arc.c revision 317470
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#include <sys/racct.h> 261#endif 262#include <sys/callb.h> 263#include <sys/kstat.h> 264#include <sys/trim_map.h> 265#include <zfs_fletcher.h> 266#include <sys/sdt.h> 267 268#include <machine/vmparam.h> 269 270#ifdef illumos 271#ifndef _KERNEL 272/* set with ZFS_DEBUG=watch, to enable watchpoints on frozen buffers */ 273boolean_t arc_watch = B_FALSE; 274int arc_procfd; 275#endif 276#endif /* illumos */ 277 278static kmutex_t arc_reclaim_lock; 279static kcondvar_t arc_reclaim_thread_cv; 280static boolean_t arc_reclaim_thread_exit; 281static kcondvar_t arc_reclaim_waiters_cv; 282 283static kmutex_t arc_dnlc_evicts_lock; 284static kcondvar_t arc_dnlc_evicts_cv; 285static boolean_t arc_dnlc_evicts_thread_exit; 286 287uint_t arc_reduce_dnlc_percent = 3; 288 289/* 290 * The number of headers to evict in arc_evict_state_impl() before 291 * dropping the sublist lock and evicting from another sublist. A lower 292 * value means we're more likely to evict the "correct" header (i.e. the 293 * oldest header in the arc state), but comes with higher overhead 294 * (i.e. more invocations of arc_evict_state_impl()). 295 */ 296int zfs_arc_evict_batch_limit = 10; 297 298/* 299 * The number of sublists used for each of the arc state lists. If this 300 * is not set to a suitable value by the user, it will be configured to 301 * the number of CPUs on the system in arc_init(). 302 */ 303int zfs_arc_num_sublists_per_state = 0; 304 305/* number of seconds before growing cache again */ 306static int arc_grow_retry = 60; 307 308/* shift of arc_c for calculating overflow limit in arc_get_data_buf */ 309int zfs_arc_overflow_shift = 8; 310 311/* shift of arc_c for calculating both min and max arc_p */ 312static int arc_p_min_shift = 4; 313 314/* log2(fraction of arc to reclaim) */ 315static int arc_shrink_shift = 7; 316 317/* 318 * log2(fraction of ARC which must be free to allow growing). 319 * I.e. If there is less than arc_c >> arc_no_grow_shift free memory, 320 * when reading a new block into the ARC, we will evict an equal-sized block 321 * from the ARC. 322 * 323 * This must be less than arc_shrink_shift, so that when we shrink the ARC, 324 * we will still not allow it to grow. 325 */ 326int arc_no_grow_shift = 5; 327 328 329/* 330 * minimum lifespan of a prefetch block in clock ticks 331 * (initialized in arc_init()) 332 */ 333static int arc_min_prefetch_lifespan; 334 335/* 336 * If this percent of memory is free, don't throttle. 337 */ 338int arc_lotsfree_percent = 10; 339 340static int arc_dead; 341extern boolean_t zfs_prefetch_disable; 342 343/* 344 * The arc has filled available memory and has now warmed up. 345 */ 346static boolean_t arc_warm; 347 348/* 349 * These tunables are for performance analysis. 350 */ 351uint64_t zfs_arc_max; 352uint64_t zfs_arc_min; 353uint64_t zfs_arc_meta_limit = 0; 354uint64_t zfs_arc_meta_min = 0; 355int zfs_arc_grow_retry = 0; 356int zfs_arc_shrink_shift = 0; 357int zfs_arc_p_min_shift = 0; 358uint64_t zfs_arc_average_blocksize = 8 * 1024; /* 8KB */ 359u_int zfs_arc_free_target = 0; 360 361/* Absolute min for arc min / max is 16MB. */ 362static uint64_t arc_abs_min = 16 << 20; 363 364boolean_t zfs_compressed_arc_enabled = B_TRUE; 365 366static int sysctl_vfs_zfs_arc_free_target(SYSCTL_HANDLER_ARGS); 367static int sysctl_vfs_zfs_arc_meta_limit(SYSCTL_HANDLER_ARGS); 368static int sysctl_vfs_zfs_arc_max(SYSCTL_HANDLER_ARGS); 369static int sysctl_vfs_zfs_arc_min(SYSCTL_HANDLER_ARGS); 370 371#if defined(__FreeBSD__) && defined(_KERNEL) 372static void 373arc_free_target_init(void *unused __unused) 374{ 375 376 zfs_arc_free_target = vm_pageout_wakeup_thresh; 377} 378SYSINIT(arc_free_target_init, SI_SUB_KTHREAD_PAGE, SI_ORDER_ANY, 379 arc_free_target_init, NULL); 380 381TUNABLE_QUAD("vfs.zfs.arc_meta_limit", &zfs_arc_meta_limit); 382TUNABLE_QUAD("vfs.zfs.arc_meta_min", &zfs_arc_meta_min); 383TUNABLE_INT("vfs.zfs.arc_shrink_shift", &zfs_arc_shrink_shift); 384SYSCTL_DECL(_vfs_zfs); 385SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_max, CTLTYPE_U64 | CTLFLAG_RWTUN, 386 0, sizeof(uint64_t), sysctl_vfs_zfs_arc_max, "QU", "Maximum ARC size"); 387SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_min, CTLTYPE_U64 | CTLFLAG_RWTUN, 388 0, sizeof(uint64_t), sysctl_vfs_zfs_arc_min, "QU", "Minimum ARC size"); 389SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_average_blocksize, CTLFLAG_RDTUN, 390 &zfs_arc_average_blocksize, 0, 391 "ARC average blocksize"); 392SYSCTL_INT(_vfs_zfs, OID_AUTO, arc_shrink_shift, CTLFLAG_RW, 393 &arc_shrink_shift, 0, 394 "log2(fraction of arc to reclaim)"); 395SYSCTL_INT(_vfs_zfs, OID_AUTO, compressed_arc_enabled, CTLFLAG_RDTUN, 396 &zfs_compressed_arc_enabled, 0, "Enable compressed ARC"); 397 398/* 399 * We don't have a tunable for arc_free_target due to the dependency on 400 * pagedaemon initialisation. 401 */ 402SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_free_target, 403 CTLTYPE_UINT | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, sizeof(u_int), 404 sysctl_vfs_zfs_arc_free_target, "IU", 405 "Desired number of free pages below which ARC triggers reclaim"); 406 407static int 408sysctl_vfs_zfs_arc_free_target(SYSCTL_HANDLER_ARGS) 409{ 410 u_int val; 411 int err; 412 413 val = zfs_arc_free_target; 414 err = sysctl_handle_int(oidp, &val, 0, req); 415 if (err != 0 || req->newptr == NULL) 416 return (err); 417 418 if (val < minfree) 419 return (EINVAL); 420 if (val > vm_cnt.v_page_count) 421 return (EINVAL); 422 423 zfs_arc_free_target = val; 424 425 return (0); 426} 427 428/* 429 * Must be declared here, before the definition of corresponding kstat 430 * macro which uses the same names will confuse the compiler. 431 */ 432SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_meta_limit, 433 CTLTYPE_U64 | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, sizeof(uint64_t), 434 sysctl_vfs_zfs_arc_meta_limit, "QU", 435 "ARC metadata limit"); 436#endif 437 438/* 439 * Note that buffers can be in one of 6 states: 440 * ARC_anon - anonymous (discussed below) 441 * ARC_mru - recently used, currently cached 442 * ARC_mru_ghost - recentely used, no longer in cache 443 * ARC_mfu - frequently used, currently cached 444 * ARC_mfu_ghost - frequently used, no longer in cache 445 * ARC_l2c_only - exists in L2ARC but not other states 446 * When there are no active references to the buffer, they are 447 * are linked onto a list in one of these arc states. These are 448 * the only buffers that can be evicted or deleted. Within each 449 * state there are multiple lists, one for meta-data and one for 450 * non-meta-data. Meta-data (indirect blocks, blocks of dnodes, 451 * etc.) is tracked separately so that it can be managed more 452 * explicitly: favored over data, limited explicitly. 453 * 454 * Anonymous buffers are buffers that are not associated with 455 * a DVA. These are buffers that hold dirty block copies 456 * before they are written to stable storage. By definition, 457 * they are "ref'd" and are considered part of arc_mru 458 * that cannot be freed. Generally, they will aquire a DVA 459 * as they are written and migrate onto the arc_mru list. 460 * 461 * The ARC_l2c_only state is for buffers that are in the second 462 * level ARC but no longer in any of the ARC_m* lists. The second 463 * level ARC itself may also contain buffers that are in any of 464 * the ARC_m* states - meaning that a buffer can exist in two 465 * places. The reason for the ARC_l2c_only state is to keep the 466 * buffer header in the hash table, so that reads that hit the 467 * second level ARC benefit from these fast lookups. 468 */ 469 470typedef struct arc_state { 471 /* 472 * list of evictable buffers 473 */ 474 multilist_t arcs_list[ARC_BUFC_NUMTYPES]; 475 /* 476 * total amount of evictable data in this state 477 */ 478 refcount_t arcs_esize[ARC_BUFC_NUMTYPES]; 479 /* 480 * total amount of data in this state; this includes: evictable, 481 * non-evictable, ARC_BUFC_DATA, and ARC_BUFC_METADATA. 482 */ 483 refcount_t arcs_size; 484} arc_state_t; 485 486/* The 6 states: */ 487static arc_state_t ARC_anon; 488static arc_state_t ARC_mru; 489static arc_state_t ARC_mru_ghost; 490static arc_state_t ARC_mfu; 491static arc_state_t ARC_mfu_ghost; 492static arc_state_t ARC_l2c_only; 493 494typedef struct arc_stats { 495 kstat_named_t arcstat_hits; 496 kstat_named_t arcstat_misses; 497 kstat_named_t arcstat_demand_data_hits; 498 kstat_named_t arcstat_demand_data_misses; 499 kstat_named_t arcstat_demand_metadata_hits; 500 kstat_named_t arcstat_demand_metadata_misses; 501 kstat_named_t arcstat_prefetch_data_hits; 502 kstat_named_t arcstat_prefetch_data_misses; 503 kstat_named_t arcstat_prefetch_metadata_hits; 504 kstat_named_t arcstat_prefetch_metadata_misses; 505 kstat_named_t arcstat_mru_hits; 506 kstat_named_t arcstat_mru_ghost_hits; 507 kstat_named_t arcstat_mfu_hits; 508 kstat_named_t arcstat_mfu_ghost_hits; 509 kstat_named_t arcstat_allocated; 510 kstat_named_t arcstat_deleted; 511 /* 512 * Number of buffers that could not be evicted because the hash lock 513 * was held by another thread. The lock may not necessarily be held 514 * by something using the same buffer, since hash locks are shared 515 * by multiple buffers. 516 */ 517 kstat_named_t arcstat_mutex_miss; 518 /* 519 * Number of buffers skipped because they have I/O in progress, are 520 * indrect prefetch buffers that have not lived long enough, or are 521 * not from the spa we're trying to evict from. 522 */ 523 kstat_named_t arcstat_evict_skip; 524 /* 525 * Number of times arc_evict_state() was unable to evict enough 526 * buffers to reach it's target amount. 527 */ 528 kstat_named_t arcstat_evict_not_enough; 529 kstat_named_t arcstat_evict_l2_cached; 530 kstat_named_t arcstat_evict_l2_eligible; 531 kstat_named_t arcstat_evict_l2_ineligible; 532 kstat_named_t arcstat_evict_l2_skip; 533 kstat_named_t arcstat_hash_elements; 534 kstat_named_t arcstat_hash_elements_max; 535 kstat_named_t arcstat_hash_collisions; 536 kstat_named_t arcstat_hash_chains; 537 kstat_named_t arcstat_hash_chain_max; 538 kstat_named_t arcstat_p; 539 kstat_named_t arcstat_c; 540 kstat_named_t arcstat_c_min; 541 kstat_named_t arcstat_c_max; 542 kstat_named_t arcstat_size; 543 /* 544 * Number of compressed bytes stored in the arc_buf_hdr_t's b_pdata. 545 * Note that the compressed bytes may match the uncompressed bytes 546 * if the block is either not compressed or compressed arc is disabled. 547 */ 548 kstat_named_t arcstat_compressed_size; 549 /* 550 * Uncompressed size of the data stored in b_pdata. If compressed 551 * arc is disabled then this value will be identical to the stat 552 * above. 553 */ 554 kstat_named_t arcstat_uncompressed_size; 555 /* 556 * Number of bytes stored in all the arc_buf_t's. This is classified 557 * as "overhead" since this data is typically short-lived and will 558 * be evicted from the arc when it becomes unreferenced unless the 559 * zfs_keep_uncompressed_metadata or zfs_keep_uncompressed_level 560 * values have been set (see comment in dbuf.c for more information). 561 */ 562 kstat_named_t arcstat_overhead_size; 563 /* 564 * Number of bytes consumed by internal ARC structures necessary 565 * for tracking purposes; these structures are not actually 566 * backed by ARC buffers. This includes arc_buf_hdr_t structures 567 * (allocated via arc_buf_hdr_t_full and arc_buf_hdr_t_l2only 568 * caches), and arc_buf_t structures (allocated via arc_buf_t 569 * cache). 570 */ 571 kstat_named_t arcstat_hdr_size; 572 /* 573 * Number of bytes consumed by ARC buffers of type equal to 574 * ARC_BUFC_DATA. This is generally consumed by buffers backing 575 * on disk user data (e.g. plain file contents). 576 */ 577 kstat_named_t arcstat_data_size; 578 /* 579 * Number of bytes consumed by ARC buffers of type equal to 580 * ARC_BUFC_METADATA. This is generally consumed by buffers 581 * backing on disk data that is used for internal ZFS 582 * structures (e.g. ZAP, dnode, indirect blocks, etc). 583 */ 584 kstat_named_t arcstat_metadata_size; 585 /* 586 * Number of bytes consumed by various buffers and structures 587 * not actually backed with ARC buffers. This includes bonus 588 * buffers (allocated directly via zio_buf_* functions), 589 * dmu_buf_impl_t structures (allocated via dmu_buf_impl_t 590 * cache), and dnode_t structures (allocated via dnode_t cache). 591 */ 592 kstat_named_t arcstat_other_size; 593 /* 594 * Total number of bytes consumed by ARC buffers residing in the 595 * arc_anon state. This includes *all* buffers in the arc_anon 596 * state; e.g. data, metadata, evictable, and unevictable buffers 597 * are all included in this value. 598 */ 599 kstat_named_t arcstat_anon_size; 600 /* 601 * Number of bytes consumed by ARC buffers that meet the 602 * following criteria: backing buffers of type ARC_BUFC_DATA, 603 * residing in the arc_anon state, and are eligible for eviction 604 * (e.g. have no outstanding holds on the buffer). 605 */ 606 kstat_named_t arcstat_anon_evictable_data; 607 /* 608 * Number of bytes consumed by ARC buffers that meet the 609 * following criteria: backing buffers of type ARC_BUFC_METADATA, 610 * residing in the arc_anon state, and are eligible for eviction 611 * (e.g. have no outstanding holds on the buffer). 612 */ 613 kstat_named_t arcstat_anon_evictable_metadata; 614 /* 615 * Total number of bytes consumed by ARC buffers residing in the 616 * arc_mru state. This includes *all* buffers in the arc_mru 617 * state; e.g. data, metadata, evictable, and unevictable buffers 618 * are all included in this value. 619 */ 620 kstat_named_t arcstat_mru_size; 621 /* 622 * Number of bytes consumed by ARC buffers that meet the 623 * following criteria: backing buffers of type ARC_BUFC_DATA, 624 * residing in the arc_mru state, and are eligible for eviction 625 * (e.g. have no outstanding holds on the buffer). 626 */ 627 kstat_named_t arcstat_mru_evictable_data; 628 /* 629 * Number of bytes consumed by ARC buffers that meet the 630 * following criteria: backing buffers of type ARC_BUFC_METADATA, 631 * residing in the arc_mru state, and are eligible for eviction 632 * (e.g. have no outstanding holds on the buffer). 633 */ 634 kstat_named_t arcstat_mru_evictable_metadata; 635 /* 636 * Total number of bytes that *would have been* consumed by ARC 637 * buffers in the arc_mru_ghost state. The key thing to note 638 * here, is the fact that this size doesn't actually indicate 639 * RAM consumption. The ghost lists only consist of headers and 640 * don't actually have ARC buffers linked off of these headers. 641 * Thus, *if* the headers had associated ARC buffers, these 642 * buffers *would have* consumed this number of bytes. 643 */ 644 kstat_named_t arcstat_mru_ghost_size; 645 /* 646 * Number of bytes that *would have been* consumed by ARC 647 * buffers that are eligible for eviction, of type 648 * ARC_BUFC_DATA, and linked off the arc_mru_ghost state. 649 */ 650 kstat_named_t arcstat_mru_ghost_evictable_data; 651 /* 652 * Number of bytes that *would have been* consumed by ARC 653 * buffers that are eligible for eviction, of type 654 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state. 655 */ 656 kstat_named_t arcstat_mru_ghost_evictable_metadata; 657 /* 658 * Total number of bytes consumed by ARC buffers residing in the 659 * arc_mfu state. This includes *all* buffers in the arc_mfu 660 * state; e.g. data, metadata, evictable, and unevictable buffers 661 * are all included in this value. 662 */ 663 kstat_named_t arcstat_mfu_size; 664 /* 665 * Number of bytes consumed by ARC buffers that are eligible for 666 * eviction, of type ARC_BUFC_DATA, and reside in the arc_mfu 667 * state. 668 */ 669 kstat_named_t arcstat_mfu_evictable_data; 670 /* 671 * Number of bytes consumed by ARC buffers that are eligible for 672 * eviction, of type ARC_BUFC_METADATA, and reside in the 673 * arc_mfu state. 674 */ 675 kstat_named_t arcstat_mfu_evictable_metadata; 676 /* 677 * Total number of bytes that *would have been* consumed by ARC 678 * buffers in the arc_mfu_ghost state. See the comment above 679 * arcstat_mru_ghost_size for more details. 680 */ 681 kstat_named_t arcstat_mfu_ghost_size; 682 /* 683 * Number of bytes that *would have been* consumed by ARC 684 * buffers that are eligible for eviction, of type 685 * ARC_BUFC_DATA, and linked off the arc_mfu_ghost state. 686 */ 687 kstat_named_t arcstat_mfu_ghost_evictable_data; 688 /* 689 * Number of bytes that *would have been* consumed by ARC 690 * buffers that are eligible for eviction, of type 691 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state. 692 */ 693 kstat_named_t arcstat_mfu_ghost_evictable_metadata; 694 kstat_named_t arcstat_l2_hits; 695 kstat_named_t arcstat_l2_misses; 696 kstat_named_t arcstat_l2_feeds; 697 kstat_named_t arcstat_l2_rw_clash; 698 kstat_named_t arcstat_l2_read_bytes; 699 kstat_named_t arcstat_l2_write_bytes; 700 kstat_named_t arcstat_l2_writes_sent; 701 kstat_named_t arcstat_l2_writes_done; 702 kstat_named_t arcstat_l2_writes_error; 703 kstat_named_t arcstat_l2_writes_lock_retry; 704 kstat_named_t arcstat_l2_evict_lock_retry; 705 kstat_named_t arcstat_l2_evict_reading; 706 kstat_named_t arcstat_l2_evict_l1cached; 707 kstat_named_t arcstat_l2_free_on_write; 708 kstat_named_t arcstat_l2_abort_lowmem; 709 kstat_named_t arcstat_l2_cksum_bad; 710 kstat_named_t arcstat_l2_io_error; 711 kstat_named_t arcstat_l2_size; 712 kstat_named_t arcstat_l2_asize; 713 kstat_named_t arcstat_l2_hdr_size; 714 kstat_named_t arcstat_l2_write_trylock_fail; 715 kstat_named_t arcstat_l2_write_passed_headroom; 716 kstat_named_t arcstat_l2_write_spa_mismatch; 717 kstat_named_t arcstat_l2_write_in_l2; 718 kstat_named_t arcstat_l2_write_hdr_io_in_progress; 719 kstat_named_t arcstat_l2_write_not_cacheable; 720 kstat_named_t arcstat_l2_write_full; 721 kstat_named_t arcstat_l2_write_buffer_iter; 722 kstat_named_t arcstat_l2_write_pios; 723 kstat_named_t arcstat_l2_write_buffer_bytes_scanned; 724 kstat_named_t arcstat_l2_write_buffer_list_iter; 725 kstat_named_t arcstat_l2_write_buffer_list_null_iter; 726 kstat_named_t arcstat_memory_throttle_count; 727 kstat_named_t arcstat_meta_used; 728 kstat_named_t arcstat_meta_limit; 729 kstat_named_t arcstat_meta_max; 730 kstat_named_t arcstat_meta_min; 731 kstat_named_t arcstat_sync_wait_for_async; 732 kstat_named_t arcstat_demand_hit_predictive_prefetch; 733} arc_stats_t; 734 735static arc_stats_t arc_stats = { 736 { "hits", KSTAT_DATA_UINT64 }, 737 { "misses", KSTAT_DATA_UINT64 }, 738 { "demand_data_hits", KSTAT_DATA_UINT64 }, 739 { "demand_data_misses", KSTAT_DATA_UINT64 }, 740 { "demand_metadata_hits", KSTAT_DATA_UINT64 }, 741 { "demand_metadata_misses", KSTAT_DATA_UINT64 }, 742 { "prefetch_data_hits", KSTAT_DATA_UINT64 }, 743 { "prefetch_data_misses", KSTAT_DATA_UINT64 }, 744 { "prefetch_metadata_hits", KSTAT_DATA_UINT64 }, 745 { "prefetch_metadata_misses", KSTAT_DATA_UINT64 }, 746 { "mru_hits", KSTAT_DATA_UINT64 }, 747 { "mru_ghost_hits", KSTAT_DATA_UINT64 }, 748 { "mfu_hits", KSTAT_DATA_UINT64 }, 749 { "mfu_ghost_hits", KSTAT_DATA_UINT64 }, 750 { "allocated", KSTAT_DATA_UINT64 }, 751 { "deleted", KSTAT_DATA_UINT64 }, 752 { "mutex_miss", KSTAT_DATA_UINT64 }, 753 { "evict_skip", KSTAT_DATA_UINT64 }, 754 { "evict_not_enough", KSTAT_DATA_UINT64 }, 755 { "evict_l2_cached", KSTAT_DATA_UINT64 }, 756 { "evict_l2_eligible", KSTAT_DATA_UINT64 }, 757 { "evict_l2_ineligible", KSTAT_DATA_UINT64 }, 758 { "evict_l2_skip", KSTAT_DATA_UINT64 }, 759 { "hash_elements", KSTAT_DATA_UINT64 }, 760 { "hash_elements_max", KSTAT_DATA_UINT64 }, 761 { "hash_collisions", KSTAT_DATA_UINT64 }, 762 { "hash_chains", KSTAT_DATA_UINT64 }, 763 { "hash_chain_max", KSTAT_DATA_UINT64 }, 764 { "p", KSTAT_DATA_UINT64 }, 765 { "c", KSTAT_DATA_UINT64 }, 766 { "c_min", KSTAT_DATA_UINT64 }, 767 { "c_max", KSTAT_DATA_UINT64 }, 768 { "size", KSTAT_DATA_UINT64 }, 769 { "compressed_size", KSTAT_DATA_UINT64 }, 770 { "uncompressed_size", KSTAT_DATA_UINT64 }, 771 { "overhead_size", KSTAT_DATA_UINT64 }, 772 { "hdr_size", KSTAT_DATA_UINT64 }, 773 { "data_size", KSTAT_DATA_UINT64 }, 774 { "metadata_size", KSTAT_DATA_UINT64 }, 775 { "other_size", KSTAT_DATA_UINT64 }, 776 { "anon_size", KSTAT_DATA_UINT64 }, 777 { "anon_evictable_data", KSTAT_DATA_UINT64 }, 778 { "anon_evictable_metadata", KSTAT_DATA_UINT64 }, 779 { "mru_size", KSTAT_DATA_UINT64 }, 780 { "mru_evictable_data", KSTAT_DATA_UINT64 }, 781 { "mru_evictable_metadata", KSTAT_DATA_UINT64 }, 782 { "mru_ghost_size", KSTAT_DATA_UINT64 }, 783 { "mru_ghost_evictable_data", KSTAT_DATA_UINT64 }, 784 { "mru_ghost_evictable_metadata", KSTAT_DATA_UINT64 }, 785 { "mfu_size", KSTAT_DATA_UINT64 }, 786 { "mfu_evictable_data", KSTAT_DATA_UINT64 }, 787 { "mfu_evictable_metadata", KSTAT_DATA_UINT64 }, 788 { "mfu_ghost_size", KSTAT_DATA_UINT64 }, 789 { "mfu_ghost_evictable_data", KSTAT_DATA_UINT64 }, 790 { "mfu_ghost_evictable_metadata", KSTAT_DATA_UINT64 }, 791 { "l2_hits", KSTAT_DATA_UINT64 }, 792 { "l2_misses", KSTAT_DATA_UINT64 }, 793 { "l2_feeds", KSTAT_DATA_UINT64 }, 794 { "l2_rw_clash", KSTAT_DATA_UINT64 }, 795 { "l2_read_bytes", KSTAT_DATA_UINT64 }, 796 { "l2_write_bytes", KSTAT_DATA_UINT64 }, 797 { "l2_writes_sent", KSTAT_DATA_UINT64 }, 798 { "l2_writes_done", KSTAT_DATA_UINT64 }, 799 { "l2_writes_error", KSTAT_DATA_UINT64 }, 800 { "l2_writes_lock_retry", KSTAT_DATA_UINT64 }, 801 { "l2_evict_lock_retry", KSTAT_DATA_UINT64 }, 802 { "l2_evict_reading", KSTAT_DATA_UINT64 }, 803 { "l2_evict_l1cached", KSTAT_DATA_UINT64 }, 804 { "l2_free_on_write", KSTAT_DATA_UINT64 }, 805 { "l2_abort_lowmem", KSTAT_DATA_UINT64 }, 806 { "l2_cksum_bad", KSTAT_DATA_UINT64 }, 807 { "l2_io_error", KSTAT_DATA_UINT64 }, 808 { "l2_size", KSTAT_DATA_UINT64 }, 809 { "l2_asize", KSTAT_DATA_UINT64 }, 810 { "l2_hdr_size", KSTAT_DATA_UINT64 }, 811 { "l2_write_trylock_fail", KSTAT_DATA_UINT64 }, 812 { "l2_write_passed_headroom", KSTAT_DATA_UINT64 }, 813 { "l2_write_spa_mismatch", KSTAT_DATA_UINT64 }, 814 { "l2_write_in_l2", KSTAT_DATA_UINT64 }, 815 { "l2_write_io_in_progress", KSTAT_DATA_UINT64 }, 816 { "l2_write_not_cacheable", KSTAT_DATA_UINT64 }, 817 { "l2_write_full", KSTAT_DATA_UINT64 }, 818 { "l2_write_buffer_iter", KSTAT_DATA_UINT64 }, 819 { "l2_write_pios", KSTAT_DATA_UINT64 }, 820 { "l2_write_buffer_bytes_scanned", KSTAT_DATA_UINT64 }, 821 { "l2_write_buffer_list_iter", KSTAT_DATA_UINT64 }, 822 { "l2_write_buffer_list_null_iter", KSTAT_DATA_UINT64 }, 823 { "memory_throttle_count", KSTAT_DATA_UINT64 }, 824 { "arc_meta_used", KSTAT_DATA_UINT64 }, 825 { "arc_meta_limit", KSTAT_DATA_UINT64 }, 826 { "arc_meta_max", KSTAT_DATA_UINT64 }, 827 { "arc_meta_min", KSTAT_DATA_UINT64 }, 828 { "sync_wait_for_async", KSTAT_DATA_UINT64 }, 829 { "demand_hit_predictive_prefetch", KSTAT_DATA_UINT64 }, 830}; 831 832#define ARCSTAT(stat) (arc_stats.stat.value.ui64) 833 834#define ARCSTAT_INCR(stat, val) \ 835 atomic_add_64(&arc_stats.stat.value.ui64, (val)) 836 837#define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1) 838#define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1) 839 840#define ARCSTAT_MAX(stat, val) { \ 841 uint64_t m; \ 842 while ((val) > (m = arc_stats.stat.value.ui64) && \ 843 (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \ 844 continue; \ 845} 846 847#define ARCSTAT_MAXSTAT(stat) \ 848 ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64) 849 850/* 851 * We define a macro to allow ARC hits/misses to be easily broken down by 852 * two separate conditions, giving a total of four different subtypes for 853 * each of hits and misses (so eight statistics total). 854 */ 855#define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \ 856 if (cond1) { \ 857 if (cond2) { \ 858 ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \ 859 } else { \ 860 ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \ 861 } \ 862 } else { \ 863 if (cond2) { \ 864 ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \ 865 } else { \ 866 ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\ 867 } \ 868 } 869 870kstat_t *arc_ksp; 871static arc_state_t *arc_anon; 872static arc_state_t *arc_mru; 873static arc_state_t *arc_mru_ghost; 874static arc_state_t *arc_mfu; 875static arc_state_t *arc_mfu_ghost; 876static arc_state_t *arc_l2c_only; 877 878/* 879 * There are several ARC variables that are critical to export as kstats -- 880 * but we don't want to have to grovel around in the kstat whenever we wish to 881 * manipulate them. For these variables, we therefore define them to be in 882 * terms of the statistic variable. This assures that we are not introducing 883 * the possibility of inconsistency by having shadow copies of the variables, 884 * while still allowing the code to be readable. 885 */ 886#define arc_size ARCSTAT(arcstat_size) /* actual total arc size */ 887#define arc_p ARCSTAT(arcstat_p) /* target size of MRU */ 888#define arc_c ARCSTAT(arcstat_c) /* target size of cache */ 889#define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */ 890#define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */ 891#define arc_meta_limit ARCSTAT(arcstat_meta_limit) /* max size for metadata */ 892#define arc_meta_min ARCSTAT(arcstat_meta_min) /* min size for metadata */ 893#define arc_meta_used ARCSTAT(arcstat_meta_used) /* size of metadata */ 894#define arc_meta_max ARCSTAT(arcstat_meta_max) /* max size of metadata */ 895 896/* compressed size of entire arc */ 897#define arc_compressed_size ARCSTAT(arcstat_compressed_size) 898/* uncompressed size of entire arc */ 899#define arc_uncompressed_size ARCSTAT(arcstat_uncompressed_size) 900/* number of bytes in the arc from arc_buf_t's */ 901#define arc_overhead_size ARCSTAT(arcstat_overhead_size) 902 903static int arc_no_grow; /* Don't try to grow cache size */ 904static uint64_t arc_tempreserve; 905static uint64_t arc_loaned_bytes; 906 907typedef struct arc_callback arc_callback_t; 908 909struct arc_callback { 910 void *acb_private; 911 arc_done_func_t *acb_done; 912 arc_buf_t *acb_buf; 913 zio_t *acb_zio_dummy; 914 arc_callback_t *acb_next; 915}; 916 917typedef struct arc_write_callback arc_write_callback_t; 918 919struct arc_write_callback { 920 void *awcb_private; 921 arc_done_func_t *awcb_ready; 922 arc_done_func_t *awcb_children_ready; 923 arc_done_func_t *awcb_physdone; 924 arc_done_func_t *awcb_done; 925 arc_buf_t *awcb_buf; 926}; 927 928/* 929 * ARC buffers are separated into multiple structs as a memory saving measure: 930 * - Common fields struct, always defined, and embedded within it: 931 * - L2-only fields, always allocated but undefined when not in L2ARC 932 * - L1-only fields, only allocated when in L1ARC 933 * 934 * Buffer in L1 Buffer only in L2 935 * +------------------------+ +------------------------+ 936 * | arc_buf_hdr_t | | arc_buf_hdr_t | 937 * | | | | 938 * | | | | 939 * | | | | 940 * +------------------------+ +------------------------+ 941 * | l2arc_buf_hdr_t | | l2arc_buf_hdr_t | 942 * | (undefined if L1-only) | | | 943 * +------------------------+ +------------------------+ 944 * | l1arc_buf_hdr_t | 945 * | | 946 * | | 947 * | | 948 * | | 949 * +------------------------+ 950 * 951 * Because it's possible for the L2ARC to become extremely large, we can wind 952 * up eating a lot of memory in L2ARC buffer headers, so the size of a header 953 * is minimized by only allocating the fields necessary for an L1-cached buffer 954 * when a header is actually in the L1 cache. The sub-headers (l1arc_buf_hdr and 955 * l2arc_buf_hdr) are embedded rather than allocated separately to save a couple 956 * words in pointers. arc_hdr_realloc() is used to switch a header between 957 * these two allocation states. 958 */ 959typedef struct l1arc_buf_hdr { 960 kmutex_t b_freeze_lock; 961 zio_cksum_t *b_freeze_cksum; 962#ifdef ZFS_DEBUG 963 /* 964 * used for debugging wtih kmem_flags - by allocating and freeing 965 * b_thawed when the buffer is thawed, we get a record of the stack 966 * trace that thawed it. 967 */ 968 void *b_thawed; 969#endif 970 971 arc_buf_t *b_buf; 972 uint32_t b_bufcnt; 973 /* for waiting on writes to complete */ 974 kcondvar_t b_cv; 975 uint8_t b_byteswap; 976 977 /* protected by arc state mutex */ 978 arc_state_t *b_state; 979 multilist_node_t b_arc_node; 980 981 /* updated atomically */ 982 clock_t b_arc_access; 983 984 /* self protecting */ 985 refcount_t b_refcnt; 986 987 arc_callback_t *b_acb; 988 void *b_pdata; 989} l1arc_buf_hdr_t; 990 991typedef struct l2arc_dev l2arc_dev_t; 992 993typedef struct l2arc_buf_hdr { 994 /* protected by arc_buf_hdr mutex */ 995 l2arc_dev_t *b_dev; /* L2ARC device */ 996 uint64_t b_daddr; /* disk address, offset byte */ 997 998 list_node_t b_l2node; 999} l2arc_buf_hdr_t; 1000 1001struct arc_buf_hdr { 1002 /* protected by hash lock */ 1003 dva_t b_dva; 1004 uint64_t b_birth; 1005 1006 arc_buf_contents_t b_type; 1007 arc_buf_hdr_t *b_hash_next; 1008 arc_flags_t b_flags; 1009 1010 /* 1011 * This field stores the size of the data buffer after 1012 * compression, and is set in the arc's zio completion handlers. 1013 * It is in units of SPA_MINBLOCKSIZE (e.g. 1 == 512 bytes). 1014 * 1015 * While the block pointers can store up to 32MB in their psize 1016 * field, we can only store up to 32MB minus 512B. This is due 1017 * to the bp using a bias of 1, whereas we use a bias of 0 (i.e. 1018 * a field of zeros represents 512B in the bp). We can't use a 1019 * bias of 1 since we need to reserve a psize of zero, here, to 1020 * represent holes and embedded blocks. 1021 * 1022 * This isn't a problem in practice, since the maximum size of a 1023 * buffer is limited to 16MB, so we never need to store 32MB in 1024 * this field. Even in the upstream illumos code base, the 1025 * maximum size of a buffer is limited to 16MB. 1026 */ 1027 uint16_t b_psize; 1028 1029 /* 1030 * This field stores the size of the data buffer before 1031 * compression, and cannot change once set. It is in units 1032 * of SPA_MINBLOCKSIZE (e.g. 2 == 1024 bytes) 1033 */ 1034 uint16_t b_lsize; /* immutable */ 1035 uint64_t b_spa; /* immutable */ 1036 1037 /* L2ARC fields. Undefined when not in L2ARC. */ 1038 l2arc_buf_hdr_t b_l2hdr; 1039 /* L1ARC fields. Undefined when in l2arc_only state */ 1040 l1arc_buf_hdr_t b_l1hdr; 1041}; 1042 1043#if defined(__FreeBSD__) && defined(_KERNEL) 1044static int 1045sysctl_vfs_zfs_arc_meta_limit(SYSCTL_HANDLER_ARGS) 1046{ 1047 uint64_t val; 1048 int err; 1049 1050 val = arc_meta_limit; 1051 err = sysctl_handle_64(oidp, &val, 0, req); 1052 if (err != 0 || req->newptr == NULL) 1053 return (err); 1054 1055 if (val <= 0 || val > arc_c_max) 1056 return (EINVAL); 1057 1058 arc_meta_limit = val; 1059 return (0); 1060} 1061 1062static int 1063sysctl_vfs_zfs_arc_max(SYSCTL_HANDLER_ARGS) 1064{ 1065 uint64_t val; 1066 int err; 1067 1068 val = zfs_arc_max; 1069 err = sysctl_handle_64(oidp, &val, 0, req); 1070 if (err != 0 || req->newptr == NULL) 1071 return (err); 1072 1073 if (zfs_arc_max == 0) { 1074 /* Loader tunable so blindly set */ 1075 zfs_arc_max = val; 1076 return (0); 1077 } 1078 1079 if (val < arc_abs_min || val > kmem_size()) 1080 return (EINVAL); 1081 if (val < arc_c_min) 1082 return (EINVAL); 1083 if (zfs_arc_meta_limit > 0 && val < zfs_arc_meta_limit) 1084 return (EINVAL); 1085 1086 arc_c_max = val; 1087 1088 arc_c = arc_c_max; 1089 arc_p = (arc_c >> 1); 1090 1091 if (zfs_arc_meta_limit == 0) { 1092 /* limit meta-data to 1/4 of the arc capacity */ 1093 arc_meta_limit = arc_c_max / 4; 1094 } 1095 1096 /* if kmem_flags are set, lets try to use less memory */ 1097 if (kmem_debugging()) 1098 arc_c = arc_c / 2; 1099 1100 zfs_arc_max = arc_c; 1101 1102 return (0); 1103} 1104 1105static int 1106sysctl_vfs_zfs_arc_min(SYSCTL_HANDLER_ARGS) 1107{ 1108 uint64_t val; 1109 int err; 1110 1111 val = zfs_arc_min; 1112 err = sysctl_handle_64(oidp, &val, 0, req); 1113 if (err != 0 || req->newptr == NULL) 1114 return (err); 1115 1116 if (zfs_arc_min == 0) { 1117 /* Loader tunable so blindly set */ 1118 zfs_arc_min = val; 1119 return (0); 1120 } 1121 1122 if (val < arc_abs_min || val > arc_c_max) 1123 return (EINVAL); 1124 1125 arc_c_min = val; 1126 1127 if (zfs_arc_meta_min == 0) 1128 arc_meta_min = arc_c_min / 2; 1129 1130 if (arc_c < arc_c_min) 1131 arc_c = arc_c_min; 1132 1133 zfs_arc_min = arc_c_min; 1134 1135 return (0); 1136} 1137#endif 1138 1139#define GHOST_STATE(state) \ 1140 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \ 1141 (state) == arc_l2c_only) 1142 1143#define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_FLAG_IN_HASH_TABLE) 1144#define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) 1145#define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_FLAG_IO_ERROR) 1146#define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_FLAG_PREFETCH) 1147#define HDR_COMPRESSION_ENABLED(hdr) \ 1148 ((hdr)->b_flags & ARC_FLAG_COMPRESSED_ARC) 1149 1150#define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_FLAG_L2CACHE) 1151#define HDR_L2_READING(hdr) \ 1152 (((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) && \ 1153 ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR)) 1154#define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITING) 1155#define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_FLAG_L2_EVICTED) 1156#define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITE_HEAD) 1157#define HDR_SHARED_DATA(hdr) ((hdr)->b_flags & ARC_FLAG_SHARED_DATA) 1158 1159#define HDR_ISTYPE_METADATA(hdr) \ 1160 ((hdr)->b_flags & ARC_FLAG_BUFC_METADATA) 1161#define HDR_ISTYPE_DATA(hdr) (!HDR_ISTYPE_METADATA(hdr)) 1162 1163#define HDR_HAS_L1HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L1HDR) 1164#define HDR_HAS_L2HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR) 1165 1166/* For storing compression mode in b_flags */ 1167#define HDR_COMPRESS_OFFSET (highbit64(ARC_FLAG_COMPRESS_0) - 1) 1168 1169#define HDR_GET_COMPRESS(hdr) ((enum zio_compress)BF32_GET((hdr)->b_flags, \ 1170 HDR_COMPRESS_OFFSET, SPA_COMPRESSBITS)) 1171#define HDR_SET_COMPRESS(hdr, cmp) BF32_SET((hdr)->b_flags, \ 1172 HDR_COMPRESS_OFFSET, SPA_COMPRESSBITS, (cmp)); 1173 1174#define ARC_BUF_LAST(buf) ((buf)->b_next == NULL) 1175 1176/* 1177 * Other sizes 1178 */ 1179 1180#define HDR_FULL_SIZE ((int64_t)sizeof (arc_buf_hdr_t)) 1181#define HDR_L2ONLY_SIZE ((int64_t)offsetof(arc_buf_hdr_t, b_l1hdr)) 1182 1183/* 1184 * Hash table routines 1185 */ 1186 1187#define HT_LOCK_PAD CACHE_LINE_SIZE 1188 1189struct ht_lock { 1190 kmutex_t ht_lock; 1191#ifdef _KERNEL 1192 unsigned char pad[(HT_LOCK_PAD - sizeof (kmutex_t))]; 1193#endif 1194}; 1195 1196#define BUF_LOCKS 256 1197typedef struct buf_hash_table { 1198 uint64_t ht_mask; 1199 arc_buf_hdr_t **ht_table; 1200 struct ht_lock ht_locks[BUF_LOCKS] __aligned(CACHE_LINE_SIZE); 1201} buf_hash_table_t; 1202 1203static buf_hash_table_t buf_hash_table; 1204 1205#define BUF_HASH_INDEX(spa, dva, birth) \ 1206 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask) 1207#define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)]) 1208#define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock)) 1209#define HDR_LOCK(hdr) \ 1210 (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth))) 1211 1212uint64_t zfs_crc64_table[256]; 1213 1214/* 1215 * Level 2 ARC 1216 */ 1217 1218#define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */ 1219#define L2ARC_HEADROOM 2 /* num of writes */ 1220/* 1221 * If we discover during ARC scan any buffers to be compressed, we boost 1222 * our headroom for the next scanning cycle by this percentage multiple. 1223 */ 1224#define L2ARC_HEADROOM_BOOST 200 1225#define L2ARC_FEED_SECS 1 /* caching interval secs */ 1226#define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */ 1227 1228#define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent) 1229#define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done) 1230 1231/* L2ARC Performance Tunables */ 1232uint64_t l2arc_write_max = L2ARC_WRITE_SIZE; /* default max write size */ 1233uint64_t l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra write during warmup */ 1234uint64_t l2arc_headroom = L2ARC_HEADROOM; /* number of dev writes */ 1235uint64_t l2arc_headroom_boost = L2ARC_HEADROOM_BOOST; 1236uint64_t l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */ 1237uint64_t l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval milliseconds */ 1238boolean_t l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */ 1239boolean_t l2arc_feed_again = B_TRUE; /* turbo warmup */ 1240boolean_t l2arc_norw = B_TRUE; /* no reads during writes */ 1241 1242SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_max, CTLFLAG_RW, 1243 &l2arc_write_max, 0, "max write size"); 1244SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_boost, CTLFLAG_RW, 1245 &l2arc_write_boost, 0, "extra write during warmup"); 1246SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_headroom, CTLFLAG_RW, 1247 &l2arc_headroom, 0, "number of dev writes"); 1248SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_secs, CTLFLAG_RW, 1249 &l2arc_feed_secs, 0, "interval seconds"); 1250SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_min_ms, CTLFLAG_RW, 1251 &l2arc_feed_min_ms, 0, "min interval milliseconds"); 1252 1253SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_noprefetch, CTLFLAG_RW, 1254 &l2arc_noprefetch, 0, "don't cache prefetch bufs"); 1255SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_feed_again, CTLFLAG_RW, 1256 &l2arc_feed_again, 0, "turbo warmup"); 1257SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_norw, CTLFLAG_RW, 1258 &l2arc_norw, 0, "no reads during writes"); 1259 1260SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_size, CTLFLAG_RD, 1261 &ARC_anon.arcs_size.rc_count, 0, "size of anonymous state"); 1262SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_metadata_esize, CTLFLAG_RD, 1263 &ARC_anon.arcs_esize[ARC_BUFC_METADATA].rc_count, 0, 1264 "size of anonymous state"); 1265SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_data_esize, CTLFLAG_RD, 1266 &ARC_anon.arcs_esize[ARC_BUFC_DATA].rc_count, 0, 1267 "size of anonymous state"); 1268 1269SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_size, CTLFLAG_RD, 1270 &ARC_mru.arcs_size.rc_count, 0, "size of mru state"); 1271SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_metadata_esize, CTLFLAG_RD, 1272 &ARC_mru.arcs_esize[ARC_BUFC_METADATA].rc_count, 0, 1273 "size of metadata in mru state"); 1274SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_data_esize, CTLFLAG_RD, 1275 &ARC_mru.arcs_esize[ARC_BUFC_DATA].rc_count, 0, 1276 "size of data in mru state"); 1277 1278SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_size, CTLFLAG_RD, 1279 &ARC_mru_ghost.arcs_size.rc_count, 0, "size of mru ghost state"); 1280SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_metadata_esize, CTLFLAG_RD, 1281 &ARC_mru_ghost.arcs_esize[ARC_BUFC_METADATA].rc_count, 0, 1282 "size of metadata in mru ghost state"); 1283SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_data_esize, CTLFLAG_RD, 1284 &ARC_mru_ghost.arcs_esize[ARC_BUFC_DATA].rc_count, 0, 1285 "size of data in mru ghost state"); 1286 1287SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_size, CTLFLAG_RD, 1288 &ARC_mfu.arcs_size.rc_count, 0, "size of mfu state"); 1289SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_metadata_esize, CTLFLAG_RD, 1290 &ARC_mfu.arcs_esize[ARC_BUFC_METADATA].rc_count, 0, 1291 "size of metadata in mfu state"); 1292SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_data_esize, CTLFLAG_RD, 1293 &ARC_mfu.arcs_esize[ARC_BUFC_DATA].rc_count, 0, 1294 "size of data in mfu state"); 1295 1296SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_size, CTLFLAG_RD, 1297 &ARC_mfu_ghost.arcs_size.rc_count, 0, "size of mfu ghost state"); 1298SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_metadata_esize, CTLFLAG_RD, 1299 &ARC_mfu_ghost.arcs_esize[ARC_BUFC_METADATA].rc_count, 0, 1300 "size of metadata in mfu ghost state"); 1301SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_data_esize, CTLFLAG_RD, 1302 &ARC_mfu_ghost.arcs_esize[ARC_BUFC_DATA].rc_count, 0, 1303 "size of data in mfu ghost state"); 1304 1305SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2c_only_size, CTLFLAG_RD, 1306 &ARC_l2c_only.arcs_size.rc_count, 0, "size of mru state"); 1307 1308/* 1309 * L2ARC Internals 1310 */ 1311struct l2arc_dev { 1312 vdev_t *l2ad_vdev; /* vdev */ 1313 spa_t *l2ad_spa; /* spa */ 1314 uint64_t l2ad_hand; /* next write location */ 1315 uint64_t l2ad_start; /* first addr on device */ 1316 uint64_t l2ad_end; /* last addr on device */ 1317 boolean_t l2ad_first; /* first sweep through */ 1318 boolean_t l2ad_writing; /* currently writing */ 1319 kmutex_t l2ad_mtx; /* lock for buffer list */ 1320 list_t l2ad_buflist; /* buffer list */ 1321 list_node_t l2ad_node; /* device list node */ 1322 refcount_t l2ad_alloc; /* allocated bytes */ 1323}; 1324 1325static list_t L2ARC_dev_list; /* device list */ 1326static list_t *l2arc_dev_list; /* device list pointer */ 1327static kmutex_t l2arc_dev_mtx; /* device list mutex */ 1328static l2arc_dev_t *l2arc_dev_last; /* last device used */ 1329static list_t L2ARC_free_on_write; /* free after write buf list */ 1330static list_t *l2arc_free_on_write; /* free after write list ptr */ 1331static kmutex_t l2arc_free_on_write_mtx; /* mutex for list */ 1332static uint64_t l2arc_ndev; /* number of devices */ 1333 1334typedef struct l2arc_read_callback { 1335 arc_buf_hdr_t *l2rcb_hdr; /* read buffer */ 1336 blkptr_t l2rcb_bp; /* original blkptr */ 1337 zbookmark_phys_t l2rcb_zb; /* original bookmark */ 1338 int l2rcb_flags; /* original flags */ 1339 void *l2rcb_data; /* temporary buffer */ 1340} l2arc_read_callback_t; 1341 1342typedef struct l2arc_write_callback { 1343 l2arc_dev_t *l2wcb_dev; /* device info */ 1344 arc_buf_hdr_t *l2wcb_head; /* head of write buflist */ 1345} l2arc_write_callback_t; 1346 1347typedef struct l2arc_data_free { 1348 /* protected by l2arc_free_on_write_mtx */ 1349 void *l2df_data; 1350 size_t l2df_size; 1351 arc_buf_contents_t l2df_type; 1352 list_node_t l2df_list_node; 1353} l2arc_data_free_t; 1354 1355static kmutex_t l2arc_feed_thr_lock; 1356static kcondvar_t l2arc_feed_thr_cv; 1357static uint8_t l2arc_thread_exit; 1358 1359static void *arc_get_data_buf(arc_buf_hdr_t *, uint64_t, void *); 1360static void arc_free_data_buf(arc_buf_hdr_t *, void *, uint64_t, void *); 1361static void arc_hdr_free_pdata(arc_buf_hdr_t *hdr); 1362static void arc_hdr_alloc_pdata(arc_buf_hdr_t *); 1363static void arc_access(arc_buf_hdr_t *, kmutex_t *); 1364static boolean_t arc_is_overflowing(); 1365static void arc_buf_watch(arc_buf_t *); 1366 1367static arc_buf_contents_t arc_buf_type(arc_buf_hdr_t *); 1368static uint32_t arc_bufc_to_flags(arc_buf_contents_t); 1369static inline void arc_hdr_set_flags(arc_buf_hdr_t *hdr, arc_flags_t flags); 1370static inline void arc_hdr_clear_flags(arc_buf_hdr_t *hdr, arc_flags_t flags); 1371 1372static boolean_t l2arc_write_eligible(uint64_t, arc_buf_hdr_t *); 1373static void l2arc_read_done(zio_t *); 1374 1375static void 1376l2arc_trim(const arc_buf_hdr_t *hdr) 1377{ 1378 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev; 1379 1380 ASSERT(HDR_HAS_L2HDR(hdr)); 1381 ASSERT(MUTEX_HELD(&dev->l2ad_mtx)); 1382 1383 if (HDR_GET_PSIZE(hdr) != 0) { 1384 trim_map_free(dev->l2ad_vdev, hdr->b_l2hdr.b_daddr, 1385 HDR_GET_PSIZE(hdr), 0); 1386 } 1387} 1388 1389static uint64_t 1390buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth) 1391{ 1392 uint8_t *vdva = (uint8_t *)dva; 1393 uint64_t crc = -1ULL; 1394 int i; 1395 1396 ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY); 1397 1398 for (i = 0; i < sizeof (dva_t); i++) 1399 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ vdva[i]) & 0xFF]; 1400 1401 crc ^= (spa>>8) ^ birth; 1402 1403 return (crc); 1404} 1405 1406#define HDR_EMPTY(hdr) \ 1407 ((hdr)->b_dva.dva_word[0] == 0 && \ 1408 (hdr)->b_dva.dva_word[1] == 0) 1409 1410#define HDR_EQUAL(spa, dva, birth, hdr) \ 1411 ((hdr)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \ 1412 ((hdr)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \ 1413 ((hdr)->b_birth == birth) && ((hdr)->b_spa == spa) 1414 1415static void 1416buf_discard_identity(arc_buf_hdr_t *hdr) 1417{ 1418 hdr->b_dva.dva_word[0] = 0; 1419 hdr->b_dva.dva_word[1] = 0; 1420 hdr->b_birth = 0; 1421} 1422 1423static arc_buf_hdr_t * 1424buf_hash_find(uint64_t spa, const blkptr_t *bp, kmutex_t **lockp) 1425{ 1426 const dva_t *dva = BP_IDENTITY(bp); 1427 uint64_t birth = BP_PHYSICAL_BIRTH(bp); 1428 uint64_t idx = BUF_HASH_INDEX(spa, dva, birth); 1429 kmutex_t *hash_lock = BUF_HASH_LOCK(idx); 1430 arc_buf_hdr_t *hdr; 1431 1432 mutex_enter(hash_lock); 1433 for (hdr = buf_hash_table.ht_table[idx]; hdr != NULL; 1434 hdr = hdr->b_hash_next) { 1435 if (HDR_EQUAL(spa, dva, birth, hdr)) { 1436 *lockp = hash_lock; 1437 return (hdr); 1438 } 1439 } 1440 mutex_exit(hash_lock); 1441 *lockp = NULL; 1442 return (NULL); 1443} 1444 1445/* 1446 * Insert an entry into the hash table. If there is already an element 1447 * equal to elem in the hash table, then the already existing element 1448 * will be returned and the new element will not be inserted. 1449 * Otherwise returns NULL. 1450 * If lockp == NULL, the caller is assumed to already hold the hash lock. 1451 */ 1452static arc_buf_hdr_t * 1453buf_hash_insert(arc_buf_hdr_t *hdr, kmutex_t **lockp) 1454{ 1455 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth); 1456 kmutex_t *hash_lock = BUF_HASH_LOCK(idx); 1457 arc_buf_hdr_t *fhdr; 1458 uint32_t i; 1459 1460 ASSERT(!DVA_IS_EMPTY(&hdr->b_dva)); 1461 ASSERT(hdr->b_birth != 0); 1462 ASSERT(!HDR_IN_HASH_TABLE(hdr)); 1463 1464 if (lockp != NULL) { 1465 *lockp = hash_lock; 1466 mutex_enter(hash_lock); 1467 } else { 1468 ASSERT(MUTEX_HELD(hash_lock)); 1469 } 1470 1471 for (fhdr = buf_hash_table.ht_table[idx], i = 0; fhdr != NULL; 1472 fhdr = fhdr->b_hash_next, i++) { 1473 if (HDR_EQUAL(hdr->b_spa, &hdr->b_dva, hdr->b_birth, fhdr)) 1474 return (fhdr); 1475 } 1476 1477 hdr->b_hash_next = buf_hash_table.ht_table[idx]; 1478 buf_hash_table.ht_table[idx] = hdr; 1479 arc_hdr_set_flags(hdr, ARC_FLAG_IN_HASH_TABLE); 1480 1481 /* collect some hash table performance data */ 1482 if (i > 0) { 1483 ARCSTAT_BUMP(arcstat_hash_collisions); 1484 if (i == 1) 1485 ARCSTAT_BUMP(arcstat_hash_chains); 1486 1487 ARCSTAT_MAX(arcstat_hash_chain_max, i); 1488 } 1489 1490 ARCSTAT_BUMP(arcstat_hash_elements); 1491 ARCSTAT_MAXSTAT(arcstat_hash_elements); 1492 1493 return (NULL); 1494} 1495 1496static void 1497buf_hash_remove(arc_buf_hdr_t *hdr) 1498{ 1499 arc_buf_hdr_t *fhdr, **hdrp; 1500 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth); 1501 1502 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx))); 1503 ASSERT(HDR_IN_HASH_TABLE(hdr)); 1504 1505 hdrp = &buf_hash_table.ht_table[idx]; 1506 while ((fhdr = *hdrp) != hdr) { 1507 ASSERT3P(fhdr, !=, NULL); 1508 hdrp = &fhdr->b_hash_next; 1509 } 1510 *hdrp = hdr->b_hash_next; 1511 hdr->b_hash_next = NULL; 1512 arc_hdr_clear_flags(hdr, ARC_FLAG_IN_HASH_TABLE); 1513 1514 /* collect some hash table performance data */ 1515 ARCSTAT_BUMPDOWN(arcstat_hash_elements); 1516 1517 if (buf_hash_table.ht_table[idx] && 1518 buf_hash_table.ht_table[idx]->b_hash_next == NULL) 1519 ARCSTAT_BUMPDOWN(arcstat_hash_chains); 1520} 1521 1522/* 1523 * Global data structures and functions for the buf kmem cache. 1524 */ 1525static kmem_cache_t *hdr_full_cache; 1526static kmem_cache_t *hdr_l2only_cache; 1527static kmem_cache_t *buf_cache; 1528 1529static void 1530buf_fini(void) 1531{ 1532 int i; 1533 1534 kmem_free(buf_hash_table.ht_table, 1535 (buf_hash_table.ht_mask + 1) * sizeof (void *)); 1536 for (i = 0; i < BUF_LOCKS; i++) 1537 mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock); 1538 kmem_cache_destroy(hdr_full_cache); 1539 kmem_cache_destroy(hdr_l2only_cache); 1540 kmem_cache_destroy(buf_cache); 1541} 1542 1543/* 1544 * Constructor callback - called when the cache is empty 1545 * and a new buf is requested. 1546 */ 1547/* ARGSUSED */ 1548static int 1549hdr_full_cons(void *vbuf, void *unused, int kmflag) 1550{ 1551 arc_buf_hdr_t *hdr = vbuf; 1552 1553 bzero(hdr, HDR_FULL_SIZE); 1554 cv_init(&hdr->b_l1hdr.b_cv, NULL, CV_DEFAULT, NULL); 1555 refcount_create(&hdr->b_l1hdr.b_refcnt); 1556 mutex_init(&hdr->b_l1hdr.b_freeze_lock, NULL, MUTEX_DEFAULT, NULL); 1557 multilist_link_init(&hdr->b_l1hdr.b_arc_node); 1558 arc_space_consume(HDR_FULL_SIZE, ARC_SPACE_HDRS); 1559 1560 return (0); 1561} 1562 1563/* ARGSUSED */ 1564static int 1565hdr_l2only_cons(void *vbuf, void *unused, int kmflag) 1566{ 1567 arc_buf_hdr_t *hdr = vbuf; 1568 1569 bzero(hdr, HDR_L2ONLY_SIZE); 1570 arc_space_consume(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS); 1571 1572 return (0); 1573} 1574 1575/* ARGSUSED */ 1576static int 1577buf_cons(void *vbuf, void *unused, int kmflag) 1578{ 1579 arc_buf_t *buf = vbuf; 1580 1581 bzero(buf, sizeof (arc_buf_t)); 1582 mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL); 1583 arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS); 1584 1585 return (0); 1586} 1587 1588/* 1589 * Destructor callback - called when a cached buf is 1590 * no longer required. 1591 */ 1592/* ARGSUSED */ 1593static void 1594hdr_full_dest(void *vbuf, void *unused) 1595{ 1596 arc_buf_hdr_t *hdr = vbuf; 1597 1598 ASSERT(HDR_EMPTY(hdr)); 1599 cv_destroy(&hdr->b_l1hdr.b_cv); 1600 refcount_destroy(&hdr->b_l1hdr.b_refcnt); 1601 mutex_destroy(&hdr->b_l1hdr.b_freeze_lock); 1602 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 1603 arc_space_return(HDR_FULL_SIZE, ARC_SPACE_HDRS); 1604} 1605 1606/* ARGSUSED */ 1607static void 1608hdr_l2only_dest(void *vbuf, void *unused) 1609{ 1610 arc_buf_hdr_t *hdr = vbuf; 1611 1612 ASSERT(HDR_EMPTY(hdr)); 1613 arc_space_return(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS); 1614} 1615 1616/* ARGSUSED */ 1617static void 1618buf_dest(void *vbuf, void *unused) 1619{ 1620 arc_buf_t *buf = vbuf; 1621 1622 mutex_destroy(&buf->b_evict_lock); 1623 arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS); 1624} 1625 1626/* 1627 * Reclaim callback -- invoked when memory is low. 1628 */ 1629/* ARGSUSED */ 1630static void 1631hdr_recl(void *unused) 1632{ 1633 dprintf("hdr_recl called\n"); 1634 /* 1635 * umem calls the reclaim func when we destroy the buf cache, 1636 * which is after we do arc_fini(). 1637 */ 1638 if (!arc_dead) 1639 cv_signal(&arc_reclaim_thread_cv); 1640} 1641 1642static void 1643buf_init(void) 1644{ 1645 uint64_t *ct; 1646 uint64_t hsize = 1ULL << 12; 1647 int i, j; 1648 1649 /* 1650 * The hash table is big enough to fill all of physical memory 1651 * with an average block size of zfs_arc_average_blocksize (default 8K). 1652 * By default, the table will take up 1653 * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers). 1654 */ 1655 while (hsize * zfs_arc_average_blocksize < (uint64_t)physmem * PAGESIZE) 1656 hsize <<= 1; 1657retry: 1658 buf_hash_table.ht_mask = hsize - 1; 1659 buf_hash_table.ht_table = 1660 kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP); 1661 if (buf_hash_table.ht_table == NULL) { 1662 ASSERT(hsize > (1ULL << 8)); 1663 hsize >>= 1; 1664 goto retry; 1665 } 1666 1667 hdr_full_cache = kmem_cache_create("arc_buf_hdr_t_full", HDR_FULL_SIZE, 1668 0, hdr_full_cons, hdr_full_dest, hdr_recl, NULL, NULL, 0); 1669 hdr_l2only_cache = kmem_cache_create("arc_buf_hdr_t_l2only", 1670 HDR_L2ONLY_SIZE, 0, hdr_l2only_cons, hdr_l2only_dest, hdr_recl, 1671 NULL, NULL, 0); 1672 buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t), 1673 0, buf_cons, buf_dest, NULL, NULL, NULL, 0); 1674 1675 for (i = 0; i < 256; i++) 1676 for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--) 1677 *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY); 1678 1679 for (i = 0; i < BUF_LOCKS; i++) { 1680 mutex_init(&buf_hash_table.ht_locks[i].ht_lock, 1681 NULL, MUTEX_DEFAULT, NULL); 1682 } 1683} 1684 1685#define ARC_MINTIME (hz>>4) /* 62 ms */ 1686 1687static inline boolean_t 1688arc_buf_is_shared(arc_buf_t *buf) 1689{ 1690 boolean_t shared = (buf->b_data != NULL && 1691 buf->b_data == buf->b_hdr->b_l1hdr.b_pdata); 1692 IMPLY(shared, HDR_SHARED_DATA(buf->b_hdr)); 1693 return (shared); 1694} 1695 1696static inline void 1697arc_cksum_free(arc_buf_hdr_t *hdr) 1698{ 1699 ASSERT(HDR_HAS_L1HDR(hdr)); 1700 mutex_enter(&hdr->b_l1hdr.b_freeze_lock); 1701 if (hdr->b_l1hdr.b_freeze_cksum != NULL) { 1702 kmem_free(hdr->b_l1hdr.b_freeze_cksum, sizeof (zio_cksum_t)); 1703 hdr->b_l1hdr.b_freeze_cksum = NULL; 1704 } 1705 mutex_exit(&hdr->b_l1hdr.b_freeze_lock); 1706} 1707 1708static void 1709arc_cksum_verify(arc_buf_t *buf) 1710{ 1711 arc_buf_hdr_t *hdr = buf->b_hdr; 1712 zio_cksum_t zc; 1713 1714 if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 1715 return; 1716 1717 ASSERT(HDR_HAS_L1HDR(hdr)); 1718 1719 mutex_enter(&hdr->b_l1hdr.b_freeze_lock); 1720 if (hdr->b_l1hdr.b_freeze_cksum == NULL || HDR_IO_ERROR(hdr)) { 1721 mutex_exit(&hdr->b_l1hdr.b_freeze_lock); 1722 return; 1723 } 1724 fletcher_2_native(buf->b_data, HDR_GET_LSIZE(hdr), NULL, &zc); 1725 if (!ZIO_CHECKSUM_EQUAL(*hdr->b_l1hdr.b_freeze_cksum, zc)) 1726 panic("buffer modified while frozen!"); 1727 mutex_exit(&hdr->b_l1hdr.b_freeze_lock); 1728} 1729 1730static boolean_t 1731arc_cksum_is_equal(arc_buf_hdr_t *hdr, zio_t *zio) 1732{ 1733 enum zio_compress compress = BP_GET_COMPRESS(zio->io_bp); 1734 boolean_t valid_cksum; 1735 1736 ASSERT(!BP_IS_EMBEDDED(zio->io_bp)); 1737 VERIFY3U(BP_GET_PSIZE(zio->io_bp), ==, HDR_GET_PSIZE(hdr)); 1738 1739 /* 1740 * We rely on the blkptr's checksum to determine if the block 1741 * is valid or not. When compressed arc is enabled, the l2arc 1742 * writes the block to the l2arc just as it appears in the pool. 1743 * This allows us to use the blkptr's checksum to validate the 1744 * data that we just read off of the l2arc without having to store 1745 * a separate checksum in the arc_buf_hdr_t. However, if compressed 1746 * arc is disabled, then the data written to the l2arc is always 1747 * uncompressed and won't match the block as it exists in the main 1748 * pool. When this is the case, we must first compress it if it is 1749 * compressed on the main pool before we can validate the checksum. 1750 */ 1751 if (!HDR_COMPRESSION_ENABLED(hdr) && compress != ZIO_COMPRESS_OFF) { 1752 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF); 1753 uint64_t lsize = HDR_GET_LSIZE(hdr); 1754 uint64_t csize; 1755 1756 void *cbuf = zio_buf_alloc(HDR_GET_PSIZE(hdr)); 1757 csize = zio_compress_data(compress, zio->io_data, cbuf, lsize); 1758 ASSERT3U(csize, <=, HDR_GET_PSIZE(hdr)); 1759 if (csize < HDR_GET_PSIZE(hdr)) { 1760 /* 1761 * Compressed blocks are always a multiple of the 1762 * smallest ashift in the pool. Ideally, we would 1763 * like to round up the csize to the next 1764 * spa_min_ashift but that value may have changed 1765 * since the block was last written. Instead, 1766 * we rely on the fact that the hdr's psize 1767 * was set to the psize of the block when it was 1768 * last written. We set the csize to that value 1769 * and zero out any part that should not contain 1770 * data. 1771 */ 1772 bzero((char *)cbuf + csize, HDR_GET_PSIZE(hdr) - csize); 1773 csize = HDR_GET_PSIZE(hdr); 1774 } 1775 zio_push_transform(zio, cbuf, csize, HDR_GET_PSIZE(hdr), NULL); 1776 } 1777 1778 /* 1779 * Block pointers always store the checksum for the logical data. 1780 * If the block pointer has the gang bit set, then the checksum 1781 * it represents is for the reconstituted data and not for an 1782 * individual gang member. The zio pipeline, however, must be able to 1783 * determine the checksum of each of the gang constituents so it 1784 * treats the checksum comparison differently than what we need 1785 * for l2arc blocks. This prevents us from using the 1786 * zio_checksum_error() interface directly. Instead we must call the 1787 * zio_checksum_error_impl() so that we can ensure the checksum is 1788 * generated using the correct checksum algorithm and accounts for the 1789 * logical I/O size and not just a gang fragment. 1790 */ 1791 valid_cksum = (zio_checksum_error_impl(zio->io_spa, zio->io_bp, 1792 BP_GET_CHECKSUM(zio->io_bp), zio->io_data, zio->io_size, 1793 zio->io_offset, NULL) == 0); 1794 zio_pop_transforms(zio); 1795 return (valid_cksum); 1796} 1797 1798static void 1799arc_cksum_compute(arc_buf_t *buf) 1800{ 1801 arc_buf_hdr_t *hdr = buf->b_hdr; 1802 1803 if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 1804 return; 1805 1806 ASSERT(HDR_HAS_L1HDR(hdr)); 1807 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1808 if (hdr->b_l1hdr.b_freeze_cksum != NULL) { 1809 mutex_exit(&hdr->b_l1hdr.b_freeze_lock); 1810 return; 1811 } 1812 hdr->b_l1hdr.b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t), 1813 KM_SLEEP); 1814 fletcher_2_native(buf->b_data, HDR_GET_LSIZE(hdr), NULL, 1815 hdr->b_l1hdr.b_freeze_cksum); 1816 mutex_exit(&hdr->b_l1hdr.b_freeze_lock); 1817#ifdef illumos 1818 arc_buf_watch(buf); 1819#endif 1820} 1821 1822#ifdef illumos 1823#ifndef _KERNEL 1824typedef struct procctl { 1825 long cmd; 1826 prwatch_t prwatch; 1827} procctl_t; 1828#endif 1829 1830/* ARGSUSED */ 1831static void 1832arc_buf_unwatch(arc_buf_t *buf) 1833{ 1834#ifndef _KERNEL 1835 if (arc_watch) { 1836 int result; 1837 procctl_t ctl; 1838 ctl.cmd = PCWATCH; 1839 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data; 1840 ctl.prwatch.pr_size = 0; 1841 ctl.prwatch.pr_wflags = 0; 1842 result = write(arc_procfd, &ctl, sizeof (ctl)); 1843 ASSERT3U(result, ==, sizeof (ctl)); 1844 } 1845#endif 1846} 1847 1848/* ARGSUSED */ 1849static void 1850arc_buf_watch(arc_buf_t *buf) 1851{ 1852#ifndef _KERNEL 1853 if (arc_watch) { 1854 int result; 1855 procctl_t ctl; 1856 ctl.cmd = PCWATCH; 1857 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data; 1858 ctl.prwatch.pr_size = HDR_GET_LSIZE(buf->b_hdr); 1859 ctl.prwatch.pr_wflags = WA_WRITE; 1860 result = write(arc_procfd, &ctl, sizeof (ctl)); 1861 ASSERT3U(result, ==, sizeof (ctl)); 1862 } 1863#endif 1864} 1865#endif /* illumos */ 1866 1867static arc_buf_contents_t 1868arc_buf_type(arc_buf_hdr_t *hdr) 1869{ 1870 arc_buf_contents_t type; 1871 if (HDR_ISTYPE_METADATA(hdr)) { 1872 type = ARC_BUFC_METADATA; 1873 } else { 1874 type = ARC_BUFC_DATA; 1875 } 1876 VERIFY3U(hdr->b_type, ==, type); 1877 return (type); 1878} 1879 1880static uint32_t 1881arc_bufc_to_flags(arc_buf_contents_t type) 1882{ 1883 switch (type) { 1884 case ARC_BUFC_DATA: 1885 /* metadata field is 0 if buffer contains normal data */ 1886 return (0); 1887 case ARC_BUFC_METADATA: 1888 return (ARC_FLAG_BUFC_METADATA); 1889 default: 1890 break; 1891 } 1892 panic("undefined ARC buffer type!"); 1893 return ((uint32_t)-1); 1894} 1895 1896void 1897arc_buf_thaw(arc_buf_t *buf) 1898{ 1899 arc_buf_hdr_t *hdr = buf->b_hdr; 1900 1901 if (zfs_flags & ZFS_DEBUG_MODIFY) { 1902 if (hdr->b_l1hdr.b_state != arc_anon) 1903 panic("modifying non-anon buffer!"); 1904 if (HDR_IO_IN_PROGRESS(hdr)) 1905 panic("modifying buffer while i/o in progress!"); 1906 arc_cksum_verify(buf); 1907 } 1908 1909 ASSERT(HDR_HAS_L1HDR(hdr)); 1910 arc_cksum_free(hdr); 1911 1912 mutex_enter(&hdr->b_l1hdr.b_freeze_lock); 1913#ifdef ZFS_DEBUG 1914 if (zfs_flags & ZFS_DEBUG_MODIFY) { 1915 if (hdr->b_l1hdr.b_thawed != NULL) 1916 kmem_free(hdr->b_l1hdr.b_thawed, 1); 1917 hdr->b_l1hdr.b_thawed = kmem_alloc(1, KM_SLEEP); 1918 } 1919#endif 1920 1921 mutex_exit(&hdr->b_l1hdr.b_freeze_lock); 1922 1923#ifdef illumos 1924 arc_buf_unwatch(buf); 1925#endif 1926} 1927 1928void 1929arc_buf_freeze(arc_buf_t *buf) 1930{ 1931 arc_buf_hdr_t *hdr = buf->b_hdr; 1932 kmutex_t *hash_lock; 1933 1934 if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 1935 return; 1936 1937 hash_lock = HDR_LOCK(hdr); 1938 mutex_enter(hash_lock); 1939 1940 ASSERT(HDR_HAS_L1HDR(hdr)); 1941 ASSERT(hdr->b_l1hdr.b_freeze_cksum != NULL || 1942 hdr->b_l1hdr.b_state == arc_anon); 1943 arc_cksum_compute(buf); 1944 mutex_exit(hash_lock); 1945 1946} 1947 1948/* 1949 * The arc_buf_hdr_t's b_flags should never be modified directly. Instead, 1950 * the following functions should be used to ensure that the flags are 1951 * updated in a thread-safe way. When manipulating the flags either 1952 * the hash_lock must be held or the hdr must be undiscoverable. This 1953 * ensures that we're not racing with any other threads when updating 1954 * the flags. 1955 */ 1956static inline void 1957arc_hdr_set_flags(arc_buf_hdr_t *hdr, arc_flags_t flags) 1958{ 1959 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); 1960 hdr->b_flags |= flags; 1961} 1962 1963static inline void 1964arc_hdr_clear_flags(arc_buf_hdr_t *hdr, arc_flags_t flags) 1965{ 1966 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); 1967 hdr->b_flags &= ~flags; 1968} 1969 1970/* 1971 * Setting the compression bits in the arc_buf_hdr_t's b_flags is 1972 * done in a special way since we have to clear and set bits 1973 * at the same time. Consumers that wish to set the compression bits 1974 * must use this function to ensure that the flags are updated in 1975 * thread-safe manner. 1976 */ 1977static void 1978arc_hdr_set_compress(arc_buf_hdr_t *hdr, enum zio_compress cmp) 1979{ 1980 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); 1981 1982 /* 1983 * Holes and embedded blocks will always have a psize = 0 so 1984 * we ignore the compression of the blkptr and set the 1985 * arc_buf_hdr_t's compression to ZIO_COMPRESS_OFF. 1986 * Holes and embedded blocks remain anonymous so we don't 1987 * want to uncompress them. Mark them as uncompressed. 1988 */ 1989 if (!zfs_compressed_arc_enabled || HDR_GET_PSIZE(hdr) == 0) { 1990 arc_hdr_clear_flags(hdr, ARC_FLAG_COMPRESSED_ARC); 1991 HDR_SET_COMPRESS(hdr, ZIO_COMPRESS_OFF); 1992 ASSERT(!HDR_COMPRESSION_ENABLED(hdr)); 1993 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF); 1994 } else { 1995 arc_hdr_set_flags(hdr, ARC_FLAG_COMPRESSED_ARC); 1996 HDR_SET_COMPRESS(hdr, cmp); 1997 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, cmp); 1998 ASSERT(HDR_COMPRESSION_ENABLED(hdr)); 1999 } 2000} 2001 2002static int 2003arc_decompress(arc_buf_t *buf) 2004{ 2005 arc_buf_hdr_t *hdr = buf->b_hdr; 2006 dmu_object_byteswap_t bswap = hdr->b_l1hdr.b_byteswap; 2007 int error; 2008 2009 if (arc_buf_is_shared(buf)) { 2010 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF); 2011 } else if (HDR_GET_COMPRESS(hdr) == ZIO_COMPRESS_OFF) { 2012 /* 2013 * The arc_buf_hdr_t is either not compressed or is 2014 * associated with an embedded block or a hole in which 2015 * case they remain anonymous. 2016 */ 2017 IMPLY(HDR_COMPRESSION_ENABLED(hdr), HDR_GET_PSIZE(hdr) == 0 || 2018 HDR_GET_PSIZE(hdr) == HDR_GET_LSIZE(hdr)); 2019 ASSERT(!HDR_SHARED_DATA(hdr)); 2020 bcopy(hdr->b_l1hdr.b_pdata, buf->b_data, HDR_GET_LSIZE(hdr)); 2021 } else { 2022 ASSERT(!HDR_SHARED_DATA(hdr)); 2023 ASSERT3U(HDR_GET_LSIZE(hdr), !=, HDR_GET_PSIZE(hdr)); 2024 error = zio_decompress_data(HDR_GET_COMPRESS(hdr), 2025 hdr->b_l1hdr.b_pdata, buf->b_data, HDR_GET_PSIZE(hdr), 2026 HDR_GET_LSIZE(hdr)); 2027 if (error != 0) { 2028 zfs_dbgmsg("hdr %p, compress %d, psize %d, lsize %d", 2029 hdr, HDR_GET_COMPRESS(hdr), HDR_GET_PSIZE(hdr), 2030 HDR_GET_LSIZE(hdr)); 2031 return (SET_ERROR(EIO)); 2032 } 2033 } 2034 if (bswap != DMU_BSWAP_NUMFUNCS) { 2035 ASSERT(!HDR_SHARED_DATA(hdr)); 2036 ASSERT3U(bswap, <, DMU_BSWAP_NUMFUNCS); 2037 dmu_ot_byteswap[bswap].ob_func(buf->b_data, HDR_GET_LSIZE(hdr)); 2038 } 2039 arc_cksum_compute(buf); 2040 return (0); 2041} 2042 2043/* 2044 * Return the size of the block, b_pdata, that is stored in the arc_buf_hdr_t. 2045 */ 2046static uint64_t 2047arc_hdr_size(arc_buf_hdr_t *hdr) 2048{ 2049 uint64_t size; 2050 2051 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF && 2052 HDR_GET_PSIZE(hdr) > 0) { 2053 size = HDR_GET_PSIZE(hdr); 2054 } else { 2055 ASSERT3U(HDR_GET_LSIZE(hdr), !=, 0); 2056 size = HDR_GET_LSIZE(hdr); 2057 } 2058 return (size); 2059} 2060 2061/* 2062 * Increment the amount of evictable space in the arc_state_t's refcount. 2063 * We account for the space used by the hdr and the arc buf individually 2064 * so that we can add and remove them from the refcount individually. 2065 */ 2066static void 2067arc_evictable_space_increment(arc_buf_hdr_t *hdr, arc_state_t *state) 2068{ 2069 arc_buf_contents_t type = arc_buf_type(hdr); 2070 uint64_t lsize = HDR_GET_LSIZE(hdr); 2071 2072 ASSERT(HDR_HAS_L1HDR(hdr)); 2073 2074 if (GHOST_STATE(state)) { 2075 ASSERT0(hdr->b_l1hdr.b_bufcnt); 2076 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 2077 ASSERT3P(hdr->b_l1hdr.b_pdata, ==, NULL); 2078 (void) refcount_add_many(&state->arcs_esize[type], lsize, hdr); 2079 return; 2080 } 2081 2082 ASSERT(!GHOST_STATE(state)); 2083 if (hdr->b_l1hdr.b_pdata != NULL) { 2084 (void) refcount_add_many(&state->arcs_esize[type], 2085 arc_hdr_size(hdr), hdr); 2086 } 2087 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL; 2088 buf = buf->b_next) { 2089 if (arc_buf_is_shared(buf)) { 2090 ASSERT(ARC_BUF_LAST(buf)); 2091 continue; 2092 } 2093 (void) refcount_add_many(&state->arcs_esize[type], lsize, buf); 2094 } 2095} 2096 2097/* 2098 * Decrement the amount of evictable space in the arc_state_t's refcount. 2099 * We account for the space used by the hdr and the arc buf individually 2100 * so that we can add and remove them from the refcount individually. 2101 */ 2102static void 2103arc_evitable_space_decrement(arc_buf_hdr_t *hdr, arc_state_t *state) 2104{ 2105 arc_buf_contents_t type = arc_buf_type(hdr); 2106 uint64_t lsize = HDR_GET_LSIZE(hdr); 2107 2108 ASSERT(HDR_HAS_L1HDR(hdr)); 2109 2110 if (GHOST_STATE(state)) { 2111 ASSERT0(hdr->b_l1hdr.b_bufcnt); 2112 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 2113 ASSERT3P(hdr->b_l1hdr.b_pdata, ==, NULL); 2114 (void) refcount_remove_many(&state->arcs_esize[type], 2115 lsize, hdr); 2116 return; 2117 } 2118 2119 ASSERT(!GHOST_STATE(state)); 2120 if (hdr->b_l1hdr.b_pdata != NULL) { 2121 (void) refcount_remove_many(&state->arcs_esize[type], 2122 arc_hdr_size(hdr), hdr); 2123 } 2124 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL; 2125 buf = buf->b_next) { 2126 if (arc_buf_is_shared(buf)) { 2127 ASSERT(ARC_BUF_LAST(buf)); 2128 continue; 2129 } 2130 (void) refcount_remove_many(&state->arcs_esize[type], 2131 lsize, buf); 2132 } 2133} 2134 2135/* 2136 * Add a reference to this hdr indicating that someone is actively 2137 * referencing that memory. When the refcount transitions from 0 to 1, 2138 * we remove it from the respective arc_state_t list to indicate that 2139 * it is not evictable. 2140 */ 2141static void 2142add_reference(arc_buf_hdr_t *hdr, void *tag) 2143{ 2144 ASSERT(HDR_HAS_L1HDR(hdr)); 2145 if (!MUTEX_HELD(HDR_LOCK(hdr))) { 2146 ASSERT(hdr->b_l1hdr.b_state == arc_anon); 2147 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 2148 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 2149 } 2150 2151 arc_state_t *state = hdr->b_l1hdr.b_state; 2152 2153 if ((refcount_add(&hdr->b_l1hdr.b_refcnt, tag) == 1) && 2154 (state != arc_anon)) { 2155 /* We don't use the L2-only state list. */ 2156 if (state != arc_l2c_only) { 2157 multilist_remove(&state->arcs_list[arc_buf_type(hdr)], 2158 hdr); 2159 arc_evitable_space_decrement(hdr, state); 2160 } 2161 /* remove the prefetch flag if we get a reference */ 2162 arc_hdr_clear_flags(hdr, ARC_FLAG_PREFETCH); 2163 } 2164} 2165 2166/* 2167 * Remove a reference from this hdr. When the reference transitions from 2168 * 1 to 0 and we're not anonymous, then we add this hdr to the arc_state_t's 2169 * list making it eligible for eviction. 2170 */ 2171static int 2172remove_reference(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, void *tag) 2173{ 2174 int cnt; 2175 arc_state_t *state = hdr->b_l1hdr.b_state; 2176 2177 ASSERT(HDR_HAS_L1HDR(hdr)); 2178 ASSERT(state == arc_anon || MUTEX_HELD(hash_lock)); 2179 ASSERT(!GHOST_STATE(state)); 2180 2181 /* 2182 * arc_l2c_only counts as a ghost state so we don't need to explicitly 2183 * check to prevent usage of the arc_l2c_only list. 2184 */ 2185 if (((cnt = refcount_remove(&hdr->b_l1hdr.b_refcnt, tag)) == 0) && 2186 (state != arc_anon)) { 2187 multilist_insert(&state->arcs_list[arc_buf_type(hdr)], hdr); 2188 ASSERT3U(hdr->b_l1hdr.b_bufcnt, >, 0); 2189 arc_evictable_space_increment(hdr, state); 2190 } 2191 return (cnt); 2192} 2193 2194/* 2195 * Move the supplied buffer to the indicated state. The hash lock 2196 * for the buffer must be held by the caller. 2197 */ 2198static void 2199arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *hdr, 2200 kmutex_t *hash_lock) 2201{ 2202 arc_state_t *old_state; 2203 int64_t refcnt; 2204 uint32_t bufcnt; 2205 boolean_t update_old, update_new; 2206 arc_buf_contents_t buftype = arc_buf_type(hdr); 2207 2208 /* 2209 * We almost always have an L1 hdr here, since we call arc_hdr_realloc() 2210 * in arc_read() when bringing a buffer out of the L2ARC. However, the 2211 * L1 hdr doesn't always exist when we change state to arc_anon before 2212 * destroying a header, in which case reallocating to add the L1 hdr is 2213 * pointless. 2214 */ 2215 if (HDR_HAS_L1HDR(hdr)) { 2216 old_state = hdr->b_l1hdr.b_state; 2217 refcnt = refcount_count(&hdr->b_l1hdr.b_refcnt); 2218 bufcnt = hdr->b_l1hdr.b_bufcnt; 2219 update_old = (bufcnt > 0 || hdr->b_l1hdr.b_pdata != NULL); 2220 } else { 2221 old_state = arc_l2c_only; 2222 refcnt = 0; 2223 bufcnt = 0; 2224 update_old = B_FALSE; 2225 } 2226 update_new = update_old; 2227 2228 ASSERT(MUTEX_HELD(hash_lock)); 2229 ASSERT3P(new_state, !=, old_state); 2230 ASSERT(!GHOST_STATE(new_state) || bufcnt == 0); 2231 ASSERT(old_state != arc_anon || bufcnt <= 1); 2232 2233 /* 2234 * If this buffer is evictable, transfer it from the 2235 * old state list to the new state list. 2236 */ 2237 if (refcnt == 0) { 2238 if (old_state != arc_anon && old_state != arc_l2c_only) { 2239 ASSERT(HDR_HAS_L1HDR(hdr)); 2240 multilist_remove(&old_state->arcs_list[buftype], hdr); 2241 2242 if (GHOST_STATE(old_state)) { 2243 ASSERT0(bufcnt); 2244 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 2245 update_old = B_TRUE; 2246 } 2247 arc_evitable_space_decrement(hdr, old_state); 2248 } 2249 if (new_state != arc_anon && new_state != arc_l2c_only) { 2250 2251 /* 2252 * An L1 header always exists here, since if we're 2253 * moving to some L1-cached state (i.e. not l2c_only or 2254 * anonymous), we realloc the header to add an L1hdr 2255 * beforehand. 2256 */ 2257 ASSERT(HDR_HAS_L1HDR(hdr)); 2258 multilist_insert(&new_state->arcs_list[buftype], hdr); 2259 2260 if (GHOST_STATE(new_state)) { 2261 ASSERT0(bufcnt); 2262 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 2263 update_new = B_TRUE; 2264 } 2265 arc_evictable_space_increment(hdr, new_state); 2266 } 2267 } 2268 2269 ASSERT(!HDR_EMPTY(hdr)); 2270 if (new_state == arc_anon && HDR_IN_HASH_TABLE(hdr)) 2271 buf_hash_remove(hdr); 2272 2273 /* adjust state sizes (ignore arc_l2c_only) */ 2274 2275 if (update_new && new_state != arc_l2c_only) { 2276 ASSERT(HDR_HAS_L1HDR(hdr)); 2277 if (GHOST_STATE(new_state)) { 2278 ASSERT0(bufcnt); 2279 2280 /* 2281 * When moving a header to a ghost state, we first 2282 * remove all arc buffers. Thus, we'll have a 2283 * bufcnt of zero, and no arc buffer to use for 2284 * the reference. As a result, we use the arc 2285 * header pointer for the reference. 2286 */ 2287 (void) refcount_add_many(&new_state->arcs_size, 2288 HDR_GET_LSIZE(hdr), hdr); 2289 ASSERT3P(hdr->b_l1hdr.b_pdata, ==, NULL); 2290 } else { 2291 uint32_t buffers = 0; 2292 2293 /* 2294 * Each individual buffer holds a unique reference, 2295 * thus we must remove each of these references one 2296 * at a time. 2297 */ 2298 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL; 2299 buf = buf->b_next) { 2300 ASSERT3U(bufcnt, !=, 0); 2301 buffers++; 2302 2303 /* 2304 * When the arc_buf_t is sharing the data 2305 * block with the hdr, the owner of the 2306 * reference belongs to the hdr. Only 2307 * add to the refcount if the arc_buf_t is 2308 * not shared. 2309 */ 2310 if (arc_buf_is_shared(buf)) { 2311 ASSERT(ARC_BUF_LAST(buf)); 2312 continue; 2313 } 2314 2315 (void) refcount_add_many(&new_state->arcs_size, 2316 HDR_GET_LSIZE(hdr), buf); 2317 } 2318 ASSERT3U(bufcnt, ==, buffers); 2319 2320 if (hdr->b_l1hdr.b_pdata != NULL) { 2321 (void) refcount_add_many(&new_state->arcs_size, 2322 arc_hdr_size(hdr), hdr); 2323 } else { 2324 ASSERT(GHOST_STATE(old_state)); 2325 } 2326 } 2327 } 2328 2329 if (update_old && old_state != arc_l2c_only) { 2330 ASSERT(HDR_HAS_L1HDR(hdr)); 2331 if (GHOST_STATE(old_state)) { 2332 ASSERT0(bufcnt); 2333 2334 /* 2335 * When moving a header off of a ghost state, 2336 * the header will not contain any arc buffers. 2337 * We use the arc header pointer for the reference 2338 * which is exactly what we did when we put the 2339 * header on the ghost state. 2340 */ 2341 2342 (void) refcount_remove_many(&old_state->arcs_size, 2343 HDR_GET_LSIZE(hdr), hdr); 2344 ASSERT3P(hdr->b_l1hdr.b_pdata, ==, NULL); 2345 } else { 2346 uint32_t buffers = 0; 2347 2348 /* 2349 * Each individual buffer holds a unique reference, 2350 * thus we must remove each of these references one 2351 * at a time. 2352 */ 2353 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL; 2354 buf = buf->b_next) { 2355 ASSERT3P(bufcnt, !=, 0); 2356 buffers++; 2357 2358 /* 2359 * When the arc_buf_t is sharing the data 2360 * block with the hdr, the owner of the 2361 * reference belongs to the hdr. Only 2362 * add to the refcount if the arc_buf_t is 2363 * not shared. 2364 */ 2365 if (arc_buf_is_shared(buf)) { 2366 ASSERT(ARC_BUF_LAST(buf)); 2367 continue; 2368 } 2369 2370 (void) refcount_remove_many( 2371 &old_state->arcs_size, HDR_GET_LSIZE(hdr), 2372 buf); 2373 } 2374 ASSERT3U(bufcnt, ==, buffers); 2375 ASSERT3P(hdr->b_l1hdr.b_pdata, !=, NULL); 2376 (void) refcount_remove_many( 2377 &old_state->arcs_size, arc_hdr_size(hdr), hdr); 2378 } 2379 } 2380 2381 if (HDR_HAS_L1HDR(hdr)) 2382 hdr->b_l1hdr.b_state = new_state; 2383 2384 /* 2385 * L2 headers should never be on the L2 state list since they don't 2386 * have L1 headers allocated. 2387 */ 2388 ASSERT(multilist_is_empty(&arc_l2c_only->arcs_list[ARC_BUFC_DATA]) && 2389 multilist_is_empty(&arc_l2c_only->arcs_list[ARC_BUFC_METADATA])); 2390} 2391 2392void 2393arc_space_consume(uint64_t space, arc_space_type_t type) 2394{ 2395 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES); 2396 2397 switch (type) { 2398 case ARC_SPACE_DATA: 2399 ARCSTAT_INCR(arcstat_data_size, space); 2400 break; 2401 case ARC_SPACE_META: 2402 ARCSTAT_INCR(arcstat_metadata_size, space); 2403 break; 2404 case ARC_SPACE_OTHER: 2405 ARCSTAT_INCR(arcstat_other_size, space); 2406 break; 2407 case ARC_SPACE_HDRS: 2408 ARCSTAT_INCR(arcstat_hdr_size, space); 2409 break; 2410 case ARC_SPACE_L2HDRS: 2411 ARCSTAT_INCR(arcstat_l2_hdr_size, space); 2412 break; 2413 } 2414 2415 if (type != ARC_SPACE_DATA) 2416 ARCSTAT_INCR(arcstat_meta_used, space); 2417 2418 atomic_add_64(&arc_size, space); 2419} 2420 2421void 2422arc_space_return(uint64_t space, arc_space_type_t type) 2423{ 2424 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES); 2425 2426 switch (type) { 2427 case ARC_SPACE_DATA: 2428 ARCSTAT_INCR(arcstat_data_size, -space); 2429 break; 2430 case ARC_SPACE_META: 2431 ARCSTAT_INCR(arcstat_metadata_size, -space); 2432 break; 2433 case ARC_SPACE_OTHER: 2434 ARCSTAT_INCR(arcstat_other_size, -space); 2435 break; 2436 case ARC_SPACE_HDRS: 2437 ARCSTAT_INCR(arcstat_hdr_size, -space); 2438 break; 2439 case ARC_SPACE_L2HDRS: 2440 ARCSTAT_INCR(arcstat_l2_hdr_size, -space); 2441 break; 2442 } 2443 2444 if (type != ARC_SPACE_DATA) { 2445 ASSERT(arc_meta_used >= space); 2446 if (arc_meta_max < arc_meta_used) 2447 arc_meta_max = arc_meta_used; 2448 ARCSTAT_INCR(arcstat_meta_used, -space); 2449 } 2450 2451 ASSERT(arc_size >= space); 2452 atomic_add_64(&arc_size, -space); 2453} 2454 2455/* 2456 * Allocate an initial buffer for this hdr, subsequent buffers will 2457 * use arc_buf_clone(). 2458 */ 2459static arc_buf_t * 2460arc_buf_alloc_impl(arc_buf_hdr_t *hdr, void *tag) 2461{ 2462 arc_buf_t *buf; 2463 2464 ASSERT(HDR_HAS_L1HDR(hdr)); 2465 ASSERT3U(HDR_GET_LSIZE(hdr), >, 0); 2466 VERIFY(hdr->b_type == ARC_BUFC_DATA || 2467 hdr->b_type == ARC_BUFC_METADATA); 2468 2469 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 2470 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 2471 ASSERT0(hdr->b_l1hdr.b_bufcnt); 2472 2473 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 2474 buf->b_hdr = hdr; 2475 buf->b_data = NULL; 2476 buf->b_next = NULL; 2477 2478 add_reference(hdr, tag); 2479 2480 /* 2481 * We're about to change the hdr's b_flags. We must either 2482 * hold the hash_lock or be undiscoverable. 2483 */ 2484 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); 2485 2486 /* 2487 * If the hdr's data can be shared (no byteswapping, hdr is 2488 * uncompressed, hdr's data is not currently being written to the 2489 * L2ARC write) then we share the data buffer and set the appropriate 2490 * bit in the hdr's b_flags to indicate the hdr is sharing it's 2491 * b_pdata with the arc_buf_t. Otherwise, we allocate a new buffer to 2492 * store the buf's data. 2493 */ 2494 if (hdr->b_l1hdr.b_byteswap == DMU_BSWAP_NUMFUNCS && 2495 HDR_GET_COMPRESS(hdr) == ZIO_COMPRESS_OFF && !HDR_L2_WRITING(hdr)) { 2496 buf->b_data = hdr->b_l1hdr.b_pdata; 2497 arc_hdr_set_flags(hdr, ARC_FLAG_SHARED_DATA); 2498 } else { 2499 buf->b_data = arc_get_data_buf(hdr, HDR_GET_LSIZE(hdr), buf); 2500 ARCSTAT_INCR(arcstat_overhead_size, HDR_GET_LSIZE(hdr)); 2501 arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA); 2502 } 2503 VERIFY3P(buf->b_data, !=, NULL); 2504 2505 hdr->b_l1hdr.b_buf = buf; 2506 hdr->b_l1hdr.b_bufcnt += 1; 2507 2508 return (buf); 2509} 2510 2511/* 2512 * Used when allocating additional buffers. 2513 */ 2514static arc_buf_t * 2515arc_buf_clone(arc_buf_t *from) 2516{ 2517 arc_buf_t *buf; 2518 arc_buf_hdr_t *hdr = from->b_hdr; 2519 uint64_t size = HDR_GET_LSIZE(hdr); 2520 2521 ASSERT(HDR_HAS_L1HDR(hdr)); 2522 ASSERT(hdr->b_l1hdr.b_state != arc_anon); 2523 2524 buf = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 2525 buf->b_hdr = hdr; 2526 buf->b_data = NULL; 2527 buf->b_next = hdr->b_l1hdr.b_buf; 2528 hdr->b_l1hdr.b_buf = buf; 2529 buf->b_data = arc_get_data_buf(hdr, HDR_GET_LSIZE(hdr), buf); 2530 bcopy(from->b_data, buf->b_data, size); 2531 hdr->b_l1hdr.b_bufcnt += 1; 2532 2533 ARCSTAT_INCR(arcstat_overhead_size, HDR_GET_LSIZE(hdr)); 2534 return (buf); 2535} 2536 2537static char *arc_onloan_tag = "onloan"; 2538 2539/* 2540 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in 2541 * flight data by arc_tempreserve_space() until they are "returned". Loaned 2542 * buffers must be returned to the arc before they can be used by the DMU or 2543 * freed. 2544 */ 2545arc_buf_t * 2546arc_loan_buf(spa_t *spa, int size) 2547{ 2548 arc_buf_t *buf; 2549 2550 buf = arc_alloc_buf(spa, size, arc_onloan_tag, ARC_BUFC_DATA); 2551 2552 atomic_add_64(&arc_loaned_bytes, size); 2553 return (buf); 2554} 2555 2556/* 2557 * Return a loaned arc buffer to the arc. 2558 */ 2559void 2560arc_return_buf(arc_buf_t *buf, void *tag) 2561{ 2562 arc_buf_hdr_t *hdr = buf->b_hdr; 2563 2564 ASSERT3P(buf->b_data, !=, NULL); 2565 ASSERT(HDR_HAS_L1HDR(hdr)); 2566 (void) refcount_add(&hdr->b_l1hdr.b_refcnt, tag); 2567 (void) refcount_remove(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag); 2568 2569 atomic_add_64(&arc_loaned_bytes, -HDR_GET_LSIZE(hdr)); 2570} 2571 2572/* Detach an arc_buf from a dbuf (tag) */ 2573void 2574arc_loan_inuse_buf(arc_buf_t *buf, void *tag) 2575{ 2576 arc_buf_hdr_t *hdr = buf->b_hdr; 2577 2578 ASSERT3P(buf->b_data, !=, NULL); 2579 ASSERT(HDR_HAS_L1HDR(hdr)); 2580 (void) refcount_add(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag); 2581 (void) refcount_remove(&hdr->b_l1hdr.b_refcnt, tag); 2582 2583 atomic_add_64(&arc_loaned_bytes, HDR_GET_LSIZE(hdr)); 2584} 2585 2586static void 2587l2arc_free_data_on_write(void *data, size_t size, arc_buf_contents_t type) 2588{ 2589 l2arc_data_free_t *df = kmem_alloc(sizeof (*df), KM_SLEEP); 2590 2591 df->l2df_data = data; 2592 df->l2df_size = size; 2593 df->l2df_type = type; 2594 mutex_enter(&l2arc_free_on_write_mtx); 2595 list_insert_head(l2arc_free_on_write, df); 2596 mutex_exit(&l2arc_free_on_write_mtx); 2597} 2598 2599static void 2600arc_hdr_free_on_write(arc_buf_hdr_t *hdr) 2601{ 2602 arc_state_t *state = hdr->b_l1hdr.b_state; 2603 arc_buf_contents_t type = arc_buf_type(hdr); 2604 uint64_t size = arc_hdr_size(hdr); 2605 2606 /* protected by hash lock, if in the hash table */ 2607 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) { 2608 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 2609 ASSERT(state != arc_anon && state != arc_l2c_only); 2610 2611 (void) refcount_remove_many(&state->arcs_esize[type], 2612 size, hdr); 2613 } 2614 (void) refcount_remove_many(&state->arcs_size, size, hdr); 2615 if (type == ARC_BUFC_METADATA) { 2616 arc_space_return(size, ARC_SPACE_META); 2617 } else { 2618 ASSERT(type == ARC_BUFC_DATA); 2619 arc_space_return(size, ARC_SPACE_DATA); 2620 } 2621 2622 l2arc_free_data_on_write(hdr->b_l1hdr.b_pdata, size, type); 2623} 2624 2625/* 2626 * Share the arc_buf_t's data with the hdr. Whenever we are sharing the 2627 * data buffer, we transfer the refcount ownership to the hdr and update 2628 * the appropriate kstats. 2629 */ 2630static void 2631arc_share_buf(arc_buf_hdr_t *hdr, arc_buf_t *buf) 2632{ 2633 arc_state_t *state = hdr->b_l1hdr.b_state; 2634 2635 ASSERT(!HDR_SHARED_DATA(hdr)); 2636 ASSERT(!arc_buf_is_shared(buf)); 2637 ASSERT3P(hdr->b_l1hdr.b_pdata, ==, NULL); 2638 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); 2639 2640 /* 2641 * Start sharing the data buffer. We transfer the 2642 * refcount ownership to the hdr since it always owns 2643 * the refcount whenever an arc_buf_t is shared. 2644 */ 2645 refcount_transfer_ownership(&state->arcs_size, buf, hdr); 2646 hdr->b_l1hdr.b_pdata = buf->b_data; 2647 arc_hdr_set_flags(hdr, ARC_FLAG_SHARED_DATA); 2648 2649 /* 2650 * Since we've transferred ownership to the hdr we need 2651 * to increment its compressed and uncompressed kstats and 2652 * decrement the overhead size. 2653 */ 2654 ARCSTAT_INCR(arcstat_compressed_size, arc_hdr_size(hdr)); 2655 ARCSTAT_INCR(arcstat_uncompressed_size, HDR_GET_LSIZE(hdr)); 2656 ARCSTAT_INCR(arcstat_overhead_size, -HDR_GET_LSIZE(hdr)); 2657} 2658 2659static void 2660arc_unshare_buf(arc_buf_hdr_t *hdr, arc_buf_t *buf) 2661{ 2662 arc_state_t *state = hdr->b_l1hdr.b_state; 2663 2664 ASSERT(HDR_SHARED_DATA(hdr)); 2665 ASSERT(arc_buf_is_shared(buf)); 2666 ASSERT3P(hdr->b_l1hdr.b_pdata, !=, NULL); 2667 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); 2668 2669 /* 2670 * We are no longer sharing this buffer so we need 2671 * to transfer its ownership to the rightful owner. 2672 */ 2673 refcount_transfer_ownership(&state->arcs_size, hdr, buf); 2674 arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA); 2675 hdr->b_l1hdr.b_pdata = NULL; 2676 2677 /* 2678 * Since the buffer is no longer shared between 2679 * the arc buf and the hdr, count it as overhead. 2680 */ 2681 ARCSTAT_INCR(arcstat_compressed_size, -arc_hdr_size(hdr)); 2682 ARCSTAT_INCR(arcstat_uncompressed_size, -HDR_GET_LSIZE(hdr)); 2683 ARCSTAT_INCR(arcstat_overhead_size, HDR_GET_LSIZE(hdr)); 2684} 2685 2686/* 2687 * Free up buf->b_data and if 'remove' is set, then pull the 2688 * arc_buf_t off of the the arc_buf_hdr_t's list and free it. 2689 */ 2690static void 2691arc_buf_destroy_impl(arc_buf_t *buf, boolean_t remove) 2692{ 2693 arc_buf_t **bufp; 2694 arc_buf_hdr_t *hdr = buf->b_hdr; 2695 uint64_t size = HDR_GET_LSIZE(hdr); 2696 boolean_t destroyed_buf_is_shared = arc_buf_is_shared(buf); 2697 2698 /* 2699 * Free up the data associated with the buf but only 2700 * if we're not sharing this with the hdr. If we are sharing 2701 * it with the hdr, then hdr will have performed the allocation 2702 * so allow it to do the free. 2703 */ 2704 if (buf->b_data != NULL) { 2705 /* 2706 * We're about to change the hdr's b_flags. We must either 2707 * hold the hash_lock or be undiscoverable. 2708 */ 2709 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); 2710 2711 arc_cksum_verify(buf); 2712#ifdef illumos 2713 arc_buf_unwatch(buf); 2714#endif 2715 2716 if (destroyed_buf_is_shared) { 2717 ASSERT(ARC_BUF_LAST(buf)); 2718 ASSERT(HDR_SHARED_DATA(hdr)); 2719 arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA); 2720 } else { 2721 arc_free_data_buf(hdr, buf->b_data, size, buf); 2722 ARCSTAT_INCR(arcstat_overhead_size, -size); 2723 } 2724 buf->b_data = NULL; 2725 2726 ASSERT(hdr->b_l1hdr.b_bufcnt > 0); 2727 hdr->b_l1hdr.b_bufcnt -= 1; 2728 } 2729 2730 /* only remove the buf if requested */ 2731 if (!remove) 2732 return; 2733 2734 /* remove the buf from the hdr list */ 2735 arc_buf_t *lastbuf = NULL; 2736 bufp = &hdr->b_l1hdr.b_buf; 2737 while (*bufp != NULL) { 2738 if (*bufp == buf) 2739 *bufp = buf->b_next; 2740 2741 /* 2742 * If we've removed a buffer in the middle of 2743 * the list then update the lastbuf and update 2744 * bufp. 2745 */ 2746 if (*bufp != NULL) { 2747 lastbuf = *bufp; 2748 bufp = &(*bufp)->b_next; 2749 } 2750 } 2751 buf->b_next = NULL; 2752 ASSERT3P(lastbuf, !=, buf); 2753 2754 /* 2755 * If the current arc_buf_t is sharing its data 2756 * buffer with the hdr, then reassign the hdr's 2757 * b_pdata to share it with the new buffer at the end 2758 * of the list. The shared buffer is always the last one 2759 * on the hdr's buffer list. 2760 */ 2761 if (destroyed_buf_is_shared && lastbuf != NULL) { 2762 ASSERT(ARC_BUF_LAST(buf)); 2763 ASSERT(ARC_BUF_LAST(lastbuf)); 2764 VERIFY(!arc_buf_is_shared(lastbuf)); 2765 2766 ASSERT3P(hdr->b_l1hdr.b_pdata, !=, NULL); 2767 arc_hdr_free_pdata(hdr); 2768 2769 /* 2770 * We must setup a new shared block between the 2771 * last buffer and the hdr. The data would have 2772 * been allocated by the arc buf so we need to transfer 2773 * ownership to the hdr since it's now being shared. 2774 */ 2775 arc_share_buf(hdr, lastbuf); 2776 } else if (HDR_SHARED_DATA(hdr)) { 2777 ASSERT(arc_buf_is_shared(lastbuf)); 2778 } 2779 2780 if (hdr->b_l1hdr.b_bufcnt == 0) 2781 arc_cksum_free(hdr); 2782 2783 /* clean up the buf */ 2784 buf->b_hdr = NULL; 2785 kmem_cache_free(buf_cache, buf); 2786} 2787 2788static void 2789arc_hdr_alloc_pdata(arc_buf_hdr_t *hdr) 2790{ 2791 ASSERT3U(HDR_GET_LSIZE(hdr), >, 0); 2792 ASSERT(HDR_HAS_L1HDR(hdr)); 2793 ASSERT(!HDR_SHARED_DATA(hdr)); 2794 2795 ASSERT3P(hdr->b_l1hdr.b_pdata, ==, NULL); 2796 hdr->b_l1hdr.b_pdata = arc_get_data_buf(hdr, arc_hdr_size(hdr), hdr); 2797 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS; 2798 ASSERT3P(hdr->b_l1hdr.b_pdata, !=, NULL); 2799 2800 ARCSTAT_INCR(arcstat_compressed_size, arc_hdr_size(hdr)); 2801 ARCSTAT_INCR(arcstat_uncompressed_size, HDR_GET_LSIZE(hdr)); 2802} 2803 2804static void 2805arc_hdr_free_pdata(arc_buf_hdr_t *hdr) 2806{ 2807 ASSERT(HDR_HAS_L1HDR(hdr)); 2808 ASSERT3P(hdr->b_l1hdr.b_pdata, !=, NULL); 2809 2810 /* 2811 * If the hdr is currently being written to the l2arc then 2812 * we defer freeing the data by adding it to the l2arc_free_on_write 2813 * list. The l2arc will free the data once it's finished 2814 * writing it to the l2arc device. 2815 */ 2816 if (HDR_L2_WRITING(hdr)) { 2817 arc_hdr_free_on_write(hdr); 2818 ARCSTAT_BUMP(arcstat_l2_free_on_write); 2819 } else { 2820 arc_free_data_buf(hdr, hdr->b_l1hdr.b_pdata, 2821 arc_hdr_size(hdr), hdr); 2822 } 2823 hdr->b_l1hdr.b_pdata = NULL; 2824 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS; 2825 2826 ARCSTAT_INCR(arcstat_compressed_size, -arc_hdr_size(hdr)); 2827 ARCSTAT_INCR(arcstat_uncompressed_size, -HDR_GET_LSIZE(hdr)); 2828} 2829 2830static arc_buf_hdr_t * 2831arc_hdr_alloc(uint64_t spa, int32_t psize, int32_t lsize, 2832 enum zio_compress compress, arc_buf_contents_t type) 2833{ 2834 arc_buf_hdr_t *hdr; 2835 2836 ASSERT3U(lsize, >, 0); 2837 VERIFY(type == ARC_BUFC_DATA || type == ARC_BUFC_METADATA); 2838 2839 hdr = kmem_cache_alloc(hdr_full_cache, KM_PUSHPAGE); 2840 ASSERT(HDR_EMPTY(hdr)); 2841 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL); 2842 ASSERT3P(hdr->b_l1hdr.b_thawed, ==, NULL); 2843 HDR_SET_PSIZE(hdr, psize); 2844 HDR_SET_LSIZE(hdr, lsize); 2845 hdr->b_spa = spa; 2846 hdr->b_type = type; 2847 hdr->b_flags = 0; 2848 arc_hdr_set_flags(hdr, arc_bufc_to_flags(type) | ARC_FLAG_HAS_L1HDR); 2849 arc_hdr_set_compress(hdr, compress); 2850 2851 hdr->b_l1hdr.b_state = arc_anon; 2852 hdr->b_l1hdr.b_arc_access = 0; 2853 hdr->b_l1hdr.b_bufcnt = 0; 2854 hdr->b_l1hdr.b_buf = NULL; 2855 2856 /* 2857 * Allocate the hdr's buffer. This will contain either 2858 * the compressed or uncompressed data depending on the block 2859 * it references and compressed arc enablement. 2860 */ 2861 arc_hdr_alloc_pdata(hdr); 2862 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 2863 2864 return (hdr); 2865} 2866 2867/* 2868 * Transition between the two allocation states for the arc_buf_hdr struct. 2869 * The arc_buf_hdr struct can be allocated with (hdr_full_cache) or without 2870 * (hdr_l2only_cache) the fields necessary for the L1 cache - the smaller 2871 * version is used when a cache buffer is only in the L2ARC in order to reduce 2872 * memory usage. 2873 */ 2874static arc_buf_hdr_t * 2875arc_hdr_realloc(arc_buf_hdr_t *hdr, kmem_cache_t *old, kmem_cache_t *new) 2876{ 2877 ASSERT(HDR_HAS_L2HDR(hdr)); 2878 2879 arc_buf_hdr_t *nhdr; 2880 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev; 2881 2882 ASSERT((old == hdr_full_cache && new == hdr_l2only_cache) || 2883 (old == hdr_l2only_cache && new == hdr_full_cache)); 2884 2885 nhdr = kmem_cache_alloc(new, KM_PUSHPAGE); 2886 2887 ASSERT(MUTEX_HELD(HDR_LOCK(hdr))); 2888 buf_hash_remove(hdr); 2889 2890 bcopy(hdr, nhdr, HDR_L2ONLY_SIZE); 2891 2892 if (new == hdr_full_cache) { 2893 arc_hdr_set_flags(nhdr, ARC_FLAG_HAS_L1HDR); 2894 /* 2895 * arc_access and arc_change_state need to be aware that a 2896 * header has just come out of L2ARC, so we set its state to 2897 * l2c_only even though it's about to change. 2898 */ 2899 nhdr->b_l1hdr.b_state = arc_l2c_only; 2900 2901 /* Verify previous threads set to NULL before freeing */ 2902 ASSERT3P(nhdr->b_l1hdr.b_pdata, ==, NULL); 2903 } else { 2904 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 2905 ASSERT0(hdr->b_l1hdr.b_bufcnt); 2906 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL); 2907 2908 /* 2909 * If we've reached here, We must have been called from 2910 * arc_evict_hdr(), as such we should have already been 2911 * removed from any ghost list we were previously on 2912 * (which protects us from racing with arc_evict_state), 2913 * thus no locking is needed during this check. 2914 */ 2915 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 2916 2917 /* 2918 * A buffer must not be moved into the arc_l2c_only 2919 * state if it's not finished being written out to the 2920 * l2arc device. Otherwise, the b_l1hdr.b_pdata field 2921 * might try to be accessed, even though it was removed. 2922 */ 2923 VERIFY(!HDR_L2_WRITING(hdr)); 2924 VERIFY3P(hdr->b_l1hdr.b_pdata, ==, NULL); 2925 2926#ifdef ZFS_DEBUG 2927 if (hdr->b_l1hdr.b_thawed != NULL) { 2928 kmem_free(hdr->b_l1hdr.b_thawed, 1); 2929 hdr->b_l1hdr.b_thawed = NULL; 2930 } 2931#endif 2932 2933 arc_hdr_clear_flags(nhdr, ARC_FLAG_HAS_L1HDR); 2934 } 2935 /* 2936 * The header has been reallocated so we need to re-insert it into any 2937 * lists it was on. 2938 */ 2939 (void) buf_hash_insert(nhdr, NULL); 2940 2941 ASSERT(list_link_active(&hdr->b_l2hdr.b_l2node)); 2942 2943 mutex_enter(&dev->l2ad_mtx); 2944 2945 /* 2946 * We must place the realloc'ed header back into the list at 2947 * the same spot. Otherwise, if it's placed earlier in the list, 2948 * l2arc_write_buffers() could find it during the function's 2949 * write phase, and try to write it out to the l2arc. 2950 */ 2951 list_insert_after(&dev->l2ad_buflist, hdr, nhdr); 2952 list_remove(&dev->l2ad_buflist, hdr); 2953 2954 mutex_exit(&dev->l2ad_mtx); 2955 2956 /* 2957 * Since we're using the pointer address as the tag when 2958 * incrementing and decrementing the l2ad_alloc refcount, we 2959 * must remove the old pointer (that we're about to destroy) and 2960 * add the new pointer to the refcount. Otherwise we'd remove 2961 * the wrong pointer address when calling arc_hdr_destroy() later. 2962 */ 2963 2964 (void) refcount_remove_many(&dev->l2ad_alloc, arc_hdr_size(hdr), hdr); 2965 (void) refcount_add_many(&dev->l2ad_alloc, arc_hdr_size(nhdr), nhdr); 2966 2967 buf_discard_identity(hdr); 2968 kmem_cache_free(old, hdr); 2969 2970 return (nhdr); 2971} 2972 2973/* 2974 * Allocate a new arc_buf_hdr_t and arc_buf_t and return the buf to the caller. 2975 * The buf is returned thawed since we expect the consumer to modify it. 2976 */ 2977arc_buf_t * 2978arc_alloc_buf(spa_t *spa, int32_t size, void *tag, arc_buf_contents_t type) 2979{ 2980 arc_buf_hdr_t *hdr = arc_hdr_alloc(spa_load_guid(spa), size, size, 2981 ZIO_COMPRESS_OFF, type); 2982 ASSERT(!MUTEX_HELD(HDR_LOCK(hdr))); 2983 arc_buf_t *buf = arc_buf_alloc_impl(hdr, tag); 2984 arc_buf_thaw(buf); 2985 return (buf); 2986} 2987 2988static void 2989arc_hdr_l2hdr_destroy(arc_buf_hdr_t *hdr) 2990{ 2991 l2arc_buf_hdr_t *l2hdr = &hdr->b_l2hdr; 2992 l2arc_dev_t *dev = l2hdr->b_dev; 2993 uint64_t asize = arc_hdr_size(hdr); 2994 2995 ASSERT(MUTEX_HELD(&dev->l2ad_mtx)); 2996 ASSERT(HDR_HAS_L2HDR(hdr)); 2997 2998 list_remove(&dev->l2ad_buflist, hdr); 2999 3000 ARCSTAT_INCR(arcstat_l2_asize, -asize); 3001 ARCSTAT_INCR(arcstat_l2_size, -HDR_GET_LSIZE(hdr)); 3002 3003 vdev_space_update(dev->l2ad_vdev, -asize, 0, 0); 3004 3005 (void) refcount_remove_many(&dev->l2ad_alloc, asize, hdr); 3006 arc_hdr_clear_flags(hdr, ARC_FLAG_HAS_L2HDR); 3007} 3008 3009static void 3010arc_hdr_destroy(arc_buf_hdr_t *hdr) 3011{ 3012 if (HDR_HAS_L1HDR(hdr)) { 3013 ASSERT(hdr->b_l1hdr.b_buf == NULL || 3014 hdr->b_l1hdr.b_bufcnt > 0); 3015 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 3016 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon); 3017 } 3018 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 3019 ASSERT(!HDR_IN_HASH_TABLE(hdr)); 3020 3021 if (!HDR_EMPTY(hdr)) 3022 buf_discard_identity(hdr); 3023 3024 if (HDR_HAS_L2HDR(hdr)) { 3025 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev; 3026 boolean_t buflist_held = MUTEX_HELD(&dev->l2ad_mtx); 3027 3028 if (!buflist_held) 3029 mutex_enter(&dev->l2ad_mtx); 3030 3031 /* 3032 * Even though we checked this conditional above, we 3033 * need to check this again now that we have the 3034 * l2ad_mtx. This is because we could be racing with 3035 * another thread calling l2arc_evict() which might have 3036 * destroyed this header's L2 portion as we were waiting 3037 * to acquire the l2ad_mtx. If that happens, we don't 3038 * want to re-destroy the header's L2 portion. 3039 */ 3040 if (HDR_HAS_L2HDR(hdr)) { 3041 l2arc_trim(hdr); 3042 arc_hdr_l2hdr_destroy(hdr); 3043 } 3044 3045 if (!buflist_held) 3046 mutex_exit(&dev->l2ad_mtx); 3047 } 3048 3049 if (HDR_HAS_L1HDR(hdr)) { 3050 arc_cksum_free(hdr); 3051 3052 while (hdr->b_l1hdr.b_buf != NULL) 3053 arc_buf_destroy_impl(hdr->b_l1hdr.b_buf, B_TRUE); 3054 3055#ifdef ZFS_DEBUG 3056 if (hdr->b_l1hdr.b_thawed != NULL) { 3057 kmem_free(hdr->b_l1hdr.b_thawed, 1); 3058 hdr->b_l1hdr.b_thawed = NULL; 3059 } 3060#endif 3061 3062 if (hdr->b_l1hdr.b_pdata != NULL) { 3063 arc_hdr_free_pdata(hdr); 3064 } 3065 } 3066 3067 ASSERT3P(hdr->b_hash_next, ==, NULL); 3068 if (HDR_HAS_L1HDR(hdr)) { 3069 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 3070 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL); 3071 kmem_cache_free(hdr_full_cache, hdr); 3072 } else { 3073 kmem_cache_free(hdr_l2only_cache, hdr); 3074 } 3075} 3076 3077void 3078arc_buf_destroy(arc_buf_t *buf, void* tag) 3079{ 3080 arc_buf_hdr_t *hdr = buf->b_hdr; 3081 kmutex_t *hash_lock = HDR_LOCK(hdr); 3082 3083 if (hdr->b_l1hdr.b_state == arc_anon) { 3084 ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1); 3085 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 3086 VERIFY0(remove_reference(hdr, NULL, tag)); 3087 arc_hdr_destroy(hdr); 3088 return; 3089 } 3090 3091 mutex_enter(hash_lock); 3092 ASSERT3P(hdr, ==, buf->b_hdr); 3093 ASSERT(hdr->b_l1hdr.b_bufcnt > 0); 3094 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 3095 ASSERT3P(hdr->b_l1hdr.b_state, !=, arc_anon); 3096 ASSERT3P(buf->b_data, !=, NULL); 3097 3098 (void) remove_reference(hdr, hash_lock, tag); 3099 arc_buf_destroy_impl(buf, B_TRUE); 3100 mutex_exit(hash_lock); 3101} 3102 3103int32_t 3104arc_buf_size(arc_buf_t *buf) 3105{ 3106 return (HDR_GET_LSIZE(buf->b_hdr)); 3107} 3108 3109/* 3110 * Evict the arc_buf_hdr that is provided as a parameter. The resultant 3111 * state of the header is dependent on it's state prior to entering this 3112 * function. The following transitions are possible: 3113 * 3114 * - arc_mru -> arc_mru_ghost 3115 * - arc_mfu -> arc_mfu_ghost 3116 * - arc_mru_ghost -> arc_l2c_only 3117 * - arc_mru_ghost -> deleted 3118 * - arc_mfu_ghost -> arc_l2c_only 3119 * - arc_mfu_ghost -> deleted 3120 */ 3121static int64_t 3122arc_evict_hdr(arc_buf_hdr_t *hdr, kmutex_t *hash_lock) 3123{ 3124 arc_state_t *evicted_state, *state; 3125 int64_t bytes_evicted = 0; 3126 3127 ASSERT(MUTEX_HELD(hash_lock)); 3128 ASSERT(HDR_HAS_L1HDR(hdr)); 3129 3130 state = hdr->b_l1hdr.b_state; 3131 if (GHOST_STATE(state)) { 3132 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 3133 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 3134 3135 /* 3136 * l2arc_write_buffers() relies on a header's L1 portion 3137 * (i.e. its b_pdata field) during its write phase. 3138 * Thus, we cannot push a header onto the arc_l2c_only 3139 * state (removing it's L1 piece) until the header is 3140 * done being written to the l2arc. 3141 */ 3142 if (HDR_HAS_L2HDR(hdr) && HDR_L2_WRITING(hdr)) { 3143 ARCSTAT_BUMP(arcstat_evict_l2_skip); 3144 return (bytes_evicted); 3145 } 3146 3147 ARCSTAT_BUMP(arcstat_deleted); 3148 bytes_evicted += HDR_GET_LSIZE(hdr); 3149 3150 DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, hdr); 3151 3152 ASSERT3P(hdr->b_l1hdr.b_pdata, ==, NULL); 3153 if (HDR_HAS_L2HDR(hdr)) { 3154 ASSERT(hdr->b_l1hdr.b_pdata == NULL); 3155 /* 3156 * This buffer is cached on the 2nd Level ARC; 3157 * don't destroy the header. 3158 */ 3159 arc_change_state(arc_l2c_only, hdr, hash_lock); 3160 /* 3161 * dropping from L1+L2 cached to L2-only, 3162 * realloc to remove the L1 header. 3163 */ 3164 hdr = arc_hdr_realloc(hdr, hdr_full_cache, 3165 hdr_l2only_cache); 3166 } else { 3167 ASSERT(hdr->b_l1hdr.b_pdata == NULL); 3168 arc_change_state(arc_anon, hdr, hash_lock); 3169 arc_hdr_destroy(hdr); 3170 } 3171 return (bytes_evicted); 3172 } 3173 3174 ASSERT(state == arc_mru || state == arc_mfu); 3175 evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost; 3176 3177 /* prefetch buffers have a minimum lifespan */ 3178 if (HDR_IO_IN_PROGRESS(hdr) || 3179 ((hdr->b_flags & (ARC_FLAG_PREFETCH | ARC_FLAG_INDIRECT)) && 3180 ddi_get_lbolt() - hdr->b_l1hdr.b_arc_access < 3181 arc_min_prefetch_lifespan)) { 3182 ARCSTAT_BUMP(arcstat_evict_skip); 3183 return (bytes_evicted); 3184 } 3185 3186 ASSERT0(refcount_count(&hdr->b_l1hdr.b_refcnt)); 3187 while (hdr->b_l1hdr.b_buf) { 3188 arc_buf_t *buf = hdr->b_l1hdr.b_buf; 3189 if (!mutex_tryenter(&buf->b_evict_lock)) { 3190 ARCSTAT_BUMP(arcstat_mutex_miss); 3191 break; 3192 } 3193 if (buf->b_data != NULL) 3194 bytes_evicted += HDR_GET_LSIZE(hdr); 3195 mutex_exit(&buf->b_evict_lock); 3196 arc_buf_destroy_impl(buf, B_TRUE); 3197 } 3198 3199 if (HDR_HAS_L2HDR(hdr)) { 3200 ARCSTAT_INCR(arcstat_evict_l2_cached, HDR_GET_LSIZE(hdr)); 3201 } else { 3202 if (l2arc_write_eligible(hdr->b_spa, hdr)) { 3203 ARCSTAT_INCR(arcstat_evict_l2_eligible, 3204 HDR_GET_LSIZE(hdr)); 3205 } else { 3206 ARCSTAT_INCR(arcstat_evict_l2_ineligible, 3207 HDR_GET_LSIZE(hdr)); 3208 } 3209 } 3210 3211 if (hdr->b_l1hdr.b_bufcnt == 0) { 3212 arc_cksum_free(hdr); 3213 3214 bytes_evicted += arc_hdr_size(hdr); 3215 3216 /* 3217 * If this hdr is being evicted and has a compressed 3218 * buffer then we discard it here before we change states. 3219 * This ensures that the accounting is updated correctly 3220 * in arc_free_data_buf(). 3221 */ 3222 arc_hdr_free_pdata(hdr); 3223 3224 arc_change_state(evicted_state, hdr, hash_lock); 3225 ASSERT(HDR_IN_HASH_TABLE(hdr)); 3226 arc_hdr_set_flags(hdr, ARC_FLAG_IN_HASH_TABLE); 3227 DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, hdr); 3228 } 3229 3230 return (bytes_evicted); 3231} 3232 3233static uint64_t 3234arc_evict_state_impl(multilist_t *ml, int idx, arc_buf_hdr_t *marker, 3235 uint64_t spa, int64_t bytes) 3236{ 3237 multilist_sublist_t *mls; 3238 uint64_t bytes_evicted = 0; 3239 arc_buf_hdr_t *hdr; 3240 kmutex_t *hash_lock; 3241 int evict_count = 0; 3242 3243 ASSERT3P(marker, !=, NULL); 3244 IMPLY(bytes < 0, bytes == ARC_EVICT_ALL); 3245 3246 mls = multilist_sublist_lock(ml, idx); 3247 3248 for (hdr = multilist_sublist_prev(mls, marker); hdr != NULL; 3249 hdr = multilist_sublist_prev(mls, marker)) { 3250 if ((bytes != ARC_EVICT_ALL && bytes_evicted >= bytes) || 3251 (evict_count >= zfs_arc_evict_batch_limit)) 3252 break; 3253 3254 /* 3255 * To keep our iteration location, move the marker 3256 * forward. Since we're not holding hdr's hash lock, we 3257 * must be very careful and not remove 'hdr' from the 3258 * sublist. Otherwise, other consumers might mistake the 3259 * 'hdr' as not being on a sublist when they call the 3260 * multilist_link_active() function (they all rely on 3261 * the hash lock protecting concurrent insertions and 3262 * removals). multilist_sublist_move_forward() was 3263 * specifically implemented to ensure this is the case 3264 * (only 'marker' will be removed and re-inserted). 3265 */ 3266 multilist_sublist_move_forward(mls, marker); 3267 3268 /* 3269 * The only case where the b_spa field should ever be 3270 * zero, is the marker headers inserted by 3271 * arc_evict_state(). It's possible for multiple threads 3272 * to be calling arc_evict_state() concurrently (e.g. 3273 * dsl_pool_close() and zio_inject_fault()), so we must 3274 * skip any markers we see from these other threads. 3275 */ 3276 if (hdr->b_spa == 0) 3277 continue; 3278 3279 /* we're only interested in evicting buffers of a certain spa */ 3280 if (spa != 0 && hdr->b_spa != spa) { 3281 ARCSTAT_BUMP(arcstat_evict_skip); 3282 continue; 3283 } 3284 3285 hash_lock = HDR_LOCK(hdr); 3286 3287 /* 3288 * We aren't calling this function from any code path 3289 * that would already be holding a hash lock, so we're 3290 * asserting on this assumption to be defensive in case 3291 * this ever changes. Without this check, it would be 3292 * possible to incorrectly increment arcstat_mutex_miss 3293 * below (e.g. if the code changed such that we called 3294 * this function with a hash lock held). 3295 */ 3296 ASSERT(!MUTEX_HELD(hash_lock)); 3297 3298 if (mutex_tryenter(hash_lock)) { 3299 uint64_t evicted = arc_evict_hdr(hdr, hash_lock); 3300 mutex_exit(hash_lock); 3301 3302 bytes_evicted += evicted; 3303 3304 /* 3305 * If evicted is zero, arc_evict_hdr() must have 3306 * decided to skip this header, don't increment 3307 * evict_count in this case. 3308 */ 3309 if (evicted != 0) 3310 evict_count++; 3311 3312 /* 3313 * If arc_size isn't overflowing, signal any 3314 * threads that might happen to be waiting. 3315 * 3316 * For each header evicted, we wake up a single 3317 * thread. If we used cv_broadcast, we could 3318 * wake up "too many" threads causing arc_size 3319 * to significantly overflow arc_c; since 3320 * arc_get_data_buf() doesn't check for overflow 3321 * when it's woken up (it doesn't because it's 3322 * possible for the ARC to be overflowing while 3323 * full of un-evictable buffers, and the 3324 * function should proceed in this case). 3325 * 3326 * If threads are left sleeping, due to not 3327 * using cv_broadcast, they will be woken up 3328 * just before arc_reclaim_thread() sleeps. 3329 */ 3330 mutex_enter(&arc_reclaim_lock); 3331 if (!arc_is_overflowing()) 3332 cv_signal(&arc_reclaim_waiters_cv); 3333 mutex_exit(&arc_reclaim_lock); 3334 } else { 3335 ARCSTAT_BUMP(arcstat_mutex_miss); 3336 } 3337 } 3338 3339 multilist_sublist_unlock(mls); 3340 3341 return (bytes_evicted); 3342} 3343 3344/* 3345 * Evict buffers from the given arc state, until we've removed the 3346 * specified number of bytes. Move the removed buffers to the 3347 * appropriate evict state. 3348 * 3349 * This function makes a "best effort". It skips over any buffers 3350 * it can't get a hash_lock on, and so, may not catch all candidates. 3351 * It may also return without evicting as much space as requested. 3352 * 3353 * If bytes is specified using the special value ARC_EVICT_ALL, this 3354 * will evict all available (i.e. unlocked and evictable) buffers from 3355 * the given arc state; which is used by arc_flush(). 3356 */ 3357static uint64_t 3358arc_evict_state(arc_state_t *state, uint64_t spa, int64_t bytes, 3359 arc_buf_contents_t type) 3360{ 3361 uint64_t total_evicted = 0; 3362 multilist_t *ml = &state->arcs_list[type]; 3363 int num_sublists; 3364 arc_buf_hdr_t **markers; 3365 3366 IMPLY(bytes < 0, bytes == ARC_EVICT_ALL); 3367 3368 num_sublists = multilist_get_num_sublists(ml); 3369 3370 /* 3371 * If we've tried to evict from each sublist, made some 3372 * progress, but still have not hit the target number of bytes 3373 * to evict, we want to keep trying. The markers allow us to 3374 * pick up where we left off for each individual sublist, rather 3375 * than starting from the tail each time. 3376 */ 3377 markers = kmem_zalloc(sizeof (*markers) * num_sublists, KM_SLEEP); 3378 for (int i = 0; i < num_sublists; i++) { 3379 markers[i] = kmem_cache_alloc(hdr_full_cache, KM_SLEEP); 3380 3381 /* 3382 * A b_spa of 0 is used to indicate that this header is 3383 * a marker. This fact is used in arc_adjust_type() and 3384 * arc_evict_state_impl(). 3385 */ 3386 markers[i]->b_spa = 0; 3387 3388 multilist_sublist_t *mls = multilist_sublist_lock(ml, i); 3389 multilist_sublist_insert_tail(mls, markers[i]); 3390 multilist_sublist_unlock(mls); 3391 } 3392 3393 /* 3394 * While we haven't hit our target number of bytes to evict, or 3395 * we're evicting all available buffers. 3396 */ 3397 while (total_evicted < bytes || bytes == ARC_EVICT_ALL) { 3398 /* 3399 * Start eviction using a randomly selected sublist, 3400 * this is to try and evenly balance eviction across all 3401 * sublists. Always starting at the same sublist 3402 * (e.g. index 0) would cause evictions to favor certain 3403 * sublists over others. 3404 */ 3405 int sublist_idx = multilist_get_random_index(ml); 3406 uint64_t scan_evicted = 0; 3407 3408 for (int i = 0; i < num_sublists; i++) { 3409 uint64_t bytes_remaining; 3410 uint64_t bytes_evicted; 3411 3412 if (bytes == ARC_EVICT_ALL) 3413 bytes_remaining = ARC_EVICT_ALL; 3414 else if (total_evicted < bytes) 3415 bytes_remaining = bytes - total_evicted; 3416 else 3417 break; 3418 3419 bytes_evicted = arc_evict_state_impl(ml, sublist_idx, 3420 markers[sublist_idx], spa, bytes_remaining); 3421 3422 scan_evicted += bytes_evicted; 3423 total_evicted += bytes_evicted; 3424 3425 /* we've reached the end, wrap to the beginning */ 3426 if (++sublist_idx >= num_sublists) 3427 sublist_idx = 0; 3428 } 3429 3430 /* 3431 * If we didn't evict anything during this scan, we have 3432 * no reason to believe we'll evict more during another 3433 * scan, so break the loop. 3434 */ 3435 if (scan_evicted == 0) { 3436 /* This isn't possible, let's make that obvious */ 3437 ASSERT3S(bytes, !=, 0); 3438 3439 /* 3440 * When bytes is ARC_EVICT_ALL, the only way to 3441 * break the loop is when scan_evicted is zero. 3442 * In that case, we actually have evicted enough, 3443 * so we don't want to increment the kstat. 3444 */ 3445 if (bytes != ARC_EVICT_ALL) { 3446 ASSERT3S(total_evicted, <, bytes); 3447 ARCSTAT_BUMP(arcstat_evict_not_enough); 3448 } 3449 3450 break; 3451 } 3452 } 3453 3454 for (int i = 0; i < num_sublists; i++) { 3455 multilist_sublist_t *mls = multilist_sublist_lock(ml, i); 3456 multilist_sublist_remove(mls, markers[i]); 3457 multilist_sublist_unlock(mls); 3458 3459 kmem_cache_free(hdr_full_cache, markers[i]); 3460 } 3461 kmem_free(markers, sizeof (*markers) * num_sublists); 3462 3463 return (total_evicted); 3464} 3465 3466/* 3467 * Flush all "evictable" data of the given type from the arc state 3468 * specified. This will not evict any "active" buffers (i.e. referenced). 3469 * 3470 * When 'retry' is set to B_FALSE, the function will make a single pass 3471 * over the state and evict any buffers that it can. Since it doesn't 3472 * continually retry the eviction, it might end up leaving some buffers 3473 * in the ARC due to lock misses. 3474 * 3475 * When 'retry' is set to B_TRUE, the function will continually retry the 3476 * eviction until *all* evictable buffers have been removed from the 3477 * state. As a result, if concurrent insertions into the state are 3478 * allowed (e.g. if the ARC isn't shutting down), this function might 3479 * wind up in an infinite loop, continually trying to evict buffers. 3480 */ 3481static uint64_t 3482arc_flush_state(arc_state_t *state, uint64_t spa, arc_buf_contents_t type, 3483 boolean_t retry) 3484{ 3485 uint64_t evicted = 0; 3486 3487 while (refcount_count(&state->arcs_esize[type]) != 0) { 3488 evicted += arc_evict_state(state, spa, ARC_EVICT_ALL, type); 3489 3490 if (!retry) 3491 break; 3492 } 3493 3494 return (evicted); 3495} 3496 3497/* 3498 * Evict the specified number of bytes from the state specified, 3499 * restricting eviction to the spa and type given. This function 3500 * prevents us from trying to evict more from a state's list than 3501 * is "evictable", and to skip evicting altogether when passed a 3502 * negative value for "bytes". In contrast, arc_evict_state() will 3503 * evict everything it can, when passed a negative value for "bytes". 3504 */ 3505static uint64_t 3506arc_adjust_impl(arc_state_t *state, uint64_t spa, int64_t bytes, 3507 arc_buf_contents_t type) 3508{ 3509 int64_t delta; 3510 3511 if (bytes > 0 && refcount_count(&state->arcs_esize[type]) > 0) { 3512 delta = MIN(refcount_count(&state->arcs_esize[type]), bytes); 3513 return (arc_evict_state(state, spa, delta, type)); 3514 } 3515 3516 return (0); 3517} 3518 3519/* 3520 * Evict metadata buffers from the cache, such that arc_meta_used is 3521 * capped by the arc_meta_limit tunable. 3522 */ 3523static uint64_t 3524arc_adjust_meta(void) 3525{ 3526 uint64_t total_evicted = 0; 3527 int64_t target; 3528 3529 /* 3530 * If we're over the meta limit, we want to evict enough 3531 * metadata to get back under the meta limit. We don't want to 3532 * evict so much that we drop the MRU below arc_p, though. If 3533 * we're over the meta limit more than we're over arc_p, we 3534 * evict some from the MRU here, and some from the MFU below. 3535 */ 3536 target = MIN((int64_t)(arc_meta_used - arc_meta_limit), 3537 (int64_t)(refcount_count(&arc_anon->arcs_size) + 3538 refcount_count(&arc_mru->arcs_size) - arc_p)); 3539 3540 total_evicted += arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA); 3541 3542 /* 3543 * Similar to the above, we want to evict enough bytes to get us 3544 * below the meta limit, but not so much as to drop us below the 3545 * space alloted to the MFU (which is defined as arc_c - arc_p). 3546 */ 3547 target = MIN((int64_t)(arc_meta_used - arc_meta_limit), 3548 (int64_t)(refcount_count(&arc_mfu->arcs_size) - (arc_c - arc_p))); 3549 3550 total_evicted += arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA); 3551 3552 return (total_evicted); 3553} 3554 3555/* 3556 * Return the type of the oldest buffer in the given arc state 3557 * 3558 * This function will select a random sublist of type ARC_BUFC_DATA and 3559 * a random sublist of type ARC_BUFC_METADATA. The tail of each sublist 3560 * is compared, and the type which contains the "older" buffer will be 3561 * returned. 3562 */ 3563static arc_buf_contents_t 3564arc_adjust_type(arc_state_t *state) 3565{ 3566 multilist_t *data_ml = &state->arcs_list[ARC_BUFC_DATA]; 3567 multilist_t *meta_ml = &state->arcs_list[ARC_BUFC_METADATA]; 3568 int data_idx = multilist_get_random_index(data_ml); 3569 int meta_idx = multilist_get_random_index(meta_ml); 3570 multilist_sublist_t *data_mls; 3571 multilist_sublist_t *meta_mls; 3572 arc_buf_contents_t type; 3573 arc_buf_hdr_t *data_hdr; 3574 arc_buf_hdr_t *meta_hdr; 3575 3576 /* 3577 * We keep the sublist lock until we're finished, to prevent 3578 * the headers from being destroyed via arc_evict_state(). 3579 */ 3580 data_mls = multilist_sublist_lock(data_ml, data_idx); 3581 meta_mls = multilist_sublist_lock(meta_ml, meta_idx); 3582 3583 /* 3584 * These two loops are to ensure we skip any markers that 3585 * might be at the tail of the lists due to arc_evict_state(). 3586 */ 3587 3588 for (data_hdr = multilist_sublist_tail(data_mls); data_hdr != NULL; 3589 data_hdr = multilist_sublist_prev(data_mls, data_hdr)) { 3590 if (data_hdr->b_spa != 0) 3591 break; 3592 } 3593 3594 for (meta_hdr = multilist_sublist_tail(meta_mls); meta_hdr != NULL; 3595 meta_hdr = multilist_sublist_prev(meta_mls, meta_hdr)) { 3596 if (meta_hdr->b_spa != 0) 3597 break; 3598 } 3599 3600 if (data_hdr == NULL && meta_hdr == NULL) { 3601 type = ARC_BUFC_DATA; 3602 } else if (data_hdr == NULL) { 3603 ASSERT3P(meta_hdr, !=, NULL); 3604 type = ARC_BUFC_METADATA; 3605 } else if (meta_hdr == NULL) { 3606 ASSERT3P(data_hdr, !=, NULL); 3607 type = ARC_BUFC_DATA; 3608 } else { 3609 ASSERT3P(data_hdr, !=, NULL); 3610 ASSERT3P(meta_hdr, !=, NULL); 3611 3612 /* The headers can't be on the sublist without an L1 header */ 3613 ASSERT(HDR_HAS_L1HDR(data_hdr)); 3614 ASSERT(HDR_HAS_L1HDR(meta_hdr)); 3615 3616 if (data_hdr->b_l1hdr.b_arc_access < 3617 meta_hdr->b_l1hdr.b_arc_access) { 3618 type = ARC_BUFC_DATA; 3619 } else { 3620 type = ARC_BUFC_METADATA; 3621 } 3622 } 3623 3624 multilist_sublist_unlock(meta_mls); 3625 multilist_sublist_unlock(data_mls); 3626 3627 return (type); 3628} 3629 3630/* 3631 * Evict buffers from the cache, such that arc_size is capped by arc_c. 3632 */ 3633static uint64_t 3634arc_adjust(void) 3635{ 3636 uint64_t total_evicted = 0; 3637 uint64_t bytes; 3638 int64_t target; 3639 3640 /* 3641 * If we're over arc_meta_limit, we want to correct that before 3642 * potentially evicting data buffers below. 3643 */ 3644 total_evicted += arc_adjust_meta(); 3645 3646 /* 3647 * Adjust MRU size 3648 * 3649 * If we're over the target cache size, we want to evict enough 3650 * from the list to get back to our target size. We don't want 3651 * to evict too much from the MRU, such that it drops below 3652 * arc_p. So, if we're over our target cache size more than 3653 * the MRU is over arc_p, we'll evict enough to get back to 3654 * arc_p here, and then evict more from the MFU below. 3655 */ 3656 target = MIN((int64_t)(arc_size - arc_c), 3657 (int64_t)(refcount_count(&arc_anon->arcs_size) + 3658 refcount_count(&arc_mru->arcs_size) + arc_meta_used - arc_p)); 3659 3660 /* 3661 * If we're below arc_meta_min, always prefer to evict data. 3662 * Otherwise, try to satisfy the requested number of bytes to 3663 * evict from the type which contains older buffers; in an 3664 * effort to keep newer buffers in the cache regardless of their 3665 * type. If we cannot satisfy the number of bytes from this 3666 * type, spill over into the next type. 3667 */ 3668 if (arc_adjust_type(arc_mru) == ARC_BUFC_METADATA && 3669 arc_meta_used > arc_meta_min) { 3670 bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA); 3671 total_evicted += bytes; 3672 3673 /* 3674 * If we couldn't evict our target number of bytes from 3675 * metadata, we try to get the rest from data. 3676 */ 3677 target -= bytes; 3678 3679 total_evicted += 3680 arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA); 3681 } else { 3682 bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA); 3683 total_evicted += bytes; 3684 3685 /* 3686 * If we couldn't evict our target number of bytes from 3687 * data, we try to get the rest from metadata. 3688 */ 3689 target -= bytes; 3690 3691 total_evicted += 3692 arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA); 3693 } 3694 3695 /* 3696 * Adjust MFU size 3697 * 3698 * Now that we've tried to evict enough from the MRU to get its 3699 * size back to arc_p, if we're still above the target cache 3700 * size, we evict the rest from the MFU. 3701 */ 3702 target = arc_size - arc_c; 3703 3704 if (arc_adjust_type(arc_mfu) == ARC_BUFC_METADATA && 3705 arc_meta_used > arc_meta_min) { 3706 bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA); 3707 total_evicted += bytes; 3708 3709 /* 3710 * If we couldn't evict our target number of bytes from 3711 * metadata, we try to get the rest from data. 3712 */ 3713 target -= bytes; 3714 3715 total_evicted += 3716 arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA); 3717 } else { 3718 bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA); 3719 total_evicted += bytes; 3720 3721 /* 3722 * If we couldn't evict our target number of bytes from 3723 * data, we try to get the rest from data. 3724 */ 3725 target -= bytes; 3726 3727 total_evicted += 3728 arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA); 3729 } 3730 3731 /* 3732 * Adjust ghost lists 3733 * 3734 * In addition to the above, the ARC also defines target values 3735 * for the ghost lists. The sum of the mru list and mru ghost 3736 * list should never exceed the target size of the cache, and 3737 * the sum of the mru list, mfu list, mru ghost list, and mfu 3738 * ghost list should never exceed twice the target size of the 3739 * cache. The following logic enforces these limits on the ghost 3740 * caches, and evicts from them as needed. 3741 */ 3742 target = refcount_count(&arc_mru->arcs_size) + 3743 refcount_count(&arc_mru_ghost->arcs_size) - arc_c; 3744 3745 bytes = arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_DATA); 3746 total_evicted += bytes; 3747 3748 target -= bytes; 3749 3750 total_evicted += 3751 arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_METADATA); 3752 3753 /* 3754 * We assume the sum of the mru list and mfu list is less than 3755 * or equal to arc_c (we enforced this above), which means we 3756 * can use the simpler of the two equations below: 3757 * 3758 * mru + mfu + mru ghost + mfu ghost <= 2 * arc_c 3759 * mru ghost + mfu ghost <= arc_c 3760 */ 3761 target = refcount_count(&arc_mru_ghost->arcs_size) + 3762 refcount_count(&arc_mfu_ghost->arcs_size) - arc_c; 3763 3764 bytes = arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_DATA); 3765 total_evicted += bytes; 3766 3767 target -= bytes; 3768 3769 total_evicted += 3770 arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_METADATA); 3771 3772 return (total_evicted); 3773} 3774 3775void 3776arc_flush(spa_t *spa, boolean_t retry) 3777{ 3778 uint64_t guid = 0; 3779 3780 /* 3781 * If retry is B_TRUE, a spa must not be specified since we have 3782 * no good way to determine if all of a spa's buffers have been 3783 * evicted from an arc state. 3784 */ 3785 ASSERT(!retry || spa == 0); 3786 3787 if (spa != NULL) 3788 guid = spa_load_guid(spa); 3789 3790 (void) arc_flush_state(arc_mru, guid, ARC_BUFC_DATA, retry); 3791 (void) arc_flush_state(arc_mru, guid, ARC_BUFC_METADATA, retry); 3792 3793 (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_DATA, retry); 3794 (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_METADATA, retry); 3795 3796 (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_DATA, retry); 3797 (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_METADATA, retry); 3798 3799 (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_DATA, retry); 3800 (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_METADATA, retry); 3801} 3802 3803void 3804arc_shrink(int64_t to_free) 3805{ 3806 if (arc_c > arc_c_min) { 3807 DTRACE_PROBE4(arc__shrink, uint64_t, arc_c, uint64_t, 3808 arc_c_min, uint64_t, arc_p, uint64_t, to_free); 3809 if (arc_c > arc_c_min + to_free) 3810 atomic_add_64(&arc_c, -to_free); 3811 else 3812 arc_c = arc_c_min; 3813 3814 atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift)); 3815 if (arc_c > arc_size) 3816 arc_c = MAX(arc_size, arc_c_min); 3817 if (arc_p > arc_c) 3818 arc_p = (arc_c >> 1); 3819 3820 DTRACE_PROBE2(arc__shrunk, uint64_t, arc_c, uint64_t, 3821 arc_p); 3822 3823 ASSERT(arc_c >= arc_c_min); 3824 ASSERT((int64_t)arc_p >= 0); 3825 } 3826 3827 if (arc_size > arc_c) { 3828 DTRACE_PROBE2(arc__shrink_adjust, uint64_t, arc_size, 3829 uint64_t, arc_c); 3830 (void) arc_adjust(); 3831 } 3832} 3833 3834static long needfree = 0; 3835 3836typedef enum free_memory_reason_t { 3837 FMR_UNKNOWN, 3838 FMR_NEEDFREE, 3839 FMR_LOTSFREE, 3840 FMR_SWAPFS_MINFREE, 3841 FMR_PAGES_PP_MAXIMUM, 3842 FMR_HEAP_ARENA, 3843 FMR_ZIO_ARENA, 3844 FMR_ZIO_FRAG, 3845} free_memory_reason_t; 3846 3847int64_t last_free_memory; 3848free_memory_reason_t last_free_reason; 3849 3850/* 3851 * Additional reserve of pages for pp_reserve. 3852 */ 3853int64_t arc_pages_pp_reserve = 64; 3854 3855/* 3856 * Additional reserve of pages for swapfs. 3857 */ 3858int64_t arc_swapfs_reserve = 64; 3859 3860/* 3861 * Return the amount of memory that can be consumed before reclaim will be 3862 * needed. Positive if there is sufficient free memory, negative indicates 3863 * the amount of memory that needs to be freed up. 3864 */ 3865static int64_t 3866arc_available_memory(void) 3867{ 3868 int64_t lowest = INT64_MAX; 3869 int64_t n; 3870 free_memory_reason_t r = FMR_UNKNOWN; 3871 3872#ifdef _KERNEL 3873 if (needfree > 0) { 3874 n = PAGESIZE * (-needfree); 3875 if (n < lowest) { 3876 lowest = n; 3877 r = FMR_NEEDFREE; 3878 } 3879 } 3880 3881 /* 3882 * Cooperate with pagedaemon when it's time for it to scan 3883 * and reclaim some pages. 3884 */ 3885 n = PAGESIZE * ((int64_t)freemem - zfs_arc_free_target); 3886 if (n < lowest) { 3887 lowest = n; 3888 r = FMR_LOTSFREE; 3889 } 3890 3891#ifdef illumos 3892 /* 3893 * check that we're out of range of the pageout scanner. It starts to 3894 * schedule paging if freemem is less than lotsfree and needfree. 3895 * lotsfree is the high-water mark for pageout, and needfree is the 3896 * number of needed free pages. We add extra pages here to make sure 3897 * the scanner doesn't start up while we're freeing memory. 3898 */ 3899 n = PAGESIZE * (freemem - lotsfree - needfree - desfree); 3900 if (n < lowest) { 3901 lowest = n; 3902 r = FMR_LOTSFREE; 3903 } 3904 3905 /* 3906 * check to make sure that swapfs has enough space so that anon 3907 * reservations can still succeed. anon_resvmem() checks that the 3908 * availrmem is greater than swapfs_minfree, and the number of reserved 3909 * swap pages. We also add a bit of extra here just to prevent 3910 * circumstances from getting really dire. 3911 */ 3912 n = PAGESIZE * (availrmem - swapfs_minfree - swapfs_reserve - 3913 desfree - arc_swapfs_reserve); 3914 if (n < lowest) { 3915 lowest = n; 3916 r = FMR_SWAPFS_MINFREE; 3917 } 3918 3919 3920 /* 3921 * Check that we have enough availrmem that memory locking (e.g., via 3922 * mlock(3C) or memcntl(2)) can still succeed. (pages_pp_maximum 3923 * stores the number of pages that cannot be locked; when availrmem 3924 * drops below pages_pp_maximum, page locking mechanisms such as 3925 * page_pp_lock() will fail.) 3926 */ 3927 n = PAGESIZE * (availrmem - pages_pp_maximum - 3928 arc_pages_pp_reserve); 3929 if (n < lowest) { 3930 lowest = n; 3931 r = FMR_PAGES_PP_MAXIMUM; 3932 } 3933 3934#endif /* illumos */ 3935#if defined(__i386) || !defined(UMA_MD_SMALL_ALLOC) 3936 /* 3937 * If we're on an i386 platform, it's possible that we'll exhaust the 3938 * kernel heap space before we ever run out of available physical 3939 * memory. Most checks of the size of the heap_area compare against 3940 * tune.t_minarmem, which is the minimum available real memory that we 3941 * can have in the system. However, this is generally fixed at 25 pages 3942 * which is so low that it's useless. In this comparison, we seek to 3943 * calculate the total heap-size, and reclaim if more than 3/4ths of the 3944 * heap is allocated. (Or, in the calculation, if less than 1/4th is 3945 * free) 3946 */ 3947 n = (int64_t)vmem_size(heap_arena, VMEM_FREE) - 3948 (vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC) >> 2); 3949 if (n < lowest) { 3950 lowest = n; 3951 r = FMR_HEAP_ARENA; 3952 } 3953#define zio_arena NULL 3954#else 3955#define zio_arena heap_arena 3956#endif 3957 3958 /* 3959 * If zio data pages are being allocated out of a separate heap segment, 3960 * then enforce that the size of available vmem for this arena remains 3961 * above about 1/16th free. 3962 * 3963 * Note: The 1/16th arena free requirement was put in place 3964 * to aggressively evict memory from the arc in order to avoid 3965 * memory fragmentation issues. 3966 */ 3967 if (zio_arena != NULL) { 3968 n = (int64_t)vmem_size(zio_arena, VMEM_FREE) - 3969 (vmem_size(zio_arena, VMEM_ALLOC) >> 4); 3970 if (n < lowest) { 3971 lowest = n; 3972 r = FMR_ZIO_ARENA; 3973 } 3974 } 3975 3976 /* 3977 * Above limits know nothing about real level of KVA fragmentation. 3978 * Start aggressive reclamation if too little sequential KVA left. 3979 */ 3980 if (lowest > 0) { 3981 n = (vmem_size(heap_arena, VMEM_MAXFREE) < SPA_MAXBLOCKSIZE) ? 3982 -((int64_t)vmem_size(heap_arena, VMEM_ALLOC) >> 4) : 3983 INT64_MAX; 3984 if (n < lowest) { 3985 lowest = n; 3986 r = FMR_ZIO_FRAG; 3987 } 3988 } 3989 3990#else /* _KERNEL */ 3991 /* Every 100 calls, free a small amount */ 3992 if (spa_get_random(100) == 0) 3993 lowest = -1024; 3994#endif /* _KERNEL */ 3995 3996 last_free_memory = lowest; 3997 last_free_reason = r; 3998 DTRACE_PROBE2(arc__available_memory, int64_t, lowest, int, r); 3999 return (lowest); 4000} 4001 4002 4003/* 4004 * Determine if the system is under memory pressure and is asking 4005 * to reclaim memory. A return value of B_TRUE indicates that the system 4006 * is under memory pressure and that the arc should adjust accordingly. 4007 */ 4008static boolean_t 4009arc_reclaim_needed(void) 4010{ 4011 return (arc_available_memory() < 0); 4012} 4013 4014extern kmem_cache_t *zio_buf_cache[]; 4015extern kmem_cache_t *zio_data_buf_cache[]; 4016extern kmem_cache_t *range_seg_cache; 4017 4018static __noinline void 4019arc_kmem_reap_now(void) 4020{ 4021 size_t i; 4022 kmem_cache_t *prev_cache = NULL; 4023 kmem_cache_t *prev_data_cache = NULL; 4024 4025 DTRACE_PROBE(arc__kmem_reap_start); 4026#ifdef _KERNEL 4027 if (arc_meta_used >= arc_meta_limit) { 4028 /* 4029 * We are exceeding our meta-data cache limit. 4030 * Purge some DNLC entries to release holds on meta-data. 4031 */ 4032 dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent); 4033 } 4034#if defined(__i386) 4035 /* 4036 * Reclaim unused memory from all kmem caches. 4037 */ 4038 kmem_reap(); 4039#endif 4040#endif 4041 4042 for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) { 4043 if (zio_buf_cache[i] != prev_cache) { 4044 prev_cache = zio_buf_cache[i]; 4045 kmem_cache_reap_now(zio_buf_cache[i]); 4046 } 4047 if (zio_data_buf_cache[i] != prev_data_cache) { 4048 prev_data_cache = zio_data_buf_cache[i]; 4049 kmem_cache_reap_now(zio_data_buf_cache[i]); 4050 } 4051 } 4052 kmem_cache_reap_now(buf_cache); 4053 kmem_cache_reap_now(hdr_full_cache); 4054 kmem_cache_reap_now(hdr_l2only_cache); 4055 kmem_cache_reap_now(range_seg_cache); 4056 4057#ifdef illumos 4058 if (zio_arena != NULL) { 4059 /* 4060 * Ask the vmem arena to reclaim unused memory from its 4061 * quantum caches. 4062 */ 4063 vmem_qcache_reap(zio_arena); 4064 } 4065#endif 4066 DTRACE_PROBE(arc__kmem_reap_end); 4067} 4068 4069/* 4070 * Threads can block in arc_get_data_buf() waiting for this thread to evict 4071 * enough data and signal them to proceed. When this happens, the threads in 4072 * arc_get_data_buf() are sleeping while holding the hash lock for their 4073 * particular arc header. Thus, we must be careful to never sleep on a 4074 * hash lock in this thread. This is to prevent the following deadlock: 4075 * 4076 * - Thread A sleeps on CV in arc_get_data_buf() holding hash lock "L", 4077 * waiting for the reclaim thread to signal it. 4078 * 4079 * - arc_reclaim_thread() tries to acquire hash lock "L" using mutex_enter, 4080 * fails, and goes to sleep forever. 4081 * 4082 * This possible deadlock is avoided by always acquiring a hash lock 4083 * using mutex_tryenter() from arc_reclaim_thread(). 4084 */ 4085static void 4086arc_reclaim_thread(void *dummy __unused) 4087{ 4088 hrtime_t growtime = 0; 4089 callb_cpr_t cpr; 4090 4091 CALLB_CPR_INIT(&cpr, &arc_reclaim_lock, callb_generic_cpr, FTAG); 4092 4093 mutex_enter(&arc_reclaim_lock); 4094 while (!arc_reclaim_thread_exit) { 4095 uint64_t evicted = 0; 4096 4097 /* 4098 * This is necessary in order for the mdb ::arc dcmd to 4099 * show up to date information. Since the ::arc command 4100 * does not call the kstat's update function, without 4101 * this call, the command may show stale stats for the 4102 * anon, mru, mru_ghost, mfu, and mfu_ghost lists. Even 4103 * with this change, the data might be up to 1 second 4104 * out of date; but that should suffice. The arc_state_t 4105 * structures can be queried directly if more accurate 4106 * information is needed. 4107 */ 4108 if (arc_ksp != NULL) 4109 arc_ksp->ks_update(arc_ksp, KSTAT_READ); 4110 4111 mutex_exit(&arc_reclaim_lock); 4112 4113 /* 4114 * We call arc_adjust() before (possibly) calling 4115 * arc_kmem_reap_now(), so that we can wake up 4116 * arc_get_data_buf() sooner. 4117 */ 4118 evicted = arc_adjust(); 4119 4120 int64_t free_memory = arc_available_memory(); 4121 if (free_memory < 0) { 4122 4123 arc_no_grow = B_TRUE; 4124 arc_warm = B_TRUE; 4125 4126 /* 4127 * Wait at least zfs_grow_retry (default 60) seconds 4128 * before considering growing. 4129 */ 4130 growtime = gethrtime() + SEC2NSEC(arc_grow_retry); 4131 4132 arc_kmem_reap_now(); 4133 4134 /* 4135 * If we are still low on memory, shrink the ARC 4136 * so that we have arc_shrink_min free space. 4137 */ 4138 free_memory = arc_available_memory(); 4139 4140 int64_t to_free = 4141 (arc_c >> arc_shrink_shift) - free_memory; 4142 if (to_free > 0) { 4143#ifdef _KERNEL 4144 to_free = MAX(to_free, ptob(needfree)); 4145#endif 4146 arc_shrink(to_free); 4147 } 4148 } else if (free_memory < arc_c >> arc_no_grow_shift) { 4149 arc_no_grow = B_TRUE; 4150 } else if (gethrtime() >= growtime) { 4151 arc_no_grow = B_FALSE; 4152 } 4153 4154 mutex_enter(&arc_reclaim_lock); 4155 4156 /* 4157 * If evicted is zero, we couldn't evict anything via 4158 * arc_adjust(). This could be due to hash lock 4159 * collisions, but more likely due to the majority of 4160 * arc buffers being unevictable. Therefore, even if 4161 * arc_size is above arc_c, another pass is unlikely to 4162 * be helpful and could potentially cause us to enter an 4163 * infinite loop. 4164 */ 4165 if (arc_size <= arc_c || evicted == 0) { 4166#ifdef _KERNEL 4167 needfree = 0; 4168#endif 4169 /* 4170 * We're either no longer overflowing, or we 4171 * can't evict anything more, so we should wake 4172 * up any threads before we go to sleep. 4173 */ 4174 cv_broadcast(&arc_reclaim_waiters_cv); 4175 4176 /* 4177 * Block until signaled, or after one second (we 4178 * might need to perform arc_kmem_reap_now() 4179 * even if we aren't being signalled) 4180 */ 4181 CALLB_CPR_SAFE_BEGIN(&cpr); 4182 (void) cv_timedwait_hires(&arc_reclaim_thread_cv, 4183 &arc_reclaim_lock, SEC2NSEC(1), MSEC2NSEC(1), 0); 4184 CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_lock); 4185 } 4186 } 4187 4188 arc_reclaim_thread_exit = B_FALSE; 4189 cv_broadcast(&arc_reclaim_thread_cv); 4190 CALLB_CPR_EXIT(&cpr); /* drops arc_reclaim_lock */ 4191 thread_exit(); 4192} 4193 4194static u_int arc_dnlc_evicts_arg; 4195extern struct vfsops zfs_vfsops; 4196 4197static void 4198arc_dnlc_evicts_thread(void *dummy __unused) 4199{ 4200 callb_cpr_t cpr; 4201 u_int percent; 4202 4203 CALLB_CPR_INIT(&cpr, &arc_dnlc_evicts_lock, callb_generic_cpr, FTAG); 4204 4205 mutex_enter(&arc_dnlc_evicts_lock); 4206 while (!arc_dnlc_evicts_thread_exit) { 4207 CALLB_CPR_SAFE_BEGIN(&cpr); 4208 (void) cv_wait(&arc_dnlc_evicts_cv, &arc_dnlc_evicts_lock); 4209 CALLB_CPR_SAFE_END(&cpr, &arc_dnlc_evicts_lock); 4210 if (arc_dnlc_evicts_arg != 0) { 4211 percent = arc_dnlc_evicts_arg; 4212 mutex_exit(&arc_dnlc_evicts_lock); 4213#ifdef _KERNEL 4214 vnlru_free(desiredvnodes * percent / 100, &zfs_vfsops); 4215#endif 4216 mutex_enter(&arc_dnlc_evicts_lock); 4217 /* 4218 * Clear our token only after vnlru_free() 4219 * pass is done, to avoid false queueing of 4220 * the requests. 4221 */ 4222 arc_dnlc_evicts_arg = 0; 4223 } 4224 } 4225 arc_dnlc_evicts_thread_exit = FALSE; 4226 cv_broadcast(&arc_dnlc_evicts_cv); 4227 CALLB_CPR_EXIT(&cpr); 4228 thread_exit(); 4229} 4230 4231void 4232dnlc_reduce_cache(void *arg) 4233{ 4234 u_int percent; 4235 4236 percent = (u_int)(uintptr_t)arg; 4237 mutex_enter(&arc_dnlc_evicts_lock); 4238 if (arc_dnlc_evicts_arg == 0) { 4239 arc_dnlc_evicts_arg = percent; 4240 cv_broadcast(&arc_dnlc_evicts_cv); 4241 } 4242 mutex_exit(&arc_dnlc_evicts_lock); 4243} 4244 4245/* 4246 * Adapt arc info given the number of bytes we are trying to add and 4247 * the state that we are comming from. This function is only called 4248 * when we are adding new content to the cache. 4249 */ 4250static void 4251arc_adapt(int bytes, arc_state_t *state) 4252{ 4253 int mult; 4254 uint64_t arc_p_min = (arc_c >> arc_p_min_shift); 4255 int64_t mrug_size = refcount_count(&arc_mru_ghost->arcs_size); 4256 int64_t mfug_size = refcount_count(&arc_mfu_ghost->arcs_size); 4257 4258 if (state == arc_l2c_only) 4259 return; 4260 4261 ASSERT(bytes > 0); 4262 /* 4263 * Adapt the target size of the MRU list: 4264 * - if we just hit in the MRU ghost list, then increase 4265 * the target size of the MRU list. 4266 * - if we just hit in the MFU ghost list, then increase 4267 * the target size of the MFU list by decreasing the 4268 * target size of the MRU list. 4269 */ 4270 if (state == arc_mru_ghost) { 4271 mult = (mrug_size >= mfug_size) ? 1 : (mfug_size / mrug_size); 4272 mult = MIN(mult, 10); /* avoid wild arc_p adjustment */ 4273 4274 arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult); 4275 } else if (state == arc_mfu_ghost) { 4276 uint64_t delta; 4277 4278 mult = (mfug_size >= mrug_size) ? 1 : (mrug_size / mfug_size); 4279 mult = MIN(mult, 10); 4280 4281 delta = MIN(bytes * mult, arc_p); 4282 arc_p = MAX(arc_p_min, arc_p - delta); 4283 } 4284 ASSERT((int64_t)arc_p >= 0); 4285 4286 if (arc_reclaim_needed()) { 4287 cv_signal(&arc_reclaim_thread_cv); 4288 return; 4289 } 4290 4291 if (arc_no_grow) 4292 return; 4293 4294 if (arc_c >= arc_c_max) 4295 return; 4296 4297 /* 4298 * If we're within (2 * maxblocksize) bytes of the target 4299 * cache size, increment the target cache size 4300 */ 4301 if (arc_size > arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) { 4302 DTRACE_PROBE1(arc__inc_adapt, int, bytes); 4303 atomic_add_64(&arc_c, (int64_t)bytes); 4304 if (arc_c > arc_c_max) 4305 arc_c = arc_c_max; 4306 else if (state == arc_anon) 4307 atomic_add_64(&arc_p, (int64_t)bytes); 4308 if (arc_p > arc_c) 4309 arc_p = arc_c; 4310 } 4311 ASSERT((int64_t)arc_p >= 0); 4312} 4313 4314/* 4315 * Check if arc_size has grown past our upper threshold, determined by 4316 * zfs_arc_overflow_shift. 4317 */ 4318static boolean_t 4319arc_is_overflowing(void) 4320{ 4321 /* Always allow at least one block of overflow */ 4322 uint64_t overflow = MAX(SPA_MAXBLOCKSIZE, 4323 arc_c >> zfs_arc_overflow_shift); 4324 4325 return (arc_size >= arc_c + overflow); 4326} 4327 4328/* 4329 * Allocate a block and return it to the caller. If we are hitting the 4330 * hard limit for the cache size, we must sleep, waiting for the eviction 4331 * thread to catch up. If we're past the target size but below the hard 4332 * limit, we'll only signal the reclaim thread and continue on. 4333 */ 4334static void * 4335arc_get_data_buf(arc_buf_hdr_t *hdr, uint64_t size, void *tag) 4336{ 4337 void *datap = NULL; 4338 arc_state_t *state = hdr->b_l1hdr.b_state; 4339 arc_buf_contents_t type = arc_buf_type(hdr); 4340 4341 arc_adapt(size, state); 4342 4343 /* 4344 * If arc_size is currently overflowing, and has grown past our 4345 * upper limit, we must be adding data faster than the evict 4346 * thread can evict. Thus, to ensure we don't compound the 4347 * problem by adding more data and forcing arc_size to grow even 4348 * further past it's target size, we halt and wait for the 4349 * eviction thread to catch up. 4350 * 4351 * It's also possible that the reclaim thread is unable to evict 4352 * enough buffers to get arc_size below the overflow limit (e.g. 4353 * due to buffers being un-evictable, or hash lock collisions). 4354 * In this case, we want to proceed regardless if we're 4355 * overflowing; thus we don't use a while loop here. 4356 */ 4357 if (arc_is_overflowing()) { 4358 mutex_enter(&arc_reclaim_lock); 4359 4360 /* 4361 * Now that we've acquired the lock, we may no longer be 4362 * over the overflow limit, lets check. 4363 * 4364 * We're ignoring the case of spurious wake ups. If that 4365 * were to happen, it'd let this thread consume an ARC 4366 * buffer before it should have (i.e. before we're under 4367 * the overflow limit and were signalled by the reclaim 4368 * thread). As long as that is a rare occurrence, it 4369 * shouldn't cause any harm. 4370 */ 4371 if (arc_is_overflowing()) { 4372 cv_signal(&arc_reclaim_thread_cv); 4373 cv_wait(&arc_reclaim_waiters_cv, &arc_reclaim_lock); 4374 } 4375 4376 mutex_exit(&arc_reclaim_lock); 4377 } 4378 4379 VERIFY3U(hdr->b_type, ==, type); 4380 if (type == ARC_BUFC_METADATA) { 4381 datap = zio_buf_alloc(size); 4382 arc_space_consume(size, ARC_SPACE_META); 4383 } else { 4384 ASSERT(type == ARC_BUFC_DATA); 4385 datap = zio_data_buf_alloc(size); 4386 arc_space_consume(size, ARC_SPACE_DATA); 4387 } 4388 4389 /* 4390 * Update the state size. Note that ghost states have a 4391 * "ghost size" and so don't need to be updated. 4392 */ 4393 if (!GHOST_STATE(state)) { 4394 4395 (void) refcount_add_many(&state->arcs_size, size, tag); 4396 4397 /* 4398 * If this is reached via arc_read, the link is 4399 * protected by the hash lock. If reached via 4400 * arc_buf_alloc, the header should not be accessed by 4401 * any other thread. And, if reached via arc_read_done, 4402 * the hash lock will protect it if it's found in the 4403 * hash table; otherwise no other thread should be 4404 * trying to [add|remove]_reference it. 4405 */ 4406 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) { 4407 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 4408 (void) refcount_add_many(&state->arcs_esize[type], 4409 size, tag); 4410 } 4411 4412 /* 4413 * If we are growing the cache, and we are adding anonymous 4414 * data, and we have outgrown arc_p, update arc_p 4415 */ 4416 if (arc_size < arc_c && hdr->b_l1hdr.b_state == arc_anon && 4417 (refcount_count(&arc_anon->arcs_size) + 4418 refcount_count(&arc_mru->arcs_size) > arc_p)) 4419 arc_p = MIN(arc_c, arc_p + size); 4420 } 4421 ARCSTAT_BUMP(arcstat_allocated); 4422 return (datap); 4423} 4424 4425/* 4426 * Free the arc data buffer. 4427 */ 4428static void 4429arc_free_data_buf(arc_buf_hdr_t *hdr, void *data, uint64_t size, void *tag) 4430{ 4431 arc_state_t *state = hdr->b_l1hdr.b_state; 4432 arc_buf_contents_t type = arc_buf_type(hdr); 4433 4434 /* protected by hash lock, if in the hash table */ 4435 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) { 4436 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 4437 ASSERT(state != arc_anon && state != arc_l2c_only); 4438 4439 (void) refcount_remove_many(&state->arcs_esize[type], 4440 size, tag); 4441 } 4442 (void) refcount_remove_many(&state->arcs_size, size, tag); 4443 4444 VERIFY3U(hdr->b_type, ==, type); 4445 if (type == ARC_BUFC_METADATA) { 4446 zio_buf_free(data, size); 4447 arc_space_return(size, ARC_SPACE_META); 4448 } else { 4449 ASSERT(type == ARC_BUFC_DATA); 4450 zio_data_buf_free(data, size); 4451 arc_space_return(size, ARC_SPACE_DATA); 4452 } 4453} 4454 4455/* 4456 * This routine is called whenever a buffer is accessed. 4457 * NOTE: the hash lock is dropped in this function. 4458 */ 4459static void 4460arc_access(arc_buf_hdr_t *hdr, kmutex_t *hash_lock) 4461{ 4462 clock_t now; 4463 4464 ASSERT(MUTEX_HELD(hash_lock)); 4465 ASSERT(HDR_HAS_L1HDR(hdr)); 4466 4467 if (hdr->b_l1hdr.b_state == arc_anon) { 4468 /* 4469 * This buffer is not in the cache, and does not 4470 * appear in our "ghost" list. Add the new buffer 4471 * to the MRU state. 4472 */ 4473 4474 ASSERT0(hdr->b_l1hdr.b_arc_access); 4475 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 4476 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr); 4477 arc_change_state(arc_mru, hdr, hash_lock); 4478 4479 } else if (hdr->b_l1hdr.b_state == arc_mru) { 4480 now = ddi_get_lbolt(); 4481 4482 /* 4483 * If this buffer is here because of a prefetch, then either: 4484 * - clear the flag if this is a "referencing" read 4485 * (any subsequent access will bump this into the MFU state). 4486 * or 4487 * - move the buffer to the head of the list if this is 4488 * another prefetch (to make it less likely to be evicted). 4489 */ 4490 if (HDR_PREFETCH(hdr)) { 4491 if (refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) { 4492 /* link protected by hash lock */ 4493 ASSERT(multilist_link_active( 4494 &hdr->b_l1hdr.b_arc_node)); 4495 } else { 4496 arc_hdr_clear_flags(hdr, ARC_FLAG_PREFETCH); 4497 ARCSTAT_BUMP(arcstat_mru_hits); 4498 } 4499 hdr->b_l1hdr.b_arc_access = now; 4500 return; 4501 } 4502 4503 /* 4504 * This buffer has been "accessed" only once so far, 4505 * but it is still in the cache. Move it to the MFU 4506 * state. 4507 */ 4508 if (now > hdr->b_l1hdr.b_arc_access + ARC_MINTIME) { 4509 /* 4510 * More than 125ms have passed since we 4511 * instantiated this buffer. Move it to the 4512 * most frequently used state. 4513 */ 4514 hdr->b_l1hdr.b_arc_access = now; 4515 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 4516 arc_change_state(arc_mfu, hdr, hash_lock); 4517 } 4518 ARCSTAT_BUMP(arcstat_mru_hits); 4519 } else if (hdr->b_l1hdr.b_state == arc_mru_ghost) { 4520 arc_state_t *new_state; 4521 /* 4522 * This buffer has been "accessed" recently, but 4523 * was evicted from the cache. Move it to the 4524 * MFU state. 4525 */ 4526 4527 if (HDR_PREFETCH(hdr)) { 4528 new_state = arc_mru; 4529 if (refcount_count(&hdr->b_l1hdr.b_refcnt) > 0) 4530 arc_hdr_clear_flags(hdr, ARC_FLAG_PREFETCH); 4531 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr); 4532 } else { 4533 new_state = arc_mfu; 4534 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 4535 } 4536 4537 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 4538 arc_change_state(new_state, hdr, hash_lock); 4539 4540 ARCSTAT_BUMP(arcstat_mru_ghost_hits); 4541 } else if (hdr->b_l1hdr.b_state == arc_mfu) { 4542 /* 4543 * This buffer has been accessed more than once and is 4544 * still in the cache. Keep it in the MFU state. 4545 * 4546 * NOTE: an add_reference() that occurred when we did 4547 * the arc_read() will have kicked this off the list. 4548 * If it was a prefetch, we will explicitly move it to 4549 * the head of the list now. 4550 */ 4551 if ((HDR_PREFETCH(hdr)) != 0) { 4552 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 4553 /* link protected by hash_lock */ 4554 ASSERT(multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 4555 } 4556 ARCSTAT_BUMP(arcstat_mfu_hits); 4557 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 4558 } else if (hdr->b_l1hdr.b_state == arc_mfu_ghost) { 4559 arc_state_t *new_state = arc_mfu; 4560 /* 4561 * This buffer has been accessed more than once but has 4562 * been evicted from the cache. Move it back to the 4563 * MFU state. 4564 */ 4565 4566 if (HDR_PREFETCH(hdr)) { 4567 /* 4568 * This is a prefetch access... 4569 * move this block back to the MRU state. 4570 */ 4571 ASSERT0(refcount_count(&hdr->b_l1hdr.b_refcnt)); 4572 new_state = arc_mru; 4573 } 4574 4575 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 4576 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 4577 arc_change_state(new_state, hdr, hash_lock); 4578 4579 ARCSTAT_BUMP(arcstat_mfu_ghost_hits); 4580 } else if (hdr->b_l1hdr.b_state == arc_l2c_only) { 4581 /* 4582 * This buffer is on the 2nd Level ARC. 4583 */ 4584 4585 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 4586 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 4587 arc_change_state(arc_mfu, hdr, hash_lock); 4588 } else { 4589 ASSERT(!"invalid arc state"); 4590 } 4591} 4592 4593/* a generic arc_done_func_t which you can use */ 4594/* ARGSUSED */ 4595void 4596arc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg) 4597{ 4598 if (zio == NULL || zio->io_error == 0) 4599 bcopy(buf->b_data, arg, HDR_GET_LSIZE(buf->b_hdr)); 4600 arc_buf_destroy(buf, arg); 4601} 4602 4603/* a generic arc_done_func_t */ 4604void 4605arc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg) 4606{ 4607 arc_buf_t **bufp = arg; 4608 if (zio && zio->io_error) { 4609 arc_buf_destroy(buf, arg); 4610 *bufp = NULL; 4611 } else { 4612 *bufp = buf; 4613 ASSERT(buf->b_data); 4614 } 4615} 4616 4617static void 4618arc_hdr_verify(arc_buf_hdr_t *hdr, blkptr_t *bp) 4619{ 4620 if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp)) { 4621 ASSERT3U(HDR_GET_PSIZE(hdr), ==, 0); 4622 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF); 4623 } else { 4624 if (HDR_COMPRESSION_ENABLED(hdr)) { 4625 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, 4626 BP_GET_COMPRESS(bp)); 4627 } 4628 ASSERT3U(HDR_GET_LSIZE(hdr), ==, BP_GET_LSIZE(bp)); 4629 ASSERT3U(HDR_GET_PSIZE(hdr), ==, BP_GET_PSIZE(bp)); 4630 } 4631} 4632 4633static void 4634arc_read_done(zio_t *zio) 4635{ 4636 arc_buf_hdr_t *hdr = zio->io_private; 4637 arc_buf_t *abuf = NULL; /* buffer we're assigning to callback */ 4638 kmutex_t *hash_lock = NULL; 4639 arc_callback_t *callback_list, *acb; 4640 int freeable = B_FALSE; 4641 4642 /* 4643 * The hdr was inserted into hash-table and removed from lists 4644 * prior to starting I/O. We should find this header, since 4645 * it's in the hash table, and it should be legit since it's 4646 * not possible to evict it during the I/O. The only possible 4647 * reason for it not to be found is if we were freed during the 4648 * read. 4649 */ 4650 if (HDR_IN_HASH_TABLE(hdr)) { 4651 ASSERT3U(hdr->b_birth, ==, BP_PHYSICAL_BIRTH(zio->io_bp)); 4652 ASSERT3U(hdr->b_dva.dva_word[0], ==, 4653 BP_IDENTITY(zio->io_bp)->dva_word[0]); 4654 ASSERT3U(hdr->b_dva.dva_word[1], ==, 4655 BP_IDENTITY(zio->io_bp)->dva_word[1]); 4656 4657 arc_buf_hdr_t *found = buf_hash_find(hdr->b_spa, zio->io_bp, 4658 &hash_lock); 4659 4660 ASSERT((found == hdr && 4661 DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) || 4662 (found == hdr && HDR_L2_READING(hdr))); 4663 ASSERT3P(hash_lock, !=, NULL); 4664 } 4665 4666 if (zio->io_error == 0) { 4667 /* byteswap if necessary */ 4668 if (BP_SHOULD_BYTESWAP(zio->io_bp)) { 4669 if (BP_GET_LEVEL(zio->io_bp) > 0) { 4670 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_UINT64; 4671 } else { 4672 hdr->b_l1hdr.b_byteswap = 4673 DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp)); 4674 } 4675 } else { 4676 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS; 4677 } 4678 } 4679 4680 arc_hdr_clear_flags(hdr, ARC_FLAG_L2_EVICTED); 4681 if (l2arc_noprefetch && HDR_PREFETCH(hdr)) 4682 arc_hdr_clear_flags(hdr, ARC_FLAG_L2CACHE); 4683 4684 callback_list = hdr->b_l1hdr.b_acb; 4685 ASSERT3P(callback_list, !=, NULL); 4686 4687 if (hash_lock && zio->io_error == 0 && 4688 hdr->b_l1hdr.b_state == arc_anon) { 4689 /* 4690 * Only call arc_access on anonymous buffers. This is because 4691 * if we've issued an I/O for an evicted buffer, we've already 4692 * called arc_access (to prevent any simultaneous readers from 4693 * getting confused). 4694 */ 4695 arc_access(hdr, hash_lock); 4696 } 4697 4698 /* create copies of the data buffer for the callers */ 4699 for (acb = callback_list; acb; acb = acb->acb_next) { 4700 if (acb->acb_done != NULL) { 4701 /* 4702 * If we're here, then this must be a demand read 4703 * since prefetch requests don't have callbacks. 4704 * If a read request has a callback (i.e. acb_done is 4705 * not NULL), then we decompress the data for the 4706 * first request and clone the rest. This avoids 4707 * having to waste cpu resources decompressing data 4708 * that nobody is explicitly waiting to read. 4709 */ 4710 if (abuf == NULL) { 4711 acb->acb_buf = arc_buf_alloc_impl(hdr, 4712 acb->acb_private); 4713 if (zio->io_error == 0) { 4714 zio->io_error = 4715 arc_decompress(acb->acb_buf); 4716 } 4717 abuf = acb->acb_buf; 4718 } else { 4719 add_reference(hdr, acb->acb_private); 4720 acb->acb_buf = arc_buf_clone(abuf); 4721 } 4722 } 4723 } 4724 hdr->b_l1hdr.b_acb = NULL; 4725 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS); 4726 if (abuf == NULL) { 4727 /* 4728 * This buffer didn't have a callback so it must 4729 * be a prefetch. 4730 */ 4731 ASSERT(HDR_PREFETCH(hdr)); 4732 ASSERT0(hdr->b_l1hdr.b_bufcnt); 4733 ASSERT3P(hdr->b_l1hdr.b_pdata, !=, NULL); 4734 } 4735 4736 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt) || 4737 callback_list != NULL); 4738 4739 if (zio->io_error == 0) { 4740 arc_hdr_verify(hdr, zio->io_bp); 4741 } else { 4742 arc_hdr_set_flags(hdr, ARC_FLAG_IO_ERROR); 4743 if (hdr->b_l1hdr.b_state != arc_anon) 4744 arc_change_state(arc_anon, hdr, hash_lock); 4745 if (HDR_IN_HASH_TABLE(hdr)) 4746 buf_hash_remove(hdr); 4747 freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt); 4748 } 4749 4750 /* 4751 * Broadcast before we drop the hash_lock to avoid the possibility 4752 * that the hdr (and hence the cv) might be freed before we get to 4753 * the cv_broadcast(). 4754 */ 4755 cv_broadcast(&hdr->b_l1hdr.b_cv); 4756 4757 if (hash_lock != NULL) { 4758 mutex_exit(hash_lock); 4759 } else { 4760 /* 4761 * This block was freed while we waited for the read to 4762 * complete. It has been removed from the hash table and 4763 * moved to the anonymous state (so that it won't show up 4764 * in the cache). 4765 */ 4766 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon); 4767 freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt); 4768 } 4769 4770 /* execute each callback and free its structure */ 4771 while ((acb = callback_list) != NULL) { 4772 if (acb->acb_done) 4773 acb->acb_done(zio, acb->acb_buf, acb->acb_private); 4774 4775 if (acb->acb_zio_dummy != NULL) { 4776 acb->acb_zio_dummy->io_error = zio->io_error; 4777 zio_nowait(acb->acb_zio_dummy); 4778 } 4779 4780 callback_list = acb->acb_next; 4781 kmem_free(acb, sizeof (arc_callback_t)); 4782 } 4783 4784 if (freeable) 4785 arc_hdr_destroy(hdr); 4786} 4787 4788/* 4789 * "Read" the block at the specified DVA (in bp) via the 4790 * cache. If the block is found in the cache, invoke the provided 4791 * callback immediately and return. Note that the `zio' parameter 4792 * in the callback will be NULL in this case, since no IO was 4793 * required. If the block is not in the cache pass the read request 4794 * on to the spa with a substitute callback function, so that the 4795 * requested block will be added to the cache. 4796 * 4797 * If a read request arrives for a block that has a read in-progress, 4798 * either wait for the in-progress read to complete (and return the 4799 * results); or, if this is a read with a "done" func, add a record 4800 * to the read to invoke the "done" func when the read completes, 4801 * and return; or just return. 4802 * 4803 * arc_read_done() will invoke all the requested "done" functions 4804 * for readers of this block. 4805 */ 4806int 4807arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_done_func_t *done, 4808 void *private, zio_priority_t priority, int zio_flags, 4809 arc_flags_t *arc_flags, const zbookmark_phys_t *zb) 4810{ 4811 arc_buf_hdr_t *hdr = NULL; 4812 kmutex_t *hash_lock = NULL; 4813 zio_t *rzio; 4814 uint64_t guid = spa_load_guid(spa); 4815 4816 ASSERT(!BP_IS_EMBEDDED(bp) || 4817 BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA); 4818 4819top: 4820 if (!BP_IS_EMBEDDED(bp)) { 4821 /* 4822 * Embedded BP's have no DVA and require no I/O to "read". 4823 * Create an anonymous arc buf to back it. 4824 */ 4825 hdr = buf_hash_find(guid, bp, &hash_lock); 4826 } 4827 4828 if (hdr != NULL && HDR_HAS_L1HDR(hdr) && hdr->b_l1hdr.b_pdata != NULL) { 4829 arc_buf_t *buf = NULL; 4830 *arc_flags |= ARC_FLAG_CACHED; 4831 4832 if (HDR_IO_IN_PROGRESS(hdr)) { 4833 4834 if ((hdr->b_flags & ARC_FLAG_PRIO_ASYNC_READ) && 4835 priority == ZIO_PRIORITY_SYNC_READ) { 4836 /* 4837 * This sync read must wait for an 4838 * in-progress async read (e.g. a predictive 4839 * prefetch). Async reads are queued 4840 * separately at the vdev_queue layer, so 4841 * this is a form of priority inversion. 4842 * Ideally, we would "inherit" the demand 4843 * i/o's priority by moving the i/o from 4844 * the async queue to the synchronous queue, 4845 * but there is currently no mechanism to do 4846 * so. Track this so that we can evaluate 4847 * the magnitude of this potential performance 4848 * problem. 4849 * 4850 * Note that if the prefetch i/o is already 4851 * active (has been issued to the device), 4852 * the prefetch improved performance, because 4853 * we issued it sooner than we would have 4854 * without the prefetch. 4855 */ 4856 DTRACE_PROBE1(arc__sync__wait__for__async, 4857 arc_buf_hdr_t *, hdr); 4858 ARCSTAT_BUMP(arcstat_sync_wait_for_async); 4859 } 4860 if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) { 4861 arc_hdr_clear_flags(hdr, 4862 ARC_FLAG_PREDICTIVE_PREFETCH); 4863 } 4864 4865 if (*arc_flags & ARC_FLAG_WAIT) { 4866 cv_wait(&hdr->b_l1hdr.b_cv, hash_lock); 4867 mutex_exit(hash_lock); 4868 goto top; 4869 } 4870 ASSERT(*arc_flags & ARC_FLAG_NOWAIT); 4871 4872 if (done) { 4873 arc_callback_t *acb = NULL; 4874 4875 acb = kmem_zalloc(sizeof (arc_callback_t), 4876 KM_SLEEP); 4877 acb->acb_done = done; 4878 acb->acb_private = private; 4879 if (pio != NULL) 4880 acb->acb_zio_dummy = zio_null(pio, 4881 spa, NULL, NULL, NULL, zio_flags); 4882 4883 ASSERT3P(acb->acb_done, !=, NULL); 4884 acb->acb_next = hdr->b_l1hdr.b_acb; 4885 hdr->b_l1hdr.b_acb = acb; 4886 mutex_exit(hash_lock); 4887 return (0); 4888 } 4889 mutex_exit(hash_lock); 4890 return (0); 4891 } 4892 4893 ASSERT(hdr->b_l1hdr.b_state == arc_mru || 4894 hdr->b_l1hdr.b_state == arc_mfu); 4895 4896 if (done) { 4897 if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) { 4898 /* 4899 * This is a demand read which does not have to 4900 * wait for i/o because we did a predictive 4901 * prefetch i/o for it, which has completed. 4902 */ 4903 DTRACE_PROBE1( 4904 arc__demand__hit__predictive__prefetch, 4905 arc_buf_hdr_t *, hdr); 4906 ARCSTAT_BUMP( 4907 arcstat_demand_hit_predictive_prefetch); 4908 arc_hdr_clear_flags(hdr, 4909 ARC_FLAG_PREDICTIVE_PREFETCH); 4910 } 4911 ASSERT(!BP_IS_EMBEDDED(bp) || !BP_IS_HOLE(bp)); 4912 4913 /* 4914 * If this block is already in use, create a new 4915 * copy of the data so that we will be guaranteed 4916 * that arc_release() will always succeed. 4917 */ 4918 buf = hdr->b_l1hdr.b_buf; 4919 if (buf == NULL) { 4920 ASSERT0(refcount_count(&hdr->b_l1hdr.b_refcnt)); 4921 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL); 4922 buf = arc_buf_alloc_impl(hdr, private); 4923 VERIFY0(arc_decompress(buf)); 4924 } else { 4925 add_reference(hdr, private); 4926 buf = arc_buf_clone(buf); 4927 } 4928 ASSERT3P(buf->b_data, !=, NULL); 4929 4930 } else if (*arc_flags & ARC_FLAG_PREFETCH && 4931 refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) { 4932 arc_hdr_set_flags(hdr, ARC_FLAG_PREFETCH); 4933 } 4934 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr); 4935 arc_access(hdr, hash_lock); 4936 if (*arc_flags & ARC_FLAG_L2CACHE) 4937 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE); 4938 mutex_exit(hash_lock); 4939 ARCSTAT_BUMP(arcstat_hits); 4940 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr), 4941 demand, prefetch, !HDR_ISTYPE_METADATA(hdr), 4942 data, metadata, hits); 4943 4944 if (done) 4945 done(NULL, buf, private); 4946 } else { 4947 uint64_t lsize = BP_GET_LSIZE(bp); 4948 uint64_t psize = BP_GET_PSIZE(bp); 4949 arc_callback_t *acb; 4950 vdev_t *vd = NULL; 4951 uint64_t addr = 0; 4952 boolean_t devw = B_FALSE; 4953 uint64_t size; 4954 4955 if (hdr == NULL) { 4956 /* this block is not in the cache */ 4957 arc_buf_hdr_t *exists = NULL; 4958 arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp); 4959 hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize, 4960 BP_GET_COMPRESS(bp), type); 4961 4962 if (!BP_IS_EMBEDDED(bp)) { 4963 hdr->b_dva = *BP_IDENTITY(bp); 4964 hdr->b_birth = BP_PHYSICAL_BIRTH(bp); 4965 exists = buf_hash_insert(hdr, &hash_lock); 4966 } 4967 if (exists != NULL) { 4968 /* somebody beat us to the hash insert */ 4969 mutex_exit(hash_lock); 4970 buf_discard_identity(hdr); 4971 arc_hdr_destroy(hdr); 4972 goto top; /* restart the IO request */ 4973 } 4974 } else { 4975 /* 4976 * This block is in the ghost cache. If it was L2-only 4977 * (and thus didn't have an L1 hdr), we realloc the 4978 * header to add an L1 hdr. 4979 */ 4980 if (!HDR_HAS_L1HDR(hdr)) { 4981 hdr = arc_hdr_realloc(hdr, hdr_l2only_cache, 4982 hdr_full_cache); 4983 } 4984 ASSERT3P(hdr->b_l1hdr.b_pdata, ==, NULL); 4985 ASSERT(GHOST_STATE(hdr->b_l1hdr.b_state)); 4986 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 4987 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 4988 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 4989 4990 /* 4991 * This is a delicate dance that we play here. 4992 * This hdr is in the ghost list so we access it 4993 * to move it out of the ghost list before we 4994 * initiate the read. If it's a prefetch then 4995 * it won't have a callback so we'll remove the 4996 * reference that arc_buf_alloc_impl() created. We 4997 * do this after we've called arc_access() to 4998 * avoid hitting an assert in remove_reference(). 4999 */ 5000 arc_access(hdr, hash_lock); 5001 arc_hdr_alloc_pdata(hdr); 5002 } 5003 ASSERT3P(hdr->b_l1hdr.b_pdata, !=, NULL); 5004 size = arc_hdr_size(hdr); 5005 5006 /* 5007 * If compression is enabled on the hdr, then will do 5008 * RAW I/O and will store the compressed data in the hdr's 5009 * data block. Otherwise, the hdr's data block will contain 5010 * the uncompressed data. 5011 */ 5012 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF) { 5013 zio_flags |= ZIO_FLAG_RAW; 5014 } 5015 5016 if (*arc_flags & ARC_FLAG_PREFETCH) 5017 arc_hdr_set_flags(hdr, ARC_FLAG_PREFETCH); 5018 if (*arc_flags & ARC_FLAG_L2CACHE) 5019 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE); 5020 if (BP_GET_LEVEL(bp) > 0) 5021 arc_hdr_set_flags(hdr, ARC_FLAG_INDIRECT); 5022 if (*arc_flags & ARC_FLAG_PREDICTIVE_PREFETCH) 5023 arc_hdr_set_flags(hdr, ARC_FLAG_PREDICTIVE_PREFETCH); 5024 ASSERT(!GHOST_STATE(hdr->b_l1hdr.b_state)); 5025 5026 acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP); 5027 acb->acb_done = done; 5028 acb->acb_private = private; 5029 5030 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL); 5031 hdr->b_l1hdr.b_acb = acb; 5032 arc_hdr_set_flags(hdr, ARC_FLAG_IO_IN_PROGRESS); 5033 5034 if (HDR_HAS_L2HDR(hdr) && 5035 (vd = hdr->b_l2hdr.b_dev->l2ad_vdev) != NULL) { 5036 devw = hdr->b_l2hdr.b_dev->l2ad_writing; 5037 addr = hdr->b_l2hdr.b_daddr; 5038 /* 5039 * Lock out device removal. 5040 */ 5041 if (vdev_is_dead(vd) || 5042 !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER)) 5043 vd = NULL; 5044 } 5045 5046 if (priority == ZIO_PRIORITY_ASYNC_READ) 5047 arc_hdr_set_flags(hdr, ARC_FLAG_PRIO_ASYNC_READ); 5048 else 5049 arc_hdr_clear_flags(hdr, ARC_FLAG_PRIO_ASYNC_READ); 5050 5051 if (hash_lock != NULL) 5052 mutex_exit(hash_lock); 5053 5054 /* 5055 * At this point, we have a level 1 cache miss. Try again in 5056 * L2ARC if possible. 5057 */ 5058 ASSERT3U(HDR_GET_LSIZE(hdr), ==, lsize); 5059 5060 DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp, 5061 uint64_t, lsize, zbookmark_phys_t *, zb); 5062 ARCSTAT_BUMP(arcstat_misses); 5063 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr), 5064 demand, prefetch, !HDR_ISTYPE_METADATA(hdr), 5065 data, metadata, misses); 5066#ifdef _KERNEL 5067#ifdef RACCT 5068 if (racct_enable) { 5069 PROC_LOCK(curproc); 5070 racct_add_force(curproc, RACCT_READBPS, size); 5071 racct_add_force(curproc, RACCT_READIOPS, 1); 5072 PROC_UNLOCK(curproc); 5073 } 5074#endif /* RACCT */ 5075 curthread->td_ru.ru_inblock++; 5076#endif 5077 5078 if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) { 5079 /* 5080 * Read from the L2ARC if the following are true: 5081 * 1. The L2ARC vdev was previously cached. 5082 * 2. This buffer still has L2ARC metadata. 5083 * 3. This buffer isn't currently writing to the L2ARC. 5084 * 4. The L2ARC entry wasn't evicted, which may 5085 * also have invalidated the vdev. 5086 * 5. This isn't prefetch and l2arc_noprefetch is set. 5087 */ 5088 if (HDR_HAS_L2HDR(hdr) && 5089 !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) && 5090 !(l2arc_noprefetch && HDR_PREFETCH(hdr))) { 5091 l2arc_read_callback_t *cb; 5092 void* b_data; 5093 5094 DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr); 5095 ARCSTAT_BUMP(arcstat_l2_hits); 5096 5097 cb = kmem_zalloc(sizeof (l2arc_read_callback_t), 5098 KM_SLEEP); 5099 cb->l2rcb_hdr = hdr; 5100 cb->l2rcb_bp = *bp; 5101 cb->l2rcb_zb = *zb; 5102 cb->l2rcb_flags = zio_flags; 5103 uint64_t asize = vdev_psize_to_asize(vd, size); 5104 if (asize != size) { 5105 b_data = zio_data_buf_alloc(asize); 5106 cb->l2rcb_data = b_data; 5107 } else { 5108 b_data = hdr->b_l1hdr.b_pdata; 5109 } 5110 5111 ASSERT(addr >= VDEV_LABEL_START_SIZE && 5112 addr + asize < vd->vdev_psize - 5113 VDEV_LABEL_END_SIZE); 5114 5115 /* 5116 * l2arc read. The SCL_L2ARC lock will be 5117 * released by l2arc_read_done(). 5118 * Issue a null zio if the underlying buffer 5119 * was squashed to zero size by compression. 5120 */ 5121 ASSERT3U(HDR_GET_COMPRESS(hdr), !=, 5122 ZIO_COMPRESS_EMPTY); 5123 rzio = zio_read_phys(pio, vd, addr, 5124 asize, b_data, 5125 ZIO_CHECKSUM_OFF, 5126 l2arc_read_done, cb, priority, 5127 zio_flags | ZIO_FLAG_DONT_CACHE | 5128 ZIO_FLAG_CANFAIL | 5129 ZIO_FLAG_DONT_PROPAGATE | 5130 ZIO_FLAG_DONT_RETRY, B_FALSE); 5131 DTRACE_PROBE2(l2arc__read, vdev_t *, vd, 5132 zio_t *, rzio); 5133 ARCSTAT_INCR(arcstat_l2_read_bytes, size); 5134 5135 if (*arc_flags & ARC_FLAG_NOWAIT) { 5136 zio_nowait(rzio); 5137 return (0); 5138 } 5139 5140 ASSERT(*arc_flags & ARC_FLAG_WAIT); 5141 if (zio_wait(rzio) == 0) 5142 return (0); 5143 5144 /* l2arc read error; goto zio_read() */ 5145 } else { 5146 DTRACE_PROBE1(l2arc__miss, 5147 arc_buf_hdr_t *, hdr); 5148 ARCSTAT_BUMP(arcstat_l2_misses); 5149 if (HDR_L2_WRITING(hdr)) 5150 ARCSTAT_BUMP(arcstat_l2_rw_clash); 5151 spa_config_exit(spa, SCL_L2ARC, vd); 5152 } 5153 } else { 5154 if (vd != NULL) 5155 spa_config_exit(spa, SCL_L2ARC, vd); 5156 if (l2arc_ndev != 0) { 5157 DTRACE_PROBE1(l2arc__miss, 5158 arc_buf_hdr_t *, hdr); 5159 ARCSTAT_BUMP(arcstat_l2_misses); 5160 } 5161 } 5162 5163 rzio = zio_read(pio, spa, bp, hdr->b_l1hdr.b_pdata, size, 5164 arc_read_done, hdr, priority, zio_flags, zb); 5165 5166 if (*arc_flags & ARC_FLAG_WAIT) 5167 return (zio_wait(rzio)); 5168 5169 ASSERT(*arc_flags & ARC_FLAG_NOWAIT); 5170 zio_nowait(rzio); 5171 } 5172 return (0); 5173} 5174 5175/* 5176 * Notify the arc that a block was freed, and thus will never be used again. 5177 */ 5178void 5179arc_freed(spa_t *spa, const blkptr_t *bp) 5180{ 5181 arc_buf_hdr_t *hdr; 5182 kmutex_t *hash_lock; 5183 uint64_t guid = spa_load_guid(spa); 5184 5185 ASSERT(!BP_IS_EMBEDDED(bp)); 5186 5187 hdr = buf_hash_find(guid, bp, &hash_lock); 5188 if (hdr == NULL) 5189 return; 5190 5191 /* 5192 * We might be trying to free a block that is still doing I/O 5193 * (i.e. prefetch) or has a reference (i.e. a dedup-ed, 5194 * dmu_sync-ed block). If this block is being prefetched, then it 5195 * would still have the ARC_FLAG_IO_IN_PROGRESS flag set on the hdr 5196 * until the I/O completes. A block may also have a reference if it is 5197 * part of a dedup-ed, dmu_synced write. The dmu_sync() function would 5198 * have written the new block to its final resting place on disk but 5199 * without the dedup flag set. This would have left the hdr in the MRU 5200 * state and discoverable. When the txg finally syncs it detects that 5201 * the block was overridden in open context and issues an override I/O. 5202 * Since this is a dedup block, the override I/O will determine if the 5203 * block is already in the DDT. If so, then it will replace the io_bp 5204 * with the bp from the DDT and allow the I/O to finish. When the I/O 5205 * reaches the done callback, dbuf_write_override_done, it will 5206 * check to see if the io_bp and io_bp_override are identical. 5207 * If they are not, then it indicates that the bp was replaced with 5208 * the bp in the DDT and the override bp is freed. This allows 5209 * us to arrive here with a reference on a block that is being 5210 * freed. So if we have an I/O in progress, or a reference to 5211 * this hdr, then we don't destroy the hdr. 5212 */ 5213 if (!HDR_HAS_L1HDR(hdr) || (!HDR_IO_IN_PROGRESS(hdr) && 5214 refcount_is_zero(&hdr->b_l1hdr.b_refcnt))) { 5215 arc_change_state(arc_anon, hdr, hash_lock); 5216 arc_hdr_destroy(hdr); 5217 mutex_exit(hash_lock); 5218 } else { 5219 mutex_exit(hash_lock); 5220 } 5221 5222} 5223 5224/* 5225 * Release this buffer from the cache, making it an anonymous buffer. This 5226 * must be done after a read and prior to modifying the buffer contents. 5227 * If the buffer has more than one reference, we must make 5228 * a new hdr for the buffer. 5229 */ 5230void 5231arc_release(arc_buf_t *buf, void *tag) 5232{ 5233 arc_buf_hdr_t *hdr = buf->b_hdr; 5234 5235 /* 5236 * It would be nice to assert that if it's DMU metadata (level > 5237 * 0 || it's the dnode file), then it must be syncing context. 5238 * But we don't know that information at this level. 5239 */ 5240 5241 mutex_enter(&buf->b_evict_lock); 5242 5243 ASSERT(HDR_HAS_L1HDR(hdr)); 5244 5245 /* 5246 * We don't grab the hash lock prior to this check, because if 5247 * the buffer's header is in the arc_anon state, it won't be 5248 * linked into the hash table. 5249 */ 5250 if (hdr->b_l1hdr.b_state == arc_anon) { 5251 mutex_exit(&buf->b_evict_lock); 5252 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 5253 ASSERT(!HDR_IN_HASH_TABLE(hdr)); 5254 ASSERT(!HDR_HAS_L2HDR(hdr)); 5255 ASSERT(HDR_EMPTY(hdr)); 5256 ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1); 5257 ASSERT3S(refcount_count(&hdr->b_l1hdr.b_refcnt), ==, 1); 5258 ASSERT(!list_link_active(&hdr->b_l1hdr.b_arc_node)); 5259 5260 hdr->b_l1hdr.b_arc_access = 0; 5261 5262 /* 5263 * If the buf is being overridden then it may already 5264 * have a hdr that is not empty. 5265 */ 5266 buf_discard_identity(hdr); 5267 arc_buf_thaw(buf); 5268 5269 return; 5270 } 5271 5272 kmutex_t *hash_lock = HDR_LOCK(hdr); 5273 mutex_enter(hash_lock); 5274 5275 /* 5276 * This assignment is only valid as long as the hash_lock is 5277 * held, we must be careful not to reference state or the 5278 * b_state field after dropping the lock. 5279 */ 5280 arc_state_t *state = hdr->b_l1hdr.b_state; 5281 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 5282 ASSERT3P(state, !=, arc_anon); 5283 5284 /* this buffer is not on any list */ 5285 ASSERT(refcount_count(&hdr->b_l1hdr.b_refcnt) > 0); 5286 5287 if (HDR_HAS_L2HDR(hdr)) { 5288 mutex_enter(&hdr->b_l2hdr.b_dev->l2ad_mtx); 5289 5290 /* 5291 * We have to recheck this conditional again now that 5292 * we're holding the l2ad_mtx to prevent a race with 5293 * another thread which might be concurrently calling 5294 * l2arc_evict(). In that case, l2arc_evict() might have 5295 * destroyed the header's L2 portion as we were waiting 5296 * to acquire the l2ad_mtx. 5297 */ 5298 if (HDR_HAS_L2HDR(hdr)) { 5299 l2arc_trim(hdr); 5300 arc_hdr_l2hdr_destroy(hdr); 5301 } 5302 5303 mutex_exit(&hdr->b_l2hdr.b_dev->l2ad_mtx); 5304 } 5305 5306 /* 5307 * Do we have more than one buf? 5308 */ 5309 if (hdr->b_l1hdr.b_bufcnt > 1) { 5310 arc_buf_hdr_t *nhdr; 5311 arc_buf_t **bufp; 5312 uint64_t spa = hdr->b_spa; 5313 uint64_t psize = HDR_GET_PSIZE(hdr); 5314 uint64_t lsize = HDR_GET_LSIZE(hdr); 5315 enum zio_compress compress = HDR_GET_COMPRESS(hdr); 5316 arc_buf_contents_t type = arc_buf_type(hdr); 5317 VERIFY3U(hdr->b_type, ==, type); 5318 5319 ASSERT(hdr->b_l1hdr.b_buf != buf || buf->b_next != NULL); 5320 (void) remove_reference(hdr, hash_lock, tag); 5321 5322 if (arc_buf_is_shared(buf)) { 5323 ASSERT(HDR_SHARED_DATA(hdr)); 5324 ASSERT3P(hdr->b_l1hdr.b_buf, !=, buf); 5325 ASSERT(ARC_BUF_LAST(buf)); 5326 } 5327 5328 /* 5329 * Pull the data off of this hdr and attach it to 5330 * a new anonymous hdr. Also find the last buffer 5331 * in the hdr's buffer list. 5332 */ 5333 arc_buf_t *lastbuf = NULL; 5334 bufp = &hdr->b_l1hdr.b_buf; 5335 while (*bufp != NULL) { 5336 if (*bufp == buf) { 5337 *bufp = buf->b_next; 5338 } 5339 5340 /* 5341 * If we've removed a buffer in the middle of 5342 * the list then update the lastbuf and update 5343 * bufp. 5344 */ 5345 if (*bufp != NULL) { 5346 lastbuf = *bufp; 5347 bufp = &(*bufp)->b_next; 5348 } 5349 } 5350 buf->b_next = NULL; 5351 ASSERT3P(lastbuf, !=, buf); 5352 ASSERT3P(lastbuf, !=, NULL); 5353 5354 /* 5355 * If the current arc_buf_t and the hdr are sharing their data 5356 * buffer, then we must stop sharing that block, transfer 5357 * ownership and setup sharing with a new arc_buf_t at the end 5358 * of the hdr's b_buf list. 5359 */ 5360 if (arc_buf_is_shared(buf)) { 5361 ASSERT3P(hdr->b_l1hdr.b_buf, !=, buf); 5362 ASSERT(ARC_BUF_LAST(lastbuf)); 5363 VERIFY(!arc_buf_is_shared(lastbuf)); 5364 5365 /* 5366 * First, sever the block sharing relationship between 5367 * buf and the arc_buf_hdr_t. Then, setup a new 5368 * block sharing relationship with the last buffer 5369 * on the arc_buf_t list. 5370 */ 5371 arc_unshare_buf(hdr, buf); 5372 arc_share_buf(hdr, lastbuf); 5373 VERIFY3P(lastbuf->b_data, !=, NULL); 5374 } else if (HDR_SHARED_DATA(hdr)) { 5375 ASSERT(arc_buf_is_shared(lastbuf)); 5376 } 5377 ASSERT3P(hdr->b_l1hdr.b_pdata, !=, NULL); 5378 ASSERT3P(state, !=, arc_l2c_only); 5379 5380 (void) refcount_remove_many(&state->arcs_size, 5381 HDR_GET_LSIZE(hdr), buf); 5382 5383 if (refcount_is_zero(&hdr->b_l1hdr.b_refcnt)) { 5384 ASSERT3P(state, !=, arc_l2c_only); 5385 (void) refcount_remove_many(&state->arcs_esize[type], 5386 HDR_GET_LSIZE(hdr), buf); 5387 } 5388 5389 hdr->b_l1hdr.b_bufcnt -= 1; 5390 arc_cksum_verify(buf); 5391#ifdef illumos 5392 arc_buf_unwatch(buf); 5393#endif 5394 5395 mutex_exit(hash_lock); 5396 5397 /* 5398 * Allocate a new hdr. The new hdr will contain a b_pdata 5399 * buffer which will be freed in arc_write(). 5400 */ 5401 nhdr = arc_hdr_alloc(spa, psize, lsize, compress, type); 5402 ASSERT3P(nhdr->b_l1hdr.b_buf, ==, NULL); 5403 ASSERT0(nhdr->b_l1hdr.b_bufcnt); 5404 ASSERT0(refcount_count(&nhdr->b_l1hdr.b_refcnt)); 5405 VERIFY3U(nhdr->b_type, ==, type); 5406 ASSERT(!HDR_SHARED_DATA(nhdr)); 5407 5408 nhdr->b_l1hdr.b_buf = buf; 5409 nhdr->b_l1hdr.b_bufcnt = 1; 5410 (void) refcount_add(&nhdr->b_l1hdr.b_refcnt, tag); 5411 buf->b_hdr = nhdr; 5412 5413 mutex_exit(&buf->b_evict_lock); 5414 (void) refcount_add_many(&arc_anon->arcs_size, 5415 HDR_GET_LSIZE(nhdr), buf); 5416 } else { 5417 mutex_exit(&buf->b_evict_lock); 5418 ASSERT(refcount_count(&hdr->b_l1hdr.b_refcnt) == 1); 5419 /* protected by hash lock, or hdr is on arc_anon */ 5420 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 5421 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 5422 arc_change_state(arc_anon, hdr, hash_lock); 5423 hdr->b_l1hdr.b_arc_access = 0; 5424 mutex_exit(hash_lock); 5425 5426 buf_discard_identity(hdr); 5427 arc_buf_thaw(buf); 5428 } 5429} 5430 5431int 5432arc_released(arc_buf_t *buf) 5433{ 5434 int released; 5435 5436 mutex_enter(&buf->b_evict_lock); 5437 released = (buf->b_data != NULL && 5438 buf->b_hdr->b_l1hdr.b_state == arc_anon); 5439 mutex_exit(&buf->b_evict_lock); 5440 return (released); 5441} 5442 5443#ifdef ZFS_DEBUG 5444int 5445arc_referenced(arc_buf_t *buf) 5446{ 5447 int referenced; 5448 5449 mutex_enter(&buf->b_evict_lock); 5450 referenced = (refcount_count(&buf->b_hdr->b_l1hdr.b_refcnt)); 5451 mutex_exit(&buf->b_evict_lock); 5452 return (referenced); 5453} 5454#endif 5455 5456static void 5457arc_write_ready(zio_t *zio) 5458{ 5459 arc_write_callback_t *callback = zio->io_private; 5460 arc_buf_t *buf = callback->awcb_buf; 5461 arc_buf_hdr_t *hdr = buf->b_hdr; 5462 uint64_t psize = BP_IS_HOLE(zio->io_bp) ? 0 : BP_GET_PSIZE(zio->io_bp); 5463 5464 ASSERT(HDR_HAS_L1HDR(hdr)); 5465 ASSERT(!refcount_is_zero(&buf->b_hdr->b_l1hdr.b_refcnt)); 5466 ASSERT(hdr->b_l1hdr.b_bufcnt > 0); 5467 5468 /* 5469 * If we're reexecuting this zio because the pool suspended, then 5470 * cleanup any state that was previously set the first time the 5471 * callback as invoked. 5472 */ 5473 if (zio->io_flags & ZIO_FLAG_REEXECUTED) { 5474 arc_cksum_free(hdr); 5475#ifdef illumos 5476 arc_buf_unwatch(buf); 5477#endif 5478 if (hdr->b_l1hdr.b_pdata != NULL) { 5479 if (arc_buf_is_shared(buf)) { 5480 ASSERT(HDR_SHARED_DATA(hdr)); 5481 5482 arc_unshare_buf(hdr, buf); 5483 } else { 5484 arc_hdr_free_pdata(hdr); 5485 } 5486 } 5487 } 5488 ASSERT3P(hdr->b_l1hdr.b_pdata, ==, NULL); 5489 ASSERT(!HDR_SHARED_DATA(hdr)); 5490 ASSERT(!arc_buf_is_shared(buf)); 5491 5492 callback->awcb_ready(zio, buf, callback->awcb_private); 5493 5494 if (HDR_IO_IN_PROGRESS(hdr)) 5495 ASSERT(zio->io_flags & ZIO_FLAG_REEXECUTED); 5496 5497 arc_cksum_compute(buf); 5498 arc_hdr_set_flags(hdr, ARC_FLAG_IO_IN_PROGRESS); 5499 5500 enum zio_compress compress; 5501 if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) { 5502 compress = ZIO_COMPRESS_OFF; 5503 } else { 5504 ASSERT3U(HDR_GET_LSIZE(hdr), ==, BP_GET_LSIZE(zio->io_bp)); 5505 compress = BP_GET_COMPRESS(zio->io_bp); 5506 } 5507 HDR_SET_PSIZE(hdr, psize); 5508 arc_hdr_set_compress(hdr, compress); 5509 5510 /* 5511 * If the hdr is compressed, then copy the compressed 5512 * zio contents into arc_buf_hdr_t. Otherwise, copy the original 5513 * data buf into the hdr. Ideally, we would like to always copy the 5514 * io_data into b_pdata but the user may have disabled compressed 5515 * arc thus the on-disk block may or may not match what we maintain 5516 * in the hdr's b_pdata field. 5517 */ 5518 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF) { 5519 ASSERT(BP_GET_COMPRESS(zio->io_bp) != ZIO_COMPRESS_OFF); 5520 ASSERT3U(psize, >, 0); 5521 arc_hdr_alloc_pdata(hdr); 5522 bcopy(zio->io_data, hdr->b_l1hdr.b_pdata, psize); 5523 } else { 5524 ASSERT3P(buf->b_data, ==, zio->io_orig_data); 5525 ASSERT3U(zio->io_orig_size, ==, HDR_GET_LSIZE(hdr)); 5526 ASSERT3U(hdr->b_l1hdr.b_byteswap, ==, DMU_BSWAP_NUMFUNCS); 5527 ASSERT(!HDR_SHARED_DATA(hdr)); 5528 ASSERT(!arc_buf_is_shared(buf)); 5529 ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1); 5530 ASSERT3P(hdr->b_l1hdr.b_pdata, ==, NULL); 5531 5532 /* 5533 * This hdr is not compressed so we're able to share 5534 * the arc_buf_t data buffer with the hdr. 5535 */ 5536 arc_share_buf(hdr, buf); 5537 VERIFY0(bcmp(zio->io_orig_data, hdr->b_l1hdr.b_pdata, 5538 HDR_GET_LSIZE(hdr))); 5539 } 5540 arc_hdr_verify(hdr, zio->io_bp); 5541} 5542 5543static void 5544arc_write_children_ready(zio_t *zio) 5545{ 5546 arc_write_callback_t *callback = zio->io_private; 5547 arc_buf_t *buf = callback->awcb_buf; 5548 5549 callback->awcb_children_ready(zio, buf, callback->awcb_private); 5550} 5551 5552/* 5553 * The SPA calls this callback for each physical write that happens on behalf 5554 * of a logical write. See the comment in dbuf_write_physdone() for details. 5555 */ 5556static void 5557arc_write_physdone(zio_t *zio) 5558{ 5559 arc_write_callback_t *cb = zio->io_private; 5560 if (cb->awcb_physdone != NULL) 5561 cb->awcb_physdone(zio, cb->awcb_buf, cb->awcb_private); 5562} 5563 5564static void 5565arc_write_done(zio_t *zio) 5566{ 5567 arc_write_callback_t *callback = zio->io_private; 5568 arc_buf_t *buf = callback->awcb_buf; 5569 arc_buf_hdr_t *hdr = buf->b_hdr; 5570 5571 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL); 5572 5573 if (zio->io_error == 0) { 5574 arc_hdr_verify(hdr, zio->io_bp); 5575 5576 if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) { 5577 buf_discard_identity(hdr); 5578 } else { 5579 hdr->b_dva = *BP_IDENTITY(zio->io_bp); 5580 hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp); 5581 } 5582 } else { 5583 ASSERT(HDR_EMPTY(hdr)); 5584 } 5585 5586 /* 5587 * If the block to be written was all-zero or compressed enough to be 5588 * embedded in the BP, no write was performed so there will be no 5589 * dva/birth/checksum. The buffer must therefore remain anonymous 5590 * (and uncached). 5591 */ 5592 if (!HDR_EMPTY(hdr)) { 5593 arc_buf_hdr_t *exists; 5594 kmutex_t *hash_lock; 5595 5596 ASSERT(zio->io_error == 0); 5597 5598 arc_cksum_verify(buf); 5599 5600 exists = buf_hash_insert(hdr, &hash_lock); 5601 if (exists != NULL) { 5602 /* 5603 * This can only happen if we overwrite for 5604 * sync-to-convergence, because we remove 5605 * buffers from the hash table when we arc_free(). 5606 */ 5607 if (zio->io_flags & ZIO_FLAG_IO_REWRITE) { 5608 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) 5609 panic("bad overwrite, hdr=%p exists=%p", 5610 (void *)hdr, (void *)exists); 5611 ASSERT(refcount_is_zero( 5612 &exists->b_l1hdr.b_refcnt)); 5613 arc_change_state(arc_anon, exists, hash_lock); 5614 mutex_exit(hash_lock); 5615 arc_hdr_destroy(exists); 5616 exists = buf_hash_insert(hdr, &hash_lock); 5617 ASSERT3P(exists, ==, NULL); 5618 } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) { 5619 /* nopwrite */ 5620 ASSERT(zio->io_prop.zp_nopwrite); 5621 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) 5622 panic("bad nopwrite, hdr=%p exists=%p", 5623 (void *)hdr, (void *)exists); 5624 } else { 5625 /* Dedup */ 5626 ASSERT(hdr->b_l1hdr.b_bufcnt == 1); 5627 ASSERT(hdr->b_l1hdr.b_state == arc_anon); 5628 ASSERT(BP_GET_DEDUP(zio->io_bp)); 5629 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0); 5630 } 5631 } 5632 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS); 5633 /* if it's not anon, we are doing a scrub */ 5634 if (exists == NULL && hdr->b_l1hdr.b_state == arc_anon) 5635 arc_access(hdr, hash_lock); 5636 mutex_exit(hash_lock); 5637 } else { 5638 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS); 5639 } 5640 5641 ASSERT(!refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 5642 callback->awcb_done(zio, buf, callback->awcb_private); 5643 5644 kmem_free(callback, sizeof (arc_write_callback_t)); 5645} 5646 5647zio_t * 5648arc_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, arc_buf_t *buf, 5649 boolean_t l2arc, const zio_prop_t *zp, arc_done_func_t *ready, 5650 arc_done_func_t *children_ready, arc_done_func_t *physdone, 5651 arc_done_func_t *done, void *private, zio_priority_t priority, 5652 int zio_flags, const zbookmark_phys_t *zb) 5653{ 5654 arc_buf_hdr_t *hdr = buf->b_hdr; 5655 arc_write_callback_t *callback; 5656 zio_t *zio; 5657 5658 ASSERT3P(ready, !=, NULL); 5659 ASSERT3P(done, !=, NULL); 5660 ASSERT(!HDR_IO_ERROR(hdr)); 5661 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 5662 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL); 5663 ASSERT3U(hdr->b_l1hdr.b_bufcnt, >, 0); 5664 if (l2arc) 5665 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE); 5666 callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP); 5667 callback->awcb_ready = ready; 5668 callback->awcb_children_ready = children_ready; 5669 callback->awcb_physdone = physdone; 5670 callback->awcb_done = done; 5671 callback->awcb_private = private; 5672 callback->awcb_buf = buf; 5673 5674 /* 5675 * The hdr's b_pdata is now stale, free it now. A new data block 5676 * will be allocated when the zio pipeline calls arc_write_ready(). 5677 */ 5678 if (hdr->b_l1hdr.b_pdata != NULL) { 5679 /* 5680 * If the buf is currently sharing the data block with 5681 * the hdr then we need to break that relationship here. 5682 * The hdr will remain with a NULL data pointer and the 5683 * buf will take sole ownership of the block. 5684 */ 5685 if (arc_buf_is_shared(buf)) { 5686 ASSERT(ARC_BUF_LAST(buf)); 5687 arc_unshare_buf(hdr, buf); 5688 } else { 5689 arc_hdr_free_pdata(hdr); 5690 } 5691 VERIFY3P(buf->b_data, !=, NULL); 5692 arc_hdr_set_compress(hdr, ZIO_COMPRESS_OFF); 5693 } 5694 ASSERT(!arc_buf_is_shared(buf)); 5695 ASSERT3P(hdr->b_l1hdr.b_pdata, ==, NULL); 5696 5697 zio = zio_write(pio, spa, txg, bp, buf->b_data, HDR_GET_LSIZE(hdr), zp, 5698 arc_write_ready, 5699 (children_ready != NULL) ? arc_write_children_ready : NULL, 5700 arc_write_physdone, arc_write_done, callback, 5701 priority, zio_flags, zb); 5702 5703 return (zio); 5704} 5705 5706static int 5707arc_memory_throttle(uint64_t reserve, uint64_t txg) 5708{ 5709#ifdef _KERNEL 5710 uint64_t available_memory = ptob(freemem); 5711 static uint64_t page_load = 0; 5712 static uint64_t last_txg = 0; 5713 5714#if defined(__i386) || !defined(UMA_MD_SMALL_ALLOC) 5715 available_memory = 5716 MIN(available_memory, ptob(vmem_size(heap_arena, VMEM_FREE))); 5717#endif 5718 5719 if (freemem > (uint64_t)physmem * arc_lotsfree_percent / 100) 5720 return (0); 5721 5722 if (txg > last_txg) { 5723 last_txg = txg; 5724 page_load = 0; 5725 } 5726 /* 5727 * If we are in pageout, we know that memory is already tight, 5728 * the arc is already going to be evicting, so we just want to 5729 * continue to let page writes occur as quickly as possible. 5730 */ 5731 if (curproc == pageproc) { 5732 if (page_load > MAX(ptob(minfree), available_memory) / 4) 5733 return (SET_ERROR(ERESTART)); 5734 /* Note: reserve is inflated, so we deflate */ 5735 page_load += reserve / 8; 5736 return (0); 5737 } else if (page_load > 0 && arc_reclaim_needed()) { 5738 /* memory is low, delay before restarting */ 5739 ARCSTAT_INCR(arcstat_memory_throttle_count, 1); 5740 return (SET_ERROR(EAGAIN)); 5741 } 5742 page_load = 0; 5743#endif 5744 return (0); 5745} 5746 5747void 5748arc_tempreserve_clear(uint64_t reserve) 5749{ 5750 atomic_add_64(&arc_tempreserve, -reserve); 5751 ASSERT((int64_t)arc_tempreserve >= 0); 5752} 5753 5754int 5755arc_tempreserve_space(uint64_t reserve, uint64_t txg) 5756{ 5757 int error; 5758 uint64_t anon_size; 5759 5760 if (reserve > arc_c/4 && !arc_no_grow) { 5761 arc_c = MIN(arc_c_max, reserve * 4); 5762 DTRACE_PROBE1(arc__set_reserve, uint64_t, arc_c); 5763 } 5764 if (reserve > arc_c) 5765 return (SET_ERROR(ENOMEM)); 5766 5767 /* 5768 * Don't count loaned bufs as in flight dirty data to prevent long 5769 * network delays from blocking transactions that are ready to be 5770 * assigned to a txg. 5771 */ 5772 anon_size = MAX((int64_t)(refcount_count(&arc_anon->arcs_size) - 5773 arc_loaned_bytes), 0); 5774 5775 /* 5776 * Writes will, almost always, require additional memory allocations 5777 * in order to compress/encrypt/etc the data. We therefore need to 5778 * make sure that there is sufficient available memory for this. 5779 */ 5780 error = arc_memory_throttle(reserve, txg); 5781 if (error != 0) 5782 return (error); 5783 5784 /* 5785 * Throttle writes when the amount of dirty data in the cache 5786 * gets too large. We try to keep the cache less than half full 5787 * of dirty blocks so that our sync times don't grow too large. 5788 * Note: if two requests come in concurrently, we might let them 5789 * both succeed, when one of them should fail. Not a huge deal. 5790 */ 5791 5792 if (reserve + arc_tempreserve + anon_size > arc_c / 2 && 5793 anon_size > arc_c / 4) { 5794 uint64_t meta_esize = 5795 refcount_count(&arc_anon->arcs_esize[ARC_BUFC_METADATA]); 5796 uint64_t data_esize = 5797 refcount_count(&arc_anon->arcs_esize[ARC_BUFC_DATA]); 5798 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK " 5799 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n", 5800 arc_tempreserve >> 10, meta_esize >> 10, 5801 data_esize >> 10, reserve >> 10, arc_c >> 10); 5802 return (SET_ERROR(ERESTART)); 5803 } 5804 atomic_add_64(&arc_tempreserve, reserve); 5805 return (0); 5806} 5807 5808static void 5809arc_kstat_update_state(arc_state_t *state, kstat_named_t *size, 5810 kstat_named_t *evict_data, kstat_named_t *evict_metadata) 5811{ 5812 size->value.ui64 = refcount_count(&state->arcs_size); 5813 evict_data->value.ui64 = 5814 refcount_count(&state->arcs_esize[ARC_BUFC_DATA]); 5815 evict_metadata->value.ui64 = 5816 refcount_count(&state->arcs_esize[ARC_BUFC_METADATA]); 5817} 5818 5819static int 5820arc_kstat_update(kstat_t *ksp, int rw) 5821{ 5822 arc_stats_t *as = ksp->ks_data; 5823 5824 if (rw == KSTAT_WRITE) { 5825 return (EACCES); 5826 } else { 5827 arc_kstat_update_state(arc_anon, 5828 &as->arcstat_anon_size, 5829 &as->arcstat_anon_evictable_data, 5830 &as->arcstat_anon_evictable_metadata); 5831 arc_kstat_update_state(arc_mru, 5832 &as->arcstat_mru_size, 5833 &as->arcstat_mru_evictable_data, 5834 &as->arcstat_mru_evictable_metadata); 5835 arc_kstat_update_state(arc_mru_ghost, 5836 &as->arcstat_mru_ghost_size, 5837 &as->arcstat_mru_ghost_evictable_data, 5838 &as->arcstat_mru_ghost_evictable_metadata); 5839 arc_kstat_update_state(arc_mfu, 5840 &as->arcstat_mfu_size, 5841 &as->arcstat_mfu_evictable_data, 5842 &as->arcstat_mfu_evictable_metadata); 5843 arc_kstat_update_state(arc_mfu_ghost, 5844 &as->arcstat_mfu_ghost_size, 5845 &as->arcstat_mfu_ghost_evictable_data, 5846 &as->arcstat_mfu_ghost_evictable_metadata); 5847 } 5848 5849 return (0); 5850} 5851 5852/* 5853 * This function *must* return indices evenly distributed between all 5854 * sublists of the multilist. This is needed due to how the ARC eviction 5855 * code is laid out; arc_evict_state() assumes ARC buffers are evenly 5856 * distributed between all sublists and uses this assumption when 5857 * deciding which sublist to evict from and how much to evict from it. 5858 */ 5859unsigned int 5860arc_state_multilist_index_func(multilist_t *ml, void *obj) 5861{ 5862 arc_buf_hdr_t *hdr = obj; 5863 5864 /* 5865 * We rely on b_dva to generate evenly distributed index 5866 * numbers using buf_hash below. So, as an added precaution, 5867 * let's make sure we never add empty buffers to the arc lists. 5868 */ 5869 ASSERT(!HDR_EMPTY(hdr)); 5870 5871 /* 5872 * The assumption here, is the hash value for a given 5873 * arc_buf_hdr_t will remain constant throughout it's lifetime 5874 * (i.e. it's b_spa, b_dva, and b_birth fields don't change). 5875 * Thus, we don't need to store the header's sublist index 5876 * on insertion, as this index can be recalculated on removal. 5877 * 5878 * Also, the low order bits of the hash value are thought to be 5879 * distributed evenly. Otherwise, in the case that the multilist 5880 * has a power of two number of sublists, each sublists' usage 5881 * would not be evenly distributed. 5882 */ 5883 return (buf_hash(hdr->b_spa, &hdr->b_dva, hdr->b_birth) % 5884 multilist_get_num_sublists(ml)); 5885} 5886 5887#ifdef _KERNEL 5888static eventhandler_tag arc_event_lowmem = NULL; 5889 5890static void 5891arc_lowmem(void *arg __unused, int howto __unused) 5892{ 5893 5894 mutex_enter(&arc_reclaim_lock); 5895 /* XXX: Memory deficit should be passed as argument. */ 5896 needfree = btoc(arc_c >> arc_shrink_shift); 5897 DTRACE_PROBE(arc__needfree); 5898 cv_signal(&arc_reclaim_thread_cv); 5899 5900 /* 5901 * It is unsafe to block here in arbitrary threads, because we can come 5902 * here from ARC itself and may hold ARC locks and thus risk a deadlock 5903 * with ARC reclaim thread. 5904 */ 5905 if (curproc == pageproc) 5906 (void) cv_wait(&arc_reclaim_waiters_cv, &arc_reclaim_lock); 5907 mutex_exit(&arc_reclaim_lock); 5908} 5909#endif 5910 5911static void 5912arc_state_init(void) 5913{ 5914 arc_anon = &ARC_anon; 5915 arc_mru = &ARC_mru; 5916 arc_mru_ghost = &ARC_mru_ghost; 5917 arc_mfu = &ARC_mfu; 5918 arc_mfu_ghost = &ARC_mfu_ghost; 5919 arc_l2c_only = &ARC_l2c_only; 5920 5921 multilist_create(&arc_mru->arcs_list[ARC_BUFC_METADATA], 5922 sizeof (arc_buf_hdr_t), 5923 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5924 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5925 multilist_create(&arc_mru->arcs_list[ARC_BUFC_DATA], 5926 sizeof (arc_buf_hdr_t), 5927 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5928 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5929 multilist_create(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA], 5930 sizeof (arc_buf_hdr_t), 5931 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5932 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5933 multilist_create(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA], 5934 sizeof (arc_buf_hdr_t), 5935 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5936 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5937 multilist_create(&arc_mfu->arcs_list[ARC_BUFC_METADATA], 5938 sizeof (arc_buf_hdr_t), 5939 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5940 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5941 multilist_create(&arc_mfu->arcs_list[ARC_BUFC_DATA], 5942 sizeof (arc_buf_hdr_t), 5943 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5944 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5945 multilist_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA], 5946 sizeof (arc_buf_hdr_t), 5947 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5948 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5949 multilist_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA], 5950 sizeof (arc_buf_hdr_t), 5951 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5952 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5953 multilist_create(&arc_l2c_only->arcs_list[ARC_BUFC_METADATA], 5954 sizeof (arc_buf_hdr_t), 5955 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5956 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5957 multilist_create(&arc_l2c_only->arcs_list[ARC_BUFC_DATA], 5958 sizeof (arc_buf_hdr_t), 5959 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5960 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5961 5962 refcount_create(&arc_anon->arcs_esize[ARC_BUFC_METADATA]); 5963 refcount_create(&arc_anon->arcs_esize[ARC_BUFC_DATA]); 5964 refcount_create(&arc_mru->arcs_esize[ARC_BUFC_METADATA]); 5965 refcount_create(&arc_mru->arcs_esize[ARC_BUFC_DATA]); 5966 refcount_create(&arc_mru_ghost->arcs_esize[ARC_BUFC_METADATA]); 5967 refcount_create(&arc_mru_ghost->arcs_esize[ARC_BUFC_DATA]); 5968 refcount_create(&arc_mfu->arcs_esize[ARC_BUFC_METADATA]); 5969 refcount_create(&arc_mfu->arcs_esize[ARC_BUFC_DATA]); 5970 refcount_create(&arc_mfu_ghost->arcs_esize[ARC_BUFC_METADATA]); 5971 refcount_create(&arc_mfu_ghost->arcs_esize[ARC_BUFC_DATA]); 5972 refcount_create(&arc_l2c_only->arcs_esize[ARC_BUFC_METADATA]); 5973 refcount_create(&arc_l2c_only->arcs_esize[ARC_BUFC_DATA]); 5974 5975 refcount_create(&arc_anon->arcs_size); 5976 refcount_create(&arc_mru->arcs_size); 5977 refcount_create(&arc_mru_ghost->arcs_size); 5978 refcount_create(&arc_mfu->arcs_size); 5979 refcount_create(&arc_mfu_ghost->arcs_size); 5980 refcount_create(&arc_l2c_only->arcs_size); 5981} 5982 5983static void 5984arc_state_fini(void) 5985{ 5986 refcount_destroy(&arc_anon->arcs_esize[ARC_BUFC_METADATA]); 5987 refcount_destroy(&arc_anon->arcs_esize[ARC_BUFC_DATA]); 5988 refcount_destroy(&arc_mru->arcs_esize[ARC_BUFC_METADATA]); 5989 refcount_destroy(&arc_mru->arcs_esize[ARC_BUFC_DATA]); 5990 refcount_destroy(&arc_mru_ghost->arcs_esize[ARC_BUFC_METADATA]); 5991 refcount_destroy(&arc_mru_ghost->arcs_esize[ARC_BUFC_DATA]); 5992 refcount_destroy(&arc_mfu->arcs_esize[ARC_BUFC_METADATA]); 5993 refcount_destroy(&arc_mfu->arcs_esize[ARC_BUFC_DATA]); 5994 refcount_destroy(&arc_mfu_ghost->arcs_esize[ARC_BUFC_METADATA]); 5995 refcount_destroy(&arc_mfu_ghost->arcs_esize[ARC_BUFC_DATA]); 5996 refcount_destroy(&arc_l2c_only->arcs_esize[ARC_BUFC_METADATA]); 5997 refcount_destroy(&arc_l2c_only->arcs_esize[ARC_BUFC_DATA]); 5998 5999 refcount_destroy(&arc_anon->arcs_size); 6000 refcount_destroy(&arc_mru->arcs_size); 6001 refcount_destroy(&arc_mru_ghost->arcs_size); 6002 refcount_destroy(&arc_mfu->arcs_size); 6003 refcount_destroy(&arc_mfu_ghost->arcs_size); 6004 refcount_destroy(&arc_l2c_only->arcs_size); 6005 6006 multilist_destroy(&arc_mru->arcs_list[ARC_BUFC_METADATA]); 6007 multilist_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA]); 6008 multilist_destroy(&arc_mfu->arcs_list[ARC_BUFC_METADATA]); 6009 multilist_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA]); 6010 multilist_destroy(&arc_mru->arcs_list[ARC_BUFC_DATA]); 6011 multilist_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA]); 6012 multilist_destroy(&arc_mfu->arcs_list[ARC_BUFC_DATA]); 6013 multilist_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA]); 6014} 6015 6016uint64_t 6017arc_max_bytes(void) 6018{ 6019 return (arc_c_max); 6020} 6021 6022void 6023arc_init(void) 6024{ 6025 int i, prefetch_tunable_set = 0; 6026 6027 mutex_init(&arc_reclaim_lock, NULL, MUTEX_DEFAULT, NULL); 6028 cv_init(&arc_reclaim_thread_cv, NULL, CV_DEFAULT, NULL); 6029 cv_init(&arc_reclaim_waiters_cv, NULL, CV_DEFAULT, NULL); 6030 6031 mutex_init(&arc_dnlc_evicts_lock, NULL, MUTEX_DEFAULT, NULL); 6032 cv_init(&arc_dnlc_evicts_cv, NULL, CV_DEFAULT, NULL); 6033 6034 /* Convert seconds to clock ticks */ 6035 arc_min_prefetch_lifespan = 1 * hz; 6036 6037 /* Start out with 1/8 of all memory */ 6038 arc_c = kmem_size() / 8; 6039 6040#ifdef illumos 6041#ifdef _KERNEL 6042 /* 6043 * On architectures where the physical memory can be larger 6044 * than the addressable space (intel in 32-bit mode), we may 6045 * need to limit the cache to 1/8 of VM size. 6046 */ 6047 arc_c = MIN(arc_c, vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 8); 6048#endif 6049#endif /* illumos */ 6050 /* set min cache to 1/32 of all memory, or arc_abs_min, whichever is more */ 6051 arc_c_min = MAX(arc_c / 4, arc_abs_min); 6052 /* set max to 1/2 of all memory, or all but 1GB, whichever is more */ 6053 if (arc_c * 8 >= 1 << 30) 6054 arc_c_max = (arc_c * 8) - (1 << 30); 6055 else 6056 arc_c_max = arc_c_min; 6057 arc_c_max = MAX(arc_c * 5, arc_c_max); 6058 6059 /* 6060 * In userland, there's only the memory pressure that we artificially 6061 * create (see arc_available_memory()). Don't let arc_c get too 6062 * small, because it can cause transactions to be larger than 6063 * arc_c, causing arc_tempreserve_space() to fail. 6064 */ 6065#ifndef _KERNEL 6066 arc_c_min = arc_c_max / 2; 6067#endif 6068 6069#ifdef _KERNEL 6070 /* 6071 * Allow the tunables to override our calculations if they are 6072 * reasonable. 6073 */ 6074 if (zfs_arc_max > arc_abs_min && zfs_arc_max < kmem_size()) { 6075 arc_c_max = zfs_arc_max; 6076 arc_c_min = MIN(arc_c_min, arc_c_max); 6077 } 6078 if (zfs_arc_min > arc_abs_min && zfs_arc_min <= arc_c_max) 6079 arc_c_min = zfs_arc_min; 6080#endif 6081 6082 arc_c = arc_c_max; 6083 arc_p = (arc_c >> 1); 6084 arc_size = 0; 6085 6086 /* limit meta-data to 1/4 of the arc capacity */ 6087 arc_meta_limit = arc_c_max / 4; 6088 6089 /* Allow the tunable to override if it is reasonable */ 6090 if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max) 6091 arc_meta_limit = zfs_arc_meta_limit; 6092 6093 if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0) 6094 arc_c_min = arc_meta_limit / 2; 6095 6096 if (zfs_arc_meta_min > 0) { 6097 arc_meta_min = zfs_arc_meta_min; 6098 } else { 6099 arc_meta_min = arc_c_min / 2; 6100 } 6101 6102 if (zfs_arc_grow_retry > 0) 6103 arc_grow_retry = zfs_arc_grow_retry; 6104 6105 if (zfs_arc_shrink_shift > 0) 6106 arc_shrink_shift = zfs_arc_shrink_shift; 6107 6108 /* 6109 * Ensure that arc_no_grow_shift is less than arc_shrink_shift. 6110 */ 6111 if (arc_no_grow_shift >= arc_shrink_shift) 6112 arc_no_grow_shift = arc_shrink_shift - 1; 6113 6114 if (zfs_arc_p_min_shift > 0) 6115 arc_p_min_shift = zfs_arc_p_min_shift; 6116 6117 if (zfs_arc_num_sublists_per_state < 1) 6118 zfs_arc_num_sublists_per_state = MAX(max_ncpus, 1); 6119 6120 /* if kmem_flags are set, lets try to use less memory */ 6121 if (kmem_debugging()) 6122 arc_c = arc_c / 2; 6123 if (arc_c < arc_c_min) 6124 arc_c = arc_c_min; 6125 6126 zfs_arc_min = arc_c_min; 6127 zfs_arc_max = arc_c_max; 6128 6129 arc_state_init(); 6130 buf_init(); 6131 6132 arc_reclaim_thread_exit = B_FALSE; 6133 arc_dnlc_evicts_thread_exit = FALSE; 6134 6135 arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED, 6136 sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL); 6137 6138 if (arc_ksp != NULL) { 6139 arc_ksp->ks_data = &arc_stats; 6140 arc_ksp->ks_update = arc_kstat_update; 6141 kstat_install(arc_ksp); 6142 } 6143 6144 (void) thread_create(NULL, 0, arc_reclaim_thread, NULL, 0, &p0, 6145 TS_RUN, minclsyspri); 6146 6147#ifdef _KERNEL 6148 arc_event_lowmem = EVENTHANDLER_REGISTER(vm_lowmem, arc_lowmem, NULL, 6149 EVENTHANDLER_PRI_FIRST); 6150#endif 6151 6152 (void) thread_create(NULL, 0, arc_dnlc_evicts_thread, NULL, 0, &p0, 6153 TS_RUN, minclsyspri); 6154 6155 arc_dead = B_FALSE; 6156 arc_warm = B_FALSE; 6157 6158 /* 6159 * Calculate maximum amount of dirty data per pool. 6160 * 6161 * If it has been set by /etc/system, take that. 6162 * Otherwise, use a percentage of physical memory defined by 6163 * zfs_dirty_data_max_percent (default 10%) with a cap at 6164 * zfs_dirty_data_max_max (default 4GB). 6165 */ 6166 if (zfs_dirty_data_max == 0) { 6167 zfs_dirty_data_max = ptob(physmem) * 6168 zfs_dirty_data_max_percent / 100; 6169 zfs_dirty_data_max = MIN(zfs_dirty_data_max, 6170 zfs_dirty_data_max_max); 6171 } 6172 6173#ifdef _KERNEL 6174 if (TUNABLE_INT_FETCH("vfs.zfs.prefetch_disable", &zfs_prefetch_disable)) 6175 prefetch_tunable_set = 1; 6176 6177#ifdef __i386__ 6178 if (prefetch_tunable_set == 0) { 6179 printf("ZFS NOTICE: Prefetch is disabled by default on i386 " 6180 "-- to enable,\n"); 6181 printf(" add \"vfs.zfs.prefetch_disable=0\" " 6182 "to /boot/loader.conf.\n"); 6183 zfs_prefetch_disable = 1; 6184 } 6185#else 6186 if ((((uint64_t)physmem * PAGESIZE) < (1ULL << 32)) && 6187 prefetch_tunable_set == 0) { 6188 printf("ZFS NOTICE: Prefetch is disabled by default if less " 6189 "than 4GB of RAM is present;\n" 6190 " to enable, add \"vfs.zfs.prefetch_disable=0\" " 6191 "to /boot/loader.conf.\n"); 6192 zfs_prefetch_disable = 1; 6193 } 6194#endif 6195 /* Warn about ZFS memory and address space requirements. */ 6196 if (((uint64_t)physmem * PAGESIZE) < (256 + 128 + 64) * (1 << 20)) { 6197 printf("ZFS WARNING: Recommended minimum RAM size is 512MB; " 6198 "expect unstable behavior.\n"); 6199 } 6200 if (kmem_size() < 512 * (1 << 20)) { 6201 printf("ZFS WARNING: Recommended minimum kmem_size is 512MB; " 6202 "expect unstable behavior.\n"); 6203 printf(" Consider tuning vm.kmem_size and " 6204 "vm.kmem_size_max\n"); 6205 printf(" in /boot/loader.conf.\n"); 6206 } 6207#endif 6208} 6209 6210void 6211arc_fini(void) 6212{ 6213 mutex_enter(&arc_reclaim_lock); 6214 arc_reclaim_thread_exit = B_TRUE; 6215 /* 6216 * The reclaim thread will set arc_reclaim_thread_exit back to 6217 * B_FALSE when it is finished exiting; we're waiting for that. 6218 */ 6219 while (arc_reclaim_thread_exit) { 6220 cv_signal(&arc_reclaim_thread_cv); 6221 cv_wait(&arc_reclaim_thread_cv, &arc_reclaim_lock); 6222 } 6223 mutex_exit(&arc_reclaim_lock); 6224 6225 /* Use B_TRUE to ensure *all* buffers are evicted */ 6226 arc_flush(NULL, B_TRUE); 6227 6228 mutex_enter(&arc_dnlc_evicts_lock); 6229 arc_dnlc_evicts_thread_exit = TRUE; 6230 /* 6231 * The user evicts thread will set arc_user_evicts_thread_exit 6232 * to FALSE when it is finished exiting; we're waiting for that. 6233 */ 6234 while (arc_dnlc_evicts_thread_exit) { 6235 cv_signal(&arc_dnlc_evicts_cv); 6236 cv_wait(&arc_dnlc_evicts_cv, &arc_dnlc_evicts_lock); 6237 } 6238 mutex_exit(&arc_dnlc_evicts_lock); 6239 6240 arc_dead = B_TRUE; 6241 6242 if (arc_ksp != NULL) { 6243 kstat_delete(arc_ksp); 6244 arc_ksp = NULL; 6245 } 6246 6247 mutex_destroy(&arc_reclaim_lock); 6248 cv_destroy(&arc_reclaim_thread_cv); 6249 cv_destroy(&arc_reclaim_waiters_cv); 6250 6251 mutex_destroy(&arc_dnlc_evicts_lock); 6252 cv_destroy(&arc_dnlc_evicts_cv); 6253 6254 arc_state_fini(); 6255 buf_fini(); 6256 6257 ASSERT0(arc_loaned_bytes); 6258 6259#ifdef _KERNEL 6260 if (arc_event_lowmem != NULL) 6261 EVENTHANDLER_DEREGISTER(vm_lowmem, arc_event_lowmem); 6262#endif 6263} 6264 6265/* 6266 * Level 2 ARC 6267 * 6268 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk. 6269 * It uses dedicated storage devices to hold cached data, which are populated 6270 * using large infrequent writes. The main role of this cache is to boost 6271 * the performance of random read workloads. The intended L2ARC devices 6272 * include short-stroked disks, solid state disks, and other media with 6273 * substantially faster read latency than disk. 6274 * 6275 * +-----------------------+ 6276 * | ARC | 6277 * +-----------------------+ 6278 * | ^ ^ 6279 * | | | 6280 * l2arc_feed_thread() arc_read() 6281 * | | | 6282 * | l2arc read | 6283 * V | | 6284 * +---------------+ | 6285 * | L2ARC | | 6286 * +---------------+ | 6287 * | ^ | 6288 * l2arc_write() | | 6289 * | | | 6290 * V | | 6291 * +-------+ +-------+ 6292 * | vdev | | vdev | 6293 * | cache | | cache | 6294 * +-------+ +-------+ 6295 * +=========+ .-----. 6296 * : L2ARC : |-_____-| 6297 * : devices : | Disks | 6298 * +=========+ `-_____-' 6299 * 6300 * Read requests are satisfied from the following sources, in order: 6301 * 6302 * 1) ARC 6303 * 2) vdev cache of L2ARC devices 6304 * 3) L2ARC devices 6305 * 4) vdev cache of disks 6306 * 5) disks 6307 * 6308 * Some L2ARC device types exhibit extremely slow write performance. 6309 * To accommodate for this there are some significant differences between 6310 * the L2ARC and traditional cache design: 6311 * 6312 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from 6313 * the ARC behave as usual, freeing buffers and placing headers on ghost 6314 * lists. The ARC does not send buffers to the L2ARC during eviction as 6315 * this would add inflated write latencies for all ARC memory pressure. 6316 * 6317 * 2. The L2ARC attempts to cache data from the ARC before it is evicted. 6318 * It does this by periodically scanning buffers from the eviction-end of 6319 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are 6320 * not already there. It scans until a headroom of buffers is satisfied, 6321 * which itself is a buffer for ARC eviction. If a compressible buffer is 6322 * found during scanning and selected for writing to an L2ARC device, we 6323 * temporarily boost scanning headroom during the next scan cycle to make 6324 * sure we adapt to compression effects (which might significantly reduce 6325 * the data volume we write to L2ARC). The thread that does this is 6326 * l2arc_feed_thread(), illustrated below; example sizes are included to 6327 * provide a better sense of ratio than this diagram: 6328 * 6329 * head --> tail 6330 * +---------------------+----------+ 6331 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC 6332 * +---------------------+----------+ | o L2ARC eligible 6333 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer 6334 * +---------------------+----------+ | 6335 * 15.9 Gbytes ^ 32 Mbytes | 6336 * headroom | 6337 * l2arc_feed_thread() 6338 * | 6339 * l2arc write hand <--[oooo]--' 6340 * | 8 Mbyte 6341 * | write max 6342 * V 6343 * +==============================+ 6344 * L2ARC dev |####|#|###|###| |####| ... | 6345 * +==============================+ 6346 * 32 Gbytes 6347 * 6348 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of 6349 * evicted, then the L2ARC has cached a buffer much sooner than it probably 6350 * needed to, potentially wasting L2ARC device bandwidth and storage. It is 6351 * safe to say that this is an uncommon case, since buffers at the end of 6352 * the ARC lists have moved there due to inactivity. 6353 * 6354 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom, 6355 * then the L2ARC simply misses copying some buffers. This serves as a 6356 * pressure valve to prevent heavy read workloads from both stalling the ARC 6357 * with waits and clogging the L2ARC with writes. This also helps prevent 6358 * the potential for the L2ARC to churn if it attempts to cache content too 6359 * quickly, such as during backups of the entire pool. 6360 * 6361 * 5. After system boot and before the ARC has filled main memory, there are 6362 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru 6363 * lists can remain mostly static. Instead of searching from tail of these 6364 * lists as pictured, the l2arc_feed_thread() will search from the list heads 6365 * for eligible buffers, greatly increasing its chance of finding them. 6366 * 6367 * The L2ARC device write speed is also boosted during this time so that 6368 * the L2ARC warms up faster. Since there have been no ARC evictions yet, 6369 * there are no L2ARC reads, and no fear of degrading read performance 6370 * through increased writes. 6371 * 6372 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that 6373 * the vdev queue can aggregate them into larger and fewer writes. Each 6374 * device is written to in a rotor fashion, sweeping writes through 6375 * available space then repeating. 6376 * 6377 * 7. The L2ARC does not store dirty content. It never needs to flush 6378 * write buffers back to disk based storage. 6379 * 6380 * 8. If an ARC buffer is written (and dirtied) which also exists in the 6381 * L2ARC, the now stale L2ARC buffer is immediately dropped. 6382 * 6383 * The performance of the L2ARC can be tweaked by a number of tunables, which 6384 * may be necessary for different workloads: 6385 * 6386 * l2arc_write_max max write bytes per interval 6387 * l2arc_write_boost extra write bytes during device warmup 6388 * l2arc_noprefetch skip caching prefetched buffers 6389 * l2arc_headroom number of max device writes to precache 6390 * l2arc_headroom_boost when we find compressed buffers during ARC 6391 * scanning, we multiply headroom by this 6392 * percentage factor for the next scan cycle, 6393 * since more compressed buffers are likely to 6394 * be present 6395 * l2arc_feed_secs seconds between L2ARC writing 6396 * 6397 * Tunables may be removed or added as future performance improvements are 6398 * integrated, and also may become zpool properties. 6399 * 6400 * There are three key functions that control how the L2ARC warms up: 6401 * 6402 * l2arc_write_eligible() check if a buffer is eligible to cache 6403 * l2arc_write_size() calculate how much to write 6404 * l2arc_write_interval() calculate sleep delay between writes 6405 * 6406 * These three functions determine what to write, how much, and how quickly 6407 * to send writes. 6408 */ 6409 6410static boolean_t 6411l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *hdr) 6412{ 6413 /* 6414 * A buffer is *not* eligible for the L2ARC if it: 6415 * 1. belongs to a different spa. 6416 * 2. is already cached on the L2ARC. 6417 * 3. has an I/O in progress (it may be an incomplete read). 6418 * 4. is flagged not eligible (zfs property). 6419 */ 6420 if (hdr->b_spa != spa_guid) { 6421 ARCSTAT_BUMP(arcstat_l2_write_spa_mismatch); 6422 return (B_FALSE); 6423 } 6424 if (HDR_HAS_L2HDR(hdr)) { 6425 ARCSTAT_BUMP(arcstat_l2_write_in_l2); 6426 return (B_FALSE); 6427 } 6428 if (HDR_IO_IN_PROGRESS(hdr)) { 6429 ARCSTAT_BUMP(arcstat_l2_write_hdr_io_in_progress); 6430 return (B_FALSE); 6431 } 6432 if (!HDR_L2CACHE(hdr)) { 6433 ARCSTAT_BUMP(arcstat_l2_write_not_cacheable); 6434 return (B_FALSE); 6435 } 6436 6437 return (B_TRUE); 6438} 6439 6440static uint64_t 6441l2arc_write_size(void) 6442{ 6443 uint64_t size; 6444 6445 /* 6446 * Make sure our globals have meaningful values in case the user 6447 * altered them. 6448 */ 6449 size = l2arc_write_max; 6450 if (size == 0) { 6451 cmn_err(CE_NOTE, "Bad value for l2arc_write_max, value must " 6452 "be greater than zero, resetting it to the default (%d)", 6453 L2ARC_WRITE_SIZE); 6454 size = l2arc_write_max = L2ARC_WRITE_SIZE; 6455 } 6456 6457 if (arc_warm == B_FALSE) 6458 size += l2arc_write_boost; 6459 6460 return (size); 6461 6462} 6463 6464static clock_t 6465l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote) 6466{ 6467 clock_t interval, next, now; 6468 6469 /* 6470 * If the ARC lists are busy, increase our write rate; if the 6471 * lists are stale, idle back. This is achieved by checking 6472 * how much we previously wrote - if it was more than half of 6473 * what we wanted, schedule the next write much sooner. 6474 */ 6475 if (l2arc_feed_again && wrote > (wanted / 2)) 6476 interval = (hz * l2arc_feed_min_ms) / 1000; 6477 else 6478 interval = hz * l2arc_feed_secs; 6479 6480 now = ddi_get_lbolt(); 6481 next = MAX(now, MIN(now + interval, began + interval)); 6482 6483 return (next); 6484} 6485 6486/* 6487 * Cycle through L2ARC devices. This is how L2ARC load balances. 6488 * If a device is returned, this also returns holding the spa config lock. 6489 */ 6490static l2arc_dev_t * 6491l2arc_dev_get_next(void) 6492{ 6493 l2arc_dev_t *first, *next = NULL; 6494 6495 /* 6496 * Lock out the removal of spas (spa_namespace_lock), then removal 6497 * of cache devices (l2arc_dev_mtx). Once a device has been selected, 6498 * both locks will be dropped and a spa config lock held instead. 6499 */ 6500 mutex_enter(&spa_namespace_lock); 6501 mutex_enter(&l2arc_dev_mtx); 6502 6503 /* if there are no vdevs, there is nothing to do */ 6504 if (l2arc_ndev == 0) 6505 goto out; 6506 6507 first = NULL; 6508 next = l2arc_dev_last; 6509 do { 6510 /* loop around the list looking for a non-faulted vdev */ 6511 if (next == NULL) { 6512 next = list_head(l2arc_dev_list); 6513 } else { 6514 next = list_next(l2arc_dev_list, next); 6515 if (next == NULL) 6516 next = list_head(l2arc_dev_list); 6517 } 6518 6519 /* if we have come back to the start, bail out */ 6520 if (first == NULL) 6521 first = next; 6522 else if (next == first) 6523 break; 6524 6525 } while (vdev_is_dead(next->l2ad_vdev)); 6526 6527 /* if we were unable to find any usable vdevs, return NULL */ 6528 if (vdev_is_dead(next->l2ad_vdev)) 6529 next = NULL; 6530 6531 l2arc_dev_last = next; 6532 6533out: 6534 mutex_exit(&l2arc_dev_mtx); 6535 6536 /* 6537 * Grab the config lock to prevent the 'next' device from being 6538 * removed while we are writing to it. 6539 */ 6540 if (next != NULL) 6541 spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER); 6542 mutex_exit(&spa_namespace_lock); 6543 6544 return (next); 6545} 6546 6547/* 6548 * Free buffers that were tagged for destruction. 6549 */ 6550static void 6551l2arc_do_free_on_write() 6552{ 6553 list_t *buflist; 6554 l2arc_data_free_t *df, *df_prev; 6555 6556 mutex_enter(&l2arc_free_on_write_mtx); 6557 buflist = l2arc_free_on_write; 6558 6559 for (df = list_tail(buflist); df; df = df_prev) { 6560 df_prev = list_prev(buflist, df); 6561 ASSERT3P(df->l2df_data, !=, NULL); 6562 if (df->l2df_type == ARC_BUFC_METADATA) { 6563 zio_buf_free(df->l2df_data, df->l2df_size); 6564 } else { 6565 ASSERT(df->l2df_type == ARC_BUFC_DATA); 6566 zio_data_buf_free(df->l2df_data, df->l2df_size); 6567 } 6568 list_remove(buflist, df); 6569 kmem_free(df, sizeof (l2arc_data_free_t)); 6570 } 6571 6572 mutex_exit(&l2arc_free_on_write_mtx); 6573} 6574 6575/* 6576 * A write to a cache device has completed. Update all headers to allow 6577 * reads from these buffers to begin. 6578 */ 6579static void 6580l2arc_write_done(zio_t *zio) 6581{ 6582 l2arc_write_callback_t *cb; 6583 l2arc_dev_t *dev; 6584 list_t *buflist; 6585 arc_buf_hdr_t *head, *hdr, *hdr_prev; 6586 kmutex_t *hash_lock; 6587 int64_t bytes_dropped = 0; 6588 6589 cb = zio->io_private; 6590 ASSERT3P(cb, !=, NULL); 6591 dev = cb->l2wcb_dev; 6592 ASSERT3P(dev, !=, NULL); 6593 head = cb->l2wcb_head; 6594 ASSERT3P(head, !=, NULL); 6595 buflist = &dev->l2ad_buflist; 6596 ASSERT3P(buflist, !=, NULL); 6597 DTRACE_PROBE2(l2arc__iodone, zio_t *, zio, 6598 l2arc_write_callback_t *, cb); 6599 6600 if (zio->io_error != 0) 6601 ARCSTAT_BUMP(arcstat_l2_writes_error); 6602 6603 /* 6604 * All writes completed, or an error was hit. 6605 */ 6606top: 6607 mutex_enter(&dev->l2ad_mtx); 6608 for (hdr = list_prev(buflist, head); hdr; hdr = hdr_prev) { 6609 hdr_prev = list_prev(buflist, hdr); 6610 6611 hash_lock = HDR_LOCK(hdr); 6612 6613 /* 6614 * We cannot use mutex_enter or else we can deadlock 6615 * with l2arc_write_buffers (due to swapping the order 6616 * the hash lock and l2ad_mtx are taken). 6617 */ 6618 if (!mutex_tryenter(hash_lock)) { 6619 /* 6620 * Missed the hash lock. We must retry so we 6621 * don't leave the ARC_FLAG_L2_WRITING bit set. 6622 */ 6623 ARCSTAT_BUMP(arcstat_l2_writes_lock_retry); 6624 6625 /* 6626 * We don't want to rescan the headers we've 6627 * already marked as having been written out, so 6628 * we reinsert the head node so we can pick up 6629 * where we left off. 6630 */ 6631 list_remove(buflist, head); 6632 list_insert_after(buflist, hdr, head); 6633 6634 mutex_exit(&dev->l2ad_mtx); 6635 6636 /* 6637 * We wait for the hash lock to become available 6638 * to try and prevent busy waiting, and increase 6639 * the chance we'll be able to acquire the lock 6640 * the next time around. 6641 */ 6642 mutex_enter(hash_lock); 6643 mutex_exit(hash_lock); 6644 goto top; 6645 } 6646 6647 /* 6648 * We could not have been moved into the arc_l2c_only 6649 * state while in-flight due to our ARC_FLAG_L2_WRITING 6650 * bit being set. Let's just ensure that's being enforced. 6651 */ 6652 ASSERT(HDR_HAS_L1HDR(hdr)); 6653 6654 if (zio->io_error != 0) { 6655 /* 6656 * Error - drop L2ARC entry. 6657 */ 6658 list_remove(buflist, hdr); 6659 l2arc_trim(hdr); 6660 arc_hdr_clear_flags(hdr, ARC_FLAG_HAS_L2HDR); 6661 6662 ARCSTAT_INCR(arcstat_l2_asize, -arc_hdr_size(hdr)); 6663 ARCSTAT_INCR(arcstat_l2_size, -HDR_GET_LSIZE(hdr)); 6664 6665 bytes_dropped += arc_hdr_size(hdr); 6666 (void) refcount_remove_many(&dev->l2ad_alloc, 6667 arc_hdr_size(hdr), hdr); 6668 } 6669 6670 /* 6671 * Allow ARC to begin reads and ghost list evictions to 6672 * this L2ARC entry. 6673 */ 6674 arc_hdr_clear_flags(hdr, ARC_FLAG_L2_WRITING); 6675 6676 mutex_exit(hash_lock); 6677 } 6678 6679 atomic_inc_64(&l2arc_writes_done); 6680 list_remove(buflist, head); 6681 ASSERT(!HDR_HAS_L1HDR(head)); 6682 kmem_cache_free(hdr_l2only_cache, head); 6683 mutex_exit(&dev->l2ad_mtx); 6684 6685 vdev_space_update(dev->l2ad_vdev, -bytes_dropped, 0, 0); 6686 6687 l2arc_do_free_on_write(); 6688 6689 kmem_free(cb, sizeof (l2arc_write_callback_t)); 6690} 6691 6692/* 6693 * A read to a cache device completed. Validate buffer contents before 6694 * handing over to the regular ARC routines. 6695 */ 6696static void 6697l2arc_read_done(zio_t *zio) 6698{ 6699 l2arc_read_callback_t *cb; 6700 arc_buf_hdr_t *hdr; 6701 kmutex_t *hash_lock; 6702 boolean_t valid_cksum; 6703 6704 ASSERT3P(zio->io_vd, !=, NULL); 6705 ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE); 6706 6707 spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd); 6708 6709 cb = zio->io_private; 6710 ASSERT3P(cb, !=, NULL); 6711 hdr = cb->l2rcb_hdr; 6712 ASSERT3P(hdr, !=, NULL); 6713 6714 hash_lock = HDR_LOCK(hdr); 6715 mutex_enter(hash_lock); 6716 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 6717 6718 /* 6719 * If the data was read into a temporary buffer, 6720 * move it and free the buffer. 6721 */ 6722 if (cb->l2rcb_data != NULL) { 6723 ASSERT3U(arc_hdr_size(hdr), <, zio->io_size); 6724 if (zio->io_error == 0) { 6725 bcopy(cb->l2rcb_data, hdr->b_l1hdr.b_pdata, 6726 arc_hdr_size(hdr)); 6727 } 6728 6729 /* 6730 * The following must be done regardless of whether 6731 * there was an error: 6732 * - free the temporary buffer 6733 * - point zio to the real ARC buffer 6734 * - set zio size accordingly 6735 * These are required because zio is either re-used for 6736 * an I/O of the block in the case of the error 6737 * or the zio is passed to arc_read_done() and it 6738 * needs real data. 6739 */ 6740 zio_data_buf_free(cb->l2rcb_data, zio->io_size); 6741 zio->io_size = zio->io_orig_size = arc_hdr_size(hdr); 6742 zio->io_data = zio->io_orig_data = hdr->b_l1hdr.b_pdata; 6743 } 6744 6745 ASSERT3P(zio->io_data, !=, NULL); 6746 6747 /* 6748 * Check this survived the L2ARC journey. 6749 */ 6750 ASSERT3P(zio->io_data, ==, hdr->b_l1hdr.b_pdata); 6751 zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */ 6752 zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */ 6753 6754 valid_cksum = arc_cksum_is_equal(hdr, zio); 6755 if (valid_cksum && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) { 6756 mutex_exit(hash_lock); 6757 zio->io_private = hdr; 6758 arc_read_done(zio); 6759 } else { 6760 mutex_exit(hash_lock); 6761 /* 6762 * Buffer didn't survive caching. Increment stats and 6763 * reissue to the original storage device. 6764 */ 6765 if (zio->io_error != 0) { 6766 ARCSTAT_BUMP(arcstat_l2_io_error); 6767 } else { 6768 zio->io_error = SET_ERROR(EIO); 6769 } 6770 if (!valid_cksum) 6771 ARCSTAT_BUMP(arcstat_l2_cksum_bad); 6772 6773 /* 6774 * If there's no waiter, issue an async i/o to the primary 6775 * storage now. If there *is* a waiter, the caller must 6776 * issue the i/o in a context where it's OK to block. 6777 */ 6778 if (zio->io_waiter == NULL) { 6779 zio_t *pio = zio_unique_parent(zio); 6780 6781 ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL); 6782 6783 zio_nowait(zio_read(pio, zio->io_spa, zio->io_bp, 6784 hdr->b_l1hdr.b_pdata, zio->io_size, arc_read_done, 6785 hdr, zio->io_priority, cb->l2rcb_flags, 6786 &cb->l2rcb_zb)); 6787 } 6788 } 6789 6790 kmem_free(cb, sizeof (l2arc_read_callback_t)); 6791} 6792 6793/* 6794 * This is the list priority from which the L2ARC will search for pages to 6795 * cache. This is used within loops (0..3) to cycle through lists in the 6796 * desired order. This order can have a significant effect on cache 6797 * performance. 6798 * 6799 * Currently the metadata lists are hit first, MFU then MRU, followed by 6800 * the data lists. This function returns a locked list, and also returns 6801 * the lock pointer. 6802 */ 6803static multilist_sublist_t * 6804l2arc_sublist_lock(int list_num) 6805{ 6806 multilist_t *ml = NULL; 6807 unsigned int idx; 6808 6809 ASSERT(list_num >= 0 && list_num <= 3); 6810 6811 switch (list_num) { 6812 case 0: 6813 ml = &arc_mfu->arcs_list[ARC_BUFC_METADATA]; 6814 break; 6815 case 1: 6816 ml = &arc_mru->arcs_list[ARC_BUFC_METADATA]; 6817 break; 6818 case 2: 6819 ml = &arc_mfu->arcs_list[ARC_BUFC_DATA]; 6820 break; 6821 case 3: 6822 ml = &arc_mru->arcs_list[ARC_BUFC_DATA]; 6823 break; 6824 } 6825 6826 /* 6827 * Return a randomly-selected sublist. This is acceptable 6828 * because the caller feeds only a little bit of data for each 6829 * call (8MB). Subsequent calls will result in different 6830 * sublists being selected. 6831 */ 6832 idx = multilist_get_random_index(ml); 6833 return (multilist_sublist_lock(ml, idx)); 6834} 6835 6836/* 6837 * Evict buffers from the device write hand to the distance specified in 6838 * bytes. This distance may span populated buffers, it may span nothing. 6839 * This is clearing a region on the L2ARC device ready for writing. 6840 * If the 'all' boolean is set, every buffer is evicted. 6841 */ 6842static void 6843l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all) 6844{ 6845 list_t *buflist; 6846 arc_buf_hdr_t *hdr, *hdr_prev; 6847 kmutex_t *hash_lock; 6848 uint64_t taddr; 6849 6850 buflist = &dev->l2ad_buflist; 6851 6852 if (!all && dev->l2ad_first) { 6853 /* 6854 * This is the first sweep through the device. There is 6855 * nothing to evict. 6856 */ 6857 return; 6858 } 6859 6860 if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) { 6861 /* 6862 * When nearing the end of the device, evict to the end 6863 * before the device write hand jumps to the start. 6864 */ 6865 taddr = dev->l2ad_end; 6866 } else { 6867 taddr = dev->l2ad_hand + distance; 6868 } 6869 DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist, 6870 uint64_t, taddr, boolean_t, all); 6871 6872top: 6873 mutex_enter(&dev->l2ad_mtx); 6874 for (hdr = list_tail(buflist); hdr; hdr = hdr_prev) { 6875 hdr_prev = list_prev(buflist, hdr); 6876 6877 hash_lock = HDR_LOCK(hdr); 6878 6879 /* 6880 * We cannot use mutex_enter or else we can deadlock 6881 * with l2arc_write_buffers (due to swapping the order 6882 * the hash lock and l2ad_mtx are taken). 6883 */ 6884 if (!mutex_tryenter(hash_lock)) { 6885 /* 6886 * Missed the hash lock. Retry. 6887 */ 6888 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry); 6889 mutex_exit(&dev->l2ad_mtx); 6890 mutex_enter(hash_lock); 6891 mutex_exit(hash_lock); 6892 goto top; 6893 } 6894 6895 if (HDR_L2_WRITE_HEAD(hdr)) { 6896 /* 6897 * We hit a write head node. Leave it for 6898 * l2arc_write_done(). 6899 */ 6900 list_remove(buflist, hdr); 6901 mutex_exit(hash_lock); 6902 continue; 6903 } 6904 6905 if (!all && HDR_HAS_L2HDR(hdr) && 6906 (hdr->b_l2hdr.b_daddr >= taddr || 6907 hdr->b_l2hdr.b_daddr < dev->l2ad_hand)) { 6908 /* 6909 * We've evicted to the target address, 6910 * or the end of the device. 6911 */ 6912 mutex_exit(hash_lock); 6913 break; 6914 } 6915 6916 ASSERT(HDR_HAS_L2HDR(hdr)); 6917 if (!HDR_HAS_L1HDR(hdr)) { 6918 ASSERT(!HDR_L2_READING(hdr)); 6919 /* 6920 * This doesn't exist in the ARC. Destroy. 6921 * arc_hdr_destroy() will call list_remove() 6922 * and decrement arcstat_l2_size. 6923 */ 6924 arc_change_state(arc_anon, hdr, hash_lock); 6925 arc_hdr_destroy(hdr); 6926 } else { 6927 ASSERT(hdr->b_l1hdr.b_state != arc_l2c_only); 6928 ARCSTAT_BUMP(arcstat_l2_evict_l1cached); 6929 /* 6930 * Invalidate issued or about to be issued 6931 * reads, since we may be about to write 6932 * over this location. 6933 */ 6934 if (HDR_L2_READING(hdr)) { 6935 ARCSTAT_BUMP(arcstat_l2_evict_reading); 6936 arc_hdr_set_flags(hdr, ARC_FLAG_L2_EVICTED); 6937 } 6938 6939 /* Ensure this header has finished being written */ 6940 ASSERT(!HDR_L2_WRITING(hdr)); 6941 6942 arc_hdr_l2hdr_destroy(hdr); 6943 } 6944 mutex_exit(hash_lock); 6945 } 6946 mutex_exit(&dev->l2ad_mtx); 6947} 6948 6949/* 6950 * Find and write ARC buffers to the L2ARC device. 6951 * 6952 * An ARC_FLAG_L2_WRITING flag is set so that the L2ARC buffers are not valid 6953 * for reading until they have completed writing. 6954 * The headroom_boost is an in-out parameter used to maintain headroom boost 6955 * state between calls to this function. 6956 * 6957 * Returns the number of bytes actually written (which may be smaller than 6958 * the delta by which the device hand has changed due to alignment). 6959 */ 6960static uint64_t 6961l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz) 6962{ 6963 arc_buf_hdr_t *hdr, *hdr_prev, *head; 6964 uint64_t write_asize, write_psize, write_sz, headroom; 6965 boolean_t full; 6966 l2arc_write_callback_t *cb; 6967 zio_t *pio, *wzio; 6968 uint64_t guid = spa_load_guid(spa); 6969 int try; 6970 6971 ASSERT3P(dev->l2ad_vdev, !=, NULL); 6972 6973 pio = NULL; 6974 write_sz = write_asize = write_psize = 0; 6975 full = B_FALSE; 6976 head = kmem_cache_alloc(hdr_l2only_cache, KM_PUSHPAGE); 6977 arc_hdr_set_flags(head, ARC_FLAG_L2_WRITE_HEAD | ARC_FLAG_HAS_L2HDR); 6978 6979 ARCSTAT_BUMP(arcstat_l2_write_buffer_iter); 6980 /* 6981 * Copy buffers for L2ARC writing. 6982 */ 6983 for (try = 0; try <= 3; try++) { 6984 multilist_sublist_t *mls = l2arc_sublist_lock(try); 6985 uint64_t passed_sz = 0; 6986 6987 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_iter); 6988 6989 /* 6990 * L2ARC fast warmup. 6991 * 6992 * Until the ARC is warm and starts to evict, read from the 6993 * head of the ARC lists rather than the tail. 6994 */ 6995 if (arc_warm == B_FALSE) 6996 hdr = multilist_sublist_head(mls); 6997 else 6998 hdr = multilist_sublist_tail(mls); 6999 if (hdr == NULL) 7000 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_null_iter); 7001 7002 headroom = target_sz * l2arc_headroom; 7003 if (zfs_compressed_arc_enabled) 7004 headroom = (headroom * l2arc_headroom_boost) / 100; 7005 7006 for (; hdr; hdr = hdr_prev) { 7007 kmutex_t *hash_lock; 7008 7009 if (arc_warm == B_FALSE) 7010 hdr_prev = multilist_sublist_next(mls, hdr); 7011 else 7012 hdr_prev = multilist_sublist_prev(mls, hdr); 7013 ARCSTAT_INCR(arcstat_l2_write_buffer_bytes_scanned, 7014 HDR_GET_LSIZE(hdr)); 7015 7016 hash_lock = HDR_LOCK(hdr); 7017 if (!mutex_tryenter(hash_lock)) { 7018 ARCSTAT_BUMP(arcstat_l2_write_trylock_fail); 7019 /* 7020 * Skip this buffer rather than waiting. 7021 */ 7022 continue; 7023 } 7024 7025 passed_sz += HDR_GET_LSIZE(hdr); 7026 if (passed_sz > headroom) { 7027 /* 7028 * Searched too far. 7029 */ 7030 mutex_exit(hash_lock); 7031 ARCSTAT_BUMP(arcstat_l2_write_passed_headroom); 7032 break; 7033 } 7034 7035 if (!l2arc_write_eligible(guid, hdr)) { 7036 mutex_exit(hash_lock); 7037 continue; 7038 } 7039 7040 /* 7041 * We rely on the L1 portion of the header below, so 7042 * it's invalid for this header to have been evicted out 7043 * of the ghost cache, prior to being written out. The 7044 * ARC_FLAG_L2_WRITING bit ensures this won't happen. 7045 */ 7046 ASSERT(HDR_HAS_L1HDR(hdr)); 7047 7048 ASSERT3U(HDR_GET_PSIZE(hdr), >, 0); 7049 ASSERT3P(hdr->b_l1hdr.b_pdata, !=, NULL); 7050 ASSERT3U(arc_hdr_size(hdr), >, 0); 7051 uint64_t size = arc_hdr_size(hdr); 7052 uint64_t asize = vdev_psize_to_asize(dev->l2ad_vdev, 7053 size); 7054 7055 if ((write_psize + asize) > target_sz) { 7056 full = B_TRUE; 7057 mutex_exit(hash_lock); 7058 ARCSTAT_BUMP(arcstat_l2_write_full); 7059 break; 7060 } 7061 7062 if (pio == NULL) { 7063 /* 7064 * Insert a dummy header on the buflist so 7065 * l2arc_write_done() can find where the 7066 * write buffers begin without searching. 7067 */ 7068 mutex_enter(&dev->l2ad_mtx); 7069 list_insert_head(&dev->l2ad_buflist, head); 7070 mutex_exit(&dev->l2ad_mtx); 7071 7072 cb = kmem_alloc( 7073 sizeof (l2arc_write_callback_t), KM_SLEEP); 7074 cb->l2wcb_dev = dev; 7075 cb->l2wcb_head = head; 7076 pio = zio_root(spa, l2arc_write_done, cb, 7077 ZIO_FLAG_CANFAIL); 7078 ARCSTAT_BUMP(arcstat_l2_write_pios); 7079 } 7080 7081 hdr->b_l2hdr.b_dev = dev; 7082 hdr->b_l2hdr.b_daddr = dev->l2ad_hand; 7083 arc_hdr_set_flags(hdr, 7084 ARC_FLAG_L2_WRITING | ARC_FLAG_HAS_L2HDR); 7085 7086 mutex_enter(&dev->l2ad_mtx); 7087 list_insert_head(&dev->l2ad_buflist, hdr); 7088 mutex_exit(&dev->l2ad_mtx); 7089 7090 (void) refcount_add_many(&dev->l2ad_alloc, size, hdr); 7091 7092 /* 7093 * Normally the L2ARC can use the hdr's data, but if 7094 * we're sharing data between the hdr and one of its 7095 * bufs, L2ARC needs its own copy of the data so that 7096 * the ZIO below can't race with the buf consumer. To 7097 * ensure that this copy will be available for the 7098 * lifetime of the ZIO and be cleaned up afterwards, we 7099 * add it to the l2arc_free_on_write queue. 7100 */ 7101 void *to_write; 7102 if (!HDR_SHARED_DATA(hdr) && size == asize) { 7103 to_write = hdr->b_l1hdr.b_pdata; 7104 } else { 7105 arc_buf_contents_t type = arc_buf_type(hdr); 7106 if (type == ARC_BUFC_METADATA) { 7107 to_write = zio_buf_alloc(asize); 7108 } else { 7109 ASSERT3U(type, ==, ARC_BUFC_DATA); 7110 to_write = zio_data_buf_alloc(asize); 7111 } 7112 7113 bcopy(hdr->b_l1hdr.b_pdata, to_write, size); 7114 if (asize != size) 7115 bzero(to_write + size, asize - size); 7116 l2arc_free_data_on_write(to_write, asize, type); 7117 } 7118 wzio = zio_write_phys(pio, dev->l2ad_vdev, 7119 hdr->b_l2hdr.b_daddr, asize, to_write, 7120 ZIO_CHECKSUM_OFF, NULL, hdr, 7121 ZIO_PRIORITY_ASYNC_WRITE, 7122 ZIO_FLAG_CANFAIL, B_FALSE); 7123 7124 write_sz += HDR_GET_LSIZE(hdr); 7125 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev, 7126 zio_t *, wzio); 7127 7128 write_asize += size; 7129 write_psize += asize; 7130 dev->l2ad_hand += asize; 7131 7132 mutex_exit(hash_lock); 7133 7134 (void) zio_nowait(wzio); 7135 } 7136 7137 multilist_sublist_unlock(mls); 7138 7139 if (full == B_TRUE) 7140 break; 7141 } 7142 7143 /* No buffers selected for writing? */ 7144 if (pio == NULL) { 7145 ASSERT0(write_sz); 7146 ASSERT(!HDR_HAS_L1HDR(head)); 7147 kmem_cache_free(hdr_l2only_cache, head); 7148 return (0); 7149 } 7150 7151 ASSERT3U(write_psize, <=, target_sz); 7152 ARCSTAT_BUMP(arcstat_l2_writes_sent); 7153 ARCSTAT_INCR(arcstat_l2_write_bytes, write_asize); 7154 ARCSTAT_INCR(arcstat_l2_size, write_sz); 7155 ARCSTAT_INCR(arcstat_l2_asize, write_asize); 7156 vdev_space_update(dev->l2ad_vdev, write_asize, 0, 0); 7157 7158 /* 7159 * Bump device hand to the device start if it is approaching the end. 7160 * l2arc_evict() will already have evicted ahead for this case. 7161 */ 7162 if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) { 7163 dev->l2ad_hand = dev->l2ad_start; 7164 dev->l2ad_first = B_FALSE; 7165 } 7166 7167 dev->l2ad_writing = B_TRUE; 7168 (void) zio_wait(pio); 7169 dev->l2ad_writing = B_FALSE; 7170 7171 return (write_asize); 7172} 7173 7174/* 7175 * This thread feeds the L2ARC at regular intervals. This is the beating 7176 * heart of the L2ARC. 7177 */ 7178static void 7179l2arc_feed_thread(void *dummy __unused) 7180{ 7181 callb_cpr_t cpr; 7182 l2arc_dev_t *dev; 7183 spa_t *spa; 7184 uint64_t size, wrote; 7185 clock_t begin, next = ddi_get_lbolt(); 7186 7187 CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG); 7188 7189 mutex_enter(&l2arc_feed_thr_lock); 7190 7191 while (l2arc_thread_exit == 0) { 7192 CALLB_CPR_SAFE_BEGIN(&cpr); 7193 (void) cv_timedwait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock, 7194 next - ddi_get_lbolt()); 7195 CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock); 7196 next = ddi_get_lbolt() + hz; 7197 7198 /* 7199 * Quick check for L2ARC devices. 7200 */ 7201 mutex_enter(&l2arc_dev_mtx); 7202 if (l2arc_ndev == 0) { 7203 mutex_exit(&l2arc_dev_mtx); 7204 continue; 7205 } 7206 mutex_exit(&l2arc_dev_mtx); 7207 begin = ddi_get_lbolt(); 7208 7209 /* 7210 * This selects the next l2arc device to write to, and in 7211 * doing so the next spa to feed from: dev->l2ad_spa. This 7212 * will return NULL if there are now no l2arc devices or if 7213 * they are all faulted. 7214 * 7215 * If a device is returned, its spa's config lock is also 7216 * held to prevent device removal. l2arc_dev_get_next() 7217 * will grab and release l2arc_dev_mtx. 7218 */ 7219 if ((dev = l2arc_dev_get_next()) == NULL) 7220 continue; 7221 7222 spa = dev->l2ad_spa; 7223 ASSERT3P(spa, !=, NULL); 7224 7225 /* 7226 * If the pool is read-only then force the feed thread to 7227 * sleep a little longer. 7228 */ 7229 if (!spa_writeable(spa)) { 7230 next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz; 7231 spa_config_exit(spa, SCL_L2ARC, dev); 7232 continue; 7233 } 7234 7235 /* 7236 * Avoid contributing to memory pressure. 7237 */ 7238 if (arc_reclaim_needed()) { 7239 ARCSTAT_BUMP(arcstat_l2_abort_lowmem); 7240 spa_config_exit(spa, SCL_L2ARC, dev); 7241 continue; 7242 } 7243 7244 ARCSTAT_BUMP(arcstat_l2_feeds); 7245 7246 size = l2arc_write_size(); 7247 7248 /* 7249 * Evict L2ARC buffers that will be overwritten. 7250 */ 7251 l2arc_evict(dev, size, B_FALSE); 7252 7253 /* 7254 * Write ARC buffers. 7255 */ 7256 wrote = l2arc_write_buffers(spa, dev, size); 7257 7258 /* 7259 * Calculate interval between writes. 7260 */ 7261 next = l2arc_write_interval(begin, size, wrote); 7262 spa_config_exit(spa, SCL_L2ARC, dev); 7263 } 7264 7265 l2arc_thread_exit = 0; 7266 cv_broadcast(&l2arc_feed_thr_cv); 7267 CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */ 7268 thread_exit(); 7269} 7270 7271boolean_t 7272l2arc_vdev_present(vdev_t *vd) 7273{ 7274 l2arc_dev_t *dev; 7275 7276 mutex_enter(&l2arc_dev_mtx); 7277 for (dev = list_head(l2arc_dev_list); dev != NULL; 7278 dev = list_next(l2arc_dev_list, dev)) { 7279 if (dev->l2ad_vdev == vd) 7280 break; 7281 } 7282 mutex_exit(&l2arc_dev_mtx); 7283 7284 return (dev != NULL); 7285} 7286 7287/* 7288 * Add a vdev for use by the L2ARC. By this point the spa has already 7289 * validated the vdev and opened it. 7290 */ 7291void 7292l2arc_add_vdev(spa_t *spa, vdev_t *vd) 7293{ 7294 l2arc_dev_t *adddev; 7295 7296 ASSERT(!l2arc_vdev_present(vd)); 7297 7298 vdev_ashift_optimize(vd); 7299 7300 /* 7301 * Create a new l2arc device entry. 7302 */ 7303 adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP); 7304 adddev->l2ad_spa = spa; 7305 adddev->l2ad_vdev = vd; 7306 adddev->l2ad_start = VDEV_LABEL_START_SIZE; 7307 adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd); 7308 adddev->l2ad_hand = adddev->l2ad_start; 7309 adddev->l2ad_first = B_TRUE; 7310 adddev->l2ad_writing = B_FALSE; 7311 7312 mutex_init(&adddev->l2ad_mtx, NULL, MUTEX_DEFAULT, NULL); 7313 /* 7314 * This is a list of all ARC buffers that are still valid on the 7315 * device. 7316 */ 7317 list_create(&adddev->l2ad_buflist, sizeof (arc_buf_hdr_t), 7318 offsetof(arc_buf_hdr_t, b_l2hdr.b_l2node)); 7319 7320 vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand); 7321 refcount_create(&adddev->l2ad_alloc); 7322 7323 /* 7324 * Add device to global list 7325 */ 7326 mutex_enter(&l2arc_dev_mtx); 7327 list_insert_head(l2arc_dev_list, adddev); 7328 atomic_inc_64(&l2arc_ndev); 7329 mutex_exit(&l2arc_dev_mtx); 7330} 7331 7332/* 7333 * Remove a vdev from the L2ARC. 7334 */ 7335void 7336l2arc_remove_vdev(vdev_t *vd) 7337{ 7338 l2arc_dev_t *dev, *nextdev, *remdev = NULL; 7339 7340 /* 7341 * Find the device by vdev 7342 */ 7343 mutex_enter(&l2arc_dev_mtx); 7344 for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) { 7345 nextdev = list_next(l2arc_dev_list, dev); 7346 if (vd == dev->l2ad_vdev) { 7347 remdev = dev; 7348 break; 7349 } 7350 } 7351 ASSERT3P(remdev, !=, NULL); 7352 7353 /* 7354 * Remove device from global list 7355 */ 7356 list_remove(l2arc_dev_list, remdev); 7357 l2arc_dev_last = NULL; /* may have been invalidated */ 7358 atomic_dec_64(&l2arc_ndev); 7359 mutex_exit(&l2arc_dev_mtx); 7360 7361 /* 7362 * Clear all buflists and ARC references. L2ARC device flush. 7363 */ 7364 l2arc_evict(remdev, 0, B_TRUE); 7365 list_destroy(&remdev->l2ad_buflist); 7366 mutex_destroy(&remdev->l2ad_mtx); 7367 refcount_destroy(&remdev->l2ad_alloc); 7368 kmem_free(remdev, sizeof (l2arc_dev_t)); 7369} 7370 7371void 7372l2arc_init(void) 7373{ 7374 l2arc_thread_exit = 0; 7375 l2arc_ndev = 0; 7376 l2arc_writes_sent = 0; 7377 l2arc_writes_done = 0; 7378 7379 mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL); 7380 cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL); 7381 mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL); 7382 mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL); 7383 7384 l2arc_dev_list = &L2ARC_dev_list; 7385 l2arc_free_on_write = &L2ARC_free_on_write; 7386 list_create(l2arc_dev_list, sizeof (l2arc_dev_t), 7387 offsetof(l2arc_dev_t, l2ad_node)); 7388 list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t), 7389 offsetof(l2arc_data_free_t, l2df_list_node)); 7390} 7391 7392void 7393l2arc_fini(void) 7394{ 7395 /* 7396 * This is called from dmu_fini(), which is called from spa_fini(); 7397 * Because of this, we can assume that all l2arc devices have 7398 * already been removed when the pools themselves were removed. 7399 */ 7400 7401 l2arc_do_free_on_write(); 7402 7403 mutex_destroy(&l2arc_feed_thr_lock); 7404 cv_destroy(&l2arc_feed_thr_cv); 7405 mutex_destroy(&l2arc_dev_mtx); 7406 mutex_destroy(&l2arc_free_on_write_mtx); 7407 7408 list_destroy(l2arc_dev_list); 7409 list_destroy(l2arc_free_on_write); 7410} 7411 7412void 7413l2arc_start(void) 7414{ 7415 if (!(spa_mode_global & FWRITE)) 7416 return; 7417 7418 (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0, 7419 TS_RUN, minclsyspri); 7420} 7421 7422void 7423l2arc_stop(void) 7424{ 7425 if (!(spa_mode_global & FWRITE)) 7426 return; 7427 7428 mutex_enter(&l2arc_feed_thr_lock); 7429 cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */ 7430 l2arc_thread_exit = 1; 7431 while (l2arc_thread_exit != 0) 7432 cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock); 7433 mutex_exit(&l2arc_feed_thr_lock); 7434} 7435