arc.c revision 314873
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 2616 l2arc_free_data_on_write(hdr->b_l1hdr.b_pdata, size, type); 2617} 2618 2619/* 2620 * Share the arc_buf_t's data with the hdr. Whenever we are sharing the 2621 * data buffer, we transfer the refcount ownership to the hdr and update 2622 * the appropriate kstats. 2623 */ 2624static void 2625arc_share_buf(arc_buf_hdr_t *hdr, arc_buf_t *buf) 2626{ 2627 arc_state_t *state = hdr->b_l1hdr.b_state; 2628 2629 ASSERT(!HDR_SHARED_DATA(hdr)); 2630 ASSERT(!arc_buf_is_shared(buf)); 2631 ASSERT3P(hdr->b_l1hdr.b_pdata, ==, NULL); 2632 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); 2633 2634 /* 2635 * Start sharing the data buffer. We transfer the 2636 * refcount ownership to the hdr since it always owns 2637 * the refcount whenever an arc_buf_t is shared. 2638 */ 2639 refcount_transfer_ownership(&state->arcs_size, buf, hdr); 2640 hdr->b_l1hdr.b_pdata = buf->b_data; 2641 arc_hdr_set_flags(hdr, ARC_FLAG_SHARED_DATA); 2642 2643 /* 2644 * Since we've transferred ownership to the hdr we need 2645 * to increment its compressed and uncompressed kstats and 2646 * decrement the overhead size. 2647 */ 2648 ARCSTAT_INCR(arcstat_compressed_size, arc_hdr_size(hdr)); 2649 ARCSTAT_INCR(arcstat_uncompressed_size, HDR_GET_LSIZE(hdr)); 2650 ARCSTAT_INCR(arcstat_overhead_size, -HDR_GET_LSIZE(hdr)); 2651} 2652 2653static void 2654arc_unshare_buf(arc_buf_hdr_t *hdr, arc_buf_t *buf) 2655{ 2656 arc_state_t *state = hdr->b_l1hdr.b_state; 2657 2658 ASSERT(HDR_SHARED_DATA(hdr)); 2659 ASSERT(arc_buf_is_shared(buf)); 2660 ASSERT3P(hdr->b_l1hdr.b_pdata, !=, NULL); 2661 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); 2662 2663 /* 2664 * We are no longer sharing this buffer so we need 2665 * to transfer its ownership to the rightful owner. 2666 */ 2667 refcount_transfer_ownership(&state->arcs_size, hdr, buf); 2668 arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA); 2669 hdr->b_l1hdr.b_pdata = NULL; 2670 2671 /* 2672 * Since the buffer is no longer shared between 2673 * the arc buf and the hdr, count it as overhead. 2674 */ 2675 ARCSTAT_INCR(arcstat_compressed_size, -arc_hdr_size(hdr)); 2676 ARCSTAT_INCR(arcstat_uncompressed_size, -HDR_GET_LSIZE(hdr)); 2677 ARCSTAT_INCR(arcstat_overhead_size, HDR_GET_LSIZE(hdr)); 2678} 2679 2680/* 2681 * Free up buf->b_data and if 'remove' is set, then pull the 2682 * arc_buf_t off of the the arc_buf_hdr_t's list and free it. 2683 */ 2684static void 2685arc_buf_destroy_impl(arc_buf_t *buf, boolean_t remove) 2686{ 2687 arc_buf_t **bufp; 2688 arc_buf_hdr_t *hdr = buf->b_hdr; 2689 uint64_t size = HDR_GET_LSIZE(hdr); 2690 boolean_t destroyed_buf_is_shared = arc_buf_is_shared(buf); 2691 2692 /* 2693 * Free up the data associated with the buf but only 2694 * if we're not sharing this with the hdr. If we are sharing 2695 * it with the hdr, then hdr will have performed the allocation 2696 * so allow it to do the free. 2697 */ 2698 if (buf->b_data != NULL) { 2699 /* 2700 * We're about to change the hdr's b_flags. We must either 2701 * hold the hash_lock or be undiscoverable. 2702 */ 2703 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); 2704 2705 arc_cksum_verify(buf); 2706#ifdef illumos 2707 arc_buf_unwatch(buf); 2708#endif 2709 2710 if (destroyed_buf_is_shared) { 2711 ASSERT(ARC_BUF_LAST(buf)); 2712 ASSERT(HDR_SHARED_DATA(hdr)); 2713 arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA); 2714 } else { 2715 arc_free_data_buf(hdr, buf->b_data, size, buf); 2716 ARCSTAT_INCR(arcstat_overhead_size, -size); 2717 } 2718 buf->b_data = NULL; 2719 2720 ASSERT(hdr->b_l1hdr.b_bufcnt > 0); 2721 hdr->b_l1hdr.b_bufcnt -= 1; 2722 } 2723 2724 /* only remove the buf if requested */ 2725 if (!remove) 2726 return; 2727 2728 /* remove the buf from the hdr list */ 2729 arc_buf_t *lastbuf = NULL; 2730 bufp = &hdr->b_l1hdr.b_buf; 2731 while (*bufp != NULL) { 2732 if (*bufp == buf) 2733 *bufp = buf->b_next; 2734 2735 /* 2736 * If we've removed a buffer in the middle of 2737 * the list then update the lastbuf and update 2738 * bufp. 2739 */ 2740 if (*bufp != NULL) { 2741 lastbuf = *bufp; 2742 bufp = &(*bufp)->b_next; 2743 } 2744 } 2745 buf->b_next = NULL; 2746 ASSERT3P(lastbuf, !=, buf); 2747 2748 /* 2749 * If the current arc_buf_t is sharing its data 2750 * buffer with the hdr, then reassign the hdr's 2751 * b_pdata to share it with the new buffer at the end 2752 * of the list. The shared buffer is always the last one 2753 * on the hdr's buffer list. 2754 */ 2755 if (destroyed_buf_is_shared && lastbuf != NULL) { 2756 ASSERT(ARC_BUF_LAST(buf)); 2757 ASSERT(ARC_BUF_LAST(lastbuf)); 2758 VERIFY(!arc_buf_is_shared(lastbuf)); 2759 2760 ASSERT3P(hdr->b_l1hdr.b_pdata, !=, NULL); 2761 arc_hdr_free_pdata(hdr); 2762 2763 /* 2764 * We must setup a new shared block between the 2765 * last buffer and the hdr. The data would have 2766 * been allocated by the arc buf so we need to transfer 2767 * ownership to the hdr since it's now being shared. 2768 */ 2769 arc_share_buf(hdr, lastbuf); 2770 } else if (HDR_SHARED_DATA(hdr)) { 2771 ASSERT(arc_buf_is_shared(lastbuf)); 2772 } 2773 2774 if (hdr->b_l1hdr.b_bufcnt == 0) 2775 arc_cksum_free(hdr); 2776 2777 /* clean up the buf */ 2778 buf->b_hdr = NULL; 2779 kmem_cache_free(buf_cache, buf); 2780} 2781 2782static void 2783arc_hdr_alloc_pdata(arc_buf_hdr_t *hdr) 2784{ 2785 ASSERT3U(HDR_GET_LSIZE(hdr), >, 0); 2786 ASSERT(HDR_HAS_L1HDR(hdr)); 2787 ASSERT(!HDR_SHARED_DATA(hdr)); 2788 2789 ASSERT3P(hdr->b_l1hdr.b_pdata, ==, NULL); 2790 hdr->b_l1hdr.b_pdata = arc_get_data_buf(hdr, arc_hdr_size(hdr), hdr); 2791 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS; 2792 ASSERT3P(hdr->b_l1hdr.b_pdata, !=, NULL); 2793 2794 ARCSTAT_INCR(arcstat_compressed_size, arc_hdr_size(hdr)); 2795 ARCSTAT_INCR(arcstat_uncompressed_size, HDR_GET_LSIZE(hdr)); 2796} 2797 2798static void 2799arc_hdr_free_pdata(arc_buf_hdr_t *hdr) 2800{ 2801 ASSERT(HDR_HAS_L1HDR(hdr)); 2802 ASSERT3P(hdr->b_l1hdr.b_pdata, !=, NULL); 2803 2804 /* 2805 * If the hdr is currently being written to the l2arc then 2806 * we defer freeing the data by adding it to the l2arc_free_on_write 2807 * list. The l2arc will free the data once it's finished 2808 * writing it to the l2arc device. 2809 */ 2810 if (HDR_L2_WRITING(hdr)) { 2811 arc_hdr_free_on_write(hdr); 2812 ARCSTAT_BUMP(arcstat_l2_free_on_write); 2813 } else { 2814 arc_free_data_buf(hdr, hdr->b_l1hdr.b_pdata, 2815 arc_hdr_size(hdr), hdr); 2816 } 2817 hdr->b_l1hdr.b_pdata = NULL; 2818 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS; 2819 2820 ARCSTAT_INCR(arcstat_compressed_size, -arc_hdr_size(hdr)); 2821 ARCSTAT_INCR(arcstat_uncompressed_size, -HDR_GET_LSIZE(hdr)); 2822} 2823 2824static arc_buf_hdr_t * 2825arc_hdr_alloc(uint64_t spa, int32_t psize, int32_t lsize, 2826 enum zio_compress compress, arc_buf_contents_t type) 2827{ 2828 arc_buf_hdr_t *hdr; 2829 2830 ASSERT3U(lsize, >, 0); 2831 VERIFY(type == ARC_BUFC_DATA || type == ARC_BUFC_METADATA); 2832 2833 hdr = kmem_cache_alloc(hdr_full_cache, KM_PUSHPAGE); 2834 ASSERT(HDR_EMPTY(hdr)); 2835 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL); 2836 ASSERT3P(hdr->b_l1hdr.b_thawed, ==, NULL); 2837 HDR_SET_PSIZE(hdr, psize); 2838 HDR_SET_LSIZE(hdr, lsize); 2839 hdr->b_spa = spa; 2840 hdr->b_type = type; 2841 hdr->b_flags = 0; 2842 arc_hdr_set_flags(hdr, arc_bufc_to_flags(type) | ARC_FLAG_HAS_L1HDR); 2843 arc_hdr_set_compress(hdr, compress); 2844 2845 hdr->b_l1hdr.b_state = arc_anon; 2846 hdr->b_l1hdr.b_arc_access = 0; 2847 hdr->b_l1hdr.b_bufcnt = 0; 2848 hdr->b_l1hdr.b_buf = NULL; 2849 2850 /* 2851 * Allocate the hdr's buffer. This will contain either 2852 * the compressed or uncompressed data depending on the block 2853 * it references and compressed arc enablement. 2854 */ 2855 arc_hdr_alloc_pdata(hdr); 2856 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 2857 2858 return (hdr); 2859} 2860 2861/* 2862 * Transition between the two allocation states for the arc_buf_hdr struct. 2863 * The arc_buf_hdr struct can be allocated with (hdr_full_cache) or without 2864 * (hdr_l2only_cache) the fields necessary for the L1 cache - the smaller 2865 * version is used when a cache buffer is only in the L2ARC in order to reduce 2866 * memory usage. 2867 */ 2868static arc_buf_hdr_t * 2869arc_hdr_realloc(arc_buf_hdr_t *hdr, kmem_cache_t *old, kmem_cache_t *new) 2870{ 2871 ASSERT(HDR_HAS_L2HDR(hdr)); 2872 2873 arc_buf_hdr_t *nhdr; 2874 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev; 2875 2876 ASSERT((old == hdr_full_cache && new == hdr_l2only_cache) || 2877 (old == hdr_l2only_cache && new == hdr_full_cache)); 2878 2879 nhdr = kmem_cache_alloc(new, KM_PUSHPAGE); 2880 2881 ASSERT(MUTEX_HELD(HDR_LOCK(hdr))); 2882 buf_hash_remove(hdr); 2883 2884 bcopy(hdr, nhdr, HDR_L2ONLY_SIZE); 2885 2886 if (new == hdr_full_cache) { 2887 arc_hdr_set_flags(nhdr, ARC_FLAG_HAS_L1HDR); 2888 /* 2889 * arc_access and arc_change_state need to be aware that a 2890 * header has just come out of L2ARC, so we set its state to 2891 * l2c_only even though it's about to change. 2892 */ 2893 nhdr->b_l1hdr.b_state = arc_l2c_only; 2894 2895 /* Verify previous threads set to NULL before freeing */ 2896 ASSERT3P(nhdr->b_l1hdr.b_pdata, ==, NULL); 2897 } else { 2898 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 2899 ASSERT0(hdr->b_l1hdr.b_bufcnt); 2900 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL); 2901 2902 /* 2903 * If we've reached here, We must have been called from 2904 * arc_evict_hdr(), as such we should have already been 2905 * removed from any ghost list we were previously on 2906 * (which protects us from racing with arc_evict_state), 2907 * thus no locking is needed during this check. 2908 */ 2909 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 2910 2911 /* 2912 * A buffer must not be moved into the arc_l2c_only 2913 * state if it's not finished being written out to the 2914 * l2arc device. Otherwise, the b_l1hdr.b_pdata field 2915 * might try to be accessed, even though it was removed. 2916 */ 2917 VERIFY(!HDR_L2_WRITING(hdr)); 2918 VERIFY3P(hdr->b_l1hdr.b_pdata, ==, NULL); 2919 2920#ifdef ZFS_DEBUG 2921 if (hdr->b_l1hdr.b_thawed != NULL) { 2922 kmem_free(hdr->b_l1hdr.b_thawed, 1); 2923 hdr->b_l1hdr.b_thawed = NULL; 2924 } 2925#endif 2926 2927 arc_hdr_clear_flags(nhdr, ARC_FLAG_HAS_L1HDR); 2928 } 2929 /* 2930 * The header has been reallocated so we need to re-insert it into any 2931 * lists it was on. 2932 */ 2933 (void) buf_hash_insert(nhdr, NULL); 2934 2935 ASSERT(list_link_active(&hdr->b_l2hdr.b_l2node)); 2936 2937 mutex_enter(&dev->l2ad_mtx); 2938 2939 /* 2940 * We must place the realloc'ed header back into the list at 2941 * the same spot. Otherwise, if it's placed earlier in the list, 2942 * l2arc_write_buffers() could find it during the function's 2943 * write phase, and try to write it out to the l2arc. 2944 */ 2945 list_insert_after(&dev->l2ad_buflist, hdr, nhdr); 2946 list_remove(&dev->l2ad_buflist, hdr); 2947 2948 mutex_exit(&dev->l2ad_mtx); 2949 2950 /* 2951 * Since we're using the pointer address as the tag when 2952 * incrementing and decrementing the l2ad_alloc refcount, we 2953 * must remove the old pointer (that we're about to destroy) and 2954 * add the new pointer to the refcount. Otherwise we'd remove 2955 * the wrong pointer address when calling arc_hdr_destroy() later. 2956 */ 2957 2958 (void) refcount_remove_many(&dev->l2ad_alloc, arc_hdr_size(hdr), hdr); 2959 (void) refcount_add_many(&dev->l2ad_alloc, arc_hdr_size(nhdr), nhdr); 2960 2961 buf_discard_identity(hdr); 2962 kmem_cache_free(old, hdr); 2963 2964 return (nhdr); 2965} 2966 2967/* 2968 * Allocate a new arc_buf_hdr_t and arc_buf_t and return the buf to the caller. 2969 * The buf is returned thawed since we expect the consumer to modify it. 2970 */ 2971arc_buf_t * 2972arc_alloc_buf(spa_t *spa, int32_t size, void *tag, arc_buf_contents_t type) 2973{ 2974 arc_buf_hdr_t *hdr = arc_hdr_alloc(spa_load_guid(spa), size, size, 2975 ZIO_COMPRESS_OFF, type); 2976 ASSERT(!MUTEX_HELD(HDR_LOCK(hdr))); 2977 arc_buf_t *buf = arc_buf_alloc_impl(hdr, tag); 2978 arc_buf_thaw(buf); 2979 return (buf); 2980} 2981 2982static void 2983arc_hdr_l2hdr_destroy(arc_buf_hdr_t *hdr) 2984{ 2985 l2arc_buf_hdr_t *l2hdr = &hdr->b_l2hdr; 2986 l2arc_dev_t *dev = l2hdr->b_dev; 2987 uint64_t asize = arc_hdr_size(hdr); 2988 2989 ASSERT(MUTEX_HELD(&dev->l2ad_mtx)); 2990 ASSERT(HDR_HAS_L2HDR(hdr)); 2991 2992 list_remove(&dev->l2ad_buflist, hdr); 2993 2994 ARCSTAT_INCR(arcstat_l2_asize, -asize); 2995 ARCSTAT_INCR(arcstat_l2_size, -HDR_GET_LSIZE(hdr)); 2996 2997 vdev_space_update(dev->l2ad_vdev, -asize, 0, 0); 2998 2999 (void) refcount_remove_many(&dev->l2ad_alloc, asize, hdr); 3000 arc_hdr_clear_flags(hdr, ARC_FLAG_HAS_L2HDR); 3001} 3002 3003static void 3004arc_hdr_destroy(arc_buf_hdr_t *hdr) 3005{ 3006 if (HDR_HAS_L1HDR(hdr)) { 3007 ASSERT(hdr->b_l1hdr.b_buf == NULL || 3008 hdr->b_l1hdr.b_bufcnt > 0); 3009 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 3010 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon); 3011 } 3012 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 3013 ASSERT(!HDR_IN_HASH_TABLE(hdr)); 3014 3015 if (!HDR_EMPTY(hdr)) 3016 buf_discard_identity(hdr); 3017 3018 if (HDR_HAS_L2HDR(hdr)) { 3019 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev; 3020 boolean_t buflist_held = MUTEX_HELD(&dev->l2ad_mtx); 3021 3022 if (!buflist_held) 3023 mutex_enter(&dev->l2ad_mtx); 3024 3025 /* 3026 * Even though we checked this conditional above, we 3027 * need to check this again now that we have the 3028 * l2ad_mtx. This is because we could be racing with 3029 * another thread calling l2arc_evict() which might have 3030 * destroyed this header's L2 portion as we were waiting 3031 * to acquire the l2ad_mtx. If that happens, we don't 3032 * want to re-destroy the header's L2 portion. 3033 */ 3034 if (HDR_HAS_L2HDR(hdr)) { 3035 l2arc_trim(hdr); 3036 arc_hdr_l2hdr_destroy(hdr); 3037 } 3038 3039 if (!buflist_held) 3040 mutex_exit(&dev->l2ad_mtx); 3041 } 3042 3043 if (HDR_HAS_L1HDR(hdr)) { 3044 arc_cksum_free(hdr); 3045 3046 while (hdr->b_l1hdr.b_buf != NULL) 3047 arc_buf_destroy_impl(hdr->b_l1hdr.b_buf, B_TRUE); 3048 3049#ifdef ZFS_DEBUG 3050 if (hdr->b_l1hdr.b_thawed != NULL) { 3051 kmem_free(hdr->b_l1hdr.b_thawed, 1); 3052 hdr->b_l1hdr.b_thawed = NULL; 3053 } 3054#endif 3055 3056 if (hdr->b_l1hdr.b_pdata != NULL) { 3057 arc_hdr_free_pdata(hdr); 3058 } 3059 } 3060 3061 ASSERT3P(hdr->b_hash_next, ==, NULL); 3062 if (HDR_HAS_L1HDR(hdr)) { 3063 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 3064 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL); 3065 kmem_cache_free(hdr_full_cache, hdr); 3066 } else { 3067 kmem_cache_free(hdr_l2only_cache, hdr); 3068 } 3069} 3070 3071void 3072arc_buf_destroy(arc_buf_t *buf, void* tag) 3073{ 3074 arc_buf_hdr_t *hdr = buf->b_hdr; 3075 kmutex_t *hash_lock = HDR_LOCK(hdr); 3076 3077 if (hdr->b_l1hdr.b_state == arc_anon) { 3078 ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1); 3079 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 3080 VERIFY0(remove_reference(hdr, NULL, tag)); 3081 arc_hdr_destroy(hdr); 3082 return; 3083 } 3084 3085 mutex_enter(hash_lock); 3086 ASSERT3P(hdr, ==, buf->b_hdr); 3087 ASSERT(hdr->b_l1hdr.b_bufcnt > 0); 3088 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 3089 ASSERT3P(hdr->b_l1hdr.b_state, !=, arc_anon); 3090 ASSERT3P(buf->b_data, !=, NULL); 3091 3092 (void) remove_reference(hdr, hash_lock, tag); 3093 arc_buf_destroy_impl(buf, B_TRUE); 3094 mutex_exit(hash_lock); 3095} 3096 3097int32_t 3098arc_buf_size(arc_buf_t *buf) 3099{ 3100 return (HDR_GET_LSIZE(buf->b_hdr)); 3101} 3102 3103/* 3104 * Evict the arc_buf_hdr that is provided as a parameter. The resultant 3105 * state of the header is dependent on it's state prior to entering this 3106 * function. The following transitions are possible: 3107 * 3108 * - arc_mru -> arc_mru_ghost 3109 * - arc_mfu -> arc_mfu_ghost 3110 * - arc_mru_ghost -> arc_l2c_only 3111 * - arc_mru_ghost -> deleted 3112 * - arc_mfu_ghost -> arc_l2c_only 3113 * - arc_mfu_ghost -> deleted 3114 */ 3115static int64_t 3116arc_evict_hdr(arc_buf_hdr_t *hdr, kmutex_t *hash_lock) 3117{ 3118 arc_state_t *evicted_state, *state; 3119 int64_t bytes_evicted = 0; 3120 3121 ASSERT(MUTEX_HELD(hash_lock)); 3122 ASSERT(HDR_HAS_L1HDR(hdr)); 3123 3124 state = hdr->b_l1hdr.b_state; 3125 if (GHOST_STATE(state)) { 3126 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 3127 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 3128 3129 /* 3130 * l2arc_write_buffers() relies on a header's L1 portion 3131 * (i.e. its b_pdata field) during its write phase. 3132 * Thus, we cannot push a header onto the arc_l2c_only 3133 * state (removing it's L1 piece) until the header is 3134 * done being written to the l2arc. 3135 */ 3136 if (HDR_HAS_L2HDR(hdr) && HDR_L2_WRITING(hdr)) { 3137 ARCSTAT_BUMP(arcstat_evict_l2_skip); 3138 return (bytes_evicted); 3139 } 3140 3141 ARCSTAT_BUMP(arcstat_deleted); 3142 bytes_evicted += HDR_GET_LSIZE(hdr); 3143 3144 DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, hdr); 3145 3146 ASSERT3P(hdr->b_l1hdr.b_pdata, ==, NULL); 3147 if (HDR_HAS_L2HDR(hdr)) { 3148 ASSERT(hdr->b_l1hdr.b_pdata == NULL); 3149 /* 3150 * This buffer is cached on the 2nd Level ARC; 3151 * don't destroy the header. 3152 */ 3153 arc_change_state(arc_l2c_only, hdr, hash_lock); 3154 /* 3155 * dropping from L1+L2 cached to L2-only, 3156 * realloc to remove the L1 header. 3157 */ 3158 hdr = arc_hdr_realloc(hdr, hdr_full_cache, 3159 hdr_l2only_cache); 3160 } else { 3161 ASSERT(hdr->b_l1hdr.b_pdata == NULL); 3162 arc_change_state(arc_anon, hdr, hash_lock); 3163 arc_hdr_destroy(hdr); 3164 } 3165 return (bytes_evicted); 3166 } 3167 3168 ASSERT(state == arc_mru || state == arc_mfu); 3169 evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost; 3170 3171 /* prefetch buffers have a minimum lifespan */ 3172 if (HDR_IO_IN_PROGRESS(hdr) || 3173 ((hdr->b_flags & (ARC_FLAG_PREFETCH | ARC_FLAG_INDIRECT)) && 3174 ddi_get_lbolt() - hdr->b_l1hdr.b_arc_access < 3175 arc_min_prefetch_lifespan)) { 3176 ARCSTAT_BUMP(arcstat_evict_skip); 3177 return (bytes_evicted); 3178 } 3179 3180 ASSERT0(refcount_count(&hdr->b_l1hdr.b_refcnt)); 3181 while (hdr->b_l1hdr.b_buf) { 3182 arc_buf_t *buf = hdr->b_l1hdr.b_buf; 3183 if (!mutex_tryenter(&buf->b_evict_lock)) { 3184 ARCSTAT_BUMP(arcstat_mutex_miss); 3185 break; 3186 } 3187 if (buf->b_data != NULL) 3188 bytes_evicted += HDR_GET_LSIZE(hdr); 3189 mutex_exit(&buf->b_evict_lock); 3190 arc_buf_destroy_impl(buf, B_TRUE); 3191 } 3192 3193 if (HDR_HAS_L2HDR(hdr)) { 3194 ARCSTAT_INCR(arcstat_evict_l2_cached, HDR_GET_LSIZE(hdr)); 3195 } else { 3196 if (l2arc_write_eligible(hdr->b_spa, hdr)) { 3197 ARCSTAT_INCR(arcstat_evict_l2_eligible, 3198 HDR_GET_LSIZE(hdr)); 3199 } else { 3200 ARCSTAT_INCR(arcstat_evict_l2_ineligible, 3201 HDR_GET_LSIZE(hdr)); 3202 } 3203 } 3204 3205 if (hdr->b_l1hdr.b_bufcnt == 0) { 3206 arc_cksum_free(hdr); 3207 3208 bytes_evicted += arc_hdr_size(hdr); 3209 3210 /* 3211 * If this hdr is being evicted and has a compressed 3212 * buffer then we discard it here before we change states. 3213 * This ensures that the accounting is updated correctly 3214 * in arc_free_data_buf(). 3215 */ 3216 arc_hdr_free_pdata(hdr); 3217 3218 arc_change_state(evicted_state, hdr, hash_lock); 3219 ASSERT(HDR_IN_HASH_TABLE(hdr)); 3220 arc_hdr_set_flags(hdr, ARC_FLAG_IN_HASH_TABLE); 3221 DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, hdr); 3222 } 3223 3224 return (bytes_evicted); 3225} 3226 3227static uint64_t 3228arc_evict_state_impl(multilist_t *ml, int idx, arc_buf_hdr_t *marker, 3229 uint64_t spa, int64_t bytes) 3230{ 3231 multilist_sublist_t *mls; 3232 uint64_t bytes_evicted = 0; 3233 arc_buf_hdr_t *hdr; 3234 kmutex_t *hash_lock; 3235 int evict_count = 0; 3236 3237 ASSERT3P(marker, !=, NULL); 3238 IMPLY(bytes < 0, bytes == ARC_EVICT_ALL); 3239 3240 mls = multilist_sublist_lock(ml, idx); 3241 3242 for (hdr = multilist_sublist_prev(mls, marker); hdr != NULL; 3243 hdr = multilist_sublist_prev(mls, marker)) { 3244 if ((bytes != ARC_EVICT_ALL && bytes_evicted >= bytes) || 3245 (evict_count >= zfs_arc_evict_batch_limit)) 3246 break; 3247 3248 /* 3249 * To keep our iteration location, move the marker 3250 * forward. Since we're not holding hdr's hash lock, we 3251 * must be very careful and not remove 'hdr' from the 3252 * sublist. Otherwise, other consumers might mistake the 3253 * 'hdr' as not being on a sublist when they call the 3254 * multilist_link_active() function (they all rely on 3255 * the hash lock protecting concurrent insertions and 3256 * removals). multilist_sublist_move_forward() was 3257 * specifically implemented to ensure this is the case 3258 * (only 'marker' will be removed and re-inserted). 3259 */ 3260 multilist_sublist_move_forward(mls, marker); 3261 3262 /* 3263 * The only case where the b_spa field should ever be 3264 * zero, is the marker headers inserted by 3265 * arc_evict_state(). It's possible for multiple threads 3266 * to be calling arc_evict_state() concurrently (e.g. 3267 * dsl_pool_close() and zio_inject_fault()), so we must 3268 * skip any markers we see from these other threads. 3269 */ 3270 if (hdr->b_spa == 0) 3271 continue; 3272 3273 /* we're only interested in evicting buffers of a certain spa */ 3274 if (spa != 0 && hdr->b_spa != spa) { 3275 ARCSTAT_BUMP(arcstat_evict_skip); 3276 continue; 3277 } 3278 3279 hash_lock = HDR_LOCK(hdr); 3280 3281 /* 3282 * We aren't calling this function from any code path 3283 * that would already be holding a hash lock, so we're 3284 * asserting on this assumption to be defensive in case 3285 * this ever changes. Without this check, it would be 3286 * possible to incorrectly increment arcstat_mutex_miss 3287 * below (e.g. if the code changed such that we called 3288 * this function with a hash lock held). 3289 */ 3290 ASSERT(!MUTEX_HELD(hash_lock)); 3291 3292 if (mutex_tryenter(hash_lock)) { 3293 uint64_t evicted = arc_evict_hdr(hdr, hash_lock); 3294 mutex_exit(hash_lock); 3295 3296 bytes_evicted += evicted; 3297 3298 /* 3299 * If evicted is zero, arc_evict_hdr() must have 3300 * decided to skip this header, don't increment 3301 * evict_count in this case. 3302 */ 3303 if (evicted != 0) 3304 evict_count++; 3305 3306 /* 3307 * If arc_size isn't overflowing, signal any 3308 * threads that might happen to be waiting. 3309 * 3310 * For each header evicted, we wake up a single 3311 * thread. If we used cv_broadcast, we could 3312 * wake up "too many" threads causing arc_size 3313 * to significantly overflow arc_c; since 3314 * arc_get_data_buf() doesn't check for overflow 3315 * when it's woken up (it doesn't because it's 3316 * possible for the ARC to be overflowing while 3317 * full of un-evictable buffers, and the 3318 * function should proceed in this case). 3319 * 3320 * If threads are left sleeping, due to not 3321 * using cv_broadcast, they will be woken up 3322 * just before arc_reclaim_thread() sleeps. 3323 */ 3324 mutex_enter(&arc_reclaim_lock); 3325 if (!arc_is_overflowing()) 3326 cv_signal(&arc_reclaim_waiters_cv); 3327 mutex_exit(&arc_reclaim_lock); 3328 } else { 3329 ARCSTAT_BUMP(arcstat_mutex_miss); 3330 } 3331 } 3332 3333 multilist_sublist_unlock(mls); 3334 3335 return (bytes_evicted); 3336} 3337 3338/* 3339 * Evict buffers from the given arc state, until we've removed the 3340 * specified number of bytes. Move the removed buffers to the 3341 * appropriate evict state. 3342 * 3343 * This function makes a "best effort". It skips over any buffers 3344 * it can't get a hash_lock on, and so, may not catch all candidates. 3345 * It may also return without evicting as much space as requested. 3346 * 3347 * If bytes is specified using the special value ARC_EVICT_ALL, this 3348 * will evict all available (i.e. unlocked and evictable) buffers from 3349 * the given arc state; which is used by arc_flush(). 3350 */ 3351static uint64_t 3352arc_evict_state(arc_state_t *state, uint64_t spa, int64_t bytes, 3353 arc_buf_contents_t type) 3354{ 3355 uint64_t total_evicted = 0; 3356 multilist_t *ml = &state->arcs_list[type]; 3357 int num_sublists; 3358 arc_buf_hdr_t **markers; 3359 3360 IMPLY(bytes < 0, bytes == ARC_EVICT_ALL); 3361 3362 num_sublists = multilist_get_num_sublists(ml); 3363 3364 /* 3365 * If we've tried to evict from each sublist, made some 3366 * progress, but still have not hit the target number of bytes 3367 * to evict, we want to keep trying. The markers allow us to 3368 * pick up where we left off for each individual sublist, rather 3369 * than starting from the tail each time. 3370 */ 3371 markers = kmem_zalloc(sizeof (*markers) * num_sublists, KM_SLEEP); 3372 for (int i = 0; i < num_sublists; i++) { 3373 markers[i] = kmem_cache_alloc(hdr_full_cache, KM_SLEEP); 3374 3375 /* 3376 * A b_spa of 0 is used to indicate that this header is 3377 * a marker. This fact is used in arc_adjust_type() and 3378 * arc_evict_state_impl(). 3379 */ 3380 markers[i]->b_spa = 0; 3381 3382 multilist_sublist_t *mls = multilist_sublist_lock(ml, i); 3383 multilist_sublist_insert_tail(mls, markers[i]); 3384 multilist_sublist_unlock(mls); 3385 } 3386 3387 /* 3388 * While we haven't hit our target number of bytes to evict, or 3389 * we're evicting all available buffers. 3390 */ 3391 while (total_evicted < bytes || bytes == ARC_EVICT_ALL) { 3392 /* 3393 * Start eviction using a randomly selected sublist, 3394 * this is to try and evenly balance eviction across all 3395 * sublists. Always starting at the same sublist 3396 * (e.g. index 0) would cause evictions to favor certain 3397 * sublists over others. 3398 */ 3399 int sublist_idx = multilist_get_random_index(ml); 3400 uint64_t scan_evicted = 0; 3401 3402 for (int i = 0; i < num_sublists; i++) { 3403 uint64_t bytes_remaining; 3404 uint64_t bytes_evicted; 3405 3406 if (bytes == ARC_EVICT_ALL) 3407 bytes_remaining = ARC_EVICT_ALL; 3408 else if (total_evicted < bytes) 3409 bytes_remaining = bytes - total_evicted; 3410 else 3411 break; 3412 3413 bytes_evicted = arc_evict_state_impl(ml, sublist_idx, 3414 markers[sublist_idx], spa, bytes_remaining); 3415 3416 scan_evicted += bytes_evicted; 3417 total_evicted += bytes_evicted; 3418 3419 /* we've reached the end, wrap to the beginning */ 3420 if (++sublist_idx >= num_sublists) 3421 sublist_idx = 0; 3422 } 3423 3424 /* 3425 * If we didn't evict anything during this scan, we have 3426 * no reason to believe we'll evict more during another 3427 * scan, so break the loop. 3428 */ 3429 if (scan_evicted == 0) { 3430 /* This isn't possible, let's make that obvious */ 3431 ASSERT3S(bytes, !=, 0); 3432 3433 /* 3434 * When bytes is ARC_EVICT_ALL, the only way to 3435 * break the loop is when scan_evicted is zero. 3436 * In that case, we actually have evicted enough, 3437 * so we don't want to increment the kstat. 3438 */ 3439 if (bytes != ARC_EVICT_ALL) { 3440 ASSERT3S(total_evicted, <, bytes); 3441 ARCSTAT_BUMP(arcstat_evict_not_enough); 3442 } 3443 3444 break; 3445 } 3446 } 3447 3448 for (int i = 0; i < num_sublists; i++) { 3449 multilist_sublist_t *mls = multilist_sublist_lock(ml, i); 3450 multilist_sublist_remove(mls, markers[i]); 3451 multilist_sublist_unlock(mls); 3452 3453 kmem_cache_free(hdr_full_cache, markers[i]); 3454 } 3455 kmem_free(markers, sizeof (*markers) * num_sublists); 3456 3457 return (total_evicted); 3458} 3459 3460/* 3461 * Flush all "evictable" data of the given type from the arc state 3462 * specified. This will not evict any "active" buffers (i.e. referenced). 3463 * 3464 * When 'retry' is set to B_FALSE, the function will make a single pass 3465 * over the state and evict any buffers that it can. Since it doesn't 3466 * continually retry the eviction, it might end up leaving some buffers 3467 * in the ARC due to lock misses. 3468 * 3469 * When 'retry' is set to B_TRUE, the function will continually retry the 3470 * eviction until *all* evictable buffers have been removed from the 3471 * state. As a result, if concurrent insertions into the state are 3472 * allowed (e.g. if the ARC isn't shutting down), this function might 3473 * wind up in an infinite loop, continually trying to evict buffers. 3474 */ 3475static uint64_t 3476arc_flush_state(arc_state_t *state, uint64_t spa, arc_buf_contents_t type, 3477 boolean_t retry) 3478{ 3479 uint64_t evicted = 0; 3480 3481 while (refcount_count(&state->arcs_esize[type]) != 0) { 3482 evicted += arc_evict_state(state, spa, ARC_EVICT_ALL, type); 3483 3484 if (!retry) 3485 break; 3486 } 3487 3488 return (evicted); 3489} 3490 3491/* 3492 * Evict the specified number of bytes from the state specified, 3493 * restricting eviction to the spa and type given. This function 3494 * prevents us from trying to evict more from a state's list than 3495 * is "evictable", and to skip evicting altogether when passed a 3496 * negative value for "bytes". In contrast, arc_evict_state() will 3497 * evict everything it can, when passed a negative value for "bytes". 3498 */ 3499static uint64_t 3500arc_adjust_impl(arc_state_t *state, uint64_t spa, int64_t bytes, 3501 arc_buf_contents_t type) 3502{ 3503 int64_t delta; 3504 3505 if (bytes > 0 && refcount_count(&state->arcs_esize[type]) > 0) { 3506 delta = MIN(refcount_count(&state->arcs_esize[type]), bytes); 3507 return (arc_evict_state(state, spa, delta, type)); 3508 } 3509 3510 return (0); 3511} 3512 3513/* 3514 * Evict metadata buffers from the cache, such that arc_meta_used is 3515 * capped by the arc_meta_limit tunable. 3516 */ 3517static uint64_t 3518arc_adjust_meta(void) 3519{ 3520 uint64_t total_evicted = 0; 3521 int64_t target; 3522 3523 /* 3524 * If we're over the meta limit, we want to evict enough 3525 * metadata to get back under the meta limit. We don't want to 3526 * evict so much that we drop the MRU below arc_p, though. If 3527 * we're over the meta limit more than we're over arc_p, we 3528 * evict some from the MRU here, and some from the MFU below. 3529 */ 3530 target = MIN((int64_t)(arc_meta_used - arc_meta_limit), 3531 (int64_t)(refcount_count(&arc_anon->arcs_size) + 3532 refcount_count(&arc_mru->arcs_size) - arc_p)); 3533 3534 total_evicted += arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA); 3535 3536 /* 3537 * Similar to the above, we want to evict enough bytes to get us 3538 * below the meta limit, but not so much as to drop us below the 3539 * space alloted to the MFU (which is defined as arc_c - arc_p). 3540 */ 3541 target = MIN((int64_t)(arc_meta_used - arc_meta_limit), 3542 (int64_t)(refcount_count(&arc_mfu->arcs_size) - (arc_c - arc_p))); 3543 3544 total_evicted += arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA); 3545 3546 return (total_evicted); 3547} 3548 3549/* 3550 * Return the type of the oldest buffer in the given arc state 3551 * 3552 * This function will select a random sublist of type ARC_BUFC_DATA and 3553 * a random sublist of type ARC_BUFC_METADATA. The tail of each sublist 3554 * is compared, and the type which contains the "older" buffer will be 3555 * returned. 3556 */ 3557static arc_buf_contents_t 3558arc_adjust_type(arc_state_t *state) 3559{ 3560 multilist_t *data_ml = &state->arcs_list[ARC_BUFC_DATA]; 3561 multilist_t *meta_ml = &state->arcs_list[ARC_BUFC_METADATA]; 3562 int data_idx = multilist_get_random_index(data_ml); 3563 int meta_idx = multilist_get_random_index(meta_ml); 3564 multilist_sublist_t *data_mls; 3565 multilist_sublist_t *meta_mls; 3566 arc_buf_contents_t type; 3567 arc_buf_hdr_t *data_hdr; 3568 arc_buf_hdr_t *meta_hdr; 3569 3570 /* 3571 * We keep the sublist lock until we're finished, to prevent 3572 * the headers from being destroyed via arc_evict_state(). 3573 */ 3574 data_mls = multilist_sublist_lock(data_ml, data_idx); 3575 meta_mls = multilist_sublist_lock(meta_ml, meta_idx); 3576 3577 /* 3578 * These two loops are to ensure we skip any markers that 3579 * might be at the tail of the lists due to arc_evict_state(). 3580 */ 3581 3582 for (data_hdr = multilist_sublist_tail(data_mls); data_hdr != NULL; 3583 data_hdr = multilist_sublist_prev(data_mls, data_hdr)) { 3584 if (data_hdr->b_spa != 0) 3585 break; 3586 } 3587 3588 for (meta_hdr = multilist_sublist_tail(meta_mls); meta_hdr != NULL; 3589 meta_hdr = multilist_sublist_prev(meta_mls, meta_hdr)) { 3590 if (meta_hdr->b_spa != 0) 3591 break; 3592 } 3593 3594 if (data_hdr == NULL && meta_hdr == NULL) { 3595 type = ARC_BUFC_DATA; 3596 } else if (data_hdr == NULL) { 3597 ASSERT3P(meta_hdr, !=, NULL); 3598 type = ARC_BUFC_METADATA; 3599 } else if (meta_hdr == NULL) { 3600 ASSERT3P(data_hdr, !=, NULL); 3601 type = ARC_BUFC_DATA; 3602 } else { 3603 ASSERT3P(data_hdr, !=, NULL); 3604 ASSERT3P(meta_hdr, !=, NULL); 3605 3606 /* The headers can't be on the sublist without an L1 header */ 3607 ASSERT(HDR_HAS_L1HDR(data_hdr)); 3608 ASSERT(HDR_HAS_L1HDR(meta_hdr)); 3609 3610 if (data_hdr->b_l1hdr.b_arc_access < 3611 meta_hdr->b_l1hdr.b_arc_access) { 3612 type = ARC_BUFC_DATA; 3613 } else { 3614 type = ARC_BUFC_METADATA; 3615 } 3616 } 3617 3618 multilist_sublist_unlock(meta_mls); 3619 multilist_sublist_unlock(data_mls); 3620 3621 return (type); 3622} 3623 3624/* 3625 * Evict buffers from the cache, such that arc_size is capped by arc_c. 3626 */ 3627static uint64_t 3628arc_adjust(void) 3629{ 3630 uint64_t total_evicted = 0; 3631 uint64_t bytes; 3632 int64_t target; 3633 3634 /* 3635 * If we're over arc_meta_limit, we want to correct that before 3636 * potentially evicting data buffers below. 3637 */ 3638 total_evicted += arc_adjust_meta(); 3639 3640 /* 3641 * Adjust MRU size 3642 * 3643 * If we're over the target cache size, we want to evict enough 3644 * from the list to get back to our target size. We don't want 3645 * to evict too much from the MRU, such that it drops below 3646 * arc_p. So, if we're over our target cache size more than 3647 * the MRU is over arc_p, we'll evict enough to get back to 3648 * arc_p here, and then evict more from the MFU below. 3649 */ 3650 target = MIN((int64_t)(arc_size - arc_c), 3651 (int64_t)(refcount_count(&arc_anon->arcs_size) + 3652 refcount_count(&arc_mru->arcs_size) + arc_meta_used - arc_p)); 3653 3654 /* 3655 * If we're below arc_meta_min, always prefer to evict data. 3656 * Otherwise, try to satisfy the requested number of bytes to 3657 * evict from the type which contains older buffers; in an 3658 * effort to keep newer buffers in the cache regardless of their 3659 * type. If we cannot satisfy the number of bytes from this 3660 * type, spill over into the next type. 3661 */ 3662 if (arc_adjust_type(arc_mru) == ARC_BUFC_METADATA && 3663 arc_meta_used > arc_meta_min) { 3664 bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA); 3665 total_evicted += bytes; 3666 3667 /* 3668 * If we couldn't evict our target number of bytes from 3669 * metadata, we try to get the rest from data. 3670 */ 3671 target -= bytes; 3672 3673 total_evicted += 3674 arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA); 3675 } else { 3676 bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA); 3677 total_evicted += bytes; 3678 3679 /* 3680 * If we couldn't evict our target number of bytes from 3681 * data, we try to get the rest from metadata. 3682 */ 3683 target -= bytes; 3684 3685 total_evicted += 3686 arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA); 3687 } 3688 3689 /* 3690 * Adjust MFU size 3691 * 3692 * Now that we've tried to evict enough from the MRU to get its 3693 * size back to arc_p, if we're still above the target cache 3694 * size, we evict the rest from the MFU. 3695 */ 3696 target = arc_size - arc_c; 3697 3698 if (arc_adjust_type(arc_mfu) == ARC_BUFC_METADATA && 3699 arc_meta_used > arc_meta_min) { 3700 bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA); 3701 total_evicted += bytes; 3702 3703 /* 3704 * If we couldn't evict our target number of bytes from 3705 * metadata, we try to get the rest from data. 3706 */ 3707 target -= bytes; 3708 3709 total_evicted += 3710 arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA); 3711 } else { 3712 bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA); 3713 total_evicted += bytes; 3714 3715 /* 3716 * If we couldn't evict our target number of bytes from 3717 * data, we try to get the rest from data. 3718 */ 3719 target -= bytes; 3720 3721 total_evicted += 3722 arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA); 3723 } 3724 3725 /* 3726 * Adjust ghost lists 3727 * 3728 * In addition to the above, the ARC also defines target values 3729 * for the ghost lists. The sum of the mru list and mru ghost 3730 * list should never exceed the target size of the cache, and 3731 * the sum of the mru list, mfu list, mru ghost list, and mfu 3732 * ghost list should never exceed twice the target size of the 3733 * cache. The following logic enforces these limits on the ghost 3734 * caches, and evicts from them as needed. 3735 */ 3736 target = refcount_count(&arc_mru->arcs_size) + 3737 refcount_count(&arc_mru_ghost->arcs_size) - arc_c; 3738 3739 bytes = arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_DATA); 3740 total_evicted += bytes; 3741 3742 target -= bytes; 3743 3744 total_evicted += 3745 arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_METADATA); 3746 3747 /* 3748 * We assume the sum of the mru list and mfu list is less than 3749 * or equal to arc_c (we enforced this above), which means we 3750 * can use the simpler of the two equations below: 3751 * 3752 * mru + mfu + mru ghost + mfu ghost <= 2 * arc_c 3753 * mru ghost + mfu ghost <= arc_c 3754 */ 3755 target = refcount_count(&arc_mru_ghost->arcs_size) + 3756 refcount_count(&arc_mfu_ghost->arcs_size) - arc_c; 3757 3758 bytes = arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_DATA); 3759 total_evicted += bytes; 3760 3761 target -= bytes; 3762 3763 total_evicted += 3764 arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_METADATA); 3765 3766 return (total_evicted); 3767} 3768 3769void 3770arc_flush(spa_t *spa, boolean_t retry) 3771{ 3772 uint64_t guid = 0; 3773 3774 /* 3775 * If retry is B_TRUE, a spa must not be specified since we have 3776 * no good way to determine if all of a spa's buffers have been 3777 * evicted from an arc state. 3778 */ 3779 ASSERT(!retry || spa == 0); 3780 3781 if (spa != NULL) 3782 guid = spa_load_guid(spa); 3783 3784 (void) arc_flush_state(arc_mru, guid, ARC_BUFC_DATA, retry); 3785 (void) arc_flush_state(arc_mru, guid, ARC_BUFC_METADATA, retry); 3786 3787 (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_DATA, retry); 3788 (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_METADATA, retry); 3789 3790 (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_DATA, retry); 3791 (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_METADATA, retry); 3792 3793 (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_DATA, retry); 3794 (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_METADATA, retry); 3795} 3796 3797void 3798arc_shrink(int64_t to_free) 3799{ 3800 if (arc_c > arc_c_min) { 3801 DTRACE_PROBE4(arc__shrink, uint64_t, arc_c, uint64_t, 3802 arc_c_min, uint64_t, arc_p, uint64_t, to_free); 3803 if (arc_c > arc_c_min + to_free) 3804 atomic_add_64(&arc_c, -to_free); 3805 else 3806 arc_c = arc_c_min; 3807 3808 atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift)); 3809 if (arc_c > arc_size) 3810 arc_c = MAX(arc_size, arc_c_min); 3811 if (arc_p > arc_c) 3812 arc_p = (arc_c >> 1); 3813 3814 DTRACE_PROBE2(arc__shrunk, uint64_t, arc_c, uint64_t, 3815 arc_p); 3816 3817 ASSERT(arc_c >= arc_c_min); 3818 ASSERT((int64_t)arc_p >= 0); 3819 } 3820 3821 if (arc_size > arc_c) { 3822 DTRACE_PROBE2(arc__shrink_adjust, uint64_t, arc_size, 3823 uint64_t, arc_c); 3824 (void) arc_adjust(); 3825 } 3826} 3827 3828static long needfree = 0; 3829 3830typedef enum free_memory_reason_t { 3831 FMR_UNKNOWN, 3832 FMR_NEEDFREE, 3833 FMR_LOTSFREE, 3834 FMR_SWAPFS_MINFREE, 3835 FMR_PAGES_PP_MAXIMUM, 3836 FMR_HEAP_ARENA, 3837 FMR_ZIO_ARENA, 3838 FMR_ZIO_FRAG, 3839} free_memory_reason_t; 3840 3841int64_t last_free_memory; 3842free_memory_reason_t last_free_reason; 3843 3844/* 3845 * Additional reserve of pages for pp_reserve. 3846 */ 3847int64_t arc_pages_pp_reserve = 64; 3848 3849/* 3850 * Additional reserve of pages for swapfs. 3851 */ 3852int64_t arc_swapfs_reserve = 64; 3853 3854/* 3855 * Return the amount of memory that can be consumed before reclaim will be 3856 * needed. Positive if there is sufficient free memory, negative indicates 3857 * the amount of memory that needs to be freed up. 3858 */ 3859static int64_t 3860arc_available_memory(void) 3861{ 3862 int64_t lowest = INT64_MAX; 3863 int64_t n; 3864 free_memory_reason_t r = FMR_UNKNOWN; 3865 3866#ifdef _KERNEL 3867 if (needfree > 0) { 3868 n = PAGESIZE * (-needfree); 3869 if (n < lowest) { 3870 lowest = n; 3871 r = FMR_NEEDFREE; 3872 } 3873 } 3874 3875 /* 3876 * Cooperate with pagedaemon when it's time for it to scan 3877 * and reclaim some pages. 3878 */ 3879 n = PAGESIZE * ((int64_t)freemem - zfs_arc_free_target); 3880 if (n < lowest) { 3881 lowest = n; 3882 r = FMR_LOTSFREE; 3883 } 3884 3885#ifdef illumos 3886 /* 3887 * check that we're out of range of the pageout scanner. It starts to 3888 * schedule paging if freemem is less than lotsfree and needfree. 3889 * lotsfree is the high-water mark for pageout, and needfree is the 3890 * number of needed free pages. We add extra pages here to make sure 3891 * the scanner doesn't start up while we're freeing memory. 3892 */ 3893 n = PAGESIZE * (freemem - lotsfree - needfree - desfree); 3894 if (n < lowest) { 3895 lowest = n; 3896 r = FMR_LOTSFREE; 3897 } 3898 3899 /* 3900 * check to make sure that swapfs has enough space so that anon 3901 * reservations can still succeed. anon_resvmem() checks that the 3902 * availrmem is greater than swapfs_minfree, and the number of reserved 3903 * swap pages. We also add a bit of extra here just to prevent 3904 * circumstances from getting really dire. 3905 */ 3906 n = PAGESIZE * (availrmem - swapfs_minfree - swapfs_reserve - 3907 desfree - arc_swapfs_reserve); 3908 if (n < lowest) { 3909 lowest = n; 3910 r = FMR_SWAPFS_MINFREE; 3911 } 3912 3913 3914 /* 3915 * Check that we have enough availrmem that memory locking (e.g., via 3916 * mlock(3C) or memcntl(2)) can still succeed. (pages_pp_maximum 3917 * stores the number of pages that cannot be locked; when availrmem 3918 * drops below pages_pp_maximum, page locking mechanisms such as 3919 * page_pp_lock() will fail.) 3920 */ 3921 n = PAGESIZE * (availrmem - pages_pp_maximum - 3922 arc_pages_pp_reserve); 3923 if (n < lowest) { 3924 lowest = n; 3925 r = FMR_PAGES_PP_MAXIMUM; 3926 } 3927 3928#endif /* illumos */ 3929#if defined(__i386) || !defined(UMA_MD_SMALL_ALLOC) 3930 /* 3931 * If we're on an i386 platform, it's possible that we'll exhaust the 3932 * kernel heap space before we ever run out of available physical 3933 * memory. Most checks of the size of the heap_area compare against 3934 * tune.t_minarmem, which is the minimum available real memory that we 3935 * can have in the system. However, this is generally fixed at 25 pages 3936 * which is so low that it's useless. In this comparison, we seek to 3937 * calculate the total heap-size, and reclaim if more than 3/4ths of the 3938 * heap is allocated. (Or, in the calculation, if less than 1/4th is 3939 * free) 3940 */ 3941 n = (int64_t)vmem_size(heap_arena, VMEM_FREE) - 3942 (vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC) >> 2); 3943 if (n < lowest) { 3944 lowest = n; 3945 r = FMR_HEAP_ARENA; 3946 } 3947#define zio_arena NULL 3948#else 3949#define zio_arena heap_arena 3950#endif 3951 3952 /* 3953 * If zio data pages are being allocated out of a separate heap segment, 3954 * then enforce that the size of available vmem for this arena remains 3955 * above about 1/16th free. 3956 * 3957 * Note: The 1/16th arena free requirement was put in place 3958 * to aggressively evict memory from the arc in order to avoid 3959 * memory fragmentation issues. 3960 */ 3961 if (zio_arena != NULL) { 3962 n = (int64_t)vmem_size(zio_arena, VMEM_FREE) - 3963 (vmem_size(zio_arena, VMEM_ALLOC) >> 4); 3964 if (n < lowest) { 3965 lowest = n; 3966 r = FMR_ZIO_ARENA; 3967 } 3968 } 3969 3970 /* 3971 * Above limits know nothing about real level of KVA fragmentation. 3972 * Start aggressive reclamation if too little sequential KVA left. 3973 */ 3974 if (lowest > 0) { 3975 n = (vmem_size(heap_arena, VMEM_MAXFREE) < zfs_max_recordsize) ? 3976 -((int64_t)vmem_size(heap_arena, VMEM_ALLOC) >> 4) : 3977 INT64_MAX; 3978 if (n < lowest) { 3979 lowest = n; 3980 r = FMR_ZIO_FRAG; 3981 } 3982 } 3983 3984#else /* _KERNEL */ 3985 /* Every 100 calls, free a small amount */ 3986 if (spa_get_random(100) == 0) 3987 lowest = -1024; 3988#endif /* _KERNEL */ 3989 3990 last_free_memory = lowest; 3991 last_free_reason = r; 3992 DTRACE_PROBE2(arc__available_memory, int64_t, lowest, int, r); 3993 return (lowest); 3994} 3995 3996 3997/* 3998 * Determine if the system is under memory pressure and is asking 3999 * to reclaim memory. A return value of B_TRUE indicates that the system 4000 * is under memory pressure and that the arc should adjust accordingly. 4001 */ 4002static boolean_t 4003arc_reclaim_needed(void) 4004{ 4005 return (arc_available_memory() < 0); 4006} 4007 4008extern kmem_cache_t *zio_buf_cache[]; 4009extern kmem_cache_t *zio_data_buf_cache[]; 4010extern kmem_cache_t *range_seg_cache; 4011 4012static __noinline void 4013arc_kmem_reap_now(void) 4014{ 4015 size_t i; 4016 kmem_cache_t *prev_cache = NULL; 4017 kmem_cache_t *prev_data_cache = NULL; 4018 4019 DTRACE_PROBE(arc__kmem_reap_start); 4020#ifdef _KERNEL 4021 if (arc_meta_used >= arc_meta_limit) { 4022 /* 4023 * We are exceeding our meta-data cache limit. 4024 * Purge some DNLC entries to release holds on meta-data. 4025 */ 4026 dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent); 4027 } 4028#if defined(__i386) 4029 /* 4030 * Reclaim unused memory from all kmem caches. 4031 */ 4032 kmem_reap(); 4033#endif 4034#endif 4035 4036 for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) { 4037 if (zio_buf_cache[i] != prev_cache) { 4038 prev_cache = zio_buf_cache[i]; 4039 kmem_cache_reap_now(zio_buf_cache[i]); 4040 } 4041 if (zio_data_buf_cache[i] != prev_data_cache) { 4042 prev_data_cache = zio_data_buf_cache[i]; 4043 kmem_cache_reap_now(zio_data_buf_cache[i]); 4044 } 4045 } 4046 kmem_cache_reap_now(buf_cache); 4047 kmem_cache_reap_now(hdr_full_cache); 4048 kmem_cache_reap_now(hdr_l2only_cache); 4049 kmem_cache_reap_now(range_seg_cache); 4050 4051#ifdef illumos 4052 if (zio_arena != NULL) { 4053 /* 4054 * Ask the vmem arena to reclaim unused memory from its 4055 * quantum caches. 4056 */ 4057 vmem_qcache_reap(zio_arena); 4058 } 4059#endif 4060 DTRACE_PROBE(arc__kmem_reap_end); 4061} 4062 4063/* 4064 * Threads can block in arc_get_data_buf() waiting for this thread to evict 4065 * enough data and signal them to proceed. When this happens, the threads in 4066 * arc_get_data_buf() are sleeping while holding the hash lock for their 4067 * particular arc header. Thus, we must be careful to never sleep on a 4068 * hash lock in this thread. This is to prevent the following deadlock: 4069 * 4070 * - Thread A sleeps on CV in arc_get_data_buf() holding hash lock "L", 4071 * waiting for the reclaim thread to signal it. 4072 * 4073 * - arc_reclaim_thread() tries to acquire hash lock "L" using mutex_enter, 4074 * fails, and goes to sleep forever. 4075 * 4076 * This possible deadlock is avoided by always acquiring a hash lock 4077 * using mutex_tryenter() from arc_reclaim_thread(). 4078 */ 4079static void 4080arc_reclaim_thread(void *dummy __unused) 4081{ 4082 hrtime_t growtime = 0; 4083 callb_cpr_t cpr; 4084 4085 CALLB_CPR_INIT(&cpr, &arc_reclaim_lock, callb_generic_cpr, FTAG); 4086 4087 mutex_enter(&arc_reclaim_lock); 4088 while (!arc_reclaim_thread_exit) { 4089 uint64_t evicted = 0; 4090 4091 /* 4092 * This is necessary in order for the mdb ::arc dcmd to 4093 * show up to date information. Since the ::arc command 4094 * does not call the kstat's update function, without 4095 * this call, the command may show stale stats for the 4096 * anon, mru, mru_ghost, mfu, and mfu_ghost lists. Even 4097 * with this change, the data might be up to 1 second 4098 * out of date; but that should suffice. The arc_state_t 4099 * structures can be queried directly if more accurate 4100 * information is needed. 4101 */ 4102 if (arc_ksp != NULL) 4103 arc_ksp->ks_update(arc_ksp, KSTAT_READ); 4104 4105 mutex_exit(&arc_reclaim_lock); 4106 4107 /* 4108 * We call arc_adjust() before (possibly) calling 4109 * arc_kmem_reap_now(), so that we can wake up 4110 * arc_get_data_buf() sooner. 4111 */ 4112 evicted = arc_adjust(); 4113 4114 int64_t free_memory = arc_available_memory(); 4115 if (free_memory < 0) { 4116 4117 arc_no_grow = B_TRUE; 4118 arc_warm = B_TRUE; 4119 4120 /* 4121 * Wait at least zfs_grow_retry (default 60) seconds 4122 * before considering growing. 4123 */ 4124 growtime = gethrtime() + SEC2NSEC(arc_grow_retry); 4125 4126 arc_kmem_reap_now(); 4127 4128 /* 4129 * If we are still low on memory, shrink the ARC 4130 * so that we have arc_shrink_min free space. 4131 */ 4132 free_memory = arc_available_memory(); 4133 4134 int64_t to_free = 4135 (arc_c >> arc_shrink_shift) - free_memory; 4136 if (to_free > 0) { 4137#ifdef _KERNEL 4138 to_free = MAX(to_free, ptob(needfree)); 4139#endif 4140 arc_shrink(to_free); 4141 } 4142 } else if (free_memory < arc_c >> arc_no_grow_shift) { 4143 arc_no_grow = B_TRUE; 4144 } else if (gethrtime() >= growtime) { 4145 arc_no_grow = B_FALSE; 4146 } 4147 4148 mutex_enter(&arc_reclaim_lock); 4149 4150 /* 4151 * If evicted is zero, we couldn't evict anything via 4152 * arc_adjust(). This could be due to hash lock 4153 * collisions, but more likely due to the majority of 4154 * arc buffers being unevictable. Therefore, even if 4155 * arc_size is above arc_c, another pass is unlikely to 4156 * be helpful and could potentially cause us to enter an 4157 * infinite loop. 4158 */ 4159 if (arc_size <= arc_c || evicted == 0) { 4160#ifdef _KERNEL 4161 needfree = 0; 4162#endif 4163 /* 4164 * We're either no longer overflowing, or we 4165 * can't evict anything more, so we should wake 4166 * up any threads before we go to sleep. 4167 */ 4168 cv_broadcast(&arc_reclaim_waiters_cv); 4169 4170 /* 4171 * Block until signaled, or after one second (we 4172 * might need to perform arc_kmem_reap_now() 4173 * even if we aren't being signalled) 4174 */ 4175 CALLB_CPR_SAFE_BEGIN(&cpr); 4176 (void) cv_timedwait_hires(&arc_reclaim_thread_cv, 4177 &arc_reclaim_lock, SEC2NSEC(1), MSEC2NSEC(1), 0); 4178 CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_lock); 4179 } 4180 } 4181 4182 arc_reclaim_thread_exit = B_FALSE; 4183 cv_broadcast(&arc_reclaim_thread_cv); 4184 CALLB_CPR_EXIT(&cpr); /* drops arc_reclaim_lock */ 4185 thread_exit(); 4186} 4187 4188static u_int arc_dnlc_evicts_arg; 4189extern struct vfsops zfs_vfsops; 4190 4191static void 4192arc_dnlc_evicts_thread(void *dummy __unused) 4193{ 4194 callb_cpr_t cpr; 4195 u_int percent; 4196 4197 CALLB_CPR_INIT(&cpr, &arc_dnlc_evicts_lock, callb_generic_cpr, FTAG); 4198 4199 mutex_enter(&arc_dnlc_evicts_lock); 4200 while (!arc_dnlc_evicts_thread_exit) { 4201 CALLB_CPR_SAFE_BEGIN(&cpr); 4202 (void) cv_wait(&arc_dnlc_evicts_cv, &arc_dnlc_evicts_lock); 4203 CALLB_CPR_SAFE_END(&cpr, &arc_dnlc_evicts_lock); 4204 if (arc_dnlc_evicts_arg != 0) { 4205 percent = arc_dnlc_evicts_arg; 4206 mutex_exit(&arc_dnlc_evicts_lock); 4207#ifdef _KERNEL 4208 vnlru_free(desiredvnodes * percent / 100, &zfs_vfsops); 4209#endif 4210 mutex_enter(&arc_dnlc_evicts_lock); 4211 /* 4212 * Clear our token only after vnlru_free() 4213 * pass is done, to avoid false queueing of 4214 * the requests. 4215 */ 4216 arc_dnlc_evicts_arg = 0; 4217 } 4218 } 4219 arc_dnlc_evicts_thread_exit = FALSE; 4220 cv_broadcast(&arc_dnlc_evicts_cv); 4221 CALLB_CPR_EXIT(&cpr); 4222 thread_exit(); 4223} 4224 4225void 4226dnlc_reduce_cache(void *arg) 4227{ 4228 u_int percent; 4229 4230 percent = (u_int)(uintptr_t)arg; 4231 mutex_enter(&arc_dnlc_evicts_lock); 4232 if (arc_dnlc_evicts_arg == 0) { 4233 arc_dnlc_evicts_arg = percent; 4234 cv_broadcast(&arc_dnlc_evicts_cv); 4235 } 4236 mutex_exit(&arc_dnlc_evicts_lock); 4237} 4238 4239/* 4240 * Adapt arc info given the number of bytes we are trying to add and 4241 * the state that we are comming from. This function is only called 4242 * when we are adding new content to the cache. 4243 */ 4244static void 4245arc_adapt(int bytes, arc_state_t *state) 4246{ 4247 int mult; 4248 uint64_t arc_p_min = (arc_c >> arc_p_min_shift); 4249 int64_t mrug_size = refcount_count(&arc_mru_ghost->arcs_size); 4250 int64_t mfug_size = refcount_count(&arc_mfu_ghost->arcs_size); 4251 4252 if (state == arc_l2c_only) 4253 return; 4254 4255 ASSERT(bytes > 0); 4256 /* 4257 * Adapt the target size of the MRU list: 4258 * - if we just hit in the MRU ghost list, then increase 4259 * the target size of the MRU list. 4260 * - if we just hit in the MFU ghost list, then increase 4261 * the target size of the MFU list by decreasing the 4262 * target size of the MRU list. 4263 */ 4264 if (state == arc_mru_ghost) { 4265 mult = (mrug_size >= mfug_size) ? 1 : (mfug_size / mrug_size); 4266 mult = MIN(mult, 10); /* avoid wild arc_p adjustment */ 4267 4268 arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult); 4269 } else if (state == arc_mfu_ghost) { 4270 uint64_t delta; 4271 4272 mult = (mfug_size >= mrug_size) ? 1 : (mrug_size / mfug_size); 4273 mult = MIN(mult, 10); 4274 4275 delta = MIN(bytes * mult, arc_p); 4276 arc_p = MAX(arc_p_min, arc_p - delta); 4277 } 4278 ASSERT((int64_t)arc_p >= 0); 4279 4280 if (arc_reclaim_needed()) { 4281 cv_signal(&arc_reclaim_thread_cv); 4282 return; 4283 } 4284 4285 if (arc_no_grow) 4286 return; 4287 4288 if (arc_c >= arc_c_max) 4289 return; 4290 4291 /* 4292 * If we're within (2 * maxblocksize) bytes of the target 4293 * cache size, increment the target cache size 4294 */ 4295 if (arc_size > arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) { 4296 DTRACE_PROBE1(arc__inc_adapt, int, bytes); 4297 atomic_add_64(&arc_c, (int64_t)bytes); 4298 if (arc_c > arc_c_max) 4299 arc_c = arc_c_max; 4300 else if (state == arc_anon) 4301 atomic_add_64(&arc_p, (int64_t)bytes); 4302 if (arc_p > arc_c) 4303 arc_p = arc_c; 4304 } 4305 ASSERT((int64_t)arc_p >= 0); 4306} 4307 4308/* 4309 * Check if arc_size has grown past our upper threshold, determined by 4310 * zfs_arc_overflow_shift. 4311 */ 4312static boolean_t 4313arc_is_overflowing(void) 4314{ 4315 /* Always allow at least one block of overflow */ 4316 uint64_t overflow = MAX(SPA_MAXBLOCKSIZE, 4317 arc_c >> zfs_arc_overflow_shift); 4318 4319 return (arc_size >= arc_c + overflow); 4320} 4321 4322/* 4323 * Allocate a block and return it to the caller. If we are hitting the 4324 * hard limit for the cache size, we must sleep, waiting for the eviction 4325 * thread to catch up. If we're past the target size but below the hard 4326 * limit, we'll only signal the reclaim thread and continue on. 4327 */ 4328static void * 4329arc_get_data_buf(arc_buf_hdr_t *hdr, uint64_t size, void *tag) 4330{ 4331 void *datap = NULL; 4332 arc_state_t *state = hdr->b_l1hdr.b_state; 4333 arc_buf_contents_t type = arc_buf_type(hdr); 4334 4335 arc_adapt(size, state); 4336 4337 /* 4338 * If arc_size is currently overflowing, and has grown past our 4339 * upper limit, we must be adding data faster than the evict 4340 * thread can evict. Thus, to ensure we don't compound the 4341 * problem by adding more data and forcing arc_size to grow even 4342 * further past it's target size, we halt and wait for the 4343 * eviction thread to catch up. 4344 * 4345 * It's also possible that the reclaim thread is unable to evict 4346 * enough buffers to get arc_size below the overflow limit (e.g. 4347 * due to buffers being un-evictable, or hash lock collisions). 4348 * In this case, we want to proceed regardless if we're 4349 * overflowing; thus we don't use a while loop here. 4350 */ 4351 if (arc_is_overflowing()) { 4352 mutex_enter(&arc_reclaim_lock); 4353 4354 /* 4355 * Now that we've acquired the lock, we may no longer be 4356 * over the overflow limit, lets check. 4357 * 4358 * We're ignoring the case of spurious wake ups. If that 4359 * were to happen, it'd let this thread consume an ARC 4360 * buffer before it should have (i.e. before we're under 4361 * the overflow limit and were signalled by the reclaim 4362 * thread). As long as that is a rare occurrence, it 4363 * shouldn't cause any harm. 4364 */ 4365 if (arc_is_overflowing()) { 4366 cv_signal(&arc_reclaim_thread_cv); 4367 cv_wait(&arc_reclaim_waiters_cv, &arc_reclaim_lock); 4368 } 4369 4370 mutex_exit(&arc_reclaim_lock); 4371 } 4372 4373 VERIFY3U(hdr->b_type, ==, type); 4374 if (type == ARC_BUFC_METADATA) { 4375 datap = zio_buf_alloc(size); 4376 arc_space_consume(size, ARC_SPACE_META); 4377 } else { 4378 ASSERT(type == ARC_BUFC_DATA); 4379 datap = zio_data_buf_alloc(size); 4380 arc_space_consume(size, ARC_SPACE_DATA); 4381 } 4382 4383 /* 4384 * Update the state size. Note that ghost states have a 4385 * "ghost size" and so don't need to be updated. 4386 */ 4387 if (!GHOST_STATE(state)) { 4388 4389 (void) refcount_add_many(&state->arcs_size, size, tag); 4390 4391 /* 4392 * If this is reached via arc_read, the link is 4393 * protected by the hash lock. If reached via 4394 * arc_buf_alloc, the header should not be accessed by 4395 * any other thread. And, if reached via arc_read_done, 4396 * the hash lock will protect it if it's found in the 4397 * hash table; otherwise no other thread should be 4398 * trying to [add|remove]_reference it. 4399 */ 4400 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) { 4401 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 4402 (void) refcount_add_many(&state->arcs_esize[type], 4403 size, tag); 4404 } 4405 4406 /* 4407 * If we are growing the cache, and we are adding anonymous 4408 * data, and we have outgrown arc_p, update arc_p 4409 */ 4410 if (arc_size < arc_c && hdr->b_l1hdr.b_state == arc_anon && 4411 (refcount_count(&arc_anon->arcs_size) + 4412 refcount_count(&arc_mru->arcs_size) > arc_p)) 4413 arc_p = MIN(arc_c, arc_p + size); 4414 } 4415 ARCSTAT_BUMP(arcstat_allocated); 4416 return (datap); 4417} 4418 4419/* 4420 * Free the arc data buffer. 4421 */ 4422static void 4423arc_free_data_buf(arc_buf_hdr_t *hdr, void *data, uint64_t size, void *tag) 4424{ 4425 arc_state_t *state = hdr->b_l1hdr.b_state; 4426 arc_buf_contents_t type = arc_buf_type(hdr); 4427 4428 /* protected by hash lock, if in the hash table */ 4429 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) { 4430 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 4431 ASSERT(state != arc_anon && state != arc_l2c_only); 4432 4433 (void) refcount_remove_many(&state->arcs_esize[type], 4434 size, tag); 4435 } 4436 (void) refcount_remove_many(&state->arcs_size, size, tag); 4437 4438 VERIFY3U(hdr->b_type, ==, type); 4439 if (type == ARC_BUFC_METADATA) { 4440 zio_buf_free(data, size); 4441 arc_space_return(size, ARC_SPACE_META); 4442 } else { 4443 ASSERT(type == ARC_BUFC_DATA); 4444 zio_data_buf_free(data, size); 4445 arc_space_return(size, ARC_SPACE_DATA); 4446 } 4447} 4448 4449/* 4450 * This routine is called whenever a buffer is accessed. 4451 * NOTE: the hash lock is dropped in this function. 4452 */ 4453static void 4454arc_access(arc_buf_hdr_t *hdr, kmutex_t *hash_lock) 4455{ 4456 clock_t now; 4457 4458 ASSERT(MUTEX_HELD(hash_lock)); 4459 ASSERT(HDR_HAS_L1HDR(hdr)); 4460 4461 if (hdr->b_l1hdr.b_state == arc_anon) { 4462 /* 4463 * This buffer is not in the cache, and does not 4464 * appear in our "ghost" list. Add the new buffer 4465 * to the MRU state. 4466 */ 4467 4468 ASSERT0(hdr->b_l1hdr.b_arc_access); 4469 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 4470 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr); 4471 arc_change_state(arc_mru, hdr, hash_lock); 4472 4473 } else if (hdr->b_l1hdr.b_state == arc_mru) { 4474 now = ddi_get_lbolt(); 4475 4476 /* 4477 * If this buffer is here because of a prefetch, then either: 4478 * - clear the flag if this is a "referencing" read 4479 * (any subsequent access will bump this into the MFU state). 4480 * or 4481 * - move the buffer to the head of the list if this is 4482 * another prefetch (to make it less likely to be evicted). 4483 */ 4484 if (HDR_PREFETCH(hdr)) { 4485 if (refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) { 4486 /* link protected by hash lock */ 4487 ASSERT(multilist_link_active( 4488 &hdr->b_l1hdr.b_arc_node)); 4489 } else { 4490 arc_hdr_clear_flags(hdr, ARC_FLAG_PREFETCH); 4491 ARCSTAT_BUMP(arcstat_mru_hits); 4492 } 4493 hdr->b_l1hdr.b_arc_access = now; 4494 return; 4495 } 4496 4497 /* 4498 * This buffer has been "accessed" only once so far, 4499 * but it is still in the cache. Move it to the MFU 4500 * state. 4501 */ 4502 if (now > hdr->b_l1hdr.b_arc_access + ARC_MINTIME) { 4503 /* 4504 * More than 125ms have passed since we 4505 * instantiated this buffer. Move it to the 4506 * most frequently used state. 4507 */ 4508 hdr->b_l1hdr.b_arc_access = now; 4509 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 4510 arc_change_state(arc_mfu, hdr, hash_lock); 4511 } 4512 ARCSTAT_BUMP(arcstat_mru_hits); 4513 } else if (hdr->b_l1hdr.b_state == arc_mru_ghost) { 4514 arc_state_t *new_state; 4515 /* 4516 * This buffer has been "accessed" recently, but 4517 * was evicted from the cache. Move it to the 4518 * MFU state. 4519 */ 4520 4521 if (HDR_PREFETCH(hdr)) { 4522 new_state = arc_mru; 4523 if (refcount_count(&hdr->b_l1hdr.b_refcnt) > 0) 4524 arc_hdr_clear_flags(hdr, ARC_FLAG_PREFETCH); 4525 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr); 4526 } else { 4527 new_state = arc_mfu; 4528 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 4529 } 4530 4531 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 4532 arc_change_state(new_state, hdr, hash_lock); 4533 4534 ARCSTAT_BUMP(arcstat_mru_ghost_hits); 4535 } else if (hdr->b_l1hdr.b_state == arc_mfu) { 4536 /* 4537 * This buffer has been accessed more than once and is 4538 * still in the cache. Keep it in the MFU state. 4539 * 4540 * NOTE: an add_reference() that occurred when we did 4541 * the arc_read() will have kicked this off the list. 4542 * If it was a prefetch, we will explicitly move it to 4543 * the head of the list now. 4544 */ 4545 if ((HDR_PREFETCH(hdr)) != 0) { 4546 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 4547 /* link protected by hash_lock */ 4548 ASSERT(multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 4549 } 4550 ARCSTAT_BUMP(arcstat_mfu_hits); 4551 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 4552 } else if (hdr->b_l1hdr.b_state == arc_mfu_ghost) { 4553 arc_state_t *new_state = arc_mfu; 4554 /* 4555 * This buffer has been accessed more than once but has 4556 * been evicted from the cache. Move it back to the 4557 * MFU state. 4558 */ 4559 4560 if (HDR_PREFETCH(hdr)) { 4561 /* 4562 * This is a prefetch access... 4563 * move this block back to the MRU state. 4564 */ 4565 ASSERT0(refcount_count(&hdr->b_l1hdr.b_refcnt)); 4566 new_state = arc_mru; 4567 } 4568 4569 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 4570 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 4571 arc_change_state(new_state, hdr, hash_lock); 4572 4573 ARCSTAT_BUMP(arcstat_mfu_ghost_hits); 4574 } else if (hdr->b_l1hdr.b_state == arc_l2c_only) { 4575 /* 4576 * This buffer is on the 2nd Level ARC. 4577 */ 4578 4579 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 4580 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 4581 arc_change_state(arc_mfu, hdr, hash_lock); 4582 } else { 4583 ASSERT(!"invalid arc state"); 4584 } 4585} 4586 4587/* a generic arc_done_func_t which you can use */ 4588/* ARGSUSED */ 4589void 4590arc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg) 4591{ 4592 if (zio == NULL || zio->io_error == 0) 4593 bcopy(buf->b_data, arg, HDR_GET_LSIZE(buf->b_hdr)); 4594 arc_buf_destroy(buf, arg); 4595} 4596 4597/* a generic arc_done_func_t */ 4598void 4599arc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg) 4600{ 4601 arc_buf_t **bufp = arg; 4602 if (zio && zio->io_error) { 4603 arc_buf_destroy(buf, arg); 4604 *bufp = NULL; 4605 } else { 4606 *bufp = buf; 4607 ASSERT(buf->b_data); 4608 } 4609} 4610 4611static void 4612arc_hdr_verify(arc_buf_hdr_t *hdr, blkptr_t *bp) 4613{ 4614 if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp)) { 4615 ASSERT3U(HDR_GET_PSIZE(hdr), ==, 0); 4616 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF); 4617 } else { 4618 if (HDR_COMPRESSION_ENABLED(hdr)) { 4619 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, 4620 BP_GET_COMPRESS(bp)); 4621 } 4622 ASSERT3U(HDR_GET_LSIZE(hdr), ==, BP_GET_LSIZE(bp)); 4623 ASSERT3U(HDR_GET_PSIZE(hdr), ==, BP_GET_PSIZE(bp)); 4624 } 4625} 4626 4627static void 4628arc_read_done(zio_t *zio) 4629{ 4630 arc_buf_hdr_t *hdr = zio->io_private; 4631 arc_buf_t *abuf = NULL; /* buffer we're assigning to callback */ 4632 kmutex_t *hash_lock = NULL; 4633 arc_callback_t *callback_list, *acb; 4634 int freeable = B_FALSE; 4635 4636 /* 4637 * The hdr was inserted into hash-table and removed from lists 4638 * prior to starting I/O. We should find this header, since 4639 * it's in the hash table, and it should be legit since it's 4640 * not possible to evict it during the I/O. The only possible 4641 * reason for it not to be found is if we were freed during the 4642 * read. 4643 */ 4644 if (HDR_IN_HASH_TABLE(hdr)) { 4645 ASSERT3U(hdr->b_birth, ==, BP_PHYSICAL_BIRTH(zio->io_bp)); 4646 ASSERT3U(hdr->b_dva.dva_word[0], ==, 4647 BP_IDENTITY(zio->io_bp)->dva_word[0]); 4648 ASSERT3U(hdr->b_dva.dva_word[1], ==, 4649 BP_IDENTITY(zio->io_bp)->dva_word[1]); 4650 4651 arc_buf_hdr_t *found = buf_hash_find(hdr->b_spa, zio->io_bp, 4652 &hash_lock); 4653 4654 ASSERT((found == hdr && 4655 DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) || 4656 (found == hdr && HDR_L2_READING(hdr))); 4657 ASSERT3P(hash_lock, !=, NULL); 4658 } 4659 4660 if (zio->io_error == 0) { 4661 /* byteswap if necessary */ 4662 if (BP_SHOULD_BYTESWAP(zio->io_bp)) { 4663 if (BP_GET_LEVEL(zio->io_bp) > 0) { 4664 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_UINT64; 4665 } else { 4666 hdr->b_l1hdr.b_byteswap = 4667 DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp)); 4668 } 4669 } else { 4670 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS; 4671 } 4672 } 4673 4674 arc_hdr_clear_flags(hdr, ARC_FLAG_L2_EVICTED); 4675 if (l2arc_noprefetch && HDR_PREFETCH(hdr)) 4676 arc_hdr_clear_flags(hdr, ARC_FLAG_L2CACHE); 4677 4678 callback_list = hdr->b_l1hdr.b_acb; 4679 ASSERT3P(callback_list, !=, NULL); 4680 4681 if (hash_lock && zio->io_error == 0 && 4682 hdr->b_l1hdr.b_state == arc_anon) { 4683 /* 4684 * Only call arc_access on anonymous buffers. This is because 4685 * if we've issued an I/O for an evicted buffer, we've already 4686 * called arc_access (to prevent any simultaneous readers from 4687 * getting confused). 4688 */ 4689 arc_access(hdr, hash_lock); 4690 } 4691 4692 /* create copies of the data buffer for the callers */ 4693 for (acb = callback_list; acb; acb = acb->acb_next) { 4694 if (acb->acb_done != NULL) { 4695 /* 4696 * If we're here, then this must be a demand read 4697 * since prefetch requests don't have callbacks. 4698 * If a read request has a callback (i.e. acb_done is 4699 * not NULL), then we decompress the data for the 4700 * first request and clone the rest. This avoids 4701 * having to waste cpu resources decompressing data 4702 * that nobody is explicitly waiting to read. 4703 */ 4704 if (abuf == NULL) { 4705 acb->acb_buf = arc_buf_alloc_impl(hdr, 4706 acb->acb_private); 4707 if (zio->io_error == 0) { 4708 zio->io_error = 4709 arc_decompress(acb->acb_buf); 4710 } 4711 abuf = acb->acb_buf; 4712 } else { 4713 add_reference(hdr, acb->acb_private); 4714 acb->acb_buf = arc_buf_clone(abuf); 4715 } 4716 } 4717 } 4718 hdr->b_l1hdr.b_acb = NULL; 4719 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS); 4720 if (abuf == NULL) { 4721 /* 4722 * This buffer didn't have a callback so it must 4723 * be a prefetch. 4724 */ 4725 ASSERT(HDR_PREFETCH(hdr)); 4726 ASSERT0(hdr->b_l1hdr.b_bufcnt); 4727 ASSERT3P(hdr->b_l1hdr.b_pdata, !=, NULL); 4728 } 4729 4730 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt) || 4731 callback_list != NULL); 4732 4733 if (zio->io_error == 0) { 4734 arc_hdr_verify(hdr, zio->io_bp); 4735 } else { 4736 arc_hdr_set_flags(hdr, ARC_FLAG_IO_ERROR); 4737 if (hdr->b_l1hdr.b_state != arc_anon) 4738 arc_change_state(arc_anon, hdr, hash_lock); 4739 if (HDR_IN_HASH_TABLE(hdr)) 4740 buf_hash_remove(hdr); 4741 freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt); 4742 } 4743 4744 /* 4745 * Broadcast before we drop the hash_lock to avoid the possibility 4746 * that the hdr (and hence the cv) might be freed before we get to 4747 * the cv_broadcast(). 4748 */ 4749 cv_broadcast(&hdr->b_l1hdr.b_cv); 4750 4751 if (hash_lock != NULL) { 4752 mutex_exit(hash_lock); 4753 } else { 4754 /* 4755 * This block was freed while we waited for the read to 4756 * complete. It has been removed from the hash table and 4757 * moved to the anonymous state (so that it won't show up 4758 * in the cache). 4759 */ 4760 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon); 4761 freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt); 4762 } 4763 4764 /* execute each callback and free its structure */ 4765 while ((acb = callback_list) != NULL) { 4766 if (acb->acb_done) 4767 acb->acb_done(zio, acb->acb_buf, acb->acb_private); 4768 4769 if (acb->acb_zio_dummy != NULL) { 4770 acb->acb_zio_dummy->io_error = zio->io_error; 4771 zio_nowait(acb->acb_zio_dummy); 4772 } 4773 4774 callback_list = acb->acb_next; 4775 kmem_free(acb, sizeof (arc_callback_t)); 4776 } 4777 4778 if (freeable) 4779 arc_hdr_destroy(hdr); 4780} 4781 4782/* 4783 * "Read" the block at the specified DVA (in bp) via the 4784 * cache. If the block is found in the cache, invoke the provided 4785 * callback immediately and return. Note that the `zio' parameter 4786 * in the callback will be NULL in this case, since no IO was 4787 * required. If the block is not in the cache pass the read request 4788 * on to the spa with a substitute callback function, so that the 4789 * requested block will be added to the cache. 4790 * 4791 * If a read request arrives for a block that has a read in-progress, 4792 * either wait for the in-progress read to complete (and return the 4793 * results); or, if this is a read with a "done" func, add a record 4794 * to the read to invoke the "done" func when the read completes, 4795 * and return; or just return. 4796 * 4797 * arc_read_done() will invoke all the requested "done" functions 4798 * for readers of this block. 4799 */ 4800int 4801arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_done_func_t *done, 4802 void *private, zio_priority_t priority, int zio_flags, 4803 arc_flags_t *arc_flags, const zbookmark_phys_t *zb) 4804{ 4805 arc_buf_hdr_t *hdr = NULL; 4806 kmutex_t *hash_lock = NULL; 4807 zio_t *rzio; 4808 uint64_t guid = spa_load_guid(spa); 4809 4810 ASSERT(!BP_IS_EMBEDDED(bp) || 4811 BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA); 4812 4813top: 4814 if (!BP_IS_EMBEDDED(bp)) { 4815 /* 4816 * Embedded BP's have no DVA and require no I/O to "read". 4817 * Create an anonymous arc buf to back it. 4818 */ 4819 hdr = buf_hash_find(guid, bp, &hash_lock); 4820 } 4821 4822 if (hdr != NULL && HDR_HAS_L1HDR(hdr) && hdr->b_l1hdr.b_pdata != NULL) { 4823 arc_buf_t *buf = NULL; 4824 *arc_flags |= ARC_FLAG_CACHED; 4825 4826 if (HDR_IO_IN_PROGRESS(hdr)) { 4827 4828 if ((hdr->b_flags & ARC_FLAG_PRIO_ASYNC_READ) && 4829 priority == ZIO_PRIORITY_SYNC_READ) { 4830 /* 4831 * This sync read must wait for an 4832 * in-progress async read (e.g. a predictive 4833 * prefetch). Async reads are queued 4834 * separately at the vdev_queue layer, so 4835 * this is a form of priority inversion. 4836 * Ideally, we would "inherit" the demand 4837 * i/o's priority by moving the i/o from 4838 * the async queue to the synchronous queue, 4839 * but there is currently no mechanism to do 4840 * so. Track this so that we can evaluate 4841 * the magnitude of this potential performance 4842 * problem. 4843 * 4844 * Note that if the prefetch i/o is already 4845 * active (has been issued to the device), 4846 * the prefetch improved performance, because 4847 * we issued it sooner than we would have 4848 * without the prefetch. 4849 */ 4850 DTRACE_PROBE1(arc__sync__wait__for__async, 4851 arc_buf_hdr_t *, hdr); 4852 ARCSTAT_BUMP(arcstat_sync_wait_for_async); 4853 } 4854 if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) { 4855 arc_hdr_clear_flags(hdr, 4856 ARC_FLAG_PREDICTIVE_PREFETCH); 4857 } 4858 4859 if (*arc_flags & ARC_FLAG_WAIT) { 4860 cv_wait(&hdr->b_l1hdr.b_cv, hash_lock); 4861 mutex_exit(hash_lock); 4862 goto top; 4863 } 4864 ASSERT(*arc_flags & ARC_FLAG_NOWAIT); 4865 4866 if (done) { 4867 arc_callback_t *acb = NULL; 4868 4869 acb = kmem_zalloc(sizeof (arc_callback_t), 4870 KM_SLEEP); 4871 acb->acb_done = done; 4872 acb->acb_private = private; 4873 if (pio != NULL) 4874 acb->acb_zio_dummy = zio_null(pio, 4875 spa, NULL, NULL, NULL, zio_flags); 4876 4877 ASSERT3P(acb->acb_done, !=, NULL); 4878 acb->acb_next = hdr->b_l1hdr.b_acb; 4879 hdr->b_l1hdr.b_acb = acb; 4880 mutex_exit(hash_lock); 4881 return (0); 4882 } 4883 mutex_exit(hash_lock); 4884 return (0); 4885 } 4886 4887 ASSERT(hdr->b_l1hdr.b_state == arc_mru || 4888 hdr->b_l1hdr.b_state == arc_mfu); 4889 4890 if (done) { 4891 if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) { 4892 /* 4893 * This is a demand read which does not have to 4894 * wait for i/o because we did a predictive 4895 * prefetch i/o for it, which has completed. 4896 */ 4897 DTRACE_PROBE1( 4898 arc__demand__hit__predictive__prefetch, 4899 arc_buf_hdr_t *, hdr); 4900 ARCSTAT_BUMP( 4901 arcstat_demand_hit_predictive_prefetch); 4902 arc_hdr_clear_flags(hdr, 4903 ARC_FLAG_PREDICTIVE_PREFETCH); 4904 } 4905 ASSERT(!BP_IS_EMBEDDED(bp) || !BP_IS_HOLE(bp)); 4906 4907 /* 4908 * If this block is already in use, create a new 4909 * copy of the data so that we will be guaranteed 4910 * that arc_release() will always succeed. 4911 */ 4912 buf = hdr->b_l1hdr.b_buf; 4913 if (buf == NULL) { 4914 ASSERT0(refcount_count(&hdr->b_l1hdr.b_refcnt)); 4915 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL); 4916 buf = arc_buf_alloc_impl(hdr, private); 4917 VERIFY0(arc_decompress(buf)); 4918 } else { 4919 add_reference(hdr, private); 4920 buf = arc_buf_clone(buf); 4921 } 4922 ASSERT3P(buf->b_data, !=, NULL); 4923 4924 } else if (*arc_flags & ARC_FLAG_PREFETCH && 4925 refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) { 4926 arc_hdr_set_flags(hdr, ARC_FLAG_PREFETCH); 4927 } 4928 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr); 4929 arc_access(hdr, hash_lock); 4930 if (*arc_flags & ARC_FLAG_L2CACHE) 4931 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE); 4932 mutex_exit(hash_lock); 4933 ARCSTAT_BUMP(arcstat_hits); 4934 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr), 4935 demand, prefetch, !HDR_ISTYPE_METADATA(hdr), 4936 data, metadata, hits); 4937 4938 if (done) 4939 done(NULL, buf, private); 4940 } else { 4941 uint64_t lsize = BP_GET_LSIZE(bp); 4942 uint64_t psize = BP_GET_PSIZE(bp); 4943 arc_callback_t *acb; 4944 vdev_t *vd = NULL; 4945 uint64_t addr = 0; 4946 boolean_t devw = B_FALSE; 4947 uint64_t size; 4948 4949 if (hdr == NULL) { 4950 /* this block is not in the cache */ 4951 arc_buf_hdr_t *exists = NULL; 4952 arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp); 4953 hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize, 4954 BP_GET_COMPRESS(bp), type); 4955 4956 if (!BP_IS_EMBEDDED(bp)) { 4957 hdr->b_dva = *BP_IDENTITY(bp); 4958 hdr->b_birth = BP_PHYSICAL_BIRTH(bp); 4959 exists = buf_hash_insert(hdr, &hash_lock); 4960 } 4961 if (exists != NULL) { 4962 /* somebody beat us to the hash insert */ 4963 mutex_exit(hash_lock); 4964 buf_discard_identity(hdr); 4965 arc_hdr_destroy(hdr); 4966 goto top; /* restart the IO request */ 4967 } 4968 } else { 4969 /* 4970 * This block is in the ghost cache. If it was L2-only 4971 * (and thus didn't have an L1 hdr), we realloc the 4972 * header to add an L1 hdr. 4973 */ 4974 if (!HDR_HAS_L1HDR(hdr)) { 4975 hdr = arc_hdr_realloc(hdr, hdr_l2only_cache, 4976 hdr_full_cache); 4977 } 4978 ASSERT3P(hdr->b_l1hdr.b_pdata, ==, NULL); 4979 ASSERT(GHOST_STATE(hdr->b_l1hdr.b_state)); 4980 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 4981 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 4982 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 4983 4984 /* 4985 * This is a delicate dance that we play here. 4986 * This hdr is in the ghost list so we access it 4987 * to move it out of the ghost list before we 4988 * initiate the read. If it's a prefetch then 4989 * it won't have a callback so we'll remove the 4990 * reference that arc_buf_alloc_impl() created. We 4991 * do this after we've called arc_access() to 4992 * avoid hitting an assert in remove_reference(). 4993 */ 4994 arc_access(hdr, hash_lock); 4995 arc_hdr_alloc_pdata(hdr); 4996 } 4997 ASSERT3P(hdr->b_l1hdr.b_pdata, !=, NULL); 4998 size = arc_hdr_size(hdr); 4999 5000 /* 5001 * If compression is enabled on the hdr, then will do 5002 * RAW I/O and will store the compressed data in the hdr's 5003 * data block. Otherwise, the hdr's data block will contain 5004 * the uncompressed data. 5005 */ 5006 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF) { 5007 zio_flags |= ZIO_FLAG_RAW; 5008 } 5009 5010 if (*arc_flags & ARC_FLAG_PREFETCH) 5011 arc_hdr_set_flags(hdr, ARC_FLAG_PREFETCH); 5012 if (*arc_flags & ARC_FLAG_L2CACHE) 5013 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE); 5014 if (BP_GET_LEVEL(bp) > 0) 5015 arc_hdr_set_flags(hdr, ARC_FLAG_INDIRECT); 5016 if (*arc_flags & ARC_FLAG_PREDICTIVE_PREFETCH) 5017 arc_hdr_set_flags(hdr, ARC_FLAG_PREDICTIVE_PREFETCH); 5018 ASSERT(!GHOST_STATE(hdr->b_l1hdr.b_state)); 5019 5020 acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP); 5021 acb->acb_done = done; 5022 acb->acb_private = private; 5023 5024 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL); 5025 hdr->b_l1hdr.b_acb = acb; 5026 arc_hdr_set_flags(hdr, ARC_FLAG_IO_IN_PROGRESS); 5027 5028 if (HDR_HAS_L2HDR(hdr) && 5029 (vd = hdr->b_l2hdr.b_dev->l2ad_vdev) != NULL) { 5030 devw = hdr->b_l2hdr.b_dev->l2ad_writing; 5031 addr = hdr->b_l2hdr.b_daddr; 5032 /* 5033 * Lock out device removal. 5034 */ 5035 if (vdev_is_dead(vd) || 5036 !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER)) 5037 vd = NULL; 5038 } 5039 5040 if (priority == ZIO_PRIORITY_ASYNC_READ) 5041 arc_hdr_set_flags(hdr, ARC_FLAG_PRIO_ASYNC_READ); 5042 else 5043 arc_hdr_clear_flags(hdr, ARC_FLAG_PRIO_ASYNC_READ); 5044 5045 if (hash_lock != NULL) 5046 mutex_exit(hash_lock); 5047 5048 /* 5049 * At this point, we have a level 1 cache miss. Try again in 5050 * L2ARC if possible. 5051 */ 5052 ASSERT3U(HDR_GET_LSIZE(hdr), ==, lsize); 5053 5054 DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp, 5055 uint64_t, lsize, zbookmark_phys_t *, zb); 5056 ARCSTAT_BUMP(arcstat_misses); 5057 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr), 5058 demand, prefetch, !HDR_ISTYPE_METADATA(hdr), 5059 data, metadata, misses); 5060#ifdef _KERNEL 5061#ifdef RACCT 5062 if (racct_enable) { 5063 PROC_LOCK(curproc); 5064 racct_add_force(curproc, RACCT_READBPS, size); 5065 racct_add_force(curproc, RACCT_READIOPS, 1); 5066 PROC_UNLOCK(curproc); 5067 } 5068#endif /* RACCT */ 5069 curthread->td_ru.ru_inblock++; 5070#endif 5071 5072 if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) { 5073 /* 5074 * Read from the L2ARC if the following are true: 5075 * 1. The L2ARC vdev was previously cached. 5076 * 2. This buffer still has L2ARC metadata. 5077 * 3. This buffer isn't currently writing to the L2ARC. 5078 * 4. The L2ARC entry wasn't evicted, which may 5079 * also have invalidated the vdev. 5080 * 5. This isn't prefetch and l2arc_noprefetch is set. 5081 */ 5082 if (HDR_HAS_L2HDR(hdr) && 5083 !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) && 5084 !(l2arc_noprefetch && HDR_PREFETCH(hdr))) { 5085 l2arc_read_callback_t *cb; 5086 void* b_data; 5087 5088 DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr); 5089 ARCSTAT_BUMP(arcstat_l2_hits); 5090 5091 cb = kmem_zalloc(sizeof (l2arc_read_callback_t), 5092 KM_SLEEP); 5093 cb->l2rcb_hdr = hdr; 5094 cb->l2rcb_bp = *bp; 5095 cb->l2rcb_zb = *zb; 5096 cb->l2rcb_flags = zio_flags; 5097 uint64_t asize = vdev_psize_to_asize(vd, size); 5098 if (asize != size) { 5099 b_data = zio_data_buf_alloc(asize); 5100 cb->l2rcb_data = b_data; 5101 } else { 5102 b_data = hdr->b_l1hdr.b_pdata; 5103 } 5104 5105 ASSERT(addr >= VDEV_LABEL_START_SIZE && 5106 addr + asize < vd->vdev_psize - 5107 VDEV_LABEL_END_SIZE); 5108 5109 /* 5110 * l2arc read. The SCL_L2ARC lock will be 5111 * released by l2arc_read_done(). 5112 * Issue a null zio if the underlying buffer 5113 * was squashed to zero size by compression. 5114 */ 5115 ASSERT3U(HDR_GET_COMPRESS(hdr), !=, 5116 ZIO_COMPRESS_EMPTY); 5117 rzio = zio_read_phys(pio, vd, addr, 5118 asize, b_data, 5119 ZIO_CHECKSUM_OFF, 5120 l2arc_read_done, cb, priority, 5121 zio_flags | ZIO_FLAG_DONT_CACHE | 5122 ZIO_FLAG_CANFAIL | 5123 ZIO_FLAG_DONT_PROPAGATE | 5124 ZIO_FLAG_DONT_RETRY, B_FALSE); 5125 DTRACE_PROBE2(l2arc__read, vdev_t *, vd, 5126 zio_t *, rzio); 5127 ARCSTAT_INCR(arcstat_l2_read_bytes, size); 5128 5129 if (*arc_flags & ARC_FLAG_NOWAIT) { 5130 zio_nowait(rzio); 5131 return (0); 5132 } 5133 5134 ASSERT(*arc_flags & ARC_FLAG_WAIT); 5135 if (zio_wait(rzio) == 0) 5136 return (0); 5137 5138 /* l2arc read error; goto zio_read() */ 5139 } else { 5140 DTRACE_PROBE1(l2arc__miss, 5141 arc_buf_hdr_t *, hdr); 5142 ARCSTAT_BUMP(arcstat_l2_misses); 5143 if (HDR_L2_WRITING(hdr)) 5144 ARCSTAT_BUMP(arcstat_l2_rw_clash); 5145 spa_config_exit(spa, SCL_L2ARC, vd); 5146 } 5147 } else { 5148 if (vd != NULL) 5149 spa_config_exit(spa, SCL_L2ARC, vd); 5150 if (l2arc_ndev != 0) { 5151 DTRACE_PROBE1(l2arc__miss, 5152 arc_buf_hdr_t *, hdr); 5153 ARCSTAT_BUMP(arcstat_l2_misses); 5154 } 5155 } 5156 5157 rzio = zio_read(pio, spa, bp, hdr->b_l1hdr.b_pdata, size, 5158 arc_read_done, hdr, priority, zio_flags, zb); 5159 5160 if (*arc_flags & ARC_FLAG_WAIT) 5161 return (zio_wait(rzio)); 5162 5163 ASSERT(*arc_flags & ARC_FLAG_NOWAIT); 5164 zio_nowait(rzio); 5165 } 5166 return (0); 5167} 5168 5169/* 5170 * Notify the arc that a block was freed, and thus will never be used again. 5171 */ 5172void 5173arc_freed(spa_t *spa, const blkptr_t *bp) 5174{ 5175 arc_buf_hdr_t *hdr; 5176 kmutex_t *hash_lock; 5177 uint64_t guid = spa_load_guid(spa); 5178 5179 ASSERT(!BP_IS_EMBEDDED(bp)); 5180 5181 hdr = buf_hash_find(guid, bp, &hash_lock); 5182 if (hdr == NULL) 5183 return; 5184 5185 /* 5186 * We might be trying to free a block that is still doing I/O 5187 * (i.e. prefetch) or has a reference (i.e. a dedup-ed, 5188 * dmu_sync-ed block). If this block is being prefetched, then it 5189 * would still have the ARC_FLAG_IO_IN_PROGRESS flag set on the hdr 5190 * until the I/O completes. A block may also have a reference if it is 5191 * part of a dedup-ed, dmu_synced write. The dmu_sync() function would 5192 * have written the new block to its final resting place on disk but 5193 * without the dedup flag set. This would have left the hdr in the MRU 5194 * state and discoverable. When the txg finally syncs it detects that 5195 * the block was overridden in open context and issues an override I/O. 5196 * Since this is a dedup block, the override I/O will determine if the 5197 * block is already in the DDT. If so, then it will replace the io_bp 5198 * with the bp from the DDT and allow the I/O to finish. When the I/O 5199 * reaches the done callback, dbuf_write_override_done, it will 5200 * check to see if the io_bp and io_bp_override are identical. 5201 * If they are not, then it indicates that the bp was replaced with 5202 * the bp in the DDT and the override bp is freed. This allows 5203 * us to arrive here with a reference on a block that is being 5204 * freed. So if we have an I/O in progress, or a reference to 5205 * this hdr, then we don't destroy the hdr. 5206 */ 5207 if (!HDR_HAS_L1HDR(hdr) || (!HDR_IO_IN_PROGRESS(hdr) && 5208 refcount_is_zero(&hdr->b_l1hdr.b_refcnt))) { 5209 arc_change_state(arc_anon, hdr, hash_lock); 5210 arc_hdr_destroy(hdr); 5211 mutex_exit(hash_lock); 5212 } else { 5213 mutex_exit(hash_lock); 5214 } 5215 5216} 5217 5218/* 5219 * Release this buffer from the cache, making it an anonymous buffer. This 5220 * must be done after a read and prior to modifying the buffer contents. 5221 * If the buffer has more than one reference, we must make 5222 * a new hdr for the buffer. 5223 */ 5224void 5225arc_release(arc_buf_t *buf, void *tag) 5226{ 5227 arc_buf_hdr_t *hdr = buf->b_hdr; 5228 5229 /* 5230 * It would be nice to assert that if it's DMU metadata (level > 5231 * 0 || it's the dnode file), then it must be syncing context. 5232 * But we don't know that information at this level. 5233 */ 5234 5235 mutex_enter(&buf->b_evict_lock); 5236 5237 ASSERT(HDR_HAS_L1HDR(hdr)); 5238 5239 /* 5240 * We don't grab the hash lock prior to this check, because if 5241 * the buffer's header is in the arc_anon state, it won't be 5242 * linked into the hash table. 5243 */ 5244 if (hdr->b_l1hdr.b_state == arc_anon) { 5245 mutex_exit(&buf->b_evict_lock); 5246 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 5247 ASSERT(!HDR_IN_HASH_TABLE(hdr)); 5248 ASSERT(!HDR_HAS_L2HDR(hdr)); 5249 ASSERT(HDR_EMPTY(hdr)); 5250 ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1); 5251 ASSERT3S(refcount_count(&hdr->b_l1hdr.b_refcnt), ==, 1); 5252 ASSERT(!list_link_active(&hdr->b_l1hdr.b_arc_node)); 5253 5254 hdr->b_l1hdr.b_arc_access = 0; 5255 5256 /* 5257 * If the buf is being overridden then it may already 5258 * have a hdr that is not empty. 5259 */ 5260 buf_discard_identity(hdr); 5261 arc_buf_thaw(buf); 5262 5263 return; 5264 } 5265 5266 kmutex_t *hash_lock = HDR_LOCK(hdr); 5267 mutex_enter(hash_lock); 5268 5269 /* 5270 * This assignment is only valid as long as the hash_lock is 5271 * held, we must be careful not to reference state or the 5272 * b_state field after dropping the lock. 5273 */ 5274 arc_state_t *state = hdr->b_l1hdr.b_state; 5275 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 5276 ASSERT3P(state, !=, arc_anon); 5277 5278 /* this buffer is not on any list */ 5279 ASSERT(refcount_count(&hdr->b_l1hdr.b_refcnt) > 0); 5280 5281 if (HDR_HAS_L2HDR(hdr)) { 5282 mutex_enter(&hdr->b_l2hdr.b_dev->l2ad_mtx); 5283 5284 /* 5285 * We have to recheck this conditional again now that 5286 * we're holding the l2ad_mtx to prevent a race with 5287 * another thread which might be concurrently calling 5288 * l2arc_evict(). In that case, l2arc_evict() might have 5289 * destroyed the header's L2 portion as we were waiting 5290 * to acquire the l2ad_mtx. 5291 */ 5292 if (HDR_HAS_L2HDR(hdr)) { 5293 l2arc_trim(hdr); 5294 arc_hdr_l2hdr_destroy(hdr); 5295 } 5296 5297 mutex_exit(&hdr->b_l2hdr.b_dev->l2ad_mtx); 5298 } 5299 5300 /* 5301 * Do we have more than one buf? 5302 */ 5303 if (hdr->b_l1hdr.b_bufcnt > 1) { 5304 arc_buf_hdr_t *nhdr; 5305 arc_buf_t **bufp; 5306 uint64_t spa = hdr->b_spa; 5307 uint64_t psize = HDR_GET_PSIZE(hdr); 5308 uint64_t lsize = HDR_GET_LSIZE(hdr); 5309 enum zio_compress compress = HDR_GET_COMPRESS(hdr); 5310 arc_buf_contents_t type = arc_buf_type(hdr); 5311 VERIFY3U(hdr->b_type, ==, type); 5312 5313 ASSERT(hdr->b_l1hdr.b_buf != buf || buf->b_next != NULL); 5314 (void) remove_reference(hdr, hash_lock, tag); 5315 5316 if (arc_buf_is_shared(buf)) { 5317 ASSERT(HDR_SHARED_DATA(hdr)); 5318 ASSERT3P(hdr->b_l1hdr.b_buf, !=, buf); 5319 ASSERT(ARC_BUF_LAST(buf)); 5320 } 5321 5322 /* 5323 * Pull the data off of this hdr and attach it to 5324 * a new anonymous hdr. Also find the last buffer 5325 * in the hdr's buffer list. 5326 */ 5327 arc_buf_t *lastbuf = NULL; 5328 bufp = &hdr->b_l1hdr.b_buf; 5329 while (*bufp != NULL) { 5330 if (*bufp == buf) { 5331 *bufp = buf->b_next; 5332 } 5333 5334 /* 5335 * If we've removed a buffer in the middle of 5336 * the list then update the lastbuf and update 5337 * bufp. 5338 */ 5339 if (*bufp != NULL) { 5340 lastbuf = *bufp; 5341 bufp = &(*bufp)->b_next; 5342 } 5343 } 5344 buf->b_next = NULL; 5345 ASSERT3P(lastbuf, !=, buf); 5346 ASSERT3P(lastbuf, !=, NULL); 5347 5348 /* 5349 * If the current arc_buf_t and the hdr are sharing their data 5350 * buffer, then we must stop sharing that block, transfer 5351 * ownership and setup sharing with a new arc_buf_t at the end 5352 * of the hdr's b_buf list. 5353 */ 5354 if (arc_buf_is_shared(buf)) { 5355 ASSERT3P(hdr->b_l1hdr.b_buf, !=, buf); 5356 ASSERT(ARC_BUF_LAST(lastbuf)); 5357 VERIFY(!arc_buf_is_shared(lastbuf)); 5358 5359 /* 5360 * First, sever the block sharing relationship between 5361 * buf and the arc_buf_hdr_t. Then, setup a new 5362 * block sharing relationship with the last buffer 5363 * on the arc_buf_t list. 5364 */ 5365 arc_unshare_buf(hdr, buf); 5366 arc_share_buf(hdr, lastbuf); 5367 VERIFY3P(lastbuf->b_data, !=, NULL); 5368 } else if (HDR_SHARED_DATA(hdr)) { 5369 ASSERT(arc_buf_is_shared(lastbuf)); 5370 } 5371 ASSERT3P(hdr->b_l1hdr.b_pdata, !=, NULL); 5372 ASSERT3P(state, !=, arc_l2c_only); 5373 5374 (void) refcount_remove_many(&state->arcs_size, 5375 HDR_GET_LSIZE(hdr), buf); 5376 5377 if (refcount_is_zero(&hdr->b_l1hdr.b_refcnt)) { 5378 ASSERT3P(state, !=, arc_l2c_only); 5379 (void) refcount_remove_many(&state->arcs_esize[type], 5380 HDR_GET_LSIZE(hdr), buf); 5381 } 5382 5383 hdr->b_l1hdr.b_bufcnt -= 1; 5384 arc_cksum_verify(buf); 5385#ifdef illumos 5386 arc_buf_unwatch(buf); 5387#endif 5388 5389 mutex_exit(hash_lock); 5390 5391 /* 5392 * Allocate a new hdr. The new hdr will contain a b_pdata 5393 * buffer which will be freed in arc_write(). 5394 */ 5395 nhdr = arc_hdr_alloc(spa, psize, lsize, compress, type); 5396 ASSERT3P(nhdr->b_l1hdr.b_buf, ==, NULL); 5397 ASSERT0(nhdr->b_l1hdr.b_bufcnt); 5398 ASSERT0(refcount_count(&nhdr->b_l1hdr.b_refcnt)); 5399 VERIFY3U(nhdr->b_type, ==, type); 5400 ASSERT(!HDR_SHARED_DATA(nhdr)); 5401 5402 nhdr->b_l1hdr.b_buf = buf; 5403 nhdr->b_l1hdr.b_bufcnt = 1; 5404 (void) refcount_add(&nhdr->b_l1hdr.b_refcnt, tag); 5405 buf->b_hdr = nhdr; 5406 5407 mutex_exit(&buf->b_evict_lock); 5408 (void) refcount_add_many(&arc_anon->arcs_size, 5409 HDR_GET_LSIZE(nhdr), buf); 5410 } else { 5411 mutex_exit(&buf->b_evict_lock); 5412 ASSERT(refcount_count(&hdr->b_l1hdr.b_refcnt) == 1); 5413 /* protected by hash lock, or hdr is on arc_anon */ 5414 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 5415 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 5416 arc_change_state(arc_anon, hdr, hash_lock); 5417 hdr->b_l1hdr.b_arc_access = 0; 5418 mutex_exit(hash_lock); 5419 5420 buf_discard_identity(hdr); 5421 arc_buf_thaw(buf); 5422 } 5423} 5424 5425int 5426arc_released(arc_buf_t *buf) 5427{ 5428 int released; 5429 5430 mutex_enter(&buf->b_evict_lock); 5431 released = (buf->b_data != NULL && 5432 buf->b_hdr->b_l1hdr.b_state == arc_anon); 5433 mutex_exit(&buf->b_evict_lock); 5434 return (released); 5435} 5436 5437#ifdef ZFS_DEBUG 5438int 5439arc_referenced(arc_buf_t *buf) 5440{ 5441 int referenced; 5442 5443 mutex_enter(&buf->b_evict_lock); 5444 referenced = (refcount_count(&buf->b_hdr->b_l1hdr.b_refcnt)); 5445 mutex_exit(&buf->b_evict_lock); 5446 return (referenced); 5447} 5448#endif 5449 5450static void 5451arc_write_ready(zio_t *zio) 5452{ 5453 arc_write_callback_t *callback = zio->io_private; 5454 arc_buf_t *buf = callback->awcb_buf; 5455 arc_buf_hdr_t *hdr = buf->b_hdr; 5456 uint64_t psize = BP_IS_HOLE(zio->io_bp) ? 0 : BP_GET_PSIZE(zio->io_bp); 5457 5458 ASSERT(HDR_HAS_L1HDR(hdr)); 5459 ASSERT(!refcount_is_zero(&buf->b_hdr->b_l1hdr.b_refcnt)); 5460 ASSERT(hdr->b_l1hdr.b_bufcnt > 0); 5461 5462 /* 5463 * If we're reexecuting this zio because the pool suspended, then 5464 * cleanup any state that was previously set the first time the 5465 * callback as invoked. 5466 */ 5467 if (zio->io_flags & ZIO_FLAG_REEXECUTED) { 5468 arc_cksum_free(hdr); 5469#ifdef illumos 5470 arc_buf_unwatch(buf); 5471#endif 5472 if (hdr->b_l1hdr.b_pdata != NULL) { 5473 if (arc_buf_is_shared(buf)) { 5474 ASSERT(HDR_SHARED_DATA(hdr)); 5475 5476 arc_unshare_buf(hdr, buf); 5477 } else { 5478 arc_hdr_free_pdata(hdr); 5479 } 5480 } 5481 } 5482 ASSERT3P(hdr->b_l1hdr.b_pdata, ==, NULL); 5483 ASSERT(!HDR_SHARED_DATA(hdr)); 5484 ASSERT(!arc_buf_is_shared(buf)); 5485 5486 callback->awcb_ready(zio, buf, callback->awcb_private); 5487 5488 if (HDR_IO_IN_PROGRESS(hdr)) 5489 ASSERT(zio->io_flags & ZIO_FLAG_REEXECUTED); 5490 5491 arc_cksum_compute(buf); 5492 arc_hdr_set_flags(hdr, ARC_FLAG_IO_IN_PROGRESS); 5493 5494 enum zio_compress compress; 5495 if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) { 5496 compress = ZIO_COMPRESS_OFF; 5497 } else { 5498 ASSERT3U(HDR_GET_LSIZE(hdr), ==, BP_GET_LSIZE(zio->io_bp)); 5499 compress = BP_GET_COMPRESS(zio->io_bp); 5500 } 5501 HDR_SET_PSIZE(hdr, psize); 5502 arc_hdr_set_compress(hdr, compress); 5503 5504 /* 5505 * If the hdr is compressed, then copy the compressed 5506 * zio contents into arc_buf_hdr_t. Otherwise, copy the original 5507 * data buf into the hdr. Ideally, we would like to always copy the 5508 * io_data into b_pdata but the user may have disabled compressed 5509 * arc thus the on-disk block may or may not match what we maintain 5510 * in the hdr's b_pdata field. 5511 */ 5512 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF) { 5513 ASSERT(BP_GET_COMPRESS(zio->io_bp) != ZIO_COMPRESS_OFF); 5514 ASSERT3U(psize, >, 0); 5515 arc_hdr_alloc_pdata(hdr); 5516 bcopy(zio->io_data, hdr->b_l1hdr.b_pdata, psize); 5517 } else { 5518 ASSERT3P(buf->b_data, ==, zio->io_orig_data); 5519 ASSERT3U(zio->io_orig_size, ==, HDR_GET_LSIZE(hdr)); 5520 ASSERT3U(hdr->b_l1hdr.b_byteswap, ==, DMU_BSWAP_NUMFUNCS); 5521 ASSERT(!HDR_SHARED_DATA(hdr)); 5522 ASSERT(!arc_buf_is_shared(buf)); 5523 ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1); 5524 ASSERT3P(hdr->b_l1hdr.b_pdata, ==, NULL); 5525 5526 /* 5527 * This hdr is not compressed so we're able to share 5528 * the arc_buf_t data buffer with the hdr. 5529 */ 5530 arc_share_buf(hdr, buf); 5531 VERIFY0(bcmp(zio->io_orig_data, hdr->b_l1hdr.b_pdata, 5532 HDR_GET_LSIZE(hdr))); 5533 } 5534 arc_hdr_verify(hdr, zio->io_bp); 5535} 5536 5537static void 5538arc_write_children_ready(zio_t *zio) 5539{ 5540 arc_write_callback_t *callback = zio->io_private; 5541 arc_buf_t *buf = callback->awcb_buf; 5542 5543 callback->awcb_children_ready(zio, buf, callback->awcb_private); 5544} 5545 5546/* 5547 * The SPA calls this callback for each physical write that happens on behalf 5548 * of a logical write. See the comment in dbuf_write_physdone() for details. 5549 */ 5550static void 5551arc_write_physdone(zio_t *zio) 5552{ 5553 arc_write_callback_t *cb = zio->io_private; 5554 if (cb->awcb_physdone != NULL) 5555 cb->awcb_physdone(zio, cb->awcb_buf, cb->awcb_private); 5556} 5557 5558static void 5559arc_write_done(zio_t *zio) 5560{ 5561 arc_write_callback_t *callback = zio->io_private; 5562 arc_buf_t *buf = callback->awcb_buf; 5563 arc_buf_hdr_t *hdr = buf->b_hdr; 5564 5565 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL); 5566 5567 if (zio->io_error == 0) { 5568 arc_hdr_verify(hdr, zio->io_bp); 5569 5570 if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) { 5571 buf_discard_identity(hdr); 5572 } else { 5573 hdr->b_dva = *BP_IDENTITY(zio->io_bp); 5574 hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp); 5575 } 5576 } else { 5577 ASSERT(HDR_EMPTY(hdr)); 5578 } 5579 5580 /* 5581 * If the block to be written was all-zero or compressed enough to be 5582 * embedded in the BP, no write was performed so there will be no 5583 * dva/birth/checksum. The buffer must therefore remain anonymous 5584 * (and uncached). 5585 */ 5586 if (!HDR_EMPTY(hdr)) { 5587 arc_buf_hdr_t *exists; 5588 kmutex_t *hash_lock; 5589 5590 ASSERT(zio->io_error == 0); 5591 5592 arc_cksum_verify(buf); 5593 5594 exists = buf_hash_insert(hdr, &hash_lock); 5595 if (exists != NULL) { 5596 /* 5597 * This can only happen if we overwrite for 5598 * sync-to-convergence, because we remove 5599 * buffers from the hash table when we arc_free(). 5600 */ 5601 if (zio->io_flags & ZIO_FLAG_IO_REWRITE) { 5602 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) 5603 panic("bad overwrite, hdr=%p exists=%p", 5604 (void *)hdr, (void *)exists); 5605 ASSERT(refcount_is_zero( 5606 &exists->b_l1hdr.b_refcnt)); 5607 arc_change_state(arc_anon, exists, hash_lock); 5608 mutex_exit(hash_lock); 5609 arc_hdr_destroy(exists); 5610 exists = buf_hash_insert(hdr, &hash_lock); 5611 ASSERT3P(exists, ==, NULL); 5612 } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) { 5613 /* nopwrite */ 5614 ASSERT(zio->io_prop.zp_nopwrite); 5615 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) 5616 panic("bad nopwrite, hdr=%p exists=%p", 5617 (void *)hdr, (void *)exists); 5618 } else { 5619 /* Dedup */ 5620 ASSERT(hdr->b_l1hdr.b_bufcnt == 1); 5621 ASSERT(hdr->b_l1hdr.b_state == arc_anon); 5622 ASSERT(BP_GET_DEDUP(zio->io_bp)); 5623 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0); 5624 } 5625 } 5626 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS); 5627 /* if it's not anon, we are doing a scrub */ 5628 if (exists == NULL && hdr->b_l1hdr.b_state == arc_anon) 5629 arc_access(hdr, hash_lock); 5630 mutex_exit(hash_lock); 5631 } else { 5632 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS); 5633 } 5634 5635 ASSERT(!refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 5636 callback->awcb_done(zio, buf, callback->awcb_private); 5637 5638 kmem_free(callback, sizeof (arc_write_callback_t)); 5639} 5640 5641zio_t * 5642arc_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, arc_buf_t *buf, 5643 boolean_t l2arc, const zio_prop_t *zp, arc_done_func_t *ready, 5644 arc_done_func_t *children_ready, arc_done_func_t *physdone, 5645 arc_done_func_t *done, void *private, zio_priority_t priority, 5646 int zio_flags, const zbookmark_phys_t *zb) 5647{ 5648 arc_buf_hdr_t *hdr = buf->b_hdr; 5649 arc_write_callback_t *callback; 5650 zio_t *zio; 5651 5652 ASSERT3P(ready, !=, NULL); 5653 ASSERT3P(done, !=, NULL); 5654 ASSERT(!HDR_IO_ERROR(hdr)); 5655 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 5656 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL); 5657 ASSERT3U(hdr->b_l1hdr.b_bufcnt, >, 0); 5658 if (l2arc) 5659 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE); 5660 callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP); 5661 callback->awcb_ready = ready; 5662 callback->awcb_children_ready = children_ready; 5663 callback->awcb_physdone = physdone; 5664 callback->awcb_done = done; 5665 callback->awcb_private = private; 5666 callback->awcb_buf = buf; 5667 5668 /* 5669 * The hdr's b_pdata is now stale, free it now. A new data block 5670 * will be allocated when the zio pipeline calls arc_write_ready(). 5671 */ 5672 if (hdr->b_l1hdr.b_pdata != NULL) { 5673 /* 5674 * If the buf is currently sharing the data block with 5675 * the hdr then we need to break that relationship here. 5676 * The hdr will remain with a NULL data pointer and the 5677 * buf will take sole ownership of the block. 5678 */ 5679 if (arc_buf_is_shared(buf)) { 5680 ASSERT(ARC_BUF_LAST(buf)); 5681 arc_unshare_buf(hdr, buf); 5682 } else { 5683 arc_hdr_free_pdata(hdr); 5684 } 5685 VERIFY3P(buf->b_data, !=, NULL); 5686 arc_hdr_set_compress(hdr, ZIO_COMPRESS_OFF); 5687 } 5688 ASSERT(!arc_buf_is_shared(buf)); 5689 ASSERT3P(hdr->b_l1hdr.b_pdata, ==, NULL); 5690 5691 zio = zio_write(pio, spa, txg, bp, buf->b_data, HDR_GET_LSIZE(hdr), zp, 5692 arc_write_ready, 5693 (children_ready != NULL) ? arc_write_children_ready : NULL, 5694 arc_write_physdone, arc_write_done, callback, 5695 priority, zio_flags, zb); 5696 5697 return (zio); 5698} 5699 5700static int 5701arc_memory_throttle(uint64_t reserve, uint64_t txg) 5702{ 5703#ifdef _KERNEL 5704 uint64_t available_memory = ptob(freemem); 5705 static uint64_t page_load = 0; 5706 static uint64_t last_txg = 0; 5707 5708#if defined(__i386) || !defined(UMA_MD_SMALL_ALLOC) 5709 available_memory = 5710 MIN(available_memory, ptob(vmem_size(heap_arena, VMEM_FREE))); 5711#endif 5712 5713 if (freemem > (uint64_t)physmem * arc_lotsfree_percent / 100) 5714 return (0); 5715 5716 if (txg > last_txg) { 5717 last_txg = txg; 5718 page_load = 0; 5719 } 5720 /* 5721 * If we are in pageout, we know that memory is already tight, 5722 * the arc is already going to be evicting, so we just want to 5723 * continue to let page writes occur as quickly as possible. 5724 */ 5725 if (curproc == pageproc) { 5726 if (page_load > MAX(ptob(minfree), available_memory) / 4) 5727 return (SET_ERROR(ERESTART)); 5728 /* Note: reserve is inflated, so we deflate */ 5729 page_load += reserve / 8; 5730 return (0); 5731 } else if (page_load > 0 && arc_reclaim_needed()) { 5732 /* memory is low, delay before restarting */ 5733 ARCSTAT_INCR(arcstat_memory_throttle_count, 1); 5734 return (SET_ERROR(EAGAIN)); 5735 } 5736 page_load = 0; 5737#endif 5738 return (0); 5739} 5740 5741void 5742arc_tempreserve_clear(uint64_t reserve) 5743{ 5744 atomic_add_64(&arc_tempreserve, -reserve); 5745 ASSERT((int64_t)arc_tempreserve >= 0); 5746} 5747 5748int 5749arc_tempreserve_space(uint64_t reserve, uint64_t txg) 5750{ 5751 int error; 5752 uint64_t anon_size; 5753 5754 if (reserve > arc_c/4 && !arc_no_grow) { 5755 arc_c = MIN(arc_c_max, reserve * 4); 5756 DTRACE_PROBE1(arc__set_reserve, uint64_t, arc_c); 5757 } 5758 if (reserve > arc_c) 5759 return (SET_ERROR(ENOMEM)); 5760 5761 /* 5762 * Don't count loaned bufs as in flight dirty data to prevent long 5763 * network delays from blocking transactions that are ready to be 5764 * assigned to a txg. 5765 */ 5766 anon_size = MAX((int64_t)(refcount_count(&arc_anon->arcs_size) - 5767 arc_loaned_bytes), 0); 5768 5769 /* 5770 * Writes will, almost always, require additional memory allocations 5771 * in order to compress/encrypt/etc the data. We therefore need to 5772 * make sure that there is sufficient available memory for this. 5773 */ 5774 error = arc_memory_throttle(reserve, txg); 5775 if (error != 0) 5776 return (error); 5777 5778 /* 5779 * Throttle writes when the amount of dirty data in the cache 5780 * gets too large. We try to keep the cache less than half full 5781 * of dirty blocks so that our sync times don't grow too large. 5782 * Note: if two requests come in concurrently, we might let them 5783 * both succeed, when one of them should fail. Not a huge deal. 5784 */ 5785 5786 if (reserve + arc_tempreserve + anon_size > arc_c / 2 && 5787 anon_size > arc_c / 4) { 5788 uint64_t meta_esize = 5789 refcount_count(&arc_anon->arcs_esize[ARC_BUFC_METADATA]); 5790 uint64_t data_esize = 5791 refcount_count(&arc_anon->arcs_esize[ARC_BUFC_DATA]); 5792 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK " 5793 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n", 5794 arc_tempreserve >> 10, meta_esize >> 10, 5795 data_esize >> 10, reserve >> 10, arc_c >> 10); 5796 return (SET_ERROR(ERESTART)); 5797 } 5798 atomic_add_64(&arc_tempreserve, reserve); 5799 return (0); 5800} 5801 5802static void 5803arc_kstat_update_state(arc_state_t *state, kstat_named_t *size, 5804 kstat_named_t *evict_data, kstat_named_t *evict_metadata) 5805{ 5806 size->value.ui64 = refcount_count(&state->arcs_size); 5807 evict_data->value.ui64 = 5808 refcount_count(&state->arcs_esize[ARC_BUFC_DATA]); 5809 evict_metadata->value.ui64 = 5810 refcount_count(&state->arcs_esize[ARC_BUFC_METADATA]); 5811} 5812 5813static int 5814arc_kstat_update(kstat_t *ksp, int rw) 5815{ 5816 arc_stats_t *as = ksp->ks_data; 5817 5818 if (rw == KSTAT_WRITE) { 5819 return (EACCES); 5820 } else { 5821 arc_kstat_update_state(arc_anon, 5822 &as->arcstat_anon_size, 5823 &as->arcstat_anon_evictable_data, 5824 &as->arcstat_anon_evictable_metadata); 5825 arc_kstat_update_state(arc_mru, 5826 &as->arcstat_mru_size, 5827 &as->arcstat_mru_evictable_data, 5828 &as->arcstat_mru_evictable_metadata); 5829 arc_kstat_update_state(arc_mru_ghost, 5830 &as->arcstat_mru_ghost_size, 5831 &as->arcstat_mru_ghost_evictable_data, 5832 &as->arcstat_mru_ghost_evictable_metadata); 5833 arc_kstat_update_state(arc_mfu, 5834 &as->arcstat_mfu_size, 5835 &as->arcstat_mfu_evictable_data, 5836 &as->arcstat_mfu_evictable_metadata); 5837 arc_kstat_update_state(arc_mfu_ghost, 5838 &as->arcstat_mfu_ghost_size, 5839 &as->arcstat_mfu_ghost_evictable_data, 5840 &as->arcstat_mfu_ghost_evictable_metadata); 5841 } 5842 5843 return (0); 5844} 5845 5846/* 5847 * This function *must* return indices evenly distributed between all 5848 * sublists of the multilist. This is needed due to how the ARC eviction 5849 * code is laid out; arc_evict_state() assumes ARC buffers are evenly 5850 * distributed between all sublists and uses this assumption when 5851 * deciding which sublist to evict from and how much to evict from it. 5852 */ 5853unsigned int 5854arc_state_multilist_index_func(multilist_t *ml, void *obj) 5855{ 5856 arc_buf_hdr_t *hdr = obj; 5857 5858 /* 5859 * We rely on b_dva to generate evenly distributed index 5860 * numbers using buf_hash below. So, as an added precaution, 5861 * let's make sure we never add empty buffers to the arc lists. 5862 */ 5863 ASSERT(!HDR_EMPTY(hdr)); 5864 5865 /* 5866 * The assumption here, is the hash value for a given 5867 * arc_buf_hdr_t will remain constant throughout it's lifetime 5868 * (i.e. it's b_spa, b_dva, and b_birth fields don't change). 5869 * Thus, we don't need to store the header's sublist index 5870 * on insertion, as this index can be recalculated on removal. 5871 * 5872 * Also, the low order bits of the hash value are thought to be 5873 * distributed evenly. Otherwise, in the case that the multilist 5874 * has a power of two number of sublists, each sublists' usage 5875 * would not be evenly distributed. 5876 */ 5877 return (buf_hash(hdr->b_spa, &hdr->b_dva, hdr->b_birth) % 5878 multilist_get_num_sublists(ml)); 5879} 5880 5881#ifdef _KERNEL 5882static eventhandler_tag arc_event_lowmem = NULL; 5883 5884static void 5885arc_lowmem(void *arg __unused, int howto __unused) 5886{ 5887 5888 mutex_enter(&arc_reclaim_lock); 5889 /* XXX: Memory deficit should be passed as argument. */ 5890 needfree = btoc(arc_c >> arc_shrink_shift); 5891 DTRACE_PROBE(arc__needfree); 5892 cv_signal(&arc_reclaim_thread_cv); 5893 5894 /* 5895 * It is unsafe to block here in arbitrary threads, because we can come 5896 * here from ARC itself and may hold ARC locks and thus risk a deadlock 5897 * with ARC reclaim thread. 5898 */ 5899 if (curproc == pageproc) 5900 (void) cv_wait(&arc_reclaim_waiters_cv, &arc_reclaim_lock); 5901 mutex_exit(&arc_reclaim_lock); 5902} 5903#endif 5904 5905static void 5906arc_state_init(void) 5907{ 5908 arc_anon = &ARC_anon; 5909 arc_mru = &ARC_mru; 5910 arc_mru_ghost = &ARC_mru_ghost; 5911 arc_mfu = &ARC_mfu; 5912 arc_mfu_ghost = &ARC_mfu_ghost; 5913 arc_l2c_only = &ARC_l2c_only; 5914 5915 multilist_create(&arc_mru->arcs_list[ARC_BUFC_METADATA], 5916 sizeof (arc_buf_hdr_t), 5917 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5918 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5919 multilist_create(&arc_mru->arcs_list[ARC_BUFC_DATA], 5920 sizeof (arc_buf_hdr_t), 5921 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5922 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5923 multilist_create(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA], 5924 sizeof (arc_buf_hdr_t), 5925 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5926 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5927 multilist_create(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA], 5928 sizeof (arc_buf_hdr_t), 5929 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5930 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5931 multilist_create(&arc_mfu->arcs_list[ARC_BUFC_METADATA], 5932 sizeof (arc_buf_hdr_t), 5933 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5934 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5935 multilist_create(&arc_mfu->arcs_list[ARC_BUFC_DATA], 5936 sizeof (arc_buf_hdr_t), 5937 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5938 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5939 multilist_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA], 5940 sizeof (arc_buf_hdr_t), 5941 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5942 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5943 multilist_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA], 5944 sizeof (arc_buf_hdr_t), 5945 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5946 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5947 multilist_create(&arc_l2c_only->arcs_list[ARC_BUFC_METADATA], 5948 sizeof (arc_buf_hdr_t), 5949 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5950 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5951 multilist_create(&arc_l2c_only->arcs_list[ARC_BUFC_DATA], 5952 sizeof (arc_buf_hdr_t), 5953 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5954 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5955 5956 refcount_create(&arc_anon->arcs_esize[ARC_BUFC_METADATA]); 5957 refcount_create(&arc_anon->arcs_esize[ARC_BUFC_DATA]); 5958 refcount_create(&arc_mru->arcs_esize[ARC_BUFC_METADATA]); 5959 refcount_create(&arc_mru->arcs_esize[ARC_BUFC_DATA]); 5960 refcount_create(&arc_mru_ghost->arcs_esize[ARC_BUFC_METADATA]); 5961 refcount_create(&arc_mru_ghost->arcs_esize[ARC_BUFC_DATA]); 5962 refcount_create(&arc_mfu->arcs_esize[ARC_BUFC_METADATA]); 5963 refcount_create(&arc_mfu->arcs_esize[ARC_BUFC_DATA]); 5964 refcount_create(&arc_mfu_ghost->arcs_esize[ARC_BUFC_METADATA]); 5965 refcount_create(&arc_mfu_ghost->arcs_esize[ARC_BUFC_DATA]); 5966 refcount_create(&arc_l2c_only->arcs_esize[ARC_BUFC_METADATA]); 5967 refcount_create(&arc_l2c_only->arcs_esize[ARC_BUFC_DATA]); 5968 5969 refcount_create(&arc_anon->arcs_size); 5970 refcount_create(&arc_mru->arcs_size); 5971 refcount_create(&arc_mru_ghost->arcs_size); 5972 refcount_create(&arc_mfu->arcs_size); 5973 refcount_create(&arc_mfu_ghost->arcs_size); 5974 refcount_create(&arc_l2c_only->arcs_size); 5975} 5976 5977static void 5978arc_state_fini(void) 5979{ 5980 refcount_destroy(&arc_anon->arcs_esize[ARC_BUFC_METADATA]); 5981 refcount_destroy(&arc_anon->arcs_esize[ARC_BUFC_DATA]); 5982 refcount_destroy(&arc_mru->arcs_esize[ARC_BUFC_METADATA]); 5983 refcount_destroy(&arc_mru->arcs_esize[ARC_BUFC_DATA]); 5984 refcount_destroy(&arc_mru_ghost->arcs_esize[ARC_BUFC_METADATA]); 5985 refcount_destroy(&arc_mru_ghost->arcs_esize[ARC_BUFC_DATA]); 5986 refcount_destroy(&arc_mfu->arcs_esize[ARC_BUFC_METADATA]); 5987 refcount_destroy(&arc_mfu->arcs_esize[ARC_BUFC_DATA]); 5988 refcount_destroy(&arc_mfu_ghost->arcs_esize[ARC_BUFC_METADATA]); 5989 refcount_destroy(&arc_mfu_ghost->arcs_esize[ARC_BUFC_DATA]); 5990 refcount_destroy(&arc_l2c_only->arcs_esize[ARC_BUFC_METADATA]); 5991 refcount_destroy(&arc_l2c_only->arcs_esize[ARC_BUFC_DATA]); 5992 5993 refcount_destroy(&arc_anon->arcs_size); 5994 refcount_destroy(&arc_mru->arcs_size); 5995 refcount_destroy(&arc_mru_ghost->arcs_size); 5996 refcount_destroy(&arc_mfu->arcs_size); 5997 refcount_destroy(&arc_mfu_ghost->arcs_size); 5998 refcount_destroy(&arc_l2c_only->arcs_size); 5999 6000 multilist_destroy(&arc_mru->arcs_list[ARC_BUFC_METADATA]); 6001 multilist_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA]); 6002 multilist_destroy(&arc_mfu->arcs_list[ARC_BUFC_METADATA]); 6003 multilist_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA]); 6004 multilist_destroy(&arc_mru->arcs_list[ARC_BUFC_DATA]); 6005 multilist_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA]); 6006 multilist_destroy(&arc_mfu->arcs_list[ARC_BUFC_DATA]); 6007 multilist_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA]); 6008} 6009 6010uint64_t 6011arc_max_bytes(void) 6012{ 6013 return (arc_c_max); 6014} 6015 6016void 6017arc_init(void) 6018{ 6019 int i, prefetch_tunable_set = 0; 6020 6021 mutex_init(&arc_reclaim_lock, NULL, MUTEX_DEFAULT, NULL); 6022 cv_init(&arc_reclaim_thread_cv, NULL, CV_DEFAULT, NULL); 6023 cv_init(&arc_reclaim_waiters_cv, NULL, CV_DEFAULT, NULL); 6024 6025 mutex_init(&arc_dnlc_evicts_lock, NULL, MUTEX_DEFAULT, NULL); 6026 cv_init(&arc_dnlc_evicts_cv, NULL, CV_DEFAULT, NULL); 6027 6028 /* Convert seconds to clock ticks */ 6029 arc_min_prefetch_lifespan = 1 * hz; 6030 6031 /* Start out with 1/8 of all memory */ 6032 arc_c = kmem_size() / 8; 6033 6034#ifdef illumos 6035#ifdef _KERNEL 6036 /* 6037 * On architectures where the physical memory can be larger 6038 * than the addressable space (intel in 32-bit mode), we may 6039 * need to limit the cache to 1/8 of VM size. 6040 */ 6041 arc_c = MIN(arc_c, vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 8); 6042#endif 6043#endif /* illumos */ 6044 /* set min cache to 1/32 of all memory, or arc_abs_min, whichever is more */ 6045 arc_c_min = MAX(arc_c / 4, arc_abs_min); 6046 /* set max to 1/2 of all memory, or all but 1GB, whichever is more */ 6047 if (arc_c * 8 >= 1 << 30) 6048 arc_c_max = (arc_c * 8) - (1 << 30); 6049 else 6050 arc_c_max = arc_c_min; 6051 arc_c_max = MAX(arc_c * 5, arc_c_max); 6052 6053 /* 6054 * In userland, there's only the memory pressure that we artificially 6055 * create (see arc_available_memory()). Don't let arc_c get too 6056 * small, because it can cause transactions to be larger than 6057 * arc_c, causing arc_tempreserve_space() to fail. 6058 */ 6059#ifndef _KERNEL 6060 arc_c_min = arc_c_max / 2; 6061#endif 6062 6063#ifdef _KERNEL 6064 /* 6065 * Allow the tunables to override our calculations if they are 6066 * reasonable. 6067 */ 6068 if (zfs_arc_max > arc_abs_min && zfs_arc_max < kmem_size()) { 6069 arc_c_max = zfs_arc_max; 6070 arc_c_min = MIN(arc_c_min, arc_c_max); 6071 } 6072 if (zfs_arc_min > arc_abs_min && zfs_arc_min <= arc_c_max) 6073 arc_c_min = zfs_arc_min; 6074#endif 6075 6076 arc_c = arc_c_max; 6077 arc_p = (arc_c >> 1); 6078 arc_size = 0; 6079 6080 /* limit meta-data to 1/4 of the arc capacity */ 6081 arc_meta_limit = arc_c_max / 4; 6082 6083 /* Allow the tunable to override if it is reasonable */ 6084 if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max) 6085 arc_meta_limit = zfs_arc_meta_limit; 6086 6087 if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0) 6088 arc_c_min = arc_meta_limit / 2; 6089 6090 if (zfs_arc_meta_min > 0) { 6091 arc_meta_min = zfs_arc_meta_min; 6092 } else { 6093 arc_meta_min = arc_c_min / 2; 6094 } 6095 6096 if (zfs_arc_grow_retry > 0) 6097 arc_grow_retry = zfs_arc_grow_retry; 6098 6099 if (zfs_arc_shrink_shift > 0) 6100 arc_shrink_shift = zfs_arc_shrink_shift; 6101 6102 /* 6103 * Ensure that arc_no_grow_shift is less than arc_shrink_shift. 6104 */ 6105 if (arc_no_grow_shift >= arc_shrink_shift) 6106 arc_no_grow_shift = arc_shrink_shift - 1; 6107 6108 if (zfs_arc_p_min_shift > 0) 6109 arc_p_min_shift = zfs_arc_p_min_shift; 6110 6111 if (zfs_arc_num_sublists_per_state < 1) 6112 zfs_arc_num_sublists_per_state = MAX(max_ncpus, 1); 6113 6114 /* if kmem_flags are set, lets try to use less memory */ 6115 if (kmem_debugging()) 6116 arc_c = arc_c / 2; 6117 if (arc_c < arc_c_min) 6118 arc_c = arc_c_min; 6119 6120 zfs_arc_min = arc_c_min; 6121 zfs_arc_max = arc_c_max; 6122 6123 arc_state_init(); 6124 buf_init(); 6125 6126 arc_reclaim_thread_exit = B_FALSE; 6127 arc_dnlc_evicts_thread_exit = FALSE; 6128 6129 arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED, 6130 sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL); 6131 6132 if (arc_ksp != NULL) { 6133 arc_ksp->ks_data = &arc_stats; 6134 arc_ksp->ks_update = arc_kstat_update; 6135 kstat_install(arc_ksp); 6136 } 6137 6138 (void) thread_create(NULL, 0, arc_reclaim_thread, NULL, 0, &p0, 6139 TS_RUN, minclsyspri); 6140 6141#ifdef _KERNEL 6142 arc_event_lowmem = EVENTHANDLER_REGISTER(vm_lowmem, arc_lowmem, NULL, 6143 EVENTHANDLER_PRI_FIRST); 6144#endif 6145 6146 (void) thread_create(NULL, 0, arc_dnlc_evicts_thread, NULL, 0, &p0, 6147 TS_RUN, minclsyspri); 6148 6149 arc_dead = B_FALSE; 6150 arc_warm = B_FALSE; 6151 6152 /* 6153 * Calculate maximum amount of dirty data per pool. 6154 * 6155 * If it has been set by /etc/system, take that. 6156 * Otherwise, use a percentage of physical memory defined by 6157 * zfs_dirty_data_max_percent (default 10%) with a cap at 6158 * zfs_dirty_data_max_max (default 4GB). 6159 */ 6160 if (zfs_dirty_data_max == 0) { 6161 zfs_dirty_data_max = ptob(physmem) * 6162 zfs_dirty_data_max_percent / 100; 6163 zfs_dirty_data_max = MIN(zfs_dirty_data_max, 6164 zfs_dirty_data_max_max); 6165 } 6166 6167#ifdef _KERNEL 6168 if (TUNABLE_INT_FETCH("vfs.zfs.prefetch_disable", &zfs_prefetch_disable)) 6169 prefetch_tunable_set = 1; 6170 6171#ifdef __i386__ 6172 if (prefetch_tunable_set == 0) { 6173 printf("ZFS NOTICE: Prefetch is disabled by default on i386 " 6174 "-- to enable,\n"); 6175 printf(" add \"vfs.zfs.prefetch_disable=0\" " 6176 "to /boot/loader.conf.\n"); 6177 zfs_prefetch_disable = 1; 6178 } 6179#else 6180 if ((((uint64_t)physmem * PAGESIZE) < (1ULL << 32)) && 6181 prefetch_tunable_set == 0) { 6182 printf("ZFS NOTICE: Prefetch is disabled by default if less " 6183 "than 4GB of RAM is present;\n" 6184 " to enable, add \"vfs.zfs.prefetch_disable=0\" " 6185 "to /boot/loader.conf.\n"); 6186 zfs_prefetch_disable = 1; 6187 } 6188#endif 6189 /* Warn about ZFS memory and address space requirements. */ 6190 if (((uint64_t)physmem * PAGESIZE) < (256 + 128 + 64) * (1 << 20)) { 6191 printf("ZFS WARNING: Recommended minimum RAM size is 512MB; " 6192 "expect unstable behavior.\n"); 6193 } 6194 if (kmem_size() < 512 * (1 << 20)) { 6195 printf("ZFS WARNING: Recommended minimum kmem_size is 512MB; " 6196 "expect unstable behavior.\n"); 6197 printf(" Consider tuning vm.kmem_size and " 6198 "vm.kmem_size_max\n"); 6199 printf(" in /boot/loader.conf.\n"); 6200 } 6201#endif 6202} 6203 6204void 6205arc_fini(void) 6206{ 6207 mutex_enter(&arc_reclaim_lock); 6208 arc_reclaim_thread_exit = B_TRUE; 6209 /* 6210 * The reclaim thread will set arc_reclaim_thread_exit back to 6211 * B_FALSE when it is finished exiting; we're waiting for that. 6212 */ 6213 while (arc_reclaim_thread_exit) { 6214 cv_signal(&arc_reclaim_thread_cv); 6215 cv_wait(&arc_reclaim_thread_cv, &arc_reclaim_lock); 6216 } 6217 mutex_exit(&arc_reclaim_lock); 6218 6219 /* Use B_TRUE to ensure *all* buffers are evicted */ 6220 arc_flush(NULL, B_TRUE); 6221 6222 mutex_enter(&arc_dnlc_evicts_lock); 6223 arc_dnlc_evicts_thread_exit = TRUE; 6224 /* 6225 * The user evicts thread will set arc_user_evicts_thread_exit 6226 * to FALSE when it is finished exiting; we're waiting for that. 6227 */ 6228 while (arc_dnlc_evicts_thread_exit) { 6229 cv_signal(&arc_dnlc_evicts_cv); 6230 cv_wait(&arc_dnlc_evicts_cv, &arc_dnlc_evicts_lock); 6231 } 6232 mutex_exit(&arc_dnlc_evicts_lock); 6233 6234 arc_dead = B_TRUE; 6235 6236 if (arc_ksp != NULL) { 6237 kstat_delete(arc_ksp); 6238 arc_ksp = NULL; 6239 } 6240 6241 mutex_destroy(&arc_reclaim_lock); 6242 cv_destroy(&arc_reclaim_thread_cv); 6243 cv_destroy(&arc_reclaim_waiters_cv); 6244 6245 mutex_destroy(&arc_dnlc_evicts_lock); 6246 cv_destroy(&arc_dnlc_evicts_cv); 6247 6248 arc_state_fini(); 6249 buf_fini(); 6250 6251 ASSERT0(arc_loaned_bytes); 6252 6253#ifdef _KERNEL 6254 if (arc_event_lowmem != NULL) 6255 EVENTHANDLER_DEREGISTER(vm_lowmem, arc_event_lowmem); 6256#endif 6257} 6258 6259/* 6260 * Level 2 ARC 6261 * 6262 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk. 6263 * It uses dedicated storage devices to hold cached data, which are populated 6264 * using large infrequent writes. The main role of this cache is to boost 6265 * the performance of random read workloads. The intended L2ARC devices 6266 * include short-stroked disks, solid state disks, and other media with 6267 * substantially faster read latency than disk. 6268 * 6269 * +-----------------------+ 6270 * | ARC | 6271 * +-----------------------+ 6272 * | ^ ^ 6273 * | | | 6274 * l2arc_feed_thread() arc_read() 6275 * | | | 6276 * | l2arc read | 6277 * V | | 6278 * +---------------+ | 6279 * | L2ARC | | 6280 * +---------------+ | 6281 * | ^ | 6282 * l2arc_write() | | 6283 * | | | 6284 * V | | 6285 * +-------+ +-------+ 6286 * | vdev | | vdev | 6287 * | cache | | cache | 6288 * +-------+ +-------+ 6289 * +=========+ .-----. 6290 * : L2ARC : |-_____-| 6291 * : devices : | Disks | 6292 * +=========+ `-_____-' 6293 * 6294 * Read requests are satisfied from the following sources, in order: 6295 * 6296 * 1) ARC 6297 * 2) vdev cache of L2ARC devices 6298 * 3) L2ARC devices 6299 * 4) vdev cache of disks 6300 * 5) disks 6301 * 6302 * Some L2ARC device types exhibit extremely slow write performance. 6303 * To accommodate for this there are some significant differences between 6304 * the L2ARC and traditional cache design: 6305 * 6306 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from 6307 * the ARC behave as usual, freeing buffers and placing headers on ghost 6308 * lists. The ARC does not send buffers to the L2ARC during eviction as 6309 * this would add inflated write latencies for all ARC memory pressure. 6310 * 6311 * 2. The L2ARC attempts to cache data from the ARC before it is evicted. 6312 * It does this by periodically scanning buffers from the eviction-end of 6313 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are 6314 * not already there. It scans until a headroom of buffers is satisfied, 6315 * which itself is a buffer for ARC eviction. If a compressible buffer is 6316 * found during scanning and selected for writing to an L2ARC device, we 6317 * temporarily boost scanning headroom during the next scan cycle to make 6318 * sure we adapt to compression effects (which might significantly reduce 6319 * the data volume we write to L2ARC). The thread that does this is 6320 * l2arc_feed_thread(), illustrated below; example sizes are included to 6321 * provide a better sense of ratio than this diagram: 6322 * 6323 * head --> tail 6324 * +---------------------+----------+ 6325 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC 6326 * +---------------------+----------+ | o L2ARC eligible 6327 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer 6328 * +---------------------+----------+ | 6329 * 15.9 Gbytes ^ 32 Mbytes | 6330 * headroom | 6331 * l2arc_feed_thread() 6332 * | 6333 * l2arc write hand <--[oooo]--' 6334 * | 8 Mbyte 6335 * | write max 6336 * V 6337 * +==============================+ 6338 * L2ARC dev |####|#|###|###| |####| ... | 6339 * +==============================+ 6340 * 32 Gbytes 6341 * 6342 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of 6343 * evicted, then the L2ARC has cached a buffer much sooner than it probably 6344 * needed to, potentially wasting L2ARC device bandwidth and storage. It is 6345 * safe to say that this is an uncommon case, since buffers at the end of 6346 * the ARC lists have moved there due to inactivity. 6347 * 6348 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom, 6349 * then the L2ARC simply misses copying some buffers. This serves as a 6350 * pressure valve to prevent heavy read workloads from both stalling the ARC 6351 * with waits and clogging the L2ARC with writes. This also helps prevent 6352 * the potential for the L2ARC to churn if it attempts to cache content too 6353 * quickly, such as during backups of the entire pool. 6354 * 6355 * 5. After system boot and before the ARC has filled main memory, there are 6356 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru 6357 * lists can remain mostly static. Instead of searching from tail of these 6358 * lists as pictured, the l2arc_feed_thread() will search from the list heads 6359 * for eligible buffers, greatly increasing its chance of finding them. 6360 * 6361 * The L2ARC device write speed is also boosted during this time so that 6362 * the L2ARC warms up faster. Since there have been no ARC evictions yet, 6363 * there are no L2ARC reads, and no fear of degrading read performance 6364 * through increased writes. 6365 * 6366 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that 6367 * the vdev queue can aggregate them into larger and fewer writes. Each 6368 * device is written to in a rotor fashion, sweeping writes through 6369 * available space then repeating. 6370 * 6371 * 7. The L2ARC does not store dirty content. It never needs to flush 6372 * write buffers back to disk based storage. 6373 * 6374 * 8. If an ARC buffer is written (and dirtied) which also exists in the 6375 * L2ARC, the now stale L2ARC buffer is immediately dropped. 6376 * 6377 * The performance of the L2ARC can be tweaked by a number of tunables, which 6378 * may be necessary for different workloads: 6379 * 6380 * l2arc_write_max max write bytes per interval 6381 * l2arc_write_boost extra write bytes during device warmup 6382 * l2arc_noprefetch skip caching prefetched buffers 6383 * l2arc_headroom number of max device writes to precache 6384 * l2arc_headroom_boost when we find compressed buffers during ARC 6385 * scanning, we multiply headroom by this 6386 * percentage factor for the next scan cycle, 6387 * since more compressed buffers are likely to 6388 * be present 6389 * l2arc_feed_secs seconds between L2ARC writing 6390 * 6391 * Tunables may be removed or added as future performance improvements are 6392 * integrated, and also may become zpool properties. 6393 * 6394 * There are three key functions that control how the L2ARC warms up: 6395 * 6396 * l2arc_write_eligible() check if a buffer is eligible to cache 6397 * l2arc_write_size() calculate how much to write 6398 * l2arc_write_interval() calculate sleep delay between writes 6399 * 6400 * These three functions determine what to write, how much, and how quickly 6401 * to send writes. 6402 */ 6403 6404static boolean_t 6405l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *hdr) 6406{ 6407 /* 6408 * A buffer is *not* eligible for the L2ARC if it: 6409 * 1. belongs to a different spa. 6410 * 2. is already cached on the L2ARC. 6411 * 3. has an I/O in progress (it may be an incomplete read). 6412 * 4. is flagged not eligible (zfs property). 6413 */ 6414 if (hdr->b_spa != spa_guid) { 6415 ARCSTAT_BUMP(arcstat_l2_write_spa_mismatch); 6416 return (B_FALSE); 6417 } 6418 if (HDR_HAS_L2HDR(hdr)) { 6419 ARCSTAT_BUMP(arcstat_l2_write_in_l2); 6420 return (B_FALSE); 6421 } 6422 if (HDR_IO_IN_PROGRESS(hdr)) { 6423 ARCSTAT_BUMP(arcstat_l2_write_hdr_io_in_progress); 6424 return (B_FALSE); 6425 } 6426 if (!HDR_L2CACHE(hdr)) { 6427 ARCSTAT_BUMP(arcstat_l2_write_not_cacheable); 6428 return (B_FALSE); 6429 } 6430 6431 return (B_TRUE); 6432} 6433 6434static uint64_t 6435l2arc_write_size(void) 6436{ 6437 uint64_t size; 6438 6439 /* 6440 * Make sure our globals have meaningful values in case the user 6441 * altered them. 6442 */ 6443 size = l2arc_write_max; 6444 if (size == 0) { 6445 cmn_err(CE_NOTE, "Bad value for l2arc_write_max, value must " 6446 "be greater than zero, resetting it to the default (%d)", 6447 L2ARC_WRITE_SIZE); 6448 size = l2arc_write_max = L2ARC_WRITE_SIZE; 6449 } 6450 6451 if (arc_warm == B_FALSE) 6452 size += l2arc_write_boost; 6453 6454 return (size); 6455 6456} 6457 6458static clock_t 6459l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote) 6460{ 6461 clock_t interval, next, now; 6462 6463 /* 6464 * If the ARC lists are busy, increase our write rate; if the 6465 * lists are stale, idle back. This is achieved by checking 6466 * how much we previously wrote - if it was more than half of 6467 * what we wanted, schedule the next write much sooner. 6468 */ 6469 if (l2arc_feed_again && wrote > (wanted / 2)) 6470 interval = (hz * l2arc_feed_min_ms) / 1000; 6471 else 6472 interval = hz * l2arc_feed_secs; 6473 6474 now = ddi_get_lbolt(); 6475 next = MAX(now, MIN(now + interval, began + interval)); 6476 6477 return (next); 6478} 6479 6480/* 6481 * Cycle through L2ARC devices. This is how L2ARC load balances. 6482 * If a device is returned, this also returns holding the spa config lock. 6483 */ 6484static l2arc_dev_t * 6485l2arc_dev_get_next(void) 6486{ 6487 l2arc_dev_t *first, *next = NULL; 6488 6489 /* 6490 * Lock out the removal of spas (spa_namespace_lock), then removal 6491 * of cache devices (l2arc_dev_mtx). Once a device has been selected, 6492 * both locks will be dropped and a spa config lock held instead. 6493 */ 6494 mutex_enter(&spa_namespace_lock); 6495 mutex_enter(&l2arc_dev_mtx); 6496 6497 /* if there are no vdevs, there is nothing to do */ 6498 if (l2arc_ndev == 0) 6499 goto out; 6500 6501 first = NULL; 6502 next = l2arc_dev_last; 6503 do { 6504 /* loop around the list looking for a non-faulted vdev */ 6505 if (next == NULL) { 6506 next = list_head(l2arc_dev_list); 6507 } else { 6508 next = list_next(l2arc_dev_list, next); 6509 if (next == NULL) 6510 next = list_head(l2arc_dev_list); 6511 } 6512 6513 /* if we have come back to the start, bail out */ 6514 if (first == NULL) 6515 first = next; 6516 else if (next == first) 6517 break; 6518 6519 } while (vdev_is_dead(next->l2ad_vdev)); 6520 6521 /* if we were unable to find any usable vdevs, return NULL */ 6522 if (vdev_is_dead(next->l2ad_vdev)) 6523 next = NULL; 6524 6525 l2arc_dev_last = next; 6526 6527out: 6528 mutex_exit(&l2arc_dev_mtx); 6529 6530 /* 6531 * Grab the config lock to prevent the 'next' device from being 6532 * removed while we are writing to it. 6533 */ 6534 if (next != NULL) 6535 spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER); 6536 mutex_exit(&spa_namespace_lock); 6537 6538 return (next); 6539} 6540 6541/* 6542 * Free buffers that were tagged for destruction. 6543 */ 6544static void 6545l2arc_do_free_on_write() 6546{ 6547 list_t *buflist; 6548 l2arc_data_free_t *df, *df_prev; 6549 6550 mutex_enter(&l2arc_free_on_write_mtx); 6551 buflist = l2arc_free_on_write; 6552 6553 for (df = list_tail(buflist); df; df = df_prev) { 6554 df_prev = list_prev(buflist, df); 6555 ASSERT3P(df->l2df_data, !=, NULL); 6556 if (df->l2df_type == ARC_BUFC_METADATA) { 6557 zio_buf_free(df->l2df_data, df->l2df_size); 6558 } else { 6559 ASSERT(df->l2df_type == ARC_BUFC_DATA); 6560 zio_data_buf_free(df->l2df_data, df->l2df_size); 6561 } 6562 list_remove(buflist, df); 6563 kmem_free(df, sizeof (l2arc_data_free_t)); 6564 } 6565 6566 mutex_exit(&l2arc_free_on_write_mtx); 6567} 6568 6569/* 6570 * A write to a cache device has completed. Update all headers to allow 6571 * reads from these buffers to begin. 6572 */ 6573static void 6574l2arc_write_done(zio_t *zio) 6575{ 6576 l2arc_write_callback_t *cb; 6577 l2arc_dev_t *dev; 6578 list_t *buflist; 6579 arc_buf_hdr_t *head, *hdr, *hdr_prev; 6580 kmutex_t *hash_lock; 6581 int64_t bytes_dropped = 0; 6582 6583 cb = zio->io_private; 6584 ASSERT3P(cb, !=, NULL); 6585 dev = cb->l2wcb_dev; 6586 ASSERT3P(dev, !=, NULL); 6587 head = cb->l2wcb_head; 6588 ASSERT3P(head, !=, NULL); 6589 buflist = &dev->l2ad_buflist; 6590 ASSERT3P(buflist, !=, NULL); 6591 DTRACE_PROBE2(l2arc__iodone, zio_t *, zio, 6592 l2arc_write_callback_t *, cb); 6593 6594 if (zio->io_error != 0) 6595 ARCSTAT_BUMP(arcstat_l2_writes_error); 6596 6597 /* 6598 * All writes completed, or an error was hit. 6599 */ 6600top: 6601 mutex_enter(&dev->l2ad_mtx); 6602 for (hdr = list_prev(buflist, head); hdr; hdr = hdr_prev) { 6603 hdr_prev = list_prev(buflist, hdr); 6604 6605 hash_lock = HDR_LOCK(hdr); 6606 6607 /* 6608 * We cannot use mutex_enter or else we can deadlock 6609 * with l2arc_write_buffers (due to swapping the order 6610 * the hash lock and l2ad_mtx are taken). 6611 */ 6612 if (!mutex_tryenter(hash_lock)) { 6613 /* 6614 * Missed the hash lock. We must retry so we 6615 * don't leave the ARC_FLAG_L2_WRITING bit set. 6616 */ 6617 ARCSTAT_BUMP(arcstat_l2_writes_lock_retry); 6618 6619 /* 6620 * We don't want to rescan the headers we've 6621 * already marked as having been written out, so 6622 * we reinsert the head node so we can pick up 6623 * where we left off. 6624 */ 6625 list_remove(buflist, head); 6626 list_insert_after(buflist, hdr, head); 6627 6628 mutex_exit(&dev->l2ad_mtx); 6629 6630 /* 6631 * We wait for the hash lock to become available 6632 * to try and prevent busy waiting, and increase 6633 * the chance we'll be able to acquire the lock 6634 * the next time around. 6635 */ 6636 mutex_enter(hash_lock); 6637 mutex_exit(hash_lock); 6638 goto top; 6639 } 6640 6641 /* 6642 * We could not have been moved into the arc_l2c_only 6643 * state while in-flight due to our ARC_FLAG_L2_WRITING 6644 * bit being set. Let's just ensure that's being enforced. 6645 */ 6646 ASSERT(HDR_HAS_L1HDR(hdr)); 6647 6648 if (zio->io_error != 0) { 6649 /* 6650 * Error - drop L2ARC entry. 6651 */ 6652 list_remove(buflist, hdr); 6653 l2arc_trim(hdr); 6654 arc_hdr_clear_flags(hdr, ARC_FLAG_HAS_L2HDR); 6655 6656 ARCSTAT_INCR(arcstat_l2_asize, -arc_hdr_size(hdr)); 6657 ARCSTAT_INCR(arcstat_l2_size, -HDR_GET_LSIZE(hdr)); 6658 6659 bytes_dropped += arc_hdr_size(hdr); 6660 (void) refcount_remove_many(&dev->l2ad_alloc, 6661 arc_hdr_size(hdr), hdr); 6662 } 6663 6664 /* 6665 * Allow ARC to begin reads and ghost list evictions to 6666 * this L2ARC entry. 6667 */ 6668 arc_hdr_clear_flags(hdr, ARC_FLAG_L2_WRITING); 6669 6670 mutex_exit(hash_lock); 6671 } 6672 6673 atomic_inc_64(&l2arc_writes_done); 6674 list_remove(buflist, head); 6675 ASSERT(!HDR_HAS_L1HDR(head)); 6676 kmem_cache_free(hdr_l2only_cache, head); 6677 mutex_exit(&dev->l2ad_mtx); 6678 6679 vdev_space_update(dev->l2ad_vdev, -bytes_dropped, 0, 0); 6680 6681 l2arc_do_free_on_write(); 6682 6683 kmem_free(cb, sizeof (l2arc_write_callback_t)); 6684} 6685 6686/* 6687 * A read to a cache device completed. Validate buffer contents before 6688 * handing over to the regular ARC routines. 6689 */ 6690static void 6691l2arc_read_done(zio_t *zio) 6692{ 6693 l2arc_read_callback_t *cb; 6694 arc_buf_hdr_t *hdr; 6695 kmutex_t *hash_lock; 6696 boolean_t valid_cksum; 6697 6698 ASSERT3P(zio->io_vd, !=, NULL); 6699 ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE); 6700 6701 spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd); 6702 6703 cb = zio->io_private; 6704 ASSERT3P(cb, !=, NULL); 6705 hdr = cb->l2rcb_hdr; 6706 ASSERT3P(hdr, !=, NULL); 6707 6708 hash_lock = HDR_LOCK(hdr); 6709 mutex_enter(hash_lock); 6710 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 6711 6712 /* 6713 * If the data was read into a temporary buffer, 6714 * move it and free the buffer. 6715 */ 6716 if (cb->l2rcb_data != NULL) { 6717 ASSERT3U(arc_hdr_size(hdr), <, zio->io_size); 6718 if (zio->io_error == 0) { 6719 bcopy(cb->l2rcb_data, hdr->b_l1hdr.b_pdata, 6720 arc_hdr_size(hdr)); 6721 } 6722 6723 /* 6724 * The following must be done regardless of whether 6725 * there was an error: 6726 * - free the temporary buffer 6727 * - point zio to the real ARC buffer 6728 * - set zio size accordingly 6729 * These are required because zio is either re-used for 6730 * an I/O of the block in the case of the error 6731 * or the zio is passed to arc_read_done() and it 6732 * needs real data. 6733 */ 6734 zio_data_buf_free(cb->l2rcb_data, zio->io_size); 6735 zio->io_size = zio->io_orig_size = arc_hdr_size(hdr); 6736 zio->io_data = zio->io_orig_data = hdr->b_l1hdr.b_pdata; 6737 } 6738 6739 ASSERT3P(zio->io_data, !=, NULL); 6740 6741 /* 6742 * Check this survived the L2ARC journey. 6743 */ 6744 ASSERT3P(zio->io_data, ==, hdr->b_l1hdr.b_pdata); 6745 zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */ 6746 zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */ 6747 6748 valid_cksum = arc_cksum_is_equal(hdr, zio); 6749 if (valid_cksum && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) { 6750 mutex_exit(hash_lock); 6751 zio->io_private = hdr; 6752 arc_read_done(zio); 6753 } else { 6754 mutex_exit(hash_lock); 6755 /* 6756 * Buffer didn't survive caching. Increment stats and 6757 * reissue to the original storage device. 6758 */ 6759 if (zio->io_error != 0) { 6760 ARCSTAT_BUMP(arcstat_l2_io_error); 6761 } else { 6762 zio->io_error = SET_ERROR(EIO); 6763 } 6764 if (!valid_cksum) 6765 ARCSTAT_BUMP(arcstat_l2_cksum_bad); 6766 6767 /* 6768 * If there's no waiter, issue an async i/o to the primary 6769 * storage now. If there *is* a waiter, the caller must 6770 * issue the i/o in a context where it's OK to block. 6771 */ 6772 if (zio->io_waiter == NULL) { 6773 zio_t *pio = zio_unique_parent(zio); 6774 6775 ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL); 6776 6777 zio_nowait(zio_read(pio, zio->io_spa, zio->io_bp, 6778 hdr->b_l1hdr.b_pdata, zio->io_size, arc_read_done, 6779 hdr, zio->io_priority, cb->l2rcb_flags, 6780 &cb->l2rcb_zb)); 6781 } 6782 } 6783 6784 kmem_free(cb, sizeof (l2arc_read_callback_t)); 6785} 6786 6787/* 6788 * This is the list priority from which the L2ARC will search for pages to 6789 * cache. This is used within loops (0..3) to cycle through lists in the 6790 * desired order. This order can have a significant effect on cache 6791 * performance. 6792 * 6793 * Currently the metadata lists are hit first, MFU then MRU, followed by 6794 * the data lists. This function returns a locked list, and also returns 6795 * the lock pointer. 6796 */ 6797static multilist_sublist_t * 6798l2arc_sublist_lock(int list_num) 6799{ 6800 multilist_t *ml = NULL; 6801 unsigned int idx; 6802 6803 ASSERT(list_num >= 0 && list_num <= 3); 6804 6805 switch (list_num) { 6806 case 0: 6807 ml = &arc_mfu->arcs_list[ARC_BUFC_METADATA]; 6808 break; 6809 case 1: 6810 ml = &arc_mru->arcs_list[ARC_BUFC_METADATA]; 6811 break; 6812 case 2: 6813 ml = &arc_mfu->arcs_list[ARC_BUFC_DATA]; 6814 break; 6815 case 3: 6816 ml = &arc_mru->arcs_list[ARC_BUFC_DATA]; 6817 break; 6818 } 6819 6820 /* 6821 * Return a randomly-selected sublist. This is acceptable 6822 * because the caller feeds only a little bit of data for each 6823 * call (8MB). Subsequent calls will result in different 6824 * sublists being selected. 6825 */ 6826 idx = multilist_get_random_index(ml); 6827 return (multilist_sublist_lock(ml, idx)); 6828} 6829 6830/* 6831 * Evict buffers from the device write hand to the distance specified in 6832 * bytes. This distance may span populated buffers, it may span nothing. 6833 * This is clearing a region on the L2ARC device ready for writing. 6834 * If the 'all' boolean is set, every buffer is evicted. 6835 */ 6836static void 6837l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all) 6838{ 6839 list_t *buflist; 6840 arc_buf_hdr_t *hdr, *hdr_prev; 6841 kmutex_t *hash_lock; 6842 uint64_t taddr; 6843 6844 buflist = &dev->l2ad_buflist; 6845 6846 if (!all && dev->l2ad_first) { 6847 /* 6848 * This is the first sweep through the device. There is 6849 * nothing to evict. 6850 */ 6851 return; 6852 } 6853 6854 if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) { 6855 /* 6856 * When nearing the end of the device, evict to the end 6857 * before the device write hand jumps to the start. 6858 */ 6859 taddr = dev->l2ad_end; 6860 } else { 6861 taddr = dev->l2ad_hand + distance; 6862 } 6863 DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist, 6864 uint64_t, taddr, boolean_t, all); 6865 6866top: 6867 mutex_enter(&dev->l2ad_mtx); 6868 for (hdr = list_tail(buflist); hdr; hdr = hdr_prev) { 6869 hdr_prev = list_prev(buflist, hdr); 6870 6871 hash_lock = HDR_LOCK(hdr); 6872 6873 /* 6874 * We cannot use mutex_enter or else we can deadlock 6875 * with l2arc_write_buffers (due to swapping the order 6876 * the hash lock and l2ad_mtx are taken). 6877 */ 6878 if (!mutex_tryenter(hash_lock)) { 6879 /* 6880 * Missed the hash lock. Retry. 6881 */ 6882 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry); 6883 mutex_exit(&dev->l2ad_mtx); 6884 mutex_enter(hash_lock); 6885 mutex_exit(hash_lock); 6886 goto top; 6887 } 6888 6889 if (HDR_L2_WRITE_HEAD(hdr)) { 6890 /* 6891 * We hit a write head node. Leave it for 6892 * l2arc_write_done(). 6893 */ 6894 list_remove(buflist, hdr); 6895 mutex_exit(hash_lock); 6896 continue; 6897 } 6898 6899 if (!all && HDR_HAS_L2HDR(hdr) && 6900 (hdr->b_l2hdr.b_daddr > taddr || 6901 hdr->b_l2hdr.b_daddr < dev->l2ad_hand)) { 6902 /* 6903 * We've evicted to the target address, 6904 * or the end of the device. 6905 */ 6906 mutex_exit(hash_lock); 6907 break; 6908 } 6909 6910 ASSERT(HDR_HAS_L2HDR(hdr)); 6911 if (!HDR_HAS_L1HDR(hdr)) { 6912 ASSERT(!HDR_L2_READING(hdr)); 6913 /* 6914 * This doesn't exist in the ARC. Destroy. 6915 * arc_hdr_destroy() will call list_remove() 6916 * and decrement arcstat_l2_size. 6917 */ 6918 arc_change_state(arc_anon, hdr, hash_lock); 6919 arc_hdr_destroy(hdr); 6920 } else { 6921 ASSERT(hdr->b_l1hdr.b_state != arc_l2c_only); 6922 ARCSTAT_BUMP(arcstat_l2_evict_l1cached); 6923 /* 6924 * Invalidate issued or about to be issued 6925 * reads, since we may be about to write 6926 * over this location. 6927 */ 6928 if (HDR_L2_READING(hdr)) { 6929 ARCSTAT_BUMP(arcstat_l2_evict_reading); 6930 arc_hdr_set_flags(hdr, ARC_FLAG_L2_EVICTED); 6931 } 6932 6933 /* Ensure this header has finished being written */ 6934 ASSERT(!HDR_L2_WRITING(hdr)); 6935 6936 arc_hdr_l2hdr_destroy(hdr); 6937 } 6938 mutex_exit(hash_lock); 6939 } 6940 mutex_exit(&dev->l2ad_mtx); 6941} 6942 6943/* 6944 * Find and write ARC buffers to the L2ARC device. 6945 * 6946 * An ARC_FLAG_L2_WRITING flag is set so that the L2ARC buffers are not valid 6947 * for reading until they have completed writing. 6948 * The headroom_boost is an in-out parameter used to maintain headroom boost 6949 * state between calls to this function. 6950 * 6951 * Returns the number of bytes actually written (which may be smaller than 6952 * the delta by which the device hand has changed due to alignment). 6953 */ 6954static uint64_t 6955l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz) 6956{ 6957 arc_buf_hdr_t *hdr, *hdr_prev, *head; 6958 uint64_t write_asize, write_psize, write_sz, headroom; 6959 boolean_t full; 6960 l2arc_write_callback_t *cb; 6961 zio_t *pio, *wzio; 6962 uint64_t guid = spa_load_guid(spa); 6963 int try; 6964 6965 ASSERT3P(dev->l2ad_vdev, !=, NULL); 6966 6967 pio = NULL; 6968 write_sz = write_asize = write_psize = 0; 6969 full = B_FALSE; 6970 head = kmem_cache_alloc(hdr_l2only_cache, KM_PUSHPAGE); 6971 arc_hdr_set_flags(head, ARC_FLAG_L2_WRITE_HEAD | ARC_FLAG_HAS_L2HDR); 6972 6973 ARCSTAT_BUMP(arcstat_l2_write_buffer_iter); 6974 /* 6975 * Copy buffers for L2ARC writing. 6976 */ 6977 for (try = 0; try <= 3; try++) { 6978 multilist_sublist_t *mls = l2arc_sublist_lock(try); 6979 uint64_t passed_sz = 0; 6980 6981 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_iter); 6982 6983 /* 6984 * L2ARC fast warmup. 6985 * 6986 * Until the ARC is warm and starts to evict, read from the 6987 * head of the ARC lists rather than the tail. 6988 */ 6989 if (arc_warm == B_FALSE) 6990 hdr = multilist_sublist_head(mls); 6991 else 6992 hdr = multilist_sublist_tail(mls); 6993 if (hdr == NULL) 6994 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_null_iter); 6995 6996 headroom = target_sz * l2arc_headroom; 6997 if (zfs_compressed_arc_enabled) 6998 headroom = (headroom * l2arc_headroom_boost) / 100; 6999 7000 for (; hdr; hdr = hdr_prev) { 7001 kmutex_t *hash_lock; 7002 7003 if (arc_warm == B_FALSE) 7004 hdr_prev = multilist_sublist_next(mls, hdr); 7005 else 7006 hdr_prev = multilist_sublist_prev(mls, hdr); 7007 ARCSTAT_INCR(arcstat_l2_write_buffer_bytes_scanned, 7008 HDR_GET_LSIZE(hdr)); 7009 7010 hash_lock = HDR_LOCK(hdr); 7011 if (!mutex_tryenter(hash_lock)) { 7012 ARCSTAT_BUMP(arcstat_l2_write_trylock_fail); 7013 /* 7014 * Skip this buffer rather than waiting. 7015 */ 7016 continue; 7017 } 7018 7019 passed_sz += HDR_GET_LSIZE(hdr); 7020 if (passed_sz > headroom) { 7021 /* 7022 * Searched too far. 7023 */ 7024 mutex_exit(hash_lock); 7025 ARCSTAT_BUMP(arcstat_l2_write_passed_headroom); 7026 break; 7027 } 7028 7029 if (!l2arc_write_eligible(guid, hdr)) { 7030 mutex_exit(hash_lock); 7031 continue; 7032 } 7033 7034 if ((write_asize + HDR_GET_LSIZE(hdr)) > target_sz) { 7035 full = B_TRUE; 7036 mutex_exit(hash_lock); 7037 ARCSTAT_BUMP(arcstat_l2_write_full); 7038 break; 7039 } 7040 7041 if (pio == NULL) { 7042 /* 7043 * Insert a dummy header on the buflist so 7044 * l2arc_write_done() can find where the 7045 * write buffers begin without searching. 7046 */ 7047 mutex_enter(&dev->l2ad_mtx); 7048 list_insert_head(&dev->l2ad_buflist, head); 7049 mutex_exit(&dev->l2ad_mtx); 7050 7051 cb = kmem_alloc( 7052 sizeof (l2arc_write_callback_t), KM_SLEEP); 7053 cb->l2wcb_dev = dev; 7054 cb->l2wcb_head = head; 7055 pio = zio_root(spa, l2arc_write_done, cb, 7056 ZIO_FLAG_CANFAIL); 7057 ARCSTAT_BUMP(arcstat_l2_write_pios); 7058 } 7059 7060 hdr->b_l2hdr.b_dev = dev; 7061 hdr->b_l2hdr.b_daddr = dev->l2ad_hand; 7062 arc_hdr_set_flags(hdr, 7063 ARC_FLAG_L2_WRITING | ARC_FLAG_HAS_L2HDR); 7064 7065 mutex_enter(&dev->l2ad_mtx); 7066 list_insert_head(&dev->l2ad_buflist, hdr); 7067 mutex_exit(&dev->l2ad_mtx); 7068 7069 /* 7070 * We rely on the L1 portion of the header below, so 7071 * it's invalid for this header to have been evicted out 7072 * of the ghost cache, prior to being written out. The 7073 * ARC_FLAG_L2_WRITING bit ensures this won't happen. 7074 */ 7075 ASSERT(HDR_HAS_L1HDR(hdr)); 7076 7077 ASSERT3U(HDR_GET_PSIZE(hdr), >, 0); 7078 ASSERT3P(hdr->b_l1hdr.b_pdata, !=, NULL); 7079 ASSERT3U(arc_hdr_size(hdr), >, 0); 7080 uint64_t size = arc_hdr_size(hdr); 7081 uint64_t asize = vdev_psize_to_asize(dev->l2ad_vdev, 7082 size); 7083 7084 (void) refcount_add_many(&dev->l2ad_alloc, size, hdr); 7085 7086 /* 7087 * Normally the L2ARC can use the hdr's data, but if 7088 * we're sharing data between the hdr and one of its 7089 * bufs, L2ARC needs its own copy of the data so that 7090 * the ZIO below can't race with the buf consumer. To 7091 * ensure that this copy will be available for the 7092 * lifetime of the ZIO and be cleaned up afterwards, we 7093 * add it to the l2arc_free_on_write queue. 7094 */ 7095 void *to_write; 7096 if (!HDR_SHARED_DATA(hdr) && size == asize) { 7097 to_write = hdr->b_l1hdr.b_pdata; 7098 } else { 7099 arc_buf_contents_t type = arc_buf_type(hdr); 7100 if (type == ARC_BUFC_METADATA) { 7101 to_write = zio_buf_alloc(asize); 7102 } else { 7103 ASSERT3U(type, ==, ARC_BUFC_DATA); 7104 to_write = zio_data_buf_alloc(asize); 7105 } 7106 7107 bcopy(hdr->b_l1hdr.b_pdata, to_write, size); 7108 if (asize != size) 7109 bzero(to_write + size, asize - size); 7110 l2arc_free_data_on_write(to_write, asize, type); 7111 } 7112 wzio = zio_write_phys(pio, dev->l2ad_vdev, 7113 hdr->b_l2hdr.b_daddr, asize, to_write, 7114 ZIO_CHECKSUM_OFF, NULL, hdr, 7115 ZIO_PRIORITY_ASYNC_WRITE, 7116 ZIO_FLAG_CANFAIL, B_FALSE); 7117 7118 write_sz += HDR_GET_LSIZE(hdr); 7119 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev, 7120 zio_t *, wzio); 7121 7122 write_asize += size; 7123 write_psize += asize; 7124 dev->l2ad_hand += asize; 7125 7126 mutex_exit(hash_lock); 7127 7128 (void) zio_nowait(wzio); 7129 } 7130 7131 multilist_sublist_unlock(mls); 7132 7133 if (full == B_TRUE) 7134 break; 7135 } 7136 7137 /* No buffers selected for writing? */ 7138 if (pio == NULL) { 7139 ASSERT0(write_sz); 7140 ASSERT(!HDR_HAS_L1HDR(head)); 7141 kmem_cache_free(hdr_l2only_cache, head); 7142 return (0); 7143 } 7144 7145 ASSERT3U(write_asize, <=, target_sz); 7146 ARCSTAT_BUMP(arcstat_l2_writes_sent); 7147 ARCSTAT_INCR(arcstat_l2_write_bytes, write_asize); 7148 ARCSTAT_INCR(arcstat_l2_size, write_sz); 7149 ARCSTAT_INCR(arcstat_l2_asize, write_asize); 7150 vdev_space_update(dev->l2ad_vdev, write_asize, 0, 0); 7151 7152 /* 7153 * Bump device hand to the device start if it is approaching the end. 7154 * l2arc_evict() will already have evicted ahead for this case. 7155 */ 7156 if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) { 7157 dev->l2ad_hand = dev->l2ad_start; 7158 dev->l2ad_first = B_FALSE; 7159 } 7160 7161 dev->l2ad_writing = B_TRUE; 7162 (void) zio_wait(pio); 7163 dev->l2ad_writing = B_FALSE; 7164 7165 return (write_asize); 7166} 7167 7168/* 7169 * This thread feeds the L2ARC at regular intervals. This is the beating 7170 * heart of the L2ARC. 7171 */ 7172static void 7173l2arc_feed_thread(void *dummy __unused) 7174{ 7175 callb_cpr_t cpr; 7176 l2arc_dev_t *dev; 7177 spa_t *spa; 7178 uint64_t size, wrote; 7179 clock_t begin, next = ddi_get_lbolt(); 7180 7181 CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG); 7182 7183 mutex_enter(&l2arc_feed_thr_lock); 7184 7185 while (l2arc_thread_exit == 0) { 7186 CALLB_CPR_SAFE_BEGIN(&cpr); 7187 (void) cv_timedwait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock, 7188 next - ddi_get_lbolt()); 7189 CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock); 7190 next = ddi_get_lbolt() + hz; 7191 7192 /* 7193 * Quick check for L2ARC devices. 7194 */ 7195 mutex_enter(&l2arc_dev_mtx); 7196 if (l2arc_ndev == 0) { 7197 mutex_exit(&l2arc_dev_mtx); 7198 continue; 7199 } 7200 mutex_exit(&l2arc_dev_mtx); 7201 begin = ddi_get_lbolt(); 7202 7203 /* 7204 * This selects the next l2arc device to write to, and in 7205 * doing so the next spa to feed from: dev->l2ad_spa. This 7206 * will return NULL if there are now no l2arc devices or if 7207 * they are all faulted. 7208 * 7209 * If a device is returned, its spa's config lock is also 7210 * held to prevent device removal. l2arc_dev_get_next() 7211 * will grab and release l2arc_dev_mtx. 7212 */ 7213 if ((dev = l2arc_dev_get_next()) == NULL) 7214 continue; 7215 7216 spa = dev->l2ad_spa; 7217 ASSERT3P(spa, !=, NULL); 7218 7219 /* 7220 * If the pool is read-only then force the feed thread to 7221 * sleep a little longer. 7222 */ 7223 if (!spa_writeable(spa)) { 7224 next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz; 7225 spa_config_exit(spa, SCL_L2ARC, dev); 7226 continue; 7227 } 7228 7229 /* 7230 * Avoid contributing to memory pressure. 7231 */ 7232 if (arc_reclaim_needed()) { 7233 ARCSTAT_BUMP(arcstat_l2_abort_lowmem); 7234 spa_config_exit(spa, SCL_L2ARC, dev); 7235 continue; 7236 } 7237 7238 ARCSTAT_BUMP(arcstat_l2_feeds); 7239 7240 size = l2arc_write_size(); 7241 7242 /* 7243 * Evict L2ARC buffers that will be overwritten. 7244 */ 7245 l2arc_evict(dev, size, B_FALSE); 7246 7247 /* 7248 * Write ARC buffers. 7249 */ 7250 wrote = l2arc_write_buffers(spa, dev, size); 7251 7252 /* 7253 * Calculate interval between writes. 7254 */ 7255 next = l2arc_write_interval(begin, size, wrote); 7256 spa_config_exit(spa, SCL_L2ARC, dev); 7257 } 7258 7259 l2arc_thread_exit = 0; 7260 cv_broadcast(&l2arc_feed_thr_cv); 7261 CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */ 7262 thread_exit(); 7263} 7264 7265boolean_t 7266l2arc_vdev_present(vdev_t *vd) 7267{ 7268 l2arc_dev_t *dev; 7269 7270 mutex_enter(&l2arc_dev_mtx); 7271 for (dev = list_head(l2arc_dev_list); dev != NULL; 7272 dev = list_next(l2arc_dev_list, dev)) { 7273 if (dev->l2ad_vdev == vd) 7274 break; 7275 } 7276 mutex_exit(&l2arc_dev_mtx); 7277 7278 return (dev != NULL); 7279} 7280 7281/* 7282 * Add a vdev for use by the L2ARC. By this point the spa has already 7283 * validated the vdev and opened it. 7284 */ 7285void 7286l2arc_add_vdev(spa_t *spa, vdev_t *vd) 7287{ 7288 l2arc_dev_t *adddev; 7289 7290 ASSERT(!l2arc_vdev_present(vd)); 7291 7292 vdev_ashift_optimize(vd); 7293 7294 /* 7295 * Create a new l2arc device entry. 7296 */ 7297 adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP); 7298 adddev->l2ad_spa = spa; 7299 adddev->l2ad_vdev = vd; 7300 adddev->l2ad_start = VDEV_LABEL_START_SIZE; 7301 adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd); 7302 adddev->l2ad_hand = adddev->l2ad_start; 7303 adddev->l2ad_first = B_TRUE; 7304 adddev->l2ad_writing = B_FALSE; 7305 7306 mutex_init(&adddev->l2ad_mtx, NULL, MUTEX_DEFAULT, NULL); 7307 /* 7308 * This is a list of all ARC buffers that are still valid on the 7309 * device. 7310 */ 7311 list_create(&adddev->l2ad_buflist, sizeof (arc_buf_hdr_t), 7312 offsetof(arc_buf_hdr_t, b_l2hdr.b_l2node)); 7313 7314 vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand); 7315 refcount_create(&adddev->l2ad_alloc); 7316 7317 /* 7318 * Add device to global list 7319 */ 7320 mutex_enter(&l2arc_dev_mtx); 7321 list_insert_head(l2arc_dev_list, adddev); 7322 atomic_inc_64(&l2arc_ndev); 7323 mutex_exit(&l2arc_dev_mtx); 7324} 7325 7326/* 7327 * Remove a vdev from the L2ARC. 7328 */ 7329void 7330l2arc_remove_vdev(vdev_t *vd) 7331{ 7332 l2arc_dev_t *dev, *nextdev, *remdev = NULL; 7333 7334 /* 7335 * Find the device by vdev 7336 */ 7337 mutex_enter(&l2arc_dev_mtx); 7338 for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) { 7339 nextdev = list_next(l2arc_dev_list, dev); 7340 if (vd == dev->l2ad_vdev) { 7341 remdev = dev; 7342 break; 7343 } 7344 } 7345 ASSERT3P(remdev, !=, NULL); 7346 7347 /* 7348 * Remove device from global list 7349 */ 7350 list_remove(l2arc_dev_list, remdev); 7351 l2arc_dev_last = NULL; /* may have been invalidated */ 7352 atomic_dec_64(&l2arc_ndev); 7353 mutex_exit(&l2arc_dev_mtx); 7354 7355 /* 7356 * Clear all buflists and ARC references. L2ARC device flush. 7357 */ 7358 l2arc_evict(remdev, 0, B_TRUE); 7359 list_destroy(&remdev->l2ad_buflist); 7360 mutex_destroy(&remdev->l2ad_mtx); 7361 refcount_destroy(&remdev->l2ad_alloc); 7362 kmem_free(remdev, sizeof (l2arc_dev_t)); 7363} 7364 7365void 7366l2arc_init(void) 7367{ 7368 l2arc_thread_exit = 0; 7369 l2arc_ndev = 0; 7370 l2arc_writes_sent = 0; 7371 l2arc_writes_done = 0; 7372 7373 mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL); 7374 cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL); 7375 mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL); 7376 mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL); 7377 7378 l2arc_dev_list = &L2ARC_dev_list; 7379 l2arc_free_on_write = &L2ARC_free_on_write; 7380 list_create(l2arc_dev_list, sizeof (l2arc_dev_t), 7381 offsetof(l2arc_dev_t, l2ad_node)); 7382 list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t), 7383 offsetof(l2arc_data_free_t, l2df_list_node)); 7384} 7385 7386void 7387l2arc_fini(void) 7388{ 7389 /* 7390 * This is called from dmu_fini(), which is called from spa_fini(); 7391 * Because of this, we can assume that all l2arc devices have 7392 * already been removed when the pools themselves were removed. 7393 */ 7394 7395 l2arc_do_free_on_write(); 7396 7397 mutex_destroy(&l2arc_feed_thr_lock); 7398 cv_destroy(&l2arc_feed_thr_cv); 7399 mutex_destroy(&l2arc_dev_mtx); 7400 mutex_destroy(&l2arc_free_on_write_mtx); 7401 7402 list_destroy(l2arc_dev_list); 7403 list_destroy(l2arc_free_on_write); 7404} 7405 7406void 7407l2arc_start(void) 7408{ 7409 if (!(spa_mode_global & FWRITE)) 7410 return; 7411 7412 (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0, 7413 TS_RUN, minclsyspri); 7414} 7415 7416void 7417l2arc_stop(void) 7418{ 7419 if (!(spa_mode_global & FWRITE)) 7420 return; 7421 7422 mutex_enter(&l2arc_feed_thr_lock); 7423 cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */ 7424 l2arc_thread_exit = 1; 7425 while (l2arc_thread_exit != 0) 7426 cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock); 7427 mutex_exit(&l2arc_feed_thr_lock); 7428} 7429