *state = hdr->b_l1hdr.b_state; 4379 arc_buf_contents_t type = arc_buf_type(hdr); 4380 4381 /* protected by hash lock, if in the hash table */ 4382 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) { 4383 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 4384 ASSERT(state != arc_anon && state != arc_l2c_only); 4385 4386 (void) refcount_remove_many(&state->arcs_esize[type], 4387 size, tag); 4388 } 4389 (void) refcount_remove_many(&state->arcs_size, size, tag); 4390 4391 VERIFY3U(hdr->b_type, ==, type); 4392 if (type == ARC_BUFC_METADATA) { 4393 zio_buf_free(data, size); 4394 arc_space_return(size, ARC_SPACE_META); 4395 } else { 4396 ASSERT(type == ARC_BUFC_DATA); 4397 zio_data_buf_free(data, size); 4398 arc_space_return(size, ARC_SPACE_DATA); 4399 } 4400} 4401 4402/* 4403 * This routine is called whenever a buffer is accessed. 4404 * NOTE: the hash lock is dropped in this function. 4405 */ 4406static void 4407arc_access(arc_buf_hdr_t *hdr, kmutex_t *hash_lock) 4408{ 4409 clock_t now; 4410 4411 ASSERT(MUTEX_HELD(hash_lock)); 4412 ASSERT(HDR_HAS_L1HDR(hdr)); 4413 4414 if (hdr->b_l1hdr.b_state == arc_anon) { 4415 /* 4416 * This buffer is not in the cache, and does not 4417 * appear in our "ghost" list. Add the new buffer 4418 * to the MRU state. 4419 */ 4420 4421 ASSERT0(hdr->b_l1hdr.b_arc_access); 4422 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 4423 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr); 4424 arc_change_state(arc_mru, hdr, hash_lock); 4425 4426 } else if (hdr->b_l1hdr.b_state == arc_mru) { 4427 now = ddi_get_lbolt(); 4428 4429 /* 4430 * If this buffer is here because of a prefetch, then either: 4431 * - clear the flag if this is a "referencing" read 4432 * (any subsequent access will bump this into the MFU state). 4433 * or 4434 * - move the buffer to the head of the list if this is 4435 * another prefetch (to make it less likely to be evicted). 4436 */ 4437 if (HDR_PREFETCH(hdr)) { 4438 if (refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) { 4439 /* link protected by hash lock */ 4440 ASSERT(multilist_link_active( 4441 &hdr->b_l1hdr.b_arc_node)); 4442 } else { 4443 arc_hdr_clear_flags(hdr, ARC_FLAG_PREFETCH); 4444 ARCSTAT_BUMP(arcstat_mru_hits); 4445 } 4446 hdr->b_l1hdr.b_arc_access = now; 4447 return; 4448 } 4449 4450 /* 4451 * This buffer has been "accessed" only once so far, 4452 * but it is still in the cache. Move it to the MFU 4453 * state. 4454 */ 4455 if (now > hdr->b_l1hdr.b_arc_access + ARC_MINTIME) { 4456 /* 4457 * More than 125ms have passed since we 4458 * instantiated this buffer. Move it to the 4459 * most frequently used state. 4460 */ 4461 hdr->b_l1hdr.b_arc_access = now; 4462 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 4463 arc_change_state(arc_mfu, hdr, hash_lock); 4464 } 4465 ARCSTAT_BUMP(arcstat_mru_hits); 4466 } else if (hdr->b_l1hdr.b_state == arc_mru_ghost) { 4467 arc_state_t *new_state; 4468 /* 4469 * This buffer has been "accessed" recently, but 4470 * was evicted from the cache. Move it to the 4471 * MFU state. 4472 */ 4473 4474 if (HDR_PREFETCH(hdr)) { 4475 new_state = arc_mru; 4476 if (refcount_count(&hdr->b_l1hdr.b_refcnt) > 0) 4477 arc_hdr_clear_flags(hdr, ARC_FLAG_PREFETCH); 4478 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr); 4479 } else { 4480 new_state = arc_mfu; 4481 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 4482 } 4483 4484 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 4485 arc_change_state(new_state, hdr, hash_lock); 4486 4487 ARCSTAT_BUMP(arcstat_mru_ghost_hits); 4488 } else if (hdr->b_l1hdr.b_state == arc_mfu) { 4489 /* 4490 * This buffer has been accessed more than once and is 4491 * still in the cache. Keep it in the MFU state. 4492 * 4493 * NOTE: an add_reference() that occurred when we did 4494 * the arc_read() will have kicked this off the list. 4495 * If it was a prefetch, we will explicitly move it to 4496 * the head of the list now. 4497 */ 4498 if ((HDR_PREFETCH(hdr)) != 0) { 4499 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 4500 /* link protected by hash_lock */ 4501 ASSERT(multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 4502 } 4503 ARCSTAT_BUMP(arcstat_mfu_hits); 4504 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 4505 } else if (hdr->b_l1hdr.b_state == arc_mfu_ghost) { 4506 arc_state_t *new_state = arc_mfu; 4507 /* 4508 * This buffer has been accessed more than once but has 4509 * been evicted from the cache. Move it back to the 4510 * MFU state. 4511 */ 4512 4513 if (HDR_PREFETCH(hdr)) { 4514 /* 4515 * This is a prefetch access... 4516 * move this block back to the MRU state. 4517 */ 4518 ASSERT0(refcount_count(&hdr->b_l1hdr.b_refcnt)); 4519 new_state = arc_mru; 4520 } 4521 4522 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 4523 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 4524 arc_change_state(new_state, hdr, hash_lock); 4525 4526 ARCSTAT_BUMP(arcstat_mfu_ghost_hits); 4527 } else if (hdr->b_l1hdr.b_state == arc_l2c_only) { 4528 /* 4529 * This buffer is on the 2nd Level ARC. 4530 */ 4531 4532 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 4533 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 4534 arc_change_state(arc_mfu, hdr, hash_lock); 4535 } else { 4536 ASSERT(!"invalid arc state"); 4537 } 4538} 4539 4540/* a generic arc_done_func_t which you can use */ 4541/* ARGSUSED */ 4542void 4543arc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg) 4544{ 4545 if (zio == NULL || zio->io_error == 0) 4546 bcopy(buf->b_data, arg, HDR_GET_LSIZE(buf->b_hdr)); 4547 arc_buf_destroy(buf, arg); 4548} 4549 4550/* a generic arc_done_func_t */ 4551void 4552arc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg) 4553{ 4554 arc_buf_t **bufp = arg; 4555 if (zio && zio->io_error) { 4556 arc_buf_destroy(buf, arg); 4557 *bufp = NULL; 4558 } else { 4559 *bufp = buf; 4560 ASSERT(buf->b_data); 4561 } 4562} 4563 4564static void 4565arc_hdr_verify(arc_buf_hdr_t *hdr, blkptr_t *bp) 4566{ 4567 if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp)) { 4568 ASSERT3U(HDR_GET_PSIZE(hdr), ==, 0); 4569 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF); 4570 } else { 4571 if (HDR_COMPRESSION_ENABLED(hdr)) { 4572 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, 4573 BP_GET_COMPRESS(bp)); 4574 } 4575 ASSERT3U(HDR_GET_LSIZE(hdr), ==, BP_GET_LSIZE(bp)); 4576 ASSERT3U(HDR_GET_PSIZE(hdr), ==, BP_GET_PSIZE(bp)); 4577 } 4578} 4579 4580static void 4581arc_read_done(zio_t *zio) 4582{ 4583 arc_buf_hdr_t *hdr = zio->io_private; 4584 arc_buf_t *abuf = NULL; /* buffer we're assigning to callback */ 4585 kmutex_t *hash_lock = NULL; 4586 arc_callback_t *callback_list, *acb; 4587 int freeable = B_FALSE; 4588 4589 /* 4590 * The hdr was inserted into hash-table and removed from lists 4591 * prior to starting I/O. We should find this header, since 4592 * it's in the hash table, and it should be legit since it's 4593 * not possible to evict it during the I/O. The only possible 4594 * reason for it not to be found is if we were freed during the 4595 * read. 4596 */ 4597 if (HDR_IN_HASH_TABLE(hdr)) { 4598 ASSERT3U(hdr->b_birth, ==, BP_PHYSICAL_BIRTH(zio->io_bp)); 4599 ASSERT3U(hdr->b_dva.dva_word[0], ==, 4600 BP_IDENTITY(zio->io_bp)->dva_word[0]); 4601 ASSERT3U(hdr->b_dva.dva_word[1], ==, 4602 BP_IDENTITY(zio->io_bp)->dva_word[1]); 4603 4604 arc_buf_hdr_t *found = buf_hash_find(hdr->b_spa, zio->io_bp, 4605 &hash_lock); 4606 4607 ASSERT((found == hdr && 4608 DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) || 4609 (found == hdr && HDR_L2_READING(hdr))); 4610 ASSERT3P(hash_lock, !=, NULL); 4611 } 4612 4613 if (zio->io_error == 0) { 4614 /* byteswap if necessary */ 4615 if (BP_SHOULD_BYTESWAP(zio->io_bp)) { 4616 if (BP_GET_LEVEL(zio->io_bp) > 0) { 4617 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_UINT64; 4618 } else { 4619 hdr->b_l1hdr.b_byteswap = 4620 DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp)); 4621 } 4622 } else { 4623 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS; 4624 } 4625 } 4626 4627 arc_hdr_clear_flags(hdr, ARC_FLAG_L2_EVICTED); 4628 if (l2arc_noprefetch && HDR_PREFETCH(hdr)) 4629 arc_hdr_clear_flags(hdr, ARC_FLAG_L2CACHE); 4630 4631 callback_list = hdr->b_l1hdr.b_acb; 4632 ASSERT3P(callback_list, !=, NULL); 4633 4634 if (hash_lock && zio->io_error == 0 && 4635 hdr->b_l1hdr.b_state == arc_anon) { 4636 /* 4637 * Only call arc_access on anonymous buffers. This is because 4638 * if we've issued an I/O for an evicted buffer, we've already 4639 * called arc_access (to prevent any simultaneous readers from 4640 * getting confused). 4641 */ 4642 arc_access(hdr, hash_lock); 4643 } 4644 4645 /* create copies of the data buffer for the callers */ 4646 for (acb = callback_list; acb; acb = acb->acb_next) { 4647 if (acb->acb_done != NULL) { 4648 /* 4649 * If we're here, then this must be a demand read 4650 * since prefetch requests don't have callbacks. 4651 * If a read request has a callback (i.e. acb_done is 4652 * not NULL), then we decompress the data for the 4653 * first request and clone the rest. This avoids 4654 * having to waste cpu resources decompressing data 4655 * that nobody is explicitly waiting to read. 4656 */ 4657 if (abuf == NULL) { 4658 acb->acb_buf = arc_buf_alloc_impl(hdr, 4659 acb->acb_private); 4660 if (zio->io_error == 0) { 4661 zio->io_error = 4662 arc_decompress(acb->acb_buf); 4663 } 4664 abuf = acb->acb_buf; 4665 } else { 4666 add_reference(hdr, acb->acb_private); 4667 acb->acb_buf = arc_buf_clone(abuf); 4668 } 4669 } 4670 } 4671 hdr->b_l1hdr.b_acb = NULL; 4672 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS); 4673 if (abuf == NULL) { 4674 /* 4675 * This buffer didn't have a callback so it must 4676 * be a prefetch. 4677 */ 4678 ASSERT(HDR_PREFETCH(hdr)); 4679 ASSERT0(hdr->b_l1hdr.b_bufcnt); 4680 ASSERT3P(hdr->b_l1hdr.b_pdata, !=, NULL); 4681 } 4682 4683 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt) || 4684 callback_list != NULL); 4685 4686 if (zio->io_error == 0) { 4687 arc_hdr_verify(hdr, zio->io_bp); 4688 } else { 4689 arc_hdr_set_flags(hdr, ARC_FLAG_IO_ERROR); 4690 if (hdr->b_l1hdr.b_state != arc_anon) 4691 arc_change_state(arc_anon, hdr, hash_lock); 4692 if (HDR_IN_HASH_TABLE(hdr)) 4693 buf_hash_remove(hdr); 4694 freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt); 4695 } 4696 4697 /* 4698 * Broadcast before we drop the hash_lock to avoid the possibility 4699 * that the hdr (and hence the cv) might be freed before we get to 4700 * the cv_broadcast(). 4701 */ 4702 cv_broadcast(&hdr->b_l1hdr.b_cv); 4703 4704 if (hash_lock != NULL) { 4705 mutex_exit(hash_lock); 4706 } else { 4707 /* 4708 * This block was freed while we waited for the read to 4709 * complete. It has been removed from the hash table and 4710 * moved to the anonymous state (so that it won't show up 4711 * in the cache). 4712 */ 4713 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon); 4714 freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt); 4715 } 4716 4717 /* execute each callback and free its structure */ 4718 while ((acb = callback_list) != NULL) { 4719 if (acb->acb_done) 4720 acb->acb_done(zio, acb->acb_buf, acb->acb_private); 4721 4722 if (acb->acb_zio_dummy != NULL) { 4723 acb->acb_zio_dummy->io_error = zio->io_error; 4724 zio_nowait(acb->acb_zio_dummy); 4725 } 4726 4727 callback_list = acb->acb_next; 4728 kmem_free(acb, sizeof (arc_callback_t)); 4729 } 4730 4731 if (freeable) 4732 arc_hdr_destroy(hdr); 4733} 4734 4735/* 4736 * "Read" the block at the specified DVA (in bp) via the 4737 * cache. If the block is found in the cache, invoke the provided 4738 * callback immediately and return. Note that the `zio' parameter 4739 * in the callback will be NULL in this case, since no IO was 4740 * required. If the block is not in the cache pass the read request 4741 * on to the spa with a substitute callback function, so that the 4742 * requested block will be added to the cache. 4743 * 4744 * If a read request arrives for a block that has a read in-progress, 4745 * either wait for the in-progress read to complete (and return the 4746 * results); or, if this is a read with a "done" func, add a record 4747 * to the read to invoke the "done" func when the read completes, 4748 * and return; or just return. 4749 * 4750 * arc_read_done() will invoke all the requested "done" functions 4751 * for readers of this block. 4752 */ 4753int 4754arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_done_func_t *done, 4755 void *private, zio_priority_t priority, int zio_flags, 4756 arc_flags_t *arc_flags, const zbookmark_phys_t *zb) 4757{ 4758 arc_buf_hdr_t *hdr = NULL; 4759 kmutex_t *hash_lock = NULL; 4760 zio_t *rzio; 4761 uint64_t guid = spa_load_guid(spa); 4762 4763 ASSERT(!BP_IS_EMBEDDED(bp) || 4764 BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA); 4765 4766top: 4767 if (!BP_IS_EMBEDDED(bp)) { 4768 /* 4769 * Embedded BP's have no DVA and require no I/O to "read". 4770 * Create an anonymous arc buf to back it. 4771 */ 4772 hdr = buf_hash_find(guid, bp, &hash_lock); 4773 } 4774 4775 if (hdr != NULL && HDR_HAS_L1HDR(hdr) && hdr->b_l1hdr.b_pdata != NULL) { 4776 arc_buf_t *buf = NULL; 4777 *arc_flags |= ARC_FLAG_CACHED; 4778 4779 if (HDR_IO_IN_PROGRESS(hdr)) { 4780 4781 if ((hdr->b_flags & ARC_FLAG_PRIO_ASYNC_READ) && 4782 priority == ZIO_PRIORITY_SYNC_READ) { 4783 /* 4784 * This sync read must wait for an 4785 * in-progress async read (e.g. a predictive 4786 * prefetch). Async reads are queued 4787 * separately at the vdev_queue layer, so 4788 * this is a form of priority inversion. 4789 * Ideally, we would "inherit" the demand 4790 * i/o's priority by moving the i/o from 4791 * the async queue to the synchronous queue, 4792 * but there is currently no mechanism to do 4793 * so. Track this so that we can evaluate 4794 * the magnitude of this potential performance 4795 * problem. 4796 * 4797 * Note that if the prefetch i/o is already 4798 * active (has been issued to the device), 4799 * the prefetch improved performance, because 4800 * we issued it sooner than we would have 4801 * without the prefetch. 4802 */ 4803 DTRACE_PROBE1(arc__sync__wait__for__async, 4804 arc_buf_hdr_t *, hdr); 4805 ARCSTAT_BUMP(arcstat_sync_wait_for_async); 4806 } 4807 if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) { 4808 arc_hdr_clear_flags(hdr, 4809 ARC_FLAG_PREDICTIVE_PREFETCH); 4810 } 4811 4812 if (*arc_flags & ARC_FLAG_WAIT) { 4813 cv_wait(&hdr->b_l1hdr.b_cv, hash_lock); 4814 mutex_exit(hash_lock); 4815 goto top; 4816 } 4817 ASSERT(*arc_flags & ARC_FLAG_NOWAIT); 4818 4819 if (done) { 4820 arc_callback_t *acb = NULL; 4821 4822 acb = kmem_zalloc(sizeof (arc_callback_t), 4823 KM_SLEEP); 4824 acb->acb_done = done; 4825 acb->acb_private = private; 4826 if (pio != NULL) 4827 acb->acb_zio_dummy = zio_null(pio, 4828 spa, NULL, NULL, NULL, zio_flags); 4829 4830 ASSERT3P(acb->acb_done, !=, NULL); 4831 acb->acb_next = hdr->b_l1hdr.b_acb; 4832 hdr->b_l1hdr.b_acb = acb; 4833 mutex_exit(hash_lock); 4834 return (0); 4835 } 4836 mutex_exit(hash_lock); 4837 return (0); 4838 } 4839 4840 ASSERT(hdr->b_l1hdr.b_state == arc_mru || 4841 hdr->b_l1hdr.b_state == arc_mfu); 4842 4843 if (done) { 4844 if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) { 4845 /* 4846 * This is a demand read which does not have to 4847 * wait for i/o because we did a predictive 4848 * prefetch i/o for it, which has completed. 4849 */ 4850 DTRACE_PROBE1( 4851 arc__demand__hit__predictive__prefetch, 4852 arc_buf_hdr_t *, hdr); 4853 ARCSTAT_BUMP( 4854 arcstat_demand_hit_predictive_prefetch); 4855 arc_hdr_clear_flags(hdr, 4856 ARC_FLAG_PREDICTIVE_PREFETCH); 4857 } 4858 ASSERT(!BP_IS_EMBEDDED(bp) || !BP_IS_HOLE(bp)); 4859 4860 /* 4861 * If this block is already in use, create a new 4862 * copy of the data so that we will be guaranteed 4863 * that arc_release() will always succeed. 4864 */ 4865 buf = hdr->b_l1hdr.b_buf; 4866 if (buf == NULL) { 4867 ASSERT0(refcount_count(&hdr->b_l1hdr.b_refcnt)); 4868 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL); 4869 buf = arc_buf_alloc_impl(hdr, private); 4870 VERIFY0(arc_decompress(buf)); 4871 } else { 4872 add_reference(hdr, private); 4873 buf = arc_buf_clone(buf); 4874 } 4875 ASSERT3P(buf->b_data, !=, NULL); 4876 4877 } else if (*arc_flags & ARC_FLAG_PREFETCH && 4878 refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) { 4879 arc_hdr_set_flags(hdr, ARC_FLAG_PREFETCH); 4880 } 4881 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr); 4882 arc_access(hdr, hash_lock); 4883 if (*arc_flags & ARC_FLAG_L2CACHE) 4884 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE); 4885 mutex_exit(hash_lock); 4886 ARCSTAT_BUMP(arcstat_hits); 4887 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr), 4888 demand, prefetch, !HDR_ISTYPE_METADATA(hdr), 4889 data, metadata, hits); 4890 4891 if (done) 4892 done(NULL, buf, private); 4893 } else { 4894 uint64_t lsize = BP_GET_LSIZE(bp); 4895 uint64_t psize = BP_GET_PSIZE(bp); 4896 arc_callback_t *acb; 4897 vdev_t *vd = NULL; 4898 uint64_t addr = 0; 4899 boolean_t devw = B_FALSE; 4900 uint64_t size; 4901 4902 if (hdr == NULL) { 4903 /* this block is not in the cache */ 4904 arc_buf_hdr_t *exists = NULL; 4905 arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp); 4906 hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize, 4907 BP_GET_COMPRESS(bp), type); 4908 4909 if (!BP_IS_EMBEDDED(bp)) { 4910 hdr->b_dva = *BP_IDENTITY(bp); 4911 hdr->b_birth = BP_PHYSICAL_BIRTH(bp); 4912 exists = buf_hash_insert(hdr, &hash_lock); 4913 } 4914 if (exists != NULL) { 4915 /* somebody beat us to the hash insert */ 4916 mutex_exit(hash_lock); 4917 buf_discard_identity(hdr); 4918 arc_hdr_destroy(hdr); 4919 goto top; /* restart the IO request */ 4920 } 4921 } else { 4922 /* 4923 * This block is in the ghost cache. If it was L2-only 4924 * (and thus didn't have an L1 hdr), we realloc the 4925 * header to add an L1 hdr. 4926 */ 4927 if (!HDR_HAS_L1HDR(hdr)) { 4928 hdr = arc_hdr_realloc(hdr, hdr_l2only_cache, 4929 hdr_full_cache); 4930 } 4931 ASSERT3P(hdr->b_l1hdr.b_pdata, ==, NULL); 4932 ASSERT(GHOST_STATE(hdr->b_l1hdr.b_state)); 4933 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 4934 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 4935 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 4936 4937 /* 4938 * This is a delicate dance that we play here. 4939 * This hdr is in the ghost list so we access it 4940 * to move it out of the ghost list before we 4941 * initiate the read. If it's a prefetch then 4942 * it won't have a callback so we'll remove the 4943 * reference that arc_buf_alloc_impl() created. We 4944 * do this after we've called arc_access() to 4945 * avoid hitting an assert in remove_reference(). 4946 */ 4947 arc_access(hdr, hash_lock); 4948 arc_hdr_alloc_pdata(hdr); 4949 } 4950 ASSERT3P(hdr->b_l1hdr.b_pdata, !=, NULL); 4951 size = arc_hdr_size(hdr); 4952 4953 /* 4954 * If compression is enabled on the hdr, then will do 4955 * RAW I/O and will store the compressed data in the hdr's 4956 * data block. Otherwise, the hdr's data block will contain 4957 * the uncompressed data. 4958 */ 4959 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF) { 4960 zio_flags |= ZIO_FLAG_RAW; 4961 } 4962 4963 if (*arc_flags & ARC_FLAG_PREFETCH) 4964 arc_hdr_set_flags(hdr, ARC_FLAG_PREFETCH); 4965 if (*arc_flags & ARC_FLAG_L2CACHE) 4966 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE); 4967 if (BP_GET_LEVEL(bp) > 0) 4968 arc_hdr_set_flags(hdr, ARC_FLAG_INDIRECT); 4969 if (*arc_flags & ARC_FLAG_PREDICTIVE_PREFETCH) 4970 arc_hdr_set_flags(hdr, ARC_FLAG_PREDICTIVE_PREFETCH); 4971 ASSERT(!GHOST_STATE(hdr->b_l1hdr.b_state)); 4972 4973 acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP); 4974 acb->acb_done = done; 4975 acb->acb_private = private; 4976 4977 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL); 4978 hdr->b_l1hdr.b_acb = acb; 4979 arc_hdr_set_flags(hdr, ARC_FLAG_IO_IN_PROGRESS); 4980 4981 if (HDR_HAS_L2HDR(hdr) && 4982 (vd = hdr->b_l2hdr.b_dev->l2ad_vdev) != NULL) { 4983 devw = hdr->b_l2hdr.b_dev->l2ad_writing; 4984 addr = hdr->b_l2hdr.b_daddr; 4985 /* 4986 * Lock out device removal. 4987 */ 4988 if (vdev_is_dead(vd) || 4989 !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER)) 4990 vd = NULL; 4991 } 4992 4993 if (priority == ZIO_PRIORITY_ASYNC_READ) 4994 arc_hdr_set_flags(hdr, ARC_FLAG_PRIO_ASYNC_READ); 4995 else 4996 arc_hdr_clear_flags(hdr, ARC_FLAG_PRIO_ASYNC_READ); 4997 4998 if (hash_lock != NULL) 4999 mutex_exit(hash_lock); 5000 5001 /* 5002 * At this point, we have a level 1 cache miss. Try again in 5003 * L2ARC if possible. 5004 */ 5005 ASSERT3U(HDR_GET_LSIZE(hdr), ==, lsize); 5006 5007 DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp, 5008 uint64_t, lsize, zbookmark_phys_t *, zb); 5009 ARCSTAT_BUMP(arcstat_misses); 5010 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr), 5011 demand, prefetch, !HDR_ISTYPE_METADATA(hdr), 5012 data, metadata, misses); 5013#ifdef _KERNEL 5014 curthread->td_ru.ru_inblock++; 5015#endif 5016 5017 if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) { 5018 /* 5019 * Read from the L2ARC if the following are true: 5020 * 1. The L2ARC vdev was previously cached. 5021 * 2. This buffer still has L2ARC metadata. 5022 * 3. This buffer isn't currently writing to the L2ARC. 5023 * 4. The L2ARC entry wasn't evicted, which may 5024 * also have invalidated the vdev. 5025 * 5. This isn't prefetch and l2arc_noprefetch is set. 5026 */ 5027 if (HDR_HAS_L2HDR(hdr) && 5028 !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) && 5029 !(l2arc_noprefetch && HDR_PREFETCH(hdr))) { 5030 l2arc_read_callback_t *cb; 5031 void* b_data; 5032 5033 DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr); 5034 ARCSTAT_BUMP(arcstat_l2_hits); 5035 5036 cb = kmem_zalloc(sizeof (l2arc_read_callback_t), 5037 KM_SLEEP); 5038 cb->l2rcb_hdr = hdr; 5039 cb->l2rcb_bp = *bp; 5040 cb->l2rcb_zb = *zb; 5041 cb->l2rcb_flags = zio_flags; 5042 uint64_t asize = vdev_psize_to_asize(vd, size); 5043 if (asize != size) { 5044 b_data = zio_data_buf_alloc(asize); 5045 cb->l2rcb_data = b_data; 5046 } else { 5047 b_data = hdr->b_l1hdr.b_pdata; 5048 } 5049 5050 ASSERT(addr >= VDEV_LABEL_START_SIZE && 5051 addr + asize < vd->vdev_psize - 5052 VDEV_LABEL_END_SIZE); 5053 5054 /* 5055 * l2arc read. The SCL_L2ARC lock will be 5056 * released by l2arc_read_done(). 5057 * Issue a null zio if the underlying buffer 5058 * was squashed to zero size by compression. 5059 */ 5060 ASSERT3U(HDR_GET_COMPRESS(hdr), !=, 5061 ZIO_COMPRESS_EMPTY); 5062 rzio = zio_read_phys(pio, vd, addr, 5063 asize, b_data, 5064 ZIO_CHECKSUM_OFF, 5065 l2arc_read_done, cb, priority, 5066 zio_flags | ZIO_FLAG_DONT_CACHE | 5067 ZIO_FLAG_CANFAIL | 5068 ZIO_FLAG_DONT_PROPAGATE | 5069 ZIO_FLAG_DONT_RETRY, B_FALSE); 5070 DTRACE_PROBE2(l2arc__read, vdev_t *, vd, 5071 zio_t *, rzio); 5072 ARCSTAT_INCR(arcstat_l2_read_bytes, size); 5073 5074 if (*arc_flags & ARC_FLAG_NOWAIT) { 5075 zio_nowait(rzio); 5076 return (0); 5077 } 5078 5079 ASSERT(*arc_flags & ARC_FLAG_WAIT); 5080 if (zio_wait(rzio) == 0) 5081 return (0); 5082 5083 /* l2arc read error; goto zio_read() */ 5084 } else { 5085 DTRACE_PROBE1(l2arc__miss, 5086 arc_buf_hdr_t *, hdr); 5087 ARCSTAT_BUMP(arcstat_l2_misses); 5088 if (HDR_L2_WRITING(hdr)) 5089 ARCSTAT_BUMP(arcstat_l2_rw_clash); 5090 spa_config_exit(spa, SCL_L2ARC, vd); 5091 } 5092 } else { 5093 if (vd != NULL) 5094 spa_config_exit(spa, SCL_L2ARC, vd); 5095 if (l2arc_ndev != 0) { 5096 DTRACE_PROBE1(l2arc__miss, 5097 arc_buf_hdr_t *, hdr); 5098 ARCSTAT_BUMP(arcstat_l2_misses); 5099 } 5100 } 5101 5102 rzio = zio_read(pio, spa, bp, hdr->b_l1hdr.b_pdata, size, 5103 arc_read_done, hdr, priority, zio_flags, zb); 5104 5105 if (*arc_flags & ARC_FLAG_WAIT) 5106 return (zio_wait(rzio)); 5107 5108 ASSERT(*arc_flags & ARC_FLAG_NOWAIT); 5109 zio_nowait(rzio); 5110 } 5111 return (0); 5112} 5113 5114/* 5115 * Notify the arc that a block was freed, and thus will never be used again. 5116 */ 5117void 5118arc_freed(spa_t *spa, const blkptr_t *bp) 5119{ 5120 arc_buf_hdr_t *hdr; 5121 kmutex_t *hash_lock; 5122 uint64_t guid = spa_load_guid(spa); 5123 5124 ASSERT(!BP_IS_EMBEDDED(bp)); 5125 5126 hdr = buf_hash_find(guid, bp, &hash_lock); 5127 if (hdr == NULL) 5128 return; 5129 5130 /* 5131 * We might be trying to free a block that is still doing I/O 5132 * (i.e. prefetch) or has a reference (i.e. a dedup-ed, 5133 * dmu_sync-ed block). If this block is being prefetched, then it 5134 * would still have the ARC_FLAG_IO_IN_PROGRESS flag set on the hdr 5135 * until the I/O completes. A block may also have a reference if it is 5136 * part of a dedup-ed, dmu_synced write. The dmu_sync() function would 5137 * have written the new block to its final resting place on disk but 5138 * without the dedup flag set. This would have left the hdr in the MRU 5139 * state and discoverable. When the txg finally syncs it detects that 5140 * the block was overridden in open context and issues an override I/O. 5141 * Since this is a dedup block, the override I/O will determine if the 5142 * block is already in the DDT. If so, then it will replace the io_bp 5143 * with the bp from the DDT and allow the I/O to finish. When the I/O 5144 * reaches the done callback, dbuf_write_override_done, it will 5145 * check to see if the io_bp and io_bp_override are identical. 5146 * If they are not, then it indicates that the bp was replaced with 5147 * the bp in the DDT and the override bp is freed. This allows 5148 * us to arrive here with a reference on a block that is being 5149 * freed. So if we have an I/O in progress, or a reference to 5150 * this hdr, then we don't destroy the hdr. 5151 */ 5152 if (!HDR_HAS_L1HDR(hdr) || (!HDR_IO_IN_PROGRESS(hdr) && 5153 refcount_is_zero(&hdr->b_l1hdr.b_refcnt))) { 5154 arc_change_state(arc_anon, hdr, hash_lock); 5155 arc_hdr_destroy(hdr); 5156 mutex_exit(hash_lock); 5157 } else { 5158 mutex_exit(hash_lock); 5159 } 5160 5161} 5162 5163/* 5164 * Release this buffer from the cache, making it an anonymous buffer. This 5165 * must be done after a read and prior to modifying the buffer contents. 5166 * If the buffer has more than one reference, we must make 5167 * a new hdr for the buffer. 5168 */ 5169void 5170arc_release(arc_buf_t *buf, void *tag) 5171{ 5172 arc_buf_hdr_t *hdr = buf->b_hdr; 5173 5174 /* 5175 * It would be nice to assert that if it's DMU metadata (level > 5176 * 0 || it's the dnode file), then it must be syncing context. 5177 * But we don't know that information at this level. 5178 */ 5179 5180 mutex_enter(&buf->b_evict_lock); 5181 5182 ASSERT(HDR_HAS_L1HDR(hdr)); 5183 5184 /* 5185 * We don't grab the hash lock prior to this check, because if 5186 * the buffer's header is in the arc_anon state, it won't be 5187 * linked into the hash table. 5188 */ 5189 if (hdr->b_l1hdr.b_state == arc_anon) { 5190 mutex_exit(&buf->b_evict_lock); 5191 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 5192 ASSERT(!HDR_IN_HASH_TABLE(hdr)); 5193 ASSERT(!HDR_HAS_L2HDR(hdr)); 5194 ASSERT(HDR_EMPTY(hdr)); 5195 ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1); 5196 ASSERT3S(refcount_count(&hdr->b_l1hdr.b_refcnt), ==, 1); 5197 ASSERT(!list_link_active(&hdr->b_l1hdr.b_arc_node)); 5198 5199 hdr->b_l1hdr.b_arc_access = 0; 5200 5201 /* 5202 * If the buf is being overridden then it may already 5203 * have a hdr that is not empty. 5204 */ 5205 buf_discard_identity(hdr); 5206 arc_buf_thaw(buf); 5207 5208 return; 5209 } 5210 5211 kmutex_t *hash_lock = HDR_LOCK(hdr); 5212 mutex_enter(hash_lock); 5213 5214 /* 5215 * This assignment is only valid as long as the hash_lock is 5216 * held, we must be careful not to reference state or the 5217 * b_state field after dropping the lock. 5218 */ 5219 arc_state_t *state = hdr->b_l1hdr.b_state; 5220 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 5221 ASSERT3P(state, !=, arc_anon); 5222 5223 /* this buffer is not on any list */ 5224 ASSERT(refcount_count(&hdr->b_l1hdr.b_refcnt) > 0); 5225 5226 if (HDR_HAS_L2HDR(hdr)) { 5227 mutex_enter(&hdr->b_l2hdr.b_dev->l2ad_mtx); 5228 5229 /* 5230 * We have to recheck this conditional again now that 5231 * we're holding the l2ad_mtx to prevent a race with 5232 * another thread which might be concurrently calling 5233 * l2arc_evict(). In that case, l2arc_evict() might have 5234 * destroyed the header's L2 portion as we were waiting 5235 * to acquire the l2ad_mtx. 5236 */ 5237 if (HDR_HAS_L2HDR(hdr)) { 5238 l2arc_trim(hdr); 5239 arc_hdr_l2hdr_destroy(hdr); 5240 } 5241 5242 mutex_exit(&hdr->b_l2hdr.b_dev->l2ad_mtx); 5243 } 5244 5245 /* 5246 * Do we have more than one buf? 5247 */ 5248 if (hdr->b_l1hdr.b_bufcnt > 1) { 5249 arc_buf_hdr_t *nhdr; 5250 arc_buf_t **bufp; 5251 uint64_t spa = hdr->b_spa; 5252 uint64_t psize = HDR_GET_PSIZE(hdr); 5253 uint64_t lsize = HDR_GET_LSIZE(hdr); 5254 enum zio_compress compress = HDR_GET_COMPRESS(hdr); 5255 arc_buf_contents_t type = arc_buf_type(hdr); 5256 VERIFY3U(hdr->b_type, ==, type); 5257 5258 ASSERT(hdr->b_l1hdr.b_buf != buf || buf->b_next != NULL); 5259 (void) remove_reference(hdr, hash_lock, tag); 5260 5261 if (arc_buf_is_shared(buf)) { 5262 ASSERT(HDR_SHARED_DATA(hdr)); 5263 ASSERT3P(hdr->b_l1hdr.b_buf, !=, buf); 5264 ASSERT(ARC_BUF_LAST(buf)); 5265 } 5266 5267 /* 5268 * Pull the data off of this hdr and attach it to 5269 * a new anonymous hdr. Also find the last buffer 5270 * in the hdr's buffer list. 5271 */ 5272 arc_buf_t *lastbuf = NULL; 5273 bufp = &hdr->b_l1hdr.b_buf; 5274 while (*bufp != NULL) { 5275 if (*bufp == buf) { 5276 *bufp = buf->b_next; 5277 } 5278 5279 /* 5280 * If we've removed a buffer in the middle of 5281 * the list then update the lastbuf and update 5282 * bufp. 5283 */ 5284 if (*bufp != NULL) { 5285 lastbuf = *bufp; 5286 bufp = &(*bufp)->b_next; 5287 } 5288 } 5289 buf->b_next = NULL; 5290 ASSERT3P(lastbuf, !=, buf); 5291 ASSERT3P(lastbuf, !=, NULL); 5292 5293 /* 5294 * If the current arc_buf_t and the hdr are sharing their data 5295 * buffer, then we must stop sharing that block, transfer 5296 * ownership and setup sharing with a new arc_buf_t at the end 5297 * of the hdr's b_buf list. 5298 */ 5299 if (arc_buf_is_shared(buf)) { 5300 ASSERT3P(hdr->b_l1hdr.b_buf, !=, buf); 5301 ASSERT(ARC_BUF_LAST(lastbuf)); 5302 VERIFY(!arc_buf_is_shared(lastbuf)); 5303 5304 /* 5305 * First, sever the block sharing relationship between 5306 * buf and the arc_buf_hdr_t. Then, setup a new 5307 * block sharing relationship with the last buffer 5308 * on the arc_buf_t list. 5309 */ 5310 arc_unshare_buf(hdr, buf); 5311 arc_share_buf(hdr, lastbuf); 5312 VERIFY3P(lastbuf->b_data, !=, NULL); 5313 } else if (HDR_SHARED_DATA(hdr)) { 5314 ASSERT(arc_buf_is_shared(lastbuf)); 5315 } 5316 ASSERT3P(hdr->b_l1hdr.b_pdata, !=, NULL); 5317 ASSERT3P(state, !=, arc_l2c_only); 5318 5319 (void) refcount_remove_many(&state->arcs_size, 5320 HDR_GET_LSIZE(hdr), buf); 5321 5322 if (refcount_is_zero(&hdr->b_l1hdr.b_refcnt)) { 5323 ASSERT3P(state, !=, arc_l2c_only); 5324 (void) refcount_remove_many(&state->arcs_esize[type], 5325 HDR_GET_LSIZE(hdr), buf); 5326 } 5327 5328 hdr->b_l1hdr.b_bufcnt -= 1; 5329 arc_cksum_verify(buf); 5330#ifdef illumos 5331 arc_buf_unwatch(buf); 5332#endif 5333 5334 mutex_exit(hash_lock); 5335 5336 /* 5337 * Allocate a new hdr. The new hdr will contain a b_pdata 5338 * buffer which will be freed in arc_write(). 5339 */ 5340 nhdr = arc_hdr_alloc(spa, psize, lsize, compress, type); 5341 ASSERT3P(nhdr->b_l1hdr.b_buf, ==, NULL); 5342 ASSERT0(nhdr->b_l1hdr.b_bufcnt); 5343 ASSERT0(refcount_count(&nhdr->b_l1hdr.b_refcnt)); 5344 VERIFY3U(nhdr->b_type, ==, type); 5345 ASSERT(!HDR_SHARED_DATA(nhdr)); 5346 5347 nhdr->b_l1hdr.b_buf = buf; 5348 nhdr->b_l1hdr.b_bufcnt = 1; 5349 (void) refcount_add(&nhdr->b_l1hdr.b_refcnt, tag); 5350 buf->b_hdr = nhdr; 5351 5352 mutex_exit(&buf->b_evict_lock); 5353 (void) refcount_add_many(&arc_anon->arcs_size, 5354 HDR_GET_LSIZE(nhdr), buf); 5355 } else { 5356 mutex_exit(&buf->b_evict_lock); 5357 ASSERT(refcount_count(&hdr->b_l1hdr.b_refcnt) == 1); 5358 /* protected by hash lock, or hdr is on arc_anon */ 5359 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 5360 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 5361 arc_change_state(arc_anon, hdr, hash_lock); 5362 hdr->b_l1hdr.b_arc_access = 0; 5363 mutex_exit(hash_lock); 5364 5365 buf_discard_identity(hdr); 5366 arc_buf_thaw(buf); 5367 } 5368} 5369 5370int 5371arc_released(arc_buf_t *buf) 5372{ 5373 int released; 5374 5375 mutex_enter(&buf->b_evict_lock); 5376 released = (buf->b_data != NULL && 5377 buf->b_hdr->b_l1hdr.b_state == arc_anon); 5378 mutex_exit(&buf->b_evict_lock); 5379 return (released); 5380} 5381 5382#ifdef ZFS_DEBUG 5383int 5384arc_referenced(arc_buf_t *buf) 5385{ 5386 int referenced; 5387 5388 mutex_enter(&buf->b_evict_lock); 5389 referenced = (refcount_count(&buf->b_hdr->b_l1hdr.b_refcnt)); 5390 mutex_exit(&buf->b_evict_lock); 5391 return (referenced); 5392} 5393#endif 5394 5395static void 5396arc_write_ready(zio_t *zio) 5397{ 5398 arc_write_callback_t *callback = zio->io_private; 5399 arc_buf_t *buf = callback->awcb_buf; 5400 arc_buf_hdr_t *hdr = buf->b_hdr; 5401 uint64_t psize = BP_IS_HOLE(zio->io_bp) ? 0 : BP_GET_PSIZE(zio->io_bp); 5402 5403 ASSERT(HDR_HAS_L1HDR(hdr)); 5404 ASSERT(!refcount_is_zero(&buf->b_hdr->b_l1hdr.b_refcnt)); 5405 ASSERT(hdr->b_l1hdr.b_bufcnt > 0); 5406 5407 /* 5408 * If we're reexecuting this zio because the pool suspended, then 5409 * cleanup any state that was previously set the first time the 5410 * callback as invoked. 5411 */ 5412 if (zio->io_flags & ZIO_FLAG_REEXECUTED) { 5413 arc_cksum_free(hdr); 5414#ifdef illumos 5415 arc_buf_unwatch(buf); 5416#endif 5417 if (hdr->b_l1hdr.b_pdata != NULL) { 5418 if (arc_buf_is_shared(buf)) { 5419 ASSERT(HDR_SHARED_DATA(hdr)); 5420 5421 arc_unshare_buf(hdr, buf); 5422 } else { 5423 arc_hdr_free_pdata(hdr); 5424 } 5425 } 5426 } 5427 ASSERT3P(hdr->b_l1hdr.b_pdata, ==, NULL); 5428 ASSERT(!HDR_SHARED_DATA(hdr)); 5429 ASSERT(!arc_buf_is_shared(buf)); 5430 5431 callback->awcb_ready(zio, buf, callback->awcb_private); 5432 5433 if (HDR_IO_IN_PROGRESS(hdr)) 5434 ASSERT(zio->io_flags & ZIO_FLAG_REEXECUTED); 5435 5436 arc_cksum_compute(buf); 5437 arc_hdr_set_flags(hdr, ARC_FLAG_IO_IN_PROGRESS); 5438 5439 enum zio_compress compress; 5440 if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) { 5441 compress = ZIO_COMPRESS_OFF; 5442 } else { 5443 ASSERT3U(HDR_GET_LSIZE(hdr), ==, BP_GET_LSIZE(zio->io_bp)); 5444 compress = BP_GET_COMPRESS(zio->io_bp); 5445 } 5446 HDR_SET_PSIZE(hdr, psize); 5447 arc_hdr_set_compress(hdr, compress); 5448 5449 /* 5450 * If the hdr is compressed, then copy the compressed 5451 * zio contents into arc_buf_hdr_t. Otherwise, copy the original 5452 * data buf into the hdr. Ideally, we would like to always copy the 5453 * io_data into b_pdata but the user may have disabled compressed 5454 * arc thus the on-disk block may or may not match what we maintain 5455 * in the hdr's b_pdata field. 5456 */ 5457 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF) { 5458 ASSERT(BP_GET_COMPRESS(zio->io_bp) != ZIO_COMPRESS_OFF); 5459 ASSERT3U(psize, >, 0); 5460 arc_hdr_alloc_pdata(hdr); 5461 bcopy(zio->io_data, hdr->b_l1hdr.b_pdata, psize); 5462 } else { 5463 ASSERT3P(buf->b_data, ==, zio->io_orig_data); 5464 ASSERT3U(zio->io_orig_size, ==, HDR_GET_LSIZE(hdr)); 5465 ASSERT3U(hdr->b_l1hdr.b_byteswap, ==, DMU_BSWAP_NUMFUNCS); 5466 ASSERT(!HDR_SHARED_DATA(hdr)); 5467 ASSERT(!arc_buf_is_shared(buf)); 5468 ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1); 5469 ASSERT3P(hdr->b_l1hdr.b_pdata, ==, NULL); 5470 5471 /* 5472 * This hdr is not compressed so we're able to share 5473 * the arc_buf_t data buffer with the hdr. 5474 */ 5475 arc_share_buf(hdr, buf); 5476 VERIFY0(bcmp(zio->io_orig_data, hdr->b_l1hdr.b_pdata, 5477 HDR_GET_LSIZE(hdr))); 5478 } 5479 arc_hdr_verify(hdr, zio->io_bp); 5480} 5481 5482static void 5483arc_write_children_ready(zio_t *zio) 5484{ 5485 arc_write_callback_t *callback = zio->io_private; 5486 arc_buf_t *buf = callback->awcb_buf; 5487 5488 callback->awcb_children_ready(zio, buf, callback->awcb_private); 5489} 5490 5491/* 5492 * The SPA calls this callback for each physical write that happens on behalf 5493 * of a logical write. See the comment in dbuf_write_physdone() for details. 5494 */ 5495static void 5496arc_write_physdone(zio_t *zio) 5497{ 5498 arc_write_callback_t *cb = zio->io_private; 5499 if (cb->awcb_physdone != NULL) 5500 cb->awcb_physdone(zio, cb->awcb_buf, cb->awcb_private); 5501} 5502 5503static void 5504arc_write_done(zio_t *zio) 5505{ 5506 arc_write_callback_t *callback = zio->io_private; 5507 arc_buf_t *buf = callback->awcb_buf; 5508 arc_buf_hdr_t *hdr = buf->b_hdr; 5509 5510 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL); 5511 5512 if (zio->io_error == 0) { 5513 arc_hdr_verify(hdr, zio->io_bp); 5514 5515 if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) { 5516 buf_discard_identity(hdr); 5517 } else { 5518 hdr->b_dva = *BP_IDENTITY(zio->io_bp); 5519 hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp); 5520 } 5521 } else { 5522 ASSERT(HDR_EMPTY(hdr)); 5523 } 5524 5525 /* 5526 * If the block to be written was all-zero or compressed enough to be 5527 * embedded in the BP, no write was performed so there will be no 5528 * dva/birth/checksum. The buffer must therefore remain anonymous 5529 * (and uncached). 5530 */ 5531 if (!HDR_EMPTY(hdr)) { 5532 arc_buf_hdr_t *exists; 5533 kmutex_t *hash_lock; 5534 5535 ASSERT(zio->io_error == 0); 5536 5537 arc_cksum_verify(buf); 5538 5539 exists = buf_hash_insert(hdr, &hash_lock); 5540 if (exists != NULL) { 5541 /* 5542 * This can only happen if we overwrite for 5543 * sync-to-convergence, because we remove 5544 * buffers from the hash table when we arc_free(). 5545 */ 5546 if (zio->io_flags & ZIO_FLAG_IO_REWRITE) { 5547 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) 5548 panic("bad overwrite, hdr=%p exists=%p", 5549 (void *)hdr, (void *)exists); 5550 ASSERT(refcount_is_zero( 5551 &exists->b_l1hdr.b_refcnt)); 5552 arc_change_state(arc_anon, exists, hash_lock); 5553 mutex_exit(hash_lock); 5554 arc_hdr_destroy(exists); 5555 exists = buf_hash_insert(hdr, &hash_lock); 5556 ASSERT3P(exists, ==, NULL); 5557 } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) { 5558 /* nopwrite */ 5559 ASSERT(zio->io_prop.zp_nopwrite); 5560 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) 5561 panic("bad nopwrite, hdr=%p exists=%p", 5562 (void *)hdr, (void *)exists); 5563 } else { 5564 /* Dedup */ 5565 ASSERT(hdr->b_l1hdr.b_bufcnt == 1); 5566 ASSERT(hdr->b_l1hdr.b_state == arc_anon); 5567 ASSERT(BP_GET_DEDUP(zio->io_bp)); 5568 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0); 5569 } 5570 } 5571 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS); 5572 /* if it's not anon, we are doing a scrub */ 5573 if (exists == NULL && hdr->b_l1hdr.b_state == arc_anon) 5574 arc_access(hdr, hash_lock); 5575 mutex_exit(hash_lock); 5576 } else { 5577 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS); 5578 } 5579 5580 ASSERT(!refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 5581 callback->awcb_done(zio, buf, callback->awcb_private); 5582 5583 kmem_free(callback, sizeof (arc_write_callback_t)); 5584} 5585 5586zio_t * 5587arc_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, arc_buf_t *buf, 5588 boolean_t l2arc, const zio_prop_t *zp, arc_done_func_t *ready, 5589 arc_done_func_t *children_ready, arc_done_func_t *physdone, 5590 arc_done_func_t *done, void *private, zio_priority_t priority, 5591 int zio_flags, const zbookmark_phys_t *zb) 5592{ 5593 arc_buf_hdr_t *hdr = buf->b_hdr; 5594 arc_write_callback_t *callback; 5595 zio_t *zio; 5596 5597 ASSERT3P(ready, !=, NULL); 5598 ASSERT3P(done, !=, NULL); 5599 ASSERT(!HDR_IO_ERROR(hdr)); 5600 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 5601 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL); 5602 ASSERT3U(hdr->b_l1hdr.b_bufcnt, >, 0); 5603 if (l2arc) 5604 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE); 5605 callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP); 5606 callback->awcb_ready = ready; 5607 callback->awcb_children_ready = children_ready; 5608 callback->awcb_physdone = physdone; 5609 callback->awcb_done = done; 5610 callback->awcb_private = private; 5611 callback->awcb_buf = buf; 5612 5613 /* 5614 * The hdr's b_pdata is now stale, free it now. A new data block 5615 * will be allocated when the zio pipeline calls arc_write_ready(). 5616 */ 5617 if (hdr->b_l1hdr.b_pdata != NULL) { 5618 /* 5619 * If the buf is currently sharing the data block with 5620 * the hdr then we need to break that relationship here. 5621 * The hdr will remain with a NULL data pointer and the 5622 * buf will take sole ownership of the block. 5623 */ 5624 if (arc_buf_is_shared(buf)) { 5625 ASSERT(ARC_BUF_LAST(buf)); 5626 arc_unshare_buf(hdr, buf); 5627 } else { 5628 arc_hdr_free_pdata(hdr); 5629 } 5630 VERIFY3P(buf->b_data, !=, NULL); 5631 arc_hdr_set_compress(hdr, ZIO_COMPRESS_OFF); 5632 } 5633 ASSERT(!arc_buf_is_shared(buf)); 5634 ASSERT3P(hdr->b_l1hdr.b_pdata, ==, NULL); 5635 5636 zio = zio_write(pio, spa, txg, bp, buf->b_data, HDR_GET_LSIZE(hdr), zp, 5637 arc_write_ready, 5638 (children_ready != NULL) ? arc_write_children_ready : NULL, 5639 arc_write_physdone, arc_write_done, callback, 5640 priority, zio_flags, zb); 5641 5642 return (zio); 5643} 5644 5645static int 5646arc_memory_throttle(uint64_t reserve, uint64_t txg) 5647{ 5648#ifdef _KERNEL 5649 uint64_t available_memory = ptob(freemem); 5650 static uint64_t page_load = 0; 5651 static uint64_t last_txg = 0; 5652 5653#if defined(__i386) || !defined(UMA_MD_SMALL_ALLOC) 5654 available_memory = 5655 MIN(available_memory, ptob(vmem_size(heap_arena, VMEM_FREE))); 5656#endif 5657 5658 if (freemem > (uint64_t)physmem * arc_lotsfree_percent / 100) 5659 return (0); 5660 5661 if (txg > last_txg) { 5662 last_txg = txg; 5663 page_load = 0; 5664 } 5665 /* 5666 * If we are in pageout, we know that memory is already tight, 5667 * the arc is already going to be evicting, so we just want to 5668 * continue to let page writes occur as quickly as possible. 5669 */ 5670 if (curproc == pageproc) { 5671 if (page_load > MAX(ptob(minfree), available_memory) / 4) 5672 return (SET_ERROR(ERESTART)); 5673 /* Note: reserve is inflated, so we deflate */ 5674 page_load += reserve / 8; 5675 return (0); 5676 } else if (page_load > 0 && arc_reclaim_needed()) { 5677 /* memory is low, delay before restarting */ 5678 ARCSTAT_INCR(arcstat_memory_throttle_count, 1); 5679 return (SET_ERROR(EAGAIN)); 5680 } 5681 page_load = 0; 5682#endif 5683 return (0); 5684} 5685 5686void 5687arc_tempreserve_clear(uint64_t reserve) 5688{ 5689 atomic_add_64(&arc_tempreserve, -reserve); 5690 ASSERT((int64_t)arc_tempreserve >= 0); 5691} 5692 5693int 5694arc_tempreserve_space(uint64_t reserve, uint64_t txg) 5695{ 5696 int error; 5697 uint64_t anon_size; 5698 5699 if (reserve > arc_c/4 && !arc_no_grow) { 5700 arc_c = MIN(arc_c_max, reserve * 4); 5701 DTRACE_PROBE1(arc__set_reserve, uint64_t, arc_c); 5702 } 5703 if (reserve > arc_c) 5704 return (SET_ERROR(ENOMEM)); 5705 5706 /* 5707 * Don't count loaned bufs as in flight dirty data to prevent long 5708 * network delays from blocking transactions that are ready to be 5709 * assigned to a txg. 5710 */ 5711 anon_size = MAX((int64_t)(refcount_count(&arc_anon->arcs_size) - 5712 arc_loaned_bytes), 0); 5713 5714 /* 5715 * Writes will, almost always, require additional memory allocations 5716 * in order to compress/encrypt/etc the data. We therefore need to 5717 * make sure that there is sufficient available memory for this. 5718 */ 5719 error = arc_memory_throttle(reserve, txg); 5720 if (error != 0) 5721 return (error); 5722 5723 /* 5724 * Throttle writes when the amount of dirty data in the cache 5725 * gets too large. We try to keep the cache less than half full 5726 * of dirty blocks so that our sync times don't grow too large. 5727 * Note: if two requests come in concurrently, we might let them 5728 * both succeed, when one of them should fail. Not a huge deal. 5729 */ 5730 5731 if (reserve + arc_tempreserve + anon_size > arc_c / 2 && 5732 anon_size > arc_c / 4) { 5733 uint64_t meta_esize = 5734 refcount_count(&arc_anon->arcs_esize[ARC_BUFC_METADATA]); 5735 uint64_t data_esize = 5736 refcount_count(&arc_anon->arcs_esize[ARC_BUFC_DATA]); 5737 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK " 5738 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n", 5739 arc_tempreserve >> 10, meta_esize >> 10, 5740 data_esize >> 10, reserve >> 10, arc_c >> 10); 5741 return (SET_ERROR(ERESTART)); 5742 } 5743 atomic_add_64(&arc_tempreserve, reserve); 5744 return (0); 5745} 5746 5747static void 5748arc_kstat_update_state(arc_state_t *state, kstat_named_t *size, 5749 kstat_named_t *evict_data, kstat_named_t *evict_metadata) 5750{ 5751 size->value.ui64 = refcount_count(&state->arcs_size); 5752 evict_data->value.ui64 = 5753 refcount_count(&state->arcs_esize[ARC_BUFC_DATA]); 5754 evict_metadata->value.ui64 = 5755 refcount_count(&state->arcs_esize[ARC_BUFC_METADATA]); 5756} 5757 5758static int 5759arc_kstat_update(kstat_t *ksp, int rw) 5760{ 5761 arc_stats_t *as = ksp->ks_data; 5762 5763 if (rw == KSTAT_WRITE) { 5764 return (EACCES); 5765 } else { 5766 arc_kstat_update_state(arc_anon, 5767 &as->arcstat_anon_size, 5768 &as->arcstat_anon_evictable_data, 5769 &as->arcstat_anon_evictable_metadata); 5770 arc_kstat_update_state(arc_mru, 5771 &as->arcstat_mru_size, 5772 &as->arcstat_mru_evictable_data, 5773 &as->arcstat_mru_evictable_metadata); 5774 arc_kstat_update_state(arc_mru_ghost, 5775 &as->arcstat_mru_ghost_size, 5776 &as->arcstat_mru_ghost_evictable_data, 5777 &as->arcstat_mru_ghost_evictable_metadata); 5778 arc_kstat_update_state(arc_mfu, 5779 &as->arcstat_mfu_size, 5780 &as->arcstat_mfu_evictable_data, 5781 &as->arcstat_mfu_evictable_metadata); 5782 arc_kstat_update_state(arc_mfu_ghost, 5783 &as->arcstat_mfu_ghost_size, 5784 &as->arcstat_mfu_ghost_evictable_data, 5785 &as->arcstat_mfu_ghost_evictable_metadata); 5786 } 5787 5788 return (0); 5789} 5790 5791/* 5792 * This function *must* return indices evenly distributed between all 5793 * sublists of the multilist. This is needed due to how the ARC eviction 5794 * code is laid out; arc_evict_state() assumes ARC buffers are evenly 5795 * distributed between all sublists and uses this assumption when 5796 * deciding which sublist to evict from and how much to evict from it. 5797 */ 5798unsigned int 5799arc_state_multilist_index_func(multilist_t *ml, void *obj) 5800{ 5801 arc_buf_hdr_t *hdr = obj; 5802 5803 /* 5804 * We rely on b_dva to generate evenly distributed index 5805 * numbers using buf_hash below. So, as an added precaution, 5806 * let's make sure we never add empty buffers to the arc lists. 5807 */ 5808 ASSERT(!HDR_EMPTY(hdr)); 5809 5810 /* 5811 * The assumption here, is the hash value for a given 5812 * arc_buf_hdr_t will remain constant throughout it's lifetime 5813 * (i.e. it's b_spa, b_dva, and b_birth fields don't change). 5814 * Thus, we don't need to store the header's sublist index 5815 * on insertion, as this index can be recalculated on removal. 5816 * 5817 * Also, the low order bits of the hash value are thought to be 5818 * distributed evenly. Otherwise, in the case that the multilist 5819 * has a power of two number of sublists, each sublists' usage 5820 * would not be evenly distributed. 5821 */ 5822 return (buf_hash(hdr->b_spa, &hdr->b_dva, hdr->b_birth) % 5823 multilist_get_num_sublists(ml)); 5824} 5825 5826#ifdef _KERNEL 5827static eventhandler_tag arc_event_lowmem = NULL; 5828 5829static void 5830arc_lowmem(void *arg __unused, int howto __unused) 5831{ 5832 5833 mutex_enter(&arc_reclaim_lock); 5834 /* XXX: Memory deficit should be passed as argument. */ 5835 needfree = btoc(arc_c >> arc_shrink_shift); 5836 DTRACE_PROBE(arc__needfree); 5837 cv_signal(&arc_reclaim_thread_cv); 5838 5839 /* 5840 * It is unsafe to block here in arbitrary threads, because we can come 5841 * here from ARC itself and may hold ARC locks and thus risk a deadlock 5842 * with ARC reclaim thread. 5843 */ 5844 if (curproc == pageproc) 5845 (void) cv_wait(&arc_reclaim_waiters_cv, &arc_reclaim_lock); 5846 mutex_exit(&arc_reclaim_lock); 5847} 5848#endif 5849 5850static void 5851arc_state_init(void) 5852{ 5853 arc_anon = &ARC_anon; 5854 arc_mru = &ARC_mru; 5855 arc_mru_ghost = &ARC_mru_ghost; 5856 arc_mfu = &ARC_mfu; 5857 arc_mfu_ghost = &ARC_mfu_ghost; 5858 arc_l2c_only = &ARC_l2c_only; 5859 5860 multilist_create(&arc_mru->arcs_list[ARC_BUFC_METADATA], 5861 sizeof (arc_buf_hdr_t), 5862 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5863 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5864 multilist_create(&arc_mru->arcs_list[ARC_BUFC_DATA], 5865 sizeof (arc_buf_hdr_t), 5866 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5867 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5868 multilist_create(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA], 5869 sizeof (arc_buf_hdr_t), 5870 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5871 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5872 multilist_create(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA], 5873 sizeof (arc_buf_hdr_t), 5874 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5875 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5876 multilist_create(&arc_mfu->arcs_list[ARC_BUFC_METADATA], 5877 sizeof (arc_buf_hdr_t), 5878 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5879 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5880 multilist_create(&arc_mfu->arcs_list[ARC_BUFC_DATA], 5881 sizeof (arc_buf_hdr_t), 5882 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5883 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5884 multilist_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA], 5885 sizeof (arc_buf_hdr_t), 5886 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5887 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5888 multilist_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA], 5889 sizeof (arc_buf_hdr_t), 5890 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5891 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5892 multilist_create(&arc_l2c_only->arcs_list[ARC_BUFC_METADATA], 5893 sizeof (arc_buf_hdr_t), 5894 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5895 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5896 multilist_create(&arc_l2c_only->arcs_list[ARC_BUFC_DATA], 5897 sizeof (arc_buf_hdr_t), 5898 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5899 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5900 5901 refcount_create(&arc_anon->arcs_esize[ARC_BUFC_METADATA]); 5902 refcount_create(&arc_anon->arcs_esize[ARC_BUFC_DATA]); 5903 refcount_create(&arc_mru->arcs_esize[ARC_BUFC_METADATA]); 5904 refcount_create(&arc_mru->arcs_esize[ARC_BUFC_DATA]); 5905 refcount_create(&arc_mru_ghost->arcs_esize[ARC_BUFC_METADATA]); 5906 refcount_create(&arc_mru_ghost->arcs_esize[ARC_BUFC_DATA]); 5907 refcount_create(&arc_mfu->arcs_esize[ARC_BUFC_METADATA]); 5908 refcount_create(&arc_mfu->arcs_esize[ARC_BUFC_DATA]); 5909 refcount_create(&arc_mfu_ghost->arcs_esize[ARC_BUFC_METADATA]); 5910 refcount_create(&arc_mfu_ghost->arcs_esize[ARC_BUFC_DATA]); 5911 refcount_create(&arc_l2c_only->arcs_esize[ARC_BUFC_METADATA]); 5912 refcount_create(&arc_l2c_only->arcs_esize[ARC_BUFC_DATA]); 5913 5914 refcount_create(&arc_anon->arcs_size); 5915 refcount_create(&arc_mru->arcs_size); 5916 refcount_create(&arc_mru_ghost->arcs_size); 5917 refcount_create(&arc_mfu->arcs_size); 5918 refcount_create(&arc_mfu_ghost->arcs_size); 5919 refcount_create(&arc_l2c_only->arcs_size); 5920} 5921 5922static void 5923arc_state_fini(void) 5924{ 5925 refcount_destroy(&arc_anon->arcs_esize[ARC_BUFC_METADATA]); 5926 refcount_destroy(&arc_anon->arcs_esize[ARC_BUFC_DATA]); 5927 refcount_destroy(&arc_mru->arcs_esize[ARC_BUFC_METADATA]); 5928 refcount_destroy(&arc_mru->arcs_esize[ARC_BUFC_DATA]); 5929 refcount_destroy(&arc_mru_ghost->arcs_esize[ARC_BUFC_METADATA]); 5930 refcount_destroy(&arc_mru_ghost->arcs_esize[ARC_BUFC_DATA]); 5931 refcount_destroy(&arc_mfu->arcs_esize[ARC_BUFC_METADATA]); 5932 refcount_destroy(&arc_mfu->arcs_esize[ARC_BUFC_DATA]); 5933 refcount_destroy(&arc_mfu_ghost->arcs_esize[ARC_BUFC_METADATA]); 5934 refcount_destroy(&arc_mfu_ghost->arcs_esize[ARC_BUFC_DATA]); 5935 refcount_destroy(&arc_l2c_only->arcs_esize[ARC_BUFC_METADATA]); 5936 refcount_destroy(&arc_l2c_only->arcs_esize[ARC_BUFC_DATA]); 5937 5938 refcount_destroy(&arc_anon->arcs_size); 5939 refcount_destroy(&arc_mru->arcs_size); 5940 refcount_destroy(&arc_mru_ghost->arcs_size); 5941 refcount_destroy(&arc_mfu->arcs_size); 5942 refcount_destroy(&arc_mfu_ghost->arcs_size); 5943 refcount_destroy(&arc_l2c_only->arcs_size); 5944 5945 multilist_destroy(&arc_mru->arcs_list[ARC_BUFC_METADATA]); 5946 multilist_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA]); 5947 multilist_destroy(&arc_mfu->arcs_list[ARC_BUFC_METADATA]); 5948 multilist_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA]); 5949 multilist_destroy(&arc_mru->arcs_list[ARC_BUFC_DATA]); 5950 multilist_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA]); 5951 multilist_destroy(&arc_mfu->arcs_list[ARC_BUFC_DATA]); 5952 multilist_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA]); 5953} 5954 5955uint64_t 5956arc_max_bytes(void) 5957{ 5958 return (arc_c_max); 5959} 5960 5961void 5962arc_init(void) 5963{ 5964 int i, prefetch_tunable_set = 0; 5965 5966 mutex_init(&arc_reclaim_lock, NULL, MUTEX_DEFAULT, NULL); 5967 cv_init(&arc_reclaim_thread_cv, NULL, CV_DEFAULT, NULL); 5968 cv_init(&arc_reclaim_waiters_cv, NULL, CV_DEFAULT, NULL); 5969 5970 /* Convert seconds to clock ticks */ 5971 arc_min_prefetch_lifespan = 1 * hz; 5972 5973 /* Start out with 1/8 of all memory */ 5974 arc_c = kmem_size() / 8; 5975 5976#ifdef illumos 5977#ifdef _KERNEL 5978 /* 5979 * On architectures where the physical memory can be larger 5980 * than the addressable space (intel in 32-bit mode), we may 5981 * need to limit the cache to 1/8 of VM size. 5982 */ 5983 arc_c = MIN(arc_c, vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 8); 5984#endif 5985#endif /* illumos */ 5986 /* set min cache to 1/32 of all memory, or arc_abs_min, whichever is more */ 5987 arc_c_min = MAX(arc_c / 4, arc_abs_min); 5988 /* set max to 1/2 of all memory, or all but 1GB, whichever is more */ 5989 if (arc_c * 8 >= 1 << 30) 5990 arc_c_max = (arc_c * 8) - (1 << 30); 5991 else 5992 arc_c_max = arc_c_min; 5993 arc_c_max = MAX(arc_c * 5, arc_c_max); 5994 5995 /* 5996 * In userland, there's only the memory pressure that we artificially 5997 * create (see arc_available_memory()). Don't let arc_c get too 5998 * small, because it can cause transactions to be larger than 5999 * arc_c, causing arc_tempreserve_space() to fail. 6000 */ 6001#ifndef _KERNEL 6002 arc_c_min = arc_c_max / 2; 6003#endif 6004 6005#ifdef _KERNEL 6006 /* 6007 * Allow the tunables to override our calculations if they are 6008 * reasonable. 6009 */ 6010 if (zfs_arc_max > arc_abs_min && zfs_arc_max < kmem_size()) { 6011 arc_c_max = zfs_arc_max; 6012 arc_c_min = MIN(arc_c_min, arc_c_max); 6013 } 6014 if (zfs_arc_min > arc_abs_min && zfs_arc_min <= arc_c_max) 6015 arc_c_min = zfs_arc_min; 6016#endif 6017 6018 arc_c = arc_c_max; 6019 arc_p = (arc_c >> 1); 6020 arc_size = 0; 6021 6022 /* limit meta-data to 1/4 of the arc capacity */ 6023 arc_meta_limit = arc_c_max / 4; 6024 6025 /* Allow the tunable to override if it is reasonable */ 6026 if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max) 6027 arc_meta_limit = zfs_arc_meta_limit; 6028 6029 if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0) 6030 arc_c_min = arc_meta_limit / 2; 6031 6032 if (zfs_arc_meta_min > 0) { 6033 arc_meta_min = zfs_arc_meta_min; 6034 } else { 6035 arc_meta_min = arc_c_min / 2; 6036 } 6037 6038 if (zfs_arc_grow_retry > 0) 6039 arc_grow_retry = zfs_arc_grow_retry; 6040 6041 if (zfs_arc_shrink_shift > 0) 6042 arc_shrink_shift = zfs_arc_shrink_shift; 6043 6044 /* 6045 * Ensure that arc_no_grow_shift is less than arc_shrink_shift. 6046 */ 6047 if (arc_no_grow_shift >= arc_shrink_shift) 6048 arc_no_grow_shift = arc_shrink_shift - 1; 6049 6050 if (zfs_arc_p_min_shift > 0) 6051 arc_p_min_shift = zfs_arc_p_min_shift; 6052 6053 if (zfs_arc_num_sublists_per_state < 1) 6054 zfs_arc_num_sublists_per_state = MAX(max_ncpus, 1); 6055 6056 /* if kmem_flags are set, lets try to use less memory */ 6057 if (kmem_debugging()) 6058 arc_c = arc_c / 2; 6059 if (arc_c < arc_c_min) 6060 arc_c = arc_c_min; 6061 6062 zfs_arc_min = arc_c_min; 6063 zfs_arc_max = arc_c_max; 6064 6065 arc_state_init(); 6066 buf_init(); 6067 6068 arc_reclaim_thread_exit = B_FALSE; 6069 6070 arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED, 6071 sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL); 6072 6073 if (arc_ksp != NULL) { 6074 arc_ksp->ks_data = &arc_stats; 6075 arc_ksp->ks_update = arc_kstat_update; 6076 kstat_install(arc_ksp); 6077 } 6078 6079 (void) thread_create(NULL, 0, arc_reclaim_thread, NULL, 0, &p0, 6080 TS_RUN, minclsyspri); 6081 6082#ifdef _KERNEL 6083 arc_event_lowmem = EVENTHANDLER_REGISTER(vm_lowmem, arc_lowmem, NULL, 6084 EVENTHANDLER_PRI_FIRST); 6085#endif 6086 6087 arc_dead = B_FALSE; 6088 arc_warm = B_FALSE; 6089 6090 /* 6091 * Calculate maximum amount of dirty data per pool. 6092 * 6093 * If it has been set by /etc/system, take that. 6094 * Otherwise, use a percentage of physical memory defined by 6095 * zfs_dirty_data_max_percent (default 10%) with a cap at 6096 * zfs_dirty_data_max_max (default 4GB). 6097 */ 6098 if (zfs_dirty_data_max == 0) { 6099 zfs_dirty_data_max = ptob(physmem) * 6100 zfs_dirty_data_max_percent / 100; 6101 zfs_dirty_data_max = MIN(zfs_dirty_data_max, 6102 zfs_dirty_data_max_max); 6103 } 6104 6105#ifdef _KERNEL 6106 if (TUNABLE_INT_FETCH("vfs.zfs.prefetch_disable", &zfs_prefetch_disable)) 6107 prefetch_tunable_set = 1; 6108 6109#ifdef __i386__ 6110 if (prefetch_tunable_set == 0) { 6111 printf("ZFS NOTICE: Prefetch is disabled by default on i386 " 6112 "-- to enable,\n"); 6113 printf(" add \"vfs.zfs.prefetch_disable=0\" " 6114 "to /boot/loader.conf.\n"); 6115 zfs_prefetch_disable = 1; 6116 } 6117#else 6118 if ((((uint64_t)physmem * PAGESIZE) < (1ULL << 32)) && 6119 prefetch_tunable_set == 0) { 6120 printf("ZFS NOTICE: Prefetch is disabled by default if less " 6121 "than 4GB of RAM is present;\n" 6122 " to enable, add \"vfs.zfs.prefetch_disable=0\" " 6123 "to /boot/loader.conf.\n"); 6124 zfs_prefetch_disable = 1; 6125 } 6126#endif 6127 /* Warn about ZFS memory and address space requirements. */ 6128 if (((uint64_t)physmem * PAGESIZE) < (256 + 128 + 64) * (1 << 20)) { 6129 printf("ZFS WARNING: Recommended minimum RAM size is 512MB; " 6130 "expect unstable behavior.\n"); 6131 } 6132 if (kmem_size() < 512 * (1 << 20)) { 6133 printf("ZFS WARNING: Recommended minimum kmem_size is 512MB; " 6134 "expect unstable behavior.\n"); 6135 printf(" Consider tuning vm.kmem_size and " 6136 "vm.kmem_size_max\n"); 6137 printf(" in /boot/loader.conf.\n"); 6138 } 6139#endif 6140} 6141 6142void 6143arc_fini(void) 6144{ 6145 mutex_enter(&arc_reclaim_lock); 6146 arc_reclaim_thread_exit = B_TRUE; 6147 /* 6148 * The reclaim thread will set arc_reclaim_thread_exit back to 6149 * B_FALSE when it is finished exiting; we're waiting for that. 6150 */ 6151 while (arc_reclaim_thread_exit) { 6152 cv_signal(&arc_reclaim_thread_cv); 6153 cv_wait(&arc_reclaim_thread_cv, &arc_reclaim_lock); 6154 } 6155 mutex_exit(&arc_reclaim_lock); 6156 6157 /* Use B_TRUE to ensure *all* buffers are evicted */ 6158 arc_flush(NULL, B_TRUE); 6159 6160 arc_dead = B_TRUE; 6161 6162 if (arc_ksp != NULL) { 6163 kstat_delete(arc_ksp); 6164 arc_ksp = NULL; 6165 } 6166 6167 mutex_destroy(&arc_reclaim_lock); 6168 cv_destroy(&arc_reclaim_thread_cv); 6169 cv_destroy(&arc_reclaim_waiters_cv); 6170 6171 arc_state_fini(); 6172 buf_fini(); 6173 6174 ASSERT0(arc_loaned_bytes); 6175 6176#ifdef _KERNEL 6177 if (arc_event_lowmem != NULL) 6178 EVENTHANDLER_DEREGISTER(vm_lowmem, arc_event_lowmem); 6179#endif 6180} 6181 6182/* 6183 * Level 2 ARC 6184 * 6185 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk. 6186 * It uses dedicated storage devices to hold cached data, which are populated 6187 * using large infrequent writes. The main role of this cache is to boost 6188 * the performance of random read workloads. The intended L2ARC devices 6189 * include short-stroked disks, solid state disks, and other media with 6190 * substantially faster read latency than disk. 6191 * 6192 * +-----------------------+ 6193 * | ARC | 6194 * +-----------------------+ 6195 * | ^ ^ 6196 * | | | 6197 * l2arc_feed_thread() arc_read() 6198 * | | | 6199 * | l2arc read | 6200 * V | | 6201 * +---------------+ | 6202 * | L2ARC | | 6203 * +---------------+ | 6204 * | ^ | 6205 * l2arc_write() | | 6206 * | | | 6207 * V | | 6208 * +-------+ +-------+ 6209 * | vdev | | vdev | 6210 * | cache | | cache | 6211 * +-------+ +-------+ 6212 * +=========+ .-----. 6213 * : L2ARC : |-_____-| 6214 * : devices : | Disks | 6215 * +=========+ `-_____-' 6216 * 6217 * Read requests are satisfied from the following sources, in order: 6218 * 6219 * 1) ARC 6220 * 2) vdev cache of L2ARC devices 6221 * 3) L2ARC devices 6222 * 4) vdev cache of disks 6223 * 5) disks 6224 * 6225 * Some L2ARC device types exhibit extremely slow write performance. 6226 * To accommodate for this there are some significant differences between 6227 * the L2ARC and traditional cache design: 6228 * 6229 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from 6230 * the ARC behave as usual, freeing buffers and placing headers on ghost 6231 * lists. The ARC does not send buffers to the L2ARC during eviction as 6232 * this would add inflated write latencies for all ARC memory pressure. 6233 * 6234 * 2. The L2ARC attempts to cache data from the ARC before it is evicted. 6235 * It does this by periodically scanning buffers from the eviction-end of 6236 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are 6237 * not already there. It scans until a headroom of buffers is satisfied, 6238 * which itself is a buffer for ARC eviction. If a compressible buffer is 6239 * found during scanning and selected for writing to an L2ARC device, we 6240 * temporarily boost scanning headroom during the next scan cycle to make 6241 * sure we adapt to compression effects (which might significantly reduce 6242 * the data volume we write to L2ARC). The thread that does this is 6243 * l2arc_feed_thread(), illustrated below; example sizes are included to 6244 * provide a better sense of ratio than this diagram: 6245 * 6246 * head --> tail 6247 * +---------------------+----------+ 6248 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC 6249 * +---------------------+----------+ | o L2ARC eligible 6250 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer 6251 * +---------------------+----------+ | 6252 * 15.9 Gbytes ^ 32 Mbytes | 6253 * headroom | 6254 * l2arc_feed_thread() 6255 * | 6256 * l2arc write hand <--[oooo]--' 6257 * | 8 Mbyte 6258 * | write max 6259 * V 6260 * +==============================+ 6261 * L2ARC dev |####|#|###|###| |####| ... | 6262 * +==============================+ 6263 * 32 Gbytes 6264 * 6265 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of 6266 * evicted, then the L2ARC has cached a buffer much sooner than it probably 6267 * needed to, potentially wasting L2ARC device bandwidth and storage. It is 6268 * safe to say that this is an uncommon case, since buffers at the end of 6269 * the ARC lists have moved there due to inactivity. 6270 * 6271 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom, 6272 * then the L2ARC simply misses copying some buffers. This serves as a 6273 * pressure valve to prevent heavy read workloads from both stalling the ARC 6274 * with waits and clogging the L2ARC with writes. This also helps prevent 6275 * the potential for the L2ARC to churn if it attempts to cache content too 6276 * quickly, such as during backups of the entire pool. 6277 * 6278 * 5. After system boot and before the ARC has filled main memory, there are 6279 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru 6280 * lists can remain mostly static. Instead of searching from tail of these 6281 * lists as pictured, the l2arc_feed_thread() will search from the list heads 6282 * for eligible buffers, greatly increasing its chance of finding them. 6283 * 6284 * The L2ARC device write speed is also boosted during this time so that 6285 * the L2ARC warms up faster. Since there have been no ARC evictions yet, 6286 * there are no L2ARC reads, and no fear of degrading read performance 6287 * through increased writes. 6288 * 6289 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that 6290 * the vdev queue can aggregate them into larger and fewer writes. Each 6291 * device is written to in a rotor fashion, sweeping writes through 6292 * available space then repeating. 6293 * 6294 * 7. The L2ARC does not store dirty content. It never needs to flush 6295 * write buffers back to disk based storage. 6296 * 6297 * 8. If an ARC buffer is written (and dirtied) which also exists in the 6298 * L2ARC, the now stale L2ARC buffer is immediately dropped. 6299 * 6300 * The performance of the L2ARC can be tweaked by a number of tunables, which 6301 * may be necessary for different workloads: 6302 * 6303 * l2arc_write_max max write bytes per interval 6304 * l2arc_write_boost extra write bytes during device warmup 6305 * l2arc_noprefetch skip caching prefetched buffers 6306 * l2arc_headroom number of max device writes to precache 6307 * l2arc_headroom_boost when we find compressed buffers during ARC 6308 * scanning, we multiply headroom by this 6309 * percentage factor for the next scan cycle, 6310 * since more compressed buffers are likely to 6311 * be present 6312 * l2arc_feed_secs seconds between L2ARC writing 6313 * 6314 * Tunables may be removed or added as future performance improvements are 6315 * integrated, and also may become zpool properties. 6316 * 6317 * There are three key functions that control how the L2ARC warms up: 6318 * 6319 * l2arc_write_eligible() check if a buffer is eligible to cache 6320 * l2arc_write_size() calculate how much to write 6321 * l2arc_write_interval() calculate sleep delay between writes 6322 * 6323 * These three functions determine what to write, how much, and how quickly 6324 * to send writes. 6325 */ 6326 6327static boolean_t 6328l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *hdr) 6329{ 6330 /* 6331 * A buffer is *not* eligible for the L2ARC if it: 6332 * 1. belongs to a different spa. 6333 * 2. is already cached on the L2ARC. 6334 * 3. has an I/O in progress (it may be an incomplete read). 6335 * 4. is flagged not eligible (zfs property). 6336 */ 6337 if (hdr->b_spa != spa_guid) { 6338 ARCSTAT_BUMP(arcstat_l2_write_spa_mismatch); 6339 return (B_FALSE); 6340 } 6341 if (HDR_HAS_L2HDR(hdr)) { 6342 ARCSTAT_BUMP(arcstat_l2_write_in_l2); 6343 return (B_FALSE); 6344 } 6345 if (HDR_IO_IN_PROGRESS(hdr)) { 6346 ARCSTAT_BUMP(arcstat_l2_write_hdr_io_in_progress); 6347 return (B_FALSE); 6348 } 6349 if (!HDR_L2CACHE(hdr)) { 6350 ARCSTAT_BUMP(arcstat_l2_write_not_cacheable); 6351 return (B_FALSE); 6352 } 6353 6354 return (B_TRUE); 6355} 6356 6357static uint64_t 6358l2arc_write_size(void) 6359{ 6360 uint64_t size; 6361 6362 /* 6363 * Make sure our globals have meaningful values in case the user 6364 * altered them. 6365 */ 6366 size = l2arc_write_max; 6367 if (size == 0) { 6368 cmn_err(CE_NOTE, "Bad value for l2arc_write_max, value must " 6369 "be greater than zero, resetting it to the default (%d)", 6370 L2ARC_WRITE_SIZE); 6371 size = l2arc_write_max = L2ARC_WRITE_SIZE; 6372 } 6373 6374 if (arc_warm == B_FALSE) 6375 size += l2arc_write_boost; 6376 6377 return (size); 6378 6379} 6380 6381static clock_t 6382l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote) 6383{ 6384 clock_t interval, next, now; 6385 6386 /* 6387 * If the ARC lists are busy, increase our write rate; if the 6388 * lists are stale, idle back. This is achieved by checking 6389 * how much we previously wrote - if it was more than half of 6390 * what we wanted, schedule the next write much sooner. 6391 */ 6392 if (l2arc_feed_again && wrote > (wanted / 2)) 6393 interval = (hz * l2arc_feed_min_ms) / 1000; 6394 else 6395 interval = hz * l2arc_feed_secs; 6396 6397 now = ddi_get_lbolt(); 6398 next = MAX(now, MIN(now + interval, began + interval)); 6399 6400 return (next); 6401} 6402 6403/* 6404 * Cycle through L2ARC devices. This is how L2ARC load balances. 6405 * If a device is returned, this also returns holding the spa config lock. 6406 */ 6407static l2arc_dev_t * 6408l2arc_dev_get_next(void) 6409{ 6410 l2arc_dev_t *first, *next = NULL; 6411 6412 /* 6413 * Lock out the removal of spas (spa_namespace_lock), then removal 6414 * of cache devices (l2arc_dev_mtx). Once a device has been selected, 6415 * both locks will be dropped and a spa config lock held instead. 6416 */ 6417 mutex_enter(&spa_namespace_lock); 6418 mutex_enter(&l2arc_dev_mtx); 6419 6420 /* if there are no vdevs, there is nothing to do */ 6421 if (l2arc_ndev == 0) 6422 goto out; 6423 6424 first = NULL; 6425 next = l2arc_dev_last; 6426 do { 6427 /* loop around the list looking for a non-faulted vdev */ 6428 if (next == NULL) { 6429 next = list_head(l2arc_dev_list); 6430 } else { 6431 next = list_next(l2arc_dev_list, next); 6432 if (next == NULL) 6433 next = list_head(l2arc_dev_list); 6434 } 6435 6436 /* if we have come back to the start, bail out */ 6437 if (first == NULL) 6438 first = next; 6439 else if (next == first) 6440 break; 6441 6442 } while (vdev_is_dead(next->l2ad_vdev)); 6443 6444 /* if we were unable to find any usable vdevs, return NULL */ 6445 if (vdev_is_dead(next->l2ad_vdev)) 6446 next = NULL; 6447 6448 l2arc_dev_last = next; 6449 6450out: 6451 mutex_exit(&l2arc_dev_mtx); 6452 6453 /* 6454 * Grab the config lock to prevent the 'next' device from being 6455 * removed while we are writing to it. 6456 */ 6457 if (next != NULL) 6458 spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER); 6459 mutex_exit(&spa_namespace_lock); 6460 6461 return (next); 6462} 6463 6464/* 6465 * Free buffers that were tagged for destruction. 6466 */ 6467static void 6468l2arc_do_free_on_write() 6469{ 6470 list_t *buflist; 6471 l2arc_data_free_t *df, *df_prev; 6472 6473 mutex_enter(&l2arc_free_on_write_mtx); 6474 buflist = l2arc_free_on_write; 6475 6476 for (df = list_tail(buflist); df; df = df_prev) { 6477 df_prev = list_prev(buflist, df); 6478 ASSERT3P(df->l2df_data, !=, NULL); 6479 if (df->l2df_type == ARC_BUFC_METADATA) { 6480 zio_buf_free(df->l2df_data, df->l2df_size); 6481 } else { 6482 ASSERT(df->l2df_type == ARC_BUFC_DATA); 6483 zio_data_buf_free(df->l2df_data, df->l2df_size); 6484 } 6485 list_remove(buflist, df); 6486 kmem_free(df, sizeof (l2arc_data_free_t)); 6487 } 6488 6489 mutex_exit(&l2arc_free_on_write_mtx); 6490} 6491 6492/* 6493 * A write to a cache device has completed. Update all headers to allow 6494 * reads from these buffers to begin. 6495 */ 6496static void 6497l2arc_write_done(zio_t *zio) 6498{ 6499 l2arc_write_callback_t *cb; 6500 l2arc_dev_t *dev; 6501 list_t *buflist; 6502 arc_buf_hdr_t *head, *hdr, *hdr_prev; 6503 kmutex_t *hash_lock; 6504 int64_t bytes_dropped = 0; 6505 6506 cb = zio->io_private; 6507 ASSERT3P(cb, !=, NULL); 6508 dev = cb->l2wcb_dev; 6509 ASSERT3P(dev, !=, NULL); 6510 head = cb->l2wcb_head; 6511 ASSERT3P(head, !=, NULL); 6512 buflist = &dev->l2ad_buflist; 6513 ASSERT3P(buflist, !=, NULL); 6514 DTRACE_PROBE2(l2arc__iodone, zio_t *, zio, 6515 l2arc_write_callback_t *, cb); 6516 6517 if (zio->io_error != 0) 6518 ARCSTAT_BUMP(arcstat_l2_writes_error); 6519 6520 /* 6521 * All writes completed, or an error was hit. 6522 */ 6523top: 6524 mutex_enter(&dev->l2ad_mtx); 6525 for (hdr = list_prev(buflist, head); hdr; hdr = hdr_prev) { 6526 hdr_prev = list_prev(buflist, hdr); 6527 6528 hash_lock = HDR_LOCK(hdr); 6529 6530 /* 6531 * We cannot use mutex_enter or else we can deadlock 6532 * with l2arc_write_buffers (due to swapping the order 6533 * the hash lock and l2ad_mtx are taken). 6534 */ 6535 if (!mutex_tryenter(hash_lock)) { 6536 /* 6537 * Missed the hash lock. We must retry so we 6538 * don't leave the ARC_FLAG_L2_WRITING bit set. 6539 */ 6540 ARCSTAT_BUMP(arcstat_l2_writes_lock_retry); 6541 6542 /* 6543 * We don't want to rescan the headers we've 6544 * already marked as having been written out, so 6545 * we reinsert the head node so we can pick up 6546 * where we left off. 6547 */ 6548 list_remove(buflist, head); 6549 list_insert_after(buflist, hdr, head); 6550 6551 mutex_exit(&dev->l2ad_mtx); 6552 6553 /* 6554 * We wait for the hash lock to become available 6555 * to try and prevent busy waiting, and increase 6556 * the chance we'll be able to acquire the lock 6557 * the next time around. 6558 */ 6559 mutex_enter(hash_lock); 6560 mutex_exit(hash_lock); 6561 goto top; 6562 } 6563 6564 /* 6565 * We could not have been moved into the arc_l2c_only 6566 * state while in-flight due to our ARC_FLAG_L2_WRITING 6567 * bit being set. Let's just ensure that's being enforced. 6568 */ 6569 ASSERT(HDR_HAS_L1HDR(hdr)); 6570 6571 if (zio->io_error != 0) { 6572 /* 6573 * Error - drop L2ARC entry. 6574 */ 6575 list_remove(buflist, hdr); 6576 l2arc_trim(hdr); 6577 arc_hdr_clear_flags(hdr, ARC_FLAG_HAS_L2HDR); 6578 6579 ARCSTAT_INCR(arcstat_l2_asize, -arc_hdr_size(hdr)); 6580 ARCSTAT_INCR(arcstat_l2_size, -HDR_GET_LSIZE(hdr)); 6581 6582 bytes_dropped += arc_hdr_size(hdr); 6583 (void) refcount_remove_many(&dev->l2ad_alloc, 6584 arc_hdr_size(hdr), hdr); 6585 } 6586 6587 /* 6588 * Allow ARC to begin reads and ghost list evictions to 6589 * this L2ARC entry. 6590 */ 6591 arc_hdr_clear_flags(hdr, ARC_FLAG_L2_WRITING); 6592 6593 mutex_exit(hash_lock); 6594 } 6595 6596 atomic_inc_64(&l2arc_writes_done); 6597 list_remove(buflist, head); 6598 ASSERT(!HDR_HAS_L1HDR(head)); 6599 kmem_cache_free(hdr_l2only_cache, head); 6600 mutex_exit(&dev->l2ad_mtx); 6601 6602 vdev_space_update(dev->l2ad_vdev, -bytes_dropped, 0, 0); 6603 6604 l2arc_do_free_on_write(); 6605 6606 kmem_free(cb, sizeof (l2arc_write_callback_t)); 6607} 6608 6609/* 6610 * A read to a cache device completed. Validate buffer contents before 6611 * handing over to the regular ARC routines. 6612 */ 6613static void 6614l2arc_read_done(zio_t *zio) 6615{ 6616 l2arc_read_callback_t *cb; 6617 arc_buf_hdr_t *hdr; 6618 kmutex_t *hash_lock; 6619 boolean_t valid_cksum; 6620 6621 ASSERT3P(zio->io_vd, !=, NULL); 6622 ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE); 6623 6624 spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd); 6625 6626 cb = zio->io_private; 6627 ASSERT3P(cb, !=, NULL); 6628 hdr = cb->l2rcb_hdr; 6629 ASSERT3P(hdr, !=, NULL); 6630 6631 hash_lock = HDR_LOCK(hdr); 6632 mutex_enter(hash_lock); 6633 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 6634 6635 /* 6636 * If the data was read into a temporary buffer, 6637 * move it and free the buffer. 6638 */ 6639 if (cb->l2rcb_data != NULL) { 6640 ASSERT3U(arc_hdr_size(hdr), <, zio->io_size); 6641 if (zio->io_error == 0) { 6642 bcopy(cb->l2rcb_data, hdr->b_l1hdr.b_pdata, 6643 arc_hdr_size(hdr)); 6644 } 6645 6646 /* 6647 * The following must be done regardless of whether 6648 * there was an error: 6649 * - free the temporary buffer 6650 * - point zio to the real ARC buffer 6651 * - set zio size accordingly 6652 * These are required because zio is either re-used for 6653 * an I/O of the block in the case of the error 6654 * or the zio is passed to arc_read_done() and it 6655 * needs real data. 6656 */ 6657 zio_data_buf_free(cb->l2rcb_data, zio->io_size); 6658 zio->io_size = zio->io_orig_size = arc_hdr_size(hdr); 6659 zio->io_data = zio->io_orig_data = hdr->b_l1hdr.b_pdata; 6660 } 6661 6662 ASSERT3P(zio->io_data, !=, NULL); 6663 6664 /* 6665 * Check this survived the L2ARC journey. 6666 */ 6667 ASSERT3P(zio->io_data, ==, hdr->b_l1hdr.b_pdata); 6668 zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */ 6669 zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */ 6670 6671 valid_cksum = arc_cksum_is_equal(hdr, zio); 6672 if (valid_cksum && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) { 6673 mutex_exit(hash_lock); 6674 zio->io_private = hdr; 6675 arc_read_done(zio); 6676 } else { 6677 mutex_exit(hash_lock); 6678 /* 6679 * Buffer didn't survive caching. Increment stats and 6680 * reissue to the original storage device. 6681 */ 6682 if (zio->io_error != 0) { 6683 ARCSTAT_BUMP(arcstat_l2_io_error); 6684 } else { 6685 zio->io_error = SET_ERROR(EIO); 6686 } 6687 if (!valid_cksum) 6688 ARCSTAT_BUMP(arcstat_l2_cksum_bad); 6689 6690 /* 6691 * If there's no waiter, issue an async i/o to the primary 6692 * storage now. If there *is* a waiter, the caller must 6693 * issue the i/o in a context where it's OK to block. 6694 */ 6695 if (zio->io_waiter == NULL) { 6696 zio_t *pio = zio_unique_parent(zio); 6697 6698 ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL); 6699 6700 zio_nowait(zio_read(pio, zio->io_spa, zio->io_bp, 6701 hdr->b_l1hdr.b_pdata, zio->io_size, arc_read_done, 6702 hdr, zio->io_priority, cb->l2rcb_flags, 6703 &cb->l2rcb_zb)); 6704 } 6705 } 6706 6707 kmem_free(cb, sizeof (l2arc_read_callback_t)); 6708} 6709 6710/* 6711 * This is the list priority from which the L2ARC will search for pages to 6712 * cache. This is used within loops (0..3) to cycle through lists in the 6713 * desired order. This order can have a significant effect on cache 6714 * performance. 6715 * 6716 * Currently the metadata lists are hit first, MFU then MRU, followed by 6717 * the data lists. This function returns a locked list, and also returns 6718 * the lock pointer. 6719 */ 6720static multilist_sublist_t * 6721l2arc_sublist_lock(int list_num) 6722{ 6723 multilist_t *ml = NULL; 6724 unsigned int idx; 6725 6726 ASSERT(list_num >= 0 && list_num <= 3); 6727 6728 switch (list_num) { 6729 case 0: 6730 ml = &arc_mfu->arcs_list[ARC_BUFC_METADATA]; 6731 break; 6732 case 1: 6733 ml = &arc_mru->arcs_list[ARC_BUFC_METADATA]; 6734 break; 6735 case 2: 6736 ml = &arc_mfu->arcs_list[ARC_BUFC_DATA]; 6737 break; 6738 case 3: 6739 ml = &arc_mru->arcs_list[ARC_BUFC_DATA]; 6740 break; 6741 } 6742 6743 /* 6744 * Return a randomly-selected sublist. This is acceptable 6745 * because the caller feeds only a little bit of data for each 6746 * call (8MB). Subsequent calls will result in different 6747 * sublists being selected. 6748 */ 6749 idx = multilist_get_random_index(ml); 6750 return (multilist_sublist_lock(ml, idx)); 6751} 6752 6753/* 6754 * Evict buffers from the device write hand to the distance specified in 6755 * bytes. This distance may span populated buffers, it may span nothing. 6756 * This is clearing a region on the L2ARC device ready for writing. 6757 * If the 'all' boolean is set, every buffer is evicted. 6758 */ 6759static void 6760l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all) 6761{ 6762 list_t *buflist; 6763 arc_buf_hdr_t *hdr, *hdr_prev; 6764 kmutex_t *hash_lock; 6765 uint64_t taddr; 6766 6767 buflist = &dev->l2ad_buflist; 6768 6769 if (!all && dev->l2ad_first) { 6770 /* 6771 * This is the first sweep through the device. There is 6772 * nothing to evict. 6773 */ 6774 return; 6775 } 6776 6777 if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) { 6778 /* 6779 * When nearing the end of the device, evict to the end 6780 * before the device write hand jumps to the start. 6781 */ 6782 taddr = dev->l2ad_end; 6783 } else { 6784 taddr = dev->l2ad_hand + distance; 6785 } 6786 DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist, 6787 uint64_t, taddr, boolean_t, all); 6788 6789top: 6790 mutex_enter(&dev->l2ad_mtx); 6791 for (hdr = list_tail(buflist); hdr; hdr = hdr_prev) { 6792 hdr_prev = list_prev(buflist, hdr); 6793 6794 hash_lock = HDR_LOCK(hdr); 6795 6796 /* 6797 * We cannot use mutex_enter or else we can deadlock 6798 * with l2arc_write_buffers (due to swapping the order 6799 * the hash lock and l2ad_mtx are taken). 6800 */ 6801 if (!mutex_tryenter(hash_lock)) { 6802 /* 6803 * Missed the hash lock. Retry. 6804 */ 6805 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry); 6806 mutex_exit(&dev->l2ad_mtx); 6807 mutex_enter(hash_lock); 6808 mutex_exit(hash_lock); 6809 goto top; 6810 } 6811 6812 if (HDR_L2_WRITE_HEAD(hdr)) { 6813 /* 6814 * We hit a write head node. Leave it for 6815 * l2arc_write_done(). 6816 */ 6817 list_remove(buflist, hdr); 6818 mutex_exit(hash_lock); 6819 continue; 6820 } 6821 6822 if (!all && HDR_HAS_L2HDR(hdr) && 6823 (hdr->b_l2hdr.b_daddr >= taddr || 6824 hdr->b_l2hdr.b_daddr < dev->l2ad_hand)) { 6825 /* 6826 * We've evicted to the target address, 6827 * or the end of the device. 6828 */ 6829 mutex_exit(hash_lock); 6830 break; 6831 } 6832 6833 ASSERT(HDR_HAS_L2HDR(hdr)); 6834 if (!HDR_HAS_L1HDR(hdr)) { 6835 ASSERT(!HDR_L2_READING(hdr)); 6836 /* 6837 * This doesn't exist in the ARC. Destroy. 6838 * arc_hdr_destroy() will call list_remove() 6839 * and decrement arcstat_l2_size. 6840 */ 6841 arc_change_state(arc_anon, hdr, hash_lock); 6842 arc_hdr_destroy(hdr); 6843 } else { 6844 ASSERT(hdr->b_l1hdr.b_state != arc_l2c_only); 6845 ARCSTAT_BUMP(arcstat_l2_evict_l1cached); 6846 /* 6847 * Invalidate issued or about to be issued 6848 * reads, since we may be about to write 6849 * over this location. 6850 */ 6851 if (HDR_L2_READING(hdr)) { 6852 ARCSTAT_BUMP(arcstat_l2_evict_reading); 6853 arc_hdr_set_flags(hdr, ARC_FLAG_L2_EVICTED); 6854 } 6855 6856 /* Ensure this header has finished being written */ 6857 ASSERT(!HDR_L2_WRITING(hdr)); 6858 6859 arc_hdr_l2hdr_destroy(hdr); 6860 } 6861 mutex_exit(hash_lock); 6862 } 6863 mutex_exit(&dev->l2ad_mtx); 6864} 6865 6866/* 6867 * Find and write ARC buffers to the L2ARC device. 6868 * 6869 * An ARC_FLAG_L2_WRITING flag is set so that the L2ARC buffers are not valid 6870 * for reading until they have completed writing. 6871 * The headroom_boost is an in-out parameter used to maintain headroom boost 6872 * state between calls to this function. 6873 * 6874 * Returns the number of bytes actually written (which may be smaller than 6875 * the delta by which the device hand has changed due to alignment). 6876 */ 6877static uint64_t 6878l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz) 6879{ 6880 arc_buf_hdr_t *hdr, *hdr_prev, *head; 6881 uint64_t write_asize, write_psize, write_sz, headroom; 6882 boolean_t full; 6883 l2arc_write_callback_t *cb; 6884 zio_t *pio, *wzio; 6885 uint64_t guid = spa_load_guid(spa); 6886 int try; 6887 6888 ASSERT3P(dev->l2ad_vdev, !=, NULL); 6889 6890 pio = NULL; 6891 write_sz = write_asize = write_psize = 0; 6892 full = B_FALSE; 6893 head = kmem_cache_alloc(hdr_l2only_cache, KM_PUSHPAGE); 6894 arc_hdr_set_flags(head, ARC_FLAG_L2_WRITE_HEAD | ARC_FLAG_HAS_L2HDR); 6895 6896 ARCSTAT_BUMP(arcstat_l2_write_buffer_iter); 6897 /* 6898 * Copy buffers for L2ARC writing. 6899 */ 6900 for (try = 0; try <= 3; try++) { 6901 multilist_sublist_t *mls = l2arc_sublist_lock(try); 6902 uint64_t passed_sz = 0; 6903 6904 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_iter); 6905 6906 /* 6907 * L2ARC fast warmup. 6908 * 6909 * Until the ARC is warm and starts to evict, read from the 6910 * head of the ARC lists rather than the tail. 6911 */ 6912 if (arc_warm == B_FALSE) 6913 hdr = multilist_sublist_head(mls); 6914 else 6915 hdr = multilist_sublist_tail(mls); 6916 if (hdr == NULL) 6917 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_null_iter); 6918 6919 headroom = target_sz * l2arc_headroom; 6920 if (zfs_compressed_arc_enabled) 6921 headroom = (headroom * l2arc_headroom_boost) / 100; 6922 6923 for (; hdr; hdr = hdr_prev) { 6924 kmutex_t *hash_lock; 6925 6926 if (arc_warm == B_FALSE) 6927 hdr_prev = multilist_sublist_next(mls, hdr); 6928 else 6929 hdr_prev = multilist_sublist_prev(mls, hdr); 6930 ARCSTAT_INCR(arcstat_l2_write_buffer_bytes_scanned, 6931 HDR_GET_LSIZE(hdr)); 6932 6933 hash_lock = HDR_LOCK(hdr); 6934 if (!mutex_tryenter(hash_lock)) { 6935 ARCSTAT_BUMP(arcstat_l2_write_trylock_fail); 6936 /* 6937 * Skip this buffer rather than waiting. 6938 */ 6939 continue; 6940 } 6941 6942 passed_sz += HDR_GET_LSIZE(hdr); 6943 if (passed_sz > headroom) { 6944 /* 6945 * Searched too far. 6946 */ 6947 mutex_exit(hash_lock); 6948 ARCSTAT_BUMP(arcstat_l2_write_passed_headroom); 6949 break; 6950 } 6951 6952 if (!l2arc_write_eligible(guid, hdr)) { 6953 mutex_exit(hash_lock); 6954 continue; 6955 } 6956 6957 /* 6958 * We rely on the L1 portion of the header below, so 6959 * it's invalid for this header to have been evicted out 6960 * of the ghost cache, prior to being written out. The 6961 * ARC_FLAG_L2_WRITING bit ensures this won't happen. 6962 */ 6963 ASSERT(HDR_HAS_L1HDR(hdr)); 6964 6965 ASSERT3U(HDR_GET_PSIZE(hdr), >, 0); 6966 ASSERT3P(hdr->b_l1hdr.b_pdata, !=, NULL); 6967 ASSERT3U(arc_hdr_size(hdr), >, 0); 6968 uint64_t size = arc_hdr_size(hdr); 6969 uint64_t asize = vdev_psize_to_asize(dev->l2ad_vdev, 6970 size); 6971 6972 if ((write_psize + asize) > target_sz) { 6973 full = B_TRUE; 6974 mutex_exit(hash_lock); 6975 ARCSTAT_BUMP(arcstat_l2_write_full); 6976 break; 6977 } 6978 6979 if (pio == NULL) { 6980 /* 6981 * Insert a dummy header on the buflist so 6982 * l2arc_write_done() can find where the 6983 * write buffers begin without searching. 6984 */ 6985 mutex_enter(&dev->l2ad_mtx); 6986 list_insert_head(&dev->l2ad_buflist, head); 6987 mutex_exit(&dev->l2ad_mtx); 6988 6989 cb = kmem_alloc( 6990 sizeof (l2arc_write_callback_t), KM_SLEEP); 6991 cb->l2wcb_dev = dev; 6992 cb->l2wcb_head = head; 6993 pio = zio_root(spa, l2arc_write_done, cb, 6994 ZIO_FLAG_CANFAIL); 6995 ARCSTAT_BUMP(arcstat_l2_write_pios); 6996 } 6997 6998 hdr->b_l2hdr.b_dev = dev; 6999 hdr->b_l2hdr.b_daddr = dev->l2ad_hand; 7000 arc_hdr_set_flags(hdr, 7001 ARC_FLAG_L2_WRITING | ARC_FLAG_HAS_L2HDR); 7002 7003 mutex_enter(&dev->l2ad_mtx); 7004 list_insert_head(&dev->l2ad_buflist, hdr); 7005 mutex_exit(&dev->l2ad_mtx); 7006 7007 (void) refcount_add_many(&dev->l2ad_alloc, size, hdr); 7008 7009 /* 7010 * Normally the L2ARC can use the hdr's data, but if 7011 * we're sharing data between the hdr and one of its 7012 * bufs, L2ARC needs its own copy of the data so that 7013 * the ZIO below can't race with the buf consumer. To 7014 * ensure that this copy will be available for the 7015 * lifetime of the ZIO and be cleaned up afterwards, we 7016 * add it to the l2arc_free_on_write queue. 7017 */ 7018 void *to_write; 7019 if (!HDR_SHARED_DATA(hdr) && size == asize) { 7020 to_write = hdr->b_l1hdr.b_pdata; 7021 } else { 7022 arc_buf_contents_t type = arc_buf_type(hdr); 7023 if (type == ARC_BUFC_METADATA) { 7024 to_write = zio_buf_alloc(asize); 7025 } else { 7026 ASSERT3U(type, ==, ARC_BUFC_DATA); 7027 to_write = zio_data_buf_alloc(asize); 7028 } 7029 7030 bcopy(hdr->b_l1hdr.b_pdata, to_write, size); 7031 if (asize != size) 7032 bzero(to_write + size, asize - size); 7033 l2arc_free_data_on_write(to_write, asize, type); 7034 } 7035 wzio = zio_write_phys(pio, dev->l2ad_vdev, 7036 hdr->b_l2hdr.b_daddr, asize, to_write, 7037 ZIO_CHECKSUM_OFF, NULL, hdr, 7038 ZIO_PRIORITY_ASYNC_WRITE, 7039 ZIO_FLAG_CANFAIL, B_FALSE); 7040 7041 write_sz += HDR_GET_LSIZE(hdr); 7042 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev, 7043 zio_t *, wzio); 7044 7045 write_asize += size; 7046 write_psize += asize; 7047 dev->l2ad_hand += asize; 7048 7049 mutex_exit(hash_lock); 7050 7051 (void) zio_nowait(wzio); 7052 } 7053 7054 multilist_sublist_unlock(mls); 7055 7056 if (full == B_TRUE) 7057 break; 7058 } 7059 7060 /* No buffers selected for writing? */ 7061 if (pio == NULL) { 7062 ASSERT0(write_sz); 7063 ASSERT(!HDR_HAS_L1HDR(head)); 7064 kmem_cache_free(hdr_l2only_cache, head); 7065 return (0); 7066 } 7067 7068 ASSERT3U(write_psize, <=, target_sz); 7069 ARCSTAT_BUMP(arcstat_l2_writes_sent); 7070 ARCSTAT_INCR(arcstat_l2_write_bytes, write_asize); 7071 ARCSTAT_INCR(arcstat_l2_size, write_sz); 7072 ARCSTAT_INCR(arcstat_l2_asize, write_asize); 7073 vdev_space_update(dev->l2ad_vdev, write_asize, 0, 0); 7074 7075 /* 7076 * Bump device hand to the device start if it is approaching the end. 7077 * l2arc_evict() will already have evicted ahead for this case. 7078 */ 7079 if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) { 7080 dev->l2ad_hand = dev->l2ad_start; 7081 dev->l2ad_first = B_FALSE; 7082 } 7083 7084 dev->l2ad_writing = B_TRUE; 7085 (void) zio_wait(pio); 7086 dev->l2ad_writing = B_FALSE; 7087 7088 return (write_asize); 7089} 7090 7091/* 7092 * This thread feeds the L2ARC at regular intervals. This is the beating 7093 * heart of the L2ARC. 7094 */ 7095static void 7096l2arc_feed_thread(void *dummy __unused) 7097{ 7098 callb_cpr_t cpr; 7099 l2arc_dev_t *dev; 7100 spa_t *spa; 7101 uint64_t size, wrote; 7102 clock_t begin, next = ddi_get_lbolt(); 7103 7104 CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG); 7105 7106 mutex_enter(&l2arc_feed_thr_lock); 7107 7108 while (l2arc_thread_exit == 0) { 7109 CALLB_CPR_SAFE_BEGIN(&cpr); 7110 (void) cv_timedwait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock, 7111 next - ddi_get_lbolt()); 7112 CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock); 7113 next = ddi_get_lbolt() + hz; 7114 7115 /* 7116 * Quick check for L2ARC devices. 7117 */ 7118 mutex_enter(&l2arc_dev_mtx); 7119 if (l2arc_ndev == 0) { 7120 mutex_exit(&l2arc_dev_mtx); 7121 continue; 7122 } 7123 mutex_exit(&l2arc_dev_mtx); 7124 begin = ddi_get_lbolt(); 7125 7126 /* 7127 * This selects the next l2arc device to write to, and in 7128 * doing so the next spa to feed from: dev->l2ad_spa. This 7129 * will return NULL if there are now no l2arc devices or if 7130 * they are all faulted. 7131 * 7132 * If a device is returned, its spa's config lock is also 7133 * held to prevent device removal. l2arc_dev_get_next() 7134 * will grab and release l2arc_dev_mtx. 7135 */ 7136 if ((dev = l2arc_dev_get_next()) == NULL) 7137 continue; 7138 7139 spa = dev->l2ad_spa; 7140 ASSERT3P(spa, !=, NULL); 7141 7142 /* 7143 * If the pool is read-only then force the feed thread to 7144 * sleep a little longer. 7145 */ 7146 if (!spa_writeable(spa)) { 7147 next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz; 7148 spa_config_exit(spa, SCL_L2ARC, dev); 7149 continue; 7150 } 7151 7152 /* 7153 * Avoid contributing to memory pressure. 7154 */ 7155 if (arc_reclaim_needed()) { 7156 ARCSTAT_BUMP(arcstat_l2_abort_lowmem); 7157 spa_config_exit(spa, SCL_L2ARC, dev); 7158 continue; 7159 } 7160 7161 ARCSTAT_BUMP(arcstat_l2_feeds); 7162 7163 size = l2arc_write_size(); 7164 7165 /* 7166 * Evict L2ARC buffers that will be overwritten. 7167 */ 7168 l2arc_evict(dev, size, B_FALSE); 7169 7170 /* 7171 * Write ARC buffers. 7172 */ 7173 wrote = l2arc_write_buffers(spa, dev, size); 7174 7175 /* 7176 * Calculate interval between writes. 7177 */ 7178 next = l2arc_write_interval(begin, size, wrote); 7179 spa_config_exit(spa, SCL_L2ARC, dev); 7180 } 7181 7182 l2arc_thread_exit = 0; 7183 cv_broadcast(&l2arc_feed_thr_cv); 7184 CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */ 7185 thread_exit(); 7186} 7187 7188boolean_t 7189l2arc_vdev_present(vdev_t *vd) 7190{ 7191 l2arc_dev_t *dev; 7192 7193 mutex_enter(&l2arc_dev_mtx); 7194 for (dev = list_head(l2arc_dev_list); dev != NULL; 7195 dev = list_next(l2arc_dev_list, dev)) { 7196 if (dev->l2ad_vdev == vd) 7197 break; 7198 } 7199 mutex_exit(&l2arc_dev_mtx); 7200 7201 return (dev != NULL); 7202} 7203 7204/* 7205 * Add a vdev for use by the L2ARC. By this point the spa has already 7206 * validated the vdev and opened it. 7207 */ 7208void 7209l2arc_add_vdev(spa_t *spa, vdev_t *vd) 7210{ 7211 l2arc_dev_t *adddev; 7212 7213 ASSERT(!l2arc_vdev_present(vd)); 7214 7215 vdev_ashift_optimize(vd); 7216 7217 /* 7218 * Create a new l2arc device entry. 7219 */ 7220 adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP); 7221 adddev->l2ad_spa = spa; 7222 adddev->l2ad_vdev = vd; 7223 adddev->l2ad_start = VDEV_LABEL_START_SIZE; 7224 adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd); 7225 adddev->l2ad_hand = adddev->l2ad_start; 7226 adddev->l2ad_first = B_TRUE; 7227 adddev->l2ad_writing = B_FALSE; 7228 7229 mutex_init(&adddev->l2ad_mtx, NULL, MUTEX_DEFAULT, NULL); 7230 /* 7231 * This is a list of all ARC buffers that are still valid on the 7232 * device. 7233 */ 7234 list_create(&adddev->l2ad_buflist, sizeof (arc_buf_hdr_t), 7235 offsetof(arc_buf_hdr_t, b_l2hdr.b_l2node)); 7236 7237 vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand); 7238 refcount_create(&adddev->l2ad_alloc); 7239 7240 /* 7241 * Add device to global list 7242 */ 7243 mutex_enter(&l2arc_dev_mtx); 7244 list_insert_head(l2arc_dev_list, adddev); 7245 atomic_inc_64(&l2arc_ndev); 7246 mutex_exit(&l2arc_dev_mtx); 7247} 7248 7249/* 7250 * Remove a vdev from the L2ARC. 7251 */ 7252void 7253l2arc_remove_vdev(vdev_t *vd) 7254{ 7255 l2arc_dev_t *dev, *nextdev, *remdev = NULL; 7256 7257 /* 7258 * Find the device by vdev 7259 */ 7260 mutex_enter(&l2arc_dev_mtx); 7261 for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) { 7262 nextdev = list_next(l2arc_dev_list, dev); 7263 if (vd == dev->l2ad_vdev) { 7264 remdev = dev; 7265 break; 7266 } 7267 } 7268 ASSERT3P(remdev, !=, NULL); 7269 7270 /* 7271 * Remove device from global list 7272 */ 7273 list_remove(l2arc_dev_list, remdev); 7274 l2arc_dev_last = NULL; /* may have been invalidated */ 7275 atomic_dec_64(&l2arc_ndev); 7276 mutex_exit(&l2arc_dev_mtx); 7277 7278 /* 7279 * Clear all buflists and ARC references. L2ARC device flush. 7280 */ 7281 l2arc_evict(remdev, 0, B_TRUE); 7282 list_destroy(&remdev->l2ad_buflist); 7283 mutex_destroy(&remdev->l2ad_mtx); 7284 refcount_destroy(&remdev->l2ad_alloc); 7285 kmem_free(remdev, sizeof (l2arc_dev_t)); 7286} 7287 7288void 7289l2arc_init(void) 7290{ 7291 l2arc_thread_exit = 0; 7292 l2arc_ndev = 0; 7293 l2arc_writes_sent = 0; 7294 l2arc_writes_done = 0; 7295 7296 mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL); 7297 cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL); 7298 mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL); 7299 mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL); 7300 7301 l2arc_dev_list = &L2ARC_dev_list; 7302 l2arc_free_on_write = &L2ARC_free_on_write; 7303 list_create(l2arc_dev_list, sizeof (l2arc_dev_t), 7304 offsetof(l2arc_dev_t, l2ad_node)); 7305 list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t), 7306 offsetof(l2arc_data_free_t, l2df_list_node)); 7307} 7308 7309void 7310l2arc_fini(void) 7311{ 7312 /* 7313 * This is called from dmu_fini(), which is called from spa_fini(); 7314 * Because of this, we can assume that all l2arc devices have 7315 * already been removed when the pools themselves were removed. 7316 */ 7317 7318 l2arc_do_free_on_write(); 7319 7320 mutex_destroy(&l2arc_feed_thr_lock); 7321 cv_destroy(&l2arc_feed_thr_cv); 7322 mutex_destroy(&l2arc_dev_mtx); 7323 mutex_destroy(&l2arc_free_on_write_mtx); 7324 7325 list_destroy(l2arc_dev_list); 7326 list_destroy(l2arc_free_on_write); 7327} 7328 7329void 7330l2arc_start(void) 7331{ 7332 if (!(spa_mode_global & FWRITE)) 7333 return; 7334 7335 (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0, 7336 TS_RUN, minclsyspri); 7337} 7338 7339void 7340l2arc_stop(void) 7341{ 7342 if (!(spa_mode_global & FWRITE)) 7343 return; 7344 7345 mutex_enter(&l2arc_feed_thr_lock); 7346 cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */ 7347 l2arc_thread_exit = 1; 7348 while (l2arc_thread_exit != 0) 7349 cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock); 7350 mutex_exit(&l2arc_feed_thr_lock); 7351} 7352