arc.c revision 338456
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) 2018, Joyent, Inc. 24 * Copyright (c) 2011, 2017 by Delphix. All rights reserved. 25 * Copyright (c) 2014 by Saso Kiselkov. All rights reserved. 26 * Copyright 2017 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_pabd). 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_pabd) associated with the DVA of the 135 * arc_buf_hdr_t. Since the b_pabd 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_pabd 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_pabd +-+ |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_pabd buffer into a 195 * new data buffer, or shares the hdr's b_pabd 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_pabd +-+ |---------| |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_pabd 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_pabd. 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_pabd. The 245 * L2ARC will always write the contents of b_pabd 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/zio_checksum.h> 267#include <sys/multilist.h> 268#include <sys/abd.h> 269#ifdef _KERNEL 270#include <sys/dnlc.h> 271#include <sys/racct.h> 272#endif 273#include <sys/callb.h> 274#include <sys/kstat.h> 275#include <sys/trim_map.h> 276#include <zfs_fletcher.h> 277#include <sys/sdt.h> 278#include <sys/aggsum.h> 279#include <sys/cityhash.h> 280 281#include <machine/vmparam.h> 282 283#ifdef illumos 284#ifndef _KERNEL 285/* set with ZFS_DEBUG=watch, to enable watchpoints on frozen buffers */ 286boolean_t arc_watch = B_FALSE; 287int arc_procfd; 288#endif 289#endif /* illumos */ 290 291static kmutex_t arc_reclaim_lock; 292static kcondvar_t arc_reclaim_thread_cv; 293static boolean_t arc_reclaim_thread_exit; 294static kcondvar_t arc_reclaim_waiters_cv; 295 296static kmutex_t arc_dnlc_evicts_lock; 297static kcondvar_t arc_dnlc_evicts_cv; 298static boolean_t arc_dnlc_evicts_thread_exit; 299 300uint_t arc_reduce_dnlc_percent = 3; 301 302/* 303 * The number of headers to evict in arc_evict_state_impl() before 304 * dropping the sublist lock and evicting from another sublist. A lower 305 * value means we're more likely to evict the "correct" header (i.e. the 306 * oldest header in the arc state), but comes with higher overhead 307 * (i.e. more invocations of arc_evict_state_impl()). 308 */ 309int zfs_arc_evict_batch_limit = 10; 310 311/* number of seconds before growing cache again */ 312static int arc_grow_retry = 60; 313 314/* number of milliseconds before attempting a kmem-cache-reap */ 315static int arc_kmem_cache_reap_retry_ms = 0; 316 317/* shift of arc_c for calculating overflow limit in arc_get_data_impl */ 318int zfs_arc_overflow_shift = 8; 319 320/* shift of arc_c for calculating both min and max arc_p */ 321static int arc_p_min_shift = 4; 322 323/* log2(fraction of arc to reclaim) */ 324static int arc_shrink_shift = 7; 325 326/* 327 * log2(fraction of ARC which must be free to allow growing). 328 * I.e. If there is less than arc_c >> arc_no_grow_shift free memory, 329 * when reading a new block into the ARC, we will evict an equal-sized block 330 * from the ARC. 331 * 332 * This must be less than arc_shrink_shift, so that when we shrink the ARC, 333 * we will still not allow it to grow. 334 */ 335int arc_no_grow_shift = 5; 336 337 338/* 339 * minimum lifespan of a prefetch block in clock ticks 340 * (initialized in arc_init()) 341 */ 342static int arc_min_prefetch_lifespan; 343 344/* 345 * If this percent of memory is free, don't throttle. 346 */ 347int arc_lotsfree_percent = 10; 348 349static int arc_dead; 350extern boolean_t zfs_prefetch_disable; 351 352/* 353 * The arc has filled available memory and has now warmed up. 354 */ 355static boolean_t arc_warm; 356 357/* 358 * log2 fraction of the zio arena to keep free. 359 */ 360int arc_zio_arena_free_shift = 2; 361 362/* 363 * These tunables are for performance analysis. 364 */ 365uint64_t zfs_arc_max; 366uint64_t zfs_arc_min; 367uint64_t zfs_arc_meta_limit = 0; 368uint64_t zfs_arc_meta_min = 0; 369int zfs_arc_grow_retry = 0; 370int zfs_arc_shrink_shift = 0; 371int zfs_arc_no_grow_shift = 0; 372int zfs_arc_p_min_shift = 0; 373uint64_t zfs_arc_average_blocksize = 8 * 1024; /* 8KB */ 374u_int zfs_arc_free_target = 0; 375 376/* Absolute min for arc min / max is 16MB. */ 377static uint64_t arc_abs_min = 16 << 20; 378 379boolean_t zfs_compressed_arc_enabled = B_TRUE; 380 381static int sysctl_vfs_zfs_arc_free_target(SYSCTL_HANDLER_ARGS); 382static int sysctl_vfs_zfs_arc_meta_limit(SYSCTL_HANDLER_ARGS); 383static int sysctl_vfs_zfs_arc_max(SYSCTL_HANDLER_ARGS); 384static int sysctl_vfs_zfs_arc_min(SYSCTL_HANDLER_ARGS); 385static int sysctl_vfs_zfs_arc_no_grow_shift(SYSCTL_HANDLER_ARGS); 386 387#if defined(__FreeBSD__) && defined(_KERNEL) 388static void 389arc_free_target_init(void *unused __unused) 390{ 391 392 zfs_arc_free_target = vm_pageout_wakeup_thresh; 393} 394SYSINIT(arc_free_target_init, SI_SUB_KTHREAD_PAGE, SI_ORDER_ANY, 395 arc_free_target_init, NULL); 396 397TUNABLE_QUAD("vfs.zfs.arc_meta_limit", &zfs_arc_meta_limit); 398TUNABLE_QUAD("vfs.zfs.arc_meta_min", &zfs_arc_meta_min); 399TUNABLE_INT("vfs.zfs.arc_shrink_shift", &zfs_arc_shrink_shift); 400TUNABLE_INT("vfs.zfs.arc_grow_retry", &zfs_arc_grow_retry); 401TUNABLE_INT("vfs.zfs.arc_no_grow_shift", &zfs_arc_no_grow_shift); 402SYSCTL_DECL(_vfs_zfs); 403SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_max, CTLTYPE_U64 | CTLFLAG_RWTUN, 404 0, sizeof(uint64_t), sysctl_vfs_zfs_arc_max, "QU", "Maximum ARC size"); 405SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_min, CTLTYPE_U64 | CTLFLAG_RWTUN, 406 0, sizeof(uint64_t), sysctl_vfs_zfs_arc_min, "QU", "Minimum ARC size"); 407SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_no_grow_shift, CTLTYPE_U32 | CTLFLAG_RWTUN, 408 0, sizeof(uint32_t), sysctl_vfs_zfs_arc_no_grow_shift, "U", 409 "log2(fraction of ARC which must be free to allow growing)"); 410SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, arc_average_blocksize, CTLFLAG_RDTUN, 411 &zfs_arc_average_blocksize, 0, 412 "ARC average blocksize"); 413SYSCTL_INT(_vfs_zfs, OID_AUTO, arc_shrink_shift, CTLFLAG_RW, 414 &arc_shrink_shift, 0, 415 "log2(fraction of arc to reclaim)"); 416SYSCTL_INT(_vfs_zfs, OID_AUTO, arc_grow_retry, CTLFLAG_RW, 417 &arc_grow_retry, 0, 418 "Wait in seconds before considering growing ARC"); 419SYSCTL_INT(_vfs_zfs, OID_AUTO, compressed_arc_enabled, CTLFLAG_RDTUN, 420 &zfs_compressed_arc_enabled, 0, 421 "Enable compressed ARC"); 422SYSCTL_INT(_vfs_zfs, OID_AUTO, arc_kmem_cache_reap_retry_ms, CTLFLAG_RWTUN, 423 &arc_kmem_cache_reap_retry_ms, 0, 424 "Interval between ARC kmem_cache reapings"); 425 426/* 427 * We don't have a tunable for arc_free_target due to the dependency on 428 * pagedaemon initialisation. 429 */ 430SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_free_target, 431 CTLTYPE_UINT | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, sizeof(u_int), 432 sysctl_vfs_zfs_arc_free_target, "IU", 433 "Desired number of free pages below which ARC triggers reclaim"); 434 435static int 436sysctl_vfs_zfs_arc_free_target(SYSCTL_HANDLER_ARGS) 437{ 438 u_int val; 439 int err; 440 441 val = zfs_arc_free_target; 442 err = sysctl_handle_int(oidp, &val, 0, req); 443 if (err != 0 || req->newptr == NULL) 444 return (err); 445 446 if (val < minfree) 447 return (EINVAL); 448 if (val > vm_cnt.v_page_count) 449 return (EINVAL); 450 451 zfs_arc_free_target = val; 452 453 return (0); 454} 455 456/* 457 * Must be declared here, before the definition of corresponding kstat 458 * macro which uses the same names will confuse the compiler. 459 */ 460SYSCTL_PROC(_vfs_zfs, OID_AUTO, arc_meta_limit, 461 CTLTYPE_U64 | CTLFLAG_MPSAFE | CTLFLAG_RW, 0, sizeof(uint64_t), 462 sysctl_vfs_zfs_arc_meta_limit, "QU", 463 "ARC metadata limit"); 464#endif 465 466/* 467 * Note that buffers can be in one of 6 states: 468 * ARC_anon - anonymous (discussed below) 469 * ARC_mru - recently used, currently cached 470 * ARC_mru_ghost - recentely used, no longer in cache 471 * ARC_mfu - frequently used, currently cached 472 * ARC_mfu_ghost - frequently used, no longer in cache 473 * ARC_l2c_only - exists in L2ARC but not other states 474 * When there are no active references to the buffer, they are 475 * are linked onto a list in one of these arc states. These are 476 * the only buffers that can be evicted or deleted. Within each 477 * state there are multiple lists, one for meta-data and one for 478 * non-meta-data. Meta-data (indirect blocks, blocks of dnodes, 479 * etc.) is tracked separately so that it can be managed more 480 * explicitly: favored over data, limited explicitly. 481 * 482 * Anonymous buffers are buffers that are not associated with 483 * a DVA. These are buffers that hold dirty block copies 484 * before they are written to stable storage. By definition, 485 * they are "ref'd" and are considered part of arc_mru 486 * that cannot be freed. Generally, they will aquire a DVA 487 * as they are written and migrate onto the arc_mru list. 488 * 489 * The ARC_l2c_only state is for buffers that are in the second 490 * level ARC but no longer in any of the ARC_m* lists. The second 491 * level ARC itself may also contain buffers that are in any of 492 * the ARC_m* states - meaning that a buffer can exist in two 493 * places. The reason for the ARC_l2c_only state is to keep the 494 * buffer header in the hash table, so that reads that hit the 495 * second level ARC benefit from these fast lookups. 496 */ 497 498typedef struct arc_state { 499 /* 500 * list of evictable buffers 501 */ 502 multilist_t *arcs_list[ARC_BUFC_NUMTYPES]; 503 /* 504 * total amount of evictable data in this state 505 */ 506 refcount_t arcs_esize[ARC_BUFC_NUMTYPES]; 507 /* 508 * total amount of data in this state; this includes: evictable, 509 * non-evictable, ARC_BUFC_DATA, and ARC_BUFC_METADATA. 510 */ 511 refcount_t arcs_size; 512} arc_state_t; 513 514/* The 6 states: */ 515static arc_state_t ARC_anon; 516static arc_state_t ARC_mru; 517static arc_state_t ARC_mru_ghost; 518static arc_state_t ARC_mfu; 519static arc_state_t ARC_mfu_ghost; 520static arc_state_t ARC_l2c_only; 521 522typedef struct arc_stats { 523 kstat_named_t arcstat_hits; 524 kstat_named_t arcstat_misses; 525 kstat_named_t arcstat_demand_data_hits; 526 kstat_named_t arcstat_demand_data_misses; 527 kstat_named_t arcstat_demand_metadata_hits; 528 kstat_named_t arcstat_demand_metadata_misses; 529 kstat_named_t arcstat_prefetch_data_hits; 530 kstat_named_t arcstat_prefetch_data_misses; 531 kstat_named_t arcstat_prefetch_metadata_hits; 532 kstat_named_t arcstat_prefetch_metadata_misses; 533 kstat_named_t arcstat_mru_hits; 534 kstat_named_t arcstat_mru_ghost_hits; 535 kstat_named_t arcstat_mfu_hits; 536 kstat_named_t arcstat_mfu_ghost_hits; 537 kstat_named_t arcstat_allocated; 538 kstat_named_t arcstat_deleted; 539 /* 540 * Number of buffers that could not be evicted because the hash lock 541 * was held by another thread. The lock may not necessarily be held 542 * by something using the same buffer, since hash locks are shared 543 * by multiple buffers. 544 */ 545 kstat_named_t arcstat_mutex_miss; 546 /* 547 * Number of buffers skipped when updating the access state due to the 548 * header having already been released after acquiring the hash lock. 549 */ 550 kstat_named_t arcstat_access_skip; 551 /* 552 * Number of buffers skipped because they have I/O in progress, are 553 * indirect prefetch buffers that have not lived long enough, or are 554 * not from the spa we're trying to evict from. 555 */ 556 kstat_named_t arcstat_evict_skip; 557 /* 558 * Number of times arc_evict_state() was unable to evict enough 559 * buffers to reach it's target amount. 560 */ 561 kstat_named_t arcstat_evict_not_enough; 562 kstat_named_t arcstat_evict_l2_cached; 563 kstat_named_t arcstat_evict_l2_eligible; 564 kstat_named_t arcstat_evict_l2_ineligible; 565 kstat_named_t arcstat_evict_l2_skip; 566 kstat_named_t arcstat_hash_elements; 567 kstat_named_t arcstat_hash_elements_max; 568 kstat_named_t arcstat_hash_collisions; 569 kstat_named_t arcstat_hash_chains; 570 kstat_named_t arcstat_hash_chain_max; 571 kstat_named_t arcstat_p; 572 kstat_named_t arcstat_c; 573 kstat_named_t arcstat_c_min; 574 kstat_named_t arcstat_c_max; 575 /* Not updated directly; only synced in arc_kstat_update. */ 576 kstat_named_t arcstat_size; 577 /* 578 * Number of compressed bytes stored in the arc_buf_hdr_t's b_pabd. 579 * Note that the compressed bytes may match the uncompressed bytes 580 * if the block is either not compressed or compressed arc is disabled. 581 */ 582 kstat_named_t arcstat_compressed_size; 583 /* 584 * Uncompressed size of the data stored in b_pabd. If compressed 585 * arc is disabled then this value will be identical to the stat 586 * above. 587 */ 588 kstat_named_t arcstat_uncompressed_size; 589 /* 590 * Number of bytes stored in all the arc_buf_t's. This is classified 591 * as "overhead" since this data is typically short-lived and will 592 * be evicted from the arc when it becomes unreferenced unless the 593 * zfs_keep_uncompressed_metadata or zfs_keep_uncompressed_level 594 * values have been set (see comment in dbuf.c for more information). 595 */ 596 kstat_named_t arcstat_overhead_size; 597 /* 598 * Number of bytes consumed by internal ARC structures necessary 599 * for tracking purposes; these structures are not actually 600 * backed by ARC buffers. This includes arc_buf_hdr_t structures 601 * (allocated via arc_buf_hdr_t_full and arc_buf_hdr_t_l2only 602 * caches), and arc_buf_t structures (allocated via arc_buf_t 603 * cache). 604 * Not updated directly; only synced in arc_kstat_update. 605 */ 606 kstat_named_t arcstat_hdr_size; 607 /* 608 * Number of bytes consumed by ARC buffers of type equal to 609 * ARC_BUFC_DATA. This is generally consumed by buffers backing 610 * on disk user data (e.g. plain file contents). 611 * Not updated directly; only synced in arc_kstat_update. 612 */ 613 kstat_named_t arcstat_data_size; 614 /* 615 * Number of bytes consumed by ARC buffers of type equal to 616 * ARC_BUFC_METADATA. This is generally consumed by buffers 617 * backing on disk data that is used for internal ZFS 618 * structures (e.g. ZAP, dnode, indirect blocks, etc). 619 * Not updated directly; only synced in arc_kstat_update. 620 */ 621 kstat_named_t arcstat_metadata_size; 622 /* 623 * Number of bytes consumed by various buffers and structures 624 * not actually backed with ARC buffers. This includes bonus 625 * buffers (allocated directly via zio_buf_* functions), 626 * dmu_buf_impl_t structures (allocated via dmu_buf_impl_t 627 * cache), and dnode_t structures (allocated via dnode_t cache). 628 * Not updated directly; only synced in arc_kstat_update. 629 */ 630 kstat_named_t arcstat_other_size; 631 /* 632 * Total number of bytes consumed by ARC buffers residing in the 633 * arc_anon state. This includes *all* buffers in the arc_anon 634 * state; e.g. data, metadata, evictable, and unevictable buffers 635 * are all included in this value. 636 * Not updated directly; only synced in arc_kstat_update. 637 */ 638 kstat_named_t arcstat_anon_size; 639 /* 640 * Number of bytes consumed by ARC buffers that meet the 641 * following criteria: backing buffers of type ARC_BUFC_DATA, 642 * residing in the arc_anon state, and are eligible for eviction 643 * (e.g. have no outstanding holds on the buffer). 644 * Not updated directly; only synced in arc_kstat_update. 645 */ 646 kstat_named_t arcstat_anon_evictable_data; 647 /* 648 * Number of bytes consumed by ARC buffers that meet the 649 * following criteria: backing buffers of type ARC_BUFC_METADATA, 650 * residing in the arc_anon state, and are eligible for eviction 651 * (e.g. have no outstanding holds on the buffer). 652 * Not updated directly; only synced in arc_kstat_update. 653 */ 654 kstat_named_t arcstat_anon_evictable_metadata; 655 /* 656 * Total number of bytes consumed by ARC buffers residing in the 657 * arc_mru state. This includes *all* buffers in the arc_mru 658 * state; e.g. data, metadata, evictable, and unevictable buffers 659 * are all included in this value. 660 * Not updated directly; only synced in arc_kstat_update. 661 */ 662 kstat_named_t arcstat_mru_size; 663 /* 664 * Number of bytes consumed by ARC buffers that meet the 665 * following criteria: backing buffers of type ARC_BUFC_DATA, 666 * residing in the arc_mru state, and are eligible for eviction 667 * (e.g. have no outstanding holds on the buffer). 668 * Not updated directly; only synced in arc_kstat_update. 669 */ 670 kstat_named_t arcstat_mru_evictable_data; 671 /* 672 * Number of bytes consumed by ARC buffers that meet the 673 * following criteria: backing buffers of type ARC_BUFC_METADATA, 674 * residing in the arc_mru state, and are eligible for eviction 675 * (e.g. have no outstanding holds on the buffer). 676 * Not updated directly; only synced in arc_kstat_update. 677 */ 678 kstat_named_t arcstat_mru_evictable_metadata; 679 /* 680 * Total number of bytes that *would have been* consumed by ARC 681 * buffers in the arc_mru_ghost state. The key thing to note 682 * here, is the fact that this size doesn't actually indicate 683 * RAM consumption. The ghost lists only consist of headers and 684 * don't actually have ARC buffers linked off of these headers. 685 * Thus, *if* the headers had associated ARC buffers, these 686 * buffers *would have* consumed this number of bytes. 687 * Not updated directly; only synced in arc_kstat_update. 688 */ 689 kstat_named_t arcstat_mru_ghost_size; 690 /* 691 * Number of bytes that *would have been* consumed by ARC 692 * buffers that are eligible for eviction, of type 693 * ARC_BUFC_DATA, and linked off the arc_mru_ghost state. 694 * Not updated directly; only synced in arc_kstat_update. 695 */ 696 kstat_named_t arcstat_mru_ghost_evictable_data; 697 /* 698 * Number of bytes that *would have been* consumed by ARC 699 * buffers that are eligible for eviction, of type 700 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state. 701 * Not updated directly; only synced in arc_kstat_update. 702 */ 703 kstat_named_t arcstat_mru_ghost_evictable_metadata; 704 /* 705 * Total number of bytes consumed by ARC buffers residing in the 706 * arc_mfu state. This includes *all* buffers in the arc_mfu 707 * state; e.g. data, metadata, evictable, and unevictable buffers 708 * are all included in this value. 709 * Not updated directly; only synced in arc_kstat_update. 710 */ 711 kstat_named_t arcstat_mfu_size; 712 /* 713 * Number of bytes consumed by ARC buffers that are eligible for 714 * eviction, of type ARC_BUFC_DATA, and reside in the arc_mfu 715 * state. 716 * Not updated directly; only synced in arc_kstat_update. 717 */ 718 kstat_named_t arcstat_mfu_evictable_data; 719 /* 720 * Number of bytes consumed by ARC buffers that are eligible for 721 * eviction, of type ARC_BUFC_METADATA, and reside in the 722 * arc_mfu state. 723 * Not updated directly; only synced in arc_kstat_update. 724 */ 725 kstat_named_t arcstat_mfu_evictable_metadata; 726 /* 727 * Total number of bytes that *would have been* consumed by ARC 728 * buffers in the arc_mfu_ghost state. See the comment above 729 * arcstat_mru_ghost_size for more details. 730 * Not updated directly; only synced in arc_kstat_update. 731 */ 732 kstat_named_t arcstat_mfu_ghost_size; 733 /* 734 * Number of bytes that *would have been* consumed by ARC 735 * buffers that are eligible for eviction, of type 736 * ARC_BUFC_DATA, and linked off the arc_mfu_ghost state. 737 * Not updated directly; only synced in arc_kstat_update. 738 */ 739 kstat_named_t arcstat_mfu_ghost_evictable_data; 740 /* 741 * Number of bytes that *would have been* consumed by ARC 742 * buffers that are eligible for eviction, of type 743 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state. 744 * Not updated directly; only synced in arc_kstat_update. 745 */ 746 kstat_named_t arcstat_mfu_ghost_evictable_metadata; 747 kstat_named_t arcstat_l2_hits; 748 kstat_named_t arcstat_l2_misses; 749 kstat_named_t arcstat_l2_feeds; 750 kstat_named_t arcstat_l2_rw_clash; 751 kstat_named_t arcstat_l2_read_bytes; 752 kstat_named_t arcstat_l2_write_bytes; 753 kstat_named_t arcstat_l2_writes_sent; 754 kstat_named_t arcstat_l2_writes_done; 755 kstat_named_t arcstat_l2_writes_error; 756 kstat_named_t arcstat_l2_writes_lock_retry; 757 kstat_named_t arcstat_l2_evict_lock_retry; 758 kstat_named_t arcstat_l2_evict_reading; 759 kstat_named_t arcstat_l2_evict_l1cached; 760 kstat_named_t arcstat_l2_free_on_write; 761 kstat_named_t arcstat_l2_abort_lowmem; 762 kstat_named_t arcstat_l2_cksum_bad; 763 kstat_named_t arcstat_l2_io_error; 764 kstat_named_t arcstat_l2_lsize; 765 kstat_named_t arcstat_l2_psize; 766 /* Not updated directly; only synced in arc_kstat_update. */ 767 kstat_named_t arcstat_l2_hdr_size; 768 kstat_named_t arcstat_l2_write_trylock_fail; 769 kstat_named_t arcstat_l2_write_passed_headroom; 770 kstat_named_t arcstat_l2_write_spa_mismatch; 771 kstat_named_t arcstat_l2_write_in_l2; 772 kstat_named_t arcstat_l2_write_hdr_io_in_progress; 773 kstat_named_t arcstat_l2_write_not_cacheable; 774 kstat_named_t arcstat_l2_write_full; 775 kstat_named_t arcstat_l2_write_buffer_iter; 776 kstat_named_t arcstat_l2_write_pios; 777 kstat_named_t arcstat_l2_write_buffer_bytes_scanned; 778 kstat_named_t arcstat_l2_write_buffer_list_iter; 779 kstat_named_t arcstat_l2_write_buffer_list_null_iter; 780 kstat_named_t arcstat_memory_throttle_count; 781 /* Not updated directly; only synced in arc_kstat_update. */ 782 kstat_named_t arcstat_meta_used; 783 kstat_named_t arcstat_meta_limit; 784 kstat_named_t arcstat_meta_max; 785 kstat_named_t arcstat_meta_min; 786 kstat_named_t arcstat_sync_wait_for_async; 787 kstat_named_t arcstat_demand_hit_predictive_prefetch; 788} arc_stats_t; 789 790static arc_stats_t arc_stats = { 791 { "hits", KSTAT_DATA_UINT64 }, 792 { "misses", KSTAT_DATA_UINT64 }, 793 { "demand_data_hits", KSTAT_DATA_UINT64 }, 794 { "demand_data_misses", KSTAT_DATA_UINT64 }, 795 { "demand_metadata_hits", KSTAT_DATA_UINT64 }, 796 { "demand_metadata_misses", KSTAT_DATA_UINT64 }, 797 { "prefetch_data_hits", KSTAT_DATA_UINT64 }, 798 { "prefetch_data_misses", KSTAT_DATA_UINT64 }, 799 { "prefetch_metadata_hits", KSTAT_DATA_UINT64 }, 800 { "prefetch_metadata_misses", KSTAT_DATA_UINT64 }, 801 { "mru_hits", KSTAT_DATA_UINT64 }, 802 { "mru_ghost_hits", KSTAT_DATA_UINT64 }, 803 { "mfu_hits", KSTAT_DATA_UINT64 }, 804 { "mfu_ghost_hits", KSTAT_DATA_UINT64 }, 805 { "allocated", KSTAT_DATA_UINT64 }, 806 { "deleted", KSTAT_DATA_UINT64 }, 807 { "mutex_miss", KSTAT_DATA_UINT64 }, 808 { "access_skip", KSTAT_DATA_UINT64 }, 809 { "evict_skip", KSTAT_DATA_UINT64 }, 810 { "evict_not_enough", KSTAT_DATA_UINT64 }, 811 { "evict_l2_cached", KSTAT_DATA_UINT64 }, 812 { "evict_l2_eligible", KSTAT_DATA_UINT64 }, 813 { "evict_l2_ineligible", KSTAT_DATA_UINT64 }, 814 { "evict_l2_skip", KSTAT_DATA_UINT64 }, 815 { "hash_elements", KSTAT_DATA_UINT64 }, 816 { "hash_elements_max", KSTAT_DATA_UINT64 }, 817 { "hash_collisions", KSTAT_DATA_UINT64 }, 818 { "hash_chains", KSTAT_DATA_UINT64 }, 819 { "hash_chain_max", KSTAT_DATA_UINT64 }, 820 { "p", KSTAT_DATA_UINT64 }, 821 { "c", KSTAT_DATA_UINT64 }, 822 { "c_min", KSTAT_DATA_UINT64 }, 823 { "c_max", KSTAT_DATA_UINT64 }, 824 { "size", KSTAT_DATA_UINT64 }, 825 { "compressed_size", KSTAT_DATA_UINT64 }, 826 { "uncompressed_size", KSTAT_DATA_UINT64 }, 827 { "overhead_size", KSTAT_DATA_UINT64 }, 828 { "hdr_size", KSTAT_DATA_UINT64 }, 829 { "data_size", KSTAT_DATA_UINT64 }, 830 { "metadata_size", KSTAT_DATA_UINT64 }, 831 { "other_size", KSTAT_DATA_UINT64 }, 832 { "anon_size", KSTAT_DATA_UINT64 }, 833 { "anon_evictable_data", KSTAT_DATA_UINT64 }, 834 { "anon_evictable_metadata", KSTAT_DATA_UINT64 }, 835 { "mru_size", KSTAT_DATA_UINT64 }, 836 { "mru_evictable_data", KSTAT_DATA_UINT64 }, 837 { "mru_evictable_metadata", KSTAT_DATA_UINT64 }, 838 { "mru_ghost_size", KSTAT_DATA_UINT64 }, 839 { "mru_ghost_evictable_data", KSTAT_DATA_UINT64 }, 840 { "mru_ghost_evictable_metadata", KSTAT_DATA_UINT64 }, 841 { "mfu_size", KSTAT_DATA_UINT64 }, 842 { "mfu_evictable_data", KSTAT_DATA_UINT64 }, 843 { "mfu_evictable_metadata", KSTAT_DATA_UINT64 }, 844 { "mfu_ghost_size", KSTAT_DATA_UINT64 }, 845 { "mfu_ghost_evictable_data", KSTAT_DATA_UINT64 }, 846 { "mfu_ghost_evictable_metadata", KSTAT_DATA_UINT64 }, 847 { "l2_hits", KSTAT_DATA_UINT64 }, 848 { "l2_misses", KSTAT_DATA_UINT64 }, 849 { "l2_feeds", KSTAT_DATA_UINT64 }, 850 { "l2_rw_clash", KSTAT_DATA_UINT64 }, 851 { "l2_read_bytes", KSTAT_DATA_UINT64 }, 852 { "l2_write_bytes", KSTAT_DATA_UINT64 }, 853 { "l2_writes_sent", KSTAT_DATA_UINT64 }, 854 { "l2_writes_done", KSTAT_DATA_UINT64 }, 855 { "l2_writes_error", KSTAT_DATA_UINT64 }, 856 { "l2_writes_lock_retry", KSTAT_DATA_UINT64 }, 857 { "l2_evict_lock_retry", KSTAT_DATA_UINT64 }, 858 { "l2_evict_reading", KSTAT_DATA_UINT64 }, 859 { "l2_evict_l1cached", KSTAT_DATA_UINT64 }, 860 { "l2_free_on_write", KSTAT_DATA_UINT64 }, 861 { "l2_abort_lowmem", KSTAT_DATA_UINT64 }, 862 { "l2_cksum_bad", KSTAT_DATA_UINT64 }, 863 { "l2_io_error", KSTAT_DATA_UINT64 }, 864 { "l2_size", KSTAT_DATA_UINT64 }, 865 { "l2_asize", KSTAT_DATA_UINT64 }, 866 { "l2_hdr_size", KSTAT_DATA_UINT64 }, 867 { "l2_write_trylock_fail", KSTAT_DATA_UINT64 }, 868 { "l2_write_passed_headroom", KSTAT_DATA_UINT64 }, 869 { "l2_write_spa_mismatch", KSTAT_DATA_UINT64 }, 870 { "l2_write_in_l2", KSTAT_DATA_UINT64 }, 871 { "l2_write_io_in_progress", KSTAT_DATA_UINT64 }, 872 { "l2_write_not_cacheable", KSTAT_DATA_UINT64 }, 873 { "l2_write_full", KSTAT_DATA_UINT64 }, 874 { "l2_write_buffer_iter", KSTAT_DATA_UINT64 }, 875 { "l2_write_pios", KSTAT_DATA_UINT64 }, 876 { "l2_write_buffer_bytes_scanned", KSTAT_DATA_UINT64 }, 877 { "l2_write_buffer_list_iter", KSTAT_DATA_UINT64 }, 878 { "l2_write_buffer_list_null_iter", KSTAT_DATA_UINT64 }, 879 { "memory_throttle_count", KSTAT_DATA_UINT64 }, 880 { "arc_meta_used", KSTAT_DATA_UINT64 }, 881 { "arc_meta_limit", KSTAT_DATA_UINT64 }, 882 { "arc_meta_max", KSTAT_DATA_UINT64 }, 883 { "arc_meta_min", KSTAT_DATA_UINT64 }, 884 { "sync_wait_for_async", KSTAT_DATA_UINT64 }, 885 { "demand_hit_predictive_prefetch", KSTAT_DATA_UINT64 }, 886}; 887 888#define ARCSTAT(stat) (arc_stats.stat.value.ui64) 889 890#define ARCSTAT_INCR(stat, val) \ 891 atomic_add_64(&arc_stats.stat.value.ui64, (val)) 892 893#define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1) 894#define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1) 895 896#define ARCSTAT_MAX(stat, val) { \ 897 uint64_t m; \ 898 while ((val) > (m = arc_stats.stat.value.ui64) && \ 899 (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \ 900 continue; \ 901} 902 903#define ARCSTAT_MAXSTAT(stat) \ 904 ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64) 905 906/* 907 * We define a macro to allow ARC hits/misses to be easily broken down by 908 * two separate conditions, giving a total of four different subtypes for 909 * each of hits and misses (so eight statistics total). 910 */ 911#define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \ 912 if (cond1) { \ 913 if (cond2) { \ 914 ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \ 915 } else { \ 916 ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \ 917 } \ 918 } else { \ 919 if (cond2) { \ 920 ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \ 921 } else { \ 922 ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\ 923 } \ 924 } 925 926kstat_t *arc_ksp; 927static arc_state_t *arc_anon; 928static arc_state_t *arc_mru; 929static arc_state_t *arc_mru_ghost; 930static arc_state_t *arc_mfu; 931static arc_state_t *arc_mfu_ghost; 932static arc_state_t *arc_l2c_only; 933 934/* 935 * There are several ARC variables that are critical to export as kstats -- 936 * but we don't want to have to grovel around in the kstat whenever we wish to 937 * manipulate them. For these variables, we therefore define them to be in 938 * terms of the statistic variable. This assures that we are not introducing 939 * the possibility of inconsistency by having shadow copies of the variables, 940 * while still allowing the code to be readable. 941 */ 942#define arc_p ARCSTAT(arcstat_p) /* target size of MRU */ 943#define arc_c ARCSTAT(arcstat_c) /* target size of cache */ 944#define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */ 945#define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */ 946#define arc_meta_limit ARCSTAT(arcstat_meta_limit) /* max size for metadata */ 947#define arc_meta_min ARCSTAT(arcstat_meta_min) /* min size for metadata */ 948#define arc_meta_max ARCSTAT(arcstat_meta_max) /* max size of metadata */ 949 950/* compressed size of entire arc */ 951#define arc_compressed_size ARCSTAT(arcstat_compressed_size) 952/* uncompressed size of entire arc */ 953#define arc_uncompressed_size ARCSTAT(arcstat_uncompressed_size) 954/* number of bytes in the arc from arc_buf_t's */ 955#define arc_overhead_size ARCSTAT(arcstat_overhead_size) 956 957/* 958 * There are also some ARC variables that we want to export, but that are 959 * updated so often that having the canonical representation be the statistic 960 * variable causes a performance bottleneck. We want to use aggsum_t's for these 961 * instead, but still be able to export the kstat in the same way as before. 962 * The solution is to always use the aggsum version, except in the kstat update 963 * callback. 964 */ 965aggsum_t arc_size; 966aggsum_t arc_meta_used; 967aggsum_t astat_data_size; 968aggsum_t astat_metadata_size; 969aggsum_t astat_hdr_size; 970aggsum_t astat_other_size; 971aggsum_t astat_l2_hdr_size; 972 973static int arc_no_grow; /* Don't try to grow cache size */ 974static uint64_t arc_tempreserve; 975static uint64_t arc_loaned_bytes; 976 977typedef struct arc_callback arc_callback_t; 978 979struct arc_callback { 980 void *acb_private; 981 arc_done_func_t *acb_done; 982 arc_buf_t *acb_buf; 983 boolean_t acb_compressed; 984 zio_t *acb_zio_dummy; 985 arc_callback_t *acb_next; 986}; 987 988typedef struct arc_write_callback arc_write_callback_t; 989 990struct arc_write_callback { 991 void *awcb_private; 992 arc_done_func_t *awcb_ready; 993 arc_done_func_t *awcb_children_ready; 994 arc_done_func_t *awcb_physdone; 995 arc_done_func_t *awcb_done; 996 arc_buf_t *awcb_buf; 997}; 998 999/* 1000 * ARC buffers are separated into multiple structs as a memory saving measure: 1001 * - Common fields struct, always defined, and embedded within it: 1002 * - L2-only fields, always allocated but undefined when not in L2ARC 1003 * - L1-only fields, only allocated when in L1ARC 1004 * 1005 * Buffer in L1 Buffer only in L2 1006 * +------------------------+ +------------------------+ 1007 * | arc_buf_hdr_t | | arc_buf_hdr_t | 1008 * | | | | 1009 * | | | | 1010 * | | | | 1011 * +------------------------+ +------------------------+ 1012 * | l2arc_buf_hdr_t | | l2arc_buf_hdr_t | 1013 * | (undefined if L1-only) | | | 1014 * +------------------------+ +------------------------+ 1015 * | l1arc_buf_hdr_t | 1016 * | | 1017 * | | 1018 * | | 1019 * | | 1020 * +------------------------+ 1021 * 1022 * Because it's possible for the L2ARC to become extremely large, we can wind 1023 * up eating a lot of memory in L2ARC buffer headers, so the size of a header 1024 * is minimized by only allocating the fields necessary for an L1-cached buffer 1025 * when a header is actually in the L1 cache. The sub-headers (l1arc_buf_hdr and 1026 * l2arc_buf_hdr) are embedded rather than allocated separately to save a couple 1027 * words in pointers. arc_hdr_realloc() is used to switch a header between 1028 * these two allocation states. 1029 */ 1030typedef struct l1arc_buf_hdr { 1031 kmutex_t b_freeze_lock; 1032 zio_cksum_t *b_freeze_cksum; 1033#ifdef ZFS_DEBUG 1034 /* 1035 * Used for debugging with kmem_flags - by allocating and freeing 1036 * b_thawed when the buffer is thawed, we get a record of the stack 1037 * trace that thawed it. 1038 */ 1039 void *b_thawed; 1040#endif 1041 1042 arc_buf_t *b_buf; 1043 uint32_t b_bufcnt; 1044 /* for waiting on writes to complete */ 1045 kcondvar_t b_cv; 1046 uint8_t b_byteswap; 1047 1048 /* protected by arc state mutex */ 1049 arc_state_t *b_state; 1050 multilist_node_t b_arc_node; 1051 1052 /* updated atomically */ 1053 clock_t b_arc_access; 1054 1055 /* self protecting */ 1056 refcount_t b_refcnt; 1057 1058 arc_callback_t *b_acb; 1059 abd_t *b_pabd; 1060} l1arc_buf_hdr_t; 1061 1062typedef struct l2arc_dev l2arc_dev_t; 1063 1064typedef struct l2arc_buf_hdr { 1065 /* protected by arc_buf_hdr mutex */ 1066 l2arc_dev_t *b_dev; /* L2ARC device */ 1067 uint64_t b_daddr; /* disk address, offset byte */ 1068 1069 list_node_t b_l2node; 1070} l2arc_buf_hdr_t; 1071 1072struct arc_buf_hdr { 1073 /* protected by hash lock */ 1074 dva_t b_dva; 1075 uint64_t b_birth; 1076 1077 arc_buf_contents_t b_type; 1078 arc_buf_hdr_t *b_hash_next; 1079 arc_flags_t b_flags; 1080 1081 /* 1082 * This field stores the size of the data buffer after 1083 * compression, and is set in the arc's zio completion handlers. 1084 * It is in units of SPA_MINBLOCKSIZE (e.g. 1 == 512 bytes). 1085 * 1086 * While the block pointers can store up to 32MB in their psize 1087 * field, we can only store up to 32MB minus 512B. This is due 1088 * to the bp using a bias of 1, whereas we use a bias of 0 (i.e. 1089 * a field of zeros represents 512B in the bp). We can't use a 1090 * bias of 1 since we need to reserve a psize of zero, here, to 1091 * represent holes and embedded blocks. 1092 * 1093 * This isn't a problem in practice, since the maximum size of a 1094 * buffer is limited to 16MB, so we never need to store 32MB in 1095 * this field. Even in the upstream illumos code base, the 1096 * maximum size of a buffer is limited to 16MB. 1097 */ 1098 uint16_t b_psize; 1099 1100 /* 1101 * This field stores the size of the data buffer before 1102 * compression, and cannot change once set. It is in units 1103 * of SPA_MINBLOCKSIZE (e.g. 2 == 1024 bytes) 1104 */ 1105 uint16_t b_lsize; /* immutable */ 1106 uint64_t b_spa; /* immutable */ 1107 1108 /* L2ARC fields. Undefined when not in L2ARC. */ 1109 l2arc_buf_hdr_t b_l2hdr; 1110 /* L1ARC fields. Undefined when in l2arc_only state */ 1111 l1arc_buf_hdr_t b_l1hdr; 1112}; 1113 1114#if defined(__FreeBSD__) && defined(_KERNEL) 1115static int 1116sysctl_vfs_zfs_arc_meta_limit(SYSCTL_HANDLER_ARGS) 1117{ 1118 uint64_t val; 1119 int err; 1120 1121 val = arc_meta_limit; 1122 err = sysctl_handle_64(oidp, &val, 0, req); 1123 if (err != 0 || req->newptr == NULL) 1124 return (err); 1125 1126 if (val <= 0 || val > arc_c_max) 1127 return (EINVAL); 1128 1129 arc_meta_limit = val; 1130 return (0); 1131} 1132 1133static int 1134sysctl_vfs_zfs_arc_no_grow_shift(SYSCTL_HANDLER_ARGS) 1135{ 1136 uint32_t val; 1137 int err; 1138 1139 val = arc_no_grow_shift; 1140 err = sysctl_handle_32(oidp, &val, 0, req); 1141 if (err != 0 || req->newptr == NULL) 1142 return (err); 1143 1144 if (val >= arc_shrink_shift) 1145 return (EINVAL); 1146 1147 arc_no_grow_shift = val; 1148 return (0); 1149} 1150 1151static int 1152sysctl_vfs_zfs_arc_max(SYSCTL_HANDLER_ARGS) 1153{ 1154 uint64_t val; 1155 int err; 1156 1157 val = zfs_arc_max; 1158 err = sysctl_handle_64(oidp, &val, 0, req); 1159 if (err != 0 || req->newptr == NULL) 1160 return (err); 1161 1162 if (zfs_arc_max == 0) { 1163 /* Loader tunable so blindly set */ 1164 zfs_arc_max = val; 1165 return (0); 1166 } 1167 1168 if (val < arc_abs_min || val > kmem_size()) 1169 return (EINVAL); 1170 if (val < arc_c_min) 1171 return (EINVAL); 1172 if (zfs_arc_meta_limit > 0 && val < zfs_arc_meta_limit) 1173 return (EINVAL); 1174 1175 arc_c_max = val; 1176 1177 arc_c = arc_c_max; 1178 arc_p = (arc_c >> 1); 1179 1180 if (zfs_arc_meta_limit == 0) { 1181 /* limit meta-data to 1/4 of the arc capacity */ 1182 arc_meta_limit = arc_c_max / 4; 1183 } 1184 1185 /* if kmem_flags are set, lets try to use less memory */ 1186 if (kmem_debugging()) 1187 arc_c = arc_c / 2; 1188 1189 zfs_arc_max = arc_c; 1190 1191 return (0); 1192} 1193 1194static int 1195sysctl_vfs_zfs_arc_min(SYSCTL_HANDLER_ARGS) 1196{ 1197 uint64_t val; 1198 int err; 1199 1200 val = zfs_arc_min; 1201 err = sysctl_handle_64(oidp, &val, 0, req); 1202 if (err != 0 || req->newptr == NULL) 1203 return (err); 1204 1205 if (zfs_arc_min == 0) { 1206 /* Loader tunable so blindly set */ 1207 zfs_arc_min = val; 1208 return (0); 1209 } 1210 1211 if (val < arc_abs_min || val > arc_c_max) 1212 return (EINVAL); 1213 1214 arc_c_min = val; 1215 1216 if (zfs_arc_meta_min == 0) 1217 arc_meta_min = arc_c_min / 2; 1218 1219 if (arc_c < arc_c_min) 1220 arc_c = arc_c_min; 1221 1222 zfs_arc_min = arc_c_min; 1223 1224 return (0); 1225} 1226#endif 1227 1228#define GHOST_STATE(state) \ 1229 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \ 1230 (state) == arc_l2c_only) 1231 1232#define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_FLAG_IN_HASH_TABLE) 1233#define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) 1234#define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_FLAG_IO_ERROR) 1235#define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_FLAG_PREFETCH) 1236#define HDR_COMPRESSION_ENABLED(hdr) \ 1237 ((hdr)->b_flags & ARC_FLAG_COMPRESSED_ARC) 1238 1239#define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_FLAG_L2CACHE) 1240#define HDR_L2_READING(hdr) \ 1241 (((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) && \ 1242 ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR)) 1243#define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITING) 1244#define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_FLAG_L2_EVICTED) 1245#define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITE_HEAD) 1246#define HDR_SHARED_DATA(hdr) ((hdr)->b_flags & ARC_FLAG_SHARED_DATA) 1247 1248#define HDR_ISTYPE_METADATA(hdr) \ 1249 ((hdr)->b_flags & ARC_FLAG_BUFC_METADATA) 1250#define HDR_ISTYPE_DATA(hdr) (!HDR_ISTYPE_METADATA(hdr)) 1251 1252#define HDR_HAS_L1HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L1HDR) 1253#define HDR_HAS_L2HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR) 1254 1255/* For storing compression mode in b_flags */ 1256#define HDR_COMPRESS_OFFSET (highbit64(ARC_FLAG_COMPRESS_0) - 1) 1257 1258#define HDR_GET_COMPRESS(hdr) ((enum zio_compress)BF32_GET((hdr)->b_flags, \ 1259 HDR_COMPRESS_OFFSET, SPA_COMPRESSBITS)) 1260#define HDR_SET_COMPRESS(hdr, cmp) BF32_SET((hdr)->b_flags, \ 1261 HDR_COMPRESS_OFFSET, SPA_COMPRESSBITS, (cmp)); 1262 1263#define ARC_BUF_LAST(buf) ((buf)->b_next == NULL) 1264#define ARC_BUF_SHARED(buf) ((buf)->b_flags & ARC_BUF_FLAG_SHARED) 1265#define ARC_BUF_COMPRESSED(buf) ((buf)->b_flags & ARC_BUF_FLAG_COMPRESSED) 1266 1267/* 1268 * Other sizes 1269 */ 1270 1271#define HDR_FULL_SIZE ((int64_t)sizeof (arc_buf_hdr_t)) 1272#define HDR_L2ONLY_SIZE ((int64_t)offsetof(arc_buf_hdr_t, b_l1hdr)) 1273 1274/* 1275 * Hash table routines 1276 */ 1277 1278#define HT_LOCK_PAD CACHE_LINE_SIZE 1279 1280struct ht_lock { 1281 kmutex_t ht_lock; 1282#ifdef _KERNEL 1283 unsigned char pad[(HT_LOCK_PAD - sizeof (kmutex_t))]; 1284#endif 1285}; 1286 1287#define BUF_LOCKS 256 1288typedef struct buf_hash_table { 1289 uint64_t ht_mask; 1290 arc_buf_hdr_t **ht_table; 1291 struct ht_lock ht_locks[BUF_LOCKS] __aligned(CACHE_LINE_SIZE); 1292} buf_hash_table_t; 1293 1294static buf_hash_table_t buf_hash_table; 1295 1296#define BUF_HASH_INDEX(spa, dva, birth) \ 1297 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask) 1298#define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)]) 1299#define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock)) 1300#define HDR_LOCK(hdr) \ 1301 (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth))) 1302 1303uint64_t zfs_crc64_table[256]; 1304 1305/* 1306 * Level 2 ARC 1307 */ 1308 1309#define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */ 1310#define L2ARC_HEADROOM 2 /* num of writes */ 1311/* 1312 * If we discover during ARC scan any buffers to be compressed, we boost 1313 * our headroom for the next scanning cycle by this percentage multiple. 1314 */ 1315#define L2ARC_HEADROOM_BOOST 200 1316#define L2ARC_FEED_SECS 1 /* caching interval secs */ 1317#define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */ 1318 1319#define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent) 1320#define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done) 1321 1322/* L2ARC Performance Tunables */ 1323uint64_t l2arc_write_max = L2ARC_WRITE_SIZE; /* default max write size */ 1324uint64_t l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra write during warmup */ 1325uint64_t l2arc_headroom = L2ARC_HEADROOM; /* number of dev writes */ 1326uint64_t l2arc_headroom_boost = L2ARC_HEADROOM_BOOST; 1327uint64_t l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */ 1328uint64_t l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval milliseconds */ 1329boolean_t l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */ 1330boolean_t l2arc_feed_again = B_TRUE; /* turbo warmup */ 1331boolean_t l2arc_norw = B_TRUE; /* no reads during writes */ 1332 1333SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_max, CTLFLAG_RW, 1334 &l2arc_write_max, 0, "max write size"); 1335SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_write_boost, CTLFLAG_RW, 1336 &l2arc_write_boost, 0, "extra write during warmup"); 1337SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_headroom, CTLFLAG_RW, 1338 &l2arc_headroom, 0, "number of dev writes"); 1339SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_secs, CTLFLAG_RW, 1340 &l2arc_feed_secs, 0, "interval seconds"); 1341SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2arc_feed_min_ms, CTLFLAG_RW, 1342 &l2arc_feed_min_ms, 0, "min interval milliseconds"); 1343 1344SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_noprefetch, CTLFLAG_RW, 1345 &l2arc_noprefetch, 0, "don't cache prefetch bufs"); 1346SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_feed_again, CTLFLAG_RW, 1347 &l2arc_feed_again, 0, "turbo warmup"); 1348SYSCTL_INT(_vfs_zfs, OID_AUTO, l2arc_norw, CTLFLAG_RW, 1349 &l2arc_norw, 0, "no reads during writes"); 1350 1351SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_size, CTLFLAG_RD, 1352 &ARC_anon.arcs_size.rc_count, 0, "size of anonymous state"); 1353SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_metadata_esize, CTLFLAG_RD, 1354 &ARC_anon.arcs_esize[ARC_BUFC_METADATA].rc_count, 0, 1355 "size of anonymous state"); 1356SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, anon_data_esize, CTLFLAG_RD, 1357 &ARC_anon.arcs_esize[ARC_BUFC_DATA].rc_count, 0, 1358 "size of anonymous state"); 1359 1360SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_size, CTLFLAG_RD, 1361 &ARC_mru.arcs_size.rc_count, 0, "size of mru state"); 1362SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_metadata_esize, CTLFLAG_RD, 1363 &ARC_mru.arcs_esize[ARC_BUFC_METADATA].rc_count, 0, 1364 "size of metadata in mru state"); 1365SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_data_esize, CTLFLAG_RD, 1366 &ARC_mru.arcs_esize[ARC_BUFC_DATA].rc_count, 0, 1367 "size of data in mru state"); 1368 1369SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_size, CTLFLAG_RD, 1370 &ARC_mru_ghost.arcs_size.rc_count, 0, "size of mru ghost state"); 1371SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_metadata_esize, CTLFLAG_RD, 1372 &ARC_mru_ghost.arcs_esize[ARC_BUFC_METADATA].rc_count, 0, 1373 "size of metadata in mru ghost state"); 1374SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mru_ghost_data_esize, CTLFLAG_RD, 1375 &ARC_mru_ghost.arcs_esize[ARC_BUFC_DATA].rc_count, 0, 1376 "size of data in mru ghost state"); 1377 1378SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_size, CTLFLAG_RD, 1379 &ARC_mfu.arcs_size.rc_count, 0, "size of mfu state"); 1380SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_metadata_esize, CTLFLAG_RD, 1381 &ARC_mfu.arcs_esize[ARC_BUFC_METADATA].rc_count, 0, 1382 "size of metadata in mfu state"); 1383SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_data_esize, CTLFLAG_RD, 1384 &ARC_mfu.arcs_esize[ARC_BUFC_DATA].rc_count, 0, 1385 "size of data in mfu state"); 1386 1387SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_size, CTLFLAG_RD, 1388 &ARC_mfu_ghost.arcs_size.rc_count, 0, "size of mfu ghost state"); 1389SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_metadata_esize, CTLFLAG_RD, 1390 &ARC_mfu_ghost.arcs_esize[ARC_BUFC_METADATA].rc_count, 0, 1391 "size of metadata in mfu ghost state"); 1392SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, mfu_ghost_data_esize, CTLFLAG_RD, 1393 &ARC_mfu_ghost.arcs_esize[ARC_BUFC_DATA].rc_count, 0, 1394 "size of data in mfu ghost state"); 1395 1396SYSCTL_UQUAD(_vfs_zfs, OID_AUTO, l2c_only_size, CTLFLAG_RD, 1397 &ARC_l2c_only.arcs_size.rc_count, 0, "size of mru state"); 1398 1399/* 1400 * L2ARC Internals 1401 */ 1402struct l2arc_dev { 1403 vdev_t *l2ad_vdev; /* vdev */ 1404 spa_t *l2ad_spa; /* spa */ 1405 uint64_t l2ad_hand; /* next write location */ 1406 uint64_t l2ad_start; /* first addr on device */ 1407 uint64_t l2ad_end; /* last addr on device */ 1408 boolean_t l2ad_first; /* first sweep through */ 1409 boolean_t l2ad_writing; /* currently writing */ 1410 kmutex_t l2ad_mtx; /* lock for buffer list */ 1411 list_t l2ad_buflist; /* buffer list */ 1412 list_node_t l2ad_node; /* device list node */ 1413 refcount_t l2ad_alloc; /* allocated bytes */ 1414}; 1415 1416static list_t L2ARC_dev_list; /* device list */ 1417static list_t *l2arc_dev_list; /* device list pointer */ 1418static kmutex_t l2arc_dev_mtx; /* device list mutex */ 1419static l2arc_dev_t *l2arc_dev_last; /* last device used */ 1420static list_t L2ARC_free_on_write; /* free after write buf list */ 1421static list_t *l2arc_free_on_write; /* free after write list ptr */ 1422static kmutex_t l2arc_free_on_write_mtx; /* mutex for list */ 1423static uint64_t l2arc_ndev; /* number of devices */ 1424 1425typedef struct l2arc_read_callback { 1426 arc_buf_hdr_t *l2rcb_hdr; /* read header */ 1427 blkptr_t l2rcb_bp; /* original blkptr */ 1428 zbookmark_phys_t l2rcb_zb; /* original bookmark */ 1429 int l2rcb_flags; /* original flags */ 1430 abd_t *l2rcb_abd; /* temporary buffer */ 1431} l2arc_read_callback_t; 1432 1433typedef struct l2arc_write_callback { 1434 l2arc_dev_t *l2wcb_dev; /* device info */ 1435 arc_buf_hdr_t *l2wcb_head; /* head of write buflist */ 1436} l2arc_write_callback_t; 1437 1438typedef struct l2arc_data_free { 1439 /* protected by l2arc_free_on_write_mtx */ 1440 abd_t *l2df_abd; 1441 size_t l2df_size; 1442 arc_buf_contents_t l2df_type; 1443 list_node_t l2df_list_node; 1444} l2arc_data_free_t; 1445 1446static kmutex_t l2arc_feed_thr_lock; 1447static kcondvar_t l2arc_feed_thr_cv; 1448static uint8_t l2arc_thread_exit; 1449 1450static abd_t *arc_get_data_abd(arc_buf_hdr_t *, uint64_t, void *); 1451static void *arc_get_data_buf(arc_buf_hdr_t *, uint64_t, void *); 1452static void arc_get_data_impl(arc_buf_hdr_t *, uint64_t, void *); 1453static void arc_free_data_abd(arc_buf_hdr_t *, abd_t *, uint64_t, void *); 1454static void arc_free_data_buf(arc_buf_hdr_t *, void *, uint64_t, void *); 1455static void arc_free_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag); 1456static void arc_hdr_free_pabd(arc_buf_hdr_t *); 1457static void arc_hdr_alloc_pabd(arc_buf_hdr_t *); 1458static void arc_access(arc_buf_hdr_t *, kmutex_t *); 1459static boolean_t arc_is_overflowing(); 1460static void arc_buf_watch(arc_buf_t *); 1461 1462static arc_buf_contents_t arc_buf_type(arc_buf_hdr_t *); 1463static uint32_t arc_bufc_to_flags(arc_buf_contents_t); 1464static inline void arc_hdr_set_flags(arc_buf_hdr_t *hdr, arc_flags_t flags); 1465static inline void arc_hdr_clear_flags(arc_buf_hdr_t *hdr, arc_flags_t flags); 1466 1467static boolean_t l2arc_write_eligible(uint64_t, arc_buf_hdr_t *); 1468static void l2arc_read_done(zio_t *); 1469 1470static void 1471l2arc_trim(const arc_buf_hdr_t *hdr) 1472{ 1473 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev; 1474 1475 ASSERT(HDR_HAS_L2HDR(hdr)); 1476 ASSERT(MUTEX_HELD(&dev->l2ad_mtx)); 1477 1478 if (HDR_GET_PSIZE(hdr) != 0) { 1479 trim_map_free(dev->l2ad_vdev, hdr->b_l2hdr.b_daddr, 1480 HDR_GET_PSIZE(hdr), 0); 1481 } 1482} 1483 1484/* 1485 * We use Cityhash for this. It's fast, and has good hash properties without 1486 * requiring any large static buffers. 1487 */ 1488static uint64_t 1489buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth) 1490{ 1491 return (cityhash4(spa, dva->dva_word[0], dva->dva_word[1], birth)); 1492} 1493 1494#define HDR_EMPTY(hdr) \ 1495 ((hdr)->b_dva.dva_word[0] == 0 && \ 1496 (hdr)->b_dva.dva_word[1] == 0) 1497 1498#define HDR_EQUAL(spa, dva, birth, hdr) \ 1499 ((hdr)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \ 1500 ((hdr)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \ 1501 ((hdr)->b_birth == birth) && ((hdr)->b_spa == spa) 1502 1503static void 1504buf_discard_identity(arc_buf_hdr_t *hdr) 1505{ 1506 hdr->b_dva.dva_word[0] = 0; 1507 hdr->b_dva.dva_word[1] = 0; 1508 hdr->b_birth = 0; 1509} 1510 1511static arc_buf_hdr_t * 1512buf_hash_find(uint64_t spa, const blkptr_t *bp, kmutex_t **lockp) 1513{ 1514 const dva_t *dva = BP_IDENTITY(bp); 1515 uint64_t birth = BP_PHYSICAL_BIRTH(bp); 1516 uint64_t idx = BUF_HASH_INDEX(spa, dva, birth); 1517 kmutex_t *hash_lock = BUF_HASH_LOCK(idx); 1518 arc_buf_hdr_t *hdr; 1519 1520 mutex_enter(hash_lock); 1521 for (hdr = buf_hash_table.ht_table[idx]; hdr != NULL; 1522 hdr = hdr->b_hash_next) { 1523 if (HDR_EQUAL(spa, dva, birth, hdr)) { 1524 *lockp = hash_lock; 1525 return (hdr); 1526 } 1527 } 1528 mutex_exit(hash_lock); 1529 *lockp = NULL; 1530 return (NULL); 1531} 1532 1533/* 1534 * Insert an entry into the hash table. If there is already an element 1535 * equal to elem in the hash table, then the already existing element 1536 * will be returned and the new element will not be inserted. 1537 * Otherwise returns NULL. 1538 * If lockp == NULL, the caller is assumed to already hold the hash lock. 1539 */ 1540static arc_buf_hdr_t * 1541buf_hash_insert(arc_buf_hdr_t *hdr, kmutex_t **lockp) 1542{ 1543 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth); 1544 kmutex_t *hash_lock = BUF_HASH_LOCK(idx); 1545 arc_buf_hdr_t *fhdr; 1546 uint32_t i; 1547 1548 ASSERT(!DVA_IS_EMPTY(&hdr->b_dva)); 1549 ASSERT(hdr->b_birth != 0); 1550 ASSERT(!HDR_IN_HASH_TABLE(hdr)); 1551 1552 if (lockp != NULL) { 1553 *lockp = hash_lock; 1554 mutex_enter(hash_lock); 1555 } else { 1556 ASSERT(MUTEX_HELD(hash_lock)); 1557 } 1558 1559 for (fhdr = buf_hash_table.ht_table[idx], i = 0; fhdr != NULL; 1560 fhdr = fhdr->b_hash_next, i++) { 1561 if (HDR_EQUAL(hdr->b_spa, &hdr->b_dva, hdr->b_birth, fhdr)) 1562 return (fhdr); 1563 } 1564 1565 hdr->b_hash_next = buf_hash_table.ht_table[idx]; 1566 buf_hash_table.ht_table[idx] = hdr; 1567 arc_hdr_set_flags(hdr, ARC_FLAG_IN_HASH_TABLE); 1568 1569 /* collect some hash table performance data */ 1570 if (i > 0) { 1571 ARCSTAT_BUMP(arcstat_hash_collisions); 1572 if (i == 1) 1573 ARCSTAT_BUMP(arcstat_hash_chains); 1574 1575 ARCSTAT_MAX(arcstat_hash_chain_max, i); 1576 } 1577 1578 ARCSTAT_BUMP(arcstat_hash_elements); 1579 ARCSTAT_MAXSTAT(arcstat_hash_elements); 1580 1581 return (NULL); 1582} 1583 1584static void 1585buf_hash_remove(arc_buf_hdr_t *hdr) 1586{ 1587 arc_buf_hdr_t *fhdr, **hdrp; 1588 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth); 1589 1590 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx))); 1591 ASSERT(HDR_IN_HASH_TABLE(hdr)); 1592 1593 hdrp = &buf_hash_table.ht_table[idx]; 1594 while ((fhdr = *hdrp) != hdr) { 1595 ASSERT3P(fhdr, !=, NULL); 1596 hdrp = &fhdr->b_hash_next; 1597 } 1598 *hdrp = hdr->b_hash_next; 1599 hdr->b_hash_next = NULL; 1600 arc_hdr_clear_flags(hdr, ARC_FLAG_IN_HASH_TABLE); 1601 1602 /* collect some hash table performance data */ 1603 ARCSTAT_BUMPDOWN(arcstat_hash_elements); 1604 1605 if (buf_hash_table.ht_table[idx] && 1606 buf_hash_table.ht_table[idx]->b_hash_next == NULL) 1607 ARCSTAT_BUMPDOWN(arcstat_hash_chains); 1608} 1609 1610/* 1611 * Global data structures and functions for the buf kmem cache. 1612 */ 1613static kmem_cache_t *hdr_full_cache; 1614static kmem_cache_t *hdr_l2only_cache; 1615static kmem_cache_t *buf_cache; 1616 1617static void 1618buf_fini(void) 1619{ 1620 int i; 1621 1622 kmem_free(buf_hash_table.ht_table, 1623 (buf_hash_table.ht_mask + 1) * sizeof (void *)); 1624 for (i = 0; i < BUF_LOCKS; i++) 1625 mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock); 1626 kmem_cache_destroy(hdr_full_cache); 1627 kmem_cache_destroy(hdr_l2only_cache); 1628 kmem_cache_destroy(buf_cache); 1629} 1630 1631/* 1632 * Constructor callback - called when the cache is empty 1633 * and a new buf is requested. 1634 */ 1635/* ARGSUSED */ 1636static int 1637hdr_full_cons(void *vbuf, void *unused, int kmflag) 1638{ 1639 arc_buf_hdr_t *hdr = vbuf; 1640 1641 bzero(hdr, HDR_FULL_SIZE); 1642 cv_init(&hdr->b_l1hdr.b_cv, NULL, CV_DEFAULT, NULL); 1643 refcount_create(&hdr->b_l1hdr.b_refcnt); 1644 mutex_init(&hdr->b_l1hdr.b_freeze_lock, NULL, MUTEX_DEFAULT, NULL); 1645 multilist_link_init(&hdr->b_l1hdr.b_arc_node); 1646 arc_space_consume(HDR_FULL_SIZE, ARC_SPACE_HDRS); 1647 1648 return (0); 1649} 1650 1651/* ARGSUSED */ 1652static int 1653hdr_l2only_cons(void *vbuf, void *unused, int kmflag) 1654{ 1655 arc_buf_hdr_t *hdr = vbuf; 1656 1657 bzero(hdr, HDR_L2ONLY_SIZE); 1658 arc_space_consume(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS); 1659 1660 return (0); 1661} 1662 1663/* ARGSUSED */ 1664static int 1665buf_cons(void *vbuf, void *unused, int kmflag) 1666{ 1667 arc_buf_t *buf = vbuf; 1668 1669 bzero(buf, sizeof (arc_buf_t)); 1670 mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL); 1671 arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS); 1672 1673 return (0); 1674} 1675 1676/* 1677 * Destructor callback - called when a cached buf is 1678 * no longer required. 1679 */ 1680/* ARGSUSED */ 1681static void 1682hdr_full_dest(void *vbuf, void *unused) 1683{ 1684 arc_buf_hdr_t *hdr = vbuf; 1685 1686 ASSERT(HDR_EMPTY(hdr)); 1687 cv_destroy(&hdr->b_l1hdr.b_cv); 1688 refcount_destroy(&hdr->b_l1hdr.b_refcnt); 1689 mutex_destroy(&hdr->b_l1hdr.b_freeze_lock); 1690 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 1691 arc_space_return(HDR_FULL_SIZE, ARC_SPACE_HDRS); 1692} 1693 1694/* ARGSUSED */ 1695static void 1696hdr_l2only_dest(void *vbuf, void *unused) 1697{ 1698 arc_buf_hdr_t *hdr = vbuf; 1699 1700 ASSERT(HDR_EMPTY(hdr)); 1701 arc_space_return(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS); 1702} 1703 1704/* ARGSUSED */ 1705static void 1706buf_dest(void *vbuf, void *unused) 1707{ 1708 arc_buf_t *buf = vbuf; 1709 1710 mutex_destroy(&buf->b_evict_lock); 1711 arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS); 1712} 1713 1714/* 1715 * Reclaim callback -- invoked when memory is low. 1716 */ 1717/* ARGSUSED */ 1718static void 1719hdr_recl(void *unused) 1720{ 1721 dprintf("hdr_recl called\n"); 1722 /* 1723 * umem calls the reclaim func when we destroy the buf cache, 1724 * which is after we do arc_fini(). 1725 */ 1726 if (!arc_dead) 1727 cv_signal(&arc_reclaim_thread_cv); 1728} 1729 1730static void 1731buf_init(void) 1732{ 1733 uint64_t *ct; 1734 uint64_t hsize = 1ULL << 12; 1735 int i, j; 1736 1737 /* 1738 * The hash table is big enough to fill all of physical memory 1739 * with an average block size of zfs_arc_average_blocksize (default 8K). 1740 * By default, the table will take up 1741 * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers). 1742 */ 1743 while (hsize * zfs_arc_average_blocksize < (uint64_t)physmem * PAGESIZE) 1744 hsize <<= 1; 1745retry: 1746 buf_hash_table.ht_mask = hsize - 1; 1747 buf_hash_table.ht_table = 1748 kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP); 1749 if (buf_hash_table.ht_table == NULL) { 1750 ASSERT(hsize > (1ULL << 8)); 1751 hsize >>= 1; 1752 goto retry; 1753 } 1754 1755 hdr_full_cache = kmem_cache_create("arc_buf_hdr_t_full", HDR_FULL_SIZE, 1756 0, hdr_full_cons, hdr_full_dest, hdr_recl, NULL, NULL, 0); 1757 hdr_l2only_cache = kmem_cache_create("arc_buf_hdr_t_l2only", 1758 HDR_L2ONLY_SIZE, 0, hdr_l2only_cons, hdr_l2only_dest, hdr_recl, 1759 NULL, NULL, 0); 1760 buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t), 1761 0, buf_cons, buf_dest, NULL, NULL, NULL, 0); 1762 1763 for (i = 0; i < 256; i++) 1764 for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--) 1765 *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY); 1766 1767 for (i = 0; i < BUF_LOCKS; i++) { 1768 mutex_init(&buf_hash_table.ht_locks[i].ht_lock, 1769 NULL, MUTEX_DEFAULT, NULL); 1770 } 1771} 1772 1773/* 1774 * This is the size that the buf occupies in memory. If the buf is compressed, 1775 * it will correspond to the compressed size. You should use this method of 1776 * getting the buf size unless you explicitly need the logical size. 1777 */ 1778int32_t 1779arc_buf_size(arc_buf_t *buf) 1780{ 1781 return (ARC_BUF_COMPRESSED(buf) ? 1782 HDR_GET_PSIZE(buf->b_hdr) : HDR_GET_LSIZE(buf->b_hdr)); 1783} 1784 1785int32_t 1786arc_buf_lsize(arc_buf_t *buf) 1787{ 1788 return (HDR_GET_LSIZE(buf->b_hdr)); 1789} 1790 1791enum zio_compress 1792arc_get_compression(arc_buf_t *buf) 1793{ 1794 return (ARC_BUF_COMPRESSED(buf) ? 1795 HDR_GET_COMPRESS(buf->b_hdr) : ZIO_COMPRESS_OFF); 1796} 1797 1798#define ARC_MINTIME (hz>>4) /* 62 ms */ 1799 1800static inline boolean_t 1801arc_buf_is_shared(arc_buf_t *buf) 1802{ 1803 boolean_t shared = (buf->b_data != NULL && 1804 buf->b_hdr->b_l1hdr.b_pabd != NULL && 1805 abd_is_linear(buf->b_hdr->b_l1hdr.b_pabd) && 1806 buf->b_data == abd_to_buf(buf->b_hdr->b_l1hdr.b_pabd)); 1807 IMPLY(shared, HDR_SHARED_DATA(buf->b_hdr)); 1808 IMPLY(shared, ARC_BUF_SHARED(buf)); 1809 IMPLY(shared, ARC_BUF_COMPRESSED(buf) || ARC_BUF_LAST(buf)); 1810 1811 /* 1812 * It would be nice to assert arc_can_share() too, but the "hdr isn't 1813 * already being shared" requirement prevents us from doing that. 1814 */ 1815 1816 return (shared); 1817} 1818 1819/* 1820 * Free the checksum associated with this header. If there is no checksum, this 1821 * is a no-op. 1822 */ 1823static inline void 1824arc_cksum_free(arc_buf_hdr_t *hdr) 1825{ 1826 ASSERT(HDR_HAS_L1HDR(hdr)); 1827 mutex_enter(&hdr->b_l1hdr.b_freeze_lock); 1828 if (hdr->b_l1hdr.b_freeze_cksum != NULL) { 1829 kmem_free(hdr->b_l1hdr.b_freeze_cksum, sizeof (zio_cksum_t)); 1830 hdr->b_l1hdr.b_freeze_cksum = NULL; 1831 } 1832 mutex_exit(&hdr->b_l1hdr.b_freeze_lock); 1833} 1834 1835/* 1836 * Return true iff at least one of the bufs on hdr is not compressed. 1837 */ 1838static boolean_t 1839arc_hdr_has_uncompressed_buf(arc_buf_hdr_t *hdr) 1840{ 1841 for (arc_buf_t *b = hdr->b_l1hdr.b_buf; b != NULL; b = b->b_next) { 1842 if (!ARC_BUF_COMPRESSED(b)) { 1843 return (B_TRUE); 1844 } 1845 } 1846 return (B_FALSE); 1847} 1848 1849/* 1850 * If we've turned on the ZFS_DEBUG_MODIFY flag, verify that the buf's data 1851 * matches the checksum that is stored in the hdr. If there is no checksum, 1852 * or if the buf is compressed, this is a no-op. 1853 */ 1854static void 1855arc_cksum_verify(arc_buf_t *buf) 1856{ 1857 arc_buf_hdr_t *hdr = buf->b_hdr; 1858 zio_cksum_t zc; 1859 1860 if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 1861 return; 1862 1863 if (ARC_BUF_COMPRESSED(buf)) { 1864 ASSERT(hdr->b_l1hdr.b_freeze_cksum == NULL || 1865 arc_hdr_has_uncompressed_buf(hdr)); 1866 return; 1867 } 1868 1869 ASSERT(HDR_HAS_L1HDR(hdr)); 1870 1871 mutex_enter(&hdr->b_l1hdr.b_freeze_lock); 1872 if (hdr->b_l1hdr.b_freeze_cksum == NULL || HDR_IO_ERROR(hdr)) { 1873 mutex_exit(&hdr->b_l1hdr.b_freeze_lock); 1874 return; 1875 } 1876 1877 fletcher_2_native(buf->b_data, arc_buf_size(buf), NULL, &zc); 1878 if (!ZIO_CHECKSUM_EQUAL(*hdr->b_l1hdr.b_freeze_cksum, zc)) 1879 panic("buffer modified while frozen!"); 1880 mutex_exit(&hdr->b_l1hdr.b_freeze_lock); 1881} 1882 1883static boolean_t 1884arc_cksum_is_equal(arc_buf_hdr_t *hdr, zio_t *zio) 1885{ 1886 enum zio_compress compress = BP_GET_COMPRESS(zio->io_bp); 1887 boolean_t valid_cksum; 1888 1889 ASSERT(!BP_IS_EMBEDDED(zio->io_bp)); 1890 VERIFY3U(BP_GET_PSIZE(zio->io_bp), ==, HDR_GET_PSIZE(hdr)); 1891 1892 /* 1893 * We rely on the blkptr's checksum to determine if the block 1894 * is valid or not. When compressed arc is enabled, the l2arc 1895 * writes the block to the l2arc just as it appears in the pool. 1896 * This allows us to use the blkptr's checksum to validate the 1897 * data that we just read off of the l2arc without having to store 1898 * a separate checksum in the arc_buf_hdr_t. However, if compressed 1899 * arc is disabled, then the data written to the l2arc is always 1900 * uncompressed and won't match the block as it exists in the main 1901 * pool. When this is the case, we must first compress it if it is 1902 * compressed on the main pool before we can validate the checksum. 1903 */ 1904 if (!HDR_COMPRESSION_ENABLED(hdr) && compress != ZIO_COMPRESS_OFF) { 1905 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF); 1906 uint64_t lsize = HDR_GET_LSIZE(hdr); 1907 uint64_t csize; 1908 1909 abd_t *cdata = abd_alloc_linear(HDR_GET_PSIZE(hdr), B_TRUE); 1910 csize = zio_compress_data(compress, zio->io_abd, 1911 abd_to_buf(cdata), lsize); 1912 1913 ASSERT3U(csize, <=, HDR_GET_PSIZE(hdr)); 1914 if (csize < HDR_GET_PSIZE(hdr)) { 1915 /* 1916 * Compressed blocks are always a multiple of the 1917 * smallest ashift in the pool. Ideally, we would 1918 * like to round up the csize to the next 1919 * spa_min_ashift but that value may have changed 1920 * since the block was last written. Instead, 1921 * we rely on the fact that the hdr's psize 1922 * was set to the psize of the block when it was 1923 * last written. We set the csize to that value 1924 * and zero out any part that should not contain 1925 * data. 1926 */ 1927 abd_zero_off(cdata, csize, HDR_GET_PSIZE(hdr) - csize); 1928 csize = HDR_GET_PSIZE(hdr); 1929 } 1930 zio_push_transform(zio, cdata, csize, HDR_GET_PSIZE(hdr), NULL); 1931 } 1932 1933 /* 1934 * Block pointers always store the checksum for the logical data. 1935 * If the block pointer has the gang bit set, then the checksum 1936 * it represents is for the reconstituted data and not for an 1937 * individual gang member. The zio pipeline, however, must be able to 1938 * determine the checksum of each of the gang constituents so it 1939 * treats the checksum comparison differently than what we need 1940 * for l2arc blocks. This prevents us from using the 1941 * zio_checksum_error() interface directly. Instead we must call the 1942 * zio_checksum_error_impl() so that we can ensure the checksum is 1943 * generated using the correct checksum algorithm and accounts for the 1944 * logical I/O size and not just a gang fragment. 1945 */ 1946 valid_cksum = (zio_checksum_error_impl(zio->io_spa, zio->io_bp, 1947 BP_GET_CHECKSUM(zio->io_bp), zio->io_abd, zio->io_size, 1948 zio->io_offset, NULL) == 0); 1949 zio_pop_transforms(zio); 1950 return (valid_cksum); 1951} 1952 1953/* 1954 * Given a buf full of data, if ZFS_DEBUG_MODIFY is enabled this computes a 1955 * checksum and attaches it to the buf's hdr so that we can ensure that the buf 1956 * isn't modified later on. If buf is compressed or there is already a checksum 1957 * on the hdr, this is a no-op (we only checksum uncompressed bufs). 1958 */ 1959static void 1960arc_cksum_compute(arc_buf_t *buf) 1961{ 1962 arc_buf_hdr_t *hdr = buf->b_hdr; 1963 1964 if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 1965 return; 1966 1967 ASSERT(HDR_HAS_L1HDR(hdr)); 1968 1969 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1970 if (hdr->b_l1hdr.b_freeze_cksum != NULL) { 1971 ASSERT(arc_hdr_has_uncompressed_buf(hdr)); 1972 mutex_exit(&hdr->b_l1hdr.b_freeze_lock); 1973 return; 1974 } else if (ARC_BUF_COMPRESSED(buf)) { 1975 mutex_exit(&hdr->b_l1hdr.b_freeze_lock); 1976 return; 1977 } 1978 1979 ASSERT(!ARC_BUF_COMPRESSED(buf)); 1980 hdr->b_l1hdr.b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t), 1981 KM_SLEEP); 1982 fletcher_2_native(buf->b_data, arc_buf_size(buf), NULL, 1983 hdr->b_l1hdr.b_freeze_cksum); 1984 mutex_exit(&hdr->b_l1hdr.b_freeze_lock); 1985#ifdef illumos 1986 arc_buf_watch(buf); 1987#endif 1988} 1989 1990#ifdef illumos 1991#ifndef _KERNEL 1992typedef struct procctl { 1993 long cmd; 1994 prwatch_t prwatch; 1995} procctl_t; 1996#endif 1997 1998/* ARGSUSED */ 1999static void 2000arc_buf_unwatch(arc_buf_t *buf) 2001{ 2002#ifndef _KERNEL 2003 if (arc_watch) { 2004 int result; 2005 procctl_t ctl; 2006 ctl.cmd = PCWATCH; 2007 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data; 2008 ctl.prwatch.pr_size = 0; 2009 ctl.prwatch.pr_wflags = 0; 2010 result = write(arc_procfd, &ctl, sizeof (ctl)); 2011 ASSERT3U(result, ==, sizeof (ctl)); 2012 } 2013#endif 2014} 2015 2016/* ARGSUSED */ 2017static void 2018arc_buf_watch(arc_buf_t *buf) 2019{ 2020#ifndef _KERNEL 2021 if (arc_watch) { 2022 int result; 2023 procctl_t ctl; 2024 ctl.cmd = PCWATCH; 2025 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data; 2026 ctl.prwatch.pr_size = arc_buf_size(buf); 2027 ctl.prwatch.pr_wflags = WA_WRITE; 2028 result = write(arc_procfd, &ctl, sizeof (ctl)); 2029 ASSERT3U(result, ==, sizeof (ctl)); 2030 } 2031#endif 2032} 2033#endif /* illumos */ 2034 2035static arc_buf_contents_t 2036arc_buf_type(arc_buf_hdr_t *hdr) 2037{ 2038 arc_buf_contents_t type; 2039 if (HDR_ISTYPE_METADATA(hdr)) { 2040 type = ARC_BUFC_METADATA; 2041 } else { 2042 type = ARC_BUFC_DATA; 2043 } 2044 VERIFY3U(hdr->b_type, ==, type); 2045 return (type); 2046} 2047 2048boolean_t 2049arc_is_metadata(arc_buf_t *buf) 2050{ 2051 return (HDR_ISTYPE_METADATA(buf->b_hdr) != 0); 2052} 2053 2054static uint32_t 2055arc_bufc_to_flags(arc_buf_contents_t type) 2056{ 2057 switch (type) { 2058 case ARC_BUFC_DATA: 2059 /* metadata field is 0 if buffer contains normal data */ 2060 return (0); 2061 case ARC_BUFC_METADATA: 2062 return (ARC_FLAG_BUFC_METADATA); 2063 default: 2064 break; 2065 } 2066 panic("undefined ARC buffer type!"); 2067 return ((uint32_t)-1); 2068} 2069 2070void 2071arc_buf_thaw(arc_buf_t *buf) 2072{ 2073 arc_buf_hdr_t *hdr = buf->b_hdr; 2074 2075 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon); 2076 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 2077 2078 arc_cksum_verify(buf); 2079 2080 /* 2081 * Compressed buffers do not manipulate the b_freeze_cksum or 2082 * allocate b_thawed. 2083 */ 2084 if (ARC_BUF_COMPRESSED(buf)) { 2085 ASSERT(hdr->b_l1hdr.b_freeze_cksum == NULL || 2086 arc_hdr_has_uncompressed_buf(hdr)); 2087 return; 2088 } 2089 2090 ASSERT(HDR_HAS_L1HDR(hdr)); 2091 arc_cksum_free(hdr); 2092 2093 mutex_enter(&hdr->b_l1hdr.b_freeze_lock); 2094#ifdef ZFS_DEBUG 2095 if (zfs_flags & ZFS_DEBUG_MODIFY) { 2096 if (hdr->b_l1hdr.b_thawed != NULL) 2097 kmem_free(hdr->b_l1hdr.b_thawed, 1); 2098 hdr->b_l1hdr.b_thawed = kmem_alloc(1, KM_SLEEP); 2099 } 2100#endif 2101 2102 mutex_exit(&hdr->b_l1hdr.b_freeze_lock); 2103 2104#ifdef illumos 2105 arc_buf_unwatch(buf); 2106#endif 2107} 2108 2109void 2110arc_buf_freeze(arc_buf_t *buf) 2111{ 2112 arc_buf_hdr_t *hdr = buf->b_hdr; 2113 kmutex_t *hash_lock; 2114 2115 if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 2116 return; 2117 2118 if (ARC_BUF_COMPRESSED(buf)) { 2119 ASSERT(hdr->b_l1hdr.b_freeze_cksum == NULL || 2120 arc_hdr_has_uncompressed_buf(hdr)); 2121 return; 2122 } 2123 2124 hash_lock = HDR_LOCK(hdr); 2125 mutex_enter(hash_lock); 2126 2127 ASSERT(HDR_HAS_L1HDR(hdr)); 2128 ASSERT(hdr->b_l1hdr.b_freeze_cksum != NULL || 2129 hdr->b_l1hdr.b_state == arc_anon); 2130 arc_cksum_compute(buf); 2131 mutex_exit(hash_lock); 2132} 2133 2134/* 2135 * The arc_buf_hdr_t's b_flags should never be modified directly. Instead, 2136 * the following functions should be used to ensure that the flags are 2137 * updated in a thread-safe way. When manipulating the flags either 2138 * the hash_lock must be held or the hdr must be undiscoverable. This 2139 * ensures that we're not racing with any other threads when updating 2140 * the flags. 2141 */ 2142static inline void 2143arc_hdr_set_flags(arc_buf_hdr_t *hdr, arc_flags_t flags) 2144{ 2145 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); 2146 hdr->b_flags |= flags; 2147} 2148 2149static inline void 2150arc_hdr_clear_flags(arc_buf_hdr_t *hdr, arc_flags_t flags) 2151{ 2152 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); 2153 hdr->b_flags &= ~flags; 2154} 2155 2156/* 2157 * Setting the compression bits in the arc_buf_hdr_t's b_flags is 2158 * done in a special way since we have to clear and set bits 2159 * at the same time. Consumers that wish to set the compression bits 2160 * must use this function to ensure that the flags are updated in 2161 * thread-safe manner. 2162 */ 2163static void 2164arc_hdr_set_compress(arc_buf_hdr_t *hdr, enum zio_compress cmp) 2165{ 2166 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); 2167 2168 /* 2169 * Holes and embedded blocks will always have a psize = 0 so 2170 * we ignore the compression of the blkptr and set the 2171 * arc_buf_hdr_t's compression to ZIO_COMPRESS_OFF. 2172 * Holes and embedded blocks remain anonymous so we don't 2173 * want to uncompress them. Mark them as uncompressed. 2174 */ 2175 if (!zfs_compressed_arc_enabled || HDR_GET_PSIZE(hdr) == 0) { 2176 arc_hdr_clear_flags(hdr, ARC_FLAG_COMPRESSED_ARC); 2177 HDR_SET_COMPRESS(hdr, ZIO_COMPRESS_OFF); 2178 ASSERT(!HDR_COMPRESSION_ENABLED(hdr)); 2179 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF); 2180 } else { 2181 arc_hdr_set_flags(hdr, ARC_FLAG_COMPRESSED_ARC); 2182 HDR_SET_COMPRESS(hdr, cmp); 2183 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, cmp); 2184 ASSERT(HDR_COMPRESSION_ENABLED(hdr)); 2185 } 2186} 2187 2188/* 2189 * Looks for another buf on the same hdr which has the data decompressed, copies 2190 * from it, and returns true. If no such buf exists, returns false. 2191 */ 2192static boolean_t 2193arc_buf_try_copy_decompressed_data(arc_buf_t *buf) 2194{ 2195 arc_buf_hdr_t *hdr = buf->b_hdr; 2196 boolean_t copied = B_FALSE; 2197 2198 ASSERT(HDR_HAS_L1HDR(hdr)); 2199 ASSERT3P(buf->b_data, !=, NULL); 2200 ASSERT(!ARC_BUF_COMPRESSED(buf)); 2201 2202 for (arc_buf_t *from = hdr->b_l1hdr.b_buf; from != NULL; 2203 from = from->b_next) { 2204 /* can't use our own data buffer */ 2205 if (from == buf) { 2206 continue; 2207 } 2208 2209 if (!ARC_BUF_COMPRESSED(from)) { 2210 bcopy(from->b_data, buf->b_data, arc_buf_size(buf)); 2211 copied = B_TRUE; 2212 break; 2213 } 2214 } 2215 2216 /* 2217 * There were no decompressed bufs, so there should not be a 2218 * checksum on the hdr either. 2219 */ 2220 EQUIV(!copied, hdr->b_l1hdr.b_freeze_cksum == NULL); 2221 2222 return (copied); 2223} 2224 2225/* 2226 * Given a buf that has a data buffer attached to it, this function will 2227 * efficiently fill the buf with data of the specified compression setting from 2228 * the hdr and update the hdr's b_freeze_cksum if necessary. If the buf and hdr 2229 * are already sharing a data buf, no copy is performed. 2230 * 2231 * If the buf is marked as compressed but uncompressed data was requested, this 2232 * will allocate a new data buffer for the buf, remove that flag, and fill the 2233 * buf with uncompressed data. You can't request a compressed buf on a hdr with 2234 * uncompressed data, and (since we haven't added support for it yet) if you 2235 * want compressed data your buf must already be marked as compressed and have 2236 * the correct-sized data buffer. 2237 */ 2238static int 2239arc_buf_fill(arc_buf_t *buf, boolean_t compressed) 2240{ 2241 arc_buf_hdr_t *hdr = buf->b_hdr; 2242 boolean_t hdr_compressed = (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF); 2243 dmu_object_byteswap_t bswap = hdr->b_l1hdr.b_byteswap; 2244 2245 ASSERT3P(buf->b_data, !=, NULL); 2246 IMPLY(compressed, hdr_compressed); 2247 IMPLY(compressed, ARC_BUF_COMPRESSED(buf)); 2248 2249 if (hdr_compressed == compressed) { 2250 if (!arc_buf_is_shared(buf)) { 2251 abd_copy_to_buf(buf->b_data, hdr->b_l1hdr.b_pabd, 2252 arc_buf_size(buf)); 2253 } 2254 } else { 2255 ASSERT(hdr_compressed); 2256 ASSERT(!compressed); 2257 ASSERT3U(HDR_GET_LSIZE(hdr), !=, HDR_GET_PSIZE(hdr)); 2258 2259 /* 2260 * If the buf is sharing its data with the hdr, unlink it and 2261 * allocate a new data buffer for the buf. 2262 */ 2263 if (arc_buf_is_shared(buf)) { 2264 ASSERT(ARC_BUF_COMPRESSED(buf)); 2265 2266 /* We need to give the buf it's own b_data */ 2267 buf->b_flags &= ~ARC_BUF_FLAG_SHARED; 2268 buf->b_data = 2269 arc_get_data_buf(hdr, HDR_GET_LSIZE(hdr), buf); 2270 arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA); 2271 2272 /* Previously overhead was 0; just add new overhead */ 2273 ARCSTAT_INCR(arcstat_overhead_size, HDR_GET_LSIZE(hdr)); 2274 } else if (ARC_BUF_COMPRESSED(buf)) { 2275 /* We need to reallocate the buf's b_data */ 2276 arc_free_data_buf(hdr, buf->b_data, HDR_GET_PSIZE(hdr), 2277 buf); 2278 buf->b_data = 2279 arc_get_data_buf(hdr, HDR_GET_LSIZE(hdr), buf); 2280 2281 /* We increased the size of b_data; update overhead */ 2282 ARCSTAT_INCR(arcstat_overhead_size, 2283 HDR_GET_LSIZE(hdr) - HDR_GET_PSIZE(hdr)); 2284 } 2285 2286 /* 2287 * Regardless of the buf's previous compression settings, it 2288 * should not be compressed at the end of this function. 2289 */ 2290 buf->b_flags &= ~ARC_BUF_FLAG_COMPRESSED; 2291 2292 /* 2293 * Try copying the data from another buf which already has a 2294 * decompressed version. If that's not possible, it's time to 2295 * bite the bullet and decompress the data from the hdr. 2296 */ 2297 if (arc_buf_try_copy_decompressed_data(buf)) { 2298 /* Skip byteswapping and checksumming (already done) */ 2299 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, !=, NULL); 2300 return (0); 2301 } else { 2302 int error = zio_decompress_data(HDR_GET_COMPRESS(hdr), 2303 hdr->b_l1hdr.b_pabd, buf->b_data, 2304 HDR_GET_PSIZE(hdr), HDR_GET_LSIZE(hdr)); 2305 2306 /* 2307 * Absent hardware errors or software bugs, this should 2308 * be impossible, but log it anyway so we can debug it. 2309 */ 2310 if (error != 0) { 2311 zfs_dbgmsg( 2312 "hdr %p, compress %d, psize %d, lsize %d", 2313 hdr, HDR_GET_COMPRESS(hdr), 2314 HDR_GET_PSIZE(hdr), HDR_GET_LSIZE(hdr)); 2315 return (SET_ERROR(EIO)); 2316 } 2317 } 2318 } 2319 2320 /* Byteswap the buf's data if necessary */ 2321 if (bswap != DMU_BSWAP_NUMFUNCS) { 2322 ASSERT(!HDR_SHARED_DATA(hdr)); 2323 ASSERT3U(bswap, <, DMU_BSWAP_NUMFUNCS); 2324 dmu_ot_byteswap[bswap].ob_func(buf->b_data, HDR_GET_LSIZE(hdr)); 2325 } 2326 2327 /* Compute the hdr's checksum if necessary */ 2328 arc_cksum_compute(buf); 2329 2330 return (0); 2331} 2332 2333int 2334arc_decompress(arc_buf_t *buf) 2335{ 2336 return (arc_buf_fill(buf, B_FALSE)); 2337} 2338 2339/* 2340 * Return the size of the block, b_pabd, that is stored in the arc_buf_hdr_t. 2341 */ 2342static uint64_t 2343arc_hdr_size(arc_buf_hdr_t *hdr) 2344{ 2345 uint64_t size; 2346 2347 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF && 2348 HDR_GET_PSIZE(hdr) > 0) { 2349 size = HDR_GET_PSIZE(hdr); 2350 } else { 2351 ASSERT3U(HDR_GET_LSIZE(hdr), !=, 0); 2352 size = HDR_GET_LSIZE(hdr); 2353 } 2354 return (size); 2355} 2356 2357/* 2358 * Increment the amount of evictable space in the arc_state_t's refcount. 2359 * We account for the space used by the hdr and the arc buf individually 2360 * so that we can add and remove them from the refcount individually. 2361 */ 2362static void 2363arc_evictable_space_increment(arc_buf_hdr_t *hdr, arc_state_t *state) 2364{ 2365 arc_buf_contents_t type = arc_buf_type(hdr); 2366 2367 ASSERT(HDR_HAS_L1HDR(hdr)); 2368 2369 if (GHOST_STATE(state)) { 2370 ASSERT0(hdr->b_l1hdr.b_bufcnt); 2371 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 2372 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL); 2373 (void) refcount_add_many(&state->arcs_esize[type], 2374 HDR_GET_LSIZE(hdr), hdr); 2375 return; 2376 } 2377 2378 ASSERT(!GHOST_STATE(state)); 2379 if (hdr->b_l1hdr.b_pabd != NULL) { 2380 (void) refcount_add_many(&state->arcs_esize[type], 2381 arc_hdr_size(hdr), hdr); 2382 } 2383 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL; 2384 buf = buf->b_next) { 2385 if (arc_buf_is_shared(buf)) 2386 continue; 2387 (void) refcount_add_many(&state->arcs_esize[type], 2388 arc_buf_size(buf), buf); 2389 } 2390} 2391 2392/* 2393 * Decrement the amount of evictable space in the arc_state_t's refcount. 2394 * We account for the space used by the hdr and the arc buf individually 2395 * so that we can add and remove them from the refcount individually. 2396 */ 2397static void 2398arc_evictable_space_decrement(arc_buf_hdr_t *hdr, arc_state_t *state) 2399{ 2400 arc_buf_contents_t type = arc_buf_type(hdr); 2401 2402 ASSERT(HDR_HAS_L1HDR(hdr)); 2403 2404 if (GHOST_STATE(state)) { 2405 ASSERT0(hdr->b_l1hdr.b_bufcnt); 2406 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 2407 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL); 2408 (void) refcount_remove_many(&state->arcs_esize[type], 2409 HDR_GET_LSIZE(hdr), hdr); 2410 return; 2411 } 2412 2413 ASSERT(!GHOST_STATE(state)); 2414 if (hdr->b_l1hdr.b_pabd != NULL) { 2415 (void) refcount_remove_many(&state->arcs_esize[type], 2416 arc_hdr_size(hdr), hdr); 2417 } 2418 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL; 2419 buf = buf->b_next) { 2420 if (arc_buf_is_shared(buf)) 2421 continue; 2422 (void) refcount_remove_many(&state->arcs_esize[type], 2423 arc_buf_size(buf), buf); 2424 } 2425} 2426 2427/* 2428 * Add a reference to this hdr indicating that someone is actively 2429 * referencing that memory. When the refcount transitions from 0 to 1, 2430 * we remove it from the respective arc_state_t list to indicate that 2431 * it is not evictable. 2432 */ 2433static void 2434add_reference(arc_buf_hdr_t *hdr, void *tag) 2435{ 2436 ASSERT(HDR_HAS_L1HDR(hdr)); 2437 if (!MUTEX_HELD(HDR_LOCK(hdr))) { 2438 ASSERT(hdr->b_l1hdr.b_state == arc_anon); 2439 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 2440 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 2441 } 2442 2443 arc_state_t *state = hdr->b_l1hdr.b_state; 2444 2445 if ((refcount_add(&hdr->b_l1hdr.b_refcnt, tag) == 1) && 2446 (state != arc_anon)) { 2447 /* We don't use the L2-only state list. */ 2448 if (state != arc_l2c_only) { 2449 multilist_remove(state->arcs_list[arc_buf_type(hdr)], 2450 hdr); 2451 arc_evictable_space_decrement(hdr, state); 2452 } 2453 /* remove the prefetch flag if we get a reference */ 2454 arc_hdr_clear_flags(hdr, ARC_FLAG_PREFETCH); 2455 } 2456} 2457 2458/* 2459 * Remove a reference from this hdr. When the reference transitions from 2460 * 1 to 0 and we're not anonymous, then we add this hdr to the arc_state_t's 2461 * list making it eligible for eviction. 2462 */ 2463static int 2464remove_reference(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, void *tag) 2465{ 2466 int cnt; 2467 arc_state_t *state = hdr->b_l1hdr.b_state; 2468 2469 ASSERT(HDR_HAS_L1HDR(hdr)); 2470 ASSERT(state == arc_anon || MUTEX_HELD(hash_lock)); 2471 ASSERT(!GHOST_STATE(state)); 2472 2473 /* 2474 * arc_l2c_only counts as a ghost state so we don't need to explicitly 2475 * check to prevent usage of the arc_l2c_only list. 2476 */ 2477 if (((cnt = refcount_remove(&hdr->b_l1hdr.b_refcnt, tag)) == 0) && 2478 (state != arc_anon)) { 2479 multilist_insert(state->arcs_list[arc_buf_type(hdr)], hdr); 2480 ASSERT3U(hdr->b_l1hdr.b_bufcnt, >, 0); 2481 arc_evictable_space_increment(hdr, state); 2482 } 2483 return (cnt); 2484} 2485 2486/* 2487 * Move the supplied buffer to the indicated state. The hash lock 2488 * for the buffer must be held by the caller. 2489 */ 2490static void 2491arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *hdr, 2492 kmutex_t *hash_lock) 2493{ 2494 arc_state_t *old_state; 2495 int64_t refcnt; 2496 uint32_t bufcnt; 2497 boolean_t update_old, update_new; 2498 arc_buf_contents_t buftype = arc_buf_type(hdr); 2499 2500 /* 2501 * We almost always have an L1 hdr here, since we call arc_hdr_realloc() 2502 * in arc_read() when bringing a buffer out of the L2ARC. However, the 2503 * L1 hdr doesn't always exist when we change state to arc_anon before 2504 * destroying a header, in which case reallocating to add the L1 hdr is 2505 * pointless. 2506 */ 2507 if (HDR_HAS_L1HDR(hdr)) { 2508 old_state = hdr->b_l1hdr.b_state; 2509 refcnt = refcount_count(&hdr->b_l1hdr.b_refcnt); 2510 bufcnt = hdr->b_l1hdr.b_bufcnt; 2511 update_old = (bufcnt > 0 || hdr->b_l1hdr.b_pabd != NULL); 2512 } else { 2513 old_state = arc_l2c_only; 2514 refcnt = 0; 2515 bufcnt = 0; 2516 update_old = B_FALSE; 2517 } 2518 update_new = update_old; 2519 2520 ASSERT(MUTEX_HELD(hash_lock)); 2521 ASSERT3P(new_state, !=, old_state); 2522 ASSERT(!GHOST_STATE(new_state) || bufcnt == 0); 2523 ASSERT(old_state != arc_anon || bufcnt <= 1); 2524 2525 /* 2526 * If this buffer is evictable, transfer it from the 2527 * old state list to the new state list. 2528 */ 2529 if (refcnt == 0) { 2530 if (old_state != arc_anon && old_state != arc_l2c_only) { 2531 ASSERT(HDR_HAS_L1HDR(hdr)); 2532 multilist_remove(old_state->arcs_list[buftype], hdr); 2533 2534 if (GHOST_STATE(old_state)) { 2535 ASSERT0(bufcnt); 2536 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 2537 update_old = B_TRUE; 2538 } 2539 arc_evictable_space_decrement(hdr, old_state); 2540 } 2541 if (new_state != arc_anon && new_state != arc_l2c_only) { 2542 2543 /* 2544 * An L1 header always exists here, since if we're 2545 * moving to some L1-cached state (i.e. not l2c_only or 2546 * anonymous), we realloc the header to add an L1hdr 2547 * beforehand. 2548 */ 2549 ASSERT(HDR_HAS_L1HDR(hdr)); 2550 multilist_insert(new_state->arcs_list[buftype], hdr); 2551 2552 if (GHOST_STATE(new_state)) { 2553 ASSERT0(bufcnt); 2554 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 2555 update_new = B_TRUE; 2556 } 2557 arc_evictable_space_increment(hdr, new_state); 2558 } 2559 } 2560 2561 ASSERT(!HDR_EMPTY(hdr)); 2562 if (new_state == arc_anon && HDR_IN_HASH_TABLE(hdr)) 2563 buf_hash_remove(hdr); 2564 2565 /* adjust state sizes (ignore arc_l2c_only) */ 2566 2567 if (update_new && new_state != arc_l2c_only) { 2568 ASSERT(HDR_HAS_L1HDR(hdr)); 2569 if (GHOST_STATE(new_state)) { 2570 ASSERT0(bufcnt); 2571 2572 /* 2573 * When moving a header to a ghost state, we first 2574 * remove all arc buffers. Thus, we'll have a 2575 * bufcnt of zero, and no arc buffer to use for 2576 * the reference. As a result, we use the arc 2577 * header pointer for the reference. 2578 */ 2579 (void) refcount_add_many(&new_state->arcs_size, 2580 HDR_GET_LSIZE(hdr), hdr); 2581 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL); 2582 } else { 2583 uint32_t buffers = 0; 2584 2585 /* 2586 * Each individual buffer holds a unique reference, 2587 * thus we must remove each of these references one 2588 * at a time. 2589 */ 2590 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL; 2591 buf = buf->b_next) { 2592 ASSERT3U(bufcnt, !=, 0); 2593 buffers++; 2594 2595 /* 2596 * When the arc_buf_t is sharing the data 2597 * block with the hdr, the owner of the 2598 * reference belongs to the hdr. Only 2599 * add to the refcount if the arc_buf_t is 2600 * not shared. 2601 */ 2602 if (arc_buf_is_shared(buf)) 2603 continue; 2604 2605 (void) refcount_add_many(&new_state->arcs_size, 2606 arc_buf_size(buf), buf); 2607 } 2608 ASSERT3U(bufcnt, ==, buffers); 2609 2610 if (hdr->b_l1hdr.b_pabd != NULL) { 2611 (void) refcount_add_many(&new_state->arcs_size, 2612 arc_hdr_size(hdr), hdr); 2613 } else { 2614 ASSERT(GHOST_STATE(old_state)); 2615 } 2616 } 2617 } 2618 2619 if (update_old && old_state != arc_l2c_only) { 2620 ASSERT(HDR_HAS_L1HDR(hdr)); 2621 if (GHOST_STATE(old_state)) { 2622 ASSERT0(bufcnt); 2623 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL); 2624 2625 /* 2626 * When moving a header off of a ghost state, 2627 * the header will not contain any arc buffers. 2628 * We use the arc header pointer for the reference 2629 * which is exactly what we did when we put the 2630 * header on the ghost state. 2631 */ 2632 2633 (void) refcount_remove_many(&old_state->arcs_size, 2634 HDR_GET_LSIZE(hdr), hdr); 2635 } else { 2636 uint32_t buffers = 0; 2637 2638 /* 2639 * Each individual buffer holds a unique reference, 2640 * thus we must remove each of these references one 2641 * at a time. 2642 */ 2643 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL; 2644 buf = buf->b_next) { 2645 ASSERT3U(bufcnt, !=, 0); 2646 buffers++; 2647 2648 /* 2649 * When the arc_buf_t is sharing the data 2650 * block with the hdr, the owner of the 2651 * reference belongs to the hdr. Only 2652 * add to the refcount if the arc_buf_t is 2653 * not shared. 2654 */ 2655 if (arc_buf_is_shared(buf)) 2656 continue; 2657 2658 (void) refcount_remove_many( 2659 &old_state->arcs_size, arc_buf_size(buf), 2660 buf); 2661 } 2662 ASSERT3U(bufcnt, ==, buffers); 2663 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL); 2664 (void) refcount_remove_many( 2665 &old_state->arcs_size, arc_hdr_size(hdr), hdr); 2666 } 2667 } 2668 2669 if (HDR_HAS_L1HDR(hdr)) 2670 hdr->b_l1hdr.b_state = new_state; 2671 2672 /* 2673 * L2 headers should never be on the L2 state list since they don't 2674 * have L1 headers allocated. 2675 */ 2676 ASSERT(multilist_is_empty(arc_l2c_only->arcs_list[ARC_BUFC_DATA]) && 2677 multilist_is_empty(arc_l2c_only->arcs_list[ARC_BUFC_METADATA])); 2678} 2679 2680void 2681arc_space_consume(uint64_t space, arc_space_type_t type) 2682{ 2683 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES); 2684 2685 switch (type) { 2686 case ARC_SPACE_DATA: 2687 aggsum_add(&astat_data_size, space); 2688 break; 2689 case ARC_SPACE_META: 2690 aggsum_add(&astat_metadata_size, space); 2691 break; 2692 case ARC_SPACE_OTHER: 2693 aggsum_add(&astat_other_size, space); 2694 break; 2695 case ARC_SPACE_HDRS: 2696 aggsum_add(&astat_hdr_size, space); 2697 break; 2698 case ARC_SPACE_L2HDRS: 2699 aggsum_add(&astat_l2_hdr_size, space); 2700 break; 2701 } 2702 2703 if (type != ARC_SPACE_DATA) 2704 aggsum_add(&arc_meta_used, space); 2705 2706 aggsum_add(&arc_size, space); 2707} 2708 2709void 2710arc_space_return(uint64_t space, arc_space_type_t type) 2711{ 2712 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES); 2713 2714 switch (type) { 2715 case ARC_SPACE_DATA: 2716 aggsum_add(&astat_data_size, -space); 2717 break; 2718 case ARC_SPACE_META: 2719 aggsum_add(&astat_metadata_size, -space); 2720 break; 2721 case ARC_SPACE_OTHER: 2722 aggsum_add(&astat_other_size, -space); 2723 break; 2724 case ARC_SPACE_HDRS: 2725 aggsum_add(&astat_hdr_size, -space); 2726 break; 2727 case ARC_SPACE_L2HDRS: 2728 aggsum_add(&astat_l2_hdr_size, -space); 2729 break; 2730 } 2731 2732 if (type != ARC_SPACE_DATA) { 2733 ASSERT(aggsum_compare(&arc_meta_used, space) >= 0); 2734 /* 2735 * We use the upper bound here rather than the precise value 2736 * because the arc_meta_max value doesn't need to be 2737 * precise. It's only consumed by humans via arcstats. 2738 */ 2739 if (arc_meta_max < aggsum_upper_bound(&arc_meta_used)) 2740 arc_meta_max = aggsum_upper_bound(&arc_meta_used); 2741 aggsum_add(&arc_meta_used, -space); 2742 } 2743 2744 ASSERT(aggsum_compare(&arc_size, space) >= 0); 2745 aggsum_add(&arc_size, -space); 2746} 2747 2748/* 2749 * Given a hdr and a buf, returns whether that buf can share its b_data buffer 2750 * with the hdr's b_pabd. 2751 */ 2752static boolean_t 2753arc_can_share(arc_buf_hdr_t *hdr, arc_buf_t *buf) 2754{ 2755 /* 2756 * The criteria for sharing a hdr's data are: 2757 * 1. the hdr's compression matches the buf's compression 2758 * 2. the hdr doesn't need to be byteswapped 2759 * 3. the hdr isn't already being shared 2760 * 4. the buf is either compressed or it is the last buf in the hdr list 2761 * 2762 * Criterion #4 maintains the invariant that shared uncompressed 2763 * bufs must be the final buf in the hdr's b_buf list. Reading this, you 2764 * might ask, "if a compressed buf is allocated first, won't that be the 2765 * last thing in the list?", but in that case it's impossible to create 2766 * a shared uncompressed buf anyway (because the hdr must be compressed 2767 * to have the compressed buf). You might also think that #3 is 2768 * sufficient to make this guarantee, however it's possible 2769 * (specifically in the rare L2ARC write race mentioned in 2770 * arc_buf_alloc_impl()) there will be an existing uncompressed buf that 2771 * is sharable, but wasn't at the time of its allocation. Rather than 2772 * allow a new shared uncompressed buf to be created and then shuffle 2773 * the list around to make it the last element, this simply disallows 2774 * sharing if the new buf isn't the first to be added. 2775 */ 2776 ASSERT3P(buf->b_hdr, ==, hdr); 2777 boolean_t hdr_compressed = HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF; 2778 boolean_t buf_compressed = ARC_BUF_COMPRESSED(buf) != 0; 2779 return (buf_compressed == hdr_compressed && 2780 hdr->b_l1hdr.b_byteswap == DMU_BSWAP_NUMFUNCS && 2781 !HDR_SHARED_DATA(hdr) && 2782 (ARC_BUF_LAST(buf) || ARC_BUF_COMPRESSED(buf))); 2783} 2784 2785/* 2786 * Allocate a buf for this hdr. If you care about the data that's in the hdr, 2787 * or if you want a compressed buffer, pass those flags in. Returns 0 if the 2788 * copy was made successfully, or an error code otherwise. 2789 */ 2790static int 2791arc_buf_alloc_impl(arc_buf_hdr_t *hdr, void *tag, boolean_t compressed, 2792 boolean_t fill, arc_buf_t **ret) 2793{ 2794 arc_buf_t *buf; 2795 2796 ASSERT(HDR_HAS_L1HDR(hdr)); 2797 ASSERT3U(HDR_GET_LSIZE(hdr), >, 0); 2798 VERIFY(hdr->b_type == ARC_BUFC_DATA || 2799 hdr->b_type == ARC_BUFC_METADATA); 2800 ASSERT3P(ret, !=, NULL); 2801 ASSERT3P(*ret, ==, NULL); 2802 2803 buf = *ret = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 2804 buf->b_hdr = hdr; 2805 buf->b_data = NULL; 2806 buf->b_next = hdr->b_l1hdr.b_buf; 2807 buf->b_flags = 0; 2808 2809 add_reference(hdr, tag); 2810 2811 /* 2812 * We're about to change the hdr's b_flags. We must either 2813 * hold the hash_lock or be undiscoverable. 2814 */ 2815 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); 2816 2817 /* 2818 * Only honor requests for compressed bufs if the hdr is actually 2819 * compressed. 2820 */ 2821 if (compressed && HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF) 2822 buf->b_flags |= ARC_BUF_FLAG_COMPRESSED; 2823 2824 /* 2825 * If the hdr's data can be shared then we share the data buffer and 2826 * set the appropriate bit in the hdr's b_flags to indicate the hdr is 2827 * sharing it's b_pabd with the arc_buf_t. Otherwise, we allocate a new 2828 * buffer to store the buf's data. 2829 * 2830 * There are two additional restrictions here because we're sharing 2831 * hdr -> buf instead of the usual buf -> hdr. First, the hdr can't be 2832 * actively involved in an L2ARC write, because if this buf is used by 2833 * an arc_write() then the hdr's data buffer will be released when the 2834 * write completes, even though the L2ARC write might still be using it. 2835 * Second, the hdr's ABD must be linear so that the buf's user doesn't 2836 * need to be ABD-aware. 2837 */ 2838 boolean_t can_share = arc_can_share(hdr, buf) && !HDR_L2_WRITING(hdr) && 2839 abd_is_linear(hdr->b_l1hdr.b_pabd); 2840 2841 /* Set up b_data and sharing */ 2842 if (can_share) { 2843 buf->b_data = abd_to_buf(hdr->b_l1hdr.b_pabd); 2844 buf->b_flags |= ARC_BUF_FLAG_SHARED; 2845 arc_hdr_set_flags(hdr, ARC_FLAG_SHARED_DATA); 2846 } else { 2847 buf->b_data = 2848 arc_get_data_buf(hdr, arc_buf_size(buf), buf); 2849 ARCSTAT_INCR(arcstat_overhead_size, arc_buf_size(buf)); 2850 } 2851 VERIFY3P(buf->b_data, !=, NULL); 2852 2853 hdr->b_l1hdr.b_buf = buf; 2854 hdr->b_l1hdr.b_bufcnt += 1; 2855 2856 /* 2857 * If the user wants the data from the hdr, we need to either copy or 2858 * decompress the data. 2859 */ 2860 if (fill) { 2861 return (arc_buf_fill(buf, ARC_BUF_COMPRESSED(buf) != 0)); 2862 } 2863 2864 return (0); 2865} 2866 2867static char *arc_onloan_tag = "onloan"; 2868 2869static inline void 2870arc_loaned_bytes_update(int64_t delta) 2871{ 2872 atomic_add_64(&arc_loaned_bytes, delta); 2873 2874 /* assert that it did not wrap around */ 2875 ASSERT3S(atomic_add_64_nv(&arc_loaned_bytes, 0), >=, 0); 2876} 2877 2878/* 2879 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in 2880 * flight data by arc_tempreserve_space() until they are "returned". Loaned 2881 * buffers must be returned to the arc before they can be used by the DMU or 2882 * freed. 2883 */ 2884arc_buf_t * 2885arc_loan_buf(spa_t *spa, boolean_t is_metadata, int size) 2886{ 2887 arc_buf_t *buf = arc_alloc_buf(spa, arc_onloan_tag, 2888 is_metadata ? ARC_BUFC_METADATA : ARC_BUFC_DATA, size); 2889 2890 arc_loaned_bytes_update(arc_buf_size(buf)); 2891 2892 return (buf); 2893} 2894 2895arc_buf_t * 2896arc_loan_compressed_buf(spa_t *spa, uint64_t psize, uint64_t lsize, 2897 enum zio_compress compression_type) 2898{ 2899 arc_buf_t *buf = arc_alloc_compressed_buf(spa, arc_onloan_tag, 2900 psize, lsize, compression_type); 2901 2902 arc_loaned_bytes_update(arc_buf_size(buf)); 2903 2904 return (buf); 2905} 2906 2907 2908/* 2909 * Return a loaned arc buffer to the arc. 2910 */ 2911void 2912arc_return_buf(arc_buf_t *buf, void *tag) 2913{ 2914 arc_buf_hdr_t *hdr = buf->b_hdr; 2915 2916 ASSERT3P(buf->b_data, !=, NULL); 2917 ASSERT(HDR_HAS_L1HDR(hdr)); 2918 (void) refcount_add(&hdr->b_l1hdr.b_refcnt, tag); 2919 (void) refcount_remove(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag); 2920 2921 arc_loaned_bytes_update(-arc_buf_size(buf)); 2922} 2923 2924/* Detach an arc_buf from a dbuf (tag) */ 2925void 2926arc_loan_inuse_buf(arc_buf_t *buf, void *tag) 2927{ 2928 arc_buf_hdr_t *hdr = buf->b_hdr; 2929 2930 ASSERT3P(buf->b_data, !=, NULL); 2931 ASSERT(HDR_HAS_L1HDR(hdr)); 2932 (void) refcount_add(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag); 2933 (void) refcount_remove(&hdr->b_l1hdr.b_refcnt, tag); 2934 2935 arc_loaned_bytes_update(arc_buf_size(buf)); 2936} 2937 2938static void 2939l2arc_free_abd_on_write(abd_t *abd, size_t size, arc_buf_contents_t type) 2940{ 2941 l2arc_data_free_t *df = kmem_alloc(sizeof (*df), KM_SLEEP); 2942 2943 df->l2df_abd = abd; 2944 df->l2df_size = size; 2945 df->l2df_type = type; 2946 mutex_enter(&l2arc_free_on_write_mtx); 2947 list_insert_head(l2arc_free_on_write, df); 2948 mutex_exit(&l2arc_free_on_write_mtx); 2949} 2950 2951static void 2952arc_hdr_free_on_write(arc_buf_hdr_t *hdr) 2953{ 2954 arc_state_t *state = hdr->b_l1hdr.b_state; 2955 arc_buf_contents_t type = arc_buf_type(hdr); 2956 uint64_t size = arc_hdr_size(hdr); 2957 2958 /* protected by hash lock, if in the hash table */ 2959 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) { 2960 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 2961 ASSERT(state != arc_anon && state != arc_l2c_only); 2962 2963 (void) refcount_remove_many(&state->arcs_esize[type], 2964 size, hdr); 2965 } 2966 (void) refcount_remove_many(&state->arcs_size, size, hdr); 2967 if (type == ARC_BUFC_METADATA) { 2968 arc_space_return(size, ARC_SPACE_META); 2969 } else { 2970 ASSERT(type == ARC_BUFC_DATA); 2971 arc_space_return(size, ARC_SPACE_DATA); 2972 } 2973 2974 l2arc_free_abd_on_write(hdr->b_l1hdr.b_pabd, size, type); 2975} 2976 2977/* 2978 * Share the arc_buf_t's data with the hdr. Whenever we are sharing the 2979 * data buffer, we transfer the refcount ownership to the hdr and update 2980 * the appropriate kstats. 2981 */ 2982static void 2983arc_share_buf(arc_buf_hdr_t *hdr, arc_buf_t *buf) 2984{ 2985 arc_state_t *state = hdr->b_l1hdr.b_state; 2986 2987 ASSERT(arc_can_share(hdr, buf)); 2988 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL); 2989 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); 2990 2991 /* 2992 * Start sharing the data buffer. We transfer the 2993 * refcount ownership to the hdr since it always owns 2994 * the refcount whenever an arc_buf_t is shared. 2995 */ 2996 refcount_transfer_ownership(&state->arcs_size, buf, hdr); 2997 hdr->b_l1hdr.b_pabd = abd_get_from_buf(buf->b_data, arc_buf_size(buf)); 2998 abd_take_ownership_of_buf(hdr->b_l1hdr.b_pabd, 2999 HDR_ISTYPE_METADATA(hdr)); 3000 arc_hdr_set_flags(hdr, ARC_FLAG_SHARED_DATA); 3001 buf->b_flags |= ARC_BUF_FLAG_SHARED; 3002 3003 /* 3004 * Since we've transferred ownership to the hdr we need 3005 * to increment its compressed and uncompressed kstats and 3006 * decrement the overhead size. 3007 */ 3008 ARCSTAT_INCR(arcstat_compressed_size, arc_hdr_size(hdr)); 3009 ARCSTAT_INCR(arcstat_uncompressed_size, HDR_GET_LSIZE(hdr)); 3010 ARCSTAT_INCR(arcstat_overhead_size, -arc_buf_size(buf)); 3011} 3012 3013static void 3014arc_unshare_buf(arc_buf_hdr_t *hdr, arc_buf_t *buf) 3015{ 3016 arc_state_t *state = hdr->b_l1hdr.b_state; 3017 3018 ASSERT(arc_buf_is_shared(buf)); 3019 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL); 3020 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); 3021 3022 /* 3023 * We are no longer sharing this buffer so we need 3024 * to transfer its ownership to the rightful owner. 3025 */ 3026 refcount_transfer_ownership(&state->arcs_size, hdr, buf); 3027 arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA); 3028 abd_release_ownership_of_buf(hdr->b_l1hdr.b_pabd); 3029 abd_put(hdr->b_l1hdr.b_pabd); 3030 hdr->b_l1hdr.b_pabd = NULL; 3031 buf->b_flags &= ~ARC_BUF_FLAG_SHARED; 3032 3033 /* 3034 * Since the buffer is no longer shared between 3035 * the arc buf and the hdr, count it as overhead. 3036 */ 3037 ARCSTAT_INCR(arcstat_compressed_size, -arc_hdr_size(hdr)); 3038 ARCSTAT_INCR(arcstat_uncompressed_size, -HDR_GET_LSIZE(hdr)); 3039 ARCSTAT_INCR(arcstat_overhead_size, arc_buf_size(buf)); 3040} 3041 3042/* 3043 * Remove an arc_buf_t from the hdr's buf list and return the last 3044 * arc_buf_t on the list. If no buffers remain on the list then return 3045 * NULL. 3046 */ 3047static arc_buf_t * 3048arc_buf_remove(arc_buf_hdr_t *hdr, arc_buf_t *buf) 3049{ 3050 ASSERT(HDR_HAS_L1HDR(hdr)); 3051 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); 3052 3053 arc_buf_t **bufp = &hdr->b_l1hdr.b_buf; 3054 arc_buf_t *lastbuf = NULL; 3055 3056 /* 3057 * Remove the buf from the hdr list and locate the last 3058 * remaining buffer on the list. 3059 */ 3060 while (*bufp != NULL) { 3061 if (*bufp == buf) 3062 *bufp = buf->b_next; 3063 3064 /* 3065 * If we've removed a buffer in the middle of 3066 * the list then update the lastbuf and update 3067 * bufp. 3068 */ 3069 if (*bufp != NULL) { 3070 lastbuf = *bufp; 3071 bufp = &(*bufp)->b_next; 3072 } 3073 } 3074 buf->b_next = NULL; 3075 ASSERT3P(lastbuf, !=, buf); 3076 IMPLY(hdr->b_l1hdr.b_bufcnt > 0, lastbuf != NULL); 3077 IMPLY(hdr->b_l1hdr.b_bufcnt > 0, hdr->b_l1hdr.b_buf != NULL); 3078 IMPLY(lastbuf != NULL, ARC_BUF_LAST(lastbuf)); 3079 3080 return (lastbuf); 3081} 3082 3083/* 3084 * Free up buf->b_data and pull the arc_buf_t off of the the arc_buf_hdr_t's 3085 * list and free it. 3086 */ 3087static void 3088arc_buf_destroy_impl(arc_buf_t *buf) 3089{ 3090 arc_buf_hdr_t *hdr = buf->b_hdr; 3091 3092 /* 3093 * Free up the data associated with the buf but only if we're not 3094 * sharing this with the hdr. If we are sharing it with the hdr, the 3095 * hdr is responsible for doing the free. 3096 */ 3097 if (buf->b_data != NULL) { 3098 /* 3099 * We're about to change the hdr's b_flags. We must either 3100 * hold the hash_lock or be undiscoverable. 3101 */ 3102 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); 3103 3104 arc_cksum_verify(buf); 3105#ifdef illumos 3106 arc_buf_unwatch(buf); 3107#endif 3108 3109 if (arc_buf_is_shared(buf)) { 3110 arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA); 3111 } else { 3112 uint64_t size = arc_buf_size(buf); 3113 arc_free_data_buf(hdr, buf->b_data, size, buf); 3114 ARCSTAT_INCR(arcstat_overhead_size, -size); 3115 } 3116 buf->b_data = NULL; 3117 3118 ASSERT(hdr->b_l1hdr.b_bufcnt > 0); 3119 hdr->b_l1hdr.b_bufcnt -= 1; 3120 } 3121 3122 arc_buf_t *lastbuf = arc_buf_remove(hdr, buf); 3123 3124 if (ARC_BUF_SHARED(buf) && !ARC_BUF_COMPRESSED(buf)) { 3125 /* 3126 * If the current arc_buf_t is sharing its data buffer with the 3127 * hdr, then reassign the hdr's b_pabd to share it with the new 3128 * buffer at the end of the list. The shared buffer is always 3129 * the last one on the hdr's buffer list. 3130 * 3131 * There is an equivalent case for compressed bufs, but since 3132 * they aren't guaranteed to be the last buf in the list and 3133 * that is an exceedingly rare case, we just allow that space be 3134 * wasted temporarily. 3135 */ 3136 if (lastbuf != NULL) { 3137 /* Only one buf can be shared at once */ 3138 VERIFY(!arc_buf_is_shared(lastbuf)); 3139 /* hdr is uncompressed so can't have compressed buf */ 3140 VERIFY(!ARC_BUF_COMPRESSED(lastbuf)); 3141 3142 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL); 3143 arc_hdr_free_pabd(hdr); 3144 3145 /* 3146 * We must setup a new shared block between the 3147 * last buffer and the hdr. The data would have 3148 * been allocated by the arc buf so we need to transfer 3149 * ownership to the hdr since it's now being shared. 3150 */ 3151 arc_share_buf(hdr, lastbuf); 3152 } 3153 } else if (HDR_SHARED_DATA(hdr)) { 3154 /* 3155 * Uncompressed shared buffers are always at the end 3156 * of the list. Compressed buffers don't have the 3157 * same requirements. This makes it hard to 3158 * simply assert that the lastbuf is shared so 3159 * we rely on the hdr's compression flags to determine 3160 * if we have a compressed, shared buffer. 3161 */ 3162 ASSERT3P(lastbuf, !=, NULL); 3163 ASSERT(arc_buf_is_shared(lastbuf) || 3164 HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF); 3165 } 3166 3167 /* 3168 * Free the checksum if we're removing the last uncompressed buf from 3169 * this hdr. 3170 */ 3171 if (!arc_hdr_has_uncompressed_buf(hdr)) { 3172 arc_cksum_free(hdr); 3173 } 3174 3175 /* clean up the buf */ 3176 buf->b_hdr = NULL; 3177 kmem_cache_free(buf_cache, buf); 3178} 3179 3180static void 3181arc_hdr_alloc_pabd(arc_buf_hdr_t *hdr) 3182{ 3183 ASSERT3U(HDR_GET_LSIZE(hdr), >, 0); 3184 ASSERT(HDR_HAS_L1HDR(hdr)); 3185 ASSERT(!HDR_SHARED_DATA(hdr)); 3186 3187 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL); 3188 hdr->b_l1hdr.b_pabd = arc_get_data_abd(hdr, arc_hdr_size(hdr), hdr); 3189 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS; 3190 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL); 3191 3192 ARCSTAT_INCR(arcstat_compressed_size, arc_hdr_size(hdr)); 3193 ARCSTAT_INCR(arcstat_uncompressed_size, HDR_GET_LSIZE(hdr)); 3194} 3195 3196static void 3197arc_hdr_free_pabd(arc_buf_hdr_t *hdr) 3198{ 3199 ASSERT(HDR_HAS_L1HDR(hdr)); 3200 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL); 3201 3202 /* 3203 * If the hdr is currently being written to the l2arc then 3204 * we defer freeing the data by adding it to the l2arc_free_on_write 3205 * list. The l2arc will free the data once it's finished 3206 * writing it to the l2arc device. 3207 */ 3208 if (HDR_L2_WRITING(hdr)) { 3209 arc_hdr_free_on_write(hdr); 3210 ARCSTAT_BUMP(arcstat_l2_free_on_write); 3211 } else { 3212 arc_free_data_abd(hdr, hdr->b_l1hdr.b_pabd, 3213 arc_hdr_size(hdr), hdr); 3214 } 3215 hdr->b_l1hdr.b_pabd = NULL; 3216 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS; 3217 3218 ARCSTAT_INCR(arcstat_compressed_size, -arc_hdr_size(hdr)); 3219 ARCSTAT_INCR(arcstat_uncompressed_size, -HDR_GET_LSIZE(hdr)); 3220} 3221 3222static arc_buf_hdr_t * 3223arc_hdr_alloc(uint64_t spa, int32_t psize, int32_t lsize, 3224 enum zio_compress compression_type, arc_buf_contents_t type) 3225{ 3226 arc_buf_hdr_t *hdr; 3227 3228 VERIFY(type == ARC_BUFC_DATA || type == ARC_BUFC_METADATA); 3229 3230 hdr = kmem_cache_alloc(hdr_full_cache, KM_PUSHPAGE); 3231 ASSERT(HDR_EMPTY(hdr)); 3232 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL); 3233 ASSERT3P(hdr->b_l1hdr.b_thawed, ==, NULL); 3234 HDR_SET_PSIZE(hdr, psize); 3235 HDR_SET_LSIZE(hdr, lsize); 3236 hdr->b_spa = spa; 3237 hdr->b_type = type; 3238 hdr->b_flags = 0; 3239 arc_hdr_set_flags(hdr, arc_bufc_to_flags(type) | ARC_FLAG_HAS_L1HDR); 3240 arc_hdr_set_compress(hdr, compression_type); 3241 3242 hdr->b_l1hdr.b_state = arc_anon; 3243 hdr->b_l1hdr.b_arc_access = 0; 3244 hdr->b_l1hdr.b_bufcnt = 0; 3245 hdr->b_l1hdr.b_buf = NULL; 3246 3247 /* 3248 * Allocate the hdr's buffer. This will contain either 3249 * the compressed or uncompressed data depending on the block 3250 * it references and compressed arc enablement. 3251 */ 3252 arc_hdr_alloc_pabd(hdr); 3253 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 3254 3255 return (hdr); 3256} 3257 3258/* 3259 * Transition between the two allocation states for the arc_buf_hdr struct. 3260 * The arc_buf_hdr struct can be allocated with (hdr_full_cache) or without 3261 * (hdr_l2only_cache) the fields necessary for the L1 cache - the smaller 3262 * version is used when a cache buffer is only in the L2ARC in order to reduce 3263 * memory usage. 3264 */ 3265static arc_buf_hdr_t * 3266arc_hdr_realloc(arc_buf_hdr_t *hdr, kmem_cache_t *old, kmem_cache_t *new) 3267{ 3268 ASSERT(HDR_HAS_L2HDR(hdr)); 3269 3270 arc_buf_hdr_t *nhdr; 3271 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev; 3272 3273 ASSERT((old == hdr_full_cache && new == hdr_l2only_cache) || 3274 (old == hdr_l2only_cache && new == hdr_full_cache)); 3275 3276 nhdr = kmem_cache_alloc(new, KM_PUSHPAGE); 3277 3278 ASSERT(MUTEX_HELD(HDR_LOCK(hdr))); 3279 buf_hash_remove(hdr); 3280 3281 bcopy(hdr, nhdr, HDR_L2ONLY_SIZE); 3282 3283 if (new == hdr_full_cache) { 3284 arc_hdr_set_flags(nhdr, ARC_FLAG_HAS_L1HDR); 3285 /* 3286 * arc_access and arc_change_state need to be aware that a 3287 * header has just come out of L2ARC, so we set its state to 3288 * l2c_only even though it's about to change. 3289 */ 3290 nhdr->b_l1hdr.b_state = arc_l2c_only; 3291 3292 /* Verify previous threads set to NULL before freeing */ 3293 ASSERT3P(nhdr->b_l1hdr.b_pabd, ==, NULL); 3294 } else { 3295 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 3296 ASSERT0(hdr->b_l1hdr.b_bufcnt); 3297 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL); 3298 3299 /* 3300 * If we've reached here, We must have been called from 3301 * arc_evict_hdr(), as such we should have already been 3302 * removed from any ghost list we were previously on 3303 * (which protects us from racing with arc_evict_state), 3304 * thus no locking is needed during this check. 3305 */ 3306 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 3307 3308 /* 3309 * A buffer must not be moved into the arc_l2c_only 3310 * state if it's not finished being written out to the 3311 * l2arc device. Otherwise, the b_l1hdr.b_pabd field 3312 * might try to be accessed, even though it was removed. 3313 */ 3314 VERIFY(!HDR_L2_WRITING(hdr)); 3315 VERIFY3P(hdr->b_l1hdr.b_pabd, ==, NULL); 3316 3317#ifdef ZFS_DEBUG 3318 if (hdr->b_l1hdr.b_thawed != NULL) { 3319 kmem_free(hdr->b_l1hdr.b_thawed, 1); 3320 hdr->b_l1hdr.b_thawed = NULL; 3321 } 3322#endif 3323 3324 arc_hdr_clear_flags(nhdr, ARC_FLAG_HAS_L1HDR); 3325 } 3326 /* 3327 * The header has been reallocated so we need to re-insert it into any 3328 * lists it was on. 3329 */ 3330 (void) buf_hash_insert(nhdr, NULL); 3331 3332 ASSERT(list_link_active(&hdr->b_l2hdr.b_l2node)); 3333 3334 mutex_enter(&dev->l2ad_mtx); 3335 3336 /* 3337 * We must place the realloc'ed header back into the list at 3338 * the same spot. Otherwise, if it's placed earlier in the list, 3339 * l2arc_write_buffers() could find it during the function's 3340 * write phase, and try to write it out to the l2arc. 3341 */ 3342 list_insert_after(&dev->l2ad_buflist, hdr, nhdr); 3343 list_remove(&dev->l2ad_buflist, hdr); 3344 3345 mutex_exit(&dev->l2ad_mtx); 3346 3347 /* 3348 * Since we're using the pointer address as the tag when 3349 * incrementing and decrementing the l2ad_alloc refcount, we 3350 * must remove the old pointer (that we're about to destroy) and 3351 * add the new pointer to the refcount. Otherwise we'd remove 3352 * the wrong pointer address when calling arc_hdr_destroy() later. 3353 */ 3354 3355 (void) refcount_remove_many(&dev->l2ad_alloc, arc_hdr_size(hdr), hdr); 3356 (void) refcount_add_many(&dev->l2ad_alloc, arc_hdr_size(nhdr), nhdr); 3357 3358 buf_discard_identity(hdr); 3359 kmem_cache_free(old, hdr); 3360 3361 return (nhdr); 3362} 3363 3364/* 3365 * Allocate a new arc_buf_hdr_t and arc_buf_t and return the buf to the caller. 3366 * The buf is returned thawed since we expect the consumer to modify it. 3367 */ 3368arc_buf_t * 3369arc_alloc_buf(spa_t *spa, void *tag, arc_buf_contents_t type, int32_t size) 3370{ 3371 arc_buf_hdr_t *hdr = arc_hdr_alloc(spa_load_guid(spa), size, size, 3372 ZIO_COMPRESS_OFF, type); 3373 ASSERT(!MUTEX_HELD(HDR_LOCK(hdr))); 3374 3375 arc_buf_t *buf = NULL; 3376 VERIFY0(arc_buf_alloc_impl(hdr, tag, B_FALSE, B_FALSE, &buf)); 3377 arc_buf_thaw(buf); 3378 3379 return (buf); 3380} 3381 3382/* 3383 * Allocate a compressed buf in the same manner as arc_alloc_buf. Don't use this 3384 * for bufs containing metadata. 3385 */ 3386arc_buf_t * 3387arc_alloc_compressed_buf(spa_t *spa, void *tag, uint64_t psize, uint64_t lsize, 3388 enum zio_compress compression_type) 3389{ 3390 ASSERT3U(lsize, >, 0); 3391 ASSERT3U(lsize, >=, psize); 3392 ASSERT(compression_type > ZIO_COMPRESS_OFF); 3393 ASSERT(compression_type < ZIO_COMPRESS_FUNCTIONS); 3394 3395 arc_buf_hdr_t *hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize, 3396 compression_type, ARC_BUFC_DATA); 3397 ASSERT(!MUTEX_HELD(HDR_LOCK(hdr))); 3398 3399 arc_buf_t *buf = NULL; 3400 VERIFY0(arc_buf_alloc_impl(hdr, tag, B_TRUE, B_FALSE, &buf)); 3401 arc_buf_thaw(buf); 3402 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL); 3403 3404 if (!arc_buf_is_shared(buf)) { 3405 /* 3406 * To ensure that the hdr has the correct data in it if we call 3407 * arc_decompress() on this buf before it's been written to 3408 * disk, it's easiest if we just set up sharing between the 3409 * buf and the hdr. 3410 */ 3411 ASSERT(!abd_is_linear(hdr->b_l1hdr.b_pabd)); 3412 arc_hdr_free_pabd(hdr); 3413 arc_share_buf(hdr, buf); 3414 } 3415 3416 return (buf); 3417} 3418 3419static void 3420arc_hdr_l2hdr_destroy(arc_buf_hdr_t *hdr) 3421{ 3422 l2arc_buf_hdr_t *l2hdr = &hdr->b_l2hdr; 3423 l2arc_dev_t *dev = l2hdr->b_dev; 3424 uint64_t psize = arc_hdr_size(hdr); 3425 3426 ASSERT(MUTEX_HELD(&dev->l2ad_mtx)); 3427 ASSERT(HDR_HAS_L2HDR(hdr)); 3428 3429 list_remove(&dev->l2ad_buflist, hdr); 3430 3431 ARCSTAT_INCR(arcstat_l2_psize, -psize); 3432 ARCSTAT_INCR(arcstat_l2_lsize, -HDR_GET_LSIZE(hdr)); 3433 3434 vdev_space_update(dev->l2ad_vdev, -psize, 0, 0); 3435 3436 (void) refcount_remove_many(&dev->l2ad_alloc, psize, hdr); 3437 arc_hdr_clear_flags(hdr, ARC_FLAG_HAS_L2HDR); 3438} 3439 3440static void 3441arc_hdr_destroy(arc_buf_hdr_t *hdr) 3442{ 3443 if (HDR_HAS_L1HDR(hdr)) { 3444 ASSERT(hdr->b_l1hdr.b_buf == NULL || 3445 hdr->b_l1hdr.b_bufcnt > 0); 3446 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 3447 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon); 3448 } 3449 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 3450 ASSERT(!HDR_IN_HASH_TABLE(hdr)); 3451 3452 if (!HDR_EMPTY(hdr)) 3453 buf_discard_identity(hdr); 3454 3455 if (HDR_HAS_L2HDR(hdr)) { 3456 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev; 3457 boolean_t buflist_held = MUTEX_HELD(&dev->l2ad_mtx); 3458 3459 if (!buflist_held) 3460 mutex_enter(&dev->l2ad_mtx); 3461 3462 /* 3463 * Even though we checked this conditional above, we 3464 * need to check this again now that we have the 3465 * l2ad_mtx. This is because we could be racing with 3466 * another thread calling l2arc_evict() which might have 3467 * destroyed this header's L2 portion as we were waiting 3468 * to acquire the l2ad_mtx. If that happens, we don't 3469 * want to re-destroy the header's L2 portion. 3470 */ 3471 if (HDR_HAS_L2HDR(hdr)) { 3472 l2arc_trim(hdr); 3473 arc_hdr_l2hdr_destroy(hdr); 3474 } 3475 3476 if (!buflist_held) 3477 mutex_exit(&dev->l2ad_mtx); 3478 } 3479 3480 if (HDR_HAS_L1HDR(hdr)) { 3481 arc_cksum_free(hdr); 3482 3483 while (hdr->b_l1hdr.b_buf != NULL) 3484 arc_buf_destroy_impl(hdr->b_l1hdr.b_buf); 3485 3486#ifdef ZFS_DEBUG 3487 if (hdr->b_l1hdr.b_thawed != NULL) { 3488 kmem_free(hdr->b_l1hdr.b_thawed, 1); 3489 hdr->b_l1hdr.b_thawed = NULL; 3490 } 3491#endif 3492 3493 if (hdr->b_l1hdr.b_pabd != NULL) { 3494 arc_hdr_free_pabd(hdr); 3495 } 3496 } 3497 3498 ASSERT3P(hdr->b_hash_next, ==, NULL); 3499 if (HDR_HAS_L1HDR(hdr)) { 3500 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 3501 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL); 3502 kmem_cache_free(hdr_full_cache, hdr); 3503 } else { 3504 kmem_cache_free(hdr_l2only_cache, hdr); 3505 } 3506} 3507 3508void 3509arc_buf_destroy(arc_buf_t *buf, void* tag) 3510{ 3511 arc_buf_hdr_t *hdr = buf->b_hdr; 3512 kmutex_t *hash_lock = HDR_LOCK(hdr); 3513 3514 if (hdr->b_l1hdr.b_state == arc_anon) { 3515 ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1); 3516 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 3517 VERIFY0(remove_reference(hdr, NULL, tag)); 3518 arc_hdr_destroy(hdr); 3519 return; 3520 } 3521 3522 mutex_enter(hash_lock); 3523 ASSERT3P(hdr, ==, buf->b_hdr); 3524 ASSERT(hdr->b_l1hdr.b_bufcnt > 0); 3525 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 3526 ASSERT3P(hdr->b_l1hdr.b_state, !=, arc_anon); 3527 ASSERT3P(buf->b_data, !=, NULL); 3528 3529 (void) remove_reference(hdr, hash_lock, tag); 3530 arc_buf_destroy_impl(buf); 3531 mutex_exit(hash_lock); 3532} 3533 3534/* 3535 * Evict the arc_buf_hdr that is provided as a parameter. The resultant 3536 * state of the header is dependent on it's state prior to entering this 3537 * function. The following transitions are possible: 3538 * 3539 * - arc_mru -> arc_mru_ghost 3540 * - arc_mfu -> arc_mfu_ghost 3541 * - arc_mru_ghost -> arc_l2c_only 3542 * - arc_mru_ghost -> deleted 3543 * - arc_mfu_ghost -> arc_l2c_only 3544 * - arc_mfu_ghost -> deleted 3545 */ 3546static int64_t 3547arc_evict_hdr(arc_buf_hdr_t *hdr, kmutex_t *hash_lock) 3548{ 3549 arc_state_t *evicted_state, *state; 3550 int64_t bytes_evicted = 0; 3551 3552 ASSERT(MUTEX_HELD(hash_lock)); 3553 ASSERT(HDR_HAS_L1HDR(hdr)); 3554 3555 state = hdr->b_l1hdr.b_state; 3556 if (GHOST_STATE(state)) { 3557 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 3558 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 3559 3560 /* 3561 * l2arc_write_buffers() relies on a header's L1 portion 3562 * (i.e. its b_pabd field) during it's write phase. 3563 * Thus, we cannot push a header onto the arc_l2c_only 3564 * state (removing it's L1 piece) until the header is 3565 * done being written to the l2arc. 3566 */ 3567 if (HDR_HAS_L2HDR(hdr) && HDR_L2_WRITING(hdr)) { 3568 ARCSTAT_BUMP(arcstat_evict_l2_skip); 3569 return (bytes_evicted); 3570 } 3571 3572 ARCSTAT_BUMP(arcstat_deleted); 3573 bytes_evicted += HDR_GET_LSIZE(hdr); 3574 3575 DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, hdr); 3576 3577 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL); 3578 if (HDR_HAS_L2HDR(hdr)) { 3579 /* 3580 * This buffer is cached on the 2nd Level ARC; 3581 * don't destroy the header. 3582 */ 3583 arc_change_state(arc_l2c_only, hdr, hash_lock); 3584 /* 3585 * dropping from L1+L2 cached to L2-only, 3586 * realloc to remove the L1 header. 3587 */ 3588 hdr = arc_hdr_realloc(hdr, hdr_full_cache, 3589 hdr_l2only_cache); 3590 } else { 3591 arc_change_state(arc_anon, hdr, hash_lock); 3592 arc_hdr_destroy(hdr); 3593 } 3594 return (bytes_evicted); 3595 } 3596 3597 ASSERT(state == arc_mru || state == arc_mfu); 3598 evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost; 3599 3600 /* prefetch buffers have a minimum lifespan */ 3601 if (HDR_IO_IN_PROGRESS(hdr) || 3602 ((hdr->b_flags & (ARC_FLAG_PREFETCH | ARC_FLAG_INDIRECT)) && 3603 ddi_get_lbolt() - hdr->b_l1hdr.b_arc_access < 3604 arc_min_prefetch_lifespan)) { 3605 ARCSTAT_BUMP(arcstat_evict_skip); 3606 return (bytes_evicted); 3607 } 3608 3609 ASSERT0(refcount_count(&hdr->b_l1hdr.b_refcnt)); 3610 while (hdr->b_l1hdr.b_buf) { 3611 arc_buf_t *buf = hdr->b_l1hdr.b_buf; 3612 if (!mutex_tryenter(&buf->b_evict_lock)) { 3613 ARCSTAT_BUMP(arcstat_mutex_miss); 3614 break; 3615 } 3616 if (buf->b_data != NULL) 3617 bytes_evicted += HDR_GET_LSIZE(hdr); 3618 mutex_exit(&buf->b_evict_lock); 3619 arc_buf_destroy_impl(buf); 3620 } 3621 3622 if (HDR_HAS_L2HDR(hdr)) { 3623 ARCSTAT_INCR(arcstat_evict_l2_cached, HDR_GET_LSIZE(hdr)); 3624 } else { 3625 if (l2arc_write_eligible(hdr->b_spa, hdr)) { 3626 ARCSTAT_INCR(arcstat_evict_l2_eligible, 3627 HDR_GET_LSIZE(hdr)); 3628 } else { 3629 ARCSTAT_INCR(arcstat_evict_l2_ineligible, 3630 HDR_GET_LSIZE(hdr)); 3631 } 3632 } 3633 3634 if (hdr->b_l1hdr.b_bufcnt == 0) { 3635 arc_cksum_free(hdr); 3636 3637 bytes_evicted += arc_hdr_size(hdr); 3638 3639 /* 3640 * If this hdr is being evicted and has a compressed 3641 * buffer then we discard it here before we change states. 3642 * This ensures that the accounting is updated correctly 3643 * in arc_free_data_impl(). 3644 */ 3645 arc_hdr_free_pabd(hdr); 3646 3647 arc_change_state(evicted_state, hdr, hash_lock); 3648 ASSERT(HDR_IN_HASH_TABLE(hdr)); 3649 arc_hdr_set_flags(hdr, ARC_FLAG_IN_HASH_TABLE); 3650 DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, hdr); 3651 } 3652 3653 return (bytes_evicted); 3654} 3655 3656static uint64_t 3657arc_evict_state_impl(multilist_t *ml, int idx, arc_buf_hdr_t *marker, 3658 uint64_t spa, int64_t bytes) 3659{ 3660 multilist_sublist_t *mls; 3661 uint64_t bytes_evicted = 0; 3662 arc_buf_hdr_t *hdr; 3663 kmutex_t *hash_lock; 3664 int evict_count = 0; 3665 3666 ASSERT3P(marker, !=, NULL); 3667 IMPLY(bytes < 0, bytes == ARC_EVICT_ALL); 3668 3669 mls = multilist_sublist_lock(ml, idx); 3670 3671 for (hdr = multilist_sublist_prev(mls, marker); hdr != NULL; 3672 hdr = multilist_sublist_prev(mls, marker)) { 3673 if ((bytes != ARC_EVICT_ALL && bytes_evicted >= bytes) || 3674 (evict_count >= zfs_arc_evict_batch_limit)) 3675 break; 3676 3677 /* 3678 * To keep our iteration location, move the marker 3679 * forward. Since we're not holding hdr's hash lock, we 3680 * must be very careful and not remove 'hdr' from the 3681 * sublist. Otherwise, other consumers might mistake the 3682 * 'hdr' as not being on a sublist when they call the 3683 * multilist_link_active() function (they all rely on 3684 * the hash lock protecting concurrent insertions and 3685 * removals). multilist_sublist_move_forward() was 3686 * specifically implemented to ensure this is the case 3687 * (only 'marker' will be removed and re-inserted). 3688 */ 3689 multilist_sublist_move_forward(mls, marker); 3690 3691 /* 3692 * The only case where the b_spa field should ever be 3693 * zero, is the marker headers inserted by 3694 * arc_evict_state(). It's possible for multiple threads 3695 * to be calling arc_evict_state() concurrently (e.g. 3696 * dsl_pool_close() and zio_inject_fault()), so we must 3697 * skip any markers we see from these other threads. 3698 */ 3699 if (hdr->b_spa == 0) 3700 continue; 3701 3702 /* we're only interested in evicting buffers of a certain spa */ 3703 if (spa != 0 && hdr->b_spa != spa) { 3704 ARCSTAT_BUMP(arcstat_evict_skip); 3705 continue; 3706 } 3707 3708 hash_lock = HDR_LOCK(hdr); 3709 3710 /* 3711 * We aren't calling this function from any code path 3712 * that would already be holding a hash lock, so we're 3713 * asserting on this assumption to be defensive in case 3714 * this ever changes. Without this check, it would be 3715 * possible to incorrectly increment arcstat_mutex_miss 3716 * below (e.g. if the code changed such that we called 3717 * this function with a hash lock held). 3718 */ 3719 ASSERT(!MUTEX_HELD(hash_lock)); 3720 3721 if (mutex_tryenter(hash_lock)) { 3722 uint64_t evicted = arc_evict_hdr(hdr, hash_lock); 3723 mutex_exit(hash_lock); 3724 3725 bytes_evicted += evicted; 3726 3727 /* 3728 * If evicted is zero, arc_evict_hdr() must have 3729 * decided to skip this header, don't increment 3730 * evict_count in this case. 3731 */ 3732 if (evicted != 0) 3733 evict_count++; 3734 3735 /* 3736 * If arc_size isn't overflowing, signal any 3737 * threads that might happen to be waiting. 3738 * 3739 * For each header evicted, we wake up a single 3740 * thread. If we used cv_broadcast, we could 3741 * wake up "too many" threads causing arc_size 3742 * to significantly overflow arc_c; since 3743 * arc_get_data_impl() doesn't check for overflow 3744 * when it's woken up (it doesn't because it's 3745 * possible for the ARC to be overflowing while 3746 * full of un-evictable buffers, and the 3747 * function should proceed in this case). 3748 * 3749 * If threads are left sleeping, due to not 3750 * using cv_broadcast, they will be woken up 3751 * just before arc_reclaim_thread() sleeps. 3752 */ 3753 mutex_enter(&arc_reclaim_lock); 3754 if (!arc_is_overflowing()) 3755 cv_signal(&arc_reclaim_waiters_cv); 3756 mutex_exit(&arc_reclaim_lock); 3757 } else { 3758 ARCSTAT_BUMP(arcstat_mutex_miss); 3759 } 3760 } 3761 3762 multilist_sublist_unlock(mls); 3763 3764 return (bytes_evicted); 3765} 3766 3767/* 3768 * Evict buffers from the given arc state, until we've removed the 3769 * specified number of bytes. Move the removed buffers to the 3770 * appropriate evict state. 3771 * 3772 * This function makes a "best effort". It skips over any buffers 3773 * it can't get a hash_lock on, and so, may not catch all candidates. 3774 * It may also return without evicting as much space as requested. 3775 * 3776 * If bytes is specified using the special value ARC_EVICT_ALL, this 3777 * will evict all available (i.e. unlocked and evictable) buffers from 3778 * the given arc state; which is used by arc_flush(). 3779 */ 3780static uint64_t 3781arc_evict_state(arc_state_t *state, uint64_t spa, int64_t bytes, 3782 arc_buf_contents_t type) 3783{ 3784 uint64_t total_evicted = 0; 3785 multilist_t *ml = state->arcs_list[type]; 3786 int num_sublists; 3787 arc_buf_hdr_t **markers; 3788 3789 IMPLY(bytes < 0, bytes == ARC_EVICT_ALL); 3790 3791 num_sublists = multilist_get_num_sublists(ml); 3792 3793 /* 3794 * If we've tried to evict from each sublist, made some 3795 * progress, but still have not hit the target number of bytes 3796 * to evict, we want to keep trying. The markers allow us to 3797 * pick up where we left off for each individual sublist, rather 3798 * than starting from the tail each time. 3799 */ 3800 markers = kmem_zalloc(sizeof (*markers) * num_sublists, KM_SLEEP); 3801 for (int i = 0; i < num_sublists; i++) { 3802 markers[i] = kmem_cache_alloc(hdr_full_cache, KM_SLEEP); 3803 3804 /* 3805 * A b_spa of 0 is used to indicate that this header is 3806 * a marker. This fact is used in arc_adjust_type() and 3807 * arc_evict_state_impl(). 3808 */ 3809 markers[i]->b_spa = 0; 3810 3811 multilist_sublist_t *mls = multilist_sublist_lock(ml, i); 3812 multilist_sublist_insert_tail(mls, markers[i]); 3813 multilist_sublist_unlock(mls); 3814 } 3815 3816 /* 3817 * While we haven't hit our target number of bytes to evict, or 3818 * we're evicting all available buffers. 3819 */ 3820 while (total_evicted < bytes || bytes == ARC_EVICT_ALL) { 3821 /* 3822 * Start eviction using a randomly selected sublist, 3823 * this is to try and evenly balance eviction across all 3824 * sublists. Always starting at the same sublist 3825 * (e.g. index 0) would cause evictions to favor certain 3826 * sublists over others. 3827 */ 3828 int sublist_idx = multilist_get_random_index(ml); 3829 uint64_t scan_evicted = 0; 3830 3831 for (int i = 0; i < num_sublists; i++) { 3832 uint64_t bytes_remaining; 3833 uint64_t bytes_evicted; 3834 3835 if (bytes == ARC_EVICT_ALL) 3836 bytes_remaining = ARC_EVICT_ALL; 3837 else if (total_evicted < bytes) 3838 bytes_remaining = bytes - total_evicted; 3839 else 3840 break; 3841 3842 bytes_evicted = arc_evict_state_impl(ml, sublist_idx, 3843 markers[sublist_idx], spa, bytes_remaining); 3844 3845 scan_evicted += bytes_evicted; 3846 total_evicted += bytes_evicted; 3847 3848 /* we've reached the end, wrap to the beginning */ 3849 if (++sublist_idx >= num_sublists) 3850 sublist_idx = 0; 3851 } 3852 3853 /* 3854 * If we didn't evict anything during this scan, we have 3855 * no reason to believe we'll evict more during another 3856 * scan, so break the loop. 3857 */ 3858 if (scan_evicted == 0) { 3859 /* This isn't possible, let's make that obvious */ 3860 ASSERT3S(bytes, !=, 0); 3861 3862 /* 3863 * When bytes is ARC_EVICT_ALL, the only way to 3864 * break the loop is when scan_evicted is zero. 3865 * In that case, we actually have evicted enough, 3866 * so we don't want to increment the kstat. 3867 */ 3868 if (bytes != ARC_EVICT_ALL) { 3869 ASSERT3S(total_evicted, <, bytes); 3870 ARCSTAT_BUMP(arcstat_evict_not_enough); 3871 } 3872 3873 break; 3874 } 3875 } 3876 3877 for (int i = 0; i < num_sublists; i++) { 3878 multilist_sublist_t *mls = multilist_sublist_lock(ml, i); 3879 multilist_sublist_remove(mls, markers[i]); 3880 multilist_sublist_unlock(mls); 3881 3882 kmem_cache_free(hdr_full_cache, markers[i]); 3883 } 3884 kmem_free(markers, sizeof (*markers) * num_sublists); 3885 3886 return (total_evicted); 3887} 3888 3889/* 3890 * Flush all "evictable" data of the given type from the arc state 3891 * specified. This will not evict any "active" buffers (i.e. referenced). 3892 * 3893 * When 'retry' is set to B_FALSE, the function will make a single pass 3894 * over the state and evict any buffers that it can. Since it doesn't 3895 * continually retry the eviction, it might end up leaving some buffers 3896 * in the ARC due to lock misses. 3897 * 3898 * When 'retry' is set to B_TRUE, the function will continually retry the 3899 * eviction until *all* evictable buffers have been removed from the 3900 * state. As a result, if concurrent insertions into the state are 3901 * allowed (e.g. if the ARC isn't shutting down), this function might 3902 * wind up in an infinite loop, continually trying to evict buffers. 3903 */ 3904static uint64_t 3905arc_flush_state(arc_state_t *state, uint64_t spa, arc_buf_contents_t type, 3906 boolean_t retry) 3907{ 3908 uint64_t evicted = 0; 3909 3910 while (refcount_count(&state->arcs_esize[type]) != 0) { 3911 evicted += arc_evict_state(state, spa, ARC_EVICT_ALL, type); 3912 3913 if (!retry) 3914 break; 3915 } 3916 3917 return (evicted); 3918} 3919 3920/* 3921 * Evict the specified number of bytes from the state specified, 3922 * restricting eviction to the spa and type given. This function 3923 * prevents us from trying to evict more from a state's list than 3924 * is "evictable", and to skip evicting altogether when passed a 3925 * negative value for "bytes". In contrast, arc_evict_state() will 3926 * evict everything it can, when passed a negative value for "bytes". 3927 */ 3928static uint64_t 3929arc_adjust_impl(arc_state_t *state, uint64_t spa, int64_t bytes, 3930 arc_buf_contents_t type) 3931{ 3932 int64_t delta; 3933 3934 if (bytes > 0 && refcount_count(&state->arcs_esize[type]) > 0) { 3935 delta = MIN(refcount_count(&state->arcs_esize[type]), bytes); 3936 return (arc_evict_state(state, spa, delta, type)); 3937 } 3938 3939 return (0); 3940} 3941 3942/* 3943 * Evict metadata buffers from the cache, such that arc_meta_used is 3944 * capped by the arc_meta_limit tunable. 3945 */ 3946static uint64_t 3947arc_adjust_meta(uint64_t meta_used) 3948{ 3949 uint64_t total_evicted = 0; 3950 int64_t target; 3951 3952 /* 3953 * If we're over the meta limit, we want to evict enough 3954 * metadata to get back under the meta limit. We don't want to 3955 * evict so much that we drop the MRU below arc_p, though. If 3956 * we're over the meta limit more than we're over arc_p, we 3957 * evict some from the MRU here, and some from the MFU below. 3958 */ 3959 target = MIN((int64_t)(meta_used - arc_meta_limit), 3960 (int64_t)(refcount_count(&arc_anon->arcs_size) + 3961 refcount_count(&arc_mru->arcs_size) - arc_p)); 3962 3963 total_evicted += arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA); 3964 3965 /* 3966 * Similar to the above, we want to evict enough bytes to get us 3967 * below the meta limit, but not so much as to drop us below the 3968 * space allotted to the MFU (which is defined as arc_c - arc_p). 3969 */ 3970 target = MIN((int64_t)(meta_used - arc_meta_limit), 3971 (int64_t)(refcount_count(&arc_mfu->arcs_size) - 3972 (arc_c - arc_p))); 3973 3974 total_evicted += arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA); 3975 3976 return (total_evicted); 3977} 3978 3979/* 3980 * Return the type of the oldest buffer in the given arc state 3981 * 3982 * This function will select a random sublist of type ARC_BUFC_DATA and 3983 * a random sublist of type ARC_BUFC_METADATA. The tail of each sublist 3984 * is compared, and the type which contains the "older" buffer will be 3985 * returned. 3986 */ 3987static arc_buf_contents_t 3988arc_adjust_type(arc_state_t *state) 3989{ 3990 multilist_t *data_ml = state->arcs_list[ARC_BUFC_DATA]; 3991 multilist_t *meta_ml = state->arcs_list[ARC_BUFC_METADATA]; 3992 int data_idx = multilist_get_random_index(data_ml); 3993 int meta_idx = multilist_get_random_index(meta_ml); 3994 multilist_sublist_t *data_mls; 3995 multilist_sublist_t *meta_mls; 3996 arc_buf_contents_t type; 3997 arc_buf_hdr_t *data_hdr; 3998 arc_buf_hdr_t *meta_hdr; 3999 4000 /* 4001 * We keep the sublist lock until we're finished, to prevent 4002 * the headers from being destroyed via arc_evict_state(). 4003 */ 4004 data_mls = multilist_sublist_lock(data_ml, data_idx); 4005 meta_mls = multilist_sublist_lock(meta_ml, meta_idx); 4006 4007 /* 4008 * These two loops are to ensure we skip any markers that 4009 * might be at the tail of the lists due to arc_evict_state(). 4010 */ 4011 4012 for (data_hdr = multilist_sublist_tail(data_mls); data_hdr != NULL; 4013 data_hdr = multilist_sublist_prev(data_mls, data_hdr)) { 4014 if (data_hdr->b_spa != 0) 4015 break; 4016 } 4017 4018 for (meta_hdr = multilist_sublist_tail(meta_mls); meta_hdr != NULL; 4019 meta_hdr = multilist_sublist_prev(meta_mls, meta_hdr)) { 4020 if (meta_hdr->b_spa != 0) 4021 break; 4022 } 4023 4024 if (data_hdr == NULL && meta_hdr == NULL) { 4025 type = ARC_BUFC_DATA; 4026 } else if (data_hdr == NULL) { 4027 ASSERT3P(meta_hdr, !=, NULL); 4028 type = ARC_BUFC_METADATA; 4029 } else if (meta_hdr == NULL) { 4030 ASSERT3P(data_hdr, !=, NULL); 4031 type = ARC_BUFC_DATA; 4032 } else { 4033 ASSERT3P(data_hdr, !=, NULL); 4034 ASSERT3P(meta_hdr, !=, NULL); 4035 4036 /* The headers can't be on the sublist without an L1 header */ 4037 ASSERT(HDR_HAS_L1HDR(data_hdr)); 4038 ASSERT(HDR_HAS_L1HDR(meta_hdr)); 4039 4040 if (data_hdr->b_l1hdr.b_arc_access < 4041 meta_hdr->b_l1hdr.b_arc_access) { 4042 type = ARC_BUFC_DATA; 4043 } else { 4044 type = ARC_BUFC_METADATA; 4045 } 4046 } 4047 4048 multilist_sublist_unlock(meta_mls); 4049 multilist_sublist_unlock(data_mls); 4050 4051 return (type); 4052} 4053 4054/* 4055 * Evict buffers from the cache, such that arc_size is capped by arc_c. 4056 */ 4057static uint64_t 4058arc_adjust(void) 4059{ 4060 uint64_t total_evicted = 0; 4061 uint64_t bytes; 4062 int64_t target; 4063 uint64_t asize = aggsum_value(&arc_size); 4064 uint64_t ameta = aggsum_value(&arc_meta_used); 4065 4066 /* 4067 * If we're over arc_meta_limit, we want to correct that before 4068 * potentially evicting data buffers below. 4069 */ 4070 total_evicted += arc_adjust_meta(ameta); 4071 4072 /* 4073 * Adjust MRU size 4074 * 4075 * If we're over the target cache size, we want to evict enough 4076 * from the list to get back to our target size. We don't want 4077 * to evict too much from the MRU, such that it drops below 4078 * arc_p. So, if we're over our target cache size more than 4079 * the MRU is over arc_p, we'll evict enough to get back to 4080 * arc_p here, and then evict more from the MFU below. 4081 */ 4082 target = MIN((int64_t)(asize - arc_c), 4083 (int64_t)(refcount_count(&arc_anon->arcs_size) + 4084 refcount_count(&arc_mru->arcs_size) + ameta - arc_p)); 4085 4086 /* 4087 * If we're below arc_meta_min, always prefer to evict data. 4088 * Otherwise, try to satisfy the requested number of bytes to 4089 * evict from the type which contains older buffers; in an 4090 * effort to keep newer buffers in the cache regardless of their 4091 * type. If we cannot satisfy the number of bytes from this 4092 * type, spill over into the next type. 4093 */ 4094 if (arc_adjust_type(arc_mru) == ARC_BUFC_METADATA && 4095 ameta > arc_meta_min) { 4096 bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA); 4097 total_evicted += bytes; 4098 4099 /* 4100 * If we couldn't evict our target number of bytes from 4101 * metadata, we try to get the rest from data. 4102 */ 4103 target -= bytes; 4104 4105 total_evicted += 4106 arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA); 4107 } else { 4108 bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA); 4109 total_evicted += bytes; 4110 4111 /* 4112 * If we couldn't evict our target number of bytes from 4113 * data, we try to get the rest from metadata. 4114 */ 4115 target -= bytes; 4116 4117 total_evicted += 4118 arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA); 4119 } 4120 4121 /* 4122 * Re-sum ARC stats after the first round of evictions. 4123 */ 4124 asize = aggsum_value(&arc_size); 4125 ameta = aggsum_value(&arc_meta_used); 4126 4127 /* 4128 * Adjust MFU size 4129 * 4130 * Now that we've tried to evict enough from the MRU to get its 4131 * size back to arc_p, if we're still above the target cache 4132 * size, we evict the rest from the MFU. 4133 */ 4134 target = asize - arc_c; 4135 4136 if (arc_adjust_type(arc_mfu) == ARC_BUFC_METADATA && 4137 ameta > arc_meta_min) { 4138 bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA); 4139 total_evicted += bytes; 4140 4141 /* 4142 * If we couldn't evict our target number of bytes from 4143 * metadata, we try to get the rest from data. 4144 */ 4145 target -= bytes; 4146 4147 total_evicted += 4148 arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA); 4149 } else { 4150 bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA); 4151 total_evicted += bytes; 4152 4153 /* 4154 * If we couldn't evict our target number of bytes from 4155 * data, we try to get the rest from data. 4156 */ 4157 target -= bytes; 4158 4159 total_evicted += 4160 arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA); 4161 } 4162 4163 /* 4164 * Adjust ghost lists 4165 * 4166 * In addition to the above, the ARC also defines target values 4167 * for the ghost lists. The sum of the mru list and mru ghost 4168 * list should never exceed the target size of the cache, and 4169 * the sum of the mru list, mfu list, mru ghost list, and mfu 4170 * ghost list should never exceed twice the target size of the 4171 * cache. The following logic enforces these limits on the ghost 4172 * caches, and evicts from them as needed. 4173 */ 4174 target = refcount_count(&arc_mru->arcs_size) + 4175 refcount_count(&arc_mru_ghost->arcs_size) - arc_c; 4176 4177 bytes = arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_DATA); 4178 total_evicted += bytes; 4179 4180 target -= bytes; 4181 4182 total_evicted += 4183 arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_METADATA); 4184 4185 /* 4186 * We assume the sum of the mru list and mfu list is less than 4187 * or equal to arc_c (we enforced this above), which means we 4188 * can use the simpler of the two equations below: 4189 * 4190 * mru + mfu + mru ghost + mfu ghost <= 2 * arc_c 4191 * mru ghost + mfu ghost <= arc_c 4192 */ 4193 target = refcount_count(&arc_mru_ghost->arcs_size) + 4194 refcount_count(&arc_mfu_ghost->arcs_size) - arc_c; 4195 4196 bytes = arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_DATA); 4197 total_evicted += bytes; 4198 4199 target -= bytes; 4200 4201 total_evicted += 4202 arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_METADATA); 4203 4204 return (total_evicted); 4205} 4206 4207void 4208arc_flush(spa_t *spa, boolean_t retry) 4209{ 4210 uint64_t guid = 0; 4211 4212 /* 4213 * If retry is B_TRUE, a spa must not be specified since we have 4214 * no good way to determine if all of a spa's buffers have been 4215 * evicted from an arc state. 4216 */ 4217 ASSERT(!retry || spa == 0); 4218 4219 if (spa != NULL) 4220 guid = spa_load_guid(spa); 4221 4222 (void) arc_flush_state(arc_mru, guid, ARC_BUFC_DATA, retry); 4223 (void) arc_flush_state(arc_mru, guid, ARC_BUFC_METADATA, retry); 4224 4225 (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_DATA, retry); 4226 (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_METADATA, retry); 4227 4228 (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_DATA, retry); 4229 (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_METADATA, retry); 4230 4231 (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_DATA, retry); 4232 (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_METADATA, retry); 4233} 4234 4235void 4236arc_shrink(int64_t to_free) 4237{ 4238 uint64_t asize = aggsum_value(&arc_size); 4239 if (arc_c > arc_c_min) { 4240 DTRACE_PROBE4(arc__shrink, uint64_t, arc_c, uint64_t, 4241 arc_c_min, uint64_t, arc_p, uint64_t, to_free); 4242 if (arc_c > arc_c_min + to_free) 4243 atomic_add_64(&arc_c, -to_free); 4244 else 4245 arc_c = arc_c_min; 4246 4247 atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift)); 4248 if (asize < arc_c) 4249 arc_c = MAX(asize, arc_c_min); 4250 if (arc_p > arc_c) 4251 arc_p = (arc_c >> 1); 4252 4253 DTRACE_PROBE2(arc__shrunk, uint64_t, arc_c, uint64_t, 4254 arc_p); 4255 4256 ASSERT(arc_c >= arc_c_min); 4257 ASSERT((int64_t)arc_p >= 0); 4258 } 4259 4260 if (asize > arc_c) { 4261 DTRACE_PROBE2(arc__shrink_adjust, uint64_t, asize, 4262 uint64_t, arc_c); 4263 (void) arc_adjust(); 4264 } 4265} 4266 4267typedef enum free_memory_reason_t { 4268 FMR_UNKNOWN, 4269 FMR_NEEDFREE, 4270 FMR_LOTSFREE, 4271 FMR_SWAPFS_MINFREE, 4272 FMR_PAGES_PP_MAXIMUM, 4273 FMR_HEAP_ARENA, 4274 FMR_ZIO_ARENA, 4275 FMR_ZIO_FRAG, 4276} free_memory_reason_t; 4277 4278int64_t last_free_memory; 4279free_memory_reason_t last_free_reason; 4280 4281/* 4282 * Additional reserve of pages for pp_reserve. 4283 */ 4284int64_t arc_pages_pp_reserve = 64; 4285 4286/* 4287 * Additional reserve of pages for swapfs. 4288 */ 4289int64_t arc_swapfs_reserve = 64; 4290 4291/* 4292 * Return the amount of memory that can be consumed before reclaim will be 4293 * needed. Positive if there is sufficient free memory, negative indicates 4294 * the amount of memory that needs to be freed up. 4295 */ 4296static int64_t 4297arc_available_memory(void) 4298{ 4299 int64_t lowest = INT64_MAX; 4300 int64_t n; 4301 free_memory_reason_t r = FMR_UNKNOWN; 4302 4303#ifdef _KERNEL 4304#ifdef __FreeBSD__ 4305 /* 4306 * Cooperate with pagedaemon when it's time for it to scan 4307 * and reclaim some pages. 4308 */ 4309 n = PAGESIZE * ((int64_t)freemem - zfs_arc_free_target); 4310 if (n < lowest) { 4311 lowest = n; 4312 r = FMR_LOTSFREE; 4313 } 4314 4315#else 4316 if (needfree > 0) { 4317 n = PAGESIZE * (-needfree); 4318 if (n < lowest) { 4319 lowest = n; 4320 r = FMR_NEEDFREE; 4321 } 4322 } 4323 4324 /* 4325 * check that we're out of range of the pageout scanner. It starts to 4326 * schedule paging if freemem is less than lotsfree and needfree. 4327 * lotsfree is the high-water mark for pageout, and needfree is the 4328 * number of needed free pages. We add extra pages here to make sure 4329 * the scanner doesn't start up while we're freeing memory. 4330 */ 4331 n = PAGESIZE * (freemem - lotsfree - needfree - desfree); 4332 if (n < lowest) { 4333 lowest = n; 4334 r = FMR_LOTSFREE; 4335 } 4336 4337 /* 4338 * check to make sure that swapfs has enough space so that anon 4339 * reservations can still succeed. anon_resvmem() checks that the 4340 * availrmem is greater than swapfs_minfree, and the number of reserved 4341 * swap pages. We also add a bit of extra here just to prevent 4342 * circumstances from getting really dire. 4343 */ 4344 n = PAGESIZE * (availrmem - swapfs_minfree - swapfs_reserve - 4345 desfree - arc_swapfs_reserve); 4346 if (n < lowest) { 4347 lowest = n; 4348 r = FMR_SWAPFS_MINFREE; 4349 } 4350 4351 4352 /* 4353 * Check that we have enough availrmem that memory locking (e.g., via 4354 * mlock(3C) or memcntl(2)) can still succeed. (pages_pp_maximum 4355 * stores the number of pages that cannot be locked; when availrmem 4356 * drops below pages_pp_maximum, page locking mechanisms such as 4357 * page_pp_lock() will fail.) 4358 */ 4359 n = PAGESIZE * (availrmem - pages_pp_maximum - 4360 arc_pages_pp_reserve); 4361 if (n < lowest) { 4362 lowest = n; 4363 r = FMR_PAGES_PP_MAXIMUM; 4364 } 4365 4366#endif /* __FreeBSD__ */ 4367#if defined(__i386) || !defined(UMA_MD_SMALL_ALLOC) 4368 /* 4369 * If we're on an i386 platform, it's possible that we'll exhaust the 4370 * kernel heap space before we ever run out of available physical 4371 * memory. Most checks of the size of the heap_area compare against 4372 * tune.t_minarmem, which is the minimum available real memory that we 4373 * can have in the system. However, this is generally fixed at 25 pages 4374 * which is so low that it's useless. In this comparison, we seek to 4375 * calculate the total heap-size, and reclaim if more than 3/4ths of the 4376 * heap is allocated. (Or, in the calculation, if less than 1/4th is 4377 * free) 4378 */ 4379 n = (int64_t)vmem_size(heap_arena, VMEM_FREE) - 4380 (vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC) >> 2); 4381 if (n < lowest) { 4382 lowest = n; 4383 r = FMR_HEAP_ARENA; 4384 } 4385#define zio_arena NULL 4386#else 4387#define zio_arena heap_arena 4388#endif 4389 4390 /* 4391 * If zio data pages are being allocated out of a separate heap segment, 4392 * then enforce that the size of available vmem for this arena remains 4393 * above about 1/4th (1/(2^arc_zio_arena_free_shift)) free. 4394 * 4395 * Note that reducing the arc_zio_arena_free_shift keeps more virtual 4396 * memory (in the zio_arena) free, which can avoid memory 4397 * fragmentation issues. 4398 */ 4399 if (zio_arena != NULL) { 4400 n = (int64_t)vmem_size(zio_arena, VMEM_FREE) - 4401 (vmem_size(zio_arena, VMEM_ALLOC) >> 4402 arc_zio_arena_free_shift); 4403 if (n < lowest) { 4404 lowest = n; 4405 r = FMR_ZIO_ARENA; 4406 } 4407 } 4408 4409 /* 4410 * Above limits know nothing about real level of KVA fragmentation. 4411 * Start aggressive reclamation if too little sequential KVA left. 4412 */ 4413 if (lowest > 0) { 4414 n = (vmem_size(heap_arena, VMEM_MAXFREE) < SPA_MAXBLOCKSIZE) ? 4415 -((int64_t)vmem_size(heap_arena, VMEM_ALLOC) >> 4) : 4416 INT64_MAX; 4417 if (n < lowest) { 4418 lowest = n; 4419 r = FMR_ZIO_FRAG; 4420 } 4421 } 4422 4423#else /* _KERNEL */ 4424 /* Every 100 calls, free a small amount */ 4425 if (spa_get_random(100) == 0) 4426 lowest = -1024; 4427#endif /* _KERNEL */ 4428 4429 last_free_memory = lowest; 4430 last_free_reason = r; 4431 DTRACE_PROBE2(arc__available_memory, int64_t, lowest, int, r); 4432 return (lowest); 4433} 4434 4435 4436/* 4437 * Determine if the system is under memory pressure and is asking 4438 * to reclaim memory. A return value of B_TRUE indicates that the system 4439 * is under memory pressure and that the arc should adjust accordingly. 4440 */ 4441static boolean_t 4442arc_reclaim_needed(void) 4443{ 4444 return (arc_available_memory() < 0); 4445} 4446 4447extern kmem_cache_t *zio_buf_cache[]; 4448extern kmem_cache_t *zio_data_buf_cache[]; 4449extern kmem_cache_t *range_seg_cache; 4450extern kmem_cache_t *abd_chunk_cache; 4451 4452static __noinline void 4453arc_kmem_reap_now(void) 4454{ 4455 size_t i; 4456 kmem_cache_t *prev_cache = NULL; 4457 kmem_cache_t *prev_data_cache = NULL; 4458 4459 DTRACE_PROBE(arc__kmem_reap_start); 4460#ifdef _KERNEL 4461 if (aggsum_compare(&arc_meta_used, arc_meta_limit) >= 0) { 4462 /* 4463 * We are exceeding our meta-data cache limit. 4464 * Purge some DNLC entries to release holds on meta-data. 4465 */ 4466 dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent); 4467 } 4468#if defined(__i386) 4469 /* 4470 * Reclaim unused memory from all kmem caches. 4471 */ 4472 kmem_reap(); 4473#endif 4474#endif 4475 4476 /* 4477 * If a kmem reap is already active, don't schedule more. We must 4478 * check for this because kmem_cache_reap_soon() won't actually 4479 * block on the cache being reaped (this is to prevent callers from 4480 * becoming implicitly blocked by a system-wide kmem reap -- which, 4481 * on a system with many, many full magazines, can take minutes). 4482 */ 4483 if (kmem_cache_reap_active()) 4484 return; 4485 4486 for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) { 4487 if (zio_buf_cache[i] != prev_cache) { 4488 prev_cache = zio_buf_cache[i]; 4489 kmem_cache_reap_soon(zio_buf_cache[i]); 4490 } 4491 if (zio_data_buf_cache[i] != prev_data_cache) { 4492 prev_data_cache = zio_data_buf_cache[i]; 4493 kmem_cache_reap_soon(zio_data_buf_cache[i]); 4494 } 4495 } 4496 kmem_cache_reap_soon(abd_chunk_cache); 4497 kmem_cache_reap_soon(buf_cache); 4498 kmem_cache_reap_soon(hdr_full_cache); 4499 kmem_cache_reap_soon(hdr_l2only_cache); 4500 kmem_cache_reap_soon(range_seg_cache); 4501 4502#ifdef illumos 4503 if (zio_arena != NULL) { 4504 /* 4505 * Ask the vmem arena to reclaim unused memory from its 4506 * quantum caches. 4507 */ 4508 vmem_qcache_reap(zio_arena); 4509 } 4510#endif 4511 DTRACE_PROBE(arc__kmem_reap_end); 4512} 4513 4514/* 4515 * Threads can block in arc_get_data_impl() waiting for this thread to evict 4516 * enough data and signal them to proceed. When this happens, the threads in 4517 * arc_get_data_impl() are sleeping while holding the hash lock for their 4518 * particular arc header. Thus, we must be careful to never sleep on a 4519 * hash lock in this thread. This is to prevent the following deadlock: 4520 * 4521 * - Thread A sleeps on CV in arc_get_data_impl() holding hash lock "L", 4522 * waiting for the reclaim thread to signal it. 4523 * 4524 * - arc_reclaim_thread() tries to acquire hash lock "L" using mutex_enter, 4525 * fails, and goes to sleep forever. 4526 * 4527 * This possible deadlock is avoided by always acquiring a hash lock 4528 * using mutex_tryenter() from arc_reclaim_thread(). 4529 */ 4530/* ARGSUSED */ 4531static void 4532arc_reclaim_thread(void *unused __unused) 4533{ 4534 hrtime_t growtime = 0; 4535 hrtime_t kmem_reap_time = 0; 4536 callb_cpr_t cpr; 4537 4538 CALLB_CPR_INIT(&cpr, &arc_reclaim_lock, callb_generic_cpr, FTAG); 4539 4540 mutex_enter(&arc_reclaim_lock); 4541 while (!arc_reclaim_thread_exit) { 4542 uint64_t evicted = 0; 4543 4544 /* 4545 * This is necessary in order for the mdb ::arc dcmd to 4546 * show up to date information. Since the ::arc command 4547 * does not call the kstat's update function, without 4548 * this call, the command may show stale stats for the 4549 * anon, mru, mru_ghost, mfu, and mfu_ghost lists. Even 4550 * with this change, the data might be up to 1 second 4551 * out of date; but that should suffice. The arc_state_t 4552 * structures can be queried directly if more accurate 4553 * information is needed. 4554 */ 4555 if (arc_ksp != NULL) 4556 arc_ksp->ks_update(arc_ksp, KSTAT_READ); 4557 4558 mutex_exit(&arc_reclaim_lock); 4559 4560 /* 4561 * We call arc_adjust() before (possibly) calling 4562 * arc_kmem_reap_now(), so that we can wake up 4563 * arc_get_data_impl() sooner. 4564 */ 4565 evicted = arc_adjust(); 4566 4567 int64_t free_memory = arc_available_memory(); 4568 if (free_memory < 0) { 4569 hrtime_t curtime = gethrtime(); 4570 arc_no_grow = B_TRUE; 4571 arc_warm = B_TRUE; 4572 4573 /* 4574 * Wait at least zfs_grow_retry (default 60) seconds 4575 * before considering growing. 4576 */ 4577 growtime = curtime + SEC2NSEC(arc_grow_retry); 4578 4579 /* 4580 * Wait at least arc_kmem_cache_reap_retry_ms 4581 * between arc_kmem_reap_now() calls. Without 4582 * this check it is possible to end up in a 4583 * situation where we spend lots of time 4584 * reaping caches, while we're near arc_c_min. 4585 */ 4586 if (curtime >= kmem_reap_time) { 4587 arc_kmem_reap_now(); 4588 kmem_reap_time = gethrtime() + 4589 MSEC2NSEC(arc_kmem_cache_reap_retry_ms); 4590 } 4591 4592 /* 4593 * If we are still low on memory, shrink the ARC 4594 * so that we have arc_shrink_min free space. 4595 */ 4596 free_memory = arc_available_memory(); 4597 4598 int64_t to_free = 4599 (arc_c >> arc_shrink_shift) - free_memory; 4600 if (to_free > 0) { 4601#ifdef _KERNEL 4602#ifdef illumos 4603 to_free = MAX(to_free, ptob(needfree)); 4604#endif 4605#endif 4606 arc_shrink(to_free); 4607 } 4608 } else if (free_memory < arc_c >> arc_no_grow_shift) { 4609 arc_no_grow = B_TRUE; 4610 } else if (gethrtime() >= growtime) { 4611 arc_no_grow = B_FALSE; 4612 } 4613 4614 mutex_enter(&arc_reclaim_lock); 4615 4616 /* 4617 * If evicted is zero, we couldn't evict anything via 4618 * arc_adjust(). This could be due to hash lock 4619 * collisions, but more likely due to the majority of 4620 * arc buffers being unevictable. Therefore, even if 4621 * arc_size is above arc_c, another pass is unlikely to 4622 * be helpful and could potentially cause us to enter an 4623 * infinite loop. 4624 */ 4625 if (aggsum_compare(&arc_size, arc_c) <= 0|| evicted == 0) { 4626 /* 4627 * We're either no longer overflowing, or we 4628 * can't evict anything more, so we should wake 4629 * up any threads before we go to sleep. 4630 */ 4631 cv_broadcast(&arc_reclaim_waiters_cv); 4632 4633 /* 4634 * Block until signaled, or after one second (we 4635 * might need to perform arc_kmem_reap_now() 4636 * even if we aren't being signalled) 4637 */ 4638 CALLB_CPR_SAFE_BEGIN(&cpr); 4639 (void) cv_timedwait_hires(&arc_reclaim_thread_cv, 4640 &arc_reclaim_lock, SEC2NSEC(1), MSEC2NSEC(1), 0); 4641 CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_lock); 4642 } 4643 } 4644 4645 arc_reclaim_thread_exit = B_FALSE; 4646 cv_broadcast(&arc_reclaim_thread_cv); 4647 CALLB_CPR_EXIT(&cpr); /* drops arc_reclaim_lock */ 4648 thread_exit(); 4649} 4650 4651static u_int arc_dnlc_evicts_arg; 4652extern struct vfsops zfs_vfsops; 4653 4654static void 4655arc_dnlc_evicts_thread(void *dummy __unused) 4656{ 4657 callb_cpr_t cpr; 4658 u_int percent; 4659 4660 CALLB_CPR_INIT(&cpr, &arc_dnlc_evicts_lock, callb_generic_cpr, FTAG); 4661 4662 mutex_enter(&arc_dnlc_evicts_lock); 4663 while (!arc_dnlc_evicts_thread_exit) { 4664 CALLB_CPR_SAFE_BEGIN(&cpr); 4665 (void) cv_wait(&arc_dnlc_evicts_cv, &arc_dnlc_evicts_lock); 4666 CALLB_CPR_SAFE_END(&cpr, &arc_dnlc_evicts_lock); 4667 if (arc_dnlc_evicts_arg != 0) { 4668 percent = arc_dnlc_evicts_arg; 4669 mutex_exit(&arc_dnlc_evicts_lock); 4670#ifdef _KERNEL 4671 vnlru_free(desiredvnodes * percent / 100, &zfs_vfsops); 4672#endif 4673 mutex_enter(&arc_dnlc_evicts_lock); 4674 /* 4675 * Clear our token only after vnlru_free() 4676 * pass is done, to avoid false queueing of 4677 * the requests. 4678 */ 4679 arc_dnlc_evicts_arg = 0; 4680 } 4681 } 4682 arc_dnlc_evicts_thread_exit = FALSE; 4683 cv_broadcast(&arc_dnlc_evicts_cv); 4684 CALLB_CPR_EXIT(&cpr); 4685 thread_exit(); 4686} 4687 4688void 4689dnlc_reduce_cache(void *arg) 4690{ 4691 u_int percent; 4692 4693 percent = (u_int)(uintptr_t)arg; 4694 mutex_enter(&arc_dnlc_evicts_lock); 4695 if (arc_dnlc_evicts_arg == 0) { 4696 arc_dnlc_evicts_arg = percent; 4697 cv_broadcast(&arc_dnlc_evicts_cv); 4698 } 4699 mutex_exit(&arc_dnlc_evicts_lock); 4700} 4701 4702/* 4703 * Adapt arc info given the number of bytes we are trying to add and 4704 * the state that we are comming from. This function is only called 4705 * when we are adding new content to the cache. 4706 */ 4707static void 4708arc_adapt(int bytes, arc_state_t *state) 4709{ 4710 int mult; 4711 uint64_t arc_p_min = (arc_c >> arc_p_min_shift); 4712 int64_t mrug_size = refcount_count(&arc_mru_ghost->arcs_size); 4713 int64_t mfug_size = refcount_count(&arc_mfu_ghost->arcs_size); 4714 4715 if (state == arc_l2c_only) 4716 return; 4717 4718 ASSERT(bytes > 0); 4719 /* 4720 * Adapt the target size of the MRU list: 4721 * - if we just hit in the MRU ghost list, then increase 4722 * the target size of the MRU list. 4723 * - if we just hit in the MFU ghost list, then increase 4724 * the target size of the MFU list by decreasing the 4725 * target size of the MRU list. 4726 */ 4727 if (state == arc_mru_ghost) { 4728 mult = (mrug_size >= mfug_size) ? 1 : (mfug_size / mrug_size); 4729 mult = MIN(mult, 10); /* avoid wild arc_p adjustment */ 4730 4731 arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult); 4732 } else if (state == arc_mfu_ghost) { 4733 uint64_t delta; 4734 4735 mult = (mfug_size >= mrug_size) ? 1 : (mrug_size / mfug_size); 4736 mult = MIN(mult, 10); 4737 4738 delta = MIN(bytes * mult, arc_p); 4739 arc_p = MAX(arc_p_min, arc_p - delta); 4740 } 4741 ASSERT((int64_t)arc_p >= 0); 4742 4743 if (arc_reclaim_needed()) { 4744 cv_signal(&arc_reclaim_thread_cv); 4745 return; 4746 } 4747 4748 if (arc_no_grow) 4749 return; 4750 4751 if (arc_c >= arc_c_max) 4752 return; 4753 4754 /* 4755 * If we're within (2 * maxblocksize) bytes of the target 4756 * cache size, increment the target cache size 4757 */ 4758 if (aggsum_compare(&arc_size, arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) > 4759 0) { 4760 DTRACE_PROBE1(arc__inc_adapt, int, bytes); 4761 atomic_add_64(&arc_c, (int64_t)bytes); 4762 if (arc_c > arc_c_max) 4763 arc_c = arc_c_max; 4764 else if (state == arc_anon) 4765 atomic_add_64(&arc_p, (int64_t)bytes); 4766 if (arc_p > arc_c) 4767 arc_p = arc_c; 4768 } 4769 ASSERT((int64_t)arc_p >= 0); 4770} 4771 4772/* 4773 * Check if arc_size has grown past our upper threshold, determined by 4774 * zfs_arc_overflow_shift. 4775 */ 4776static boolean_t 4777arc_is_overflowing(void) 4778{ 4779 /* Always allow at least one block of overflow */ 4780 uint64_t overflow = MAX(SPA_MAXBLOCKSIZE, 4781 arc_c >> zfs_arc_overflow_shift); 4782 4783 /* 4784 * We just compare the lower bound here for performance reasons. Our 4785 * primary goals are to make sure that the arc never grows without 4786 * bound, and that it can reach its maximum size. This check 4787 * accomplishes both goals. The maximum amount we could run over by is 4788 * 2 * aggsum_borrow_multiplier * NUM_CPUS * the average size of a block 4789 * in the ARC. In practice, that's in the tens of MB, which is low 4790 * enough to be safe. 4791 */ 4792 return (aggsum_lower_bound(&arc_size) >= arc_c + overflow); 4793} 4794 4795static abd_t * 4796arc_get_data_abd(arc_buf_hdr_t *hdr, uint64_t size, void *tag) 4797{ 4798 arc_buf_contents_t type = arc_buf_type(hdr); 4799 4800 arc_get_data_impl(hdr, size, tag); 4801 if (type == ARC_BUFC_METADATA) { 4802 return (abd_alloc(size, B_TRUE)); 4803 } else { 4804 ASSERT(type == ARC_BUFC_DATA); 4805 return (abd_alloc(size, B_FALSE)); 4806 } 4807} 4808 4809static void * 4810arc_get_data_buf(arc_buf_hdr_t *hdr, uint64_t size, void *tag) 4811{ 4812 arc_buf_contents_t type = arc_buf_type(hdr); 4813 4814 arc_get_data_impl(hdr, size, tag); 4815 if (type == ARC_BUFC_METADATA) { 4816 return (zio_buf_alloc(size)); 4817 } else { 4818 ASSERT(type == ARC_BUFC_DATA); 4819 return (zio_data_buf_alloc(size)); 4820 } 4821} 4822 4823/* 4824 * Allocate a block and return it to the caller. If we are hitting the 4825 * hard limit for the cache size, we must sleep, waiting for the eviction 4826 * thread to catch up. If we're past the target size but below the hard 4827 * limit, we'll only signal the reclaim thread and continue on. 4828 */ 4829static void 4830arc_get_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag) 4831{ 4832 arc_state_t *state = hdr->b_l1hdr.b_state; 4833 arc_buf_contents_t type = arc_buf_type(hdr); 4834 4835 arc_adapt(size, state); 4836 4837 /* 4838 * If arc_size is currently overflowing, and has grown past our 4839 * upper limit, we must be adding data faster than the evict 4840 * thread can evict. Thus, to ensure we don't compound the 4841 * problem by adding more data and forcing arc_size to grow even 4842 * further past it's target size, we halt and wait for the 4843 * eviction thread to catch up. 4844 * 4845 * It's also possible that the reclaim thread is unable to evict 4846 * enough buffers to get arc_size below the overflow limit (e.g. 4847 * due to buffers being un-evictable, or hash lock collisions). 4848 * In this case, we want to proceed regardless if we're 4849 * overflowing; thus we don't use a while loop here. 4850 */ 4851 if (arc_is_overflowing()) { 4852 mutex_enter(&arc_reclaim_lock); 4853 4854 /* 4855 * Now that we've acquired the lock, we may no longer be 4856 * over the overflow limit, lets check. 4857 * 4858 * We're ignoring the case of spurious wake ups. If that 4859 * were to happen, it'd let this thread consume an ARC 4860 * buffer before it should have (i.e. before we're under 4861 * the overflow limit and were signalled by the reclaim 4862 * thread). As long as that is a rare occurrence, it 4863 * shouldn't cause any harm. 4864 */ 4865 if (arc_is_overflowing()) { 4866 cv_signal(&arc_reclaim_thread_cv); 4867 cv_wait(&arc_reclaim_waiters_cv, &arc_reclaim_lock); 4868 } 4869 4870 mutex_exit(&arc_reclaim_lock); 4871 } 4872 4873 VERIFY3U(hdr->b_type, ==, type); 4874 if (type == ARC_BUFC_METADATA) { 4875 arc_space_consume(size, ARC_SPACE_META); 4876 } else { 4877 arc_space_consume(size, ARC_SPACE_DATA); 4878 } 4879 4880 /* 4881 * Update the state size. Note that ghost states have a 4882 * "ghost size" and so don't need to be updated. 4883 */ 4884 if (!GHOST_STATE(state)) { 4885 4886 (void) refcount_add_many(&state->arcs_size, size, tag); 4887 4888 /* 4889 * If this is reached via arc_read, the link is 4890 * protected by the hash lock. If reached via 4891 * arc_buf_alloc, the header should not be accessed by 4892 * any other thread. And, if reached via arc_read_done, 4893 * the hash lock will protect it if it's found in the 4894 * hash table; otherwise no other thread should be 4895 * trying to [add|remove]_reference it. 4896 */ 4897 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) { 4898 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 4899 (void) refcount_add_many(&state->arcs_esize[type], 4900 size, tag); 4901 } 4902 4903 /* 4904 * If we are growing the cache, and we are adding anonymous 4905 * data, and we have outgrown arc_p, update arc_p 4906 */ 4907 if (aggsum_compare(&arc_size, arc_c) < 0 && 4908 hdr->b_l1hdr.b_state == arc_anon && 4909 (refcount_count(&arc_anon->arcs_size) + 4910 refcount_count(&arc_mru->arcs_size) > arc_p)) 4911 arc_p = MIN(arc_c, arc_p + size); 4912 } 4913 ARCSTAT_BUMP(arcstat_allocated); 4914} 4915 4916static void 4917arc_free_data_abd(arc_buf_hdr_t *hdr, abd_t *abd, uint64_t size, void *tag) 4918{ 4919 arc_free_data_impl(hdr, size, tag); 4920 abd_free(abd); 4921} 4922 4923static void 4924arc_free_data_buf(arc_buf_hdr_t *hdr, void *buf, uint64_t size, void *tag) 4925{ 4926 arc_buf_contents_t type = arc_buf_type(hdr); 4927 4928 arc_free_data_impl(hdr, size, tag); 4929 if (type == ARC_BUFC_METADATA) { 4930 zio_buf_free(buf, size); 4931 } else { 4932 ASSERT(type == ARC_BUFC_DATA); 4933 zio_data_buf_free(buf, size); 4934 } 4935} 4936 4937/* 4938 * Free the arc data buffer. 4939 */ 4940static void 4941arc_free_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag) 4942{ 4943 arc_state_t *state = hdr->b_l1hdr.b_state; 4944 arc_buf_contents_t type = arc_buf_type(hdr); 4945 4946 /* protected by hash lock, if in the hash table */ 4947 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) { 4948 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 4949 ASSERT(state != arc_anon && state != arc_l2c_only); 4950 4951 (void) refcount_remove_many(&state->arcs_esize[type], 4952 size, tag); 4953 } 4954 (void) refcount_remove_many(&state->arcs_size, size, tag); 4955 4956 VERIFY3U(hdr->b_type, ==, type); 4957 if (type == ARC_BUFC_METADATA) { 4958 arc_space_return(size, ARC_SPACE_META); 4959 } else { 4960 ASSERT(type == ARC_BUFC_DATA); 4961 arc_space_return(size, ARC_SPACE_DATA); 4962 } 4963} 4964 4965/* 4966 * This routine is called whenever a buffer is accessed. 4967 * NOTE: the hash lock is dropped in this function. 4968 */ 4969static void 4970arc_access(arc_buf_hdr_t *hdr, kmutex_t *hash_lock) 4971{ 4972 clock_t now; 4973 4974 ASSERT(MUTEX_HELD(hash_lock)); 4975 ASSERT(HDR_HAS_L1HDR(hdr)); 4976 4977 if (hdr->b_l1hdr.b_state == arc_anon) { 4978 /* 4979 * This buffer is not in the cache, and does not 4980 * appear in our "ghost" list. Add the new buffer 4981 * to the MRU state. 4982 */ 4983 4984 ASSERT0(hdr->b_l1hdr.b_arc_access); 4985 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 4986 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr); 4987 arc_change_state(arc_mru, hdr, hash_lock); 4988 4989 } else if (hdr->b_l1hdr.b_state == arc_mru) { 4990 now = ddi_get_lbolt(); 4991 4992 /* 4993 * If this buffer is here because of a prefetch, then either: 4994 * - clear the flag if this is a "referencing" read 4995 * (any subsequent access will bump this into the MFU state). 4996 * or 4997 * - move the buffer to the head of the list if this is 4998 * another prefetch (to make it less likely to be evicted). 4999 */ 5000 if (HDR_PREFETCH(hdr)) { 5001 if (refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) { 5002 /* link protected by hash lock */ 5003 ASSERT(multilist_link_active( 5004 &hdr->b_l1hdr.b_arc_node)); 5005 } else { 5006 arc_hdr_clear_flags(hdr, ARC_FLAG_PREFETCH); 5007 ARCSTAT_BUMP(arcstat_mru_hits); 5008 } 5009 hdr->b_l1hdr.b_arc_access = now; 5010 return; 5011 } 5012 5013 /* 5014 * This buffer has been "accessed" only once so far, 5015 * but it is still in the cache. Move it to the MFU 5016 * state. 5017 */ 5018 if (now > hdr->b_l1hdr.b_arc_access + ARC_MINTIME) { 5019 /* 5020 * More than 125ms have passed since we 5021 * instantiated this buffer. Move it to the 5022 * most frequently used state. 5023 */ 5024 hdr->b_l1hdr.b_arc_access = now; 5025 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 5026 arc_change_state(arc_mfu, hdr, hash_lock); 5027 } 5028 ARCSTAT_BUMP(arcstat_mru_hits); 5029 } else if (hdr->b_l1hdr.b_state == arc_mru_ghost) { 5030 arc_state_t *new_state; 5031 /* 5032 * This buffer has been "accessed" recently, but 5033 * was evicted from the cache. Move it to the 5034 * MFU state. 5035 */ 5036 5037 if (HDR_PREFETCH(hdr)) { 5038 new_state = arc_mru; 5039 if (refcount_count(&hdr->b_l1hdr.b_refcnt) > 0) 5040 arc_hdr_clear_flags(hdr, ARC_FLAG_PREFETCH); 5041 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr); 5042 } else { 5043 new_state = arc_mfu; 5044 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 5045 } 5046 5047 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 5048 arc_change_state(new_state, hdr, hash_lock); 5049 5050 ARCSTAT_BUMP(arcstat_mru_ghost_hits); 5051 } else if (hdr->b_l1hdr.b_state == arc_mfu) { 5052 /* 5053 * This buffer has been accessed more than once and is 5054 * still in the cache. Keep it in the MFU state. 5055 * 5056 * NOTE: an add_reference() that occurred when we did 5057 * the arc_read() will have kicked this off the list. 5058 * If it was a prefetch, we will explicitly move it to 5059 * the head of the list now. 5060 */ 5061 if ((HDR_PREFETCH(hdr)) != 0) { 5062 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 5063 /* link protected by hash_lock */ 5064 ASSERT(multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 5065 } 5066 ARCSTAT_BUMP(arcstat_mfu_hits); 5067 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 5068 } else if (hdr->b_l1hdr.b_state == arc_mfu_ghost) { 5069 arc_state_t *new_state = arc_mfu; 5070 /* 5071 * This buffer has been accessed more than once but has 5072 * been evicted from the cache. Move it back to the 5073 * MFU state. 5074 */ 5075 5076 if (HDR_PREFETCH(hdr)) { 5077 /* 5078 * This is a prefetch access... 5079 * move this block back to the MRU state. 5080 */ 5081 ASSERT0(refcount_count(&hdr->b_l1hdr.b_refcnt)); 5082 new_state = arc_mru; 5083 } 5084 5085 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 5086 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 5087 arc_change_state(new_state, hdr, hash_lock); 5088 5089 ARCSTAT_BUMP(arcstat_mfu_ghost_hits); 5090 } else if (hdr->b_l1hdr.b_state == arc_l2c_only) { 5091 /* 5092 * This buffer is on the 2nd Level ARC. 5093 */ 5094 5095 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 5096 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 5097 arc_change_state(arc_mfu, hdr, hash_lock); 5098 } else { 5099 ASSERT(!"invalid arc state"); 5100 } 5101} 5102 5103/* 5104 * This routine is called by dbuf_hold() to update the arc_access() state 5105 * which otherwise would be skipped for entries in the dbuf cache. 5106 */ 5107void 5108arc_buf_access(arc_buf_t *buf) 5109{ 5110 mutex_enter(&buf->b_evict_lock); 5111 arc_buf_hdr_t *hdr = buf->b_hdr; 5112 5113 /* 5114 * Avoid taking the hash_lock when possible as an optimization. 5115 * The header must be checked again under the hash_lock in order 5116 * to handle the case where it is concurrently being released. 5117 */ 5118 if (hdr->b_l1hdr.b_state == arc_anon || HDR_EMPTY(hdr)) { 5119 mutex_exit(&buf->b_evict_lock); 5120 ARCSTAT_BUMP(arcstat_access_skip); 5121 return; 5122 } 5123 5124 kmutex_t *hash_lock = HDR_LOCK(hdr); 5125 mutex_enter(hash_lock); 5126 5127 if (hdr->b_l1hdr.b_state == arc_anon || HDR_EMPTY(hdr)) { 5128 mutex_exit(hash_lock); 5129 mutex_exit(&buf->b_evict_lock); 5130 ARCSTAT_BUMP(arcstat_access_skip); 5131 return; 5132 } 5133 5134 mutex_exit(&buf->b_evict_lock); 5135 5136 ASSERT(hdr->b_l1hdr.b_state == arc_mru || 5137 hdr->b_l1hdr.b_state == arc_mfu); 5138 5139 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr); 5140 arc_access(hdr, hash_lock); 5141 mutex_exit(hash_lock); 5142 5143 ARCSTAT_BUMP(arcstat_hits); 5144 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr), 5145 demand, prefetch, !HDR_ISTYPE_METADATA(hdr), data, metadata, hits); 5146} 5147 5148/* a generic arc_done_func_t which you can use */ 5149/* ARGSUSED */ 5150void 5151arc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg) 5152{ 5153 if (zio == NULL || zio->io_error == 0) 5154 bcopy(buf->b_data, arg, arc_buf_size(buf)); 5155 arc_buf_destroy(buf, arg); 5156} 5157 5158/* a generic arc_done_func_t */ 5159void 5160arc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg) 5161{ 5162 arc_buf_t **bufp = arg; 5163 if (zio && zio->io_error) { 5164 arc_buf_destroy(buf, arg); 5165 *bufp = NULL; 5166 } else { 5167 *bufp = buf; 5168 ASSERT(buf->b_data); 5169 } 5170} 5171 5172static void 5173arc_hdr_verify(arc_buf_hdr_t *hdr, blkptr_t *bp) 5174{ 5175 if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp)) { 5176 ASSERT3U(HDR_GET_PSIZE(hdr), ==, 0); 5177 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF); 5178 } else { 5179 if (HDR_COMPRESSION_ENABLED(hdr)) { 5180 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, 5181 BP_GET_COMPRESS(bp)); 5182 } 5183 ASSERT3U(HDR_GET_LSIZE(hdr), ==, BP_GET_LSIZE(bp)); 5184 ASSERT3U(HDR_GET_PSIZE(hdr), ==, BP_GET_PSIZE(bp)); 5185 } 5186} 5187 5188static void 5189arc_read_done(zio_t *zio) 5190{ 5191 arc_buf_hdr_t *hdr = zio->io_private; 5192 kmutex_t *hash_lock = NULL; 5193 arc_callback_t *callback_list; 5194 arc_callback_t *acb; 5195 boolean_t freeable = B_FALSE; 5196 boolean_t no_zio_error = (zio->io_error == 0); 5197 5198 /* 5199 * The hdr was inserted into hash-table and removed from lists 5200 * prior to starting I/O. We should find this header, since 5201 * it's in the hash table, and it should be legit since it's 5202 * not possible to evict it during the I/O. The only possible 5203 * reason for it not to be found is if we were freed during the 5204 * read. 5205 */ 5206 if (HDR_IN_HASH_TABLE(hdr)) { 5207 ASSERT3U(hdr->b_birth, ==, BP_PHYSICAL_BIRTH(zio->io_bp)); 5208 ASSERT3U(hdr->b_dva.dva_word[0], ==, 5209 BP_IDENTITY(zio->io_bp)->dva_word[0]); 5210 ASSERT3U(hdr->b_dva.dva_word[1], ==, 5211 BP_IDENTITY(zio->io_bp)->dva_word[1]); 5212 5213 arc_buf_hdr_t *found = buf_hash_find(hdr->b_spa, zio->io_bp, 5214 &hash_lock); 5215 5216 ASSERT((found == hdr && 5217 DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) || 5218 (found == hdr && HDR_L2_READING(hdr))); 5219 ASSERT3P(hash_lock, !=, NULL); 5220 } 5221 5222 if (no_zio_error) { 5223 /* byteswap if necessary */ 5224 if (BP_SHOULD_BYTESWAP(zio->io_bp)) { 5225 if (BP_GET_LEVEL(zio->io_bp) > 0) { 5226 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_UINT64; 5227 } else { 5228 hdr->b_l1hdr.b_byteswap = 5229 DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp)); 5230 } 5231 } else { 5232 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS; 5233 } 5234 } 5235 5236 arc_hdr_clear_flags(hdr, ARC_FLAG_L2_EVICTED); 5237 if (l2arc_noprefetch && HDR_PREFETCH(hdr)) 5238 arc_hdr_clear_flags(hdr, ARC_FLAG_L2CACHE); 5239 5240 callback_list = hdr->b_l1hdr.b_acb; 5241 ASSERT3P(callback_list, !=, NULL); 5242 5243 if (hash_lock && no_zio_error && hdr->b_l1hdr.b_state == arc_anon) { 5244 /* 5245 * Only call arc_access on anonymous buffers. This is because 5246 * if we've issued an I/O for an evicted buffer, we've already 5247 * called arc_access (to prevent any simultaneous readers from 5248 * getting confused). 5249 */ 5250 arc_access(hdr, hash_lock); 5251 } 5252 5253 /* 5254 * If a read request has a callback (i.e. acb_done is not NULL), then we 5255 * make a buf containing the data according to the parameters which were 5256 * passed in. The implementation of arc_buf_alloc_impl() ensures that we 5257 * aren't needlessly decompressing the data multiple times. 5258 */ 5259 int callback_cnt = 0; 5260 for (acb = callback_list; acb != NULL; acb = acb->acb_next) { 5261 if (!acb->acb_done) 5262 continue; 5263 5264 /* This is a demand read since prefetches don't use callbacks */ 5265 callback_cnt++; 5266 5267 int error = arc_buf_alloc_impl(hdr, acb->acb_private, 5268 acb->acb_compressed, no_zio_error, &acb->acb_buf); 5269 if (no_zio_error) { 5270 zio->io_error = error; 5271 } 5272 } 5273 hdr->b_l1hdr.b_acb = NULL; 5274 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS); 5275 if (callback_cnt == 0) { 5276 ASSERT(HDR_PREFETCH(hdr)); 5277 ASSERT0(hdr->b_l1hdr.b_bufcnt); 5278 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL); 5279 } 5280 5281 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt) || 5282 callback_list != NULL); 5283 5284 if (no_zio_error) { 5285 arc_hdr_verify(hdr, zio->io_bp); 5286 } else { 5287 arc_hdr_set_flags(hdr, ARC_FLAG_IO_ERROR); 5288 if (hdr->b_l1hdr.b_state != arc_anon) 5289 arc_change_state(arc_anon, hdr, hash_lock); 5290 if (HDR_IN_HASH_TABLE(hdr)) 5291 buf_hash_remove(hdr); 5292 freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt); 5293 } 5294 5295 /* 5296 * Broadcast before we drop the hash_lock to avoid the possibility 5297 * that the hdr (and hence the cv) might be freed before we get to 5298 * the cv_broadcast(). 5299 */ 5300 cv_broadcast(&hdr->b_l1hdr.b_cv); 5301 5302 if (hash_lock != NULL) { 5303 mutex_exit(hash_lock); 5304 } else { 5305 /* 5306 * This block was freed while we waited for the read to 5307 * complete. It has been removed from the hash table and 5308 * moved to the anonymous state (so that it won't show up 5309 * in the cache). 5310 */ 5311 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon); 5312 freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt); 5313 } 5314 5315 /* execute each callback and free its structure */ 5316 while ((acb = callback_list) != NULL) { 5317 if (acb->acb_done) 5318 acb->acb_done(zio, acb->acb_buf, acb->acb_private); 5319 5320 if (acb->acb_zio_dummy != NULL) { 5321 acb->acb_zio_dummy->io_error = zio->io_error; 5322 zio_nowait(acb->acb_zio_dummy); 5323 } 5324 5325 callback_list = acb->acb_next; 5326 kmem_free(acb, sizeof (arc_callback_t)); 5327 } 5328 5329 if (freeable) 5330 arc_hdr_destroy(hdr); 5331} 5332 5333/* 5334 * "Read" the block at the specified DVA (in bp) via the 5335 * cache. If the block is found in the cache, invoke the provided 5336 * callback immediately and return. Note that the `zio' parameter 5337 * in the callback will be NULL in this case, since no IO was 5338 * required. If the block is not in the cache pass the read request 5339 * on to the spa with a substitute callback function, so that the 5340 * requested block will be added to the cache. 5341 * 5342 * If a read request arrives for a block that has a read in-progress, 5343 * either wait for the in-progress read to complete (and return the 5344 * results); or, if this is a read with a "done" func, add a record 5345 * to the read to invoke the "done" func when the read completes, 5346 * and return; or just return. 5347 * 5348 * arc_read_done() will invoke all the requested "done" functions 5349 * for readers of this block. 5350 */ 5351int 5352arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_done_func_t *done, 5353 void *private, zio_priority_t priority, int zio_flags, 5354 arc_flags_t *arc_flags, const zbookmark_phys_t *zb) 5355{ 5356 arc_buf_hdr_t *hdr = NULL; 5357 kmutex_t *hash_lock = NULL; 5358 zio_t *rzio; 5359 uint64_t guid = spa_load_guid(spa); 5360 boolean_t compressed_read = (zio_flags & ZIO_FLAG_RAW) != 0; 5361 5362 ASSERT(!BP_IS_EMBEDDED(bp) || 5363 BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA); 5364 5365top: 5366 if (!BP_IS_EMBEDDED(bp)) { 5367 /* 5368 * Embedded BP's have no DVA and require no I/O to "read". 5369 * Create an anonymous arc buf to back it. 5370 */ 5371 hdr = buf_hash_find(guid, bp, &hash_lock); 5372 } 5373 5374 if (hdr != NULL && HDR_HAS_L1HDR(hdr) && hdr->b_l1hdr.b_pabd != NULL) { 5375 arc_buf_t *buf = NULL; 5376 *arc_flags |= ARC_FLAG_CACHED; 5377 5378 if (HDR_IO_IN_PROGRESS(hdr)) { 5379 5380 if ((hdr->b_flags & ARC_FLAG_PRIO_ASYNC_READ) && 5381 priority == ZIO_PRIORITY_SYNC_READ) { 5382 /* 5383 * This sync read must wait for an 5384 * in-progress async read (e.g. a predictive 5385 * prefetch). Async reads are queued 5386 * separately at the vdev_queue layer, so 5387 * this is a form of priority inversion. 5388 * Ideally, we would "inherit" the demand 5389 * i/o's priority by moving the i/o from 5390 * the async queue to the synchronous queue, 5391 * but there is currently no mechanism to do 5392 * so. Track this so that we can evaluate 5393 * the magnitude of this potential performance 5394 * problem. 5395 * 5396 * Note that if the prefetch i/o is already 5397 * active (has been issued to the device), 5398 * the prefetch improved performance, because 5399 * we issued it sooner than we would have 5400 * without the prefetch. 5401 */ 5402 DTRACE_PROBE1(arc__sync__wait__for__async, 5403 arc_buf_hdr_t *, hdr); 5404 ARCSTAT_BUMP(arcstat_sync_wait_for_async); 5405 } 5406 if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) { 5407 arc_hdr_clear_flags(hdr, 5408 ARC_FLAG_PREDICTIVE_PREFETCH); 5409 } 5410 5411 if (*arc_flags & ARC_FLAG_WAIT) { 5412 cv_wait(&hdr->b_l1hdr.b_cv, hash_lock); 5413 mutex_exit(hash_lock); 5414 goto top; 5415 } 5416 ASSERT(*arc_flags & ARC_FLAG_NOWAIT); 5417 5418 if (done) { 5419 arc_callback_t *acb = NULL; 5420 5421 acb = kmem_zalloc(sizeof (arc_callback_t), 5422 KM_SLEEP); 5423 acb->acb_done = done; 5424 acb->acb_private = private; 5425 acb->acb_compressed = compressed_read; 5426 if (pio != NULL) 5427 acb->acb_zio_dummy = zio_null(pio, 5428 spa, NULL, NULL, NULL, zio_flags); 5429 5430 ASSERT3P(acb->acb_done, !=, NULL); 5431 acb->acb_next = hdr->b_l1hdr.b_acb; 5432 hdr->b_l1hdr.b_acb = acb; 5433 mutex_exit(hash_lock); 5434 return (0); 5435 } 5436 mutex_exit(hash_lock); 5437 return (0); 5438 } 5439 5440 ASSERT(hdr->b_l1hdr.b_state == arc_mru || 5441 hdr->b_l1hdr.b_state == arc_mfu); 5442 5443 if (done) { 5444 if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) { 5445 /* 5446 * This is a demand read which does not have to 5447 * wait for i/o because we did a predictive 5448 * prefetch i/o for it, which has completed. 5449 */ 5450 DTRACE_PROBE1( 5451 arc__demand__hit__predictive__prefetch, 5452 arc_buf_hdr_t *, hdr); 5453 ARCSTAT_BUMP( 5454 arcstat_demand_hit_predictive_prefetch); 5455 arc_hdr_clear_flags(hdr, 5456 ARC_FLAG_PREDICTIVE_PREFETCH); 5457 } 5458 ASSERT(!BP_IS_EMBEDDED(bp) || !BP_IS_HOLE(bp)); 5459 5460 /* Get a buf with the desired data in it. */ 5461 VERIFY0(arc_buf_alloc_impl(hdr, private, 5462 compressed_read, B_TRUE, &buf)); 5463 } else if (*arc_flags & ARC_FLAG_PREFETCH && 5464 refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) { 5465 arc_hdr_set_flags(hdr, ARC_FLAG_PREFETCH); 5466 } 5467 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr); 5468 arc_access(hdr, hash_lock); 5469 if (*arc_flags & ARC_FLAG_L2CACHE) 5470 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE); 5471 mutex_exit(hash_lock); 5472 ARCSTAT_BUMP(arcstat_hits); 5473 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr), 5474 demand, prefetch, !HDR_ISTYPE_METADATA(hdr), 5475 data, metadata, hits); 5476 5477 if (done) 5478 done(NULL, buf, private); 5479 } else { 5480 uint64_t lsize = BP_GET_LSIZE(bp); 5481 uint64_t psize = BP_GET_PSIZE(bp); 5482 arc_callback_t *acb; 5483 vdev_t *vd = NULL; 5484 uint64_t addr = 0; 5485 boolean_t devw = B_FALSE; 5486 uint64_t size; 5487 5488 if (hdr == NULL) { 5489 /* this block is not in the cache */ 5490 arc_buf_hdr_t *exists = NULL; 5491 arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp); 5492 hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize, 5493 BP_GET_COMPRESS(bp), type); 5494 5495 if (!BP_IS_EMBEDDED(bp)) { 5496 hdr->b_dva = *BP_IDENTITY(bp); 5497 hdr->b_birth = BP_PHYSICAL_BIRTH(bp); 5498 exists = buf_hash_insert(hdr, &hash_lock); 5499 } 5500 if (exists != NULL) { 5501 /* somebody beat us to the hash insert */ 5502 mutex_exit(hash_lock); 5503 buf_discard_identity(hdr); 5504 arc_hdr_destroy(hdr); 5505 goto top; /* restart the IO request */ 5506 } 5507 } else { 5508 /* 5509 * This block is in the ghost cache. If it was L2-only 5510 * (and thus didn't have an L1 hdr), we realloc the 5511 * header to add an L1 hdr. 5512 */ 5513 if (!HDR_HAS_L1HDR(hdr)) { 5514 hdr = arc_hdr_realloc(hdr, hdr_l2only_cache, 5515 hdr_full_cache); 5516 } 5517 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL); 5518 ASSERT(GHOST_STATE(hdr->b_l1hdr.b_state)); 5519 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 5520 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 5521 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 5522 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL); 5523 5524 /* 5525 * This is a delicate dance that we play here. 5526 * This hdr is in the ghost list so we access it 5527 * to move it out of the ghost list before we 5528 * initiate the read. If it's a prefetch then 5529 * it won't have a callback so we'll remove the 5530 * reference that arc_buf_alloc_impl() created. We 5531 * do this after we've called arc_access() to 5532 * avoid hitting an assert in remove_reference(). 5533 */ 5534 arc_access(hdr, hash_lock); 5535 arc_hdr_alloc_pabd(hdr); 5536 } 5537 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL); 5538 size = arc_hdr_size(hdr); 5539 5540 /* 5541 * If compression is enabled on the hdr, then will do 5542 * RAW I/O and will store the compressed data in the hdr's 5543 * data block. Otherwise, the hdr's data block will contain 5544 * the uncompressed data. 5545 */ 5546 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF) { 5547 zio_flags |= ZIO_FLAG_RAW; 5548 } 5549 5550 if (*arc_flags & ARC_FLAG_PREFETCH) 5551 arc_hdr_set_flags(hdr, ARC_FLAG_PREFETCH); 5552 if (*arc_flags & ARC_FLAG_L2CACHE) 5553 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE); 5554 if (BP_GET_LEVEL(bp) > 0) 5555 arc_hdr_set_flags(hdr, ARC_FLAG_INDIRECT); 5556 if (*arc_flags & ARC_FLAG_PREDICTIVE_PREFETCH) 5557 arc_hdr_set_flags(hdr, ARC_FLAG_PREDICTIVE_PREFETCH); 5558 ASSERT(!GHOST_STATE(hdr->b_l1hdr.b_state)); 5559 5560 acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP); 5561 acb->acb_done = done; 5562 acb->acb_private = private; 5563 acb->acb_compressed = compressed_read; 5564 5565 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL); 5566 hdr->b_l1hdr.b_acb = acb; 5567 arc_hdr_set_flags(hdr, ARC_FLAG_IO_IN_PROGRESS); 5568 5569 if (HDR_HAS_L2HDR(hdr) && 5570 (vd = hdr->b_l2hdr.b_dev->l2ad_vdev) != NULL) { 5571 devw = hdr->b_l2hdr.b_dev->l2ad_writing; 5572 addr = hdr->b_l2hdr.b_daddr; 5573 /* 5574 * Lock out L2ARC device removal. 5575 */ 5576 if (vdev_is_dead(vd) || 5577 !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER)) 5578 vd = NULL; 5579 } 5580 5581 if (priority == ZIO_PRIORITY_ASYNC_READ) 5582 arc_hdr_set_flags(hdr, ARC_FLAG_PRIO_ASYNC_READ); 5583 else 5584 arc_hdr_clear_flags(hdr, ARC_FLAG_PRIO_ASYNC_READ); 5585 5586 if (hash_lock != NULL) 5587 mutex_exit(hash_lock); 5588 5589 /* 5590 * At this point, we have a level 1 cache miss. Try again in 5591 * L2ARC if possible. 5592 */ 5593 ASSERT3U(HDR_GET_LSIZE(hdr), ==, lsize); 5594 5595 DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp, 5596 uint64_t, lsize, zbookmark_phys_t *, zb); 5597 ARCSTAT_BUMP(arcstat_misses); 5598 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr), 5599 demand, prefetch, !HDR_ISTYPE_METADATA(hdr), 5600 data, metadata, misses); 5601#ifdef _KERNEL 5602#ifdef RACCT 5603 if (racct_enable) { 5604 PROC_LOCK(curproc); 5605 racct_add_force(curproc, RACCT_READBPS, size); 5606 racct_add_force(curproc, RACCT_READIOPS, 1); 5607 PROC_UNLOCK(curproc); 5608 } 5609#endif /* RACCT */ 5610 curthread->td_ru.ru_inblock++; 5611#endif 5612 5613 if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) { 5614 /* 5615 * Read from the L2ARC if the following are true: 5616 * 1. The L2ARC vdev was previously cached. 5617 * 2. This buffer still has L2ARC metadata. 5618 * 3. This buffer isn't currently writing to the L2ARC. 5619 * 4. The L2ARC entry wasn't evicted, which may 5620 * also have invalidated the vdev. 5621 * 5. This isn't prefetch and l2arc_noprefetch is set. 5622 */ 5623 if (HDR_HAS_L2HDR(hdr) && 5624 !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) && 5625 !(l2arc_noprefetch && HDR_PREFETCH(hdr))) { 5626 l2arc_read_callback_t *cb; 5627 abd_t *abd; 5628 uint64_t asize; 5629 5630 DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr); 5631 ARCSTAT_BUMP(arcstat_l2_hits); 5632 5633 cb = kmem_zalloc(sizeof (l2arc_read_callback_t), 5634 KM_SLEEP); 5635 cb->l2rcb_hdr = hdr; 5636 cb->l2rcb_bp = *bp; 5637 cb->l2rcb_zb = *zb; 5638 cb->l2rcb_flags = zio_flags; 5639 5640 asize = vdev_psize_to_asize(vd, size); 5641 if (asize != size) { 5642 abd = abd_alloc_for_io(asize, 5643 HDR_ISTYPE_METADATA(hdr)); 5644 cb->l2rcb_abd = abd; 5645 } else { 5646 abd = hdr->b_l1hdr.b_pabd; 5647 } 5648 5649 ASSERT(addr >= VDEV_LABEL_START_SIZE && 5650 addr + asize <= vd->vdev_psize - 5651 VDEV_LABEL_END_SIZE); 5652 5653 /* 5654 * l2arc read. The SCL_L2ARC lock will be 5655 * released by l2arc_read_done(). 5656 * Issue a null zio if the underlying buffer 5657 * was squashed to zero size by compression. 5658 */ 5659 ASSERT3U(HDR_GET_COMPRESS(hdr), !=, 5660 ZIO_COMPRESS_EMPTY); 5661 rzio = zio_read_phys(pio, vd, addr, 5662 asize, abd, 5663 ZIO_CHECKSUM_OFF, 5664 l2arc_read_done, cb, priority, 5665 zio_flags | ZIO_FLAG_DONT_CACHE | 5666 ZIO_FLAG_CANFAIL | 5667 ZIO_FLAG_DONT_PROPAGATE | 5668 ZIO_FLAG_DONT_RETRY, B_FALSE); 5669 DTRACE_PROBE2(l2arc__read, vdev_t *, vd, 5670 zio_t *, rzio); 5671 ARCSTAT_INCR(arcstat_l2_read_bytes, size); 5672 5673 if (*arc_flags & ARC_FLAG_NOWAIT) { 5674 zio_nowait(rzio); 5675 return (0); 5676 } 5677 5678 ASSERT(*arc_flags & ARC_FLAG_WAIT); 5679 if (zio_wait(rzio) == 0) 5680 return (0); 5681 5682 /* l2arc read error; goto zio_read() */ 5683 } else { 5684 DTRACE_PROBE1(l2arc__miss, 5685 arc_buf_hdr_t *, hdr); 5686 ARCSTAT_BUMP(arcstat_l2_misses); 5687 if (HDR_L2_WRITING(hdr)) 5688 ARCSTAT_BUMP(arcstat_l2_rw_clash); 5689 spa_config_exit(spa, SCL_L2ARC, vd); 5690 } 5691 } else { 5692 if (vd != NULL) 5693 spa_config_exit(spa, SCL_L2ARC, vd); 5694 if (l2arc_ndev != 0) { 5695 DTRACE_PROBE1(l2arc__miss, 5696 arc_buf_hdr_t *, hdr); 5697 ARCSTAT_BUMP(arcstat_l2_misses); 5698 } 5699 } 5700 5701 rzio = zio_read(pio, spa, bp, hdr->b_l1hdr.b_pabd, size, 5702 arc_read_done, hdr, priority, zio_flags, zb); 5703 5704 if (*arc_flags & ARC_FLAG_WAIT) 5705 return (zio_wait(rzio)); 5706 5707 ASSERT(*arc_flags & ARC_FLAG_NOWAIT); 5708 zio_nowait(rzio); 5709 } 5710 return (0); 5711} 5712 5713/* 5714 * Notify the arc that a block was freed, and thus will never be used again. 5715 */ 5716void 5717arc_freed(spa_t *spa, const blkptr_t *bp) 5718{ 5719 arc_buf_hdr_t *hdr; 5720 kmutex_t *hash_lock; 5721 uint64_t guid = spa_load_guid(spa); 5722 5723 ASSERT(!BP_IS_EMBEDDED(bp)); 5724 5725 hdr = buf_hash_find(guid, bp, &hash_lock); 5726 if (hdr == NULL) 5727 return; 5728 5729 /* 5730 * We might be trying to free a block that is still doing I/O 5731 * (i.e. prefetch) or has a reference (i.e. a dedup-ed, 5732 * dmu_sync-ed block). If this block is being prefetched, then it 5733 * would still have the ARC_FLAG_IO_IN_PROGRESS flag set on the hdr 5734 * until the I/O completes. A block may also have a reference if it is 5735 * part of a dedup-ed, dmu_synced write. The dmu_sync() function would 5736 * have written the new block to its final resting place on disk but 5737 * without the dedup flag set. This would have left the hdr in the MRU 5738 * state and discoverable. When the txg finally syncs it detects that 5739 * the block was overridden in open context and issues an override I/O. 5740 * Since this is a dedup block, the override I/O will determine if the 5741 * block is already in the DDT. If so, then it will replace the io_bp 5742 * with the bp from the DDT and allow the I/O to finish. When the I/O 5743 * reaches the done callback, dbuf_write_override_done, it will 5744 * check to see if the io_bp and io_bp_override are identical. 5745 * If they are not, then it indicates that the bp was replaced with 5746 * the bp in the DDT and the override bp is freed. This allows 5747 * us to arrive here with a reference on a block that is being 5748 * freed. So if we have an I/O in progress, or a reference to 5749 * this hdr, then we don't destroy the hdr. 5750 */ 5751 if (!HDR_HAS_L1HDR(hdr) || (!HDR_IO_IN_PROGRESS(hdr) && 5752 refcount_is_zero(&hdr->b_l1hdr.b_refcnt))) { 5753 arc_change_state(arc_anon, hdr, hash_lock); 5754 arc_hdr_destroy(hdr); 5755 mutex_exit(hash_lock); 5756 } else { 5757 mutex_exit(hash_lock); 5758 } 5759 5760} 5761 5762/* 5763 * Release this buffer from the cache, making it an anonymous buffer. This 5764 * must be done after a read and prior to modifying the buffer contents. 5765 * If the buffer has more than one reference, we must make 5766 * a new hdr for the buffer. 5767 */ 5768void 5769arc_release(arc_buf_t *buf, void *tag) 5770{ 5771 arc_buf_hdr_t *hdr = buf->b_hdr; 5772 5773 /* 5774 * It would be nice to assert that if it's DMU metadata (level > 5775 * 0 || it's the dnode file), then it must be syncing context. 5776 * But we don't know that information at this level. 5777 */ 5778 5779 mutex_enter(&buf->b_evict_lock); 5780 5781 ASSERT(HDR_HAS_L1HDR(hdr)); 5782 5783 /* 5784 * We don't grab the hash lock prior to this check, because if 5785 * the buffer's header is in the arc_anon state, it won't be 5786 * linked into the hash table. 5787 */ 5788 if (hdr->b_l1hdr.b_state == arc_anon) { 5789 mutex_exit(&buf->b_evict_lock); 5790 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 5791 ASSERT(!HDR_IN_HASH_TABLE(hdr)); 5792 ASSERT(!HDR_HAS_L2HDR(hdr)); 5793 ASSERT(HDR_EMPTY(hdr)); 5794 ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1); 5795 ASSERT3S(refcount_count(&hdr->b_l1hdr.b_refcnt), ==, 1); 5796 ASSERT(!list_link_active(&hdr->b_l1hdr.b_arc_node)); 5797 5798 hdr->b_l1hdr.b_arc_access = 0; 5799 5800 /* 5801 * If the buf is being overridden then it may already 5802 * have a hdr that is not empty. 5803 */ 5804 buf_discard_identity(hdr); 5805 arc_buf_thaw(buf); 5806 5807 return; 5808 } 5809 5810 kmutex_t *hash_lock = HDR_LOCK(hdr); 5811 mutex_enter(hash_lock); 5812 5813 /* 5814 * This assignment is only valid as long as the hash_lock is 5815 * held, we must be careful not to reference state or the 5816 * b_state field after dropping the lock. 5817 */ 5818 arc_state_t *state = hdr->b_l1hdr.b_state; 5819 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 5820 ASSERT3P(state, !=, arc_anon); 5821 5822 /* this buffer is not on any list */ 5823 ASSERT3S(refcount_count(&hdr->b_l1hdr.b_refcnt), >, 0); 5824 5825 if (HDR_HAS_L2HDR(hdr)) { 5826 mutex_enter(&hdr->b_l2hdr.b_dev->l2ad_mtx); 5827 5828 /* 5829 * We have to recheck this conditional again now that 5830 * we're holding the l2ad_mtx to prevent a race with 5831 * another thread which might be concurrently calling 5832 * l2arc_evict(). In that case, l2arc_evict() might have 5833 * destroyed the header's L2 portion as we were waiting 5834 * to acquire the l2ad_mtx. 5835 */ 5836 if (HDR_HAS_L2HDR(hdr)) { 5837 l2arc_trim(hdr); 5838 arc_hdr_l2hdr_destroy(hdr); 5839 } 5840 5841 mutex_exit(&hdr->b_l2hdr.b_dev->l2ad_mtx); 5842 } 5843 5844 /* 5845 * Do we have more than one buf? 5846 */ 5847 if (hdr->b_l1hdr.b_bufcnt > 1) { 5848 arc_buf_hdr_t *nhdr; 5849 uint64_t spa = hdr->b_spa; 5850 uint64_t psize = HDR_GET_PSIZE(hdr); 5851 uint64_t lsize = HDR_GET_LSIZE(hdr); 5852 enum zio_compress compress = HDR_GET_COMPRESS(hdr); 5853 arc_buf_contents_t type = arc_buf_type(hdr); 5854 VERIFY3U(hdr->b_type, ==, type); 5855 5856 ASSERT(hdr->b_l1hdr.b_buf != buf || buf->b_next != NULL); 5857 (void) remove_reference(hdr, hash_lock, tag); 5858 5859 if (arc_buf_is_shared(buf) && !ARC_BUF_COMPRESSED(buf)) { 5860 ASSERT3P(hdr->b_l1hdr.b_buf, !=, buf); 5861 ASSERT(ARC_BUF_LAST(buf)); 5862 } 5863 5864 /* 5865 * Pull the data off of this hdr and attach it to 5866 * a new anonymous hdr. Also find the last buffer 5867 * in the hdr's buffer list. 5868 */ 5869 arc_buf_t *lastbuf = arc_buf_remove(hdr, buf); 5870 ASSERT3P(lastbuf, !=, NULL); 5871 5872 /* 5873 * If the current arc_buf_t and the hdr are sharing their data 5874 * buffer, then we must stop sharing that block. 5875 */ 5876 if (arc_buf_is_shared(buf)) { 5877 VERIFY(!arc_buf_is_shared(lastbuf)); 5878 5879 /* 5880 * First, sever the block sharing relationship between 5881 * buf and the arc_buf_hdr_t. 5882 */ 5883 arc_unshare_buf(hdr, buf); 5884 5885 /* 5886 * Now we need to recreate the hdr's b_pabd. Since we 5887 * have lastbuf handy, we try to share with it, but if 5888 * we can't then we allocate a new b_pabd and copy the 5889 * data from buf into it. 5890 */ 5891 if (arc_can_share(hdr, lastbuf)) { 5892 arc_share_buf(hdr, lastbuf); 5893 } else { 5894 arc_hdr_alloc_pabd(hdr); 5895 abd_copy_from_buf(hdr->b_l1hdr.b_pabd, 5896 buf->b_data, psize); 5897 } 5898 VERIFY3P(lastbuf->b_data, !=, NULL); 5899 } else if (HDR_SHARED_DATA(hdr)) { 5900 /* 5901 * Uncompressed shared buffers are always at the end 5902 * of the list. Compressed buffers don't have the 5903 * same requirements. This makes it hard to 5904 * simply assert that the lastbuf is shared so 5905 * we rely on the hdr's compression flags to determine 5906 * if we have a compressed, shared buffer. 5907 */ 5908 ASSERT(arc_buf_is_shared(lastbuf) || 5909 HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF); 5910 ASSERT(!ARC_BUF_SHARED(buf)); 5911 } 5912 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL); 5913 ASSERT3P(state, !=, arc_l2c_only); 5914 5915 (void) refcount_remove_many(&state->arcs_size, 5916 arc_buf_size(buf), buf); 5917 5918 if (refcount_is_zero(&hdr->b_l1hdr.b_refcnt)) { 5919 ASSERT3P(state, !=, arc_l2c_only); 5920 (void) refcount_remove_many(&state->arcs_esize[type], 5921 arc_buf_size(buf), buf); 5922 } 5923 5924 hdr->b_l1hdr.b_bufcnt -= 1; 5925 arc_cksum_verify(buf); 5926#ifdef illumos 5927 arc_buf_unwatch(buf); 5928#endif 5929 5930 mutex_exit(hash_lock); 5931 5932 /* 5933 * Allocate a new hdr. The new hdr will contain a b_pabd 5934 * buffer which will be freed in arc_write(). 5935 */ 5936 nhdr = arc_hdr_alloc(spa, psize, lsize, compress, type); 5937 ASSERT3P(nhdr->b_l1hdr.b_buf, ==, NULL); 5938 ASSERT0(nhdr->b_l1hdr.b_bufcnt); 5939 ASSERT0(refcount_count(&nhdr->b_l1hdr.b_refcnt)); 5940 VERIFY3U(nhdr->b_type, ==, type); 5941 ASSERT(!HDR_SHARED_DATA(nhdr)); 5942 5943 nhdr->b_l1hdr.b_buf = buf; 5944 nhdr->b_l1hdr.b_bufcnt = 1; 5945 (void) refcount_add(&nhdr->b_l1hdr.b_refcnt, tag); 5946 buf->b_hdr = nhdr; 5947 5948 mutex_exit(&buf->b_evict_lock); 5949 (void) refcount_add_many(&arc_anon->arcs_size, 5950 arc_buf_size(buf), buf); 5951 } else { 5952 mutex_exit(&buf->b_evict_lock); 5953 ASSERT(refcount_count(&hdr->b_l1hdr.b_refcnt) == 1); 5954 /* protected by hash lock, or hdr is on arc_anon */ 5955 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 5956 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 5957 arc_change_state(arc_anon, hdr, hash_lock); 5958 hdr->b_l1hdr.b_arc_access = 0; 5959 mutex_exit(hash_lock); 5960 5961 buf_discard_identity(hdr); 5962 arc_buf_thaw(buf); 5963 } 5964} 5965 5966int 5967arc_released(arc_buf_t *buf) 5968{ 5969 int released; 5970 5971 mutex_enter(&buf->b_evict_lock); 5972 released = (buf->b_data != NULL && 5973 buf->b_hdr->b_l1hdr.b_state == arc_anon); 5974 mutex_exit(&buf->b_evict_lock); 5975 return (released); 5976} 5977 5978#ifdef ZFS_DEBUG 5979int 5980arc_referenced(arc_buf_t *buf) 5981{ 5982 int referenced; 5983 5984 mutex_enter(&buf->b_evict_lock); 5985 referenced = (refcount_count(&buf->b_hdr->b_l1hdr.b_refcnt)); 5986 mutex_exit(&buf->b_evict_lock); 5987 return (referenced); 5988} 5989#endif 5990 5991static void 5992arc_write_ready(zio_t *zio) 5993{ 5994 arc_write_callback_t *callback = zio->io_private; 5995 arc_buf_t *buf = callback->awcb_buf; 5996 arc_buf_hdr_t *hdr = buf->b_hdr; 5997 uint64_t psize = BP_IS_HOLE(zio->io_bp) ? 0 : BP_GET_PSIZE(zio->io_bp); 5998 5999 ASSERT(HDR_HAS_L1HDR(hdr)); 6000 ASSERT(!refcount_is_zero(&buf->b_hdr->b_l1hdr.b_refcnt)); 6001 ASSERT(hdr->b_l1hdr.b_bufcnt > 0); 6002 6003 /* 6004 * If we're reexecuting this zio because the pool suspended, then 6005 * cleanup any state that was previously set the first time the 6006 * callback was invoked. 6007 */ 6008 if (zio->io_flags & ZIO_FLAG_REEXECUTED) { 6009 arc_cksum_free(hdr); 6010#ifdef illumos 6011 arc_buf_unwatch(buf); 6012#endif 6013 if (hdr->b_l1hdr.b_pabd != NULL) { 6014 if (arc_buf_is_shared(buf)) { 6015 arc_unshare_buf(hdr, buf); 6016 } else { 6017 arc_hdr_free_pabd(hdr); 6018 } 6019 } 6020 } 6021 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL); 6022 ASSERT(!HDR_SHARED_DATA(hdr)); 6023 ASSERT(!arc_buf_is_shared(buf)); 6024 6025 callback->awcb_ready(zio, buf, callback->awcb_private); 6026 6027 if (HDR_IO_IN_PROGRESS(hdr)) 6028 ASSERT(zio->io_flags & ZIO_FLAG_REEXECUTED); 6029 6030 arc_cksum_compute(buf); 6031 arc_hdr_set_flags(hdr, ARC_FLAG_IO_IN_PROGRESS); 6032 6033 enum zio_compress compress; 6034 if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) { 6035 compress = ZIO_COMPRESS_OFF; 6036 } else { 6037 ASSERT3U(HDR_GET_LSIZE(hdr), ==, BP_GET_LSIZE(zio->io_bp)); 6038 compress = BP_GET_COMPRESS(zio->io_bp); 6039 } 6040 HDR_SET_PSIZE(hdr, psize); 6041 arc_hdr_set_compress(hdr, compress); 6042 6043 6044 /* 6045 * Fill the hdr with data. If the hdr is compressed, the data we want 6046 * is available from the zio, otherwise we can take it from the buf. 6047 * 6048 * We might be able to share the buf's data with the hdr here. However, 6049 * doing so would cause the ARC to be full of linear ABDs if we write a 6050 * lot of shareable data. As a compromise, we check whether scattered 6051 * ABDs are allowed, and assume that if they are then the user wants 6052 * the ARC to be primarily filled with them regardless of the data being 6053 * written. Therefore, if they're allowed then we allocate one and copy 6054 * the data into it; otherwise, we share the data directly if we can. 6055 */ 6056 if (zfs_abd_scatter_enabled || !arc_can_share(hdr, buf)) { 6057 arc_hdr_alloc_pabd(hdr); 6058 6059 /* 6060 * Ideally, we would always copy the io_abd into b_pabd, but the 6061 * user may have disabled compressed ARC, thus we must check the 6062 * hdr's compression setting rather than the io_bp's. 6063 */ 6064 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF) { 6065 ASSERT3U(BP_GET_COMPRESS(zio->io_bp), !=, 6066 ZIO_COMPRESS_OFF); 6067 ASSERT3U(psize, >, 0); 6068 6069 abd_copy(hdr->b_l1hdr.b_pabd, zio->io_abd, psize); 6070 } else { 6071 ASSERT3U(zio->io_orig_size, ==, arc_hdr_size(hdr)); 6072 6073 abd_copy_from_buf(hdr->b_l1hdr.b_pabd, buf->b_data, 6074 arc_buf_size(buf)); 6075 } 6076 } else { 6077 ASSERT3P(buf->b_data, ==, abd_to_buf(zio->io_orig_abd)); 6078 ASSERT3U(zio->io_orig_size, ==, arc_buf_size(buf)); 6079 ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1); 6080 6081 arc_share_buf(hdr, buf); 6082 } 6083 6084 arc_hdr_verify(hdr, zio->io_bp); 6085} 6086 6087static void 6088arc_write_children_ready(zio_t *zio) 6089{ 6090 arc_write_callback_t *callback = zio->io_private; 6091 arc_buf_t *buf = callback->awcb_buf; 6092 6093 callback->awcb_children_ready(zio, buf, callback->awcb_private); 6094} 6095 6096/* 6097 * The SPA calls this callback for each physical write that happens on behalf 6098 * of a logical write. See the comment in dbuf_write_physdone() for details. 6099 */ 6100static void 6101arc_write_physdone(zio_t *zio) 6102{ 6103 arc_write_callback_t *cb = zio->io_private; 6104 if (cb->awcb_physdone != NULL) 6105 cb->awcb_physdone(zio, cb->awcb_buf, cb->awcb_private); 6106} 6107 6108static void 6109arc_write_done(zio_t *zio) 6110{ 6111 arc_write_callback_t *callback = zio->io_private; 6112 arc_buf_t *buf = callback->awcb_buf; 6113 arc_buf_hdr_t *hdr = buf->b_hdr; 6114 6115 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL); 6116 6117 if (zio->io_error == 0) { 6118 arc_hdr_verify(hdr, zio->io_bp); 6119 6120 if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) { 6121 buf_discard_identity(hdr); 6122 } else { 6123 hdr->b_dva = *BP_IDENTITY(zio->io_bp); 6124 hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp); 6125 } 6126 } else { 6127 ASSERT(HDR_EMPTY(hdr)); 6128 } 6129 6130 /* 6131 * If the block to be written was all-zero or compressed enough to be 6132 * embedded in the BP, no write was performed so there will be no 6133 * dva/birth/checksum. The buffer must therefore remain anonymous 6134 * (and uncached). 6135 */ 6136 if (!HDR_EMPTY(hdr)) { 6137 arc_buf_hdr_t *exists; 6138 kmutex_t *hash_lock; 6139 6140 ASSERT3U(zio->io_error, ==, 0); 6141 6142 arc_cksum_verify(buf); 6143 6144 exists = buf_hash_insert(hdr, &hash_lock); 6145 if (exists != NULL) { 6146 /* 6147 * This can only happen if we overwrite for 6148 * sync-to-convergence, because we remove 6149 * buffers from the hash table when we arc_free(). 6150 */ 6151 if (zio->io_flags & ZIO_FLAG_IO_REWRITE) { 6152 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) 6153 panic("bad overwrite, hdr=%p exists=%p", 6154 (void *)hdr, (void *)exists); 6155 ASSERT(refcount_is_zero( 6156 &exists->b_l1hdr.b_refcnt)); 6157 arc_change_state(arc_anon, exists, hash_lock); 6158 mutex_exit(hash_lock); 6159 arc_hdr_destroy(exists); 6160 exists = buf_hash_insert(hdr, &hash_lock); 6161 ASSERT3P(exists, ==, NULL); 6162 } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) { 6163 /* nopwrite */ 6164 ASSERT(zio->io_prop.zp_nopwrite); 6165 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) 6166 panic("bad nopwrite, hdr=%p exists=%p", 6167 (void *)hdr, (void *)exists); 6168 } else { 6169 /* Dedup */ 6170 ASSERT(hdr->b_l1hdr.b_bufcnt == 1); 6171 ASSERT(hdr->b_l1hdr.b_state == arc_anon); 6172 ASSERT(BP_GET_DEDUP(zio->io_bp)); 6173 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0); 6174 } 6175 } 6176 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS); 6177 /* if it's not anon, we are doing a scrub */ 6178 if (exists == NULL && hdr->b_l1hdr.b_state == arc_anon) 6179 arc_access(hdr, hash_lock); 6180 mutex_exit(hash_lock); 6181 } else { 6182 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS); 6183 } 6184 6185 ASSERT(!refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 6186 callback->awcb_done(zio, buf, callback->awcb_private); 6187 6188 abd_put(zio->io_abd); 6189 kmem_free(callback, sizeof (arc_write_callback_t)); 6190} 6191 6192zio_t * 6193arc_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, arc_buf_t *buf, 6194 boolean_t l2arc, const zio_prop_t *zp, arc_done_func_t *ready, 6195 arc_done_func_t *children_ready, arc_done_func_t *physdone, 6196 arc_done_func_t *done, void *private, zio_priority_t priority, 6197 int zio_flags, const zbookmark_phys_t *zb) 6198{ 6199 arc_buf_hdr_t *hdr = buf->b_hdr; 6200 arc_write_callback_t *callback; 6201 zio_t *zio; 6202 zio_prop_t localprop = *zp; 6203 6204 ASSERT3P(ready, !=, NULL); 6205 ASSERT3P(done, !=, NULL); 6206 ASSERT(!HDR_IO_ERROR(hdr)); 6207 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 6208 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL); 6209 ASSERT3U(hdr->b_l1hdr.b_bufcnt, >, 0); 6210 if (l2arc) 6211 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE); 6212 if (ARC_BUF_COMPRESSED(buf)) { 6213 /* 6214 * We're writing a pre-compressed buffer. Make the 6215 * compression algorithm requested by the zio_prop_t match 6216 * the pre-compressed buffer's compression algorithm. 6217 */ 6218 localprop.zp_compress = HDR_GET_COMPRESS(hdr); 6219 6220 ASSERT3U(HDR_GET_LSIZE(hdr), !=, arc_buf_size(buf)); 6221 zio_flags |= ZIO_FLAG_RAW; 6222 } 6223 callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP); 6224 callback->awcb_ready = ready; 6225 callback->awcb_children_ready = children_ready; 6226 callback->awcb_physdone = physdone; 6227 callback->awcb_done = done; 6228 callback->awcb_private = private; 6229 callback->awcb_buf = buf; 6230 6231 /* 6232 * The hdr's b_pabd is now stale, free it now. A new data block 6233 * will be allocated when the zio pipeline calls arc_write_ready(). 6234 */ 6235 if (hdr->b_l1hdr.b_pabd != NULL) { 6236 /* 6237 * If the buf is currently sharing the data block with 6238 * the hdr then we need to break that relationship here. 6239 * The hdr will remain with a NULL data pointer and the 6240 * buf will take sole ownership of the block. 6241 */ 6242 if (arc_buf_is_shared(buf)) { 6243 arc_unshare_buf(hdr, buf); 6244 } else { 6245 arc_hdr_free_pabd(hdr); 6246 } 6247 VERIFY3P(buf->b_data, !=, NULL); 6248 arc_hdr_set_compress(hdr, ZIO_COMPRESS_OFF); 6249 } 6250 ASSERT(!arc_buf_is_shared(buf)); 6251 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL); 6252 6253 zio = zio_write(pio, spa, txg, bp, 6254 abd_get_from_buf(buf->b_data, HDR_GET_LSIZE(hdr)), 6255 HDR_GET_LSIZE(hdr), arc_buf_size(buf), &localprop, arc_write_ready, 6256 (children_ready != NULL) ? arc_write_children_ready : NULL, 6257 arc_write_physdone, arc_write_done, callback, 6258 priority, zio_flags, zb); 6259 6260 return (zio); 6261} 6262 6263static int 6264arc_memory_throttle(uint64_t reserve, uint64_t txg) 6265{ 6266#ifdef _KERNEL 6267 uint64_t available_memory = ptob(freemem); 6268 static uint64_t page_load = 0; 6269 static uint64_t last_txg = 0; 6270 6271#if defined(__i386) || !defined(UMA_MD_SMALL_ALLOC) 6272 available_memory = 6273 MIN(available_memory, ptob(vmem_size(heap_arena, VMEM_FREE))); 6274#endif 6275 6276 if (freemem > (uint64_t)physmem * arc_lotsfree_percent / 100) 6277 return (0); 6278 6279 if (txg > last_txg) { 6280 last_txg = txg; 6281 page_load = 0; 6282 } 6283 /* 6284 * If we are in pageout, we know that memory is already tight, 6285 * the arc is already going to be evicting, so we just want to 6286 * continue to let page writes occur as quickly as possible. 6287 */ 6288 if (curproc == pageproc) { 6289 if (page_load > MAX(ptob(minfree), available_memory) / 4) 6290 return (SET_ERROR(ERESTART)); 6291 /* Note: reserve is inflated, so we deflate */ 6292 page_load += reserve / 8; 6293 return (0); 6294 } else if (page_load > 0 && arc_reclaim_needed()) { 6295 /* memory is low, delay before restarting */ 6296 ARCSTAT_INCR(arcstat_memory_throttle_count, 1); 6297 return (SET_ERROR(EAGAIN)); 6298 } 6299 page_load = 0; 6300#endif 6301 return (0); 6302} 6303 6304void 6305arc_tempreserve_clear(uint64_t reserve) 6306{ 6307 atomic_add_64(&arc_tempreserve, -reserve); 6308 ASSERT((int64_t)arc_tempreserve >= 0); 6309} 6310 6311int 6312arc_tempreserve_space(uint64_t reserve, uint64_t txg) 6313{ 6314 int error; 6315 uint64_t anon_size; 6316 6317 if (reserve > arc_c/4 && !arc_no_grow) { 6318 arc_c = MIN(arc_c_max, reserve * 4); 6319 DTRACE_PROBE1(arc__set_reserve, uint64_t, arc_c); 6320 } 6321 if (reserve > arc_c) 6322 return (SET_ERROR(ENOMEM)); 6323 6324 /* 6325 * Don't count loaned bufs as in flight dirty data to prevent long 6326 * network delays from blocking transactions that are ready to be 6327 * assigned to a txg. 6328 */ 6329 6330 /* assert that it has not wrapped around */ 6331 ASSERT3S(atomic_add_64_nv(&arc_loaned_bytes, 0), >=, 0); 6332 6333 anon_size = MAX((int64_t)(refcount_count(&arc_anon->arcs_size) - 6334 arc_loaned_bytes), 0); 6335 6336 /* 6337 * Writes will, almost always, require additional memory allocations 6338 * in order to compress/encrypt/etc the data. We therefore need to 6339 * make sure that there is sufficient available memory for this. 6340 */ 6341 error = arc_memory_throttle(reserve, txg); 6342 if (error != 0) 6343 return (error); 6344 6345 /* 6346 * Throttle writes when the amount of dirty data in the cache 6347 * gets too large. We try to keep the cache less than half full 6348 * of dirty blocks so that our sync times don't grow too large. 6349 * Note: if two requests come in concurrently, we might let them 6350 * both succeed, when one of them should fail. Not a huge deal. 6351 */ 6352 6353 if (reserve + arc_tempreserve + anon_size > arc_c / 2 && 6354 anon_size > arc_c / 4) { 6355 uint64_t meta_esize = 6356 refcount_count(&arc_anon->arcs_esize[ARC_BUFC_METADATA]); 6357 uint64_t data_esize = 6358 refcount_count(&arc_anon->arcs_esize[ARC_BUFC_DATA]); 6359 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK " 6360 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n", 6361 arc_tempreserve >> 10, meta_esize >> 10, 6362 data_esize >> 10, reserve >> 10, arc_c >> 10); 6363 return (SET_ERROR(ERESTART)); 6364 } 6365 atomic_add_64(&arc_tempreserve, reserve); 6366 return (0); 6367} 6368 6369static void 6370arc_kstat_update_state(arc_state_t *state, kstat_named_t *size, 6371 kstat_named_t *evict_data, kstat_named_t *evict_metadata) 6372{ 6373 size->value.ui64 = refcount_count(&state->arcs_size); 6374 evict_data->value.ui64 = 6375 refcount_count(&state->arcs_esize[ARC_BUFC_DATA]); 6376 evict_metadata->value.ui64 = 6377 refcount_count(&state->arcs_esize[ARC_BUFC_METADATA]); 6378} 6379 6380static int 6381arc_kstat_update(kstat_t *ksp, int rw) 6382{ 6383 arc_stats_t *as = ksp->ks_data; 6384 6385 if (rw == KSTAT_WRITE) { 6386 return (EACCES); 6387 } else { 6388 arc_kstat_update_state(arc_anon, 6389 &as->arcstat_anon_size, 6390 &as->arcstat_anon_evictable_data, 6391 &as->arcstat_anon_evictable_metadata); 6392 arc_kstat_update_state(arc_mru, 6393 &as->arcstat_mru_size, 6394 &as->arcstat_mru_evictable_data, 6395 &as->arcstat_mru_evictable_metadata); 6396 arc_kstat_update_state(arc_mru_ghost, 6397 &as->arcstat_mru_ghost_size, 6398 &as->arcstat_mru_ghost_evictable_data, 6399 &as->arcstat_mru_ghost_evictable_metadata); 6400 arc_kstat_update_state(arc_mfu, 6401 &as->arcstat_mfu_size, 6402 &as->arcstat_mfu_evictable_data, 6403 &as->arcstat_mfu_evictable_metadata); 6404 arc_kstat_update_state(arc_mfu_ghost, 6405 &as->arcstat_mfu_ghost_size, 6406 &as->arcstat_mfu_ghost_evictable_data, 6407 &as->arcstat_mfu_ghost_evictable_metadata); 6408 6409 ARCSTAT(arcstat_size) = aggsum_value(&arc_size); 6410 ARCSTAT(arcstat_meta_used) = aggsum_value(&arc_meta_used); 6411 ARCSTAT(arcstat_data_size) = aggsum_value(&astat_data_size); 6412 ARCSTAT(arcstat_metadata_size) = 6413 aggsum_value(&astat_metadata_size); 6414 ARCSTAT(arcstat_hdr_size) = aggsum_value(&astat_hdr_size); 6415 ARCSTAT(arcstat_other_size) = aggsum_value(&astat_other_size); 6416 ARCSTAT(arcstat_l2_hdr_size) = aggsum_value(&astat_l2_hdr_size); 6417 } 6418 6419 return (0); 6420} 6421 6422/* 6423 * This function *must* return indices evenly distributed between all 6424 * sublists of the multilist. This is needed due to how the ARC eviction 6425 * code is laid out; arc_evict_state() assumes ARC buffers are evenly 6426 * distributed between all sublists and uses this assumption when 6427 * deciding which sublist to evict from and how much to evict from it. 6428 */ 6429unsigned int 6430arc_state_multilist_index_func(multilist_t *ml, void *obj) 6431{ 6432 arc_buf_hdr_t *hdr = obj; 6433 6434 /* 6435 * We rely on b_dva to generate evenly distributed index 6436 * numbers using buf_hash below. So, as an added precaution, 6437 * let's make sure we never add empty buffers to the arc lists. 6438 */ 6439 ASSERT(!HDR_EMPTY(hdr)); 6440 6441 /* 6442 * The assumption here, is the hash value for a given 6443 * arc_buf_hdr_t will remain constant throughout it's lifetime 6444 * (i.e. it's b_spa, b_dva, and b_birth fields don't change). 6445 * Thus, we don't need to store the header's sublist index 6446 * on insertion, as this index can be recalculated on removal. 6447 * 6448 * Also, the low order bits of the hash value are thought to be 6449 * distributed evenly. Otherwise, in the case that the multilist 6450 * has a power of two number of sublists, each sublists' usage 6451 * would not be evenly distributed. 6452 */ 6453 return (buf_hash(hdr->b_spa, &hdr->b_dva, hdr->b_birth) % 6454 multilist_get_num_sublists(ml)); 6455} 6456 6457#ifdef _KERNEL 6458static eventhandler_tag arc_event_lowmem = NULL; 6459 6460static void 6461arc_lowmem(void *arg __unused, int howto __unused) 6462{ 6463 6464 mutex_enter(&arc_reclaim_lock); 6465 DTRACE_PROBE1(arc__needfree, int64_t, ((int64_t)freemem - zfs_arc_free_target) * PAGESIZE); 6466 cv_signal(&arc_reclaim_thread_cv); 6467 6468 /* 6469 * It is unsafe to block here in arbitrary threads, because we can come 6470 * here from ARC itself and may hold ARC locks and thus risk a deadlock 6471 * with ARC reclaim thread. 6472 */ 6473 if (curproc == pageproc) 6474 (void) cv_wait(&arc_reclaim_waiters_cv, &arc_reclaim_lock); 6475 mutex_exit(&arc_reclaim_lock); 6476} 6477#endif 6478 6479static void 6480arc_state_init(void) 6481{ 6482 arc_anon = &ARC_anon; 6483 arc_mru = &ARC_mru; 6484 arc_mru_ghost = &ARC_mru_ghost; 6485 arc_mfu = &ARC_mfu; 6486 arc_mfu_ghost = &ARC_mfu_ghost; 6487 arc_l2c_only = &ARC_l2c_only; 6488 6489 arc_mru->arcs_list[ARC_BUFC_METADATA] = 6490 multilist_create(sizeof (arc_buf_hdr_t), 6491 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 6492 arc_state_multilist_index_func); 6493 arc_mru->arcs_list[ARC_BUFC_DATA] = 6494 multilist_create(sizeof (arc_buf_hdr_t), 6495 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 6496 arc_state_multilist_index_func); 6497 arc_mru_ghost->arcs_list[ARC_BUFC_METADATA] = 6498 multilist_create(sizeof (arc_buf_hdr_t), 6499 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 6500 arc_state_multilist_index_func); 6501 arc_mru_ghost->arcs_list[ARC_BUFC_DATA] = 6502 multilist_create(sizeof (arc_buf_hdr_t), 6503 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 6504 arc_state_multilist_index_func); 6505 arc_mfu->arcs_list[ARC_BUFC_METADATA] = 6506 multilist_create(sizeof (arc_buf_hdr_t), 6507 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 6508 arc_state_multilist_index_func); 6509 arc_mfu->arcs_list[ARC_BUFC_DATA] = 6510 multilist_create(sizeof (arc_buf_hdr_t), 6511 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 6512 arc_state_multilist_index_func); 6513 arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA] = 6514 multilist_create(sizeof (arc_buf_hdr_t), 6515 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 6516 arc_state_multilist_index_func); 6517 arc_mfu_ghost->arcs_list[ARC_BUFC_DATA] = 6518 multilist_create(sizeof (arc_buf_hdr_t), 6519 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 6520 arc_state_multilist_index_func); 6521 arc_l2c_only->arcs_list[ARC_BUFC_METADATA] = 6522 multilist_create(sizeof (arc_buf_hdr_t), 6523 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 6524 arc_state_multilist_index_func); 6525 arc_l2c_only->arcs_list[ARC_BUFC_DATA] = 6526 multilist_create(sizeof (arc_buf_hdr_t), 6527 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 6528 arc_state_multilist_index_func); 6529 6530 refcount_create(&arc_anon->arcs_esize[ARC_BUFC_METADATA]); 6531 refcount_create(&arc_anon->arcs_esize[ARC_BUFC_DATA]); 6532 refcount_create(&arc_mru->arcs_esize[ARC_BUFC_METADATA]); 6533 refcount_create(&arc_mru->arcs_esize[ARC_BUFC_DATA]); 6534 refcount_create(&arc_mru_ghost->arcs_esize[ARC_BUFC_METADATA]); 6535 refcount_create(&arc_mru_ghost->arcs_esize[ARC_BUFC_DATA]); 6536 refcount_create(&arc_mfu->arcs_esize[ARC_BUFC_METADATA]); 6537 refcount_create(&arc_mfu->arcs_esize[ARC_BUFC_DATA]); 6538 refcount_create(&arc_mfu_ghost->arcs_esize[ARC_BUFC_METADATA]); 6539 refcount_create(&arc_mfu_ghost->arcs_esize[ARC_BUFC_DATA]); 6540 refcount_create(&arc_l2c_only->arcs_esize[ARC_BUFC_METADATA]); 6541 refcount_create(&arc_l2c_only->arcs_esize[ARC_BUFC_DATA]); 6542 6543 refcount_create(&arc_anon->arcs_size); 6544 refcount_create(&arc_mru->arcs_size); 6545 refcount_create(&arc_mru_ghost->arcs_size); 6546 refcount_create(&arc_mfu->arcs_size); 6547 refcount_create(&arc_mfu_ghost->arcs_size); 6548 refcount_create(&arc_l2c_only->arcs_size); 6549 6550 aggsum_init(&arc_meta_used, 0); 6551 aggsum_init(&arc_size, 0); 6552 aggsum_init(&astat_data_size, 0); 6553 aggsum_init(&astat_metadata_size, 0); 6554 aggsum_init(&astat_hdr_size, 0); 6555 aggsum_init(&astat_other_size, 0); 6556 aggsum_init(&astat_l2_hdr_size, 0); 6557} 6558 6559static void 6560arc_state_fini(void) 6561{ 6562 refcount_destroy(&arc_anon->arcs_esize[ARC_BUFC_METADATA]); 6563 refcount_destroy(&arc_anon->arcs_esize[ARC_BUFC_DATA]); 6564 refcount_destroy(&arc_mru->arcs_esize[ARC_BUFC_METADATA]); 6565 refcount_destroy(&arc_mru->arcs_esize[ARC_BUFC_DATA]); 6566 refcount_destroy(&arc_mru_ghost->arcs_esize[ARC_BUFC_METADATA]); 6567 refcount_destroy(&arc_mru_ghost->arcs_esize[ARC_BUFC_DATA]); 6568 refcount_destroy(&arc_mfu->arcs_esize[ARC_BUFC_METADATA]); 6569 refcount_destroy(&arc_mfu->arcs_esize[ARC_BUFC_DATA]); 6570 refcount_destroy(&arc_mfu_ghost->arcs_esize[ARC_BUFC_METADATA]); 6571 refcount_destroy(&arc_mfu_ghost->arcs_esize[ARC_BUFC_DATA]); 6572 refcount_destroy(&arc_l2c_only->arcs_esize[ARC_BUFC_METADATA]); 6573 refcount_destroy(&arc_l2c_only->arcs_esize[ARC_BUFC_DATA]); 6574 6575 refcount_destroy(&arc_anon->arcs_size); 6576 refcount_destroy(&arc_mru->arcs_size); 6577 refcount_destroy(&arc_mru_ghost->arcs_size); 6578 refcount_destroy(&arc_mfu->arcs_size); 6579 refcount_destroy(&arc_mfu_ghost->arcs_size); 6580 refcount_destroy(&arc_l2c_only->arcs_size); 6581 6582 multilist_destroy(arc_mru->arcs_list[ARC_BUFC_METADATA]); 6583 multilist_destroy(arc_mru_ghost->arcs_list[ARC_BUFC_METADATA]); 6584 multilist_destroy(arc_mfu->arcs_list[ARC_BUFC_METADATA]); 6585 multilist_destroy(arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA]); 6586 multilist_destroy(arc_mru->arcs_list[ARC_BUFC_DATA]); 6587 multilist_destroy(arc_mru_ghost->arcs_list[ARC_BUFC_DATA]); 6588 multilist_destroy(arc_mfu->arcs_list[ARC_BUFC_DATA]); 6589 multilist_destroy(arc_mfu_ghost->arcs_list[ARC_BUFC_DATA]); 6590} 6591 6592uint64_t 6593arc_max_bytes(void) 6594{ 6595 return (arc_c_max); 6596} 6597 6598void 6599arc_init(void) 6600{ 6601 int i, prefetch_tunable_set = 0; 6602 6603 /* 6604 * allmem is "all memory that we could possibly use". 6605 */ 6606#ifdef illumos 6607#ifdef _KERNEL 6608 uint64_t allmem = ptob(physmem - swapfs_minfree); 6609#else 6610 uint64_t allmem = (physmem * PAGESIZE) / 2; 6611#endif 6612#else 6613 uint64_t allmem = kmem_size(); 6614#endif 6615 6616 6617 mutex_init(&arc_reclaim_lock, NULL, MUTEX_DEFAULT, NULL); 6618 cv_init(&arc_reclaim_thread_cv, NULL, CV_DEFAULT, NULL); 6619 cv_init(&arc_reclaim_waiters_cv, NULL, CV_DEFAULT, NULL); 6620 6621 mutex_init(&arc_dnlc_evicts_lock, NULL, MUTEX_DEFAULT, NULL); 6622 cv_init(&arc_dnlc_evicts_cv, NULL, CV_DEFAULT, NULL); 6623 6624 /* Convert seconds to clock ticks */ 6625 arc_min_prefetch_lifespan = 1 * hz; 6626 6627 /* set min cache to 1/32 of all memory, or arc_abs_min, whichever is more */ 6628 arc_c_min = MAX(allmem / 32, arc_abs_min); 6629 /* set max to 5/8 of all memory, or all but 1GB, whichever is more */ 6630 if (allmem >= 1 << 30) 6631 arc_c_max = allmem - (1 << 30); 6632 else 6633 arc_c_max = arc_c_min; 6634 arc_c_max = MAX(allmem * 5 / 8, arc_c_max); 6635 6636 /* 6637 * In userland, there's only the memory pressure that we artificially 6638 * create (see arc_available_memory()). Don't let arc_c get too 6639 * small, because it can cause transactions to be larger than 6640 * arc_c, causing arc_tempreserve_space() to fail. 6641 */ 6642#ifndef _KERNEL 6643 arc_c_min = arc_c_max / 2; 6644#endif 6645 6646#ifdef _KERNEL 6647 /* 6648 * Allow the tunables to override our calculations if they are 6649 * reasonable. 6650 */ 6651 if (zfs_arc_max > arc_abs_min && zfs_arc_max < allmem) { 6652 arc_c_max = zfs_arc_max; 6653 arc_c_min = MIN(arc_c_min, arc_c_max); 6654 } 6655 if (zfs_arc_min > arc_abs_min && zfs_arc_min <= arc_c_max) 6656 arc_c_min = zfs_arc_min; 6657#endif 6658 6659 arc_c = arc_c_max; 6660 arc_p = (arc_c >> 1); 6661 6662 /* limit meta-data to 1/4 of the arc capacity */ 6663 arc_meta_limit = arc_c_max / 4; 6664 6665#ifdef _KERNEL 6666 /* 6667 * Metadata is stored in the kernel's heap. Don't let us 6668 * use more than half the heap for the ARC. 6669 */ 6670 arc_meta_limit = MIN(arc_meta_limit, 6671 vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 2); 6672#endif 6673 6674 /* Allow the tunable to override if it is reasonable */ 6675 if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max) 6676 arc_meta_limit = zfs_arc_meta_limit; 6677 6678 if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0) 6679 arc_c_min = arc_meta_limit / 2; 6680 6681 if (zfs_arc_meta_min > 0) { 6682 arc_meta_min = zfs_arc_meta_min; 6683 } else { 6684 arc_meta_min = arc_c_min / 2; 6685 } 6686 6687 if (zfs_arc_grow_retry > 0) 6688 arc_grow_retry = zfs_arc_grow_retry; 6689 6690 if (zfs_arc_shrink_shift > 0) 6691 arc_shrink_shift = zfs_arc_shrink_shift; 6692 6693 if (zfs_arc_no_grow_shift > 0) 6694 arc_no_grow_shift = zfs_arc_no_grow_shift; 6695 /* 6696 * Ensure that arc_no_grow_shift is less than arc_shrink_shift. 6697 */ 6698 if (arc_no_grow_shift >= arc_shrink_shift) 6699 arc_no_grow_shift = arc_shrink_shift - 1; 6700 6701 if (zfs_arc_p_min_shift > 0) 6702 arc_p_min_shift = zfs_arc_p_min_shift; 6703 6704 /* if kmem_flags are set, lets try to use less memory */ 6705 if (kmem_debugging()) 6706 arc_c = arc_c / 2; 6707 if (arc_c < arc_c_min) 6708 arc_c = arc_c_min; 6709 6710 zfs_arc_min = arc_c_min; 6711 zfs_arc_max = arc_c_max; 6712 6713 arc_state_init(); 6714 buf_init(); 6715 6716 arc_reclaim_thread_exit = B_FALSE; 6717 arc_dnlc_evicts_thread_exit = FALSE; 6718 6719 arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED, 6720 sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL); 6721 6722 if (arc_ksp != NULL) { 6723 arc_ksp->ks_data = &arc_stats; 6724 arc_ksp->ks_update = arc_kstat_update; 6725 kstat_install(arc_ksp); 6726 } 6727 6728 (void) thread_create(NULL, 0, arc_reclaim_thread, NULL, 0, &p0, 6729 TS_RUN, minclsyspri); 6730 6731#ifdef _KERNEL 6732 arc_event_lowmem = EVENTHANDLER_REGISTER(vm_lowmem, arc_lowmem, NULL, 6733 EVENTHANDLER_PRI_FIRST); 6734#endif 6735 6736 (void) thread_create(NULL, 0, arc_dnlc_evicts_thread, NULL, 0, &p0, 6737 TS_RUN, minclsyspri); 6738 6739 arc_dead = B_FALSE; 6740 arc_warm = B_FALSE; 6741 6742 /* 6743 * Calculate maximum amount of dirty data per pool. 6744 * 6745 * If it has been set by /etc/system, take that. 6746 * Otherwise, use a percentage of physical memory defined by 6747 * zfs_dirty_data_max_percent (default 10%) with a cap at 6748 * zfs_dirty_data_max_max (default 4GB). 6749 */ 6750 if (zfs_dirty_data_max == 0) { 6751 zfs_dirty_data_max = ptob(physmem) * 6752 zfs_dirty_data_max_percent / 100; 6753 zfs_dirty_data_max = MIN(zfs_dirty_data_max, 6754 zfs_dirty_data_max_max); 6755 } 6756 6757#ifdef _KERNEL 6758 if (TUNABLE_INT_FETCH("vfs.zfs.prefetch_disable", &zfs_prefetch_disable)) 6759 prefetch_tunable_set = 1; 6760 6761#ifdef __i386__ 6762 if (prefetch_tunable_set == 0) { 6763 printf("ZFS NOTICE: Prefetch is disabled by default on i386 " 6764 "-- to enable,\n"); 6765 printf(" add \"vfs.zfs.prefetch_disable=0\" " 6766 "to /boot/loader.conf.\n"); 6767 zfs_prefetch_disable = 1; 6768 } 6769#else 6770 if ((((uint64_t)physmem * PAGESIZE) < (1ULL << 32)) && 6771 prefetch_tunable_set == 0) { 6772 printf("ZFS NOTICE: Prefetch is disabled by default if less " 6773 "than 4GB of RAM is present;\n" 6774 " to enable, add \"vfs.zfs.prefetch_disable=0\" " 6775 "to /boot/loader.conf.\n"); 6776 zfs_prefetch_disable = 1; 6777 } 6778#endif 6779 /* Warn about ZFS memory and address space requirements. */ 6780 if (((uint64_t)physmem * PAGESIZE) < (256 + 128 + 64) * (1 << 20)) { 6781 printf("ZFS WARNING: Recommended minimum RAM size is 512MB; " 6782 "expect unstable behavior.\n"); 6783 } 6784 if (allmem < 512 * (1 << 20)) { 6785 printf("ZFS WARNING: Recommended minimum kmem_size is 512MB; " 6786 "expect unstable behavior.\n"); 6787 printf(" Consider tuning vm.kmem_size and " 6788 "vm.kmem_size_max\n"); 6789 printf(" in /boot/loader.conf.\n"); 6790 } 6791#endif 6792} 6793 6794void 6795arc_fini(void) 6796{ 6797#ifdef _KERNEL 6798 if (arc_event_lowmem != NULL) 6799 EVENTHANDLER_DEREGISTER(vm_lowmem, arc_event_lowmem); 6800#endif 6801 6802 mutex_enter(&arc_reclaim_lock); 6803 arc_reclaim_thread_exit = B_TRUE; 6804 /* 6805 * The reclaim thread will set arc_reclaim_thread_exit back to 6806 * B_FALSE when it is finished exiting; we're waiting for that. 6807 */ 6808 while (arc_reclaim_thread_exit) { 6809 cv_signal(&arc_reclaim_thread_cv); 6810 cv_wait(&arc_reclaim_thread_cv, &arc_reclaim_lock); 6811 } 6812 mutex_exit(&arc_reclaim_lock); 6813 6814 /* Use B_TRUE to ensure *all* buffers are evicted */ 6815 arc_flush(NULL, B_TRUE); 6816 6817 mutex_enter(&arc_dnlc_evicts_lock); 6818 arc_dnlc_evicts_thread_exit = TRUE; 6819 /* 6820 * The user evicts thread will set arc_user_evicts_thread_exit 6821 * to FALSE when it is finished exiting; we're waiting for that. 6822 */ 6823 while (arc_dnlc_evicts_thread_exit) { 6824 cv_signal(&arc_dnlc_evicts_cv); 6825 cv_wait(&arc_dnlc_evicts_cv, &arc_dnlc_evicts_lock); 6826 } 6827 mutex_exit(&arc_dnlc_evicts_lock); 6828 6829 arc_dead = B_TRUE; 6830 6831 if (arc_ksp != NULL) { 6832 kstat_delete(arc_ksp); 6833 arc_ksp = NULL; 6834 } 6835 6836 mutex_destroy(&arc_reclaim_lock); 6837 cv_destroy(&arc_reclaim_thread_cv); 6838 cv_destroy(&arc_reclaim_waiters_cv); 6839 6840 mutex_destroy(&arc_dnlc_evicts_lock); 6841 cv_destroy(&arc_dnlc_evicts_cv); 6842 6843 arc_state_fini(); 6844 buf_fini(); 6845 6846 ASSERT0(arc_loaned_bytes); 6847} 6848 6849/* 6850 * Level 2 ARC 6851 * 6852 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk. 6853 * It uses dedicated storage devices to hold cached data, which are populated 6854 * using large infrequent writes. The main role of this cache is to boost 6855 * the performance of random read workloads. The intended L2ARC devices 6856 * include short-stroked disks, solid state disks, and other media with 6857 * substantially faster read latency than disk. 6858 * 6859 * +-----------------------+ 6860 * | ARC | 6861 * +-----------------------+ 6862 * | ^ ^ 6863 * | | | 6864 * l2arc_feed_thread() arc_read() 6865 * | | | 6866 * | l2arc read | 6867 * V | | 6868 * +---------------+ | 6869 * | L2ARC | | 6870 * +---------------+ | 6871 * | ^ | 6872 * l2arc_write() | | 6873 * | | | 6874 * V | | 6875 * +-------+ +-------+ 6876 * | vdev | | vdev | 6877 * | cache | | cache | 6878 * +-------+ +-------+ 6879 * +=========+ .-----. 6880 * : L2ARC : |-_____-| 6881 * : devices : | Disks | 6882 * +=========+ `-_____-' 6883 * 6884 * Read requests are satisfied from the following sources, in order: 6885 * 6886 * 1) ARC 6887 * 2) vdev cache of L2ARC devices 6888 * 3) L2ARC devices 6889 * 4) vdev cache of disks 6890 * 5) disks 6891 * 6892 * Some L2ARC device types exhibit extremely slow write performance. 6893 * To accommodate for this there are some significant differences between 6894 * the L2ARC and traditional cache design: 6895 * 6896 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from 6897 * the ARC behave as usual, freeing buffers and placing headers on ghost 6898 * lists. The ARC does not send buffers to the L2ARC during eviction as 6899 * this would add inflated write latencies for all ARC memory pressure. 6900 * 6901 * 2. The L2ARC attempts to cache data from the ARC before it is evicted. 6902 * It does this by periodically scanning buffers from the eviction-end of 6903 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are 6904 * not already there. It scans until a headroom of buffers is satisfied, 6905 * which itself is a buffer for ARC eviction. If a compressible buffer is 6906 * found during scanning and selected for writing to an L2ARC device, we 6907 * temporarily boost scanning headroom during the next scan cycle to make 6908 * sure we adapt to compression effects (which might significantly reduce 6909 * the data volume we write to L2ARC). The thread that does this is 6910 * l2arc_feed_thread(), illustrated below; example sizes are included to 6911 * provide a better sense of ratio than this diagram: 6912 * 6913 * head --> tail 6914 * +---------------------+----------+ 6915 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC 6916 * +---------------------+----------+ | o L2ARC eligible 6917 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer 6918 * +---------------------+----------+ | 6919 * 15.9 Gbytes ^ 32 Mbytes | 6920 * headroom | 6921 * l2arc_feed_thread() 6922 * | 6923 * l2arc write hand <--[oooo]--' 6924 * | 8 Mbyte 6925 * | write max 6926 * V 6927 * +==============================+ 6928 * L2ARC dev |####|#|###|###| |####| ... | 6929 * +==============================+ 6930 * 32 Gbytes 6931 * 6932 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of 6933 * evicted, then the L2ARC has cached a buffer much sooner than it probably 6934 * needed to, potentially wasting L2ARC device bandwidth and storage. It is 6935 * safe to say that this is an uncommon case, since buffers at the end of 6936 * the ARC lists have moved there due to inactivity. 6937 * 6938 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom, 6939 * then the L2ARC simply misses copying some buffers. This serves as a 6940 * pressure valve to prevent heavy read workloads from both stalling the ARC 6941 * with waits and clogging the L2ARC with writes. This also helps prevent 6942 * the potential for the L2ARC to churn if it attempts to cache content too 6943 * quickly, such as during backups of the entire pool. 6944 * 6945 * 5. After system boot and before the ARC has filled main memory, there are 6946 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru 6947 * lists can remain mostly static. Instead of searching from tail of these 6948 * lists as pictured, the l2arc_feed_thread() will search from the list heads 6949 * for eligible buffers, greatly increasing its chance of finding them. 6950 * 6951 * The L2ARC device write speed is also boosted during this time so that 6952 * the L2ARC warms up faster. Since there have been no ARC evictions yet, 6953 * there are no L2ARC reads, and no fear of degrading read performance 6954 * through increased writes. 6955 * 6956 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that 6957 * the vdev queue can aggregate them into larger and fewer writes. Each 6958 * device is written to in a rotor fashion, sweeping writes through 6959 * available space then repeating. 6960 * 6961 * 7. The L2ARC does not store dirty content. It never needs to flush 6962 * write buffers back to disk based storage. 6963 * 6964 * 8. If an ARC buffer is written (and dirtied) which also exists in the 6965 * L2ARC, the now stale L2ARC buffer is immediately dropped. 6966 * 6967 * The performance of the L2ARC can be tweaked by a number of tunables, which 6968 * may be necessary for different workloads: 6969 * 6970 * l2arc_write_max max write bytes per interval 6971 * l2arc_write_boost extra write bytes during device warmup 6972 * l2arc_noprefetch skip caching prefetched buffers 6973 * l2arc_headroom number of max device writes to precache 6974 * l2arc_headroom_boost when we find compressed buffers during ARC 6975 * scanning, we multiply headroom by this 6976 * percentage factor for the next scan cycle, 6977 * since more compressed buffers are likely to 6978 * be present 6979 * l2arc_feed_secs seconds between L2ARC writing 6980 * 6981 * Tunables may be removed or added as future performance improvements are 6982 * integrated, and also may become zpool properties. 6983 * 6984 * There are three key functions that control how the L2ARC warms up: 6985 * 6986 * l2arc_write_eligible() check if a buffer is eligible to cache 6987 * l2arc_write_size() calculate how much to write 6988 * l2arc_write_interval() calculate sleep delay between writes 6989 * 6990 * These three functions determine what to write, how much, and how quickly 6991 * to send writes. 6992 */ 6993 6994static boolean_t 6995l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *hdr) 6996{ 6997 /* 6998 * A buffer is *not* eligible for the L2ARC if it: 6999 * 1. belongs to a different spa. 7000 * 2. is already cached on the L2ARC. 7001 * 3. has an I/O in progress (it may be an incomplete read). 7002 * 4. is flagged not eligible (zfs property). 7003 */ 7004 if (hdr->b_spa != spa_guid) { 7005 ARCSTAT_BUMP(arcstat_l2_write_spa_mismatch); 7006 return (B_FALSE); 7007 } 7008 if (HDR_HAS_L2HDR(hdr)) { 7009 ARCSTAT_BUMP(arcstat_l2_write_in_l2); 7010 return (B_FALSE); 7011 } 7012 if (HDR_IO_IN_PROGRESS(hdr)) { 7013 ARCSTAT_BUMP(arcstat_l2_write_hdr_io_in_progress); 7014 return (B_FALSE); 7015 } 7016 if (!HDR_L2CACHE(hdr)) { 7017 ARCSTAT_BUMP(arcstat_l2_write_not_cacheable); 7018 return (B_FALSE); 7019 } 7020 7021 return (B_TRUE); 7022} 7023 7024static uint64_t 7025l2arc_write_size(void) 7026{ 7027 uint64_t size; 7028 7029 /* 7030 * Make sure our globals have meaningful values in case the user 7031 * altered them. 7032 */ 7033 size = l2arc_write_max; 7034 if (size == 0) { 7035 cmn_err(CE_NOTE, "Bad value for l2arc_write_max, value must " 7036 "be greater than zero, resetting it to the default (%d)", 7037 L2ARC_WRITE_SIZE); 7038 size = l2arc_write_max = L2ARC_WRITE_SIZE; 7039 } 7040 7041 if (arc_warm == B_FALSE) 7042 size += l2arc_write_boost; 7043 7044 return (size); 7045 7046} 7047 7048static clock_t 7049l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote) 7050{ 7051 clock_t interval, next, now; 7052 7053 /* 7054 * If the ARC lists are busy, increase our write rate; if the 7055 * lists are stale, idle back. This is achieved by checking 7056 * how much we previously wrote - if it was more than half of 7057 * what we wanted, schedule the next write much sooner. 7058 */ 7059 if (l2arc_feed_again && wrote > (wanted / 2)) 7060 interval = (hz * l2arc_feed_min_ms) / 1000; 7061 else 7062 interval = hz * l2arc_feed_secs; 7063 7064 now = ddi_get_lbolt(); 7065 next = MAX(now, MIN(now + interval, began + interval)); 7066 7067 return (next); 7068} 7069 7070/* 7071 * Cycle through L2ARC devices. This is how L2ARC load balances. 7072 * If a device is returned, this also returns holding the spa config lock. 7073 */ 7074static l2arc_dev_t * 7075l2arc_dev_get_next(void) 7076{ 7077 l2arc_dev_t *first, *next = NULL; 7078 7079 /* 7080 * Lock out the removal of spas (spa_namespace_lock), then removal 7081 * of cache devices (l2arc_dev_mtx). Once a device has been selected, 7082 * both locks will be dropped and a spa config lock held instead. 7083 */ 7084 mutex_enter(&spa_namespace_lock); 7085 mutex_enter(&l2arc_dev_mtx); 7086 7087 /* if there are no vdevs, there is nothing to do */ 7088 if (l2arc_ndev == 0) 7089 goto out; 7090 7091 first = NULL; 7092 next = l2arc_dev_last; 7093 do { 7094 /* loop around the list looking for a non-faulted vdev */ 7095 if (next == NULL) { 7096 next = list_head(l2arc_dev_list); 7097 } else { 7098 next = list_next(l2arc_dev_list, next); 7099 if (next == NULL) 7100 next = list_head(l2arc_dev_list); 7101 } 7102 7103 /* if we have come back to the start, bail out */ 7104 if (first == NULL) 7105 first = next; 7106 else if (next == first) 7107 break; 7108 7109 } while (vdev_is_dead(next->l2ad_vdev)); 7110 7111 /* if we were unable to find any usable vdevs, return NULL */ 7112 if (vdev_is_dead(next->l2ad_vdev)) 7113 next = NULL; 7114 7115 l2arc_dev_last = next; 7116 7117out: 7118 mutex_exit(&l2arc_dev_mtx); 7119 7120 /* 7121 * Grab the config lock to prevent the 'next' device from being 7122 * removed while we are writing to it. 7123 */ 7124 if (next != NULL) 7125 spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER); 7126 mutex_exit(&spa_namespace_lock); 7127 7128 return (next); 7129} 7130 7131/* 7132 * Free buffers that were tagged for destruction. 7133 */ 7134static void 7135l2arc_do_free_on_write() 7136{ 7137 list_t *buflist; 7138 l2arc_data_free_t *df, *df_prev; 7139 7140 mutex_enter(&l2arc_free_on_write_mtx); 7141 buflist = l2arc_free_on_write; 7142 7143 for (df = list_tail(buflist); df; df = df_prev) { 7144 df_prev = list_prev(buflist, df); 7145 ASSERT3P(df->l2df_abd, !=, NULL); 7146 abd_free(df->l2df_abd); 7147 list_remove(buflist, df); 7148 kmem_free(df, sizeof (l2arc_data_free_t)); 7149 } 7150 7151 mutex_exit(&l2arc_free_on_write_mtx); 7152} 7153 7154/* 7155 * A write to a cache device has completed. Update all headers to allow 7156 * reads from these buffers to begin. 7157 */ 7158static void 7159l2arc_write_done(zio_t *zio) 7160{ 7161 l2arc_write_callback_t *cb; 7162 l2arc_dev_t *dev; 7163 list_t *buflist; 7164 arc_buf_hdr_t *head, *hdr, *hdr_prev; 7165 kmutex_t *hash_lock; 7166 int64_t bytes_dropped = 0; 7167 7168 cb = zio->io_private; 7169 ASSERT3P(cb, !=, NULL); 7170 dev = cb->l2wcb_dev; 7171 ASSERT3P(dev, !=, NULL); 7172 head = cb->l2wcb_head; 7173 ASSERT3P(head, !=, NULL); 7174 buflist = &dev->l2ad_buflist; 7175 ASSERT3P(buflist, !=, NULL); 7176 DTRACE_PROBE2(l2arc__iodone, zio_t *, zio, 7177 l2arc_write_callback_t *, cb); 7178 7179 if (zio->io_error != 0) 7180 ARCSTAT_BUMP(arcstat_l2_writes_error); 7181 7182 /* 7183 * All writes completed, or an error was hit. 7184 */ 7185top: 7186 mutex_enter(&dev->l2ad_mtx); 7187 for (hdr = list_prev(buflist, head); hdr; hdr = hdr_prev) { 7188 hdr_prev = list_prev(buflist, hdr); 7189 7190 hash_lock = HDR_LOCK(hdr); 7191 7192 /* 7193 * We cannot use mutex_enter or else we can deadlock 7194 * with l2arc_write_buffers (due to swapping the order 7195 * the hash lock and l2ad_mtx are taken). 7196 */ 7197 if (!mutex_tryenter(hash_lock)) { 7198 /* 7199 * Missed the hash lock. We must retry so we 7200 * don't leave the ARC_FLAG_L2_WRITING bit set. 7201 */ 7202 ARCSTAT_BUMP(arcstat_l2_writes_lock_retry); 7203 7204 /* 7205 * We don't want to rescan the headers we've 7206 * already marked as having been written out, so 7207 * we reinsert the head node so we can pick up 7208 * where we left off. 7209 */ 7210 list_remove(buflist, head); 7211 list_insert_after(buflist, hdr, head); 7212 7213 mutex_exit(&dev->l2ad_mtx); 7214 7215 /* 7216 * We wait for the hash lock to become available 7217 * to try and prevent busy waiting, and increase 7218 * the chance we'll be able to acquire the lock 7219 * the next time around. 7220 */ 7221 mutex_enter(hash_lock); 7222 mutex_exit(hash_lock); 7223 goto top; 7224 } 7225 7226 /* 7227 * We could not have been moved into the arc_l2c_only 7228 * state while in-flight due to our ARC_FLAG_L2_WRITING 7229 * bit being set. Let's just ensure that's being enforced. 7230 */ 7231 ASSERT(HDR_HAS_L1HDR(hdr)); 7232 7233 if (zio->io_error != 0) { 7234 /* 7235 * Error - drop L2ARC entry. 7236 */ 7237 list_remove(buflist, hdr); 7238 l2arc_trim(hdr); 7239 arc_hdr_clear_flags(hdr, ARC_FLAG_HAS_L2HDR); 7240 7241 ARCSTAT_INCR(arcstat_l2_psize, -arc_hdr_size(hdr)); 7242 ARCSTAT_INCR(arcstat_l2_lsize, -HDR_GET_LSIZE(hdr)); 7243 7244 bytes_dropped += arc_hdr_size(hdr); 7245 (void) refcount_remove_many(&dev->l2ad_alloc, 7246 arc_hdr_size(hdr), hdr); 7247 } 7248 7249 /* 7250 * Allow ARC to begin reads and ghost list evictions to 7251 * this L2ARC entry. 7252 */ 7253 arc_hdr_clear_flags(hdr, ARC_FLAG_L2_WRITING); 7254 7255 mutex_exit(hash_lock); 7256 } 7257 7258 atomic_inc_64(&l2arc_writes_done); 7259 list_remove(buflist, head); 7260 ASSERT(!HDR_HAS_L1HDR(head)); 7261 kmem_cache_free(hdr_l2only_cache, head); 7262 mutex_exit(&dev->l2ad_mtx); 7263 7264 vdev_space_update(dev->l2ad_vdev, -bytes_dropped, 0, 0); 7265 7266 l2arc_do_free_on_write(); 7267 7268 kmem_free(cb, sizeof (l2arc_write_callback_t)); 7269} 7270 7271/* 7272 * A read to a cache device completed. Validate buffer contents before 7273 * handing over to the regular ARC routines. 7274 */ 7275static void 7276l2arc_read_done(zio_t *zio) 7277{ 7278 l2arc_read_callback_t *cb; 7279 arc_buf_hdr_t *hdr; 7280 kmutex_t *hash_lock; 7281 boolean_t valid_cksum; 7282 7283 ASSERT3P(zio->io_vd, !=, NULL); 7284 ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE); 7285 7286 spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd); 7287 7288 cb = zio->io_private; 7289 ASSERT3P(cb, !=, NULL); 7290 hdr = cb->l2rcb_hdr; 7291 ASSERT3P(hdr, !=, NULL); 7292 7293 hash_lock = HDR_LOCK(hdr); 7294 mutex_enter(hash_lock); 7295 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 7296 7297 /* 7298 * If the data was read into a temporary buffer, 7299 * move it and free the buffer. 7300 */ 7301 if (cb->l2rcb_abd != NULL) { 7302 ASSERT3U(arc_hdr_size(hdr), <, zio->io_size); 7303 if (zio->io_error == 0) { 7304 abd_copy(hdr->b_l1hdr.b_pabd, cb->l2rcb_abd, 7305 arc_hdr_size(hdr)); 7306 } 7307 7308 /* 7309 * The following must be done regardless of whether 7310 * there was an error: 7311 * - free the temporary buffer 7312 * - point zio to the real ARC buffer 7313 * - set zio size accordingly 7314 * These are required because zio is either re-used for 7315 * an I/O of the block in the case of the error 7316 * or the zio is passed to arc_read_done() and it 7317 * needs real data. 7318 */ 7319 abd_free(cb->l2rcb_abd); 7320 zio->io_size = zio->io_orig_size = arc_hdr_size(hdr); 7321 zio->io_abd = zio->io_orig_abd = hdr->b_l1hdr.b_pabd; 7322 } 7323 7324 ASSERT3P(zio->io_abd, !=, NULL); 7325 7326 /* 7327 * Check this survived the L2ARC journey. 7328 */ 7329 ASSERT3P(zio->io_abd, ==, hdr->b_l1hdr.b_pabd); 7330 zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */ 7331 zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */ 7332 7333 valid_cksum = arc_cksum_is_equal(hdr, zio); 7334 if (valid_cksum && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) { 7335 mutex_exit(hash_lock); 7336 zio->io_private = hdr; 7337 arc_read_done(zio); 7338 } else { 7339 mutex_exit(hash_lock); 7340 /* 7341 * Buffer didn't survive caching. Increment stats and 7342 * reissue to the original storage device. 7343 */ 7344 if (zio->io_error != 0) { 7345 ARCSTAT_BUMP(arcstat_l2_io_error); 7346 } else { 7347 zio->io_error = SET_ERROR(EIO); 7348 } 7349 if (!valid_cksum) 7350 ARCSTAT_BUMP(arcstat_l2_cksum_bad); 7351 7352 /* 7353 * If there's no waiter, issue an async i/o to the primary 7354 * storage now. If there *is* a waiter, the caller must 7355 * issue the i/o in a context where it's OK to block. 7356 */ 7357 if (zio->io_waiter == NULL) { 7358 zio_t *pio = zio_unique_parent(zio); 7359 7360 ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL); 7361 7362 zio_nowait(zio_read(pio, zio->io_spa, zio->io_bp, 7363 hdr->b_l1hdr.b_pabd, zio->io_size, arc_read_done, 7364 hdr, zio->io_priority, cb->l2rcb_flags, 7365 &cb->l2rcb_zb)); 7366 } 7367 } 7368 7369 kmem_free(cb, sizeof (l2arc_read_callback_t)); 7370} 7371 7372/* 7373 * This is the list priority from which the L2ARC will search for pages to 7374 * cache. This is used within loops (0..3) to cycle through lists in the 7375 * desired order. This order can have a significant effect on cache 7376 * performance. 7377 * 7378 * Currently the metadata lists are hit first, MFU then MRU, followed by 7379 * the data lists. This function returns a locked list, and also returns 7380 * the lock pointer. 7381 */ 7382static multilist_sublist_t * 7383l2arc_sublist_lock(int list_num) 7384{ 7385 multilist_t *ml = NULL; 7386 unsigned int idx; 7387 7388 ASSERT(list_num >= 0 && list_num <= 3); 7389 7390 switch (list_num) { 7391 case 0: 7392 ml = arc_mfu->arcs_list[ARC_BUFC_METADATA]; 7393 break; 7394 case 1: 7395 ml = arc_mru->arcs_list[ARC_BUFC_METADATA]; 7396 break; 7397 case 2: 7398 ml = arc_mfu->arcs_list[ARC_BUFC_DATA]; 7399 break; 7400 case 3: 7401 ml = arc_mru->arcs_list[ARC_BUFC_DATA]; 7402 break; 7403 } 7404 7405 /* 7406 * Return a randomly-selected sublist. This is acceptable 7407 * because the caller feeds only a little bit of data for each 7408 * call (8MB). Subsequent calls will result in different 7409 * sublists being selected. 7410 */ 7411 idx = multilist_get_random_index(ml); 7412 return (multilist_sublist_lock(ml, idx)); 7413} 7414 7415/* 7416 * Evict buffers from the device write hand to the distance specified in 7417 * bytes. This distance may span populated buffers, it may span nothing. 7418 * This is clearing a region on the L2ARC device ready for writing. 7419 * If the 'all' boolean is set, every buffer is evicted. 7420 */ 7421static void 7422l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all) 7423{ 7424 list_t *buflist; 7425 arc_buf_hdr_t *hdr, *hdr_prev; 7426 kmutex_t *hash_lock; 7427 uint64_t taddr; 7428 7429 buflist = &dev->l2ad_buflist; 7430 7431 if (!all && dev->l2ad_first) { 7432 /* 7433 * This is the first sweep through the device. There is 7434 * nothing to evict. 7435 */ 7436 return; 7437 } 7438 7439 if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) { 7440 /* 7441 * When nearing the end of the device, evict to the end 7442 * before the device write hand jumps to the start. 7443 */ 7444 taddr = dev->l2ad_end; 7445 } else { 7446 taddr = dev->l2ad_hand + distance; 7447 } 7448 DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist, 7449 uint64_t, taddr, boolean_t, all); 7450 7451top: 7452 mutex_enter(&dev->l2ad_mtx); 7453 for (hdr = list_tail(buflist); hdr; hdr = hdr_prev) { 7454 hdr_prev = list_prev(buflist, hdr); 7455 7456 hash_lock = HDR_LOCK(hdr); 7457 7458 /* 7459 * We cannot use mutex_enter or else we can deadlock 7460 * with l2arc_write_buffers (due to swapping the order 7461 * the hash lock and l2ad_mtx are taken). 7462 */ 7463 if (!mutex_tryenter(hash_lock)) { 7464 /* 7465 * Missed the hash lock. Retry. 7466 */ 7467 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry); 7468 mutex_exit(&dev->l2ad_mtx); 7469 mutex_enter(hash_lock); 7470 mutex_exit(hash_lock); 7471 goto top; 7472 } 7473 7474 /* 7475 * A header can't be on this list if it doesn't have L2 header. 7476 */ 7477 ASSERT(HDR_HAS_L2HDR(hdr)); 7478 7479 /* Ensure this header has finished being written. */ 7480 ASSERT(!HDR_L2_WRITING(hdr)); 7481 ASSERT(!HDR_L2_WRITE_HEAD(hdr)); 7482 7483 if (!all && (hdr->b_l2hdr.b_daddr >= taddr || 7484 hdr->b_l2hdr.b_daddr < dev->l2ad_hand)) { 7485 /* 7486 * We've evicted to the target address, 7487 * or the end of the device. 7488 */ 7489 mutex_exit(hash_lock); 7490 break; 7491 } 7492 7493 if (!HDR_HAS_L1HDR(hdr)) { 7494 ASSERT(!HDR_L2_READING(hdr)); 7495 /* 7496 * This doesn't exist in the ARC. Destroy. 7497 * arc_hdr_destroy() will call list_remove() 7498 * and decrement arcstat_l2_lsize. 7499 */ 7500 arc_change_state(arc_anon, hdr, hash_lock); 7501 arc_hdr_destroy(hdr); 7502 } else { 7503 ASSERT(hdr->b_l1hdr.b_state != arc_l2c_only); 7504 ARCSTAT_BUMP(arcstat_l2_evict_l1cached); 7505 /* 7506 * Invalidate issued or about to be issued 7507 * reads, since we may be about to write 7508 * over this location. 7509 */ 7510 if (HDR_L2_READING(hdr)) { 7511 ARCSTAT_BUMP(arcstat_l2_evict_reading); 7512 arc_hdr_set_flags(hdr, ARC_FLAG_L2_EVICTED); 7513 } 7514 7515 arc_hdr_l2hdr_destroy(hdr); 7516 } 7517 mutex_exit(hash_lock); 7518 } 7519 mutex_exit(&dev->l2ad_mtx); 7520} 7521 7522/* 7523 * Find and write ARC buffers to the L2ARC device. 7524 * 7525 * An ARC_FLAG_L2_WRITING flag is set so that the L2ARC buffers are not valid 7526 * for reading until they have completed writing. 7527 * The headroom_boost is an in-out parameter used to maintain headroom boost 7528 * state between calls to this function. 7529 * 7530 * Returns the number of bytes actually written (which may be smaller than 7531 * the delta by which the device hand has changed due to alignment). 7532 */ 7533static uint64_t 7534l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz) 7535{ 7536 arc_buf_hdr_t *hdr, *hdr_prev, *head; 7537 uint64_t write_asize, write_psize, write_lsize, headroom; 7538 boolean_t full; 7539 l2arc_write_callback_t *cb; 7540 zio_t *pio, *wzio; 7541 uint64_t guid = spa_load_guid(spa); 7542 int try; 7543 7544 ASSERT3P(dev->l2ad_vdev, !=, NULL); 7545 7546 pio = NULL; 7547 write_lsize = write_asize = write_psize = 0; 7548 full = B_FALSE; 7549 head = kmem_cache_alloc(hdr_l2only_cache, KM_PUSHPAGE); 7550 arc_hdr_set_flags(head, ARC_FLAG_L2_WRITE_HEAD | ARC_FLAG_HAS_L2HDR); 7551 7552 ARCSTAT_BUMP(arcstat_l2_write_buffer_iter); 7553 /* 7554 * Copy buffers for L2ARC writing. 7555 */ 7556 for (try = 0; try <= 3; try++) { 7557 multilist_sublist_t *mls = l2arc_sublist_lock(try); 7558 uint64_t passed_sz = 0; 7559 7560 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_iter); 7561 7562 /* 7563 * L2ARC fast warmup. 7564 * 7565 * Until the ARC is warm and starts to evict, read from the 7566 * head of the ARC lists rather than the tail. 7567 */ 7568 if (arc_warm == B_FALSE) 7569 hdr = multilist_sublist_head(mls); 7570 else 7571 hdr = multilist_sublist_tail(mls); 7572 if (hdr == NULL) 7573 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_null_iter); 7574 7575 headroom = target_sz * l2arc_headroom; 7576 if (zfs_compressed_arc_enabled) 7577 headroom = (headroom * l2arc_headroom_boost) / 100; 7578 7579 for (; hdr; hdr = hdr_prev) { 7580 kmutex_t *hash_lock; 7581 7582 if (arc_warm == B_FALSE) 7583 hdr_prev = multilist_sublist_next(mls, hdr); 7584 else 7585 hdr_prev = multilist_sublist_prev(mls, hdr); 7586 ARCSTAT_INCR(arcstat_l2_write_buffer_bytes_scanned, 7587 HDR_GET_LSIZE(hdr)); 7588 7589 hash_lock = HDR_LOCK(hdr); 7590 if (!mutex_tryenter(hash_lock)) { 7591 ARCSTAT_BUMP(arcstat_l2_write_trylock_fail); 7592 /* 7593 * Skip this buffer rather than waiting. 7594 */ 7595 continue; 7596 } 7597 7598 passed_sz += HDR_GET_LSIZE(hdr); 7599 if (passed_sz > headroom) { 7600 /* 7601 * Searched too far. 7602 */ 7603 mutex_exit(hash_lock); 7604 ARCSTAT_BUMP(arcstat_l2_write_passed_headroom); 7605 break; 7606 } 7607 7608 if (!l2arc_write_eligible(guid, hdr)) { 7609 mutex_exit(hash_lock); 7610 continue; 7611 } 7612 7613 /* 7614 * We rely on the L1 portion of the header below, so 7615 * it's invalid for this header to have been evicted out 7616 * of the ghost cache, prior to being written out. The 7617 * ARC_FLAG_L2_WRITING bit ensures this won't happen. 7618 */ 7619 ASSERT(HDR_HAS_L1HDR(hdr)); 7620 7621 ASSERT3U(HDR_GET_PSIZE(hdr), >, 0); 7622 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL); 7623 ASSERT3U(arc_hdr_size(hdr), >, 0); 7624 uint64_t psize = arc_hdr_size(hdr); 7625 uint64_t asize = vdev_psize_to_asize(dev->l2ad_vdev, 7626 psize); 7627 7628 if ((write_asize + asize) > target_sz) { 7629 full = B_TRUE; 7630 mutex_exit(hash_lock); 7631 ARCSTAT_BUMP(arcstat_l2_write_full); 7632 break; 7633 } 7634 7635 if (pio == NULL) { 7636 /* 7637 * Insert a dummy header on the buflist so 7638 * l2arc_write_done() can find where the 7639 * write buffers begin without searching. 7640 */ 7641 mutex_enter(&dev->l2ad_mtx); 7642 list_insert_head(&dev->l2ad_buflist, head); 7643 mutex_exit(&dev->l2ad_mtx); 7644 7645 cb = kmem_alloc( 7646 sizeof (l2arc_write_callback_t), KM_SLEEP); 7647 cb->l2wcb_dev = dev; 7648 cb->l2wcb_head = head; 7649 pio = zio_root(spa, l2arc_write_done, cb, 7650 ZIO_FLAG_CANFAIL); 7651 ARCSTAT_BUMP(arcstat_l2_write_pios); 7652 } 7653 7654 hdr->b_l2hdr.b_dev = dev; 7655 hdr->b_l2hdr.b_daddr = dev->l2ad_hand; 7656 arc_hdr_set_flags(hdr, 7657 ARC_FLAG_L2_WRITING | ARC_FLAG_HAS_L2HDR); 7658 7659 mutex_enter(&dev->l2ad_mtx); 7660 list_insert_head(&dev->l2ad_buflist, hdr); 7661 mutex_exit(&dev->l2ad_mtx); 7662 7663 (void) refcount_add_many(&dev->l2ad_alloc, psize, hdr); 7664 7665 /* 7666 * Normally the L2ARC can use the hdr's data, but if 7667 * we're sharing data between the hdr and one of its 7668 * bufs, L2ARC needs its own copy of the data so that 7669 * the ZIO below can't race with the buf consumer. 7670 * Another case where we need to create a copy of the 7671 * data is when the buffer size is not device-aligned 7672 * and we need to pad the block to make it such. 7673 * That also keeps the clock hand suitably aligned. 7674 * 7675 * To ensure that the copy will be available for the 7676 * lifetime of the ZIO and be cleaned up afterwards, we 7677 * add it to the l2arc_free_on_write queue. 7678 */ 7679 abd_t *to_write; 7680 if (!HDR_SHARED_DATA(hdr) && psize == asize) { 7681 to_write = hdr->b_l1hdr.b_pabd; 7682 } else { 7683 to_write = abd_alloc_for_io(asize, 7684 HDR_ISTYPE_METADATA(hdr)); 7685 abd_copy(to_write, hdr->b_l1hdr.b_pabd, psize); 7686 if (asize != psize) { 7687 abd_zero_off(to_write, psize, 7688 asize - psize); 7689 } 7690 l2arc_free_abd_on_write(to_write, asize, 7691 arc_buf_type(hdr)); 7692 } 7693 wzio = zio_write_phys(pio, dev->l2ad_vdev, 7694 hdr->b_l2hdr.b_daddr, asize, to_write, 7695 ZIO_CHECKSUM_OFF, NULL, hdr, 7696 ZIO_PRIORITY_ASYNC_WRITE, 7697 ZIO_FLAG_CANFAIL, B_FALSE); 7698 7699 write_lsize += HDR_GET_LSIZE(hdr); 7700 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev, 7701 zio_t *, wzio); 7702 7703 write_psize += psize; 7704 write_asize += asize; 7705 dev->l2ad_hand += asize; 7706 7707 mutex_exit(hash_lock); 7708 7709 (void) zio_nowait(wzio); 7710 } 7711 7712 multilist_sublist_unlock(mls); 7713 7714 if (full == B_TRUE) 7715 break; 7716 } 7717 7718 /* No buffers selected for writing? */ 7719 if (pio == NULL) { 7720 ASSERT0(write_lsize); 7721 ASSERT(!HDR_HAS_L1HDR(head)); 7722 kmem_cache_free(hdr_l2only_cache, head); 7723 return (0); 7724 } 7725 7726 ASSERT3U(write_psize, <=, target_sz); 7727 ARCSTAT_BUMP(arcstat_l2_writes_sent); 7728 ARCSTAT_INCR(arcstat_l2_write_bytes, write_psize); 7729 ARCSTAT_INCR(arcstat_l2_lsize, write_lsize); 7730 ARCSTAT_INCR(arcstat_l2_psize, write_psize); 7731 vdev_space_update(dev->l2ad_vdev, write_psize, 0, 0); 7732 7733 /* 7734 * Bump device hand to the device start if it is approaching the end. 7735 * l2arc_evict() will already have evicted ahead for this case. 7736 */ 7737 if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) { 7738 dev->l2ad_hand = dev->l2ad_start; 7739 dev->l2ad_first = B_FALSE; 7740 } 7741 7742 dev->l2ad_writing = B_TRUE; 7743 (void) zio_wait(pio); 7744 dev->l2ad_writing = B_FALSE; 7745 7746 return (write_asize); 7747} 7748 7749/* 7750 * This thread feeds the L2ARC at regular intervals. This is the beating 7751 * heart of the L2ARC. 7752 */ 7753/* ARGSUSED */ 7754static void 7755l2arc_feed_thread(void *unused __unused) 7756{ 7757 callb_cpr_t cpr; 7758 l2arc_dev_t *dev; 7759 spa_t *spa; 7760 uint64_t size, wrote; 7761 clock_t begin, next = ddi_get_lbolt(); 7762 7763 CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG); 7764 7765 mutex_enter(&l2arc_feed_thr_lock); 7766 7767 while (l2arc_thread_exit == 0) { 7768 CALLB_CPR_SAFE_BEGIN(&cpr); 7769 (void) cv_timedwait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock, 7770 next - ddi_get_lbolt()); 7771 CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock); 7772 next = ddi_get_lbolt() + hz; 7773 7774 /* 7775 * Quick check for L2ARC devices. 7776 */ 7777 mutex_enter(&l2arc_dev_mtx); 7778 if (l2arc_ndev == 0) { 7779 mutex_exit(&l2arc_dev_mtx); 7780 continue; 7781 } 7782 mutex_exit(&l2arc_dev_mtx); 7783 begin = ddi_get_lbolt(); 7784 7785 /* 7786 * This selects the next l2arc device to write to, and in 7787 * doing so the next spa to feed from: dev->l2ad_spa. This 7788 * will return NULL if there are now no l2arc devices or if 7789 * they are all faulted. 7790 * 7791 * If a device is returned, its spa's config lock is also 7792 * held to prevent device removal. l2arc_dev_get_next() 7793 * will grab and release l2arc_dev_mtx. 7794 */ 7795 if ((dev = l2arc_dev_get_next()) == NULL) 7796 continue; 7797 7798 spa = dev->l2ad_spa; 7799 ASSERT3P(spa, !=, NULL); 7800 7801 /* 7802 * If the pool is read-only then force the feed thread to 7803 * sleep a little longer. 7804 */ 7805 if (!spa_writeable(spa)) { 7806 next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz; 7807 spa_config_exit(spa, SCL_L2ARC, dev); 7808 continue; 7809 } 7810 7811 /* 7812 * Avoid contributing to memory pressure. 7813 */ 7814 if (arc_reclaim_needed()) { 7815 ARCSTAT_BUMP(arcstat_l2_abort_lowmem); 7816 spa_config_exit(spa, SCL_L2ARC, dev); 7817 continue; 7818 } 7819 7820 ARCSTAT_BUMP(arcstat_l2_feeds); 7821 7822 size = l2arc_write_size(); 7823 7824 /* 7825 * Evict L2ARC buffers that will be overwritten. 7826 */ 7827 l2arc_evict(dev, size, B_FALSE); 7828 7829 /* 7830 * Write ARC buffers. 7831 */ 7832 wrote = l2arc_write_buffers(spa, dev, size); 7833 7834 /* 7835 * Calculate interval between writes. 7836 */ 7837 next = l2arc_write_interval(begin, size, wrote); 7838 spa_config_exit(spa, SCL_L2ARC, dev); 7839 } 7840 7841 l2arc_thread_exit = 0; 7842 cv_broadcast(&l2arc_feed_thr_cv); 7843 CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */ 7844 thread_exit(); 7845} 7846 7847boolean_t 7848l2arc_vdev_present(vdev_t *vd) 7849{ 7850 l2arc_dev_t *dev; 7851 7852 mutex_enter(&l2arc_dev_mtx); 7853 for (dev = list_head(l2arc_dev_list); dev != NULL; 7854 dev = list_next(l2arc_dev_list, dev)) { 7855 if (dev->l2ad_vdev == vd) 7856 break; 7857 } 7858 mutex_exit(&l2arc_dev_mtx); 7859 7860 return (dev != NULL); 7861} 7862 7863/* 7864 * Add a vdev for use by the L2ARC. By this point the spa has already 7865 * validated the vdev and opened it. 7866 */ 7867void 7868l2arc_add_vdev(spa_t *spa, vdev_t *vd) 7869{ 7870 l2arc_dev_t *adddev; 7871 7872 ASSERT(!l2arc_vdev_present(vd)); 7873 7874 vdev_ashift_optimize(vd); 7875 7876 /* 7877 * Create a new l2arc device entry. 7878 */ 7879 adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP); 7880 adddev->l2ad_spa = spa; 7881 adddev->l2ad_vdev = vd; 7882 adddev->l2ad_start = VDEV_LABEL_START_SIZE; 7883 adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd); 7884 adddev->l2ad_hand = adddev->l2ad_start; 7885 adddev->l2ad_first = B_TRUE; 7886 adddev->l2ad_writing = B_FALSE; 7887 7888 mutex_init(&adddev->l2ad_mtx, NULL, MUTEX_DEFAULT, NULL); 7889 /* 7890 * This is a list of all ARC buffers that are still valid on the 7891 * device. 7892 */ 7893 list_create(&adddev->l2ad_buflist, sizeof (arc_buf_hdr_t), 7894 offsetof(arc_buf_hdr_t, b_l2hdr.b_l2node)); 7895 7896 vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand); 7897 refcount_create(&adddev->l2ad_alloc); 7898 7899 /* 7900 * Add device to global list 7901 */ 7902 mutex_enter(&l2arc_dev_mtx); 7903 list_insert_head(l2arc_dev_list, adddev); 7904 atomic_inc_64(&l2arc_ndev); 7905 mutex_exit(&l2arc_dev_mtx); 7906} 7907 7908/* 7909 * Remove a vdev from the L2ARC. 7910 */ 7911void 7912l2arc_remove_vdev(vdev_t *vd) 7913{ 7914 l2arc_dev_t *dev, *nextdev, *remdev = NULL; 7915 7916 /* 7917 * Find the device by vdev 7918 */ 7919 mutex_enter(&l2arc_dev_mtx); 7920 for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) { 7921 nextdev = list_next(l2arc_dev_list, dev); 7922 if (vd == dev->l2ad_vdev) { 7923 remdev = dev; 7924 break; 7925 } 7926 } 7927 ASSERT3P(remdev, !=, NULL); 7928 7929 /* 7930 * Remove device from global list 7931 */ 7932 list_remove(l2arc_dev_list, remdev); 7933 l2arc_dev_last = NULL; /* may have been invalidated */ 7934 atomic_dec_64(&l2arc_ndev); 7935 mutex_exit(&l2arc_dev_mtx); 7936 7937 /* 7938 * Clear all buflists and ARC references. L2ARC device flush. 7939 */ 7940 l2arc_evict(remdev, 0, B_TRUE); 7941 list_destroy(&remdev->l2ad_buflist); 7942 mutex_destroy(&remdev->l2ad_mtx); 7943 refcount_destroy(&remdev->l2ad_alloc); 7944 kmem_free(remdev, sizeof (l2arc_dev_t)); 7945} 7946 7947void 7948l2arc_init(void) 7949{ 7950 l2arc_thread_exit = 0; 7951 l2arc_ndev = 0; 7952 l2arc_writes_sent = 0; 7953 l2arc_writes_done = 0; 7954 7955 mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL); 7956 cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL); 7957 mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL); 7958 mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL); 7959 7960 l2arc_dev_list = &L2ARC_dev_list; 7961 l2arc_free_on_write = &L2ARC_free_on_write; 7962 list_create(l2arc_dev_list, sizeof (l2arc_dev_t), 7963 offsetof(l2arc_dev_t, l2ad_node)); 7964 list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t), 7965 offsetof(l2arc_data_free_t, l2df_list_node)); 7966} 7967 7968void 7969l2arc_fini(void) 7970{ 7971 /* 7972 * This is called from dmu_fini(), which is called from spa_fini(); 7973 * Because of this, we can assume that all l2arc devices have 7974 * already been removed when the pools themselves were removed. 7975 */ 7976 7977 l2arc_do_free_on_write(); 7978 7979 mutex_destroy(&l2arc_feed_thr_lock); 7980 cv_destroy(&l2arc_feed_thr_cv); 7981 mutex_destroy(&l2arc_dev_mtx); 7982 mutex_destroy(&l2arc_free_on_write_mtx); 7983 7984 list_destroy(l2arc_dev_list); 7985 list_destroy(l2arc_free_on_write); 7986} 7987 7988void 7989l2arc_start(void) 7990{ 7991 if (!(spa_mode_global & FWRITE)) 7992 return; 7993 7994 (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0, 7995 TS_RUN, minclsyspri); 7996} 7997 7998void 7999l2arc_stop(void) 8000{ 8001 if (!(spa_mode_global & FWRITE)) 8002 return; 8003 8004 mutex_enter(&l2arc_feed_thr_lock); 8005 cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */ 8006 l2arc_thread_exit = 1; 8007 while (l2arc_thread_exit != 0) 8008 cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock); 8009 mutex_exit(&l2arc_feed_thr_lock); 8010} 8011