arc.c revision 321573
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, 2017 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 * Note that the majority of the performance stats are manipulated 103 * with atomic operations. 104 * 105 * The L2ARC uses the l2ad_mtx on each vdev for the following: 106 * 107 * - L2ARC buflist creation 108 * - L2ARC buflist eviction 109 * - L2ARC write completion, which walks L2ARC buflists 110 * - ARC header destruction, as it removes from L2ARC buflists 111 * - ARC header release, as it removes from L2ARC buflists 112 */ 113 114/* 115 * ARC operation: 116 * 117 * Every block that is in the ARC is tracked by an arc_buf_hdr_t structure. 118 * This structure can point either to a block that is still in the cache or to 119 * one that is only accessible in an L2 ARC device, or it can provide 120 * information about a block that was recently evicted. If a block is 121 * only accessible in the L2ARC, then the arc_buf_hdr_t only has enough 122 * information to retrieve it from the L2ARC device. This information is 123 * stored in the l2arc_buf_hdr_t sub-structure of the arc_buf_hdr_t. A block 124 * that is in this state cannot access the data directly. 125 * 126 * Blocks that are actively being referenced or have not been evicted 127 * are cached in the L1ARC. The L1ARC (l1arc_buf_hdr_t) is a structure within 128 * the arc_buf_hdr_t that will point to the data block in memory. A block can 129 * only be read by a consumer if it has an l1arc_buf_hdr_t. The L1ARC 130 * caches data in two ways -- in a list of ARC buffers (arc_buf_t) and 131 * also in the arc_buf_hdr_t's private physical data block pointer (b_pdata). 132 * 133 * The L1ARC's data pointer may or may not be uncompressed. The ARC has the 134 * ability to store the physical data (b_pdata) associated with the DVA of the 135 * arc_buf_hdr_t. Since the b_pdata is a copy of the on-disk physical block, 136 * it will match its on-disk compression characteristics. This behavior can be 137 * disabled by setting 'zfs_compressed_arc_enabled' to B_FALSE. When the 138 * compressed ARC functionality is disabled, the b_pdata will point to an 139 * uncompressed version of the on-disk data. 140 * 141 * Data in the L1ARC is not accessed by consumers of the ARC directly. Each 142 * arc_buf_hdr_t can have multiple ARC buffers (arc_buf_t) which reference it. 143 * Each ARC buffer (arc_buf_t) is being actively accessed by a specific ARC 144 * consumer. The ARC will provide references to this data and will keep it 145 * cached until it is no longer in use. The ARC caches only the L1ARC's physical 146 * data block and will evict any arc_buf_t that is no longer referenced. The 147 * amount of memory consumed by the arc_buf_ts' data buffers can be seen via the 148 * "overhead_size" kstat. 149 * 150 * Depending on the consumer, an arc_buf_t can be requested in uncompressed or 151 * compressed form. The typical case is that consumers will want uncompressed 152 * data, and when that happens a new data buffer is allocated where the data is 153 * decompressed for them to use. Currently the only consumer who wants 154 * compressed arc_buf_t's is "zfs send", when it streams data exactly as it 155 * exists on disk. When this happens, the arc_buf_t's data buffer is shared 156 * with the arc_buf_hdr_t. 157 * 158 * Here is a diagram showing an arc_buf_hdr_t referenced by two arc_buf_t's. The 159 * first one is owned by a compressed send consumer (and therefore references 160 * the same compressed data buffer as the arc_buf_hdr_t) and the second could be 161 * used by any other consumer (and has its own uncompressed copy of the data 162 * buffer). 163 * 164 * arc_buf_hdr_t 165 * +-----------+ 166 * | fields | 167 * | common to | 168 * | L1- and | 169 * | L2ARC | 170 * +-----------+ 171 * | l2arc_buf_hdr_t 172 * | | 173 * +-----------+ 174 * | l1arc_buf_hdr_t 175 * | | arc_buf_t 176 * | b_buf +------------>+-----------+ arc_buf_t 177 * | b_pdata +-+ |b_next +---->+-----------+ 178 * +-----------+ | |-----------| |b_next +-->NULL 179 * | |b_comp = T | +-----------+ 180 * | |b_data +-+ |b_comp = F | 181 * | +-----------+ | |b_data +-+ 182 * +->+------+ | +-----------+ | 183 * compressed | | | | 184 * data | |<--------------+ | uncompressed 185 * +------+ compressed, | data 186 * shared +-->+------+ 187 * data | | 188 * | | 189 * +------+ 190 * 191 * When a consumer reads a block, the ARC must first look to see if the 192 * arc_buf_hdr_t is cached. If the hdr is cached then the ARC allocates a new 193 * arc_buf_t and either copies uncompressed data into a new data buffer from an 194 * existing uncompressed arc_buf_t, decompresses the hdr's b_pdata buffer into a 195 * new data buffer, or shares the hdr's b_pdata buffer, depending on whether the 196 * hdr is compressed and the desired compression characteristics of the 197 * arc_buf_t consumer. If the arc_buf_t ends up sharing data with the 198 * arc_buf_hdr_t and both of them are uncompressed then the arc_buf_t must be 199 * the last buffer in the hdr's b_buf list, however a shared compressed buf can 200 * be anywhere in the hdr's list. 201 * 202 * The diagram below shows an example of an uncompressed ARC hdr that is 203 * sharing its data with an arc_buf_t (note that the shared uncompressed buf is 204 * the last element in the buf list): 205 * 206 * arc_buf_hdr_t 207 * +-----------+ 208 * | | 209 * | | 210 * | | 211 * +-----------+ 212 * l2arc_buf_hdr_t| | 213 * | | 214 * +-----------+ 215 * l1arc_buf_hdr_t| | 216 * | | arc_buf_t (shared) 217 * | b_buf +------------>+---------+ arc_buf_t 218 * | | |b_next +---->+---------+ 219 * | b_pdata +-+ |---------| |b_next +-->NULL 220 * +-----------+ | | | +---------+ 221 * | |b_data +-+ | | 222 * | +---------+ | |b_data +-+ 223 * +->+------+ | +---------+ | 224 * | | | | 225 * uncompressed | | | | 226 * data +------+ | | 227 * ^ +->+------+ | 228 * | uncompressed | | | 229 * | data | | | 230 * | +------+ | 231 * +---------------------------------+ 232 * 233 * Writing to the ARC requires that the ARC first discard the hdr's b_pdata 234 * since the physical block is about to be rewritten. The new data contents 235 * will be contained in the arc_buf_t. As the I/O pipeline performs the write, 236 * it may compress the data before writing it to disk. The ARC will be called 237 * with the transformed data and will bcopy the transformed on-disk block into 238 * a newly allocated b_pdata. Writes are always done into buffers which have 239 * either been loaned (and hence are new and don't have other readers) or 240 * buffers which have been released (and hence have their own hdr, if there 241 * were originally other readers of the buf's original hdr). This ensures that 242 * the ARC only needs to update a single buf and its hdr after a write occurs. 243 * 244 * When the L2ARC is in use, it will also take advantage of the b_pdata. The 245 * L2ARC will always write the contents of b_pdata to the L2ARC. This means 246 * that when compressed ARC is enabled that the L2ARC blocks are identical 247 * to the on-disk block in the main data pool. This provides a significant 248 * advantage since the ARC can leverage the bp's checksum when reading from the 249 * L2ARC to determine if the contents are valid. However, if the compressed 250 * ARC is disabled, then the L2ARC's block must be transformed to look 251 * like the physical block in the main data pool before comparing the 252 * checksum and determining its validity. 253 */ 254 255#include <sys/spa.h> 256#include <sys/zio.h> 257#include <sys/spa_impl.h> 258#include <sys/zio_compress.h> 259#include <sys/zio_checksum.h> 260#include <sys/zfs_context.h> 261#include <sys/arc.h> 262#include <sys/refcount.h> 263#include <sys/vdev.h> 264#include <sys/vdev_impl.h> 265#include <sys/dsl_pool.h> 266#include <sys/multilist.h> 267#ifdef _KERNEL 268#include <sys/dnlc.h> 269#include <sys/racct.h> 270#endif 271#include <sys/callb.h> 272#include <sys/kstat.h> 273#include <sys/trim_map.h> 274#include <zfs_fletcher.h> 275#include <sys/sdt.h> 276 277#include <machine/vmparam.h> 278 279#ifdef illumos 280#ifndef _KERNEL 281/* set with ZFS_DEBUG=watch, to enable watchpoints on frozen buffers */ 282boolean_t arc_watch = B_FALSE; 283int arc_procfd; 284#endif 285#endif /* illumos */ 286 287static kmutex_t arc_reclaim_lock; 288static kcondvar_t arc_reclaim_thread_cv; 289static boolean_t arc_reclaim_thread_exit; 290static kcondvar_t arc_reclaim_waiters_cv; 291 292static kmutex_t arc_dnlc_evicts_lock; 293static kcondvar_t arc_dnlc_evicts_cv; 294static boolean_t arc_dnlc_evicts_thread_exit; 295 296uint_t arc_reduce_dnlc_percent = 3; 297 298/* 299 * The number of headers to evict in arc_evict_state_impl() before 300 * dropping the sublist lock and evicting from another sublist. A lower 301 * value means we're more likely to evict the "correct" header (i.e. the 302 * oldest header in the arc state), but comes with higher overhead 303 * (i.e. more invocations of arc_evict_state_impl()). 304 */ 305int zfs_arc_evict_batch_limit = 10; 306 307/* number of seconds before growing cache again */ 308static int arc_grow_retry = 60; 309 310/* shift of arc_c for calculating overflow limit in arc_get_data_buf */ 311int zfs_arc_overflow_shift = 8; 312 313/* shift of arc_c for calculating both min and max arc_p */ 314static int arc_p_min_shift = 4; 315 316/* log2(fraction of arc to reclaim) */ 317static int arc_shrink_shift = 7; 318 319/* 320 * log2(fraction of ARC which must be free to allow growing). 321 * I.e. If there is less than arc_c >> arc_no_grow_shift free memory, 322 * when reading a new block into the ARC, we will evict an equal-sized block 323 * from the ARC. 324 * 325 * This must be less than arc_shrink_shift, so that when we shrink the ARC, 326 * we will still not allow it to grow. 327 */ 328int arc_no_grow_shift = 5; 329 330 331/* 332 * minimum lifespan of a prefetch block in clock ticks 333 * (initialized in arc_init()) 334 */ 335static int arc_min_prefetch_lifespan; 336 337/* 338 * If this percent of memory is free, don't throttle. 339 */ 340int arc_lotsfree_percent = 10; 341 342static int arc_dead; 343extern boolean_t zfs_prefetch_disable; 344 345/* 346 * The arc has filled available memory and has now warmed up. 347 */ 348static boolean_t arc_warm; 349 350/* 351 * These tunables are for performance analysis. 352 */ 353uint64_t zfs_arc_max; 354uint64_t zfs_arc_min; 355uint64_t zfs_arc_meta_limit = 0; 356uint64_t zfs_arc_meta_min = 0; 357int zfs_arc_grow_retry = 0; 358int zfs_arc_shrink_shift = 0; 359int zfs_arc_p_min_shift = 0; 360uint64_t zfs_arc_average_blocksize = 8 * 1024; /* 8KB */ 361u_int zfs_arc_free_target = 0; 362 363/* Absolute min for arc min / max is 16MB. */ 364static uint64_t arc_abs_min = 16 << 20; 365 366boolean_t zfs_compressed_arc_enabled = B_TRUE; 367 368static int sysctl_vfs_zfs_arc_free_target(SYSCTL_HANDLER_ARGS); 369static int sysctl_vfs_zfs_arc_meta_limit(SYSCTL_HANDLER_ARGS); 370static int sysctl_vfs_zfs_arc_max(SYSCTL_HANDLER_ARGS); 371static int sysctl_vfs_zfs_arc_min(SYSCTL_HANDLER_ARGS); 372 373#if defined(__FreeBSD__) && defined(_KERNEL) 374static void 375arc_free_target_init(void *unused __unused) 376{ 377 378 zfs_arc_free_target = vm_pageout_wakeup_thresh; 379} 380SYSINIT(arc_free_target_init, SI_SUB_KTHREAD_PAGE, SI_ORDER_ANY, 381 arc_free_target_init, NULL); 382 383TUNABLE_QUAD("vfs.zfs.arc_meta_limit", &zfs_arc_meta_limit); 384TUNABLE_QUAD("vfs.zfs.arc_meta_min", &zfs_arc_meta_min); 385TUNABLE_INT("vfs.zfs.arc_shrink_shift", &zfs_arc_shrink_shift); 386SYSCTL_DECL(_vfs_zfs); 387SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_max, CTLTYPE_U64 | CTLFLAG_RWTUN, 388 0, sizeof(uint64_t), sysctl_vfs_zfs_arc_max, "QU", "Maximum ARC size"); 389SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_min, CTLTYPE_U64 | CTLFLAG_RWTUN, 390 0, sizeof(uint64_t), sysctl_vfs_zfs_arc_min, "QU", "Minimum ARC size"); 391SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_average_blocksize, CTLFLAG_RDTUN, 392 &zfs_arc_average_blocksize, 0, 393 "ARC average blocksize"); 394SYSCTL_INT(_vfs_zfs, OID_AUTO, arc_shrink_shift, CTLFLAG_RW, 395 &arc_shrink_shift, 0, 396 "log2(fraction of arc to reclaim)"); 397SYSCTL_INT(_vfs_zfs, OID_AUTO, compressed_arc_enabled, CTLFLAG_RDTUN, 398 &zfs_compressed_arc_enabled, 0, "Enable compressed ARC"); 399 400/* 401 * We don't have a tunable for arc_free_target due to the dependency on 402 * pagedaemon initialisation. 403 */ 404SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_free_target, 405 CTLTYPE_UINT | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, sizeof(u_int), 406 sysctl_vfs_zfs_arc_free_target, "IU", 407 "Desired number of free pages below which ARC triggers reclaim"); 408 409static int 410sysctl_vfs_zfs_arc_free_target(SYSCTL_HANDLER_ARGS) 411{ 412 u_int val; 413 int err; 414 415 val = zfs_arc_free_target; 416 err = sysctl_handle_int(oidp, &val, 0, req); 417 if (err != 0 || req->newptr == NULL) 418 return (err); 419 420 if (val < minfree) 421 return (EINVAL); 422 if (val > vm_cnt.v_page_count) 423 return (EINVAL); 424 425 zfs_arc_free_target = val; 426 427 return (0); 428} 429 430/* 431 * Must be declared here, before the definition of corresponding kstat 432 * macro which uses the same names will confuse the compiler. 433 */ 434SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_meta_limit, 435 CTLTYPE_U64 | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, sizeof(uint64_t), 436 sysctl_vfs_zfs_arc_meta_limit, "QU", 437 "ARC metadata limit"); 438#endif 439 440/* 441 * Note that buffers can be in one of 6 states: 442 * ARC_anon - anonymous (discussed below) 443 * ARC_mru - recently used, currently cached 444 * ARC_mru_ghost - recentely used, no longer in cache 445 * ARC_mfu - frequently used, currently cached 446 * ARC_mfu_ghost - frequently used, no longer in cache 447 * ARC_l2c_only - exists in L2ARC but not other states 448 * When there are no active references to the buffer, they are 449 * are linked onto a list in one of these arc states. These are 450 * the only buffers that can be evicted or deleted. Within each 451 * state there are multiple lists, one for meta-data and one for 452 * non-meta-data. Meta-data (indirect blocks, blocks of dnodes, 453 * etc.) is tracked separately so that it can be managed more 454 * explicitly: favored over data, limited explicitly. 455 * 456 * Anonymous buffers are buffers that are not associated with 457 * a DVA. These are buffers that hold dirty block copies 458 * before they are written to stable storage. By definition, 459 * they are "ref'd" and are considered part of arc_mru 460 * that cannot be freed. Generally, they will aquire a DVA 461 * as they are written and migrate onto the arc_mru list. 462 * 463 * The ARC_l2c_only state is for buffers that are in the second 464 * level ARC but no longer in any of the ARC_m* lists. The second 465 * level ARC itself may also contain buffers that are in any of 466 * the ARC_m* states - meaning that a buffer can exist in two 467 * places. The reason for the ARC_l2c_only state is to keep the 468 * buffer header in the hash table, so that reads that hit the 469 * second level ARC benefit from these fast lookups. 470 */ 471 472typedef struct arc_state { 473 /* 474 * list of evictable buffers 475 */ 476 multilist_t *arcs_list[ARC_BUFC_NUMTYPES]; 477 /* 478 * total amount of evictable data in this state 479 */ 480 refcount_t arcs_esize[ARC_BUFC_NUMTYPES]; 481 /* 482 * total amount of data in this state; this includes: evictable, 483 * non-evictable, ARC_BUFC_DATA, and ARC_BUFC_METADATA. 484 */ 485 refcount_t arcs_size; 486} arc_state_t; 487 488/* The 6 states: */ 489static arc_state_t ARC_anon; 490static arc_state_t ARC_mru; 491static arc_state_t ARC_mru_ghost; 492static arc_state_t ARC_mfu; 493static arc_state_t ARC_mfu_ghost; 494static arc_state_t ARC_l2c_only; 495 496typedef struct arc_stats { 497 kstat_named_t arcstat_hits; 498 kstat_named_t arcstat_misses; 499 kstat_named_t arcstat_demand_data_hits; 500 kstat_named_t arcstat_demand_data_misses; 501 kstat_named_t arcstat_demand_metadata_hits; 502 kstat_named_t arcstat_demand_metadata_misses; 503 kstat_named_t arcstat_prefetch_data_hits; 504 kstat_named_t arcstat_prefetch_data_misses; 505 kstat_named_t arcstat_prefetch_metadata_hits; 506 kstat_named_t arcstat_prefetch_metadata_misses; 507 kstat_named_t arcstat_mru_hits; 508 kstat_named_t arcstat_mru_ghost_hits; 509 kstat_named_t arcstat_mfu_hits; 510 kstat_named_t arcstat_mfu_ghost_hits; 511 kstat_named_t arcstat_allocated; 512 kstat_named_t arcstat_deleted; 513 /* 514 * Number of buffers that could not be evicted because the hash lock 515 * was held by another thread. The lock may not necessarily be held 516 * by something using the same buffer, since hash locks are shared 517 * by multiple buffers. 518 */ 519 kstat_named_t arcstat_mutex_miss; 520 /* 521 * Number of buffers skipped because they have I/O in progress, are 522 * indrect prefetch buffers that have not lived long enough, or are 523 * not from the spa we're trying to evict from. 524 */ 525 kstat_named_t arcstat_evict_skip; 526 /* 527 * Number of times arc_evict_state() was unable to evict enough 528 * buffers to reach it's target amount. 529 */ 530 kstat_named_t arcstat_evict_not_enough; 531 kstat_named_t arcstat_evict_l2_cached; 532 kstat_named_t arcstat_evict_l2_eligible; 533 kstat_named_t arcstat_evict_l2_ineligible; 534 kstat_named_t arcstat_evict_l2_skip; 535 kstat_named_t arcstat_hash_elements; 536 kstat_named_t arcstat_hash_elements_max; 537 kstat_named_t arcstat_hash_collisions; 538 kstat_named_t arcstat_hash_chains; 539 kstat_named_t arcstat_hash_chain_max; 540 kstat_named_t arcstat_p; 541 kstat_named_t arcstat_c; 542 kstat_named_t arcstat_c_min; 543 kstat_named_t arcstat_c_max; 544 kstat_named_t arcstat_size; 545 /* 546 * Number of compressed bytes stored in the arc_buf_hdr_t's b_pdata. 547 * Note that the compressed bytes may match the uncompressed bytes 548 * if the block is either not compressed or compressed arc is disabled. 549 */ 550 kstat_named_t arcstat_compressed_size; 551 /* 552 * Uncompressed size of the data stored in b_pdata. If compressed 553 * arc is disabled then this value will be identical to the stat 554 * above. 555 */ 556 kstat_named_t arcstat_uncompressed_size; 557 /* 558 * Number of bytes stored in all the arc_buf_t's. This is classified 559 * as "overhead" since this data is typically short-lived and will 560 * be evicted from the arc when it becomes unreferenced unless the 561 * zfs_keep_uncompressed_metadata or zfs_keep_uncompressed_level 562 * values have been set (see comment in dbuf.c for more information). 563 */ 564 kstat_named_t arcstat_overhead_size; 565 /* 566 * Number of bytes consumed by internal ARC structures necessary 567 * for tracking purposes; these structures are not actually 568 * backed by ARC buffers. This includes arc_buf_hdr_t structures 569 * (allocated via arc_buf_hdr_t_full and arc_buf_hdr_t_l2only 570 * caches), and arc_buf_t structures (allocated via arc_buf_t 571 * cache). 572 */ 573 kstat_named_t arcstat_hdr_size; 574 /* 575 * Number of bytes consumed by ARC buffers of type equal to 576 * ARC_BUFC_DATA. This is generally consumed by buffers backing 577 * on disk user data (e.g. plain file contents). 578 */ 579 kstat_named_t arcstat_data_size; 580 /* 581 * Number of bytes consumed by ARC buffers of type equal to 582 * ARC_BUFC_METADATA. This is generally consumed by buffers 583 * backing on disk data that is used for internal ZFS 584 * structures (e.g. ZAP, dnode, indirect blocks, etc). 585 */ 586 kstat_named_t arcstat_metadata_size; 587 /* 588 * Number of bytes consumed by various buffers and structures 589 * not actually backed with ARC buffers. This includes bonus 590 * buffers (allocated directly via zio_buf_* functions), 591 * dmu_buf_impl_t structures (allocated via dmu_buf_impl_t 592 * cache), and dnode_t structures (allocated via dnode_t cache). 593 */ 594 kstat_named_t arcstat_other_size; 595 /* 596 * Total number of bytes consumed by ARC buffers residing in the 597 * arc_anon state. This includes *all* buffers in the arc_anon 598 * state; e.g. data, metadata, evictable, and unevictable buffers 599 * are all included in this value. 600 */ 601 kstat_named_t arcstat_anon_size; 602 /* 603 * Number of bytes consumed by ARC buffers that meet the 604 * following criteria: backing buffers of type ARC_BUFC_DATA, 605 * residing in the arc_anon state, and are eligible for eviction 606 * (e.g. have no outstanding holds on the buffer). 607 */ 608 kstat_named_t arcstat_anon_evictable_data; 609 /* 610 * Number of bytes consumed by ARC buffers that meet the 611 * following criteria: backing buffers of type ARC_BUFC_METADATA, 612 * residing in the arc_anon state, and are eligible for eviction 613 * (e.g. have no outstanding holds on the buffer). 614 */ 615 kstat_named_t arcstat_anon_evictable_metadata; 616 /* 617 * Total number of bytes consumed by ARC buffers residing in the 618 * arc_mru state. This includes *all* buffers in the arc_mru 619 * state; e.g. data, metadata, evictable, and unevictable buffers 620 * are all included in this value. 621 */ 622 kstat_named_t arcstat_mru_size; 623 /* 624 * Number of bytes consumed by ARC buffers that meet the 625 * following criteria: backing buffers of type ARC_BUFC_DATA, 626 * residing in the arc_mru state, and are eligible for eviction 627 * (e.g. have no outstanding holds on the buffer). 628 */ 629 kstat_named_t arcstat_mru_evictable_data; 630 /* 631 * Number of bytes consumed by ARC buffers that meet the 632 * following criteria: backing buffers of type ARC_BUFC_METADATA, 633 * residing in the arc_mru state, and are eligible for eviction 634 * (e.g. have no outstanding holds on the buffer). 635 */ 636 kstat_named_t arcstat_mru_evictable_metadata; 637 /* 638 * Total number of bytes that *would have been* consumed by ARC 639 * buffers in the arc_mru_ghost state. The key thing to note 640 * here, is the fact that this size doesn't actually indicate 641 * RAM consumption. The ghost lists only consist of headers and 642 * don't actually have ARC buffers linked off of these headers. 643 * Thus, *if* the headers had associated ARC buffers, these 644 * buffers *would have* consumed this number of bytes. 645 */ 646 kstat_named_t arcstat_mru_ghost_size; 647 /* 648 * Number of bytes that *would have been* consumed by ARC 649 * buffers that are eligible for eviction, of type 650 * ARC_BUFC_DATA, and linked off the arc_mru_ghost state. 651 */ 652 kstat_named_t arcstat_mru_ghost_evictable_data; 653 /* 654 * Number of bytes that *would have been* consumed by ARC 655 * buffers that are eligible for eviction, of type 656 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state. 657 */ 658 kstat_named_t arcstat_mru_ghost_evictable_metadata; 659 /* 660 * Total number of bytes consumed by ARC buffers residing in the 661 * arc_mfu state. This includes *all* buffers in the arc_mfu 662 * state; e.g. data, metadata, evictable, and unevictable buffers 663 * are all included in this value. 664 */ 665 kstat_named_t arcstat_mfu_size; 666 /* 667 * Number of bytes consumed by ARC buffers that are eligible for 668 * eviction, of type ARC_BUFC_DATA, and reside in the arc_mfu 669 * state. 670 */ 671 kstat_named_t arcstat_mfu_evictable_data; 672 /* 673 * Number of bytes consumed by ARC buffers that are eligible for 674 * eviction, of type ARC_BUFC_METADATA, and reside in the 675 * arc_mfu state. 676 */ 677 kstat_named_t arcstat_mfu_evictable_metadata; 678 /* 679 * Total number of bytes that *would have been* consumed by ARC 680 * buffers in the arc_mfu_ghost state. See the comment above 681 * arcstat_mru_ghost_size for more details. 682 */ 683 kstat_named_t arcstat_mfu_ghost_size; 684 /* 685 * Number of bytes that *would have been* consumed by ARC 686 * buffers that are eligible for eviction, of type 687 * ARC_BUFC_DATA, and linked off the arc_mfu_ghost state. 688 */ 689 kstat_named_t arcstat_mfu_ghost_evictable_data; 690 /* 691 * Number of bytes that *would have been* consumed by ARC 692 * buffers that are eligible for eviction, of type 693 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state. 694 */ 695 kstat_named_t arcstat_mfu_ghost_evictable_metadata; 696 kstat_named_t arcstat_l2_hits; 697 kstat_named_t arcstat_l2_misses; 698 kstat_named_t arcstat_l2_feeds; 699 kstat_named_t arcstat_l2_rw_clash; 700 kstat_named_t arcstat_l2_read_bytes; 701 kstat_named_t arcstat_l2_write_bytes; 702 kstat_named_t arcstat_l2_writes_sent; 703 kstat_named_t arcstat_l2_writes_done; 704 kstat_named_t arcstat_l2_writes_error; 705 kstat_named_t arcstat_l2_writes_lock_retry; 706 kstat_named_t arcstat_l2_evict_lock_retry; 707 kstat_named_t arcstat_l2_evict_reading; 708 kstat_named_t arcstat_l2_evict_l1cached; 709 kstat_named_t arcstat_l2_free_on_write; 710 kstat_named_t arcstat_l2_abort_lowmem; 711 kstat_named_t arcstat_l2_cksum_bad; 712 kstat_named_t arcstat_l2_io_error; 713 kstat_named_t arcstat_l2_size; 714 kstat_named_t arcstat_l2_asize; 715 kstat_named_t arcstat_l2_hdr_size; 716 kstat_named_t arcstat_l2_write_trylock_fail; 717 kstat_named_t arcstat_l2_write_passed_headroom; 718 kstat_named_t arcstat_l2_write_spa_mismatch; 719 kstat_named_t arcstat_l2_write_in_l2; 720 kstat_named_t arcstat_l2_write_hdr_io_in_progress; 721 kstat_named_t arcstat_l2_write_not_cacheable; 722 kstat_named_t arcstat_l2_write_full; 723 kstat_named_t arcstat_l2_write_buffer_iter; 724 kstat_named_t arcstat_l2_write_pios; 725 kstat_named_t arcstat_l2_write_buffer_bytes_scanned; 726 kstat_named_t arcstat_l2_write_buffer_list_iter; 727 kstat_named_t arcstat_l2_write_buffer_list_null_iter; 728 kstat_named_t arcstat_memory_throttle_count; 729 kstat_named_t arcstat_meta_used; 730 kstat_named_t arcstat_meta_limit; 731 kstat_named_t arcstat_meta_max; 732 kstat_named_t arcstat_meta_min; 733 kstat_named_t arcstat_sync_wait_for_async; 734 kstat_named_t arcstat_demand_hit_predictive_prefetch; 735} arc_stats_t; 736 737static arc_stats_t arc_stats = { 738 { "hits", KSTAT_DATA_UINT64 }, 739 { "misses", KSTAT_DATA_UINT64 }, 740 { "demand_data_hits", KSTAT_DATA_UINT64 }, 741 { "demand_data_misses", KSTAT_DATA_UINT64 }, 742 { "demand_metadata_hits", KSTAT_DATA_UINT64 }, 743 { "demand_metadata_misses", KSTAT_DATA_UINT64 }, 744 { "prefetch_data_hits", KSTAT_DATA_UINT64 }, 745 { "prefetch_data_misses", KSTAT_DATA_UINT64 }, 746 { "prefetch_metadata_hits", KSTAT_DATA_UINT64 }, 747 { "prefetch_metadata_misses", KSTAT_DATA_UINT64 }, 748 { "mru_hits", KSTAT_DATA_UINT64 }, 749 { "mru_ghost_hits", KSTAT_DATA_UINT64 }, 750 { "mfu_hits", KSTAT_DATA_UINT64 }, 751 { "mfu_ghost_hits", KSTAT_DATA_UINT64 }, 752 { "allocated", KSTAT_DATA_UINT64 }, 753 { "deleted", KSTAT_DATA_UINT64 }, 754 { "mutex_miss", KSTAT_DATA_UINT64 }, 755 { "evict_skip", KSTAT_DATA_UINT64 }, 756 { "evict_not_enough", KSTAT_DATA_UINT64 }, 757 { "evict_l2_cached", KSTAT_DATA_UINT64 }, 758 { "evict_l2_eligible", KSTAT_DATA_UINT64 }, 759 { "evict_l2_ineligible", KSTAT_DATA_UINT64 }, 760 { "evict_l2_skip", KSTAT_DATA_UINT64 }, 761 { "hash_elements", KSTAT_DATA_UINT64 }, 762 { "hash_elements_max", KSTAT_DATA_UINT64 }, 763 { "hash_collisions", KSTAT_DATA_UINT64 }, 764 { "hash_chains", KSTAT_DATA_UINT64 }, 765 { "hash_chain_max", KSTAT_DATA_UINT64 }, 766 { "p", KSTAT_DATA_UINT64 }, 767 { "c", KSTAT_DATA_UINT64 }, 768 { "c_min", KSTAT_DATA_UINT64 }, 769 { "c_max", KSTAT_DATA_UINT64 }, 770 { "size", KSTAT_DATA_UINT64 }, 771 { "compressed_size", KSTAT_DATA_UINT64 }, 772 { "uncompressed_size", KSTAT_DATA_UINT64 }, 773 { "overhead_size", KSTAT_DATA_UINT64 }, 774 { "hdr_size", KSTAT_DATA_UINT64 }, 775 { "data_size", KSTAT_DATA_UINT64 }, 776 { "metadata_size", KSTAT_DATA_UINT64 }, 777 { "other_size", KSTAT_DATA_UINT64 }, 778 { "anon_size", KSTAT_DATA_UINT64 }, 779 { "anon_evictable_data", KSTAT_DATA_UINT64 }, 780 { "anon_evictable_metadata", KSTAT_DATA_UINT64 }, 781 { "mru_size", KSTAT_DATA_UINT64 }, 782 { "mru_evictable_data", KSTAT_DATA_UINT64 }, 783 { "mru_evictable_metadata", KSTAT_DATA_UINT64 }, 784 { "mru_ghost_size", KSTAT_DATA_UINT64 }, 785 { "mru_ghost_evictable_data", KSTAT_DATA_UINT64 }, 786 { "mru_ghost_evictable_metadata", KSTAT_DATA_UINT64 }, 787 { "mfu_size", KSTAT_DATA_UINT64 }, 788 { "mfu_evictable_data", KSTAT_DATA_UINT64 }, 789 { "mfu_evictable_metadata", KSTAT_DATA_UINT64 }, 790 { "mfu_ghost_size", KSTAT_DATA_UINT64 }, 791 { "mfu_ghost_evictable_data", KSTAT_DATA_UINT64 }, 792 { "mfu_ghost_evictable_metadata", KSTAT_DATA_UINT64 }, 793 { "l2_hits", KSTAT_DATA_UINT64 }, 794 { "l2_misses", KSTAT_DATA_UINT64 }, 795 { "l2_feeds", KSTAT_DATA_UINT64 }, 796 { "l2_rw_clash", KSTAT_DATA_UINT64 }, 797 { "l2_read_bytes", KSTAT_DATA_UINT64 }, 798 { "l2_write_bytes", KSTAT_DATA_UINT64 }, 799 { "l2_writes_sent", KSTAT_DATA_UINT64 }, 800 { "l2_writes_done", KSTAT_DATA_UINT64 }, 801 { "l2_writes_error", KSTAT_DATA_UINT64 }, 802 { "l2_writes_lock_retry", KSTAT_DATA_UINT64 }, 803 { "l2_evict_lock_retry", KSTAT_DATA_UINT64 }, 804 { "l2_evict_reading", KSTAT_DATA_UINT64 }, 805 { "l2_evict_l1cached", KSTAT_DATA_UINT64 }, 806 { "l2_free_on_write", KSTAT_DATA_UINT64 }, 807 { "l2_abort_lowmem", KSTAT_DATA_UINT64 }, 808 { "l2_cksum_bad", KSTAT_DATA_UINT64 }, 809 { "l2_io_error", KSTAT_DATA_UINT64 }, 810 { "l2_size", KSTAT_DATA_UINT64 }, 811 { "l2_asize", KSTAT_DATA_UINT64 }, 812 { "l2_hdr_size", KSTAT_DATA_UINT64 }, 813 { "l2_write_trylock_fail", KSTAT_DATA_UINT64 }, 814 { "l2_write_passed_headroom", KSTAT_DATA_UINT64 }, 815 { "l2_write_spa_mismatch", KSTAT_DATA_UINT64 }, 816 { "l2_write_in_l2", KSTAT_DATA_UINT64 }, 817 { "l2_write_io_in_progress", KSTAT_DATA_UINT64 }, 818 { "l2_write_not_cacheable", KSTAT_DATA_UINT64 }, 819 { "l2_write_full", KSTAT_DATA_UINT64 }, 820 { "l2_write_buffer_iter", KSTAT_DATA_UINT64 }, 821 { "l2_write_pios", KSTAT_DATA_UINT64 }, 822 { "l2_write_buffer_bytes_scanned", KSTAT_DATA_UINT64 }, 823 { "l2_write_buffer_list_iter", KSTAT_DATA_UINT64 }, 824 { "l2_write_buffer_list_null_iter", KSTAT_DATA_UINT64 }, 825 { "memory_throttle_count", KSTAT_DATA_UINT64 }, 826 { "arc_meta_used", KSTAT_DATA_UINT64 }, 827 { "arc_meta_limit", KSTAT_DATA_UINT64 }, 828 { "arc_meta_max", KSTAT_DATA_UINT64 }, 829 { "arc_meta_min", KSTAT_DATA_UINT64 }, 830 { "sync_wait_for_async", KSTAT_DATA_UINT64 }, 831 { "demand_hit_predictive_prefetch", KSTAT_DATA_UINT64 }, 832}; 833 834#define ARCSTAT(stat) (arc_stats.stat.value.ui64) 835 836#define ARCSTAT_INCR(stat, val) \ 837 atomic_add_64(&arc_stats.stat.value.ui64, (val)) 838 839#define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1) 840#define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1) 841 842#define ARCSTAT_MAX(stat, val) { \ 843 uint64_t m; \ 844 while ((val) > (m = arc_stats.stat.value.ui64) && \ 845 (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \ 846 continue; \ 847} 848 849#define ARCSTAT_MAXSTAT(stat) \ 850 ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64) 851 852/* 853 * We define a macro to allow ARC hits/misses to be easily broken down by 854 * two separate conditions, giving a total of four different subtypes for 855 * each of hits and misses (so eight statistics total). 856 */ 857#define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \ 858 if (cond1) { \ 859 if (cond2) { \ 860 ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \ 861 } else { \ 862 ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \ 863 } \ 864 } else { \ 865 if (cond2) { \ 866 ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \ 867 } else { \ 868 ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\ 869 } \ 870 } 871 872kstat_t *arc_ksp; 873static arc_state_t *arc_anon; 874static arc_state_t *arc_mru; 875static arc_state_t *arc_mru_ghost; 876static arc_state_t *arc_mfu; 877static arc_state_t *arc_mfu_ghost; 878static arc_state_t *arc_l2c_only; 879 880/* 881 * There are several ARC variables that are critical to export as kstats -- 882 * but we don't want to have to grovel around in the kstat whenever we wish to 883 * manipulate them. For these variables, we therefore define them to be in 884 * terms of the statistic variable. This assures that we are not introducing 885 * the possibility of inconsistency by having shadow copies of the variables, 886 * while still allowing the code to be readable. 887 */ 888#define arc_size ARCSTAT(arcstat_size) /* actual total arc size */ 889#define arc_p ARCSTAT(arcstat_p) /* target size of MRU */ 890#define arc_c ARCSTAT(arcstat_c) /* target size of cache */ 891#define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */ 892#define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */ 893#define arc_meta_limit ARCSTAT(arcstat_meta_limit) /* max size for metadata */ 894#define arc_meta_min ARCSTAT(arcstat_meta_min) /* min size for metadata */ 895#define arc_meta_used ARCSTAT(arcstat_meta_used) /* size of metadata */ 896#define arc_meta_max ARCSTAT(arcstat_meta_max) /* max size of metadata */ 897 898/* compressed size of entire arc */ 899#define arc_compressed_size ARCSTAT(arcstat_compressed_size) 900/* uncompressed size of entire arc */ 901#define arc_uncompressed_size ARCSTAT(arcstat_uncompressed_size) 902/* number of bytes in the arc from arc_buf_t's */ 903#define arc_overhead_size ARCSTAT(arcstat_overhead_size) 904 905static int arc_no_grow; /* Don't try to grow cache size */ 906static uint64_t arc_tempreserve; 907static uint64_t arc_loaned_bytes; 908 909typedef struct arc_callback arc_callback_t; 910 911struct arc_callback { 912 void *acb_private; 913 arc_done_func_t *acb_done; 914 arc_buf_t *acb_buf; 915 boolean_t acb_compressed; 916 zio_t *acb_zio_dummy; 917 arc_callback_t *acb_next; 918}; 919 920typedef struct arc_write_callback arc_write_callback_t; 921 922struct arc_write_callback { 923 void *awcb_private; 924 arc_done_func_t *awcb_ready; 925 arc_done_func_t *awcb_children_ready; 926 arc_done_func_t *awcb_physdone; 927 arc_done_func_t *awcb_done; 928 arc_buf_t *awcb_buf; 929}; 930 931/* 932 * ARC buffers are separated into multiple structs as a memory saving measure: 933 * - Common fields struct, always defined, and embedded within it: 934 * - L2-only fields, always allocated but undefined when not in L2ARC 935 * - L1-only fields, only allocated when in L1ARC 936 * 937 * Buffer in L1 Buffer only in L2 938 * +------------------------+ +------------------------+ 939 * | arc_buf_hdr_t | | arc_buf_hdr_t | 940 * | | | | 941 * | | | | 942 * | | | | 943 * +------------------------+ +------------------------+ 944 * | l2arc_buf_hdr_t | | l2arc_buf_hdr_t | 945 * | (undefined if L1-only) | | | 946 * +------------------------+ +------------------------+ 947 * | l1arc_buf_hdr_t | 948 * | | 949 * | | 950 * | | 951 * | | 952 * +------------------------+ 953 * 954 * Because it's possible for the L2ARC to become extremely large, we can wind 955 * up eating a lot of memory in L2ARC buffer headers, so the size of a header 956 * is minimized by only allocating the fields necessary for an L1-cached buffer 957 * when a header is actually in the L1 cache. The sub-headers (l1arc_buf_hdr and 958 * l2arc_buf_hdr) are embedded rather than allocated separately to save a couple 959 * words in pointers. arc_hdr_realloc() is used to switch a header between 960 * these two allocation states. 961 */ 962typedef struct l1arc_buf_hdr { 963 kmutex_t b_freeze_lock; 964 zio_cksum_t *b_freeze_cksum; 965#ifdef ZFS_DEBUG 966 /* 967 * Used for debugging with kmem_flags - by allocating and freeing 968 * b_thawed when the buffer is thawed, we get a record of the stack 969 * trace that thawed it. 970 */ 971 void *b_thawed; 972#endif 973 974 arc_buf_t *b_buf; 975 uint32_t b_bufcnt; 976 /* for waiting on writes to complete */ 977 kcondvar_t b_cv; 978 uint8_t b_byteswap; 979 980 /* protected by arc state mutex */ 981 arc_state_t *b_state; 982 multilist_node_t b_arc_node; 983 984 /* updated atomically */ 985 clock_t b_arc_access; 986 987 /* self protecting */ 988 refcount_t b_refcnt; 989 990 arc_callback_t *b_acb; 991 void *b_pdata; 992} l1arc_buf_hdr_t; 993 994typedef struct l2arc_dev l2arc_dev_t; 995 996typedef struct l2arc_buf_hdr { 997 /* protected by arc_buf_hdr mutex */ 998 l2arc_dev_t *b_dev; /* L2ARC device */ 999 uint64_t b_daddr; /* disk address, offset byte */ 1000 1001 list_node_t b_l2node; 1002} l2arc_buf_hdr_t; 1003 1004struct arc_buf_hdr { 1005 /* protected by hash lock */ 1006 dva_t b_dva; 1007 uint64_t b_birth; 1008 1009 arc_buf_contents_t b_type; 1010 arc_buf_hdr_t *b_hash_next; 1011 arc_flags_t b_flags; 1012 1013 /* 1014 * This field stores the size of the data buffer after 1015 * compression, and is set in the arc's zio completion handlers. 1016 * It is in units of SPA_MINBLOCKSIZE (e.g. 1 == 512 bytes). 1017 * 1018 * While the block pointers can store up to 32MB in their psize 1019 * field, we can only store up to 32MB minus 512B. This is due 1020 * to the bp using a bias of 1, whereas we use a bias of 0 (i.e. 1021 * a field of zeros represents 512B in the bp). We can't use a 1022 * bias of 1 since we need to reserve a psize of zero, here, to 1023 * represent holes and embedded blocks. 1024 * 1025 * This isn't a problem in practice, since the maximum size of a 1026 * buffer is limited to 16MB, so we never need to store 32MB in 1027 * this field. Even in the upstream illumos code base, the 1028 * maximum size of a buffer is limited to 16MB. 1029 */ 1030 uint16_t b_psize; 1031 1032 /* 1033 * This field stores the size of the data buffer before 1034 * compression, and cannot change once set. It is in units 1035 * of SPA_MINBLOCKSIZE (e.g. 2 == 1024 bytes) 1036 */ 1037 uint16_t b_lsize; /* immutable */ 1038 uint64_t b_spa; /* immutable */ 1039 1040 /* L2ARC fields. Undefined when not in L2ARC. */ 1041 l2arc_buf_hdr_t b_l2hdr; 1042 /* L1ARC fields. Undefined when in l2arc_only state */ 1043 l1arc_buf_hdr_t b_l1hdr; 1044}; 1045 1046#if defined(__FreeBSD__) && defined(_KERNEL) 1047static int 1048sysctl_vfs_zfs_arc_meta_limit(SYSCTL_HANDLER_ARGS) 1049{ 1050 uint64_t val; 1051 int err; 1052 1053 val = arc_meta_limit; 1054 err = sysctl_handle_64(oidp, &val, 0, req); 1055 if (err != 0 || req->newptr == NULL) 1056 return (err); 1057 1058 if (val <= 0 || val > arc_c_max) 1059 return (EINVAL); 1060 1061 arc_meta_limit = val; 1062 return (0); 1063} 1064 1065static int 1066sysctl_vfs_zfs_arc_max(SYSCTL_HANDLER_ARGS) 1067{ 1068 uint64_t val; 1069 int err; 1070 1071 val = zfs_arc_max; 1072 err = sysctl_handle_64(oidp, &val, 0, req); 1073 if (err != 0 || req->newptr == NULL) 1074 return (err); 1075 1076 if (zfs_arc_max == 0) { 1077 /* Loader tunable so blindly set */ 1078 zfs_arc_max = val; 1079 return (0); 1080 } 1081 1082 if (val < arc_abs_min || val > kmem_size()) 1083 return (EINVAL); 1084 if (val < arc_c_min) 1085 return (EINVAL); 1086 if (zfs_arc_meta_limit > 0 && val < zfs_arc_meta_limit) 1087 return (EINVAL); 1088 1089 arc_c_max = val; 1090 1091 arc_c = arc_c_max; 1092 arc_p = (arc_c >> 1); 1093 1094 if (zfs_arc_meta_limit == 0) { 1095 /* limit meta-data to 1/4 of the arc capacity */ 1096 arc_meta_limit = arc_c_max / 4; 1097 } 1098 1099 /* if kmem_flags are set, lets try to use less memory */ 1100 if (kmem_debugging()) 1101 arc_c = arc_c / 2; 1102 1103 zfs_arc_max = arc_c; 1104 1105 return (0); 1106} 1107 1108static int 1109sysctl_vfs_zfs_arc_min(SYSCTL_HANDLER_ARGS) 1110{ 1111 uint64_t val; 1112 int err; 1113 1114 val = zfs_arc_min; 1115 err = sysctl_handle_64(oidp, &val, 0, req); 1116 if (err != 0 || req->newptr == NULL) 1117 return (err); 1118 1119 if (zfs_arc_min == 0) { 1120 /* Loader tunable so blindly set */ 1121 zfs_arc_min = val; 1122 return (0); 1123 } 1124 1125 if (val < arc_abs_min || val > arc_c_max) 1126 return (EINVAL); 1127 1128 arc_c_min = val; 1129 1130 if (zfs_arc_meta_min == 0) 1131 arc_meta_min = arc_c_min / 2; 1132 1133 if (arc_c < arc_c_min) 1134 arc_c = arc_c_min; 1135 1136 zfs_arc_min = arc_c_min; 1137 1138 return (0); 1139} 1140#endif 1141 1142#define GHOST_STATE(state) \ 1143 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \ 1144 (state) == arc_l2c_only) 1145 1146#define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_FLAG_IN_HASH_TABLE) 1147#define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) 1148#define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_FLAG_IO_ERROR) 1149#define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_FLAG_PREFETCH) 1150#define HDR_COMPRESSION_ENABLED(hdr) \ 1151 ((hdr)->b_flags & ARC_FLAG_COMPRESSED_ARC) 1152 1153#define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_FLAG_L2CACHE) 1154#define HDR_L2_READING(hdr) \ 1155 (((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) && \ 1156 ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR)) 1157#define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITING) 1158#define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_FLAG_L2_EVICTED) 1159#define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITE_HEAD) 1160#define HDR_SHARED_DATA(hdr) ((hdr)->b_flags & ARC_FLAG_SHARED_DATA) 1161 1162#define HDR_ISTYPE_METADATA(hdr) \ 1163 ((hdr)->b_flags & ARC_FLAG_BUFC_METADATA) 1164#define HDR_ISTYPE_DATA(hdr) (!HDR_ISTYPE_METADATA(hdr)) 1165 1166#define HDR_HAS_L1HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L1HDR) 1167#define HDR_HAS_L2HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR) 1168 1169/* For storing compression mode in b_flags */ 1170#define HDR_COMPRESS_OFFSET (highbit64(ARC_FLAG_COMPRESS_0) - 1) 1171 1172#define HDR_GET_COMPRESS(hdr) ((enum zio_compress)BF32_GET((hdr)->b_flags, \ 1173 HDR_COMPRESS_OFFSET, SPA_COMPRESSBITS)) 1174#define HDR_SET_COMPRESS(hdr, cmp) BF32_SET((hdr)->b_flags, \ 1175 HDR_COMPRESS_OFFSET, SPA_COMPRESSBITS, (cmp)); 1176 1177#define ARC_BUF_LAST(buf) ((buf)->b_next == NULL) 1178#define ARC_BUF_SHARED(buf) ((buf)->b_flags & ARC_BUF_FLAG_SHARED) 1179#define ARC_BUF_COMPRESSED(buf) ((buf)->b_flags & ARC_BUF_FLAG_COMPRESSED) 1180 1181/* 1182 * Other sizes 1183 */ 1184 1185#define HDR_FULL_SIZE ((int64_t)sizeof (arc_buf_hdr_t)) 1186#define HDR_L2ONLY_SIZE ((int64_t)offsetof(arc_buf_hdr_t, b_l1hdr)) 1187 1188/* 1189 * Hash table routines 1190 */ 1191 1192#define HT_LOCK_PAD CACHE_LINE_SIZE 1193 1194struct ht_lock { 1195 kmutex_t ht_lock; 1196#ifdef _KERNEL 1197 unsigned char pad[(HT_LOCK_PAD - sizeof (kmutex_t))]; 1198#endif 1199}; 1200 1201#define BUF_LOCKS 256 1202typedef struct buf_hash_table { 1203 uint64_t ht_mask; 1204 arc_buf_hdr_t **ht_table; 1205 struct ht_lock ht_locks[BUF_LOCKS] __aligned(CACHE_LINE_SIZE); 1206} buf_hash_table_t; 1207 1208static buf_hash_table_t buf_hash_table; 1209 1210#define BUF_HASH_INDEX(spa, dva, birth) \ 1211 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask) 1212#define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)]) 1213#define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock)) 1214#define HDR_LOCK(hdr) \ 1215 (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth))) 1216 1217uint64_t zfs_crc64_table[256]; 1218 1219/* 1220 * Level 2 ARC 1221 */ 1222 1223#define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */ 1224#define L2ARC_HEADROOM 2 /* num of writes */ 1225/* 1226 * If we discover during ARC scan any buffers to be compressed, we boost 1227 * our headroom for the next scanning cycle by this percentage multiple. 1228 */ 1229#define L2ARC_HEADROOM_BOOST 200 1230#define L2ARC_FEED_SECS 1 /* caching interval secs */ 1231#define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */ 1232 1233#define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent) 1234#define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done) 1235 1236/* L2ARC Performance Tunables */ 1237uint64_t l2arc_write_max = L2ARC_WRITE_SIZE; /* default max write size */ 1238uint64_t l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra write during warmup */ 1239uint64_t l2arc_headroom = L2ARC_HEADROOM; /* number of dev writes */ 1240uint64_t l2arc_headroom_boost = L2ARC_HEADROOM_BOOST; 1241uint64_t l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */ 1242uint64_t l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval milliseconds */ 1243boolean_t l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */ 1244boolean_t l2arc_feed_again = B_TRUE; /* turbo warmup */ 1245boolean_t l2arc_norw = B_TRUE; /* no reads during writes */ 1246 1247SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_max, CTLFLAG_RW, 1248 &l2arc_write_max, 0, "max write size"); 1249SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_boost, CTLFLAG_RW, 1250 &l2arc_write_boost, 0, "extra write during warmup"); 1251SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_headroom, CTLFLAG_RW, 1252 &l2arc_headroom, 0, "number of dev writes"); 1253SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_secs, CTLFLAG_RW, 1254 &l2arc_feed_secs, 0, "interval seconds"); 1255SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_min_ms, CTLFLAG_RW, 1256 &l2arc_feed_min_ms, 0, "min interval milliseconds"); 1257 1258SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_noprefetch, CTLFLAG_RW, 1259 &l2arc_noprefetch, 0, "don't cache prefetch bufs"); 1260SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_feed_again, CTLFLAG_RW, 1261 &l2arc_feed_again, 0, "turbo warmup"); 1262SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_norw, CTLFLAG_RW, 1263 &l2arc_norw, 0, "no reads during writes"); 1264 1265SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_size, CTLFLAG_RD, 1266 &ARC_anon.arcs_size.rc_count, 0, "size of anonymous state"); 1267SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_metadata_esize, CTLFLAG_RD, 1268 &ARC_anon.arcs_esize[ARC_BUFC_METADATA].rc_count, 0, 1269 "size of anonymous state"); 1270SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_data_esize, CTLFLAG_RD, 1271 &ARC_anon.arcs_esize[ARC_BUFC_DATA].rc_count, 0, 1272 "size of anonymous state"); 1273 1274SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_size, CTLFLAG_RD, 1275 &ARC_mru.arcs_size.rc_count, 0, "size of mru state"); 1276SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_metadata_esize, CTLFLAG_RD, 1277 &ARC_mru.arcs_esize[ARC_BUFC_METADATA].rc_count, 0, 1278 "size of metadata in mru state"); 1279SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_data_esize, CTLFLAG_RD, 1280 &ARC_mru.arcs_esize[ARC_BUFC_DATA].rc_count, 0, 1281 "size of data in mru state"); 1282 1283SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_size, CTLFLAG_RD, 1284 &ARC_mru_ghost.arcs_size.rc_count, 0, "size of mru ghost state"); 1285SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_metadata_esize, CTLFLAG_RD, 1286 &ARC_mru_ghost.arcs_esize[ARC_BUFC_METADATA].rc_count, 0, 1287 "size of metadata in mru ghost state"); 1288SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_data_esize, CTLFLAG_RD, 1289 &ARC_mru_ghost.arcs_esize[ARC_BUFC_DATA].rc_count, 0, 1290 "size of data in mru ghost state"); 1291 1292SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_size, CTLFLAG_RD, 1293 &ARC_mfu.arcs_size.rc_count, 0, "size of mfu state"); 1294SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_metadata_esize, CTLFLAG_RD, 1295 &ARC_mfu.arcs_esize[ARC_BUFC_METADATA].rc_count, 0, 1296 "size of metadata in mfu state"); 1297SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_data_esize, CTLFLAG_RD, 1298 &ARC_mfu.arcs_esize[ARC_BUFC_DATA].rc_count, 0, 1299 "size of data in mfu state"); 1300 1301SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_size, CTLFLAG_RD, 1302 &ARC_mfu_ghost.arcs_size.rc_count, 0, "size of mfu ghost state"); 1303SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_metadata_esize, CTLFLAG_RD, 1304 &ARC_mfu_ghost.arcs_esize[ARC_BUFC_METADATA].rc_count, 0, 1305 "size of metadata in mfu ghost state"); 1306SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_data_esize, CTLFLAG_RD, 1307 &ARC_mfu_ghost.arcs_esize[ARC_BUFC_DATA].rc_count, 0, 1308 "size of data in mfu ghost state"); 1309 1310SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2c_only_size, CTLFLAG_RD, 1311 &ARC_l2c_only.arcs_size.rc_count, 0, "size of mru state"); 1312 1313/* 1314 * L2ARC Internals 1315 */ 1316struct l2arc_dev { 1317 vdev_t *l2ad_vdev; /* vdev */ 1318 spa_t *l2ad_spa; /* spa */ 1319 uint64_t l2ad_hand; /* next write location */ 1320 uint64_t l2ad_start; /* first addr on device */ 1321 uint64_t l2ad_end; /* last addr on device */ 1322 boolean_t l2ad_first; /* first sweep through */ 1323 boolean_t l2ad_writing; /* currently writing */ 1324 kmutex_t l2ad_mtx; /* lock for buffer list */ 1325 list_t l2ad_buflist; /* buffer list */ 1326 list_node_t l2ad_node; /* device list node */ 1327 refcount_t l2ad_alloc; /* allocated bytes */ 1328}; 1329 1330static list_t L2ARC_dev_list; /* device list */ 1331static list_t *l2arc_dev_list; /* device list pointer */ 1332static kmutex_t l2arc_dev_mtx; /* device list mutex */ 1333static l2arc_dev_t *l2arc_dev_last; /* last device used */ 1334static list_t L2ARC_free_on_write; /* free after write buf list */ 1335static list_t *l2arc_free_on_write; /* free after write list ptr */ 1336static kmutex_t l2arc_free_on_write_mtx; /* mutex for list */ 1337static uint64_t l2arc_ndev; /* number of devices */ 1338 1339typedef struct l2arc_read_callback { 1340 arc_buf_hdr_t *l2rcb_hdr; /* read header */ 1341 blkptr_t l2rcb_bp; /* original blkptr */ 1342 zbookmark_phys_t l2rcb_zb; /* original bookmark */ 1343 int l2rcb_flags; /* original flags */ 1344 void *l2rcb_data; /* temporary buffer */ 1345} l2arc_read_callback_t; 1346 1347typedef struct l2arc_write_callback { 1348 l2arc_dev_t *l2wcb_dev; /* device info */ 1349 arc_buf_hdr_t *l2wcb_head; /* head of write buflist */ 1350} l2arc_write_callback_t; 1351 1352typedef struct l2arc_data_free { 1353 /* protected by l2arc_free_on_write_mtx */ 1354 void *l2df_data; 1355 size_t l2df_size; 1356 arc_buf_contents_t l2df_type; 1357 list_node_t l2df_list_node; 1358} l2arc_data_free_t; 1359 1360static kmutex_t l2arc_feed_thr_lock; 1361static kcondvar_t l2arc_feed_thr_cv; 1362static uint8_t l2arc_thread_exit; 1363 1364static void *arc_get_data_buf(arc_buf_hdr_t *, uint64_t, void *); 1365static void arc_free_data_buf(arc_buf_hdr_t *, void *, uint64_t, void *); 1366static void arc_hdr_free_pdata(arc_buf_hdr_t *hdr); 1367static void arc_hdr_alloc_pdata(arc_buf_hdr_t *); 1368static void arc_access(arc_buf_hdr_t *, kmutex_t *); 1369static boolean_t arc_is_overflowing(); 1370static void arc_buf_watch(arc_buf_t *); 1371 1372static arc_buf_contents_t arc_buf_type(arc_buf_hdr_t *); 1373static uint32_t arc_bufc_to_flags(arc_buf_contents_t); 1374static inline void arc_hdr_set_flags(arc_buf_hdr_t *hdr, arc_flags_t flags); 1375static inline void arc_hdr_clear_flags(arc_buf_hdr_t *hdr, arc_flags_t flags); 1376 1377static boolean_t l2arc_write_eligible(uint64_t, arc_buf_hdr_t *); 1378static void l2arc_read_done(zio_t *); 1379 1380static void 1381l2arc_trim(const arc_buf_hdr_t *hdr) 1382{ 1383 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev; 1384 1385 ASSERT(HDR_HAS_L2HDR(hdr)); 1386 ASSERT(MUTEX_HELD(&dev->l2ad_mtx)); 1387 1388 if (HDR_GET_PSIZE(hdr) != 0) { 1389 trim_map_free(dev->l2ad_vdev, hdr->b_l2hdr.b_daddr, 1390 HDR_GET_PSIZE(hdr), 0); 1391 } 1392} 1393 1394static uint64_t 1395buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth) 1396{ 1397 uint8_t *vdva = (uint8_t *)dva; 1398 uint64_t crc = -1ULL; 1399 int i; 1400 1401 ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY); 1402 1403 for (i = 0; i < sizeof (dva_t); i++) 1404 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ vdva[i]) & 0xFF]; 1405 1406 crc ^= (spa>>8) ^ birth; 1407 1408 return (crc); 1409} 1410 1411#define HDR_EMPTY(hdr) \ 1412 ((hdr)->b_dva.dva_word[0] == 0 && \ 1413 (hdr)->b_dva.dva_word[1] == 0) 1414 1415#define HDR_EQUAL(spa, dva, birth, hdr) \ 1416 ((hdr)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \ 1417 ((hdr)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \ 1418 ((hdr)->b_birth == birth) && ((hdr)->b_spa == spa) 1419 1420static void 1421buf_discard_identity(arc_buf_hdr_t *hdr) 1422{ 1423 hdr->b_dva.dva_word[0] = 0; 1424 hdr->b_dva.dva_word[1] = 0; 1425 hdr->b_birth = 0; 1426} 1427 1428static arc_buf_hdr_t * 1429buf_hash_find(uint64_t spa, const blkptr_t *bp, kmutex_t **lockp) 1430{ 1431 const dva_t *dva = BP_IDENTITY(bp); 1432 uint64_t birth = BP_PHYSICAL_BIRTH(bp); 1433 uint64_t idx = BUF_HASH_INDEX(spa, dva, birth); 1434 kmutex_t *hash_lock = BUF_HASH_LOCK(idx); 1435 arc_buf_hdr_t *hdr; 1436 1437 mutex_enter(hash_lock); 1438 for (hdr = buf_hash_table.ht_table[idx]; hdr != NULL; 1439 hdr = hdr->b_hash_next) { 1440 if (HDR_EQUAL(spa, dva, birth, hdr)) { 1441 *lockp = hash_lock; 1442 return (hdr); 1443 } 1444 } 1445 mutex_exit(hash_lock); 1446 *lockp = NULL; 1447 return (NULL); 1448} 1449 1450/* 1451 * Insert an entry into the hash table. If there is already an element 1452 * equal to elem in the hash table, then the already existing element 1453 * will be returned and the new element will not be inserted. 1454 * Otherwise returns NULL. 1455 * If lockp == NULL, the caller is assumed to already hold the hash lock. 1456 */ 1457static arc_buf_hdr_t * 1458buf_hash_insert(arc_buf_hdr_t *hdr, kmutex_t **lockp) 1459{ 1460 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth); 1461 kmutex_t *hash_lock = BUF_HASH_LOCK(idx); 1462 arc_buf_hdr_t *fhdr; 1463 uint32_t i; 1464 1465 ASSERT(!DVA_IS_EMPTY(&hdr->b_dva)); 1466 ASSERT(hdr->b_birth != 0); 1467 ASSERT(!HDR_IN_HASH_TABLE(hdr)); 1468 1469 if (lockp != NULL) { 1470 *lockp = hash_lock; 1471 mutex_enter(hash_lock); 1472 } else { 1473 ASSERT(MUTEX_HELD(hash_lock)); 1474 } 1475 1476 for (fhdr = buf_hash_table.ht_table[idx], i = 0; fhdr != NULL; 1477 fhdr = fhdr->b_hash_next, i++) { 1478 if (HDR_EQUAL(hdr->b_spa, &hdr->b_dva, hdr->b_birth, fhdr)) 1479 return (fhdr); 1480 } 1481 1482 hdr->b_hash_next = buf_hash_table.ht_table[idx]; 1483 buf_hash_table.ht_table[idx] = hdr; 1484 arc_hdr_set_flags(hdr, ARC_FLAG_IN_HASH_TABLE); 1485 1486 /* collect some hash table performance data */ 1487 if (i > 0) { 1488 ARCSTAT_BUMP(arcstat_hash_collisions); 1489 if (i == 1) 1490 ARCSTAT_BUMP(arcstat_hash_chains); 1491 1492 ARCSTAT_MAX(arcstat_hash_chain_max, i); 1493 } 1494 1495 ARCSTAT_BUMP(arcstat_hash_elements); 1496 ARCSTAT_MAXSTAT(arcstat_hash_elements); 1497 1498 return (NULL); 1499} 1500 1501static void 1502buf_hash_remove(arc_buf_hdr_t *hdr) 1503{ 1504 arc_buf_hdr_t *fhdr, **hdrp; 1505 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth); 1506 1507 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx))); 1508 ASSERT(HDR_IN_HASH_TABLE(hdr)); 1509 1510 hdrp = &buf_hash_table.ht_table[idx]; 1511 while ((fhdr = *hdrp) != hdr) { 1512 ASSERT3P(fhdr, !=, NULL); 1513 hdrp = &fhdr->b_hash_next; 1514 } 1515 *hdrp = hdr->b_hash_next; 1516 hdr->b_hash_next = NULL; 1517 arc_hdr_clear_flags(hdr, ARC_FLAG_IN_HASH_TABLE); 1518 1519 /* collect some hash table performance data */ 1520 ARCSTAT_BUMPDOWN(arcstat_hash_elements); 1521 1522 if (buf_hash_table.ht_table[idx] && 1523 buf_hash_table.ht_table[idx]->b_hash_next == NULL) 1524 ARCSTAT_BUMPDOWN(arcstat_hash_chains); 1525} 1526 1527/* 1528 * Global data structures and functions for the buf kmem cache. 1529 */ 1530static kmem_cache_t *hdr_full_cache; 1531static kmem_cache_t *hdr_l2only_cache; 1532static kmem_cache_t *buf_cache; 1533 1534static void 1535buf_fini(void) 1536{ 1537 int i; 1538 1539 kmem_free(buf_hash_table.ht_table, 1540 (buf_hash_table.ht_mask + 1) * sizeof (void *)); 1541 for (i = 0; i < BUF_LOCKS; i++) 1542 mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock); 1543 kmem_cache_destroy(hdr_full_cache); 1544 kmem_cache_destroy(hdr_l2only_cache); 1545 kmem_cache_destroy(buf_cache); 1546} 1547 1548/* 1549 * Constructor callback - called when the cache is empty 1550 * and a new buf is requested. 1551 */ 1552/* ARGSUSED */ 1553static int 1554hdr_full_cons(void *vbuf, void *unused, int kmflag) 1555{ 1556 arc_buf_hdr_t *hdr = vbuf; 1557 1558 bzero(hdr, HDR_FULL_SIZE); 1559 cv_init(&hdr->b_l1hdr.b_cv, NULL, CV_DEFAULT, NULL); 1560 refcount_create(&hdr->b_l1hdr.b_refcnt); 1561 mutex_init(&hdr->b_l1hdr.b_freeze_lock, NULL, MUTEX_DEFAULT, NULL); 1562 multilist_link_init(&hdr->b_l1hdr.b_arc_node); 1563 arc_space_consume(HDR_FULL_SIZE, ARC_SPACE_HDRS); 1564 1565 return (0); 1566} 1567 1568/* ARGSUSED */ 1569static int 1570hdr_l2only_cons(void *vbuf, void *unused, int kmflag) 1571{ 1572 arc_buf_hdr_t *hdr = vbuf; 1573 1574 bzero(hdr, HDR_L2ONLY_SIZE); 1575 arc_space_consume(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS); 1576 1577 return (0); 1578} 1579 1580/* ARGSUSED */ 1581static int 1582buf_cons(void *vbuf, void *unused, int kmflag) 1583{ 1584 arc_buf_t *buf = vbuf; 1585 1586 bzero(buf, sizeof (arc_buf_t)); 1587 mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL); 1588 arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS); 1589 1590 return (0); 1591} 1592 1593/* 1594 * Destructor callback - called when a cached buf is 1595 * no longer required. 1596 */ 1597/* ARGSUSED */ 1598static void 1599hdr_full_dest(void *vbuf, void *unused) 1600{ 1601 arc_buf_hdr_t *hdr = vbuf; 1602 1603 ASSERT(HDR_EMPTY(hdr)); 1604 cv_destroy(&hdr->b_l1hdr.b_cv); 1605 refcount_destroy(&hdr->b_l1hdr.b_refcnt); 1606 mutex_destroy(&hdr->b_l1hdr.b_freeze_lock); 1607 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 1608 arc_space_return(HDR_FULL_SIZE, ARC_SPACE_HDRS); 1609} 1610 1611/* ARGSUSED */ 1612static void 1613hdr_l2only_dest(void *vbuf, void *unused) 1614{ 1615 arc_buf_hdr_t *hdr = vbuf; 1616 1617 ASSERT(HDR_EMPTY(hdr)); 1618 arc_space_return(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS); 1619} 1620 1621/* ARGSUSED */ 1622static void 1623buf_dest(void *vbuf, void *unused) 1624{ 1625 arc_buf_t *buf = vbuf; 1626 1627 mutex_destroy(&buf->b_evict_lock); 1628 arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS); 1629} 1630 1631/* 1632 * Reclaim callback -- invoked when memory is low. 1633 */ 1634/* ARGSUSED */ 1635static void 1636hdr_recl(void *unused) 1637{ 1638 dprintf("hdr_recl called\n"); 1639 /* 1640 * umem calls the reclaim func when we destroy the buf cache, 1641 * which is after we do arc_fini(). 1642 */ 1643 if (!arc_dead) 1644 cv_signal(&arc_reclaim_thread_cv); 1645} 1646 1647static void 1648buf_init(void) 1649{ 1650 uint64_t *ct; 1651 uint64_t hsize = 1ULL << 12; 1652 int i, j; 1653 1654 /* 1655 * The hash table is big enough to fill all of physical memory 1656 * with an average block size of zfs_arc_average_blocksize (default 8K). 1657 * By default, the table will take up 1658 * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers). 1659 */ 1660 while (hsize * zfs_arc_average_blocksize < (uint64_t)physmem * PAGESIZE) 1661 hsize <<= 1; 1662retry: 1663 buf_hash_table.ht_mask = hsize - 1; 1664 buf_hash_table.ht_table = 1665 kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP); 1666 if (buf_hash_table.ht_table == NULL) { 1667 ASSERT(hsize > (1ULL << 8)); 1668 hsize >>= 1; 1669 goto retry; 1670 } 1671 1672 hdr_full_cache = kmem_cache_create("arc_buf_hdr_t_full", HDR_FULL_SIZE, 1673 0, hdr_full_cons, hdr_full_dest, hdr_recl, NULL, NULL, 0); 1674 hdr_l2only_cache = kmem_cache_create("arc_buf_hdr_t_l2only", 1675 HDR_L2ONLY_SIZE, 0, hdr_l2only_cons, hdr_l2only_dest, hdr_recl, 1676 NULL, NULL, 0); 1677 buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t), 1678 0, buf_cons, buf_dest, NULL, NULL, NULL, 0); 1679 1680 for (i = 0; i < 256; i++) 1681 for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--) 1682 *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY); 1683 1684 for (i = 0; i < BUF_LOCKS; i++) { 1685 mutex_init(&buf_hash_table.ht_locks[i].ht_lock, 1686 NULL, MUTEX_DEFAULT, NULL); 1687 } 1688} 1689 1690/* 1691 * This is the size that the buf occupies in memory. If the buf is compressed, 1692 * it will correspond to the compressed size. You should use this method of 1693 * getting the buf size unless you explicitly need the logical size. 1694 */ 1695int32_t 1696arc_buf_size(arc_buf_t *buf) 1697{ 1698 return (ARC_BUF_COMPRESSED(buf) ? 1699 HDR_GET_PSIZE(buf->b_hdr) : HDR_GET_LSIZE(buf->b_hdr)); 1700} 1701 1702int32_t 1703arc_buf_lsize(arc_buf_t *buf) 1704{ 1705 return (HDR_GET_LSIZE(buf->b_hdr)); 1706} 1707 1708enum zio_compress 1709arc_get_compression(arc_buf_t *buf) 1710{ 1711 return (ARC_BUF_COMPRESSED(buf) ? 1712 HDR_GET_COMPRESS(buf->b_hdr) : ZIO_COMPRESS_OFF); 1713} 1714 1715#define ARC_MINTIME (hz>>4) /* 62 ms */ 1716 1717static inline boolean_t 1718arc_buf_is_shared(arc_buf_t *buf) 1719{ 1720 boolean_t shared = (buf->b_data != NULL && 1721 buf->b_data == buf->b_hdr->b_l1hdr.b_pdata); 1722 IMPLY(shared, HDR_SHARED_DATA(buf->b_hdr)); 1723 IMPLY(shared, ARC_BUF_SHARED(buf)); 1724 IMPLY(shared, ARC_BUF_COMPRESSED(buf) || ARC_BUF_LAST(buf)); 1725 1726 /* 1727 * It would be nice to assert arc_can_share() too, but the "hdr isn't 1728 * already being shared" requirement prevents us from doing that. 1729 */ 1730 1731 return (shared); 1732} 1733 1734/* 1735 * Free the checksum associated with this header. If there is no checksum, this 1736 * is a no-op. 1737 */ 1738static inline void 1739arc_cksum_free(arc_buf_hdr_t *hdr) 1740{ 1741 ASSERT(HDR_HAS_L1HDR(hdr)); 1742 mutex_enter(&hdr->b_l1hdr.b_freeze_lock); 1743 if (hdr->b_l1hdr.b_freeze_cksum != NULL) { 1744 kmem_free(hdr->b_l1hdr.b_freeze_cksum, sizeof (zio_cksum_t)); 1745 hdr->b_l1hdr.b_freeze_cksum = NULL; 1746 } 1747 mutex_exit(&hdr->b_l1hdr.b_freeze_lock); 1748} 1749 1750/* 1751 * Return true iff at least one of the bufs on hdr is not compressed. 1752 */ 1753static boolean_t 1754arc_hdr_has_uncompressed_buf(arc_buf_hdr_t *hdr) 1755{ 1756 for (arc_buf_t *b = hdr->b_l1hdr.b_buf; b != NULL; b = b->b_next) { 1757 if (!ARC_BUF_COMPRESSED(b)) { 1758 return (B_TRUE); 1759 } 1760 } 1761 return (B_FALSE); 1762} 1763 1764/* 1765 * If we've turned on the ZFS_DEBUG_MODIFY flag, verify that the buf's data 1766 * matches the checksum that is stored in the hdr. If there is no checksum, 1767 * or if the buf is compressed, this is a no-op. 1768 */ 1769static void 1770arc_cksum_verify(arc_buf_t *buf) 1771{ 1772 arc_buf_hdr_t *hdr = buf->b_hdr; 1773 zio_cksum_t zc; 1774 1775 if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 1776 return; 1777 1778 if (ARC_BUF_COMPRESSED(buf)) { 1779 ASSERT(hdr->b_l1hdr.b_freeze_cksum == NULL || 1780 arc_hdr_has_uncompressed_buf(hdr)); 1781 return; 1782 } 1783 1784 ASSERT(HDR_HAS_L1HDR(hdr)); 1785 1786 mutex_enter(&hdr->b_l1hdr.b_freeze_lock); 1787 if (hdr->b_l1hdr.b_freeze_cksum == NULL || HDR_IO_ERROR(hdr)) { 1788 mutex_exit(&hdr->b_l1hdr.b_freeze_lock); 1789 return; 1790 } 1791 1792 fletcher_2_native(buf->b_data, arc_buf_size(buf), NULL, &zc); 1793 if (!ZIO_CHECKSUM_EQUAL(*hdr->b_l1hdr.b_freeze_cksum, zc)) 1794 panic("buffer modified while frozen!"); 1795 mutex_exit(&hdr->b_l1hdr.b_freeze_lock); 1796} 1797 1798static boolean_t 1799arc_cksum_is_equal(arc_buf_hdr_t *hdr, zio_t *zio) 1800{ 1801 enum zio_compress compress = BP_GET_COMPRESS(zio->io_bp); 1802 boolean_t valid_cksum; 1803 1804 ASSERT(!BP_IS_EMBEDDED(zio->io_bp)); 1805 VERIFY3U(BP_GET_PSIZE(zio->io_bp), ==, HDR_GET_PSIZE(hdr)); 1806 1807 /* 1808 * We rely on the blkptr's checksum to determine if the block 1809 * is valid or not. When compressed arc is enabled, the l2arc 1810 * writes the block to the l2arc just as it appears in the pool. 1811 * This allows us to use the blkptr's checksum to validate the 1812 * data that we just read off of the l2arc without having to store 1813 * a separate checksum in the arc_buf_hdr_t. However, if compressed 1814 * arc is disabled, then the data written to the l2arc is always 1815 * uncompressed and won't match the block as it exists in the main 1816 * pool. When this is the case, we must first compress it if it is 1817 * compressed on the main pool before we can validate the checksum. 1818 */ 1819 if (!HDR_COMPRESSION_ENABLED(hdr) && compress != ZIO_COMPRESS_OFF) { 1820 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF); 1821 uint64_t lsize = HDR_GET_LSIZE(hdr); 1822 uint64_t csize; 1823 1824 void *cbuf = zio_buf_alloc(HDR_GET_PSIZE(hdr)); 1825 csize = zio_compress_data(compress, zio->io_data, cbuf, lsize); 1826 ASSERT3U(csize, <=, HDR_GET_PSIZE(hdr)); 1827 if (csize < HDR_GET_PSIZE(hdr)) { 1828 /* 1829 * Compressed blocks are always a multiple of the 1830 * smallest ashift in the pool. Ideally, we would 1831 * like to round up the csize to the next 1832 * spa_min_ashift but that value may have changed 1833 * since the block was last written. Instead, 1834 * we rely on the fact that the hdr's psize 1835 * was set to the psize of the block when it was 1836 * last written. We set the csize to that value 1837 * and zero out any part that should not contain 1838 * data. 1839 */ 1840 bzero((char *)cbuf + csize, HDR_GET_PSIZE(hdr) - csize); 1841 csize = HDR_GET_PSIZE(hdr); 1842 } 1843 zio_push_transform(zio, cbuf, csize, HDR_GET_PSIZE(hdr), NULL); 1844 } 1845 1846 /* 1847 * Block pointers always store the checksum for the logical data. 1848 * If the block pointer has the gang bit set, then the checksum 1849 * it represents is for the reconstituted data and not for an 1850 * individual gang member. The zio pipeline, however, must be able to 1851 * determine the checksum of each of the gang constituents so it 1852 * treats the checksum comparison differently than what we need 1853 * for l2arc blocks. This prevents us from using the 1854 * zio_checksum_error() interface directly. Instead we must call the 1855 * zio_checksum_error_impl() so that we can ensure the checksum is 1856 * generated using the correct checksum algorithm and accounts for the 1857 * logical I/O size and not just a gang fragment. 1858 */ 1859 valid_cksum = (zio_checksum_error_impl(zio->io_spa, zio->io_bp, 1860 BP_GET_CHECKSUM(zio->io_bp), zio->io_data, zio->io_size, 1861 zio->io_offset, NULL) == 0); 1862 zio_pop_transforms(zio); 1863 return (valid_cksum); 1864} 1865 1866/* 1867 * Given a buf full of data, if ZFS_DEBUG_MODIFY is enabled this computes a 1868 * checksum and attaches it to the buf's hdr so that we can ensure that the buf 1869 * isn't modified later on. If buf is compressed or there is already a checksum 1870 * on the hdr, this is a no-op (we only checksum uncompressed bufs). 1871 */ 1872static void 1873arc_cksum_compute(arc_buf_t *buf) 1874{ 1875 arc_buf_hdr_t *hdr = buf->b_hdr; 1876 1877 if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 1878 return; 1879 1880 ASSERT(HDR_HAS_L1HDR(hdr)); 1881 1882 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1883 if (hdr->b_l1hdr.b_freeze_cksum != NULL) { 1884 ASSERT(arc_hdr_has_uncompressed_buf(hdr)); 1885 mutex_exit(&hdr->b_l1hdr.b_freeze_lock); 1886 return; 1887 } else if (ARC_BUF_COMPRESSED(buf)) { 1888 mutex_exit(&hdr->b_l1hdr.b_freeze_lock); 1889 return; 1890 } 1891 1892 ASSERT(!ARC_BUF_COMPRESSED(buf)); 1893 hdr->b_l1hdr.b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t), 1894 KM_SLEEP); 1895 fletcher_2_native(buf->b_data, arc_buf_size(buf), NULL, 1896 hdr->b_l1hdr.b_freeze_cksum); 1897 mutex_exit(&hdr->b_l1hdr.b_freeze_lock); 1898#ifdef illumos 1899 arc_buf_watch(buf); 1900#endif 1901} 1902 1903#ifdef illumos 1904#ifndef _KERNEL 1905typedef struct procctl { 1906 long cmd; 1907 prwatch_t prwatch; 1908} procctl_t; 1909#endif 1910 1911/* ARGSUSED */ 1912static void 1913arc_buf_unwatch(arc_buf_t *buf) 1914{ 1915#ifndef _KERNEL 1916 if (arc_watch) { 1917 int result; 1918 procctl_t ctl; 1919 ctl.cmd = PCWATCH; 1920 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data; 1921 ctl.prwatch.pr_size = 0; 1922 ctl.prwatch.pr_wflags = 0; 1923 result = write(arc_procfd, &ctl, sizeof (ctl)); 1924 ASSERT3U(result, ==, sizeof (ctl)); 1925 } 1926#endif 1927} 1928 1929/* ARGSUSED */ 1930static void 1931arc_buf_watch(arc_buf_t *buf) 1932{ 1933#ifndef _KERNEL 1934 if (arc_watch) { 1935 int result; 1936 procctl_t ctl; 1937 ctl.cmd = PCWATCH; 1938 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data; 1939 ctl.prwatch.pr_size = arc_buf_size(buf); 1940 ctl.prwatch.pr_wflags = WA_WRITE; 1941 result = write(arc_procfd, &ctl, sizeof (ctl)); 1942 ASSERT3U(result, ==, sizeof (ctl)); 1943 } 1944#endif 1945} 1946#endif /* illumos */ 1947 1948static arc_buf_contents_t 1949arc_buf_type(arc_buf_hdr_t *hdr) 1950{ 1951 arc_buf_contents_t type; 1952 if (HDR_ISTYPE_METADATA(hdr)) { 1953 type = ARC_BUFC_METADATA; 1954 } else { 1955 type = ARC_BUFC_DATA; 1956 } 1957 VERIFY3U(hdr->b_type, ==, type); 1958 return (type); 1959} 1960 1961boolean_t 1962arc_is_metadata(arc_buf_t *buf) 1963{ 1964 return (HDR_ISTYPE_METADATA(buf->b_hdr) != 0); 1965} 1966 1967static uint32_t 1968arc_bufc_to_flags(arc_buf_contents_t type) 1969{ 1970 switch (type) { 1971 case ARC_BUFC_DATA: 1972 /* metadata field is 0 if buffer contains normal data */ 1973 return (0); 1974 case ARC_BUFC_METADATA: 1975 return (ARC_FLAG_BUFC_METADATA); 1976 default: 1977 break; 1978 } 1979 panic("undefined ARC buffer type!"); 1980 return ((uint32_t)-1); 1981} 1982 1983void 1984arc_buf_thaw(arc_buf_t *buf) 1985{ 1986 arc_buf_hdr_t *hdr = buf->b_hdr; 1987 1988 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon); 1989 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 1990 1991 arc_cksum_verify(buf); 1992 1993 /* 1994 * Compressed buffers do not manipulate the b_freeze_cksum or 1995 * allocate b_thawed. 1996 */ 1997 if (ARC_BUF_COMPRESSED(buf)) { 1998 ASSERT(hdr->b_l1hdr.b_freeze_cksum == NULL || 1999 arc_hdr_has_uncompressed_buf(hdr)); 2000 return; 2001 } 2002 2003 ASSERT(HDR_HAS_L1HDR(hdr)); 2004 arc_cksum_free(hdr); 2005 2006 mutex_enter(&hdr->b_l1hdr.b_freeze_lock); 2007#ifdef ZFS_DEBUG 2008 if (zfs_flags & ZFS_DEBUG_MODIFY) { 2009 if (hdr->b_l1hdr.b_thawed != NULL) 2010 kmem_free(hdr->b_l1hdr.b_thawed, 1); 2011 hdr->b_l1hdr.b_thawed = kmem_alloc(1, KM_SLEEP); 2012 } 2013#endif 2014 2015 mutex_exit(&hdr->b_l1hdr.b_freeze_lock); 2016 2017#ifdef illumos 2018 arc_buf_unwatch(buf); 2019#endif 2020} 2021 2022void 2023arc_buf_freeze(arc_buf_t *buf) 2024{ 2025 arc_buf_hdr_t *hdr = buf->b_hdr; 2026 kmutex_t *hash_lock; 2027 2028 if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 2029 return; 2030 2031 if (ARC_BUF_COMPRESSED(buf)) { 2032 ASSERT(hdr->b_l1hdr.b_freeze_cksum == NULL || 2033 arc_hdr_has_uncompressed_buf(hdr)); 2034 return; 2035 } 2036 2037 hash_lock = HDR_LOCK(hdr); 2038 mutex_enter(hash_lock); 2039 2040 ASSERT(HDR_HAS_L1HDR(hdr)); 2041 ASSERT(hdr->b_l1hdr.b_freeze_cksum != NULL || 2042 hdr->b_l1hdr.b_state == arc_anon); 2043 arc_cksum_compute(buf); 2044 mutex_exit(hash_lock); 2045} 2046 2047/* 2048 * The arc_buf_hdr_t's b_flags should never be modified directly. Instead, 2049 * the following functions should be used to ensure that the flags are 2050 * updated in a thread-safe way. When manipulating the flags either 2051 * the hash_lock must be held or the hdr must be undiscoverable. This 2052 * ensures that we're not racing with any other threads when updating 2053 * the flags. 2054 */ 2055static inline void 2056arc_hdr_set_flags(arc_buf_hdr_t *hdr, arc_flags_t flags) 2057{ 2058 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); 2059 hdr->b_flags |= flags; 2060} 2061 2062static inline void 2063arc_hdr_clear_flags(arc_buf_hdr_t *hdr, arc_flags_t flags) 2064{ 2065 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); 2066 hdr->b_flags &= ~flags; 2067} 2068 2069/* 2070 * Setting the compression bits in the arc_buf_hdr_t's b_flags is 2071 * done in a special way since we have to clear and set bits 2072 * at the same time. Consumers that wish to set the compression bits 2073 * must use this function to ensure that the flags are updated in 2074 * thread-safe manner. 2075 */ 2076static void 2077arc_hdr_set_compress(arc_buf_hdr_t *hdr, enum zio_compress cmp) 2078{ 2079 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); 2080 2081 /* 2082 * Holes and embedded blocks will always have a psize = 0 so 2083 * we ignore the compression of the blkptr and set the 2084 * arc_buf_hdr_t's compression to ZIO_COMPRESS_OFF. 2085 * Holes and embedded blocks remain anonymous so we don't 2086 * want to uncompress them. Mark them as uncompressed. 2087 */ 2088 if (!zfs_compressed_arc_enabled || HDR_GET_PSIZE(hdr) == 0) { 2089 arc_hdr_clear_flags(hdr, ARC_FLAG_COMPRESSED_ARC); 2090 HDR_SET_COMPRESS(hdr, ZIO_COMPRESS_OFF); 2091 ASSERT(!HDR_COMPRESSION_ENABLED(hdr)); 2092 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF); 2093 } else { 2094 arc_hdr_set_flags(hdr, ARC_FLAG_COMPRESSED_ARC); 2095 HDR_SET_COMPRESS(hdr, cmp); 2096 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, cmp); 2097 ASSERT(HDR_COMPRESSION_ENABLED(hdr)); 2098 } 2099} 2100 2101/* 2102 * Looks for another buf on the same hdr which has the data decompressed, copies 2103 * from it, and returns true. If no such buf exists, returns false. 2104 */ 2105static boolean_t 2106arc_buf_try_copy_decompressed_data(arc_buf_t *buf) 2107{ 2108 arc_buf_hdr_t *hdr = buf->b_hdr; 2109 boolean_t copied = B_FALSE; 2110 2111 ASSERT(HDR_HAS_L1HDR(hdr)); 2112 ASSERT3P(buf->b_data, !=, NULL); 2113 ASSERT(!ARC_BUF_COMPRESSED(buf)); 2114 2115 for (arc_buf_t *from = hdr->b_l1hdr.b_buf; from != NULL; 2116 from = from->b_next) { 2117 /* can't use our own data buffer */ 2118 if (from == buf) { 2119 continue; 2120 } 2121 2122 if (!ARC_BUF_COMPRESSED(from)) { 2123 bcopy(from->b_data, buf->b_data, arc_buf_size(buf)); 2124 copied = B_TRUE; 2125 break; 2126 } 2127 } 2128 2129 /* 2130 * There were no decompressed bufs, so there should not be a 2131 * checksum on the hdr either. 2132 */ 2133 EQUIV(!copied, hdr->b_l1hdr.b_freeze_cksum == NULL); 2134 2135 return (copied); 2136} 2137 2138/* 2139 * Given a buf that has a data buffer attached to it, this function will 2140 * efficiently fill the buf with data of the specified compression setting from 2141 * the hdr and update the hdr's b_freeze_cksum if necessary. If the buf and hdr 2142 * are already sharing a data buf, no copy is performed. 2143 * 2144 * If the buf is marked as compressed but uncompressed data was requested, this 2145 * will allocate a new data buffer for the buf, remove that flag, and fill the 2146 * buf with uncompressed data. You can't request a compressed buf on a hdr with 2147 * uncompressed data, and (since we haven't added support for it yet) if you 2148 * want compressed data your buf must already be marked as compressed and have 2149 * the correct-sized data buffer. 2150 */ 2151static int 2152arc_buf_fill(arc_buf_t *buf, boolean_t compressed) 2153{ 2154 arc_buf_hdr_t *hdr = buf->b_hdr; 2155 boolean_t hdr_compressed = (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF); 2156 dmu_object_byteswap_t bswap = hdr->b_l1hdr.b_byteswap; 2157 2158 ASSERT3P(buf->b_data, !=, NULL); 2159 IMPLY(compressed, hdr_compressed); 2160 IMPLY(compressed, ARC_BUF_COMPRESSED(buf)); 2161 2162 if (hdr_compressed == compressed) { 2163 if (!arc_buf_is_shared(buf)) { 2164 bcopy(hdr->b_l1hdr.b_pdata, buf->b_data, 2165 arc_buf_size(buf)); 2166 } 2167 } else { 2168 ASSERT(hdr_compressed); 2169 ASSERT(!compressed); 2170 ASSERT3U(HDR_GET_LSIZE(hdr), !=, HDR_GET_PSIZE(hdr)); 2171 2172 /* 2173 * If the buf is sharing its data with the hdr, unlink it and 2174 * allocate a new data buffer for the buf. 2175 */ 2176 if (arc_buf_is_shared(buf)) { 2177 ASSERT(ARC_BUF_COMPRESSED(buf)); 2178 2179 /* We need to give the buf it's own b_data */ 2180 buf->b_flags &= ~ARC_BUF_FLAG_SHARED; 2181 buf->b_data = 2182 arc_get_data_buf(hdr, HDR_GET_LSIZE(hdr), buf); 2183 arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA); 2184 2185 /* Previously overhead was 0; just add new overhead */ 2186 ARCSTAT_INCR(arcstat_overhead_size, HDR_GET_LSIZE(hdr)); 2187 } else if (ARC_BUF_COMPRESSED(buf)) { 2188 /* We need to reallocate the buf's b_data */ 2189 arc_free_data_buf(hdr, buf->b_data, HDR_GET_PSIZE(hdr), 2190 buf); 2191 buf->b_data = 2192 arc_get_data_buf(hdr, HDR_GET_LSIZE(hdr), buf); 2193 2194 /* We increased the size of b_data; update overhead */ 2195 ARCSTAT_INCR(arcstat_overhead_size, 2196 HDR_GET_LSIZE(hdr) - HDR_GET_PSIZE(hdr)); 2197 } 2198 2199 /* 2200 * Regardless of the buf's previous compression settings, it 2201 * should not be compressed at the end of this function. 2202 */ 2203 buf->b_flags &= ~ARC_BUF_FLAG_COMPRESSED; 2204 2205 /* 2206 * Try copying the data from another buf which already has a 2207 * decompressed version. If that's not possible, it's time to 2208 * bite the bullet and decompress the data from the hdr. 2209 */ 2210 if (arc_buf_try_copy_decompressed_data(buf)) { 2211 /* Skip byteswapping and checksumming (already done) */ 2212 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, !=, NULL); 2213 return (0); 2214 } else { 2215 int error = zio_decompress_data(HDR_GET_COMPRESS(hdr), 2216 hdr->b_l1hdr.b_pdata, buf->b_data, 2217 HDR_GET_PSIZE(hdr), HDR_GET_LSIZE(hdr)); 2218 2219 /* 2220 * Absent hardware errors or software bugs, this should 2221 * be impossible, but log it anyway so we can debug it. 2222 */ 2223 if (error != 0) { 2224 zfs_dbgmsg( 2225 "hdr %p, compress %d, psize %d, lsize %d", 2226 hdr, HDR_GET_COMPRESS(hdr), 2227 HDR_GET_PSIZE(hdr), HDR_GET_LSIZE(hdr)); 2228 return (SET_ERROR(EIO)); 2229 } 2230 } 2231 } 2232 2233 /* Byteswap the buf's data if necessary */ 2234 if (bswap != DMU_BSWAP_NUMFUNCS) { 2235 ASSERT(!HDR_SHARED_DATA(hdr)); 2236 ASSERT3U(bswap, <, DMU_BSWAP_NUMFUNCS); 2237 dmu_ot_byteswap[bswap].ob_func(buf->b_data, HDR_GET_LSIZE(hdr)); 2238 } 2239 2240 /* Compute the hdr's checksum if necessary */ 2241 arc_cksum_compute(buf); 2242 2243 return (0); 2244} 2245 2246int 2247arc_decompress(arc_buf_t *buf) 2248{ 2249 return (arc_buf_fill(buf, B_FALSE)); 2250} 2251 2252/* 2253 * Return the size of the block, b_pdata, that is stored in the arc_buf_hdr_t. 2254 */ 2255static uint64_t 2256arc_hdr_size(arc_buf_hdr_t *hdr) 2257{ 2258 uint64_t size; 2259 2260 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF && 2261 HDR_GET_PSIZE(hdr) > 0) { 2262 size = HDR_GET_PSIZE(hdr); 2263 } else { 2264 ASSERT3U(HDR_GET_LSIZE(hdr), !=, 0); 2265 size = HDR_GET_LSIZE(hdr); 2266 } 2267 return (size); 2268} 2269 2270/* 2271 * Increment the amount of evictable space in the arc_state_t's refcount. 2272 * We account for the space used by the hdr and the arc buf individually 2273 * so that we can add and remove them from the refcount individually. 2274 */ 2275static void 2276arc_evictable_space_increment(arc_buf_hdr_t *hdr, arc_state_t *state) 2277{ 2278 arc_buf_contents_t type = arc_buf_type(hdr); 2279 2280 ASSERT(HDR_HAS_L1HDR(hdr)); 2281 2282 if (GHOST_STATE(state)) { 2283 ASSERT0(hdr->b_l1hdr.b_bufcnt); 2284 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 2285 ASSERT3P(hdr->b_l1hdr.b_pdata, ==, NULL); 2286 (void) refcount_add_many(&state->arcs_esize[type], 2287 HDR_GET_LSIZE(hdr), hdr); 2288 return; 2289 } 2290 2291 ASSERT(!GHOST_STATE(state)); 2292 if (hdr->b_l1hdr.b_pdata != NULL) { 2293 (void) refcount_add_many(&state->arcs_esize[type], 2294 arc_hdr_size(hdr), hdr); 2295 } 2296 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL; 2297 buf = buf->b_next) { 2298 if (arc_buf_is_shared(buf)) 2299 continue; 2300 (void) refcount_add_many(&state->arcs_esize[type], 2301 arc_buf_size(buf), buf); 2302 } 2303} 2304 2305/* 2306 * Decrement the amount of evictable space in the arc_state_t's refcount. 2307 * We account for the space used by the hdr and the arc buf individually 2308 * so that we can add and remove them from the refcount individually. 2309 */ 2310static void 2311arc_evictable_space_decrement(arc_buf_hdr_t *hdr, arc_state_t *state) 2312{ 2313 arc_buf_contents_t type = arc_buf_type(hdr); 2314 2315 ASSERT(HDR_HAS_L1HDR(hdr)); 2316 2317 if (GHOST_STATE(state)) { 2318 ASSERT0(hdr->b_l1hdr.b_bufcnt); 2319 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 2320 ASSERT3P(hdr->b_l1hdr.b_pdata, ==, NULL); 2321 (void) refcount_remove_many(&state->arcs_esize[type], 2322 HDR_GET_LSIZE(hdr), hdr); 2323 return; 2324 } 2325 2326 ASSERT(!GHOST_STATE(state)); 2327 if (hdr->b_l1hdr.b_pdata != NULL) { 2328 (void) refcount_remove_many(&state->arcs_esize[type], 2329 arc_hdr_size(hdr), hdr); 2330 } 2331 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL; 2332 buf = buf->b_next) { 2333 if (arc_buf_is_shared(buf)) 2334 continue; 2335 (void) refcount_remove_many(&state->arcs_esize[type], 2336 arc_buf_size(buf), buf); 2337 } 2338} 2339 2340/* 2341 * Add a reference to this hdr indicating that someone is actively 2342 * referencing that memory. When the refcount transitions from 0 to 1, 2343 * we remove it from the respective arc_state_t list to indicate that 2344 * it is not evictable. 2345 */ 2346static void 2347add_reference(arc_buf_hdr_t *hdr, void *tag) 2348{ 2349 ASSERT(HDR_HAS_L1HDR(hdr)); 2350 if (!MUTEX_HELD(HDR_LOCK(hdr))) { 2351 ASSERT(hdr->b_l1hdr.b_state == arc_anon); 2352 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 2353 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 2354 } 2355 2356 arc_state_t *state = hdr->b_l1hdr.b_state; 2357 2358 if ((refcount_add(&hdr->b_l1hdr.b_refcnt, tag) == 1) && 2359 (state != arc_anon)) { 2360 /* We don't use the L2-only state list. */ 2361 if (state != arc_l2c_only) { 2362 multilist_remove(state->arcs_list[arc_buf_type(hdr)], 2363 hdr); 2364 arc_evictable_space_decrement(hdr, state); 2365 } 2366 /* remove the prefetch flag if we get a reference */ 2367 arc_hdr_clear_flags(hdr, ARC_FLAG_PREFETCH); 2368 } 2369} 2370 2371/* 2372 * Remove a reference from this hdr. When the reference transitions from 2373 * 1 to 0 and we're not anonymous, then we add this hdr to the arc_state_t's 2374 * list making it eligible for eviction. 2375 */ 2376static int 2377remove_reference(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, void *tag) 2378{ 2379 int cnt; 2380 arc_state_t *state = hdr->b_l1hdr.b_state; 2381 2382 ASSERT(HDR_HAS_L1HDR(hdr)); 2383 ASSERT(state == arc_anon || MUTEX_HELD(hash_lock)); 2384 ASSERT(!GHOST_STATE(state)); 2385 2386 /* 2387 * arc_l2c_only counts as a ghost state so we don't need to explicitly 2388 * check to prevent usage of the arc_l2c_only list. 2389 */ 2390 if (((cnt = refcount_remove(&hdr->b_l1hdr.b_refcnt, tag)) == 0) && 2391 (state != arc_anon)) { 2392 multilist_insert(state->arcs_list[arc_buf_type(hdr)], hdr); 2393 ASSERT3U(hdr->b_l1hdr.b_bufcnt, >, 0); 2394 arc_evictable_space_increment(hdr, state); 2395 } 2396 return (cnt); 2397} 2398 2399/* 2400 * Move the supplied buffer to the indicated state. The hash lock 2401 * for the buffer must be held by the caller. 2402 */ 2403static void 2404arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *hdr, 2405 kmutex_t *hash_lock) 2406{ 2407 arc_state_t *old_state; 2408 int64_t refcnt; 2409 uint32_t bufcnt; 2410 boolean_t update_old, update_new; 2411 arc_buf_contents_t buftype = arc_buf_type(hdr); 2412 2413 /* 2414 * We almost always have an L1 hdr here, since we call arc_hdr_realloc() 2415 * in arc_read() when bringing a buffer out of the L2ARC. However, the 2416 * L1 hdr doesn't always exist when we change state to arc_anon before 2417 * destroying a header, in which case reallocating to add the L1 hdr is 2418 * pointless. 2419 */ 2420 if (HDR_HAS_L1HDR(hdr)) { 2421 old_state = hdr->b_l1hdr.b_state; 2422 refcnt = refcount_count(&hdr->b_l1hdr.b_refcnt); 2423 bufcnt = hdr->b_l1hdr.b_bufcnt; 2424 update_old = (bufcnt > 0 || hdr->b_l1hdr.b_pdata != NULL); 2425 } else { 2426 old_state = arc_l2c_only; 2427 refcnt = 0; 2428 bufcnt = 0; 2429 update_old = B_FALSE; 2430 } 2431 update_new = update_old; 2432 2433 ASSERT(MUTEX_HELD(hash_lock)); 2434 ASSERT3P(new_state, !=, old_state); 2435 ASSERT(!GHOST_STATE(new_state) || bufcnt == 0); 2436 ASSERT(old_state != arc_anon || bufcnt <= 1); 2437 2438 /* 2439 * If this buffer is evictable, transfer it from the 2440 * old state list to the new state list. 2441 */ 2442 if (refcnt == 0) { 2443 if (old_state != arc_anon && old_state != arc_l2c_only) { 2444 ASSERT(HDR_HAS_L1HDR(hdr)); 2445 multilist_remove(old_state->arcs_list[buftype], hdr); 2446 2447 if (GHOST_STATE(old_state)) { 2448 ASSERT0(bufcnt); 2449 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 2450 update_old = B_TRUE; 2451 } 2452 arc_evictable_space_decrement(hdr, old_state); 2453 } 2454 if (new_state != arc_anon && new_state != arc_l2c_only) { 2455 2456 /* 2457 * An L1 header always exists here, since if we're 2458 * moving to some L1-cached state (i.e. not l2c_only or 2459 * anonymous), we realloc the header to add an L1hdr 2460 * beforehand. 2461 */ 2462 ASSERT(HDR_HAS_L1HDR(hdr)); 2463 multilist_insert(new_state->arcs_list[buftype], hdr); 2464 2465 if (GHOST_STATE(new_state)) { 2466 ASSERT0(bufcnt); 2467 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 2468 update_new = B_TRUE; 2469 } 2470 arc_evictable_space_increment(hdr, new_state); 2471 } 2472 } 2473 2474 ASSERT(!HDR_EMPTY(hdr)); 2475 if (new_state == arc_anon && HDR_IN_HASH_TABLE(hdr)) 2476 buf_hash_remove(hdr); 2477 2478 /* adjust state sizes (ignore arc_l2c_only) */ 2479 2480 if (update_new && new_state != arc_l2c_only) { 2481 ASSERT(HDR_HAS_L1HDR(hdr)); 2482 if (GHOST_STATE(new_state)) { 2483 ASSERT0(bufcnt); 2484 2485 /* 2486 * When moving a header to a ghost state, we first 2487 * remove all arc buffers. Thus, we'll have a 2488 * bufcnt of zero, and no arc buffer to use for 2489 * the reference. As a result, we use the arc 2490 * header pointer for the reference. 2491 */ 2492 (void) refcount_add_many(&new_state->arcs_size, 2493 HDR_GET_LSIZE(hdr), hdr); 2494 ASSERT3P(hdr->b_l1hdr.b_pdata, ==, NULL); 2495 } else { 2496 uint32_t buffers = 0; 2497 2498 /* 2499 * Each individual buffer holds a unique reference, 2500 * thus we must remove each of these references one 2501 * at a time. 2502 */ 2503 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL; 2504 buf = buf->b_next) { 2505 ASSERT3U(bufcnt, !=, 0); 2506 buffers++; 2507 2508 /* 2509 * When the arc_buf_t is sharing the data 2510 * block with the hdr, the owner of the 2511 * reference belongs to the hdr. Only 2512 * add to the refcount if the arc_buf_t is 2513 * not shared. 2514 */ 2515 if (arc_buf_is_shared(buf)) 2516 continue; 2517 2518 (void) refcount_add_many(&new_state->arcs_size, 2519 arc_buf_size(buf), buf); 2520 } 2521 ASSERT3U(bufcnt, ==, buffers); 2522 2523 if (hdr->b_l1hdr.b_pdata != NULL) { 2524 (void) refcount_add_many(&new_state->arcs_size, 2525 arc_hdr_size(hdr), hdr); 2526 } else { 2527 ASSERT(GHOST_STATE(old_state)); 2528 } 2529 } 2530 } 2531 2532 if (update_old && old_state != arc_l2c_only) { 2533 ASSERT(HDR_HAS_L1HDR(hdr)); 2534 if (GHOST_STATE(old_state)) { 2535 ASSERT0(bufcnt); 2536 ASSERT3P(hdr->b_l1hdr.b_pdata, ==, NULL); 2537 2538 /* 2539 * When moving a header off of a ghost state, 2540 * the header will not contain any arc buffers. 2541 * We use the arc header pointer for the reference 2542 * which is exactly what we did when we put the 2543 * header on the ghost state. 2544 */ 2545 2546 (void) refcount_remove_many(&old_state->arcs_size, 2547 HDR_GET_LSIZE(hdr), hdr); 2548 } else { 2549 uint32_t buffers = 0; 2550 2551 /* 2552 * Each individual buffer holds a unique reference, 2553 * thus we must remove each of these references one 2554 * at a time. 2555 */ 2556 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL; 2557 buf = buf->b_next) { 2558 ASSERT3U(bufcnt, !=, 0); 2559 buffers++; 2560 2561 /* 2562 * When the arc_buf_t is sharing the data 2563 * block with the hdr, the owner of the 2564 * reference belongs to the hdr. Only 2565 * add to the refcount if the arc_buf_t is 2566 * not shared. 2567 */ 2568 if (arc_buf_is_shared(buf)) 2569 continue; 2570 2571 (void) refcount_remove_many( 2572 &old_state->arcs_size, arc_buf_size(buf), 2573 buf); 2574 } 2575 ASSERT3U(bufcnt, ==, buffers); 2576 ASSERT3P(hdr->b_l1hdr.b_pdata, !=, NULL); 2577 (void) refcount_remove_many( 2578 &old_state->arcs_size, arc_hdr_size(hdr), hdr); 2579 } 2580 } 2581 2582 if (HDR_HAS_L1HDR(hdr)) 2583 hdr->b_l1hdr.b_state = new_state; 2584 2585 /* 2586 * L2 headers should never be on the L2 state list since they don't 2587 * have L1 headers allocated. 2588 */ 2589 ASSERT(multilist_is_empty(arc_l2c_only->arcs_list[ARC_BUFC_DATA]) && 2590 multilist_is_empty(arc_l2c_only->arcs_list[ARC_BUFC_METADATA])); 2591} 2592 2593void 2594arc_space_consume(uint64_t space, arc_space_type_t type) 2595{ 2596 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES); 2597 2598 switch (type) { 2599 case ARC_SPACE_DATA: 2600 ARCSTAT_INCR(arcstat_data_size, space); 2601 break; 2602 case ARC_SPACE_META: 2603 ARCSTAT_INCR(arcstat_metadata_size, space); 2604 break; 2605 case ARC_SPACE_OTHER: 2606 ARCSTAT_INCR(arcstat_other_size, space); 2607 break; 2608 case ARC_SPACE_HDRS: 2609 ARCSTAT_INCR(arcstat_hdr_size, space); 2610 break; 2611 case ARC_SPACE_L2HDRS: 2612 ARCSTAT_INCR(arcstat_l2_hdr_size, space); 2613 break; 2614 } 2615 2616 if (type != ARC_SPACE_DATA) 2617 ARCSTAT_INCR(arcstat_meta_used, space); 2618 2619 atomic_add_64(&arc_size, space); 2620} 2621 2622void 2623arc_space_return(uint64_t space, arc_space_type_t type) 2624{ 2625 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES); 2626 2627 switch (type) { 2628 case ARC_SPACE_DATA: 2629 ARCSTAT_INCR(arcstat_data_size, -space); 2630 break; 2631 case ARC_SPACE_META: 2632 ARCSTAT_INCR(arcstat_metadata_size, -space); 2633 break; 2634 case ARC_SPACE_OTHER: 2635 ARCSTAT_INCR(arcstat_other_size, -space); 2636 break; 2637 case ARC_SPACE_HDRS: 2638 ARCSTAT_INCR(arcstat_hdr_size, -space); 2639 break; 2640 case ARC_SPACE_L2HDRS: 2641 ARCSTAT_INCR(arcstat_l2_hdr_size, -space); 2642 break; 2643 } 2644 2645 if (type != ARC_SPACE_DATA) { 2646 ASSERT(arc_meta_used >= space); 2647 if (arc_meta_max < arc_meta_used) 2648 arc_meta_max = arc_meta_used; 2649 ARCSTAT_INCR(arcstat_meta_used, -space); 2650 } 2651 2652 ASSERT(arc_size >= space); 2653 atomic_add_64(&arc_size, -space); 2654} 2655 2656/* 2657 * Given a hdr and a buf, returns whether that buf can share its b_data buffer 2658 * with the hdr's b_pdata. 2659 */ 2660static boolean_t 2661arc_can_share(arc_buf_hdr_t *hdr, arc_buf_t *buf) 2662{ 2663 /* 2664 * The criteria for sharing a hdr's data are: 2665 * 1. the hdr's compression matches the buf's compression 2666 * 2. the hdr doesn't need to be byteswapped 2667 * 3. the hdr isn't already being shared 2668 * 4. the buf is either compressed or it is the last buf in the hdr list 2669 * 2670 * Criterion #4 maintains the invariant that shared uncompressed 2671 * bufs must be the final buf in the hdr's b_buf list. Reading this, you 2672 * might ask, "if a compressed buf is allocated first, won't that be the 2673 * last thing in the list?", but in that case it's impossible to create 2674 * a shared uncompressed buf anyway (because the hdr must be compressed 2675 * to have the compressed buf). You might also think that #3 is 2676 * sufficient to make this guarantee, however it's possible 2677 * (specifically in the rare L2ARC write race mentioned in 2678 * arc_buf_alloc_impl()) there will be an existing uncompressed buf that 2679 * is sharable, but wasn't at the time of its allocation. Rather than 2680 * allow a new shared uncompressed buf to be created and then shuffle 2681 * the list around to make it the last element, this simply disallows 2682 * sharing if the new buf isn't the first to be added. 2683 */ 2684 ASSERT3P(buf->b_hdr, ==, hdr); 2685 boolean_t hdr_compressed = HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF; 2686 boolean_t buf_compressed = ARC_BUF_COMPRESSED(buf) != 0; 2687 return (buf_compressed == hdr_compressed && 2688 hdr->b_l1hdr.b_byteswap == DMU_BSWAP_NUMFUNCS && 2689 !HDR_SHARED_DATA(hdr) && 2690 (ARC_BUF_LAST(buf) || ARC_BUF_COMPRESSED(buf))); 2691} 2692 2693/* 2694 * Allocate a buf for this hdr. If you care about the data that's in the hdr, 2695 * or if you want a compressed buffer, pass those flags in. Returns 0 if the 2696 * copy was made successfully, or an error code otherwise. 2697 */ 2698static int 2699arc_buf_alloc_impl(arc_buf_hdr_t *hdr, void *tag, boolean_t compressed, 2700 boolean_t fill, arc_buf_t **ret) 2701{ 2702 arc_buf_t *buf; 2703 2704 ASSERT(HDR_HAS_L1HDR(hdr)); 2705 ASSERT3U(HDR_GET_LSIZE(hdr), >, 0); 2706 VERIFY(hdr->b_type == ARC_BUFC_DATA || 2707 hdr->b_type == ARC_BUFC_METADATA); 2708 ASSERT3P(ret, !=, NULL); 2709 ASSERT3P(*ret, ==, NULL); 2710 2711 buf = *ret = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 2712 buf->b_hdr = hdr; 2713 buf->b_data = NULL; 2714 buf->b_next = hdr->b_l1hdr.b_buf; 2715 buf->b_flags = 0; 2716 2717 add_reference(hdr, tag); 2718 2719 /* 2720 * We're about to change the hdr's b_flags. We must either 2721 * hold the hash_lock or be undiscoverable. 2722 */ 2723 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); 2724 2725 /* 2726 * Only honor requests for compressed bufs if the hdr is actually 2727 * compressed. 2728 */ 2729 if (compressed && HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF) 2730 buf->b_flags |= ARC_BUF_FLAG_COMPRESSED; 2731 2732 /* 2733 * If the hdr's data can be shared then we share the data buffer and 2734 * set the appropriate bit in the hdr's b_flags to indicate the hdr is 2735 * sharing it's b_pdata with the arc_buf_t. Otherwise, we allocate a new 2736 * buffer to store the buf's data. 2737 * 2738 * There is one additional restriction here because we're sharing 2739 * hdr -> buf instead of the usual buf -> hdr: the hdr can't be actively 2740 * involved in an L2ARC write, because if this buf is used by an 2741 * arc_write() then the hdr's data buffer will be released when the 2742 * write completes, even though the L2ARC write might still be using it. 2743 */ 2744 boolean_t can_share = arc_can_share(hdr, buf) && !HDR_L2_WRITING(hdr); 2745 2746 /* Set up b_data and sharing */ 2747 if (can_share) { 2748 buf->b_data = hdr->b_l1hdr.b_pdata; 2749 buf->b_flags |= ARC_BUF_FLAG_SHARED; 2750 arc_hdr_set_flags(hdr, ARC_FLAG_SHARED_DATA); 2751 } else { 2752 buf->b_data = 2753 arc_get_data_buf(hdr, arc_buf_size(buf), buf); 2754 ARCSTAT_INCR(arcstat_overhead_size, arc_buf_size(buf)); 2755 } 2756 VERIFY3P(buf->b_data, !=, NULL); 2757 2758 hdr->b_l1hdr.b_buf = buf; 2759 hdr->b_l1hdr.b_bufcnt += 1; 2760 2761 /* 2762 * If the user wants the data from the hdr, we need to either copy or 2763 * decompress the data. 2764 */ 2765 if (fill) { 2766 return (arc_buf_fill(buf, ARC_BUF_COMPRESSED(buf) != 0)); 2767 } 2768 2769 return (0); 2770} 2771 2772static char *arc_onloan_tag = "onloan"; 2773 2774static inline void 2775arc_loaned_bytes_update(int64_t delta) 2776{ 2777 atomic_add_64(&arc_loaned_bytes, delta); 2778 2779 /* assert that it did not wrap around */ 2780 ASSERT3S(atomic_add_64_nv(&arc_loaned_bytes, 0), >=, 0); 2781} 2782 2783/* 2784 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in 2785 * flight data by arc_tempreserve_space() until they are "returned". Loaned 2786 * buffers must be returned to the arc before they can be used by the DMU or 2787 * freed. 2788 */ 2789arc_buf_t * 2790arc_loan_buf(spa_t *spa, boolean_t is_metadata, int size) 2791{ 2792 arc_buf_t *buf = arc_alloc_buf(spa, arc_onloan_tag, 2793 is_metadata ? ARC_BUFC_METADATA : ARC_BUFC_DATA, size); 2794 2795 arc_loaned_bytes_update(size); 2796 2797 return (buf); 2798} 2799 2800arc_buf_t * 2801arc_loan_compressed_buf(spa_t *spa, uint64_t psize, uint64_t lsize, 2802 enum zio_compress compression_type) 2803{ 2804 arc_buf_t *buf = arc_alloc_compressed_buf(spa, arc_onloan_tag, 2805 psize, lsize, compression_type); 2806 2807 arc_loaned_bytes_update(psize); 2808 2809 return (buf); 2810} 2811 2812 2813/* 2814 * Return a loaned arc buffer to the arc. 2815 */ 2816void 2817arc_return_buf(arc_buf_t *buf, void *tag) 2818{ 2819 arc_buf_hdr_t *hdr = buf->b_hdr; 2820 2821 ASSERT3P(buf->b_data, !=, NULL); 2822 ASSERT(HDR_HAS_L1HDR(hdr)); 2823 (void) refcount_add(&hdr->b_l1hdr.b_refcnt, tag); 2824 (void) refcount_remove(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag); 2825 2826 arc_loaned_bytes_update(-arc_buf_size(buf)); 2827} 2828 2829/* Detach an arc_buf from a dbuf (tag) */ 2830void 2831arc_loan_inuse_buf(arc_buf_t *buf, void *tag) 2832{ 2833 arc_buf_hdr_t *hdr = buf->b_hdr; 2834 2835 ASSERT3P(buf->b_data, !=, NULL); 2836 ASSERT(HDR_HAS_L1HDR(hdr)); 2837 (void) refcount_add(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag); 2838 (void) refcount_remove(&hdr->b_l1hdr.b_refcnt, tag); 2839 2840 arc_loaned_bytes_update(arc_buf_size(buf)); 2841} 2842 2843static void 2844l2arc_free_data_on_write(void *data, size_t size, arc_buf_contents_t type) 2845{ 2846 l2arc_data_free_t *df = kmem_alloc(sizeof (*df), KM_SLEEP); 2847 2848 df->l2df_data = data; 2849 df->l2df_size = size; 2850 df->l2df_type = type; 2851 mutex_enter(&l2arc_free_on_write_mtx); 2852 list_insert_head(l2arc_free_on_write, df); 2853 mutex_exit(&l2arc_free_on_write_mtx); 2854} 2855 2856static void 2857arc_hdr_free_on_write(arc_buf_hdr_t *hdr) 2858{ 2859 arc_state_t *state = hdr->b_l1hdr.b_state; 2860 arc_buf_contents_t type = arc_buf_type(hdr); 2861 uint64_t size = arc_hdr_size(hdr); 2862 2863 /* protected by hash lock, if in the hash table */ 2864 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) { 2865 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 2866 ASSERT(state != arc_anon && state != arc_l2c_only); 2867 2868 (void) refcount_remove_many(&state->arcs_esize[type], 2869 size, hdr); 2870 } 2871 (void) refcount_remove_many(&state->arcs_size, size, hdr); 2872 if (type == ARC_BUFC_METADATA) { 2873 arc_space_return(size, ARC_SPACE_META); 2874 } else { 2875 ASSERT(type == ARC_BUFC_DATA); 2876 arc_space_return(size, ARC_SPACE_DATA); 2877 } 2878 2879 l2arc_free_data_on_write(hdr->b_l1hdr.b_pdata, size, type); 2880} 2881 2882/* 2883 * Share the arc_buf_t's data with the hdr. Whenever we are sharing the 2884 * data buffer, we transfer the refcount ownership to the hdr and update 2885 * the appropriate kstats. 2886 */ 2887static void 2888arc_share_buf(arc_buf_hdr_t *hdr, arc_buf_t *buf) 2889{ 2890 arc_state_t *state = hdr->b_l1hdr.b_state; 2891 2892 ASSERT(arc_can_share(hdr, buf)); 2893 ASSERT3P(hdr->b_l1hdr.b_pdata, ==, NULL); 2894 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); 2895 2896 /* 2897 * Start sharing the data buffer. We transfer the 2898 * refcount ownership to the hdr since it always owns 2899 * the refcount whenever an arc_buf_t is shared. 2900 */ 2901 refcount_transfer_ownership(&state->arcs_size, buf, hdr); 2902 hdr->b_l1hdr.b_pdata = buf->b_data; 2903 arc_hdr_set_flags(hdr, ARC_FLAG_SHARED_DATA); 2904 buf->b_flags |= ARC_BUF_FLAG_SHARED; 2905 2906 /* 2907 * Since we've transferred ownership to the hdr we need 2908 * to increment its compressed and uncompressed kstats and 2909 * decrement the overhead size. 2910 */ 2911 ARCSTAT_INCR(arcstat_compressed_size, arc_hdr_size(hdr)); 2912 ARCSTAT_INCR(arcstat_uncompressed_size, HDR_GET_LSIZE(hdr)); 2913 ARCSTAT_INCR(arcstat_overhead_size, -arc_buf_size(buf)); 2914} 2915 2916static void 2917arc_unshare_buf(arc_buf_hdr_t *hdr, arc_buf_t *buf) 2918{ 2919 arc_state_t *state = hdr->b_l1hdr.b_state; 2920 2921 ASSERT(arc_buf_is_shared(buf)); 2922 ASSERT3P(hdr->b_l1hdr.b_pdata, !=, NULL); 2923 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); 2924 2925 /* 2926 * We are no longer sharing this buffer so we need 2927 * to transfer its ownership to the rightful owner. 2928 */ 2929 refcount_transfer_ownership(&state->arcs_size, hdr, buf); 2930 arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA); 2931 hdr->b_l1hdr.b_pdata = NULL; 2932 buf->b_flags &= ~ARC_BUF_FLAG_SHARED; 2933 2934 /* 2935 * Since the buffer is no longer shared between 2936 * the arc buf and the hdr, count it as overhead. 2937 */ 2938 ARCSTAT_INCR(arcstat_compressed_size, -arc_hdr_size(hdr)); 2939 ARCSTAT_INCR(arcstat_uncompressed_size, -HDR_GET_LSIZE(hdr)); 2940 ARCSTAT_INCR(arcstat_overhead_size, arc_buf_size(buf)); 2941} 2942 2943/* 2944 * Remove an arc_buf_t from the hdr's buf list and return the last 2945 * arc_buf_t on the list. If no buffers remain on the list then return 2946 * NULL. 2947 */ 2948static arc_buf_t * 2949arc_buf_remove(arc_buf_hdr_t *hdr, arc_buf_t *buf) 2950{ 2951 ASSERT(HDR_HAS_L1HDR(hdr)); 2952 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); 2953 2954 arc_buf_t **bufp = &hdr->b_l1hdr.b_buf; 2955 arc_buf_t *lastbuf = NULL; 2956 2957 /* 2958 * Remove the buf from the hdr list and locate the last 2959 * remaining buffer on the list. 2960 */ 2961 while (*bufp != NULL) { 2962 if (*bufp == buf) 2963 *bufp = buf->b_next; 2964 2965 /* 2966 * If we've removed a buffer in the middle of 2967 * the list then update the lastbuf and update 2968 * bufp. 2969 */ 2970 if (*bufp != NULL) { 2971 lastbuf = *bufp; 2972 bufp = &(*bufp)->b_next; 2973 } 2974 } 2975 buf->b_next = NULL; 2976 ASSERT3P(lastbuf, !=, buf); 2977 IMPLY(hdr->b_l1hdr.b_bufcnt > 0, lastbuf != NULL); 2978 IMPLY(hdr->b_l1hdr.b_bufcnt > 0, hdr->b_l1hdr.b_buf != NULL); 2979 IMPLY(lastbuf != NULL, ARC_BUF_LAST(lastbuf)); 2980 2981 return (lastbuf); 2982} 2983 2984/* 2985 * Free up buf->b_data and pull the arc_buf_t off of the the arc_buf_hdr_t's 2986 * list and free it. 2987 */ 2988static void 2989arc_buf_destroy_impl(arc_buf_t *buf) 2990{ 2991 arc_buf_hdr_t *hdr = buf->b_hdr; 2992 2993 /* 2994 * Free up the data associated with the buf but only if we're not 2995 * sharing this with the hdr. If we are sharing it with the hdr, the 2996 * hdr is responsible for doing the free. 2997 */ 2998 if (buf->b_data != NULL) { 2999 /* 3000 * We're about to change the hdr's b_flags. We must either 3001 * hold the hash_lock or be undiscoverable. 3002 */ 3003 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); 3004 3005 arc_cksum_verify(buf); 3006#ifdef illumos 3007 arc_buf_unwatch(buf); 3008#endif 3009 3010 if (arc_buf_is_shared(buf)) { 3011 arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA); 3012 } else { 3013 uint64_t size = arc_buf_size(buf); 3014 arc_free_data_buf(hdr, buf->b_data, size, buf); 3015 ARCSTAT_INCR(arcstat_overhead_size, -size); 3016 } 3017 buf->b_data = NULL; 3018 3019 ASSERT(hdr->b_l1hdr.b_bufcnt > 0); 3020 hdr->b_l1hdr.b_bufcnt -= 1; 3021 } 3022 3023 arc_buf_t *lastbuf = arc_buf_remove(hdr, buf); 3024 3025 if (ARC_BUF_SHARED(buf) && !ARC_BUF_COMPRESSED(buf)) { 3026 /* 3027 * If the current arc_buf_t is sharing its data buffer with the 3028 * hdr, then reassign the hdr's b_pdata to share it with the new 3029 * buffer at the end of the list. The shared buffer is always 3030 * the last one on the hdr's buffer list. 3031 * 3032 * There is an equivalent case for compressed bufs, but since 3033 * they aren't guaranteed to be the last buf in the list and 3034 * that is an exceedingly rare case, we just allow that space be 3035 * wasted temporarily. 3036 */ 3037 if (lastbuf != NULL) { 3038 /* Only one buf can be shared at once */ 3039 VERIFY(!arc_buf_is_shared(lastbuf)); 3040 /* hdr is uncompressed so can't have compressed buf */ 3041 VERIFY(!ARC_BUF_COMPRESSED(lastbuf)); 3042 3043 ASSERT3P(hdr->b_l1hdr.b_pdata, !=, NULL); 3044 arc_hdr_free_pdata(hdr); 3045 3046 /* 3047 * We must setup a new shared block between the 3048 * last buffer and the hdr. The data would have 3049 * been allocated by the arc buf so we need to transfer 3050 * ownership to the hdr since it's now being shared. 3051 */ 3052 arc_share_buf(hdr, lastbuf); 3053 } 3054 } else if (HDR_SHARED_DATA(hdr)) { 3055 /* 3056 * Uncompressed shared buffers are always at the end 3057 * of the list. Compressed buffers don't have the 3058 * same requirements. This makes it hard to 3059 * simply assert that the lastbuf is shared so 3060 * we rely on the hdr's compression flags to determine 3061 * if we have a compressed, shared buffer. 3062 */ 3063 ASSERT3P(lastbuf, !=, NULL); 3064 ASSERT(arc_buf_is_shared(lastbuf) || 3065 HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF); 3066 } 3067 3068 /* 3069 * Free the checksum if we're removing the last uncompressed buf from 3070 * this hdr. 3071 */ 3072 if (!arc_hdr_has_uncompressed_buf(hdr)) { 3073 arc_cksum_free(hdr); 3074 } 3075 3076 /* clean up the buf */ 3077 buf->b_hdr = NULL; 3078 kmem_cache_free(buf_cache, buf); 3079} 3080 3081static void 3082arc_hdr_alloc_pdata(arc_buf_hdr_t *hdr) 3083{ 3084 ASSERT3U(HDR_GET_LSIZE(hdr), >, 0); 3085 ASSERT(HDR_HAS_L1HDR(hdr)); 3086 ASSERT(!HDR_SHARED_DATA(hdr)); 3087 3088 ASSERT3P(hdr->b_l1hdr.b_pdata, ==, NULL); 3089 hdr->b_l1hdr.b_pdata = arc_get_data_buf(hdr, arc_hdr_size(hdr), hdr); 3090 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS; 3091 ASSERT3P(hdr->b_l1hdr.b_pdata, !=, NULL); 3092 3093 ARCSTAT_INCR(arcstat_compressed_size, arc_hdr_size(hdr)); 3094 ARCSTAT_INCR(arcstat_uncompressed_size, HDR_GET_LSIZE(hdr)); 3095} 3096 3097static void 3098arc_hdr_free_pdata(arc_buf_hdr_t *hdr) 3099{ 3100 ASSERT(HDR_HAS_L1HDR(hdr)); 3101 ASSERT3P(hdr->b_l1hdr.b_pdata, !=, NULL); 3102 3103 /* 3104 * If the hdr is currently being written to the l2arc then 3105 * we defer freeing the data by adding it to the l2arc_free_on_write 3106 * list. The l2arc will free the data once it's finished 3107 * writing it to the l2arc device. 3108 */ 3109 if (HDR_L2_WRITING(hdr)) { 3110 arc_hdr_free_on_write(hdr); 3111 ARCSTAT_BUMP(arcstat_l2_free_on_write); 3112 } else { 3113 arc_free_data_buf(hdr, hdr->b_l1hdr.b_pdata, 3114 arc_hdr_size(hdr), hdr); 3115 } 3116 hdr->b_l1hdr.b_pdata = NULL; 3117 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS; 3118 3119 ARCSTAT_INCR(arcstat_compressed_size, -arc_hdr_size(hdr)); 3120 ARCSTAT_INCR(arcstat_uncompressed_size, -HDR_GET_LSIZE(hdr)); 3121} 3122 3123static arc_buf_hdr_t * 3124arc_hdr_alloc(uint64_t spa, int32_t psize, int32_t lsize, 3125 enum zio_compress compression_type, arc_buf_contents_t type) 3126{ 3127 arc_buf_hdr_t *hdr; 3128 3129 VERIFY(type == ARC_BUFC_DATA || type == ARC_BUFC_METADATA); 3130 3131 hdr = kmem_cache_alloc(hdr_full_cache, KM_PUSHPAGE); 3132 ASSERT(HDR_EMPTY(hdr)); 3133 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL); 3134 ASSERT3P(hdr->b_l1hdr.b_thawed, ==, NULL); 3135 HDR_SET_PSIZE(hdr, psize); 3136 HDR_SET_LSIZE(hdr, lsize); 3137 hdr->b_spa = spa; 3138 hdr->b_type = type; 3139 hdr->b_flags = 0; 3140 arc_hdr_set_flags(hdr, arc_bufc_to_flags(type) | ARC_FLAG_HAS_L1HDR); 3141 arc_hdr_set_compress(hdr, compression_type); 3142 3143 hdr->b_l1hdr.b_state = arc_anon; 3144 hdr->b_l1hdr.b_arc_access = 0; 3145 hdr->b_l1hdr.b_bufcnt = 0; 3146 hdr->b_l1hdr.b_buf = NULL; 3147 3148 /* 3149 * Allocate the hdr's buffer. This will contain either 3150 * the compressed or uncompressed data depending on the block 3151 * it references and compressed arc enablement. 3152 */ 3153 arc_hdr_alloc_pdata(hdr); 3154 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 3155 3156 return (hdr); 3157} 3158 3159/* 3160 * Transition between the two allocation states for the arc_buf_hdr struct. 3161 * The arc_buf_hdr struct can be allocated with (hdr_full_cache) or without 3162 * (hdr_l2only_cache) the fields necessary for the L1 cache - the smaller 3163 * version is used when a cache buffer is only in the L2ARC in order to reduce 3164 * memory usage. 3165 */ 3166static arc_buf_hdr_t * 3167arc_hdr_realloc(arc_buf_hdr_t *hdr, kmem_cache_t *old, kmem_cache_t *new) 3168{ 3169 ASSERT(HDR_HAS_L2HDR(hdr)); 3170 3171 arc_buf_hdr_t *nhdr; 3172 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev; 3173 3174 ASSERT((old == hdr_full_cache && new == hdr_l2only_cache) || 3175 (old == hdr_l2only_cache && new == hdr_full_cache)); 3176 3177 nhdr = kmem_cache_alloc(new, KM_PUSHPAGE); 3178 3179 ASSERT(MUTEX_HELD(HDR_LOCK(hdr))); 3180 buf_hash_remove(hdr); 3181 3182 bcopy(hdr, nhdr, HDR_L2ONLY_SIZE); 3183 3184 if (new == hdr_full_cache) { 3185 arc_hdr_set_flags(nhdr, ARC_FLAG_HAS_L1HDR); 3186 /* 3187 * arc_access and arc_change_state need to be aware that a 3188 * header has just come out of L2ARC, so we set its state to 3189 * l2c_only even though it's about to change. 3190 */ 3191 nhdr->b_l1hdr.b_state = arc_l2c_only; 3192 3193 /* Verify previous threads set to NULL before freeing */ 3194 ASSERT3P(nhdr->b_l1hdr.b_pdata, ==, NULL); 3195 } else { 3196 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 3197 ASSERT0(hdr->b_l1hdr.b_bufcnt); 3198 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL); 3199 3200 /* 3201 * If we've reached here, We must have been called from 3202 * arc_evict_hdr(), as such we should have already been 3203 * removed from any ghost list we were previously on 3204 * (which protects us from racing with arc_evict_state), 3205 * thus no locking is needed during this check. 3206 */ 3207 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 3208 3209 /* 3210 * A buffer must not be moved into the arc_l2c_only 3211 * state if it's not finished being written out to the 3212 * l2arc device. Otherwise, the b_l1hdr.b_pdata field 3213 * might try to be accessed, even though it was removed. 3214 */ 3215 VERIFY(!HDR_L2_WRITING(hdr)); 3216 VERIFY3P(hdr->b_l1hdr.b_pdata, ==, NULL); 3217 3218#ifdef ZFS_DEBUG 3219 if (hdr->b_l1hdr.b_thawed != NULL) { 3220 kmem_free(hdr->b_l1hdr.b_thawed, 1); 3221 hdr->b_l1hdr.b_thawed = NULL; 3222 } 3223#endif 3224 3225 arc_hdr_clear_flags(nhdr, ARC_FLAG_HAS_L1HDR); 3226 } 3227 /* 3228 * The header has been reallocated so we need to re-insert it into any 3229 * lists it was on. 3230 */ 3231 (void) buf_hash_insert(nhdr, NULL); 3232 3233 ASSERT(list_link_active(&hdr->b_l2hdr.b_l2node)); 3234 3235 mutex_enter(&dev->l2ad_mtx); 3236 3237 /* 3238 * We must place the realloc'ed header back into the list at 3239 * the same spot. Otherwise, if it's placed earlier in the list, 3240 * l2arc_write_buffers() could find it during the function's 3241 * write phase, and try to write it out to the l2arc. 3242 */ 3243 list_insert_after(&dev->l2ad_buflist, hdr, nhdr); 3244 list_remove(&dev->l2ad_buflist, hdr); 3245 3246 mutex_exit(&dev->l2ad_mtx); 3247 3248 /* 3249 * Since we're using the pointer address as the tag when 3250 * incrementing and decrementing the l2ad_alloc refcount, we 3251 * must remove the old pointer (that we're about to destroy) and 3252 * add the new pointer to the refcount. Otherwise we'd remove 3253 * the wrong pointer address when calling arc_hdr_destroy() later. 3254 */ 3255 3256 (void) refcount_remove_many(&dev->l2ad_alloc, arc_hdr_size(hdr), hdr); 3257 (void) refcount_add_many(&dev->l2ad_alloc, arc_hdr_size(nhdr), nhdr); 3258 3259 buf_discard_identity(hdr); 3260 kmem_cache_free(old, hdr); 3261 3262 return (nhdr); 3263} 3264 3265/* 3266 * Allocate a new arc_buf_hdr_t and arc_buf_t and return the buf to the caller. 3267 * The buf is returned thawed since we expect the consumer to modify it. 3268 */ 3269arc_buf_t * 3270arc_alloc_buf(spa_t *spa, void *tag, arc_buf_contents_t type, int32_t size) 3271{ 3272 arc_buf_hdr_t *hdr = arc_hdr_alloc(spa_load_guid(spa), size, size, 3273 ZIO_COMPRESS_OFF, type); 3274 ASSERT(!MUTEX_HELD(HDR_LOCK(hdr))); 3275 3276 arc_buf_t *buf = NULL; 3277 VERIFY0(arc_buf_alloc_impl(hdr, tag, B_FALSE, B_FALSE, &buf)); 3278 arc_buf_thaw(buf); 3279 3280 return (buf); 3281} 3282 3283/* 3284 * Allocate a compressed buf in the same manner as arc_alloc_buf. Don't use this 3285 * for bufs containing metadata. 3286 */ 3287arc_buf_t * 3288arc_alloc_compressed_buf(spa_t *spa, void *tag, uint64_t psize, uint64_t lsize, 3289 enum zio_compress compression_type) 3290{ 3291 ASSERT3U(lsize, >, 0); 3292 ASSERT3U(lsize, >=, psize); 3293 ASSERT(compression_type > ZIO_COMPRESS_OFF); 3294 ASSERT(compression_type < ZIO_COMPRESS_FUNCTIONS); 3295 3296 arc_buf_hdr_t *hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize, 3297 compression_type, ARC_BUFC_DATA); 3298 ASSERT(!MUTEX_HELD(HDR_LOCK(hdr))); 3299 3300 arc_buf_t *buf = NULL; 3301 VERIFY0(arc_buf_alloc_impl(hdr, tag, B_TRUE, B_FALSE, &buf)); 3302 arc_buf_thaw(buf); 3303 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL); 3304 3305 return (buf); 3306} 3307 3308static void 3309arc_hdr_l2hdr_destroy(arc_buf_hdr_t *hdr) 3310{ 3311 l2arc_buf_hdr_t *l2hdr = &hdr->b_l2hdr; 3312 l2arc_dev_t *dev = l2hdr->b_dev; 3313 uint64_t asize = arc_hdr_size(hdr); 3314 3315 ASSERT(MUTEX_HELD(&dev->l2ad_mtx)); 3316 ASSERT(HDR_HAS_L2HDR(hdr)); 3317 3318 list_remove(&dev->l2ad_buflist, hdr); 3319 3320 ARCSTAT_INCR(arcstat_l2_asize, -asize); 3321 ARCSTAT_INCR(arcstat_l2_size, -HDR_GET_LSIZE(hdr)); 3322 3323 vdev_space_update(dev->l2ad_vdev, -asize, 0, 0); 3324 3325 (void) refcount_remove_many(&dev->l2ad_alloc, asize, hdr); 3326 arc_hdr_clear_flags(hdr, ARC_FLAG_HAS_L2HDR); 3327} 3328 3329static void 3330arc_hdr_destroy(arc_buf_hdr_t *hdr) 3331{ 3332 if (HDR_HAS_L1HDR(hdr)) { 3333 ASSERT(hdr->b_l1hdr.b_buf == NULL || 3334 hdr->b_l1hdr.b_bufcnt > 0); 3335 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 3336 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon); 3337 } 3338 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 3339 ASSERT(!HDR_IN_HASH_TABLE(hdr)); 3340 3341 if (!HDR_EMPTY(hdr)) 3342 buf_discard_identity(hdr); 3343 3344 if (HDR_HAS_L2HDR(hdr)) { 3345 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev; 3346 boolean_t buflist_held = MUTEX_HELD(&dev->l2ad_mtx); 3347 3348 if (!buflist_held) 3349 mutex_enter(&dev->l2ad_mtx); 3350 3351 /* 3352 * Even though we checked this conditional above, we 3353 * need to check this again now that we have the 3354 * l2ad_mtx. This is because we could be racing with 3355 * another thread calling l2arc_evict() which might have 3356 * destroyed this header's L2 portion as we were waiting 3357 * to acquire the l2ad_mtx. If that happens, we don't 3358 * want to re-destroy the header's L2 portion. 3359 */ 3360 if (HDR_HAS_L2HDR(hdr)) { 3361 l2arc_trim(hdr); 3362 arc_hdr_l2hdr_destroy(hdr); 3363 } 3364 3365 if (!buflist_held) 3366 mutex_exit(&dev->l2ad_mtx); 3367 } 3368 3369 if (HDR_HAS_L1HDR(hdr)) { 3370 arc_cksum_free(hdr); 3371 3372 while (hdr->b_l1hdr.b_buf != NULL) 3373 arc_buf_destroy_impl(hdr->b_l1hdr.b_buf); 3374 3375#ifdef ZFS_DEBUG 3376 if (hdr->b_l1hdr.b_thawed != NULL) { 3377 kmem_free(hdr->b_l1hdr.b_thawed, 1); 3378 hdr->b_l1hdr.b_thawed = NULL; 3379 } 3380#endif 3381 3382 if (hdr->b_l1hdr.b_pdata != NULL) { 3383 arc_hdr_free_pdata(hdr); 3384 } 3385 } 3386 3387 ASSERT3P(hdr->b_hash_next, ==, NULL); 3388 if (HDR_HAS_L1HDR(hdr)) { 3389 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 3390 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL); 3391 kmem_cache_free(hdr_full_cache, hdr); 3392 } else { 3393 kmem_cache_free(hdr_l2only_cache, hdr); 3394 } 3395} 3396 3397void 3398arc_buf_destroy(arc_buf_t *buf, void* tag) 3399{ 3400 arc_buf_hdr_t *hdr = buf->b_hdr; 3401 kmutex_t *hash_lock = HDR_LOCK(hdr); 3402 3403 if (hdr->b_l1hdr.b_state == arc_anon) { 3404 ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1); 3405 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 3406 VERIFY0(remove_reference(hdr, NULL, tag)); 3407 arc_hdr_destroy(hdr); 3408 return; 3409 } 3410 3411 mutex_enter(hash_lock); 3412 ASSERT3P(hdr, ==, buf->b_hdr); 3413 ASSERT(hdr->b_l1hdr.b_bufcnt > 0); 3414 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 3415 ASSERT3P(hdr->b_l1hdr.b_state, !=, arc_anon); 3416 ASSERT3P(buf->b_data, !=, NULL); 3417 3418 (void) remove_reference(hdr, hash_lock, tag); 3419 arc_buf_destroy_impl(buf); 3420 mutex_exit(hash_lock); 3421} 3422 3423/* 3424 * Evict the arc_buf_hdr that is provided as a parameter. The resultant 3425 * state of the header is dependent on it's state prior to entering this 3426 * function. The following transitions are possible: 3427 * 3428 * - arc_mru -> arc_mru_ghost 3429 * - arc_mfu -> arc_mfu_ghost 3430 * - arc_mru_ghost -> arc_l2c_only 3431 * - arc_mru_ghost -> deleted 3432 * - arc_mfu_ghost -> arc_l2c_only 3433 * - arc_mfu_ghost -> deleted 3434 */ 3435static int64_t 3436arc_evict_hdr(arc_buf_hdr_t *hdr, kmutex_t *hash_lock) 3437{ 3438 arc_state_t *evicted_state, *state; 3439 int64_t bytes_evicted = 0; 3440 3441 ASSERT(MUTEX_HELD(hash_lock)); 3442 ASSERT(HDR_HAS_L1HDR(hdr)); 3443 3444 state = hdr->b_l1hdr.b_state; 3445 if (GHOST_STATE(state)) { 3446 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 3447 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 3448 3449 /* 3450 * l2arc_write_buffers() relies on a header's L1 portion 3451 * (i.e. its b_pdata field) during its write phase. 3452 * Thus, we cannot push a header onto the arc_l2c_only 3453 * state (removing it's L1 piece) until the header is 3454 * done being written to the l2arc. 3455 */ 3456 if (HDR_HAS_L2HDR(hdr) && HDR_L2_WRITING(hdr)) { 3457 ARCSTAT_BUMP(arcstat_evict_l2_skip); 3458 return (bytes_evicted); 3459 } 3460 3461 ARCSTAT_BUMP(arcstat_deleted); 3462 bytes_evicted += HDR_GET_LSIZE(hdr); 3463 3464 DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, hdr); 3465 3466 ASSERT3P(hdr->b_l1hdr.b_pdata, ==, NULL); 3467 if (HDR_HAS_L2HDR(hdr)) { 3468 /* 3469 * This buffer is cached on the 2nd Level ARC; 3470 * don't destroy the header. 3471 */ 3472 arc_change_state(arc_l2c_only, hdr, hash_lock); 3473 /* 3474 * dropping from L1+L2 cached to L2-only, 3475 * realloc to remove the L1 header. 3476 */ 3477 hdr = arc_hdr_realloc(hdr, hdr_full_cache, 3478 hdr_l2only_cache); 3479 } else { 3480 arc_change_state(arc_anon, hdr, hash_lock); 3481 arc_hdr_destroy(hdr); 3482 } 3483 return (bytes_evicted); 3484 } 3485 3486 ASSERT(state == arc_mru || state == arc_mfu); 3487 evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost; 3488 3489 /* prefetch buffers have a minimum lifespan */ 3490 if (HDR_IO_IN_PROGRESS(hdr) || 3491 ((hdr->b_flags & (ARC_FLAG_PREFETCH | ARC_FLAG_INDIRECT)) && 3492 ddi_get_lbolt() - hdr->b_l1hdr.b_arc_access < 3493 arc_min_prefetch_lifespan)) { 3494 ARCSTAT_BUMP(arcstat_evict_skip); 3495 return (bytes_evicted); 3496 } 3497 3498 ASSERT0(refcount_count(&hdr->b_l1hdr.b_refcnt)); 3499 while (hdr->b_l1hdr.b_buf) { 3500 arc_buf_t *buf = hdr->b_l1hdr.b_buf; 3501 if (!mutex_tryenter(&buf->b_evict_lock)) { 3502 ARCSTAT_BUMP(arcstat_mutex_miss); 3503 break; 3504 } 3505 if (buf->b_data != NULL) 3506 bytes_evicted += HDR_GET_LSIZE(hdr); 3507 mutex_exit(&buf->b_evict_lock); 3508 arc_buf_destroy_impl(buf); 3509 } 3510 3511 if (HDR_HAS_L2HDR(hdr)) { 3512 ARCSTAT_INCR(arcstat_evict_l2_cached, HDR_GET_LSIZE(hdr)); 3513 } else { 3514 if (l2arc_write_eligible(hdr->b_spa, hdr)) { 3515 ARCSTAT_INCR(arcstat_evict_l2_eligible, 3516 HDR_GET_LSIZE(hdr)); 3517 } else { 3518 ARCSTAT_INCR(arcstat_evict_l2_ineligible, 3519 HDR_GET_LSIZE(hdr)); 3520 } 3521 } 3522 3523 if (hdr->b_l1hdr.b_bufcnt == 0) { 3524 arc_cksum_free(hdr); 3525 3526 bytes_evicted += arc_hdr_size(hdr); 3527 3528 /* 3529 * If this hdr is being evicted and has a compressed 3530 * buffer then we discard it here before we change states. 3531 * This ensures that the accounting is updated correctly 3532 * in arc_free_data_buf(). 3533 */ 3534 arc_hdr_free_pdata(hdr); 3535 3536 arc_change_state(evicted_state, hdr, hash_lock); 3537 ASSERT(HDR_IN_HASH_TABLE(hdr)); 3538 arc_hdr_set_flags(hdr, ARC_FLAG_IN_HASH_TABLE); 3539 DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, hdr); 3540 } 3541 3542 return (bytes_evicted); 3543} 3544 3545static uint64_t 3546arc_evict_state_impl(multilist_t *ml, int idx, arc_buf_hdr_t *marker, 3547 uint64_t spa, int64_t bytes) 3548{ 3549 multilist_sublist_t *mls; 3550 uint64_t bytes_evicted = 0; 3551 arc_buf_hdr_t *hdr; 3552 kmutex_t *hash_lock; 3553 int evict_count = 0; 3554 3555 ASSERT3P(marker, !=, NULL); 3556 IMPLY(bytes < 0, bytes == ARC_EVICT_ALL); 3557 3558 mls = multilist_sublist_lock(ml, idx); 3559 3560 for (hdr = multilist_sublist_prev(mls, marker); hdr != NULL; 3561 hdr = multilist_sublist_prev(mls, marker)) { 3562 if ((bytes != ARC_EVICT_ALL && bytes_evicted >= bytes) || 3563 (evict_count >= zfs_arc_evict_batch_limit)) 3564 break; 3565 3566 /* 3567 * To keep our iteration location, move the marker 3568 * forward. Since we're not holding hdr's hash lock, we 3569 * must be very careful and not remove 'hdr' from the 3570 * sublist. Otherwise, other consumers might mistake the 3571 * 'hdr' as not being on a sublist when they call the 3572 * multilist_link_active() function (they all rely on 3573 * the hash lock protecting concurrent insertions and 3574 * removals). multilist_sublist_move_forward() was 3575 * specifically implemented to ensure this is the case 3576 * (only 'marker' will be removed and re-inserted). 3577 */ 3578 multilist_sublist_move_forward(mls, marker); 3579 3580 /* 3581 * The only case where the b_spa field should ever be 3582 * zero, is the marker headers inserted by 3583 * arc_evict_state(). It's possible for multiple threads 3584 * to be calling arc_evict_state() concurrently (e.g. 3585 * dsl_pool_close() and zio_inject_fault()), so we must 3586 * skip any markers we see from these other threads. 3587 */ 3588 if (hdr->b_spa == 0) 3589 continue; 3590 3591 /* we're only interested in evicting buffers of a certain spa */ 3592 if (spa != 0 && hdr->b_spa != spa) { 3593 ARCSTAT_BUMP(arcstat_evict_skip); 3594 continue; 3595 } 3596 3597 hash_lock = HDR_LOCK(hdr); 3598 3599 /* 3600 * We aren't calling this function from any code path 3601 * that would already be holding a hash lock, so we're 3602 * asserting on this assumption to be defensive in case 3603 * this ever changes. Without this check, it would be 3604 * possible to incorrectly increment arcstat_mutex_miss 3605 * below (e.g. if the code changed such that we called 3606 * this function with a hash lock held). 3607 */ 3608 ASSERT(!MUTEX_HELD(hash_lock)); 3609 3610 if (mutex_tryenter(hash_lock)) { 3611 uint64_t evicted = arc_evict_hdr(hdr, hash_lock); 3612 mutex_exit(hash_lock); 3613 3614 bytes_evicted += evicted; 3615 3616 /* 3617 * If evicted is zero, arc_evict_hdr() must have 3618 * decided to skip this header, don't increment 3619 * evict_count in this case. 3620 */ 3621 if (evicted != 0) 3622 evict_count++; 3623 3624 /* 3625 * If arc_size isn't overflowing, signal any 3626 * threads that might happen to be waiting. 3627 * 3628 * For each header evicted, we wake up a single 3629 * thread. If we used cv_broadcast, we could 3630 * wake up "too many" threads causing arc_size 3631 * to significantly overflow arc_c; since 3632 * arc_get_data_buf() doesn't check for overflow 3633 * when it's woken up (it doesn't because it's 3634 * possible for the ARC to be overflowing while 3635 * full of un-evictable buffers, and the 3636 * function should proceed in this case). 3637 * 3638 * If threads are left sleeping, due to not 3639 * using cv_broadcast, they will be woken up 3640 * just before arc_reclaim_thread() sleeps. 3641 */ 3642 mutex_enter(&arc_reclaim_lock); 3643 if (!arc_is_overflowing()) 3644 cv_signal(&arc_reclaim_waiters_cv); 3645 mutex_exit(&arc_reclaim_lock); 3646 } else { 3647 ARCSTAT_BUMP(arcstat_mutex_miss); 3648 } 3649 } 3650 3651 multilist_sublist_unlock(mls); 3652 3653 return (bytes_evicted); 3654} 3655 3656/* 3657 * Evict buffers from the given arc state, until we've removed the 3658 * specified number of bytes. Move the removed buffers to the 3659 * appropriate evict state. 3660 * 3661 * This function makes a "best effort". It skips over any buffers 3662 * it can't get a hash_lock on, and so, may not catch all candidates. 3663 * It may also return without evicting as much space as requested. 3664 * 3665 * If bytes is specified using the special value ARC_EVICT_ALL, this 3666 * will evict all available (i.e. unlocked and evictable) buffers from 3667 * the given arc state; which is used by arc_flush(). 3668 */ 3669static uint64_t 3670arc_evict_state(arc_state_t *state, uint64_t spa, int64_t bytes, 3671 arc_buf_contents_t type) 3672{ 3673 uint64_t total_evicted = 0; 3674 multilist_t *ml = state->arcs_list[type]; 3675 int num_sublists; 3676 arc_buf_hdr_t **markers; 3677 3678 IMPLY(bytes < 0, bytes == ARC_EVICT_ALL); 3679 3680 num_sublists = multilist_get_num_sublists(ml); 3681 3682 /* 3683 * If we've tried to evict from each sublist, made some 3684 * progress, but still have not hit the target number of bytes 3685 * to evict, we want to keep trying. The markers allow us to 3686 * pick up where we left off for each individual sublist, rather 3687 * than starting from the tail each time. 3688 */ 3689 markers = kmem_zalloc(sizeof (*markers) * num_sublists, KM_SLEEP); 3690 for (int i = 0; i < num_sublists; i++) { 3691 markers[i] = kmem_cache_alloc(hdr_full_cache, KM_SLEEP); 3692 3693 /* 3694 * A b_spa of 0 is used to indicate that this header is 3695 * a marker. This fact is used in arc_adjust_type() and 3696 * arc_evict_state_impl(). 3697 */ 3698 markers[i]->b_spa = 0; 3699 3700 multilist_sublist_t *mls = multilist_sublist_lock(ml, i); 3701 multilist_sublist_insert_tail(mls, markers[i]); 3702 multilist_sublist_unlock(mls); 3703 } 3704 3705 /* 3706 * While we haven't hit our target number of bytes to evict, or 3707 * we're evicting all available buffers. 3708 */ 3709 while (total_evicted < bytes || bytes == ARC_EVICT_ALL) { 3710 /* 3711 * Start eviction using a randomly selected sublist, 3712 * this is to try and evenly balance eviction across all 3713 * sublists. Always starting at the same sublist 3714 * (e.g. index 0) would cause evictions to favor certain 3715 * sublists over others. 3716 */ 3717 int sublist_idx = multilist_get_random_index(ml); 3718 uint64_t scan_evicted = 0; 3719 3720 for (int i = 0; i < num_sublists; i++) { 3721 uint64_t bytes_remaining; 3722 uint64_t bytes_evicted; 3723 3724 if (bytes == ARC_EVICT_ALL) 3725 bytes_remaining = ARC_EVICT_ALL; 3726 else if (total_evicted < bytes) 3727 bytes_remaining = bytes - total_evicted; 3728 else 3729 break; 3730 3731 bytes_evicted = arc_evict_state_impl(ml, sublist_idx, 3732 markers[sublist_idx], spa, bytes_remaining); 3733 3734 scan_evicted += bytes_evicted; 3735 total_evicted += bytes_evicted; 3736 3737 /* we've reached the end, wrap to the beginning */ 3738 if (++sublist_idx >= num_sublists) 3739 sublist_idx = 0; 3740 } 3741 3742 /* 3743 * If we didn't evict anything during this scan, we have 3744 * no reason to believe we'll evict more during another 3745 * scan, so break the loop. 3746 */ 3747 if (scan_evicted == 0) { 3748 /* This isn't possible, let's make that obvious */ 3749 ASSERT3S(bytes, !=, 0); 3750 3751 /* 3752 * When bytes is ARC_EVICT_ALL, the only way to 3753 * break the loop is when scan_evicted is zero. 3754 * In that case, we actually have evicted enough, 3755 * so we don't want to increment the kstat. 3756 */ 3757 if (bytes != ARC_EVICT_ALL) { 3758 ASSERT3S(total_evicted, <, bytes); 3759 ARCSTAT_BUMP(arcstat_evict_not_enough); 3760 } 3761 3762 break; 3763 } 3764 } 3765 3766 for (int i = 0; i < num_sublists; i++) { 3767 multilist_sublist_t *mls = multilist_sublist_lock(ml, i); 3768 multilist_sublist_remove(mls, markers[i]); 3769 multilist_sublist_unlock(mls); 3770 3771 kmem_cache_free(hdr_full_cache, markers[i]); 3772 } 3773 kmem_free(markers, sizeof (*markers) * num_sublists); 3774 3775 return (total_evicted); 3776} 3777 3778/* 3779 * Flush all "evictable" data of the given type from the arc state 3780 * specified. This will not evict any "active" buffers (i.e. referenced). 3781 * 3782 * When 'retry' is set to B_FALSE, the function will make a single pass 3783 * over the state and evict any buffers that it can. Since it doesn't 3784 * continually retry the eviction, it might end up leaving some buffers 3785 * in the ARC due to lock misses. 3786 * 3787 * When 'retry' is set to B_TRUE, the function will continually retry the 3788 * eviction until *all* evictable buffers have been removed from the 3789 * state. As a result, if concurrent insertions into the state are 3790 * allowed (e.g. if the ARC isn't shutting down), this function might 3791 * wind up in an infinite loop, continually trying to evict buffers. 3792 */ 3793static uint64_t 3794arc_flush_state(arc_state_t *state, uint64_t spa, arc_buf_contents_t type, 3795 boolean_t retry) 3796{ 3797 uint64_t evicted = 0; 3798 3799 while (refcount_count(&state->arcs_esize[type]) != 0) { 3800 evicted += arc_evict_state(state, spa, ARC_EVICT_ALL, type); 3801 3802 if (!retry) 3803 break; 3804 } 3805 3806 return (evicted); 3807} 3808 3809/* 3810 * Evict the specified number of bytes from the state specified, 3811 * restricting eviction to the spa and type given. This function 3812 * prevents us from trying to evict more from a state's list than 3813 * is "evictable", and to skip evicting altogether when passed a 3814 * negative value for "bytes". In contrast, arc_evict_state() will 3815 * evict everything it can, when passed a negative value for "bytes". 3816 */ 3817static uint64_t 3818arc_adjust_impl(arc_state_t *state, uint64_t spa, int64_t bytes, 3819 arc_buf_contents_t type) 3820{ 3821 int64_t delta; 3822 3823 if (bytes > 0 && refcount_count(&state->arcs_esize[type]) > 0) { 3824 delta = MIN(refcount_count(&state->arcs_esize[type]), bytes); 3825 return (arc_evict_state(state, spa, delta, type)); 3826 } 3827 3828 return (0); 3829} 3830 3831/* 3832 * Evict metadata buffers from the cache, such that arc_meta_used is 3833 * capped by the arc_meta_limit tunable. 3834 */ 3835static uint64_t 3836arc_adjust_meta(void) 3837{ 3838 uint64_t total_evicted = 0; 3839 int64_t target; 3840 3841 /* 3842 * If we're over the meta limit, we want to evict enough 3843 * metadata to get back under the meta limit. We don't want to 3844 * evict so much that we drop the MRU below arc_p, though. If 3845 * we're over the meta limit more than we're over arc_p, we 3846 * evict some from the MRU here, and some from the MFU below. 3847 */ 3848 target = MIN((int64_t)(arc_meta_used - arc_meta_limit), 3849 (int64_t)(refcount_count(&arc_anon->arcs_size) + 3850 refcount_count(&arc_mru->arcs_size) - arc_p)); 3851 3852 total_evicted += arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA); 3853 3854 /* 3855 * Similar to the above, we want to evict enough bytes to get us 3856 * below the meta limit, but not so much as to drop us below the 3857 * space allotted to the MFU (which is defined as arc_c - arc_p). 3858 */ 3859 target = MIN((int64_t)(arc_meta_used - arc_meta_limit), 3860 (int64_t)(refcount_count(&arc_mfu->arcs_size) - (arc_c - arc_p))); 3861 3862 total_evicted += arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA); 3863 3864 return (total_evicted); 3865} 3866 3867/* 3868 * Return the type of the oldest buffer in the given arc state 3869 * 3870 * This function will select a random sublist of type ARC_BUFC_DATA and 3871 * a random sublist of type ARC_BUFC_METADATA. The tail of each sublist 3872 * is compared, and the type which contains the "older" buffer will be 3873 * returned. 3874 */ 3875static arc_buf_contents_t 3876arc_adjust_type(arc_state_t *state) 3877{ 3878 multilist_t *data_ml = state->arcs_list[ARC_BUFC_DATA]; 3879 multilist_t *meta_ml = state->arcs_list[ARC_BUFC_METADATA]; 3880 int data_idx = multilist_get_random_index(data_ml); 3881 int meta_idx = multilist_get_random_index(meta_ml); 3882 multilist_sublist_t *data_mls; 3883 multilist_sublist_t *meta_mls; 3884 arc_buf_contents_t type; 3885 arc_buf_hdr_t *data_hdr; 3886 arc_buf_hdr_t *meta_hdr; 3887 3888 /* 3889 * We keep the sublist lock until we're finished, to prevent 3890 * the headers from being destroyed via arc_evict_state(). 3891 */ 3892 data_mls = multilist_sublist_lock(data_ml, data_idx); 3893 meta_mls = multilist_sublist_lock(meta_ml, meta_idx); 3894 3895 /* 3896 * These two loops are to ensure we skip any markers that 3897 * might be at the tail of the lists due to arc_evict_state(). 3898 */ 3899 3900 for (data_hdr = multilist_sublist_tail(data_mls); data_hdr != NULL; 3901 data_hdr = multilist_sublist_prev(data_mls, data_hdr)) { 3902 if (data_hdr->b_spa != 0) 3903 break; 3904 } 3905 3906 for (meta_hdr = multilist_sublist_tail(meta_mls); meta_hdr != NULL; 3907 meta_hdr = multilist_sublist_prev(meta_mls, meta_hdr)) { 3908 if (meta_hdr->b_spa != 0) 3909 break; 3910 } 3911 3912 if (data_hdr == NULL && meta_hdr == NULL) { 3913 type = ARC_BUFC_DATA; 3914 } else if (data_hdr == NULL) { 3915 ASSERT3P(meta_hdr, !=, NULL); 3916 type = ARC_BUFC_METADATA; 3917 } else if (meta_hdr == NULL) { 3918 ASSERT3P(data_hdr, !=, NULL); 3919 type = ARC_BUFC_DATA; 3920 } else { 3921 ASSERT3P(data_hdr, !=, NULL); 3922 ASSERT3P(meta_hdr, !=, NULL); 3923 3924 /* The headers can't be on the sublist without an L1 header */ 3925 ASSERT(HDR_HAS_L1HDR(data_hdr)); 3926 ASSERT(HDR_HAS_L1HDR(meta_hdr)); 3927 3928 if (data_hdr->b_l1hdr.b_arc_access < 3929 meta_hdr->b_l1hdr.b_arc_access) { 3930 type = ARC_BUFC_DATA; 3931 } else { 3932 type = ARC_BUFC_METADATA; 3933 } 3934 } 3935 3936 multilist_sublist_unlock(meta_mls); 3937 multilist_sublist_unlock(data_mls); 3938 3939 return (type); 3940} 3941 3942/* 3943 * Evict buffers from the cache, such that arc_size is capped by arc_c. 3944 */ 3945static uint64_t 3946arc_adjust(void) 3947{ 3948 uint64_t total_evicted = 0; 3949 uint64_t bytes; 3950 int64_t target; 3951 3952 /* 3953 * If we're over arc_meta_limit, we want to correct that before 3954 * potentially evicting data buffers below. 3955 */ 3956 total_evicted += arc_adjust_meta(); 3957 3958 /* 3959 * Adjust MRU size 3960 * 3961 * If we're over the target cache size, we want to evict enough 3962 * from the list to get back to our target size. We don't want 3963 * to evict too much from the MRU, such that it drops below 3964 * arc_p. So, if we're over our target cache size more than 3965 * the MRU is over arc_p, we'll evict enough to get back to 3966 * arc_p here, and then evict more from the MFU below. 3967 */ 3968 target = MIN((int64_t)(arc_size - arc_c), 3969 (int64_t)(refcount_count(&arc_anon->arcs_size) + 3970 refcount_count(&arc_mru->arcs_size) + arc_meta_used - arc_p)); 3971 3972 /* 3973 * If we're below arc_meta_min, always prefer to evict data. 3974 * Otherwise, try to satisfy the requested number of bytes to 3975 * evict from the type which contains older buffers; in an 3976 * effort to keep newer buffers in the cache regardless of their 3977 * type. If we cannot satisfy the number of bytes from this 3978 * type, spill over into the next type. 3979 */ 3980 if (arc_adjust_type(arc_mru) == ARC_BUFC_METADATA && 3981 arc_meta_used > arc_meta_min) { 3982 bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA); 3983 total_evicted += bytes; 3984 3985 /* 3986 * If we couldn't evict our target number of bytes from 3987 * metadata, we try to get the rest from data. 3988 */ 3989 target -= bytes; 3990 3991 total_evicted += 3992 arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA); 3993 } else { 3994 bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA); 3995 total_evicted += bytes; 3996 3997 /* 3998 * If we couldn't evict our target number of bytes from 3999 * data, we try to get the rest from metadata. 4000 */ 4001 target -= bytes; 4002 4003 total_evicted += 4004 arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA); 4005 } 4006 4007 /* 4008 * Adjust MFU size 4009 * 4010 * Now that we've tried to evict enough from the MRU to get its 4011 * size back to arc_p, if we're still above the target cache 4012 * size, we evict the rest from the MFU. 4013 */ 4014 target = arc_size - arc_c; 4015 4016 if (arc_adjust_type(arc_mfu) == ARC_BUFC_METADATA && 4017 arc_meta_used > arc_meta_min) { 4018 bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA); 4019 total_evicted += bytes; 4020 4021 /* 4022 * If we couldn't evict our target number of bytes from 4023 * metadata, we try to get the rest from data. 4024 */ 4025 target -= bytes; 4026 4027 total_evicted += 4028 arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA); 4029 } else { 4030 bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA); 4031 total_evicted += bytes; 4032 4033 /* 4034 * If we couldn't evict our target number of bytes from 4035 * data, we try to get the rest from data. 4036 */ 4037 target -= bytes; 4038 4039 total_evicted += 4040 arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA); 4041 } 4042 4043 /* 4044 * Adjust ghost lists 4045 * 4046 * In addition to the above, the ARC also defines target values 4047 * for the ghost lists. The sum of the mru list and mru ghost 4048 * list should never exceed the target size of the cache, and 4049 * the sum of the mru list, mfu list, mru ghost list, and mfu 4050 * ghost list should never exceed twice the target size of the 4051 * cache. The following logic enforces these limits on the ghost 4052 * caches, and evicts from them as needed. 4053 */ 4054 target = refcount_count(&arc_mru->arcs_size) + 4055 refcount_count(&arc_mru_ghost->arcs_size) - arc_c; 4056 4057 bytes = arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_DATA); 4058 total_evicted += bytes; 4059 4060 target -= bytes; 4061 4062 total_evicted += 4063 arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_METADATA); 4064 4065 /* 4066 * We assume the sum of the mru list and mfu list is less than 4067 * or equal to arc_c (we enforced this above), which means we 4068 * can use the simpler of the two equations below: 4069 * 4070 * mru + mfu + mru ghost + mfu ghost <= 2 * arc_c 4071 * mru ghost + mfu ghost <= arc_c 4072 */ 4073 target = refcount_count(&arc_mru_ghost->arcs_size) + 4074 refcount_count(&arc_mfu_ghost->arcs_size) - arc_c; 4075 4076 bytes = arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_DATA); 4077 total_evicted += bytes; 4078 4079 target -= bytes; 4080 4081 total_evicted += 4082 arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_METADATA); 4083 4084 return (total_evicted); 4085} 4086 4087void 4088arc_flush(spa_t *spa, boolean_t retry) 4089{ 4090 uint64_t guid = 0; 4091 4092 /* 4093 * If retry is B_TRUE, a spa must not be specified since we have 4094 * no good way to determine if all of a spa's buffers have been 4095 * evicted from an arc state. 4096 */ 4097 ASSERT(!retry || spa == 0); 4098 4099 if (spa != NULL) 4100 guid = spa_load_guid(spa); 4101 4102 (void) arc_flush_state(arc_mru, guid, ARC_BUFC_DATA, retry); 4103 (void) arc_flush_state(arc_mru, guid, ARC_BUFC_METADATA, retry); 4104 4105 (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_DATA, retry); 4106 (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_METADATA, retry); 4107 4108 (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_DATA, retry); 4109 (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_METADATA, retry); 4110 4111 (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_DATA, retry); 4112 (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_METADATA, retry); 4113} 4114 4115void 4116arc_shrink(int64_t to_free) 4117{ 4118 if (arc_c > arc_c_min) { 4119 DTRACE_PROBE4(arc__shrink, uint64_t, arc_c, uint64_t, 4120 arc_c_min, uint64_t, arc_p, uint64_t, to_free); 4121 if (arc_c > arc_c_min + to_free) 4122 atomic_add_64(&arc_c, -to_free); 4123 else 4124 arc_c = arc_c_min; 4125 4126 atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift)); 4127 if (arc_c > arc_size) 4128 arc_c = MAX(arc_size, arc_c_min); 4129 if (arc_p > arc_c) 4130 arc_p = (arc_c >> 1); 4131 4132 DTRACE_PROBE2(arc__shrunk, uint64_t, arc_c, uint64_t, 4133 arc_p); 4134 4135 ASSERT(arc_c >= arc_c_min); 4136 ASSERT((int64_t)arc_p >= 0); 4137 } 4138 4139 if (arc_size > arc_c) { 4140 DTRACE_PROBE2(arc__shrink_adjust, uint64_t, arc_size, 4141 uint64_t, arc_c); 4142 (void) arc_adjust(); 4143 } 4144} 4145 4146static long needfree = 0; 4147 4148typedef enum free_memory_reason_t { 4149 FMR_UNKNOWN, 4150 FMR_NEEDFREE, 4151 FMR_LOTSFREE, 4152 FMR_SWAPFS_MINFREE, 4153 FMR_PAGES_PP_MAXIMUM, 4154 FMR_HEAP_ARENA, 4155 FMR_ZIO_ARENA, 4156 FMR_ZIO_FRAG, 4157} free_memory_reason_t; 4158 4159int64_t last_free_memory; 4160free_memory_reason_t last_free_reason; 4161 4162/* 4163 * Additional reserve of pages for pp_reserve. 4164 */ 4165int64_t arc_pages_pp_reserve = 64; 4166 4167/* 4168 * Additional reserve of pages for swapfs. 4169 */ 4170int64_t arc_swapfs_reserve = 64; 4171 4172/* 4173 * Return the amount of memory that can be consumed before reclaim will be 4174 * needed. Positive if there is sufficient free memory, negative indicates 4175 * the amount of memory that needs to be freed up. 4176 */ 4177static int64_t 4178arc_available_memory(void) 4179{ 4180 int64_t lowest = INT64_MAX; 4181 int64_t n; 4182 free_memory_reason_t r = FMR_UNKNOWN; 4183 4184#ifdef _KERNEL 4185 if (needfree > 0) { 4186 n = PAGESIZE * (-needfree); 4187 if (n < lowest) { 4188 lowest = n; 4189 r = FMR_NEEDFREE; 4190 } 4191 } 4192 4193 /* 4194 * Cooperate with pagedaemon when it's time for it to scan 4195 * and reclaim some pages. 4196 */ 4197 n = PAGESIZE * ((int64_t)freemem - zfs_arc_free_target); 4198 if (n < lowest) { 4199 lowest = n; 4200 r = FMR_LOTSFREE; 4201 } 4202 4203#ifdef illumos 4204 /* 4205 * check that we're out of range of the pageout scanner. It starts to 4206 * schedule paging if freemem is less than lotsfree and needfree. 4207 * lotsfree is the high-water mark for pageout, and needfree is the 4208 * number of needed free pages. We add extra pages here to make sure 4209 * the scanner doesn't start up while we're freeing memory. 4210 */ 4211 n = PAGESIZE * (freemem - lotsfree - needfree - desfree); 4212 if (n < lowest) { 4213 lowest = n; 4214 r = FMR_LOTSFREE; 4215 } 4216 4217 /* 4218 * check to make sure that swapfs has enough space so that anon 4219 * reservations can still succeed. anon_resvmem() checks that the 4220 * availrmem is greater than swapfs_minfree, and the number of reserved 4221 * swap pages. We also add a bit of extra here just to prevent 4222 * circumstances from getting really dire. 4223 */ 4224 n = PAGESIZE * (availrmem - swapfs_minfree - swapfs_reserve - 4225 desfree - arc_swapfs_reserve); 4226 if (n < lowest) { 4227 lowest = n; 4228 r = FMR_SWAPFS_MINFREE; 4229 } 4230 4231 4232 /* 4233 * Check that we have enough availrmem that memory locking (e.g., via 4234 * mlock(3C) or memcntl(2)) can still succeed. (pages_pp_maximum 4235 * stores the number of pages that cannot be locked; when availrmem 4236 * drops below pages_pp_maximum, page locking mechanisms such as 4237 * page_pp_lock() will fail.) 4238 */ 4239 n = PAGESIZE * (availrmem - pages_pp_maximum - 4240 arc_pages_pp_reserve); 4241 if (n < lowest) { 4242 lowest = n; 4243 r = FMR_PAGES_PP_MAXIMUM; 4244 } 4245 4246#endif /* illumos */ 4247#if defined(__i386) || !defined(UMA_MD_SMALL_ALLOC) 4248 /* 4249 * If we're on an i386 platform, it's possible that we'll exhaust the 4250 * kernel heap space before we ever run out of available physical 4251 * memory. Most checks of the size of the heap_area compare against 4252 * tune.t_minarmem, which is the minimum available real memory that we 4253 * can have in the system. However, this is generally fixed at 25 pages 4254 * which is so low that it's useless. In this comparison, we seek to 4255 * calculate the total heap-size, and reclaim if more than 3/4ths of the 4256 * heap is allocated. (Or, in the calculation, if less than 1/4th is 4257 * free) 4258 */ 4259 n = (int64_t)vmem_size(heap_arena, VMEM_FREE) - 4260 (vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC) >> 2); 4261 if (n < lowest) { 4262 lowest = n; 4263 r = FMR_HEAP_ARENA; 4264 } 4265#define zio_arena NULL 4266#else 4267#define zio_arena heap_arena 4268#endif 4269 4270 /* 4271 * If zio data pages are being allocated out of a separate heap segment, 4272 * then enforce that the size of available vmem for this arena remains 4273 * above about 1/16th free. 4274 * 4275 * Note: The 1/16th arena free requirement was put in place 4276 * to aggressively evict memory from the arc in order to avoid 4277 * memory fragmentation issues. 4278 */ 4279 if (zio_arena != NULL) { 4280 n = (int64_t)vmem_size(zio_arena, VMEM_FREE) - 4281 (vmem_size(zio_arena, VMEM_ALLOC) >> 4); 4282 if (n < lowest) { 4283 lowest = n; 4284 r = FMR_ZIO_ARENA; 4285 } 4286 } 4287 4288 /* 4289 * Above limits know nothing about real level of KVA fragmentation. 4290 * Start aggressive reclamation if too little sequential KVA left. 4291 */ 4292 if (lowest > 0) { 4293 n = (vmem_size(heap_arena, VMEM_MAXFREE) < SPA_MAXBLOCKSIZE) ? 4294 -((int64_t)vmem_size(heap_arena, VMEM_ALLOC) >> 4) : 4295 INT64_MAX; 4296 if (n < lowest) { 4297 lowest = n; 4298 r = FMR_ZIO_FRAG; 4299 } 4300 } 4301 4302#else /* _KERNEL */ 4303 /* Every 100 calls, free a small amount */ 4304 if (spa_get_random(100) == 0) 4305 lowest = -1024; 4306#endif /* _KERNEL */ 4307 4308 last_free_memory = lowest; 4309 last_free_reason = r; 4310 DTRACE_PROBE2(arc__available_memory, int64_t, lowest, int, r); 4311 return (lowest); 4312} 4313 4314 4315/* 4316 * Determine if the system is under memory pressure and is asking 4317 * to reclaim memory. A return value of B_TRUE indicates that the system 4318 * is under memory pressure and that the arc should adjust accordingly. 4319 */ 4320static boolean_t 4321arc_reclaim_needed(void) 4322{ 4323 return (arc_available_memory() < 0); 4324} 4325 4326extern kmem_cache_t *zio_buf_cache[]; 4327extern kmem_cache_t *zio_data_buf_cache[]; 4328extern kmem_cache_t *range_seg_cache; 4329 4330static __noinline void 4331arc_kmem_reap_now(void) 4332{ 4333 size_t i; 4334 kmem_cache_t *prev_cache = NULL; 4335 kmem_cache_t *prev_data_cache = NULL; 4336 4337 DTRACE_PROBE(arc__kmem_reap_start); 4338#ifdef _KERNEL 4339 if (arc_meta_used >= arc_meta_limit) { 4340 /* 4341 * We are exceeding our meta-data cache limit. 4342 * Purge some DNLC entries to release holds on meta-data. 4343 */ 4344 dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent); 4345 } 4346#if defined(__i386) 4347 /* 4348 * Reclaim unused memory from all kmem caches. 4349 */ 4350 kmem_reap(); 4351#endif 4352#endif 4353 4354 for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) { 4355 if (zio_buf_cache[i] != prev_cache) { 4356 prev_cache = zio_buf_cache[i]; 4357 kmem_cache_reap_now(zio_buf_cache[i]); 4358 } 4359 if (zio_data_buf_cache[i] != prev_data_cache) { 4360 prev_data_cache = zio_data_buf_cache[i]; 4361 kmem_cache_reap_now(zio_data_buf_cache[i]); 4362 } 4363 } 4364 kmem_cache_reap_now(buf_cache); 4365 kmem_cache_reap_now(hdr_full_cache); 4366 kmem_cache_reap_now(hdr_l2only_cache); 4367 kmem_cache_reap_now(range_seg_cache); 4368 4369#ifdef illumos 4370 if (zio_arena != NULL) { 4371 /* 4372 * Ask the vmem arena to reclaim unused memory from its 4373 * quantum caches. 4374 */ 4375 vmem_qcache_reap(zio_arena); 4376 } 4377#endif 4378 DTRACE_PROBE(arc__kmem_reap_end); 4379} 4380 4381/* 4382 * Threads can block in arc_get_data_buf() waiting for this thread to evict 4383 * enough data and signal them to proceed. When this happens, the threads in 4384 * arc_get_data_buf() are sleeping while holding the hash lock for their 4385 * particular arc header. Thus, we must be careful to never sleep on a 4386 * hash lock in this thread. This is to prevent the following deadlock: 4387 * 4388 * - Thread A sleeps on CV in arc_get_data_buf() holding hash lock "L", 4389 * waiting for the reclaim thread to signal it. 4390 * 4391 * - arc_reclaim_thread() tries to acquire hash lock "L" using mutex_enter, 4392 * fails, and goes to sleep forever. 4393 * 4394 * This possible deadlock is avoided by always acquiring a hash lock 4395 * using mutex_tryenter() from arc_reclaim_thread(). 4396 */ 4397static void 4398arc_reclaim_thread(void *dummy __unused) 4399{ 4400 hrtime_t growtime = 0; 4401 callb_cpr_t cpr; 4402 4403 CALLB_CPR_INIT(&cpr, &arc_reclaim_lock, callb_generic_cpr, FTAG); 4404 4405 mutex_enter(&arc_reclaim_lock); 4406 while (!arc_reclaim_thread_exit) { 4407 uint64_t evicted = 0; 4408 4409 /* 4410 * This is necessary in order for the mdb ::arc dcmd to 4411 * show up to date information. Since the ::arc command 4412 * does not call the kstat's update function, without 4413 * this call, the command may show stale stats for the 4414 * anon, mru, mru_ghost, mfu, and mfu_ghost lists. Even 4415 * with this change, the data might be up to 1 second 4416 * out of date; but that should suffice. The arc_state_t 4417 * structures can be queried directly if more accurate 4418 * information is needed. 4419 */ 4420 if (arc_ksp != NULL) 4421 arc_ksp->ks_update(arc_ksp, KSTAT_READ); 4422 4423 mutex_exit(&arc_reclaim_lock); 4424 4425 /* 4426 * We call arc_adjust() before (possibly) calling 4427 * arc_kmem_reap_now(), so that we can wake up 4428 * arc_get_data_buf() sooner. 4429 */ 4430 evicted = arc_adjust(); 4431 4432 int64_t free_memory = arc_available_memory(); 4433 if (free_memory < 0) { 4434 4435 arc_no_grow = B_TRUE; 4436 arc_warm = B_TRUE; 4437 4438 /* 4439 * Wait at least zfs_grow_retry (default 60) seconds 4440 * before considering growing. 4441 */ 4442 growtime = gethrtime() + SEC2NSEC(arc_grow_retry); 4443 4444 arc_kmem_reap_now(); 4445 4446 /* 4447 * If we are still low on memory, shrink the ARC 4448 * so that we have arc_shrink_min free space. 4449 */ 4450 free_memory = arc_available_memory(); 4451 4452 int64_t to_free = 4453 (arc_c >> arc_shrink_shift) - free_memory; 4454 if (to_free > 0) { 4455#ifdef _KERNEL 4456 to_free = MAX(to_free, ptob(needfree)); 4457#endif 4458 arc_shrink(to_free); 4459 } 4460 } else if (free_memory < arc_c >> arc_no_grow_shift) { 4461 arc_no_grow = B_TRUE; 4462 } else if (gethrtime() >= growtime) { 4463 arc_no_grow = B_FALSE; 4464 } 4465 4466 mutex_enter(&arc_reclaim_lock); 4467 4468 /* 4469 * If evicted is zero, we couldn't evict anything via 4470 * arc_adjust(). This could be due to hash lock 4471 * collisions, but more likely due to the majority of 4472 * arc buffers being unevictable. Therefore, even if 4473 * arc_size is above arc_c, another pass is unlikely to 4474 * be helpful and could potentially cause us to enter an 4475 * infinite loop. 4476 */ 4477 if (arc_size <= arc_c || evicted == 0) { 4478#ifdef _KERNEL 4479 needfree = 0; 4480#endif 4481 /* 4482 * We're either no longer overflowing, or we 4483 * can't evict anything more, so we should wake 4484 * up any threads before we go to sleep. 4485 */ 4486 cv_broadcast(&arc_reclaim_waiters_cv); 4487 4488 /* 4489 * Block until signaled, or after one second (we 4490 * might need to perform arc_kmem_reap_now() 4491 * even if we aren't being signalled) 4492 */ 4493 CALLB_CPR_SAFE_BEGIN(&cpr); 4494 (void) cv_timedwait_hires(&arc_reclaim_thread_cv, 4495 &arc_reclaim_lock, SEC2NSEC(1), MSEC2NSEC(1), 0); 4496 CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_lock); 4497 } 4498 } 4499 4500 arc_reclaim_thread_exit = B_FALSE; 4501 cv_broadcast(&arc_reclaim_thread_cv); 4502 CALLB_CPR_EXIT(&cpr); /* drops arc_reclaim_lock */ 4503 thread_exit(); 4504} 4505 4506static u_int arc_dnlc_evicts_arg; 4507extern struct vfsops zfs_vfsops; 4508 4509static void 4510arc_dnlc_evicts_thread(void *dummy __unused) 4511{ 4512 callb_cpr_t cpr; 4513 u_int percent; 4514 4515 CALLB_CPR_INIT(&cpr, &arc_dnlc_evicts_lock, callb_generic_cpr, FTAG); 4516 4517 mutex_enter(&arc_dnlc_evicts_lock); 4518 while (!arc_dnlc_evicts_thread_exit) { 4519 CALLB_CPR_SAFE_BEGIN(&cpr); 4520 (void) cv_wait(&arc_dnlc_evicts_cv, &arc_dnlc_evicts_lock); 4521 CALLB_CPR_SAFE_END(&cpr, &arc_dnlc_evicts_lock); 4522 if (arc_dnlc_evicts_arg != 0) { 4523 percent = arc_dnlc_evicts_arg; 4524 mutex_exit(&arc_dnlc_evicts_lock); 4525#ifdef _KERNEL 4526 vnlru_free(desiredvnodes * percent / 100, &zfs_vfsops); 4527#endif 4528 mutex_enter(&arc_dnlc_evicts_lock); 4529 /* 4530 * Clear our token only after vnlru_free() 4531 * pass is done, to avoid false queueing of 4532 * the requests. 4533 */ 4534 arc_dnlc_evicts_arg = 0; 4535 } 4536 } 4537 arc_dnlc_evicts_thread_exit = FALSE; 4538 cv_broadcast(&arc_dnlc_evicts_cv); 4539 CALLB_CPR_EXIT(&cpr); 4540 thread_exit(); 4541} 4542 4543void 4544dnlc_reduce_cache(void *arg) 4545{ 4546 u_int percent; 4547 4548 percent = (u_int)(uintptr_t)arg; 4549 mutex_enter(&arc_dnlc_evicts_lock); 4550 if (arc_dnlc_evicts_arg == 0) { 4551 arc_dnlc_evicts_arg = percent; 4552 cv_broadcast(&arc_dnlc_evicts_cv); 4553 } 4554 mutex_exit(&arc_dnlc_evicts_lock); 4555} 4556 4557/* 4558 * Adapt arc info given the number of bytes we are trying to add and 4559 * the state that we are comming from. This function is only called 4560 * when we are adding new content to the cache. 4561 */ 4562static void 4563arc_adapt(int bytes, arc_state_t *state) 4564{ 4565 int mult; 4566 uint64_t arc_p_min = (arc_c >> arc_p_min_shift); 4567 int64_t mrug_size = refcount_count(&arc_mru_ghost->arcs_size); 4568 int64_t mfug_size = refcount_count(&arc_mfu_ghost->arcs_size); 4569 4570 if (state == arc_l2c_only) 4571 return; 4572 4573 ASSERT(bytes > 0); 4574 /* 4575 * Adapt the target size of the MRU list: 4576 * - if we just hit in the MRU ghost list, then increase 4577 * the target size of the MRU list. 4578 * - if we just hit in the MFU ghost list, then increase 4579 * the target size of the MFU list by decreasing the 4580 * target size of the MRU list. 4581 */ 4582 if (state == arc_mru_ghost) { 4583 mult = (mrug_size >= mfug_size) ? 1 : (mfug_size / mrug_size); 4584 mult = MIN(mult, 10); /* avoid wild arc_p adjustment */ 4585 4586 arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult); 4587 } else if (state == arc_mfu_ghost) { 4588 uint64_t delta; 4589 4590 mult = (mfug_size >= mrug_size) ? 1 : (mrug_size / mfug_size); 4591 mult = MIN(mult, 10); 4592 4593 delta = MIN(bytes * mult, arc_p); 4594 arc_p = MAX(arc_p_min, arc_p - delta); 4595 } 4596 ASSERT((int64_t)arc_p >= 0); 4597 4598 if (arc_reclaim_needed()) { 4599 cv_signal(&arc_reclaim_thread_cv); 4600 return; 4601 } 4602 4603 if (arc_no_grow) 4604 return; 4605 4606 if (arc_c >= arc_c_max) 4607 return; 4608 4609 /* 4610 * If we're within (2 * maxblocksize) bytes of the target 4611 * cache size, increment the target cache size 4612 */ 4613 if (arc_size > arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) { 4614 DTRACE_PROBE1(arc__inc_adapt, int, bytes); 4615 atomic_add_64(&arc_c, (int64_t)bytes); 4616 if (arc_c > arc_c_max) 4617 arc_c = arc_c_max; 4618 else if (state == arc_anon) 4619 atomic_add_64(&arc_p, (int64_t)bytes); 4620 if (arc_p > arc_c) 4621 arc_p = arc_c; 4622 } 4623 ASSERT((int64_t)arc_p >= 0); 4624} 4625 4626/* 4627 * Check if arc_size has grown past our upper threshold, determined by 4628 * zfs_arc_overflow_shift. 4629 */ 4630static boolean_t 4631arc_is_overflowing(void) 4632{ 4633 /* Always allow at least one block of overflow */ 4634 uint64_t overflow = MAX(SPA_MAXBLOCKSIZE, 4635 arc_c >> zfs_arc_overflow_shift); 4636 4637 return (arc_size >= arc_c + overflow); 4638} 4639 4640/* 4641 * Allocate a block and return it to the caller. If we are hitting the 4642 * hard limit for the cache size, we must sleep, waiting for the eviction 4643 * thread to catch up. If we're past the target size but below the hard 4644 * limit, we'll only signal the reclaim thread and continue on. 4645 */ 4646static void * 4647arc_get_data_buf(arc_buf_hdr_t *hdr, uint64_t size, void *tag) 4648{ 4649 void *datap = NULL; 4650 arc_state_t *state = hdr->b_l1hdr.b_state; 4651 arc_buf_contents_t type = arc_buf_type(hdr); 4652 4653 arc_adapt(size, state); 4654 4655 /* 4656 * If arc_size is currently overflowing, and has grown past our 4657 * upper limit, we must be adding data faster than the evict 4658 * thread can evict. Thus, to ensure we don't compound the 4659 * problem by adding more data and forcing arc_size to grow even 4660 * further past it's target size, we halt and wait for the 4661 * eviction thread to catch up. 4662 * 4663 * It's also possible that the reclaim thread is unable to evict 4664 * enough buffers to get arc_size below the overflow limit (e.g. 4665 * due to buffers being un-evictable, or hash lock collisions). 4666 * In this case, we want to proceed regardless if we're 4667 * overflowing; thus we don't use a while loop here. 4668 */ 4669 if (arc_is_overflowing()) { 4670 mutex_enter(&arc_reclaim_lock); 4671 4672 /* 4673 * Now that we've acquired the lock, we may no longer be 4674 * over the overflow limit, lets check. 4675 * 4676 * We're ignoring the case of spurious wake ups. If that 4677 * were to happen, it'd let this thread consume an ARC 4678 * buffer before it should have (i.e. before we're under 4679 * the overflow limit and were signalled by the reclaim 4680 * thread). As long as that is a rare occurrence, it 4681 * shouldn't cause any harm. 4682 */ 4683 if (arc_is_overflowing()) { 4684 cv_signal(&arc_reclaim_thread_cv); 4685 cv_wait(&arc_reclaim_waiters_cv, &arc_reclaim_lock); 4686 } 4687 4688 mutex_exit(&arc_reclaim_lock); 4689 } 4690 4691 VERIFY3U(hdr->b_type, ==, type); 4692 if (type == ARC_BUFC_METADATA) { 4693 datap = zio_buf_alloc(size); 4694 arc_space_consume(size, ARC_SPACE_META); 4695 } else { 4696 ASSERT(type == ARC_BUFC_DATA); 4697 datap = zio_data_buf_alloc(size); 4698 arc_space_consume(size, ARC_SPACE_DATA); 4699 } 4700 4701 /* 4702 * Update the state size. Note that ghost states have a 4703 * "ghost size" and so don't need to be updated. 4704 */ 4705 if (!GHOST_STATE(state)) { 4706 4707 (void) refcount_add_many(&state->arcs_size, size, tag); 4708 4709 /* 4710 * If this is reached via arc_read, the link is 4711 * protected by the hash lock. If reached via 4712 * arc_buf_alloc, the header should not be accessed by 4713 * any other thread. And, if reached via arc_read_done, 4714 * the hash lock will protect it if it's found in the 4715 * hash table; otherwise no other thread should be 4716 * trying to [add|remove]_reference it. 4717 */ 4718 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) { 4719 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 4720 (void) refcount_add_many(&state->arcs_esize[type], 4721 size, tag); 4722 } 4723 4724 /* 4725 * If we are growing the cache, and we are adding anonymous 4726 * data, and we have outgrown arc_p, update arc_p 4727 */ 4728 if (arc_size < arc_c && hdr->b_l1hdr.b_state == arc_anon && 4729 (refcount_count(&arc_anon->arcs_size) + 4730 refcount_count(&arc_mru->arcs_size) > arc_p)) 4731 arc_p = MIN(arc_c, arc_p + size); 4732 } 4733 ARCSTAT_BUMP(arcstat_allocated); 4734 return (datap); 4735} 4736 4737/* 4738 * Free the arc data buffer. 4739 */ 4740static void 4741arc_free_data_buf(arc_buf_hdr_t *hdr, void *data, uint64_t size, void *tag) 4742{ 4743 arc_state_t *state = hdr->b_l1hdr.b_state; 4744 arc_buf_contents_t type = arc_buf_type(hdr); 4745 4746 /* protected by hash lock, if in the hash table */ 4747 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) { 4748 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 4749 ASSERT(state != arc_anon && state != arc_l2c_only); 4750 4751 (void) refcount_remove_many(&state->arcs_esize[type], 4752 size, tag); 4753 } 4754 (void) refcount_remove_many(&state->arcs_size, size, tag); 4755 4756 VERIFY3U(hdr->b_type, ==, type); 4757 if (type == ARC_BUFC_METADATA) { 4758 zio_buf_free(data, size); 4759 arc_space_return(size, ARC_SPACE_META); 4760 } else { 4761 ASSERT(type == ARC_BUFC_DATA); 4762 zio_data_buf_free(data, size); 4763 arc_space_return(size, ARC_SPACE_DATA); 4764 } 4765} 4766 4767/* 4768 * This routine is called whenever a buffer is accessed. 4769 * NOTE: the hash lock is dropped in this function. 4770 */ 4771static void 4772arc_access(arc_buf_hdr_t *hdr, kmutex_t *hash_lock) 4773{ 4774 clock_t now; 4775 4776 ASSERT(MUTEX_HELD(hash_lock)); 4777 ASSERT(HDR_HAS_L1HDR(hdr)); 4778 4779 if (hdr->b_l1hdr.b_state == arc_anon) { 4780 /* 4781 * This buffer is not in the cache, and does not 4782 * appear in our "ghost" list. Add the new buffer 4783 * to the MRU state. 4784 */ 4785 4786 ASSERT0(hdr->b_l1hdr.b_arc_access); 4787 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 4788 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr); 4789 arc_change_state(arc_mru, hdr, hash_lock); 4790 4791 } else if (hdr->b_l1hdr.b_state == arc_mru) { 4792 now = ddi_get_lbolt(); 4793 4794 /* 4795 * If this buffer is here because of a prefetch, then either: 4796 * - clear the flag if this is a "referencing" read 4797 * (any subsequent access will bump this into the MFU state). 4798 * or 4799 * - move the buffer to the head of the list if this is 4800 * another prefetch (to make it less likely to be evicted). 4801 */ 4802 if (HDR_PREFETCH(hdr)) { 4803 if (refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) { 4804 /* link protected by hash lock */ 4805 ASSERT(multilist_link_active( 4806 &hdr->b_l1hdr.b_arc_node)); 4807 } else { 4808 arc_hdr_clear_flags(hdr, ARC_FLAG_PREFETCH); 4809 ARCSTAT_BUMP(arcstat_mru_hits); 4810 } 4811 hdr->b_l1hdr.b_arc_access = now; 4812 return; 4813 } 4814 4815 /* 4816 * This buffer has been "accessed" only once so far, 4817 * but it is still in the cache. Move it to the MFU 4818 * state. 4819 */ 4820 if (now > hdr->b_l1hdr.b_arc_access + ARC_MINTIME) { 4821 /* 4822 * More than 125ms have passed since we 4823 * instantiated this buffer. Move it to the 4824 * most frequently used state. 4825 */ 4826 hdr->b_l1hdr.b_arc_access = now; 4827 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 4828 arc_change_state(arc_mfu, hdr, hash_lock); 4829 } 4830 ARCSTAT_BUMP(arcstat_mru_hits); 4831 } else if (hdr->b_l1hdr.b_state == arc_mru_ghost) { 4832 arc_state_t *new_state; 4833 /* 4834 * This buffer has been "accessed" recently, but 4835 * was evicted from the cache. Move it to the 4836 * MFU state. 4837 */ 4838 4839 if (HDR_PREFETCH(hdr)) { 4840 new_state = arc_mru; 4841 if (refcount_count(&hdr->b_l1hdr.b_refcnt) > 0) 4842 arc_hdr_clear_flags(hdr, ARC_FLAG_PREFETCH); 4843 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr); 4844 } else { 4845 new_state = arc_mfu; 4846 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 4847 } 4848 4849 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 4850 arc_change_state(new_state, hdr, hash_lock); 4851 4852 ARCSTAT_BUMP(arcstat_mru_ghost_hits); 4853 } else if (hdr->b_l1hdr.b_state == arc_mfu) { 4854 /* 4855 * This buffer has been accessed more than once and is 4856 * still in the cache. Keep it in the MFU state. 4857 * 4858 * NOTE: an add_reference() that occurred when we did 4859 * the arc_read() will have kicked this off the list. 4860 * If it was a prefetch, we will explicitly move it to 4861 * the head of the list now. 4862 */ 4863 if ((HDR_PREFETCH(hdr)) != 0) { 4864 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 4865 /* link protected by hash_lock */ 4866 ASSERT(multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 4867 } 4868 ARCSTAT_BUMP(arcstat_mfu_hits); 4869 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 4870 } else if (hdr->b_l1hdr.b_state == arc_mfu_ghost) { 4871 arc_state_t *new_state = arc_mfu; 4872 /* 4873 * This buffer has been accessed more than once but has 4874 * been evicted from the cache. Move it back to the 4875 * MFU state. 4876 */ 4877 4878 if (HDR_PREFETCH(hdr)) { 4879 /* 4880 * This is a prefetch access... 4881 * move this block back to the MRU state. 4882 */ 4883 ASSERT0(refcount_count(&hdr->b_l1hdr.b_refcnt)); 4884 new_state = arc_mru; 4885 } 4886 4887 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 4888 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 4889 arc_change_state(new_state, hdr, hash_lock); 4890 4891 ARCSTAT_BUMP(arcstat_mfu_ghost_hits); 4892 } else if (hdr->b_l1hdr.b_state == arc_l2c_only) { 4893 /* 4894 * This buffer is on the 2nd Level ARC. 4895 */ 4896 4897 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 4898 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 4899 arc_change_state(arc_mfu, hdr, hash_lock); 4900 } else { 4901 ASSERT(!"invalid arc state"); 4902 } 4903} 4904 4905/* a generic arc_done_func_t which you can use */ 4906/* ARGSUSED */ 4907void 4908arc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg) 4909{ 4910 if (zio == NULL || zio->io_error == 0) 4911 bcopy(buf->b_data, arg, arc_buf_size(buf)); 4912 arc_buf_destroy(buf, arg); 4913} 4914 4915/* a generic arc_done_func_t */ 4916void 4917arc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg) 4918{ 4919 arc_buf_t **bufp = arg; 4920 if (zio && zio->io_error) { 4921 arc_buf_destroy(buf, arg); 4922 *bufp = NULL; 4923 } else { 4924 *bufp = buf; 4925 ASSERT(buf->b_data); 4926 } 4927} 4928 4929static void 4930arc_hdr_verify(arc_buf_hdr_t *hdr, blkptr_t *bp) 4931{ 4932 if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp)) { 4933 ASSERT3U(HDR_GET_PSIZE(hdr), ==, 0); 4934 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF); 4935 } else { 4936 if (HDR_COMPRESSION_ENABLED(hdr)) { 4937 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, 4938 BP_GET_COMPRESS(bp)); 4939 } 4940 ASSERT3U(HDR_GET_LSIZE(hdr), ==, BP_GET_LSIZE(bp)); 4941 ASSERT3U(HDR_GET_PSIZE(hdr), ==, BP_GET_PSIZE(bp)); 4942 } 4943} 4944 4945static void 4946arc_read_done(zio_t *zio) 4947{ 4948 arc_buf_hdr_t *hdr = zio->io_private; 4949 kmutex_t *hash_lock = NULL; 4950 arc_callback_t *callback_list; 4951 arc_callback_t *acb; 4952 boolean_t freeable = B_FALSE; 4953 boolean_t no_zio_error = (zio->io_error == 0); 4954 4955 /* 4956 * The hdr was inserted into hash-table and removed from lists 4957 * prior to starting I/O. We should find this header, since 4958 * it's in the hash table, and it should be legit since it's 4959 * not possible to evict it during the I/O. The only possible 4960 * reason for it not to be found is if we were freed during the 4961 * read. 4962 */ 4963 if (HDR_IN_HASH_TABLE(hdr)) { 4964 ASSERT3U(hdr->b_birth, ==, BP_PHYSICAL_BIRTH(zio->io_bp)); 4965 ASSERT3U(hdr->b_dva.dva_word[0], ==, 4966 BP_IDENTITY(zio->io_bp)->dva_word[0]); 4967 ASSERT3U(hdr->b_dva.dva_word[1], ==, 4968 BP_IDENTITY(zio->io_bp)->dva_word[1]); 4969 4970 arc_buf_hdr_t *found = buf_hash_find(hdr->b_spa, zio->io_bp, 4971 &hash_lock); 4972 4973 ASSERT((found == hdr && 4974 DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) || 4975 (found == hdr && HDR_L2_READING(hdr))); 4976 ASSERT3P(hash_lock, !=, NULL); 4977 } 4978 4979 if (no_zio_error) { 4980 /* byteswap if necessary */ 4981 if (BP_SHOULD_BYTESWAP(zio->io_bp)) { 4982 if (BP_GET_LEVEL(zio->io_bp) > 0) { 4983 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_UINT64; 4984 } else { 4985 hdr->b_l1hdr.b_byteswap = 4986 DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp)); 4987 } 4988 } else { 4989 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS; 4990 } 4991 } 4992 4993 arc_hdr_clear_flags(hdr, ARC_FLAG_L2_EVICTED); 4994 if (l2arc_noprefetch && HDR_PREFETCH(hdr)) 4995 arc_hdr_clear_flags(hdr, ARC_FLAG_L2CACHE); 4996 4997 callback_list = hdr->b_l1hdr.b_acb; 4998 ASSERT3P(callback_list, !=, NULL); 4999 5000 if (hash_lock && no_zio_error && hdr->b_l1hdr.b_state == arc_anon) { 5001 /* 5002 * Only call arc_access on anonymous buffers. This is because 5003 * if we've issued an I/O for an evicted buffer, we've already 5004 * called arc_access (to prevent any simultaneous readers from 5005 * getting confused). 5006 */ 5007 arc_access(hdr, hash_lock); 5008 } 5009 5010 /* 5011 * If a read request has a callback (i.e. acb_done is not NULL), then we 5012 * make a buf containing the data according to the parameters which were 5013 * passed in. The implementation of arc_buf_alloc_impl() ensures that we 5014 * aren't needlessly decompressing the data multiple times. 5015 */ 5016 int callback_cnt = 0; 5017 for (acb = callback_list; acb != NULL; acb = acb->acb_next) { 5018 if (!acb->acb_done) 5019 continue; 5020 5021 /* This is a demand read since prefetches don't use callbacks */ 5022 callback_cnt++; 5023 5024 int error = arc_buf_alloc_impl(hdr, acb->acb_private, 5025 acb->acb_compressed, no_zio_error, &acb->acb_buf); 5026 if (no_zio_error) { 5027 zio->io_error = error; 5028 } 5029 } 5030 hdr->b_l1hdr.b_acb = NULL; 5031 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS); 5032 if (callback_cnt == 0) { 5033 ASSERT(HDR_PREFETCH(hdr)); 5034 ASSERT0(hdr->b_l1hdr.b_bufcnt); 5035 ASSERT3P(hdr->b_l1hdr.b_pdata, !=, NULL); 5036 } 5037 5038 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt) || 5039 callback_list != NULL); 5040 5041 if (no_zio_error) { 5042 arc_hdr_verify(hdr, zio->io_bp); 5043 } else { 5044 arc_hdr_set_flags(hdr, ARC_FLAG_IO_ERROR); 5045 if (hdr->b_l1hdr.b_state != arc_anon) 5046 arc_change_state(arc_anon, hdr, hash_lock); 5047 if (HDR_IN_HASH_TABLE(hdr)) 5048 buf_hash_remove(hdr); 5049 freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt); 5050 } 5051 5052 /* 5053 * Broadcast before we drop the hash_lock to avoid the possibility 5054 * that the hdr (and hence the cv) might be freed before we get to 5055 * the cv_broadcast(). 5056 */ 5057 cv_broadcast(&hdr->b_l1hdr.b_cv); 5058 5059 if (hash_lock != NULL) { 5060 mutex_exit(hash_lock); 5061 } else { 5062 /* 5063 * This block was freed while we waited for the read to 5064 * complete. It has been removed from the hash table and 5065 * moved to the anonymous state (so that it won't show up 5066 * in the cache). 5067 */ 5068 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon); 5069 freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt); 5070 } 5071 5072 /* execute each callback and free its structure */ 5073 while ((acb = callback_list) != NULL) { 5074 if (acb->acb_done) 5075 acb->acb_done(zio, acb->acb_buf, acb->acb_private); 5076 5077 if (acb->acb_zio_dummy != NULL) { 5078 acb->acb_zio_dummy->io_error = zio->io_error; 5079 zio_nowait(acb->acb_zio_dummy); 5080 } 5081 5082 callback_list = acb->acb_next; 5083 kmem_free(acb, sizeof (arc_callback_t)); 5084 } 5085 5086 if (freeable) 5087 arc_hdr_destroy(hdr); 5088} 5089 5090/* 5091 * "Read" the block at the specified DVA (in bp) via the 5092 * cache. If the block is found in the cache, invoke the provided 5093 * callback immediately and return. Note that the `zio' parameter 5094 * in the callback will be NULL in this case, since no IO was 5095 * required. If the block is not in the cache pass the read request 5096 * on to the spa with a substitute callback function, so that the 5097 * requested block will be added to the cache. 5098 * 5099 * If a read request arrives for a block that has a read in-progress, 5100 * either wait for the in-progress read to complete (and return the 5101 * results); or, if this is a read with a "done" func, add a record 5102 * to the read to invoke the "done" func when the read completes, 5103 * and return; or just return. 5104 * 5105 * arc_read_done() will invoke all the requested "done" functions 5106 * for readers of this block. 5107 */ 5108int 5109arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_done_func_t *done, 5110 void *private, zio_priority_t priority, int zio_flags, 5111 arc_flags_t *arc_flags, const zbookmark_phys_t *zb) 5112{ 5113 arc_buf_hdr_t *hdr = NULL; 5114 kmutex_t *hash_lock = NULL; 5115 zio_t *rzio; 5116 uint64_t guid = spa_load_guid(spa); 5117 boolean_t compressed_read = (zio_flags & ZIO_FLAG_RAW) != 0; 5118 5119 ASSERT(!BP_IS_EMBEDDED(bp) || 5120 BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA); 5121 5122top: 5123 if (!BP_IS_EMBEDDED(bp)) { 5124 /* 5125 * Embedded BP's have no DVA and require no I/O to "read". 5126 * Create an anonymous arc buf to back it. 5127 */ 5128 hdr = buf_hash_find(guid, bp, &hash_lock); 5129 } 5130 5131 if (hdr != NULL && HDR_HAS_L1HDR(hdr) && hdr->b_l1hdr.b_pdata != NULL) { 5132 arc_buf_t *buf = NULL; 5133 *arc_flags |= ARC_FLAG_CACHED; 5134 5135 if (HDR_IO_IN_PROGRESS(hdr)) { 5136 5137 if ((hdr->b_flags & ARC_FLAG_PRIO_ASYNC_READ) && 5138 priority == ZIO_PRIORITY_SYNC_READ) { 5139 /* 5140 * This sync read must wait for an 5141 * in-progress async read (e.g. a predictive 5142 * prefetch). Async reads are queued 5143 * separately at the vdev_queue layer, so 5144 * this is a form of priority inversion. 5145 * Ideally, we would "inherit" the demand 5146 * i/o's priority by moving the i/o from 5147 * the async queue to the synchronous queue, 5148 * but there is currently no mechanism to do 5149 * so. Track this so that we can evaluate 5150 * the magnitude of this potential performance 5151 * problem. 5152 * 5153 * Note that if the prefetch i/o is already 5154 * active (has been issued to the device), 5155 * the prefetch improved performance, because 5156 * we issued it sooner than we would have 5157 * without the prefetch. 5158 */ 5159 DTRACE_PROBE1(arc__sync__wait__for__async, 5160 arc_buf_hdr_t *, hdr); 5161 ARCSTAT_BUMP(arcstat_sync_wait_for_async); 5162 } 5163 if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) { 5164 arc_hdr_clear_flags(hdr, 5165 ARC_FLAG_PREDICTIVE_PREFETCH); 5166 } 5167 5168 if (*arc_flags & ARC_FLAG_WAIT) { 5169 cv_wait(&hdr->b_l1hdr.b_cv, hash_lock); 5170 mutex_exit(hash_lock); 5171 goto top; 5172 } 5173 ASSERT(*arc_flags & ARC_FLAG_NOWAIT); 5174 5175 if (done) { 5176 arc_callback_t *acb = NULL; 5177 5178 acb = kmem_zalloc(sizeof (arc_callback_t), 5179 KM_SLEEP); 5180 acb->acb_done = done; 5181 acb->acb_private = private; 5182 acb->acb_compressed = compressed_read; 5183 if (pio != NULL) 5184 acb->acb_zio_dummy = zio_null(pio, 5185 spa, NULL, NULL, NULL, zio_flags); 5186 5187 ASSERT3P(acb->acb_done, !=, NULL); 5188 acb->acb_next = hdr->b_l1hdr.b_acb; 5189 hdr->b_l1hdr.b_acb = acb; 5190 mutex_exit(hash_lock); 5191 return (0); 5192 } 5193 mutex_exit(hash_lock); 5194 return (0); 5195 } 5196 5197 ASSERT(hdr->b_l1hdr.b_state == arc_mru || 5198 hdr->b_l1hdr.b_state == arc_mfu); 5199 5200 if (done) { 5201 if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) { 5202 /* 5203 * This is a demand read which does not have to 5204 * wait for i/o because we did a predictive 5205 * prefetch i/o for it, which has completed. 5206 */ 5207 DTRACE_PROBE1( 5208 arc__demand__hit__predictive__prefetch, 5209 arc_buf_hdr_t *, hdr); 5210 ARCSTAT_BUMP( 5211 arcstat_demand_hit_predictive_prefetch); 5212 arc_hdr_clear_flags(hdr, 5213 ARC_FLAG_PREDICTIVE_PREFETCH); 5214 } 5215 ASSERT(!BP_IS_EMBEDDED(bp) || !BP_IS_HOLE(bp)); 5216 5217 /* Get a buf with the desired data in it. */ 5218 VERIFY0(arc_buf_alloc_impl(hdr, private, 5219 compressed_read, B_TRUE, &buf)); 5220 } else if (*arc_flags & ARC_FLAG_PREFETCH && 5221 refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) { 5222 arc_hdr_set_flags(hdr, ARC_FLAG_PREFETCH); 5223 } 5224 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr); 5225 arc_access(hdr, hash_lock); 5226 if (*arc_flags & ARC_FLAG_L2CACHE) 5227 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE); 5228 mutex_exit(hash_lock); 5229 ARCSTAT_BUMP(arcstat_hits); 5230 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr), 5231 demand, prefetch, !HDR_ISTYPE_METADATA(hdr), 5232 data, metadata, hits); 5233 5234 if (done) 5235 done(NULL, buf, private); 5236 } else { 5237 uint64_t lsize = BP_GET_LSIZE(bp); 5238 uint64_t psize = BP_GET_PSIZE(bp); 5239 arc_callback_t *acb; 5240 vdev_t *vd = NULL; 5241 uint64_t addr = 0; 5242 boolean_t devw = B_FALSE; 5243 uint64_t size; 5244 5245 if (hdr == NULL) { 5246 /* this block is not in the cache */ 5247 arc_buf_hdr_t *exists = NULL; 5248 arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp); 5249 hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize, 5250 BP_GET_COMPRESS(bp), type); 5251 5252 if (!BP_IS_EMBEDDED(bp)) { 5253 hdr->b_dva = *BP_IDENTITY(bp); 5254 hdr->b_birth = BP_PHYSICAL_BIRTH(bp); 5255 exists = buf_hash_insert(hdr, &hash_lock); 5256 } 5257 if (exists != NULL) { 5258 /* somebody beat us to the hash insert */ 5259 mutex_exit(hash_lock); 5260 buf_discard_identity(hdr); 5261 arc_hdr_destroy(hdr); 5262 goto top; /* restart the IO request */ 5263 } 5264 } else { 5265 /* 5266 * This block is in the ghost cache. If it was L2-only 5267 * (and thus didn't have an L1 hdr), we realloc the 5268 * header to add an L1 hdr. 5269 */ 5270 if (!HDR_HAS_L1HDR(hdr)) { 5271 hdr = arc_hdr_realloc(hdr, hdr_l2only_cache, 5272 hdr_full_cache); 5273 } 5274 ASSERT3P(hdr->b_l1hdr.b_pdata, ==, NULL); 5275 ASSERT(GHOST_STATE(hdr->b_l1hdr.b_state)); 5276 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 5277 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 5278 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 5279 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL); 5280 5281 /* 5282 * This is a delicate dance that we play here. 5283 * This hdr is in the ghost list so we access it 5284 * to move it out of the ghost list before we 5285 * initiate the read. If it's a prefetch then 5286 * it won't have a callback so we'll remove the 5287 * reference that arc_buf_alloc_impl() created. We 5288 * do this after we've called arc_access() to 5289 * avoid hitting an assert in remove_reference(). 5290 */ 5291 arc_access(hdr, hash_lock); 5292 arc_hdr_alloc_pdata(hdr); 5293 } 5294 ASSERT3P(hdr->b_l1hdr.b_pdata, !=, NULL); 5295 size = arc_hdr_size(hdr); 5296 5297 /* 5298 * If compression is enabled on the hdr, then will do 5299 * RAW I/O and will store the compressed data in the hdr's 5300 * data block. Otherwise, the hdr's data block will contain 5301 * the uncompressed data. 5302 */ 5303 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF) { 5304 zio_flags |= ZIO_FLAG_RAW; 5305 } 5306 5307 if (*arc_flags & ARC_FLAG_PREFETCH) 5308 arc_hdr_set_flags(hdr, ARC_FLAG_PREFETCH); 5309 if (*arc_flags & ARC_FLAG_L2CACHE) 5310 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE); 5311 if (BP_GET_LEVEL(bp) > 0) 5312 arc_hdr_set_flags(hdr, ARC_FLAG_INDIRECT); 5313 if (*arc_flags & ARC_FLAG_PREDICTIVE_PREFETCH) 5314 arc_hdr_set_flags(hdr, ARC_FLAG_PREDICTIVE_PREFETCH); 5315 ASSERT(!GHOST_STATE(hdr->b_l1hdr.b_state)); 5316 5317 acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP); 5318 acb->acb_done = done; 5319 acb->acb_private = private; 5320 acb->acb_compressed = compressed_read; 5321 5322 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL); 5323 hdr->b_l1hdr.b_acb = acb; 5324 arc_hdr_set_flags(hdr, ARC_FLAG_IO_IN_PROGRESS); 5325 5326 if (HDR_HAS_L2HDR(hdr) && 5327 (vd = hdr->b_l2hdr.b_dev->l2ad_vdev) != NULL) { 5328 devw = hdr->b_l2hdr.b_dev->l2ad_writing; 5329 addr = hdr->b_l2hdr.b_daddr; 5330 /* 5331 * Lock out device removal. 5332 */ 5333 if (vdev_is_dead(vd) || 5334 !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER)) 5335 vd = NULL; 5336 } 5337 5338 if (priority == ZIO_PRIORITY_ASYNC_READ) 5339 arc_hdr_set_flags(hdr, ARC_FLAG_PRIO_ASYNC_READ); 5340 else 5341 arc_hdr_clear_flags(hdr, ARC_FLAG_PRIO_ASYNC_READ); 5342 5343 if (hash_lock != NULL) 5344 mutex_exit(hash_lock); 5345 5346 /* 5347 * At this point, we have a level 1 cache miss. Try again in 5348 * L2ARC if possible. 5349 */ 5350 ASSERT3U(HDR_GET_LSIZE(hdr), ==, lsize); 5351 5352 DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp, 5353 uint64_t, lsize, zbookmark_phys_t *, zb); 5354 ARCSTAT_BUMP(arcstat_misses); 5355 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr), 5356 demand, prefetch, !HDR_ISTYPE_METADATA(hdr), 5357 data, metadata, misses); 5358#ifdef _KERNEL 5359#ifdef RACCT 5360 if (racct_enable) { 5361 PROC_LOCK(curproc); 5362 racct_add_force(curproc, RACCT_READBPS, size); 5363 racct_add_force(curproc, RACCT_READIOPS, 1); 5364 PROC_UNLOCK(curproc); 5365 } 5366#endif /* RACCT */ 5367 curthread->td_ru.ru_inblock++; 5368#endif 5369 5370 if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) { 5371 /* 5372 * Read from the L2ARC if the following are true: 5373 * 1. The L2ARC vdev was previously cached. 5374 * 2. This buffer still has L2ARC metadata. 5375 * 3. This buffer isn't currently writing to the L2ARC. 5376 * 4. The L2ARC entry wasn't evicted, which may 5377 * also have invalidated the vdev. 5378 * 5. This isn't prefetch and l2arc_noprefetch is set. 5379 */ 5380 if (HDR_HAS_L2HDR(hdr) && 5381 !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) && 5382 !(l2arc_noprefetch && HDR_PREFETCH(hdr))) { 5383 l2arc_read_callback_t *cb; 5384 void* b_data; 5385 5386 DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr); 5387 ARCSTAT_BUMP(arcstat_l2_hits); 5388 5389 cb = kmem_zalloc(sizeof (l2arc_read_callback_t), 5390 KM_SLEEP); 5391 cb->l2rcb_hdr = hdr; 5392 cb->l2rcb_bp = *bp; 5393 cb->l2rcb_zb = *zb; 5394 cb->l2rcb_flags = zio_flags; 5395 uint64_t asize = vdev_psize_to_asize(vd, size); 5396 if (asize != size) { 5397 b_data = zio_data_buf_alloc(asize); 5398 cb->l2rcb_data = b_data; 5399 } else { 5400 b_data = hdr->b_l1hdr.b_pdata; 5401 } 5402 5403 ASSERT(addr >= VDEV_LABEL_START_SIZE && 5404 addr + asize < vd->vdev_psize - 5405 VDEV_LABEL_END_SIZE); 5406 5407 /* 5408 * l2arc read. The SCL_L2ARC lock will be 5409 * released by l2arc_read_done(). 5410 * Issue a null zio if the underlying buffer 5411 * was squashed to zero size by compression. 5412 */ 5413 ASSERT3U(HDR_GET_COMPRESS(hdr), !=, 5414 ZIO_COMPRESS_EMPTY); 5415 rzio = zio_read_phys(pio, vd, addr, 5416 asize, b_data, 5417 ZIO_CHECKSUM_OFF, 5418 l2arc_read_done, cb, priority, 5419 zio_flags | ZIO_FLAG_DONT_CACHE | 5420 ZIO_FLAG_CANFAIL | 5421 ZIO_FLAG_DONT_PROPAGATE | 5422 ZIO_FLAG_DONT_RETRY, B_FALSE); 5423 DTRACE_PROBE2(l2arc__read, vdev_t *, vd, 5424 zio_t *, rzio); 5425 ARCSTAT_INCR(arcstat_l2_read_bytes, size); 5426 5427 if (*arc_flags & ARC_FLAG_NOWAIT) { 5428 zio_nowait(rzio); 5429 return (0); 5430 } 5431 5432 ASSERT(*arc_flags & ARC_FLAG_WAIT); 5433 if (zio_wait(rzio) == 0) 5434 return (0); 5435 5436 /* l2arc read error; goto zio_read() */ 5437 } else { 5438 DTRACE_PROBE1(l2arc__miss, 5439 arc_buf_hdr_t *, hdr); 5440 ARCSTAT_BUMP(arcstat_l2_misses); 5441 if (HDR_L2_WRITING(hdr)) 5442 ARCSTAT_BUMP(arcstat_l2_rw_clash); 5443 spa_config_exit(spa, SCL_L2ARC, vd); 5444 } 5445 } else { 5446 if (vd != NULL) 5447 spa_config_exit(spa, SCL_L2ARC, vd); 5448 if (l2arc_ndev != 0) { 5449 DTRACE_PROBE1(l2arc__miss, 5450 arc_buf_hdr_t *, hdr); 5451 ARCSTAT_BUMP(arcstat_l2_misses); 5452 } 5453 } 5454 5455 rzio = zio_read(pio, spa, bp, hdr->b_l1hdr.b_pdata, size, 5456 arc_read_done, hdr, priority, zio_flags, zb); 5457 5458 if (*arc_flags & ARC_FLAG_WAIT) 5459 return (zio_wait(rzio)); 5460 5461 ASSERT(*arc_flags & ARC_FLAG_NOWAIT); 5462 zio_nowait(rzio); 5463 } 5464 return (0); 5465} 5466 5467/* 5468 * Notify the arc that a block was freed, and thus will never be used again. 5469 */ 5470void 5471arc_freed(spa_t *spa, const blkptr_t *bp) 5472{ 5473 arc_buf_hdr_t *hdr; 5474 kmutex_t *hash_lock; 5475 uint64_t guid = spa_load_guid(spa); 5476 5477 ASSERT(!BP_IS_EMBEDDED(bp)); 5478 5479 hdr = buf_hash_find(guid, bp, &hash_lock); 5480 if (hdr == NULL) 5481 return; 5482 5483 /* 5484 * We might be trying to free a block that is still doing I/O 5485 * (i.e. prefetch) or has a reference (i.e. a dedup-ed, 5486 * dmu_sync-ed block). If this block is being prefetched, then it 5487 * would still have the ARC_FLAG_IO_IN_PROGRESS flag set on the hdr 5488 * until the I/O completes. A block may also have a reference if it is 5489 * part of a dedup-ed, dmu_synced write. The dmu_sync() function would 5490 * have written the new block to its final resting place on disk but 5491 * without the dedup flag set. This would have left the hdr in the MRU 5492 * state and discoverable. When the txg finally syncs it detects that 5493 * the block was overridden in open context and issues an override I/O. 5494 * Since this is a dedup block, the override I/O will determine if the 5495 * block is already in the DDT. If so, then it will replace the io_bp 5496 * with the bp from the DDT and allow the I/O to finish. When the I/O 5497 * reaches the done callback, dbuf_write_override_done, it will 5498 * check to see if the io_bp and io_bp_override are identical. 5499 * If they are not, then it indicates that the bp was replaced with 5500 * the bp in the DDT and the override bp is freed. This allows 5501 * us to arrive here with a reference on a block that is being 5502 * freed. So if we have an I/O in progress, or a reference to 5503 * this hdr, then we don't destroy the hdr. 5504 */ 5505 if (!HDR_HAS_L1HDR(hdr) || (!HDR_IO_IN_PROGRESS(hdr) && 5506 refcount_is_zero(&hdr->b_l1hdr.b_refcnt))) { 5507 arc_change_state(arc_anon, hdr, hash_lock); 5508 arc_hdr_destroy(hdr); 5509 mutex_exit(hash_lock); 5510 } else { 5511 mutex_exit(hash_lock); 5512 } 5513 5514} 5515 5516/* 5517 * Release this buffer from the cache, making it an anonymous buffer. This 5518 * must be done after a read and prior to modifying the buffer contents. 5519 * If the buffer has more than one reference, we must make 5520 * a new hdr for the buffer. 5521 */ 5522void 5523arc_release(arc_buf_t *buf, void *tag) 5524{ 5525 arc_buf_hdr_t *hdr = buf->b_hdr; 5526 5527 /* 5528 * It would be nice to assert that if it's DMU metadata (level > 5529 * 0 || it's the dnode file), then it must be syncing context. 5530 * But we don't know that information at this level. 5531 */ 5532 5533 mutex_enter(&buf->b_evict_lock); 5534 5535 ASSERT(HDR_HAS_L1HDR(hdr)); 5536 5537 /* 5538 * We don't grab the hash lock prior to this check, because if 5539 * the buffer's header is in the arc_anon state, it won't be 5540 * linked into the hash table. 5541 */ 5542 if (hdr->b_l1hdr.b_state == arc_anon) { 5543 mutex_exit(&buf->b_evict_lock); 5544 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 5545 ASSERT(!HDR_IN_HASH_TABLE(hdr)); 5546 ASSERT(!HDR_HAS_L2HDR(hdr)); 5547 ASSERT(HDR_EMPTY(hdr)); 5548 ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1); 5549 ASSERT3S(refcount_count(&hdr->b_l1hdr.b_refcnt), ==, 1); 5550 ASSERT(!list_link_active(&hdr->b_l1hdr.b_arc_node)); 5551 5552 hdr->b_l1hdr.b_arc_access = 0; 5553 5554 /* 5555 * If the buf is being overridden then it may already 5556 * have a hdr that is not empty. 5557 */ 5558 buf_discard_identity(hdr); 5559 arc_buf_thaw(buf); 5560 5561 return; 5562 } 5563 5564 kmutex_t *hash_lock = HDR_LOCK(hdr); 5565 mutex_enter(hash_lock); 5566 5567 /* 5568 * This assignment is only valid as long as the hash_lock is 5569 * held, we must be careful not to reference state or the 5570 * b_state field after dropping the lock. 5571 */ 5572 arc_state_t *state = hdr->b_l1hdr.b_state; 5573 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 5574 ASSERT3P(state, !=, arc_anon); 5575 5576 /* this buffer is not on any list */ 5577 ASSERT3S(refcount_count(&hdr->b_l1hdr.b_refcnt), >, 0); 5578 5579 if (HDR_HAS_L2HDR(hdr)) { 5580 mutex_enter(&hdr->b_l2hdr.b_dev->l2ad_mtx); 5581 5582 /* 5583 * We have to recheck this conditional again now that 5584 * we're holding the l2ad_mtx to prevent a race with 5585 * another thread which might be concurrently calling 5586 * l2arc_evict(). In that case, l2arc_evict() might have 5587 * destroyed the header's L2 portion as we were waiting 5588 * to acquire the l2ad_mtx. 5589 */ 5590 if (HDR_HAS_L2HDR(hdr)) { 5591 l2arc_trim(hdr); 5592 arc_hdr_l2hdr_destroy(hdr); 5593 } 5594 5595 mutex_exit(&hdr->b_l2hdr.b_dev->l2ad_mtx); 5596 } 5597 5598 /* 5599 * Do we have more than one buf? 5600 */ 5601 if (hdr->b_l1hdr.b_bufcnt > 1) { 5602 arc_buf_hdr_t *nhdr; 5603 uint64_t spa = hdr->b_spa; 5604 uint64_t psize = HDR_GET_PSIZE(hdr); 5605 uint64_t lsize = HDR_GET_LSIZE(hdr); 5606 enum zio_compress compress = HDR_GET_COMPRESS(hdr); 5607 arc_buf_contents_t type = arc_buf_type(hdr); 5608 VERIFY3U(hdr->b_type, ==, type); 5609 5610 ASSERT(hdr->b_l1hdr.b_buf != buf || buf->b_next != NULL); 5611 (void) remove_reference(hdr, hash_lock, tag); 5612 5613 if (arc_buf_is_shared(buf) && !ARC_BUF_COMPRESSED(buf)) { 5614 ASSERT3P(hdr->b_l1hdr.b_buf, !=, buf); 5615 ASSERT(ARC_BUF_LAST(buf)); 5616 } 5617 5618 /* 5619 * Pull the data off of this hdr and attach it to 5620 * a new anonymous hdr. Also find the last buffer 5621 * in the hdr's buffer list. 5622 */ 5623 arc_buf_t *lastbuf = arc_buf_remove(hdr, buf); 5624 ASSERT3P(lastbuf, !=, NULL); 5625 5626 /* 5627 * If the current arc_buf_t and the hdr are sharing their data 5628 * buffer, then we must stop sharing that block. 5629 */ 5630 if (arc_buf_is_shared(buf)) { 5631 VERIFY(!arc_buf_is_shared(lastbuf)); 5632 5633 /* 5634 * First, sever the block sharing relationship between 5635 * buf and the arc_buf_hdr_t. 5636 */ 5637 arc_unshare_buf(hdr, buf); 5638 5639 /* 5640 * Now we need to recreate the hdr's b_pdata. Since we 5641 * have lastbuf handy, we try to share with it, but if 5642 * we can't then we allocate a new b_pdata and copy the 5643 * data from buf into it. 5644 */ 5645 if (arc_can_share(hdr, lastbuf)) { 5646 arc_share_buf(hdr, lastbuf); 5647 } else { 5648 arc_hdr_alloc_pdata(hdr); 5649 bcopy(buf->b_data, hdr->b_l1hdr.b_pdata, psize); 5650 } 5651 VERIFY3P(lastbuf->b_data, !=, NULL); 5652 } else if (HDR_SHARED_DATA(hdr)) { 5653 /* 5654 * Uncompressed shared buffers are always at the end 5655 * of the list. Compressed buffers don't have the 5656 * same requirements. This makes it hard to 5657 * simply assert that the lastbuf is shared so 5658 * we rely on the hdr's compression flags to determine 5659 * if we have a compressed, shared buffer. 5660 */ 5661 ASSERT(arc_buf_is_shared(lastbuf) || 5662 HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF); 5663 ASSERT(!ARC_BUF_SHARED(buf)); 5664 } 5665 ASSERT3P(hdr->b_l1hdr.b_pdata, !=, NULL); 5666 ASSERT3P(state, !=, arc_l2c_only); 5667 5668 (void) refcount_remove_many(&state->arcs_size, 5669 arc_buf_size(buf), buf); 5670 5671 if (refcount_is_zero(&hdr->b_l1hdr.b_refcnt)) { 5672 ASSERT3P(state, !=, arc_l2c_only); 5673 (void) refcount_remove_many(&state->arcs_esize[type], 5674 arc_buf_size(buf), buf); 5675 } 5676 5677 hdr->b_l1hdr.b_bufcnt -= 1; 5678 arc_cksum_verify(buf); 5679#ifdef illumos 5680 arc_buf_unwatch(buf); 5681#endif 5682 5683 mutex_exit(hash_lock); 5684 5685 /* 5686 * Allocate a new hdr. The new hdr will contain a b_pdata 5687 * buffer which will be freed in arc_write(). 5688 */ 5689 nhdr = arc_hdr_alloc(spa, psize, lsize, compress, type); 5690 ASSERT3P(nhdr->b_l1hdr.b_buf, ==, NULL); 5691 ASSERT0(nhdr->b_l1hdr.b_bufcnt); 5692 ASSERT0(refcount_count(&nhdr->b_l1hdr.b_refcnt)); 5693 VERIFY3U(nhdr->b_type, ==, type); 5694 ASSERT(!HDR_SHARED_DATA(nhdr)); 5695 5696 nhdr->b_l1hdr.b_buf = buf; 5697 nhdr->b_l1hdr.b_bufcnt = 1; 5698 (void) refcount_add(&nhdr->b_l1hdr.b_refcnt, tag); 5699 buf->b_hdr = nhdr; 5700 5701 mutex_exit(&buf->b_evict_lock); 5702 (void) refcount_add_many(&arc_anon->arcs_size, 5703 arc_buf_size(buf), buf); 5704 } else { 5705 mutex_exit(&buf->b_evict_lock); 5706 ASSERT(refcount_count(&hdr->b_l1hdr.b_refcnt) == 1); 5707 /* protected by hash lock, or hdr is on arc_anon */ 5708 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 5709 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 5710 arc_change_state(arc_anon, hdr, hash_lock); 5711 hdr->b_l1hdr.b_arc_access = 0; 5712 mutex_exit(hash_lock); 5713 5714 buf_discard_identity(hdr); 5715 arc_buf_thaw(buf); 5716 } 5717} 5718 5719int 5720arc_released(arc_buf_t *buf) 5721{ 5722 int released; 5723 5724 mutex_enter(&buf->b_evict_lock); 5725 released = (buf->b_data != NULL && 5726 buf->b_hdr->b_l1hdr.b_state == arc_anon); 5727 mutex_exit(&buf->b_evict_lock); 5728 return (released); 5729} 5730 5731#ifdef ZFS_DEBUG 5732int 5733arc_referenced(arc_buf_t *buf) 5734{ 5735 int referenced; 5736 5737 mutex_enter(&buf->b_evict_lock); 5738 referenced = (refcount_count(&buf->b_hdr->b_l1hdr.b_refcnt)); 5739 mutex_exit(&buf->b_evict_lock); 5740 return (referenced); 5741} 5742#endif 5743 5744static void 5745arc_write_ready(zio_t *zio) 5746{ 5747 arc_write_callback_t *callback = zio->io_private; 5748 arc_buf_t *buf = callback->awcb_buf; 5749 arc_buf_hdr_t *hdr = buf->b_hdr; 5750 uint64_t psize = BP_IS_HOLE(zio->io_bp) ? 0 : BP_GET_PSIZE(zio->io_bp); 5751 5752 ASSERT(HDR_HAS_L1HDR(hdr)); 5753 ASSERT(!refcount_is_zero(&buf->b_hdr->b_l1hdr.b_refcnt)); 5754 ASSERT(hdr->b_l1hdr.b_bufcnt > 0); 5755 5756 /* 5757 * If we're reexecuting this zio because the pool suspended, then 5758 * cleanup any state that was previously set the first time the 5759 * callback was invoked. 5760 */ 5761 if (zio->io_flags & ZIO_FLAG_REEXECUTED) { 5762 arc_cksum_free(hdr); 5763#ifdef illumos 5764 arc_buf_unwatch(buf); 5765#endif 5766 if (hdr->b_l1hdr.b_pdata != NULL) { 5767 if (arc_buf_is_shared(buf)) { 5768 arc_unshare_buf(hdr, buf); 5769 } else { 5770 arc_hdr_free_pdata(hdr); 5771 } 5772 } 5773 } 5774 ASSERT3P(hdr->b_l1hdr.b_pdata, ==, NULL); 5775 ASSERT(!HDR_SHARED_DATA(hdr)); 5776 ASSERT(!arc_buf_is_shared(buf)); 5777 5778 callback->awcb_ready(zio, buf, callback->awcb_private); 5779 5780 if (HDR_IO_IN_PROGRESS(hdr)) 5781 ASSERT(zio->io_flags & ZIO_FLAG_REEXECUTED); 5782 5783 arc_cksum_compute(buf); 5784 arc_hdr_set_flags(hdr, ARC_FLAG_IO_IN_PROGRESS); 5785 5786 enum zio_compress compress; 5787 if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) { 5788 compress = ZIO_COMPRESS_OFF; 5789 } else { 5790 ASSERT3U(HDR_GET_LSIZE(hdr), ==, BP_GET_LSIZE(zio->io_bp)); 5791 compress = BP_GET_COMPRESS(zio->io_bp); 5792 } 5793 HDR_SET_PSIZE(hdr, psize); 5794 arc_hdr_set_compress(hdr, compress); 5795 5796 /* 5797 * If the hdr is compressed, then copy the compressed 5798 * zio contents into arc_buf_hdr_t. Otherwise, copy the original 5799 * data buf into the hdr. Ideally, we would like to always copy the 5800 * io_data into b_pdata but the user may have disabled compressed 5801 * arc thus the on-disk block may or may not match what we maintain 5802 * in the hdr's b_pdata field. 5803 */ 5804 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF && 5805 !ARC_BUF_COMPRESSED(buf)) { 5806 ASSERT3U(BP_GET_COMPRESS(zio->io_bp), !=, ZIO_COMPRESS_OFF); 5807 ASSERT3U(psize, >, 0); 5808 arc_hdr_alloc_pdata(hdr); 5809 bcopy(zio->io_data, hdr->b_l1hdr.b_pdata, psize); 5810 } else { 5811 ASSERT3P(buf->b_data, ==, zio->io_orig_data); 5812 ASSERT3U(zio->io_orig_size, ==, arc_buf_size(buf)); 5813 ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1); 5814 5815 /* 5816 * This hdr is not compressed so we're able to share 5817 * the arc_buf_t data buffer with the hdr. 5818 */ 5819 arc_share_buf(hdr, buf); 5820 ASSERT0(bcmp(zio->io_orig_data, hdr->b_l1hdr.b_pdata, 5821 arc_buf_size(buf))); 5822 } 5823 arc_hdr_verify(hdr, zio->io_bp); 5824} 5825 5826static void 5827arc_write_children_ready(zio_t *zio) 5828{ 5829 arc_write_callback_t *callback = zio->io_private; 5830 arc_buf_t *buf = callback->awcb_buf; 5831 5832 callback->awcb_children_ready(zio, buf, callback->awcb_private); 5833} 5834 5835/* 5836 * The SPA calls this callback for each physical write that happens on behalf 5837 * of a logical write. See the comment in dbuf_write_physdone() for details. 5838 */ 5839static void 5840arc_write_physdone(zio_t *zio) 5841{ 5842 arc_write_callback_t *cb = zio->io_private; 5843 if (cb->awcb_physdone != NULL) 5844 cb->awcb_physdone(zio, cb->awcb_buf, cb->awcb_private); 5845} 5846 5847static void 5848arc_write_done(zio_t *zio) 5849{ 5850 arc_write_callback_t *callback = zio->io_private; 5851 arc_buf_t *buf = callback->awcb_buf; 5852 arc_buf_hdr_t *hdr = buf->b_hdr; 5853 5854 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL); 5855 5856 if (zio->io_error == 0) { 5857 arc_hdr_verify(hdr, zio->io_bp); 5858 5859 if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) { 5860 buf_discard_identity(hdr); 5861 } else { 5862 hdr->b_dva = *BP_IDENTITY(zio->io_bp); 5863 hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp); 5864 } 5865 } else { 5866 ASSERT(HDR_EMPTY(hdr)); 5867 } 5868 5869 /* 5870 * If the block to be written was all-zero or compressed enough to be 5871 * embedded in the BP, no write was performed so there will be no 5872 * dva/birth/checksum. The buffer must therefore remain anonymous 5873 * (and uncached). 5874 */ 5875 if (!HDR_EMPTY(hdr)) { 5876 arc_buf_hdr_t *exists; 5877 kmutex_t *hash_lock; 5878 5879 ASSERT3U(zio->io_error, ==, 0); 5880 5881 arc_cksum_verify(buf); 5882 5883 exists = buf_hash_insert(hdr, &hash_lock); 5884 if (exists != NULL) { 5885 /* 5886 * This can only happen if we overwrite for 5887 * sync-to-convergence, because we remove 5888 * buffers from the hash table when we arc_free(). 5889 */ 5890 if (zio->io_flags & ZIO_FLAG_IO_REWRITE) { 5891 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) 5892 panic("bad overwrite, hdr=%p exists=%p", 5893 (void *)hdr, (void *)exists); 5894 ASSERT(refcount_is_zero( 5895 &exists->b_l1hdr.b_refcnt)); 5896 arc_change_state(arc_anon, exists, hash_lock); 5897 mutex_exit(hash_lock); 5898 arc_hdr_destroy(exists); 5899 exists = buf_hash_insert(hdr, &hash_lock); 5900 ASSERT3P(exists, ==, NULL); 5901 } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) { 5902 /* nopwrite */ 5903 ASSERT(zio->io_prop.zp_nopwrite); 5904 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) 5905 panic("bad nopwrite, hdr=%p exists=%p", 5906 (void *)hdr, (void *)exists); 5907 } else { 5908 /* Dedup */ 5909 ASSERT(hdr->b_l1hdr.b_bufcnt == 1); 5910 ASSERT(hdr->b_l1hdr.b_state == arc_anon); 5911 ASSERT(BP_GET_DEDUP(zio->io_bp)); 5912 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0); 5913 } 5914 } 5915 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS); 5916 /* if it's not anon, we are doing a scrub */ 5917 if (exists == NULL && hdr->b_l1hdr.b_state == arc_anon) 5918 arc_access(hdr, hash_lock); 5919 mutex_exit(hash_lock); 5920 } else { 5921 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS); 5922 } 5923 5924 ASSERT(!refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 5925 callback->awcb_done(zio, buf, callback->awcb_private); 5926 5927 kmem_free(callback, sizeof (arc_write_callback_t)); 5928} 5929 5930zio_t * 5931arc_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, arc_buf_t *buf, 5932 boolean_t l2arc, const zio_prop_t *zp, arc_done_func_t *ready, 5933 arc_done_func_t *children_ready, arc_done_func_t *physdone, 5934 arc_done_func_t *done, void *private, zio_priority_t priority, 5935 int zio_flags, const zbookmark_phys_t *zb) 5936{ 5937 arc_buf_hdr_t *hdr = buf->b_hdr; 5938 arc_write_callback_t *callback; 5939 zio_t *zio; 5940 zio_prop_t localprop = *zp; 5941 5942 ASSERT3P(ready, !=, NULL); 5943 ASSERT3P(done, !=, NULL); 5944 ASSERT(!HDR_IO_ERROR(hdr)); 5945 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 5946 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL); 5947 ASSERT3U(hdr->b_l1hdr.b_bufcnt, >, 0); 5948 if (l2arc) 5949 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE); 5950 if (ARC_BUF_COMPRESSED(buf)) { 5951 /* 5952 * We're writing a pre-compressed buffer. Make the 5953 * compression algorithm requested by the zio_prop_t match 5954 * the pre-compressed buffer's compression algorithm. 5955 */ 5956 localprop.zp_compress = HDR_GET_COMPRESS(hdr); 5957 5958 ASSERT3U(HDR_GET_LSIZE(hdr), !=, arc_buf_size(buf)); 5959 zio_flags |= ZIO_FLAG_RAW; 5960 } 5961 callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP); 5962 callback->awcb_ready = ready; 5963 callback->awcb_children_ready = children_ready; 5964 callback->awcb_physdone = physdone; 5965 callback->awcb_done = done; 5966 callback->awcb_private = private; 5967 callback->awcb_buf = buf; 5968 5969 /* 5970 * The hdr's b_pdata is now stale, free it now. A new data block 5971 * will be allocated when the zio pipeline calls arc_write_ready(). 5972 */ 5973 if (hdr->b_l1hdr.b_pdata != NULL) { 5974 /* 5975 * If the buf is currently sharing the data block with 5976 * the hdr then we need to break that relationship here. 5977 * The hdr will remain with a NULL data pointer and the 5978 * buf will take sole ownership of the block. 5979 */ 5980 if (arc_buf_is_shared(buf)) { 5981 arc_unshare_buf(hdr, buf); 5982 } else { 5983 arc_hdr_free_pdata(hdr); 5984 } 5985 VERIFY3P(buf->b_data, !=, NULL); 5986 arc_hdr_set_compress(hdr, ZIO_COMPRESS_OFF); 5987 } 5988 ASSERT(!arc_buf_is_shared(buf)); 5989 ASSERT3P(hdr->b_l1hdr.b_pdata, ==, NULL); 5990 5991 zio = zio_write(pio, spa, txg, bp, buf->b_data, 5992 HDR_GET_LSIZE(hdr), arc_buf_size(buf), &localprop, arc_write_ready, 5993 (children_ready != NULL) ? arc_write_children_ready : NULL, 5994 arc_write_physdone, arc_write_done, callback, 5995 priority, zio_flags, zb); 5996 5997 return (zio); 5998} 5999 6000static int 6001arc_memory_throttle(uint64_t reserve, uint64_t txg) 6002{ 6003#ifdef _KERNEL 6004 uint64_t available_memory = ptob(freemem); 6005 static uint64_t page_load = 0; 6006 static uint64_t last_txg = 0; 6007 6008#if defined(__i386) || !defined(UMA_MD_SMALL_ALLOC) 6009 available_memory = 6010 MIN(available_memory, ptob(vmem_size(heap_arena, VMEM_FREE))); 6011#endif 6012 6013 if (freemem > (uint64_t)physmem * arc_lotsfree_percent / 100) 6014 return (0); 6015 6016 if (txg > last_txg) { 6017 last_txg = txg; 6018 page_load = 0; 6019 } 6020 /* 6021 * If we are in pageout, we know that memory is already tight, 6022 * the arc is already going to be evicting, so we just want to 6023 * continue to let page writes occur as quickly as possible. 6024 */ 6025 if (curproc == pageproc) { 6026 if (page_load > MAX(ptob(minfree), available_memory) / 4) 6027 return (SET_ERROR(ERESTART)); 6028 /* Note: reserve is inflated, so we deflate */ 6029 page_load += reserve / 8; 6030 return (0); 6031 } else if (page_load > 0 && arc_reclaim_needed()) { 6032 /* memory is low, delay before restarting */ 6033 ARCSTAT_INCR(arcstat_memory_throttle_count, 1); 6034 return (SET_ERROR(EAGAIN)); 6035 } 6036 page_load = 0; 6037#endif 6038 return (0); 6039} 6040 6041void 6042arc_tempreserve_clear(uint64_t reserve) 6043{ 6044 atomic_add_64(&arc_tempreserve, -reserve); 6045 ASSERT((int64_t)arc_tempreserve >= 0); 6046} 6047 6048int 6049arc_tempreserve_space(uint64_t reserve, uint64_t txg) 6050{ 6051 int error; 6052 uint64_t anon_size; 6053 6054 if (reserve > arc_c/4 && !arc_no_grow) { 6055 arc_c = MIN(arc_c_max, reserve * 4); 6056 DTRACE_PROBE1(arc__set_reserve, uint64_t, arc_c); 6057 } 6058 if (reserve > arc_c) 6059 return (SET_ERROR(ENOMEM)); 6060 6061 /* 6062 * Don't count loaned bufs as in flight dirty data to prevent long 6063 * network delays from blocking transactions that are ready to be 6064 * assigned to a txg. 6065 */ 6066 6067 /* assert that it has not wrapped around */ 6068 ASSERT3S(atomic_add_64_nv(&arc_loaned_bytes, 0), >=, 0); 6069 6070 anon_size = MAX((int64_t)(refcount_count(&arc_anon->arcs_size) - 6071 arc_loaned_bytes), 0); 6072 6073 /* 6074 * Writes will, almost always, require additional memory allocations 6075 * in order to compress/encrypt/etc the data. We therefore need to 6076 * make sure that there is sufficient available memory for this. 6077 */ 6078 error = arc_memory_throttle(reserve, txg); 6079 if (error != 0) 6080 return (error); 6081 6082 /* 6083 * Throttle writes when the amount of dirty data in the cache 6084 * gets too large. We try to keep the cache less than half full 6085 * of dirty blocks so that our sync times don't grow too large. 6086 * Note: if two requests come in concurrently, we might let them 6087 * both succeed, when one of them should fail. Not a huge deal. 6088 */ 6089 6090 if (reserve + arc_tempreserve + anon_size > arc_c / 2 && 6091 anon_size > arc_c / 4) { 6092 uint64_t meta_esize = 6093 refcount_count(&arc_anon->arcs_esize[ARC_BUFC_METADATA]); 6094 uint64_t data_esize = 6095 refcount_count(&arc_anon->arcs_esize[ARC_BUFC_DATA]); 6096 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK " 6097 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n", 6098 arc_tempreserve >> 10, meta_esize >> 10, 6099 data_esize >> 10, reserve >> 10, arc_c >> 10); 6100 return (SET_ERROR(ERESTART)); 6101 } 6102 atomic_add_64(&arc_tempreserve, reserve); 6103 return (0); 6104} 6105 6106static void 6107arc_kstat_update_state(arc_state_t *state, kstat_named_t *size, 6108 kstat_named_t *evict_data, kstat_named_t *evict_metadata) 6109{ 6110 size->value.ui64 = refcount_count(&state->arcs_size); 6111 evict_data->value.ui64 = 6112 refcount_count(&state->arcs_esize[ARC_BUFC_DATA]); 6113 evict_metadata->value.ui64 = 6114 refcount_count(&state->arcs_esize[ARC_BUFC_METADATA]); 6115} 6116 6117static int 6118arc_kstat_update(kstat_t *ksp, int rw) 6119{ 6120 arc_stats_t *as = ksp->ks_data; 6121 6122 if (rw == KSTAT_WRITE) { 6123 return (EACCES); 6124 } else { 6125 arc_kstat_update_state(arc_anon, 6126 &as->arcstat_anon_size, 6127 &as->arcstat_anon_evictable_data, 6128 &as->arcstat_anon_evictable_metadata); 6129 arc_kstat_update_state(arc_mru, 6130 &as->arcstat_mru_size, 6131 &as->arcstat_mru_evictable_data, 6132 &as->arcstat_mru_evictable_metadata); 6133 arc_kstat_update_state(arc_mru_ghost, 6134 &as->arcstat_mru_ghost_size, 6135 &as->arcstat_mru_ghost_evictable_data, 6136 &as->arcstat_mru_ghost_evictable_metadata); 6137 arc_kstat_update_state(arc_mfu, 6138 &as->arcstat_mfu_size, 6139 &as->arcstat_mfu_evictable_data, 6140 &as->arcstat_mfu_evictable_metadata); 6141 arc_kstat_update_state(arc_mfu_ghost, 6142 &as->arcstat_mfu_ghost_size, 6143 &as->arcstat_mfu_ghost_evictable_data, 6144 &as->arcstat_mfu_ghost_evictable_metadata); 6145 } 6146 6147 return (0); 6148} 6149 6150/* 6151 * This function *must* return indices evenly distributed between all 6152 * sublists of the multilist. This is needed due to how the ARC eviction 6153 * code is laid out; arc_evict_state() assumes ARC buffers are evenly 6154 * distributed between all sublists and uses this assumption when 6155 * deciding which sublist to evict from and how much to evict from it. 6156 */ 6157unsigned int 6158arc_state_multilist_index_func(multilist_t *ml, void *obj) 6159{ 6160 arc_buf_hdr_t *hdr = obj; 6161 6162 /* 6163 * We rely on b_dva to generate evenly distributed index 6164 * numbers using buf_hash below. So, as an added precaution, 6165 * let's make sure we never add empty buffers to the arc lists. 6166 */ 6167 ASSERT(!HDR_EMPTY(hdr)); 6168 6169 /* 6170 * The assumption here, is the hash value for a given 6171 * arc_buf_hdr_t will remain constant throughout it's lifetime 6172 * (i.e. it's b_spa, b_dva, and b_birth fields don't change). 6173 * Thus, we don't need to store the header's sublist index 6174 * on insertion, as this index can be recalculated on removal. 6175 * 6176 * Also, the low order bits of the hash value are thought to be 6177 * distributed evenly. Otherwise, in the case that the multilist 6178 * has a power of two number of sublists, each sublists' usage 6179 * would not be evenly distributed. 6180 */ 6181 return (buf_hash(hdr->b_spa, &hdr->b_dva, hdr->b_birth) % 6182 multilist_get_num_sublists(ml)); 6183} 6184 6185#ifdef _KERNEL 6186static eventhandler_tag arc_event_lowmem = NULL; 6187 6188static void 6189arc_lowmem(void *arg __unused, int howto __unused) 6190{ 6191 6192 mutex_enter(&arc_reclaim_lock); 6193 /* XXX: Memory deficit should be passed as argument. */ 6194 needfree = btoc(arc_c >> arc_shrink_shift); 6195 DTRACE_PROBE(arc__needfree); 6196 cv_signal(&arc_reclaim_thread_cv); 6197 6198 /* 6199 * It is unsafe to block here in arbitrary threads, because we can come 6200 * here from ARC itself and may hold ARC locks and thus risk a deadlock 6201 * with ARC reclaim thread. 6202 */ 6203 if (curproc == pageproc) 6204 (void) cv_wait(&arc_reclaim_waiters_cv, &arc_reclaim_lock); 6205 mutex_exit(&arc_reclaim_lock); 6206} 6207#endif 6208 6209static void 6210arc_state_init(void) 6211{ 6212 arc_anon = &ARC_anon; 6213 arc_mru = &ARC_mru; 6214 arc_mru_ghost = &ARC_mru_ghost; 6215 arc_mfu = &ARC_mfu; 6216 arc_mfu_ghost = &ARC_mfu_ghost; 6217 arc_l2c_only = &ARC_l2c_only; 6218 6219 arc_mru->arcs_list[ARC_BUFC_METADATA] = 6220 multilist_create(sizeof (arc_buf_hdr_t), 6221 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 6222 arc_state_multilist_index_func); 6223 arc_mru->arcs_list[ARC_BUFC_DATA] = 6224 multilist_create(sizeof (arc_buf_hdr_t), 6225 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 6226 arc_state_multilist_index_func); 6227 arc_mru_ghost->arcs_list[ARC_BUFC_METADATA] = 6228 multilist_create(sizeof (arc_buf_hdr_t), 6229 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 6230 arc_state_multilist_index_func); 6231 arc_mru_ghost->arcs_list[ARC_BUFC_DATA] = 6232 multilist_create(sizeof (arc_buf_hdr_t), 6233 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 6234 arc_state_multilist_index_func); 6235 arc_mfu->arcs_list[ARC_BUFC_METADATA] = 6236 multilist_create(sizeof (arc_buf_hdr_t), 6237 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 6238 arc_state_multilist_index_func); 6239 arc_mfu->arcs_list[ARC_BUFC_DATA] = 6240 multilist_create(sizeof (arc_buf_hdr_t), 6241 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 6242 arc_state_multilist_index_func); 6243 arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA] = 6244 multilist_create(sizeof (arc_buf_hdr_t), 6245 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 6246 arc_state_multilist_index_func); 6247 arc_mfu_ghost->arcs_list[ARC_BUFC_DATA] = 6248 multilist_create(sizeof (arc_buf_hdr_t), 6249 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 6250 arc_state_multilist_index_func); 6251 arc_l2c_only->arcs_list[ARC_BUFC_METADATA] = 6252 multilist_create(sizeof (arc_buf_hdr_t), 6253 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 6254 arc_state_multilist_index_func); 6255 arc_l2c_only->arcs_list[ARC_BUFC_DATA] = 6256 multilist_create(sizeof (arc_buf_hdr_t), 6257 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 6258 arc_state_multilist_index_func); 6259 6260 refcount_create(&arc_anon->arcs_esize[ARC_BUFC_METADATA]); 6261 refcount_create(&arc_anon->arcs_esize[ARC_BUFC_DATA]); 6262 refcount_create(&arc_mru->arcs_esize[ARC_BUFC_METADATA]); 6263 refcount_create(&arc_mru->arcs_esize[ARC_BUFC_DATA]); 6264 refcount_create(&arc_mru_ghost->arcs_esize[ARC_BUFC_METADATA]); 6265 refcount_create(&arc_mru_ghost->arcs_esize[ARC_BUFC_DATA]); 6266 refcount_create(&arc_mfu->arcs_esize[ARC_BUFC_METADATA]); 6267 refcount_create(&arc_mfu->arcs_esize[ARC_BUFC_DATA]); 6268 refcount_create(&arc_mfu_ghost->arcs_esize[ARC_BUFC_METADATA]); 6269 refcount_create(&arc_mfu_ghost->arcs_esize[ARC_BUFC_DATA]); 6270 refcount_create(&arc_l2c_only->arcs_esize[ARC_BUFC_METADATA]); 6271 refcount_create(&arc_l2c_only->arcs_esize[ARC_BUFC_DATA]); 6272 6273 refcount_create(&arc_anon->arcs_size); 6274 refcount_create(&arc_mru->arcs_size); 6275 refcount_create(&arc_mru_ghost->arcs_size); 6276 refcount_create(&arc_mfu->arcs_size); 6277 refcount_create(&arc_mfu_ghost->arcs_size); 6278 refcount_create(&arc_l2c_only->arcs_size); 6279} 6280 6281static void 6282arc_state_fini(void) 6283{ 6284 refcount_destroy(&arc_anon->arcs_esize[ARC_BUFC_METADATA]); 6285 refcount_destroy(&arc_anon->arcs_esize[ARC_BUFC_DATA]); 6286 refcount_destroy(&arc_mru->arcs_esize[ARC_BUFC_METADATA]); 6287 refcount_destroy(&arc_mru->arcs_esize[ARC_BUFC_DATA]); 6288 refcount_destroy(&arc_mru_ghost->arcs_esize[ARC_BUFC_METADATA]); 6289 refcount_destroy(&arc_mru_ghost->arcs_esize[ARC_BUFC_DATA]); 6290 refcount_destroy(&arc_mfu->arcs_esize[ARC_BUFC_METADATA]); 6291 refcount_destroy(&arc_mfu->arcs_esize[ARC_BUFC_DATA]); 6292 refcount_destroy(&arc_mfu_ghost->arcs_esize[ARC_BUFC_METADATA]); 6293 refcount_destroy(&arc_mfu_ghost->arcs_esize[ARC_BUFC_DATA]); 6294 refcount_destroy(&arc_l2c_only->arcs_esize[ARC_BUFC_METADATA]); 6295 refcount_destroy(&arc_l2c_only->arcs_esize[ARC_BUFC_DATA]); 6296 6297 refcount_destroy(&arc_anon->arcs_size); 6298 refcount_destroy(&arc_mru->arcs_size); 6299 refcount_destroy(&arc_mru_ghost->arcs_size); 6300 refcount_destroy(&arc_mfu->arcs_size); 6301 refcount_destroy(&arc_mfu_ghost->arcs_size); 6302 refcount_destroy(&arc_l2c_only->arcs_size); 6303 6304 multilist_destroy(arc_mru->arcs_list[ARC_BUFC_METADATA]); 6305 multilist_destroy(arc_mru_ghost->arcs_list[ARC_BUFC_METADATA]); 6306 multilist_destroy(arc_mfu->arcs_list[ARC_BUFC_METADATA]); 6307 multilist_destroy(arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA]); 6308 multilist_destroy(arc_mru->arcs_list[ARC_BUFC_DATA]); 6309 multilist_destroy(arc_mru_ghost->arcs_list[ARC_BUFC_DATA]); 6310 multilist_destroy(arc_mfu->arcs_list[ARC_BUFC_DATA]); 6311 multilist_destroy(arc_mfu_ghost->arcs_list[ARC_BUFC_DATA]); 6312} 6313 6314uint64_t 6315arc_max_bytes(void) 6316{ 6317 return (arc_c_max); 6318} 6319 6320void 6321arc_init(void) 6322{ 6323 int i, prefetch_tunable_set = 0; 6324 6325 /* 6326 * allmem is "all memory that we could possibly use". 6327 */ 6328#ifdef illumos 6329#ifdef _KERNEL 6330 uint64_t allmem = ptob(physmem - swapfs_minfree); 6331#else 6332 uint64_t allmem = (physmem * PAGESIZE) / 2; 6333#endif 6334#else 6335 uint64_t allmem = kmem_size(); 6336#endif 6337 6338 6339 mutex_init(&arc_reclaim_lock, NULL, MUTEX_DEFAULT, NULL); 6340 cv_init(&arc_reclaim_thread_cv, NULL, CV_DEFAULT, NULL); 6341 cv_init(&arc_reclaim_waiters_cv, NULL, CV_DEFAULT, NULL); 6342 6343 mutex_init(&arc_dnlc_evicts_lock, NULL, MUTEX_DEFAULT, NULL); 6344 cv_init(&arc_dnlc_evicts_cv, NULL, CV_DEFAULT, NULL); 6345 6346 /* Convert seconds to clock ticks */ 6347 arc_min_prefetch_lifespan = 1 * hz; 6348 6349 /* set min cache to 1/32 of all memory, or arc_abs_min, whichever is more */ 6350 arc_c_min = MAX(allmem / 32, arc_abs_min); 6351 /* set max to 5/8 of all memory, or all but 1GB, whichever is more */ 6352 if (allmem >= 1 << 30) 6353 arc_c_max = allmem - (1 << 30); 6354 else 6355 arc_c_max = arc_c_min; 6356 arc_c_max = MAX(allmem * 5 / 8, arc_c_max); 6357 6358 /* 6359 * In userland, there's only the memory pressure that we artificially 6360 * create (see arc_available_memory()). Don't let arc_c get too 6361 * small, because it can cause transactions to be larger than 6362 * arc_c, causing arc_tempreserve_space() to fail. 6363 */ 6364#ifndef _KERNEL 6365 arc_c_min = arc_c_max / 2; 6366#endif 6367 6368#ifdef _KERNEL 6369 /* 6370 * Allow the tunables to override our calculations if they are 6371 * reasonable. 6372 */ 6373 if (zfs_arc_max > arc_abs_min && zfs_arc_max < allmem) { 6374 arc_c_max = zfs_arc_max; 6375 arc_c_min = MIN(arc_c_min, arc_c_max); 6376 } 6377 if (zfs_arc_min > arc_abs_min && zfs_arc_min <= arc_c_max) 6378 arc_c_min = zfs_arc_min; 6379#endif 6380 6381 arc_c = arc_c_max; 6382 arc_p = (arc_c >> 1); 6383 arc_size = 0; 6384 6385 /* limit meta-data to 1/4 of the arc capacity */ 6386 arc_meta_limit = arc_c_max / 4; 6387 6388#ifdef _KERNEL 6389 /* 6390 * Metadata is stored in the kernel's heap. Don't let us 6391 * use more than half the heap for the ARC. 6392 */ 6393 arc_meta_limit = MIN(arc_meta_limit, 6394 vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 2); 6395#endif 6396 6397 /* Allow the tunable to override if it is reasonable */ 6398 if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max) 6399 arc_meta_limit = zfs_arc_meta_limit; 6400 6401 if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0) 6402 arc_c_min = arc_meta_limit / 2; 6403 6404 if (zfs_arc_meta_min > 0) { 6405 arc_meta_min = zfs_arc_meta_min; 6406 } else { 6407 arc_meta_min = arc_c_min / 2; 6408 } 6409 6410 if (zfs_arc_grow_retry > 0) 6411 arc_grow_retry = zfs_arc_grow_retry; 6412 6413 if (zfs_arc_shrink_shift > 0) 6414 arc_shrink_shift = zfs_arc_shrink_shift; 6415 6416 /* 6417 * Ensure that arc_no_grow_shift is less than arc_shrink_shift. 6418 */ 6419 if (arc_no_grow_shift >= arc_shrink_shift) 6420 arc_no_grow_shift = arc_shrink_shift - 1; 6421 6422 if (zfs_arc_p_min_shift > 0) 6423 arc_p_min_shift = zfs_arc_p_min_shift; 6424 6425 /* if kmem_flags are set, lets try to use less memory */ 6426 if (kmem_debugging()) 6427 arc_c = arc_c / 2; 6428 if (arc_c < arc_c_min) 6429 arc_c = arc_c_min; 6430 6431 zfs_arc_min = arc_c_min; 6432 zfs_arc_max = arc_c_max; 6433 6434 arc_state_init(); 6435 buf_init(); 6436 6437 arc_reclaim_thread_exit = B_FALSE; 6438 arc_dnlc_evicts_thread_exit = FALSE; 6439 6440 arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED, 6441 sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL); 6442 6443 if (arc_ksp != NULL) { 6444 arc_ksp->ks_data = &arc_stats; 6445 arc_ksp->ks_update = arc_kstat_update; 6446 kstat_install(arc_ksp); 6447 } 6448 6449 (void) thread_create(NULL, 0, arc_reclaim_thread, NULL, 0, &p0, 6450 TS_RUN, minclsyspri); 6451 6452#ifdef _KERNEL 6453 arc_event_lowmem = EVENTHANDLER_REGISTER(vm_lowmem, arc_lowmem, NULL, 6454 EVENTHANDLER_PRI_FIRST); 6455#endif 6456 6457 (void) thread_create(NULL, 0, arc_dnlc_evicts_thread, NULL, 0, &p0, 6458 TS_RUN, minclsyspri); 6459 6460 arc_dead = B_FALSE; 6461 arc_warm = B_FALSE; 6462 6463 /* 6464 * Calculate maximum amount of dirty data per pool. 6465 * 6466 * If it has been set by /etc/system, take that. 6467 * Otherwise, use a percentage of physical memory defined by 6468 * zfs_dirty_data_max_percent (default 10%) with a cap at 6469 * zfs_dirty_data_max_max (default 4GB). 6470 */ 6471 if (zfs_dirty_data_max == 0) { 6472 zfs_dirty_data_max = ptob(physmem) * 6473 zfs_dirty_data_max_percent / 100; 6474 zfs_dirty_data_max = MIN(zfs_dirty_data_max, 6475 zfs_dirty_data_max_max); 6476 } 6477 6478#ifdef _KERNEL 6479 if (TUNABLE_INT_FETCH("vfs.zfs.prefetch_disable", &zfs_prefetch_disable)) 6480 prefetch_tunable_set = 1; 6481 6482#ifdef __i386__ 6483 if (prefetch_tunable_set == 0) { 6484 printf("ZFS NOTICE: Prefetch is disabled by default on i386 " 6485 "-- to enable,\n"); 6486 printf(" add \"vfs.zfs.prefetch_disable=0\" " 6487 "to /boot/loader.conf.\n"); 6488 zfs_prefetch_disable = 1; 6489 } 6490#else 6491 if ((((uint64_t)physmem * PAGESIZE) < (1ULL << 32)) && 6492 prefetch_tunable_set == 0) { 6493 printf("ZFS NOTICE: Prefetch is disabled by default if less " 6494 "than 4GB of RAM is present;\n" 6495 " to enable, add \"vfs.zfs.prefetch_disable=0\" " 6496 "to /boot/loader.conf.\n"); 6497 zfs_prefetch_disable = 1; 6498 } 6499#endif 6500 /* Warn about ZFS memory and address space requirements. */ 6501 if (((uint64_t)physmem * PAGESIZE) < (256 + 128 + 64) * (1 << 20)) { 6502 printf("ZFS WARNING: Recommended minimum RAM size is 512MB; " 6503 "expect unstable behavior.\n"); 6504 } 6505 if (allmem < 512 * (1 << 20)) { 6506 printf("ZFS WARNING: Recommended minimum kmem_size is 512MB; " 6507 "expect unstable behavior.\n"); 6508 printf(" Consider tuning vm.kmem_size and " 6509 "vm.kmem_size_max\n"); 6510 printf(" in /boot/loader.conf.\n"); 6511 } 6512#endif 6513} 6514 6515void 6516arc_fini(void) 6517{ 6518 mutex_enter(&arc_reclaim_lock); 6519 arc_reclaim_thread_exit = B_TRUE; 6520 /* 6521 * The reclaim thread will set arc_reclaim_thread_exit back to 6522 * B_FALSE when it is finished exiting; we're waiting for that. 6523 */ 6524 while (arc_reclaim_thread_exit) { 6525 cv_signal(&arc_reclaim_thread_cv); 6526 cv_wait(&arc_reclaim_thread_cv, &arc_reclaim_lock); 6527 } 6528 mutex_exit(&arc_reclaim_lock); 6529 6530 /* Use B_TRUE to ensure *all* buffers are evicted */ 6531 arc_flush(NULL, B_TRUE); 6532 6533 mutex_enter(&arc_dnlc_evicts_lock); 6534 arc_dnlc_evicts_thread_exit = TRUE; 6535 /* 6536 * The user evicts thread will set arc_user_evicts_thread_exit 6537 * to FALSE when it is finished exiting; we're waiting for that. 6538 */ 6539 while (arc_dnlc_evicts_thread_exit) { 6540 cv_signal(&arc_dnlc_evicts_cv); 6541 cv_wait(&arc_dnlc_evicts_cv, &arc_dnlc_evicts_lock); 6542 } 6543 mutex_exit(&arc_dnlc_evicts_lock); 6544 6545 arc_dead = B_TRUE; 6546 6547 if (arc_ksp != NULL) { 6548 kstat_delete(arc_ksp); 6549 arc_ksp = NULL; 6550 } 6551 6552 mutex_destroy(&arc_reclaim_lock); 6553 cv_destroy(&arc_reclaim_thread_cv); 6554 cv_destroy(&arc_reclaim_waiters_cv); 6555 6556 mutex_destroy(&arc_dnlc_evicts_lock); 6557 cv_destroy(&arc_dnlc_evicts_cv); 6558 6559 arc_state_fini(); 6560 buf_fini(); 6561 6562 ASSERT0(arc_loaned_bytes); 6563 6564#ifdef _KERNEL 6565 if (arc_event_lowmem != NULL) 6566 EVENTHANDLER_DEREGISTER(vm_lowmem, arc_event_lowmem); 6567#endif 6568} 6569 6570/* 6571 * Level 2 ARC 6572 * 6573 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk. 6574 * It uses dedicated storage devices to hold cached data, which are populated 6575 * using large infrequent writes. The main role of this cache is to boost 6576 * the performance of random read workloads. The intended L2ARC devices 6577 * include short-stroked disks, solid state disks, and other media with 6578 * substantially faster read latency than disk. 6579 * 6580 * +-----------------------+ 6581 * | ARC | 6582 * +-----------------------+ 6583 * | ^ ^ 6584 * | | | 6585 * l2arc_feed_thread() arc_read() 6586 * | | | 6587 * | l2arc read | 6588 * V | | 6589 * +---------------+ | 6590 * | L2ARC | | 6591 * +---------------+ | 6592 * | ^ | 6593 * l2arc_write() | | 6594 * | | | 6595 * V | | 6596 * +-------+ +-------+ 6597 * | vdev | | vdev | 6598 * | cache | | cache | 6599 * +-------+ +-------+ 6600 * +=========+ .-----. 6601 * : L2ARC : |-_____-| 6602 * : devices : | Disks | 6603 * +=========+ `-_____-' 6604 * 6605 * Read requests are satisfied from the following sources, in order: 6606 * 6607 * 1) ARC 6608 * 2) vdev cache of L2ARC devices 6609 * 3) L2ARC devices 6610 * 4) vdev cache of disks 6611 * 5) disks 6612 * 6613 * Some L2ARC device types exhibit extremely slow write performance. 6614 * To accommodate for this there are some significant differences between 6615 * the L2ARC and traditional cache design: 6616 * 6617 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from 6618 * the ARC behave as usual, freeing buffers and placing headers on ghost 6619 * lists. The ARC does not send buffers to the L2ARC during eviction as 6620 * this would add inflated write latencies for all ARC memory pressure. 6621 * 6622 * 2. The L2ARC attempts to cache data from the ARC before it is evicted. 6623 * It does this by periodically scanning buffers from the eviction-end of 6624 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are 6625 * not already there. It scans until a headroom of buffers is satisfied, 6626 * which itself is a buffer for ARC eviction. If a compressible buffer is 6627 * found during scanning and selected for writing to an L2ARC device, we 6628 * temporarily boost scanning headroom during the next scan cycle to make 6629 * sure we adapt to compression effects (which might significantly reduce 6630 * the data volume we write to L2ARC). The thread that does this is 6631 * l2arc_feed_thread(), illustrated below; example sizes are included to 6632 * provide a better sense of ratio than this diagram: 6633 * 6634 * head --> tail 6635 * +---------------------+----------+ 6636 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC 6637 * +---------------------+----------+ | o L2ARC eligible 6638 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer 6639 * +---------------------+----------+ | 6640 * 15.9 Gbytes ^ 32 Mbytes | 6641 * headroom | 6642 * l2arc_feed_thread() 6643 * | 6644 * l2arc write hand <--[oooo]--' 6645 * | 8 Mbyte 6646 * | write max 6647 * V 6648 * +==============================+ 6649 * L2ARC dev |####|#|###|###| |####| ... | 6650 * +==============================+ 6651 * 32 Gbytes 6652 * 6653 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of 6654 * evicted, then the L2ARC has cached a buffer much sooner than it probably 6655 * needed to, potentially wasting L2ARC device bandwidth and storage. It is 6656 * safe to say that this is an uncommon case, since buffers at the end of 6657 * the ARC lists have moved there due to inactivity. 6658 * 6659 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom, 6660 * then the L2ARC simply misses copying some buffers. This serves as a 6661 * pressure valve to prevent heavy read workloads from both stalling the ARC 6662 * with waits and clogging the L2ARC with writes. This also helps prevent 6663 * the potential for the L2ARC to churn if it attempts to cache content too 6664 * quickly, such as during backups of the entire pool. 6665 * 6666 * 5. After system boot and before the ARC has filled main memory, there are 6667 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru 6668 * lists can remain mostly static. Instead of searching from tail of these 6669 * lists as pictured, the l2arc_feed_thread() will search from the list heads 6670 * for eligible buffers, greatly increasing its chance of finding them. 6671 * 6672 * The L2ARC device write speed is also boosted during this time so that 6673 * the L2ARC warms up faster. Since there have been no ARC evictions yet, 6674 * there are no L2ARC reads, and no fear of degrading read performance 6675 * through increased writes. 6676 * 6677 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that 6678 * the vdev queue can aggregate them into larger and fewer writes. Each 6679 * device is written to in a rotor fashion, sweeping writes through 6680 * available space then repeating. 6681 * 6682 * 7. The L2ARC does not store dirty content. It never needs to flush 6683 * write buffers back to disk based storage. 6684 * 6685 * 8. If an ARC buffer is written (and dirtied) which also exists in the 6686 * L2ARC, the now stale L2ARC buffer is immediately dropped. 6687 * 6688 * The performance of the L2ARC can be tweaked by a number of tunables, which 6689 * may be necessary for different workloads: 6690 * 6691 * l2arc_write_max max write bytes per interval 6692 * l2arc_write_boost extra write bytes during device warmup 6693 * l2arc_noprefetch skip caching prefetched buffers 6694 * l2arc_headroom number of max device writes to precache 6695 * l2arc_headroom_boost when we find compressed buffers during ARC 6696 * scanning, we multiply headroom by this 6697 * percentage factor for the next scan cycle, 6698 * since more compressed buffers are likely to 6699 * be present 6700 * l2arc_feed_secs seconds between L2ARC writing 6701 * 6702 * Tunables may be removed or added as future performance improvements are 6703 * integrated, and also may become zpool properties. 6704 * 6705 * There are three key functions that control how the L2ARC warms up: 6706 * 6707 * l2arc_write_eligible() check if a buffer is eligible to cache 6708 * l2arc_write_size() calculate how much to write 6709 * l2arc_write_interval() calculate sleep delay between writes 6710 * 6711 * These three functions determine what to write, how much, and how quickly 6712 * to send writes. 6713 */ 6714 6715static boolean_t 6716l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *hdr) 6717{ 6718 /* 6719 * A buffer is *not* eligible for the L2ARC if it: 6720 * 1. belongs to a different spa. 6721 * 2. is already cached on the L2ARC. 6722 * 3. has an I/O in progress (it may be an incomplete read). 6723 * 4. is flagged not eligible (zfs property). 6724 */ 6725 if (hdr->b_spa != spa_guid) { 6726 ARCSTAT_BUMP(arcstat_l2_write_spa_mismatch); 6727 return (B_FALSE); 6728 } 6729 if (HDR_HAS_L2HDR(hdr)) { 6730 ARCSTAT_BUMP(arcstat_l2_write_in_l2); 6731 return (B_FALSE); 6732 } 6733 if (HDR_IO_IN_PROGRESS(hdr)) { 6734 ARCSTAT_BUMP(arcstat_l2_write_hdr_io_in_progress); 6735 return (B_FALSE); 6736 } 6737 if (!HDR_L2CACHE(hdr)) { 6738 ARCSTAT_BUMP(arcstat_l2_write_not_cacheable); 6739 return (B_FALSE); 6740 } 6741 6742 return (B_TRUE); 6743} 6744 6745static uint64_t 6746l2arc_write_size(void) 6747{ 6748 uint64_t size; 6749 6750 /* 6751 * Make sure our globals have meaningful values in case the user 6752 * altered them. 6753 */ 6754 size = l2arc_write_max; 6755 if (size == 0) { 6756 cmn_err(CE_NOTE, "Bad value for l2arc_write_max, value must " 6757 "be greater than zero, resetting it to the default (%d)", 6758 L2ARC_WRITE_SIZE); 6759 size = l2arc_write_max = L2ARC_WRITE_SIZE; 6760 } 6761 6762 if (arc_warm == B_FALSE) 6763 size += l2arc_write_boost; 6764 6765 return (size); 6766 6767} 6768 6769static clock_t 6770l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote) 6771{ 6772 clock_t interval, next, now; 6773 6774 /* 6775 * If the ARC lists are busy, increase our write rate; if the 6776 * lists are stale, idle back. This is achieved by checking 6777 * how much we previously wrote - if it was more than half of 6778 * what we wanted, schedule the next write much sooner. 6779 */ 6780 if (l2arc_feed_again && wrote > (wanted / 2)) 6781 interval = (hz * l2arc_feed_min_ms) / 1000; 6782 else 6783 interval = hz * l2arc_feed_secs; 6784 6785 now = ddi_get_lbolt(); 6786 next = MAX(now, MIN(now + interval, began + interval)); 6787 6788 return (next); 6789} 6790 6791/* 6792 * Cycle through L2ARC devices. This is how L2ARC load balances. 6793 * If a device is returned, this also returns holding the spa config lock. 6794 */ 6795static l2arc_dev_t * 6796l2arc_dev_get_next(void) 6797{ 6798 l2arc_dev_t *first, *next = NULL; 6799 6800 /* 6801 * Lock out the removal of spas (spa_namespace_lock), then removal 6802 * of cache devices (l2arc_dev_mtx). Once a device has been selected, 6803 * both locks will be dropped and a spa config lock held instead. 6804 */ 6805 mutex_enter(&spa_namespace_lock); 6806 mutex_enter(&l2arc_dev_mtx); 6807 6808 /* if there are no vdevs, there is nothing to do */ 6809 if (l2arc_ndev == 0) 6810 goto out; 6811 6812 first = NULL; 6813 next = l2arc_dev_last; 6814 do { 6815 /* loop around the list looking for a non-faulted vdev */ 6816 if (next == NULL) { 6817 next = list_head(l2arc_dev_list); 6818 } else { 6819 next = list_next(l2arc_dev_list, next); 6820 if (next == NULL) 6821 next = list_head(l2arc_dev_list); 6822 } 6823 6824 /* if we have come back to the start, bail out */ 6825 if (first == NULL) 6826 first = next; 6827 else if (next == first) 6828 break; 6829 6830 } while (vdev_is_dead(next->l2ad_vdev)); 6831 6832 /* if we were unable to find any usable vdevs, return NULL */ 6833 if (vdev_is_dead(next->l2ad_vdev)) 6834 next = NULL; 6835 6836 l2arc_dev_last = next; 6837 6838out: 6839 mutex_exit(&l2arc_dev_mtx); 6840 6841 /* 6842 * Grab the config lock to prevent the 'next' device from being 6843 * removed while we are writing to it. 6844 */ 6845 if (next != NULL) 6846 spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER); 6847 mutex_exit(&spa_namespace_lock); 6848 6849 return (next); 6850} 6851 6852/* 6853 * Free buffers that were tagged for destruction. 6854 */ 6855static void 6856l2arc_do_free_on_write() 6857{ 6858 list_t *buflist; 6859 l2arc_data_free_t *df, *df_prev; 6860 6861 mutex_enter(&l2arc_free_on_write_mtx); 6862 buflist = l2arc_free_on_write; 6863 6864 for (df = list_tail(buflist); df; df = df_prev) { 6865 df_prev = list_prev(buflist, df); 6866 ASSERT3P(df->l2df_data, !=, NULL); 6867 if (df->l2df_type == ARC_BUFC_METADATA) { 6868 zio_buf_free(df->l2df_data, df->l2df_size); 6869 } else { 6870 ASSERT(df->l2df_type == ARC_BUFC_DATA); 6871 zio_data_buf_free(df->l2df_data, df->l2df_size); 6872 } 6873 list_remove(buflist, df); 6874 kmem_free(df, sizeof (l2arc_data_free_t)); 6875 } 6876 6877 mutex_exit(&l2arc_free_on_write_mtx); 6878} 6879 6880/* 6881 * A write to a cache device has completed. Update all headers to allow 6882 * reads from these buffers to begin. 6883 */ 6884static void 6885l2arc_write_done(zio_t *zio) 6886{ 6887 l2arc_write_callback_t *cb; 6888 l2arc_dev_t *dev; 6889 list_t *buflist; 6890 arc_buf_hdr_t *head, *hdr, *hdr_prev; 6891 kmutex_t *hash_lock; 6892 int64_t bytes_dropped = 0; 6893 6894 cb = zio->io_private; 6895 ASSERT3P(cb, !=, NULL); 6896 dev = cb->l2wcb_dev; 6897 ASSERT3P(dev, !=, NULL); 6898 head = cb->l2wcb_head; 6899 ASSERT3P(head, !=, NULL); 6900 buflist = &dev->l2ad_buflist; 6901 ASSERT3P(buflist, !=, NULL); 6902 DTRACE_PROBE2(l2arc__iodone, zio_t *, zio, 6903 l2arc_write_callback_t *, cb); 6904 6905 if (zio->io_error != 0) 6906 ARCSTAT_BUMP(arcstat_l2_writes_error); 6907 6908 /* 6909 * All writes completed, or an error was hit. 6910 */ 6911top: 6912 mutex_enter(&dev->l2ad_mtx); 6913 for (hdr = list_prev(buflist, head); hdr; hdr = hdr_prev) { 6914 hdr_prev = list_prev(buflist, hdr); 6915 6916 hash_lock = HDR_LOCK(hdr); 6917 6918 /* 6919 * We cannot use mutex_enter or else we can deadlock 6920 * with l2arc_write_buffers (due to swapping the order 6921 * the hash lock and l2ad_mtx are taken). 6922 */ 6923 if (!mutex_tryenter(hash_lock)) { 6924 /* 6925 * Missed the hash lock. We must retry so we 6926 * don't leave the ARC_FLAG_L2_WRITING bit set. 6927 */ 6928 ARCSTAT_BUMP(arcstat_l2_writes_lock_retry); 6929 6930 /* 6931 * We don't want to rescan the headers we've 6932 * already marked as having been written out, so 6933 * we reinsert the head node so we can pick up 6934 * where we left off. 6935 */ 6936 list_remove(buflist, head); 6937 list_insert_after(buflist, hdr, head); 6938 6939 mutex_exit(&dev->l2ad_mtx); 6940 6941 /* 6942 * We wait for the hash lock to become available 6943 * to try and prevent busy waiting, and increase 6944 * the chance we'll be able to acquire the lock 6945 * the next time around. 6946 */ 6947 mutex_enter(hash_lock); 6948 mutex_exit(hash_lock); 6949 goto top; 6950 } 6951 6952 /* 6953 * We could not have been moved into the arc_l2c_only 6954 * state while in-flight due to our ARC_FLAG_L2_WRITING 6955 * bit being set. Let's just ensure that's being enforced. 6956 */ 6957 ASSERT(HDR_HAS_L1HDR(hdr)); 6958 6959 if (zio->io_error != 0) { 6960 /* 6961 * Error - drop L2ARC entry. 6962 */ 6963 list_remove(buflist, hdr); 6964 l2arc_trim(hdr); 6965 arc_hdr_clear_flags(hdr, ARC_FLAG_HAS_L2HDR); 6966 6967 ARCSTAT_INCR(arcstat_l2_asize, -arc_hdr_size(hdr)); 6968 ARCSTAT_INCR(arcstat_l2_size, -HDR_GET_LSIZE(hdr)); 6969 6970 bytes_dropped += arc_hdr_size(hdr); 6971 (void) refcount_remove_many(&dev->l2ad_alloc, 6972 arc_hdr_size(hdr), hdr); 6973 } 6974 6975 /* 6976 * Allow ARC to begin reads and ghost list evictions to 6977 * this L2ARC entry. 6978 */ 6979 arc_hdr_clear_flags(hdr, ARC_FLAG_L2_WRITING); 6980 6981 mutex_exit(hash_lock); 6982 } 6983 6984 atomic_inc_64(&l2arc_writes_done); 6985 list_remove(buflist, head); 6986 ASSERT(!HDR_HAS_L1HDR(head)); 6987 kmem_cache_free(hdr_l2only_cache, head); 6988 mutex_exit(&dev->l2ad_mtx); 6989 6990 vdev_space_update(dev->l2ad_vdev, -bytes_dropped, 0, 0); 6991 6992 l2arc_do_free_on_write(); 6993 6994 kmem_free(cb, sizeof (l2arc_write_callback_t)); 6995} 6996 6997/* 6998 * A read to a cache device completed. Validate buffer contents before 6999 * handing over to the regular ARC routines. 7000 */ 7001static void 7002l2arc_read_done(zio_t *zio) 7003{ 7004 l2arc_read_callback_t *cb; 7005 arc_buf_hdr_t *hdr; 7006 kmutex_t *hash_lock; 7007 boolean_t valid_cksum; 7008 7009 ASSERT3P(zio->io_vd, !=, NULL); 7010 ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE); 7011 7012 spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd); 7013 7014 cb = zio->io_private; 7015 ASSERT3P(cb, !=, NULL); 7016 hdr = cb->l2rcb_hdr; 7017 ASSERT3P(hdr, !=, NULL); 7018 7019 hash_lock = HDR_LOCK(hdr); 7020 mutex_enter(hash_lock); 7021 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 7022 7023 /* 7024 * If the data was read into a temporary buffer, 7025 * move it and free the buffer. 7026 */ 7027 if (cb->l2rcb_data != NULL) { 7028 ASSERT3U(arc_hdr_size(hdr), <, zio->io_size); 7029 if (zio->io_error == 0) { 7030 bcopy(cb->l2rcb_data, hdr->b_l1hdr.b_pdata, 7031 arc_hdr_size(hdr)); 7032 } 7033 7034 /* 7035 * The following must be done regardless of whether 7036 * there was an error: 7037 * - free the temporary buffer 7038 * - point zio to the real ARC buffer 7039 * - set zio size accordingly 7040 * These are required because zio is either re-used for 7041 * an I/O of the block in the case of the error 7042 * or the zio is passed to arc_read_done() and it 7043 * needs real data. 7044 */ 7045 zio_data_buf_free(cb->l2rcb_data, zio->io_size); 7046 zio->io_size = zio->io_orig_size = arc_hdr_size(hdr); 7047 zio->io_data = zio->io_orig_data = hdr->b_l1hdr.b_pdata; 7048 } 7049 7050 ASSERT3P(zio->io_data, !=, NULL); 7051 7052 /* 7053 * Check this survived the L2ARC journey. 7054 */ 7055 ASSERT3P(zio->io_data, ==, hdr->b_l1hdr.b_pdata); 7056 zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */ 7057 zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */ 7058 7059 valid_cksum = arc_cksum_is_equal(hdr, zio); 7060 if (valid_cksum && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) { 7061 mutex_exit(hash_lock); 7062 zio->io_private = hdr; 7063 arc_read_done(zio); 7064 } else { 7065 mutex_exit(hash_lock); 7066 /* 7067 * Buffer didn't survive caching. Increment stats and 7068 * reissue to the original storage device. 7069 */ 7070 if (zio->io_error != 0) { 7071 ARCSTAT_BUMP(arcstat_l2_io_error); 7072 } else { 7073 zio->io_error = SET_ERROR(EIO); 7074 } 7075 if (!valid_cksum) 7076 ARCSTAT_BUMP(arcstat_l2_cksum_bad); 7077 7078 /* 7079 * If there's no waiter, issue an async i/o to the primary 7080 * storage now. If there *is* a waiter, the caller must 7081 * issue the i/o in a context where it's OK to block. 7082 */ 7083 if (zio->io_waiter == NULL) { 7084 zio_t *pio = zio_unique_parent(zio); 7085 7086 ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL); 7087 7088 zio_nowait(zio_read(pio, zio->io_spa, zio->io_bp, 7089 hdr->b_l1hdr.b_pdata, zio->io_size, arc_read_done, 7090 hdr, zio->io_priority, cb->l2rcb_flags, 7091 &cb->l2rcb_zb)); 7092 } 7093 } 7094 7095 kmem_free(cb, sizeof (l2arc_read_callback_t)); 7096} 7097 7098/* 7099 * This is the list priority from which the L2ARC will search for pages to 7100 * cache. This is used within loops (0..3) to cycle through lists in the 7101 * desired order. This order can have a significant effect on cache 7102 * performance. 7103 * 7104 * Currently the metadata lists are hit first, MFU then MRU, followed by 7105 * the data lists. This function returns a locked list, and also returns 7106 * the lock pointer. 7107 */ 7108static multilist_sublist_t * 7109l2arc_sublist_lock(int list_num) 7110{ 7111 multilist_t *ml = NULL; 7112 unsigned int idx; 7113 7114 ASSERT(list_num >= 0 && list_num <= 3); 7115 7116 switch (list_num) { 7117 case 0: 7118 ml = arc_mfu->arcs_list[ARC_BUFC_METADATA]; 7119 break; 7120 case 1: 7121 ml = arc_mru->arcs_list[ARC_BUFC_METADATA]; 7122 break; 7123 case 2: 7124 ml = arc_mfu->arcs_list[ARC_BUFC_DATA]; 7125 break; 7126 case 3: 7127 ml = arc_mru->arcs_list[ARC_BUFC_DATA]; 7128 break; 7129 } 7130 7131 /* 7132 * Return a randomly-selected sublist. This is acceptable 7133 * because the caller feeds only a little bit of data for each 7134 * call (8MB). Subsequent calls will result in different 7135 * sublists being selected. 7136 */ 7137 idx = multilist_get_random_index(ml); 7138 return (multilist_sublist_lock(ml, idx)); 7139} 7140 7141/* 7142 * Evict buffers from the device write hand to the distance specified in 7143 * bytes. This distance may span populated buffers, it may span nothing. 7144 * This is clearing a region on the L2ARC device ready for writing. 7145 * If the 'all' boolean is set, every buffer is evicted. 7146 */ 7147static void 7148l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all) 7149{ 7150 list_t *buflist; 7151 arc_buf_hdr_t *hdr, *hdr_prev; 7152 kmutex_t *hash_lock; 7153 uint64_t taddr; 7154 7155 buflist = &dev->l2ad_buflist; 7156 7157 if (!all && dev->l2ad_first) { 7158 /* 7159 * This is the first sweep through the device. There is 7160 * nothing to evict. 7161 */ 7162 return; 7163 } 7164 7165 if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) { 7166 /* 7167 * When nearing the end of the device, evict to the end 7168 * before the device write hand jumps to the start. 7169 */ 7170 taddr = dev->l2ad_end; 7171 } else { 7172 taddr = dev->l2ad_hand + distance; 7173 } 7174 DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist, 7175 uint64_t, taddr, boolean_t, all); 7176 7177top: 7178 mutex_enter(&dev->l2ad_mtx); 7179 for (hdr = list_tail(buflist); hdr; hdr = hdr_prev) { 7180 hdr_prev = list_prev(buflist, hdr); 7181 7182 hash_lock = HDR_LOCK(hdr); 7183 7184 /* 7185 * We cannot use mutex_enter or else we can deadlock 7186 * with l2arc_write_buffers (due to swapping the order 7187 * the hash lock and l2ad_mtx are taken). 7188 */ 7189 if (!mutex_tryenter(hash_lock)) { 7190 /* 7191 * Missed the hash lock. Retry. 7192 */ 7193 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry); 7194 mutex_exit(&dev->l2ad_mtx); 7195 mutex_enter(hash_lock); 7196 mutex_exit(hash_lock); 7197 goto top; 7198 } 7199 7200 if (HDR_L2_WRITE_HEAD(hdr)) { 7201 /* 7202 * We hit a write head node. Leave it for 7203 * l2arc_write_done(). 7204 */ 7205 list_remove(buflist, hdr); 7206 mutex_exit(hash_lock); 7207 continue; 7208 } 7209 7210 if (!all && HDR_HAS_L2HDR(hdr) && 7211 (hdr->b_l2hdr.b_daddr >= taddr || 7212 hdr->b_l2hdr.b_daddr < dev->l2ad_hand)) { 7213 /* 7214 * We've evicted to the target address, 7215 * or the end of the device. 7216 */ 7217 mutex_exit(hash_lock); 7218 break; 7219 } 7220 7221 ASSERT(HDR_HAS_L2HDR(hdr)); 7222 if (!HDR_HAS_L1HDR(hdr)) { 7223 ASSERT(!HDR_L2_READING(hdr)); 7224 /* 7225 * This doesn't exist in the ARC. Destroy. 7226 * arc_hdr_destroy() will call list_remove() 7227 * and decrement arcstat_l2_size. 7228 */ 7229 arc_change_state(arc_anon, hdr, hash_lock); 7230 arc_hdr_destroy(hdr); 7231 } else { 7232 ASSERT(hdr->b_l1hdr.b_state != arc_l2c_only); 7233 ARCSTAT_BUMP(arcstat_l2_evict_l1cached); 7234 /* 7235 * Invalidate issued or about to be issued 7236 * reads, since we may be about to write 7237 * over this location. 7238 */ 7239 if (HDR_L2_READING(hdr)) { 7240 ARCSTAT_BUMP(arcstat_l2_evict_reading); 7241 arc_hdr_set_flags(hdr, ARC_FLAG_L2_EVICTED); 7242 } 7243 7244 /* Ensure this header has finished being written */ 7245 ASSERT(!HDR_L2_WRITING(hdr)); 7246 7247 arc_hdr_l2hdr_destroy(hdr); 7248 } 7249 mutex_exit(hash_lock); 7250 } 7251 mutex_exit(&dev->l2ad_mtx); 7252} 7253 7254/* 7255 * Find and write ARC buffers to the L2ARC device. 7256 * 7257 * An ARC_FLAG_L2_WRITING flag is set so that the L2ARC buffers are not valid 7258 * for reading until they have completed writing. 7259 * The headroom_boost is an in-out parameter used to maintain headroom boost 7260 * state between calls to this function. 7261 * 7262 * Returns the number of bytes actually written (which may be smaller than 7263 * the delta by which the device hand has changed due to alignment). 7264 */ 7265static uint64_t 7266l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz) 7267{ 7268 arc_buf_hdr_t *hdr, *hdr_prev, *head; 7269 uint64_t write_asize, write_psize, write_sz, headroom; 7270 boolean_t full; 7271 l2arc_write_callback_t *cb; 7272 zio_t *pio, *wzio; 7273 uint64_t guid = spa_load_guid(spa); 7274 int try; 7275 7276 ASSERT3P(dev->l2ad_vdev, !=, NULL); 7277 7278 pio = NULL; 7279 write_sz = write_asize = write_psize = 0; 7280 full = B_FALSE; 7281 head = kmem_cache_alloc(hdr_l2only_cache, KM_PUSHPAGE); 7282 arc_hdr_set_flags(head, ARC_FLAG_L2_WRITE_HEAD | ARC_FLAG_HAS_L2HDR); 7283 7284 ARCSTAT_BUMP(arcstat_l2_write_buffer_iter); 7285 /* 7286 * Copy buffers for L2ARC writing. 7287 */ 7288 for (try = 0; try <= 3; try++) { 7289 multilist_sublist_t *mls = l2arc_sublist_lock(try); 7290 uint64_t passed_sz = 0; 7291 7292 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_iter); 7293 7294 /* 7295 * L2ARC fast warmup. 7296 * 7297 * Until the ARC is warm and starts to evict, read from the 7298 * head of the ARC lists rather than the tail. 7299 */ 7300 if (arc_warm == B_FALSE) 7301 hdr = multilist_sublist_head(mls); 7302 else 7303 hdr = multilist_sublist_tail(mls); 7304 if (hdr == NULL) 7305 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_null_iter); 7306 7307 headroom = target_sz * l2arc_headroom; 7308 if (zfs_compressed_arc_enabled) 7309 headroom = (headroom * l2arc_headroom_boost) / 100; 7310 7311 for (; hdr; hdr = hdr_prev) { 7312 kmutex_t *hash_lock; 7313 7314 if (arc_warm == B_FALSE) 7315 hdr_prev = multilist_sublist_next(mls, hdr); 7316 else 7317 hdr_prev = multilist_sublist_prev(mls, hdr); 7318 ARCSTAT_INCR(arcstat_l2_write_buffer_bytes_scanned, 7319 HDR_GET_LSIZE(hdr)); 7320 7321 hash_lock = HDR_LOCK(hdr); 7322 if (!mutex_tryenter(hash_lock)) { 7323 ARCSTAT_BUMP(arcstat_l2_write_trylock_fail); 7324 /* 7325 * Skip this buffer rather than waiting. 7326 */ 7327 continue; 7328 } 7329 7330 passed_sz += HDR_GET_LSIZE(hdr); 7331 if (passed_sz > headroom) { 7332 /* 7333 * Searched too far. 7334 */ 7335 mutex_exit(hash_lock); 7336 ARCSTAT_BUMP(arcstat_l2_write_passed_headroom); 7337 break; 7338 } 7339 7340 if (!l2arc_write_eligible(guid, hdr)) { 7341 mutex_exit(hash_lock); 7342 continue; 7343 } 7344 7345 /* 7346 * We rely on the L1 portion of the header below, so 7347 * it's invalid for this header to have been evicted out 7348 * of the ghost cache, prior to being written out. The 7349 * ARC_FLAG_L2_WRITING bit ensures this won't happen. 7350 */ 7351 ASSERT(HDR_HAS_L1HDR(hdr)); 7352 7353 ASSERT3U(HDR_GET_PSIZE(hdr), >, 0); 7354 ASSERT3P(hdr->b_l1hdr.b_pdata, !=, NULL); 7355 ASSERT3U(arc_hdr_size(hdr), >, 0); 7356 uint64_t size = arc_hdr_size(hdr); 7357 uint64_t asize = vdev_psize_to_asize(dev->l2ad_vdev, 7358 size); 7359 7360 if ((write_psize + asize) > target_sz) { 7361 full = B_TRUE; 7362 mutex_exit(hash_lock); 7363 ARCSTAT_BUMP(arcstat_l2_write_full); 7364 break; 7365 } 7366 7367 if (pio == NULL) { 7368 /* 7369 * Insert a dummy header on the buflist so 7370 * l2arc_write_done() can find where the 7371 * write buffers begin without searching. 7372 */ 7373 mutex_enter(&dev->l2ad_mtx); 7374 list_insert_head(&dev->l2ad_buflist, head); 7375 mutex_exit(&dev->l2ad_mtx); 7376 7377 cb = kmem_alloc( 7378 sizeof (l2arc_write_callback_t), KM_SLEEP); 7379 cb->l2wcb_dev = dev; 7380 cb->l2wcb_head = head; 7381 pio = zio_root(spa, l2arc_write_done, cb, 7382 ZIO_FLAG_CANFAIL); 7383 ARCSTAT_BUMP(arcstat_l2_write_pios); 7384 } 7385 7386 hdr->b_l2hdr.b_dev = dev; 7387 hdr->b_l2hdr.b_daddr = dev->l2ad_hand; 7388 arc_hdr_set_flags(hdr, 7389 ARC_FLAG_L2_WRITING | ARC_FLAG_HAS_L2HDR); 7390 7391 mutex_enter(&dev->l2ad_mtx); 7392 list_insert_head(&dev->l2ad_buflist, hdr); 7393 mutex_exit(&dev->l2ad_mtx); 7394 7395 (void) refcount_add_many(&dev->l2ad_alloc, size, hdr); 7396 7397 /* 7398 * Normally the L2ARC can use the hdr's data, but if 7399 * we're sharing data between the hdr and one of its 7400 * bufs, L2ARC needs its own copy of the data so that 7401 * the ZIO below can't race with the buf consumer. To 7402 * ensure that this copy will be available for the 7403 * lifetime of the ZIO and be cleaned up afterwards, we 7404 * add it to the l2arc_free_on_write queue. 7405 */ 7406 void *to_write; 7407 if (!HDR_SHARED_DATA(hdr) && size == asize) { 7408 to_write = hdr->b_l1hdr.b_pdata; 7409 } else { 7410 arc_buf_contents_t type = arc_buf_type(hdr); 7411 if (type == ARC_BUFC_METADATA) { 7412 to_write = zio_buf_alloc(asize); 7413 } else { 7414 ASSERT3U(type, ==, ARC_BUFC_DATA); 7415 to_write = zio_data_buf_alloc(asize); 7416 } 7417 7418 bcopy(hdr->b_l1hdr.b_pdata, to_write, size); 7419 if (asize != size) 7420 bzero(to_write + size, asize - size); 7421 l2arc_free_data_on_write(to_write, asize, type); 7422 } 7423 wzio = zio_write_phys(pio, dev->l2ad_vdev, 7424 hdr->b_l2hdr.b_daddr, asize, to_write, 7425 ZIO_CHECKSUM_OFF, NULL, hdr, 7426 ZIO_PRIORITY_ASYNC_WRITE, 7427 ZIO_FLAG_CANFAIL, B_FALSE); 7428 7429 write_sz += HDR_GET_LSIZE(hdr); 7430 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev, 7431 zio_t *, wzio); 7432 7433 write_asize += size; 7434 write_psize += asize; 7435 dev->l2ad_hand += asize; 7436 7437 mutex_exit(hash_lock); 7438 7439 (void) zio_nowait(wzio); 7440 } 7441 7442 multilist_sublist_unlock(mls); 7443 7444 if (full == B_TRUE) 7445 break; 7446 } 7447 7448 /* No buffers selected for writing? */ 7449 if (pio == NULL) { 7450 ASSERT0(write_sz); 7451 ASSERT(!HDR_HAS_L1HDR(head)); 7452 kmem_cache_free(hdr_l2only_cache, head); 7453 return (0); 7454 } 7455 7456 ASSERT3U(write_psize, <=, target_sz); 7457 ARCSTAT_BUMP(arcstat_l2_writes_sent); 7458 ARCSTAT_INCR(arcstat_l2_write_bytes, write_asize); 7459 ARCSTAT_INCR(arcstat_l2_size, write_sz); 7460 ARCSTAT_INCR(arcstat_l2_asize, write_asize); 7461 vdev_space_update(dev->l2ad_vdev, write_asize, 0, 0); 7462 7463 /* 7464 * Bump device hand to the device start if it is approaching the end. 7465 * l2arc_evict() will already have evicted ahead for this case. 7466 */ 7467 if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) { 7468 dev->l2ad_hand = dev->l2ad_start; 7469 dev->l2ad_first = B_FALSE; 7470 } 7471 7472 dev->l2ad_writing = B_TRUE; 7473 (void) zio_wait(pio); 7474 dev->l2ad_writing = B_FALSE; 7475 7476 return (write_asize); 7477} 7478 7479/* 7480 * This thread feeds the L2ARC at regular intervals. This is the beating 7481 * heart of the L2ARC. 7482 */ 7483static void 7484l2arc_feed_thread(void *dummy __unused) 7485{ 7486 callb_cpr_t cpr; 7487 l2arc_dev_t *dev; 7488 spa_t *spa; 7489 uint64_t size, wrote; 7490 clock_t begin, next = ddi_get_lbolt(); 7491 7492 CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG); 7493 7494 mutex_enter(&l2arc_feed_thr_lock); 7495 7496 while (l2arc_thread_exit == 0) { 7497 CALLB_CPR_SAFE_BEGIN(&cpr); 7498 (void) cv_timedwait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock, 7499 next - ddi_get_lbolt()); 7500 CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock); 7501 next = ddi_get_lbolt() + hz; 7502 7503 /* 7504 * Quick check for L2ARC devices. 7505 */ 7506 mutex_enter(&l2arc_dev_mtx); 7507 if (l2arc_ndev == 0) { 7508 mutex_exit(&l2arc_dev_mtx); 7509 continue; 7510 } 7511 mutex_exit(&l2arc_dev_mtx); 7512 begin = ddi_get_lbolt(); 7513 7514 /* 7515 * This selects the next l2arc device to write to, and in 7516 * doing so the next spa to feed from: dev->l2ad_spa. This 7517 * will return NULL if there are now no l2arc devices or if 7518 * they are all faulted. 7519 * 7520 * If a device is returned, its spa's config lock is also 7521 * held to prevent device removal. l2arc_dev_get_next() 7522 * will grab and release l2arc_dev_mtx. 7523 */ 7524 if ((dev = l2arc_dev_get_next()) == NULL) 7525 continue; 7526 7527 spa = dev->l2ad_spa; 7528 ASSERT3P(spa, !=, NULL); 7529 7530 /* 7531 * If the pool is read-only then force the feed thread to 7532 * sleep a little longer. 7533 */ 7534 if (!spa_writeable(spa)) { 7535 next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz; 7536 spa_config_exit(spa, SCL_L2ARC, dev); 7537 continue; 7538 } 7539 7540 /* 7541 * Avoid contributing to memory pressure. 7542 */ 7543 if (arc_reclaim_needed()) { 7544 ARCSTAT_BUMP(arcstat_l2_abort_lowmem); 7545 spa_config_exit(spa, SCL_L2ARC, dev); 7546 continue; 7547 } 7548 7549 ARCSTAT_BUMP(arcstat_l2_feeds); 7550 7551 size = l2arc_write_size(); 7552 7553 /* 7554 * Evict L2ARC buffers that will be overwritten. 7555 */ 7556 l2arc_evict(dev, size, B_FALSE); 7557 7558 /* 7559 * Write ARC buffers. 7560 */ 7561 wrote = l2arc_write_buffers(spa, dev, size); 7562 7563 /* 7564 * Calculate interval between writes. 7565 */ 7566 next = l2arc_write_interval(begin, size, wrote); 7567 spa_config_exit(spa, SCL_L2ARC, dev); 7568 } 7569 7570 l2arc_thread_exit = 0; 7571 cv_broadcast(&l2arc_feed_thr_cv); 7572 CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */ 7573 thread_exit(); 7574} 7575 7576boolean_t 7577l2arc_vdev_present(vdev_t *vd) 7578{ 7579 l2arc_dev_t *dev; 7580 7581 mutex_enter(&l2arc_dev_mtx); 7582 for (dev = list_head(l2arc_dev_list); dev != NULL; 7583 dev = list_next(l2arc_dev_list, dev)) { 7584 if (dev->l2ad_vdev == vd) 7585 break; 7586 } 7587 mutex_exit(&l2arc_dev_mtx); 7588 7589 return (dev != NULL); 7590} 7591 7592/* 7593 * Add a vdev for use by the L2ARC. By this point the spa has already 7594 * validated the vdev and opened it. 7595 */ 7596void 7597l2arc_add_vdev(spa_t *spa, vdev_t *vd) 7598{ 7599 l2arc_dev_t *adddev; 7600 7601 ASSERT(!l2arc_vdev_present(vd)); 7602 7603 vdev_ashift_optimize(vd); 7604 7605 /* 7606 * Create a new l2arc device entry. 7607 */ 7608 adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP); 7609 adddev->l2ad_spa = spa; 7610 adddev->l2ad_vdev = vd; 7611 adddev->l2ad_start = VDEV_LABEL_START_SIZE; 7612 adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd); 7613 adddev->l2ad_hand = adddev->l2ad_start; 7614 adddev->l2ad_first = B_TRUE; 7615 adddev->l2ad_writing = B_FALSE; 7616 7617 mutex_init(&adddev->l2ad_mtx, NULL, MUTEX_DEFAULT, NULL); 7618 /* 7619 * This is a list of all ARC buffers that are still valid on the 7620 * device. 7621 */ 7622 list_create(&adddev->l2ad_buflist, sizeof (arc_buf_hdr_t), 7623 offsetof(arc_buf_hdr_t, b_l2hdr.b_l2node)); 7624 7625 vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand); 7626 refcount_create(&adddev->l2ad_alloc); 7627 7628 /* 7629 * Add device to global list 7630 */ 7631 mutex_enter(&l2arc_dev_mtx); 7632 list_insert_head(l2arc_dev_list, adddev); 7633 atomic_inc_64(&l2arc_ndev); 7634 mutex_exit(&l2arc_dev_mtx); 7635} 7636 7637/* 7638 * Remove a vdev from the L2ARC. 7639 */ 7640void 7641l2arc_remove_vdev(vdev_t *vd) 7642{ 7643 l2arc_dev_t *dev, *nextdev, *remdev = NULL; 7644 7645 /* 7646 * Find the device by vdev 7647 */ 7648 mutex_enter(&l2arc_dev_mtx); 7649 for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) { 7650 nextdev = list_next(l2arc_dev_list, dev); 7651 if (vd == dev->l2ad_vdev) { 7652 remdev = dev; 7653 break; 7654 } 7655 } 7656 ASSERT3P(remdev, !=, NULL); 7657 7658 /* 7659 * Remove device from global list 7660 */ 7661 list_remove(l2arc_dev_list, remdev); 7662 l2arc_dev_last = NULL; /* may have been invalidated */ 7663 atomic_dec_64(&l2arc_ndev); 7664 mutex_exit(&l2arc_dev_mtx); 7665 7666 /* 7667 * Clear all buflists and ARC references. L2ARC device flush. 7668 */ 7669 l2arc_evict(remdev, 0, B_TRUE); 7670 list_destroy(&remdev->l2ad_buflist); 7671 mutex_destroy(&remdev->l2ad_mtx); 7672 refcount_destroy(&remdev->l2ad_alloc); 7673 kmem_free(remdev, sizeof (l2arc_dev_t)); 7674} 7675 7676void 7677l2arc_init(void) 7678{ 7679 l2arc_thread_exit = 0; 7680 l2arc_ndev = 0; 7681 l2arc_writes_sent = 0; 7682 l2arc_writes_done = 0; 7683 7684 mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL); 7685 cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL); 7686 mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL); 7687 mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL); 7688 7689 l2arc_dev_list = &L2ARC_dev_list; 7690 l2arc_free_on_write = &L2ARC_free_on_write; 7691 list_create(l2arc_dev_list, sizeof (l2arc_dev_t), 7692 offsetof(l2arc_dev_t, l2ad_node)); 7693 list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t), 7694 offsetof(l2arc_data_free_t, l2df_list_node)); 7695} 7696 7697void 7698l2arc_fini(void) 7699{ 7700 /* 7701 * This is called from dmu_fini(), which is called from spa_fini(); 7702 * Because of this, we can assume that all l2arc devices have 7703 * already been removed when the pools themselves were removed. 7704 */ 7705 7706 l2arc_do_free_on_write(); 7707 7708 mutex_destroy(&l2arc_feed_thr_lock); 7709 cv_destroy(&l2arc_feed_thr_cv); 7710 mutex_destroy(&l2arc_dev_mtx); 7711 mutex_destroy(&l2arc_free_on_write_mtx); 7712 7713 list_destroy(l2arc_dev_list); 7714 list_destroy(l2arc_free_on_write); 7715} 7716 7717void 7718l2arc_start(void) 7719{ 7720 if (!(spa_mode_global & FWRITE)) 7721 return; 7722 7723 (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0, 7724 TS_RUN, minclsyspri); 7725} 7726 7727void 7728l2arc_stop(void) 7729{ 7730 if (!(spa_mode_global & FWRITE)) 7731 return; 7732 7733 mutex_enter(&l2arc_feed_thr_lock); 7734 cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */ 7735 l2arc_thread_exit = 1; 7736 while (l2arc_thread_exit != 0) 7737 cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock); 7738 mutex_exit(&l2arc_feed_thr_lock); 7739} 7740