arc.c revision 338346
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 * Adjust MFU size 4123 * 4124 * Now that we've tried to evict enough from the MRU to get its 4125 * size back to arc_p, if we're still above the target cache 4126 * size, we evict the rest from the MFU. 4127 */ 4128 target = asize - arc_c; 4129 4130 if (arc_adjust_type(arc_mfu) == ARC_BUFC_METADATA && 4131 ameta > arc_meta_min) { 4132 bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA); 4133 total_evicted += bytes; 4134 4135 /* 4136 * If we couldn't evict our target number of bytes from 4137 * metadata, we try to get the rest from data. 4138 */ 4139 target -= bytes; 4140 4141 total_evicted += 4142 arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA); 4143 } else { 4144 bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA); 4145 total_evicted += bytes; 4146 4147 /* 4148 * If we couldn't evict our target number of bytes from 4149 * data, we try to get the rest from data. 4150 */ 4151 target -= bytes; 4152 4153 total_evicted += 4154 arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA); 4155 } 4156 4157 /* 4158 * Adjust ghost lists 4159 * 4160 * In addition to the above, the ARC also defines target values 4161 * for the ghost lists. The sum of the mru list and mru ghost 4162 * list should never exceed the target size of the cache, and 4163 * the sum of the mru list, mfu list, mru ghost list, and mfu 4164 * ghost list should never exceed twice the target size of the 4165 * cache. The following logic enforces these limits on the ghost 4166 * caches, and evicts from them as needed. 4167 */ 4168 target = refcount_count(&arc_mru->arcs_size) + 4169 refcount_count(&arc_mru_ghost->arcs_size) - arc_c; 4170 4171 bytes = arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_DATA); 4172 total_evicted += bytes; 4173 4174 target -= bytes; 4175 4176 total_evicted += 4177 arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_METADATA); 4178 4179 /* 4180 * We assume the sum of the mru list and mfu list is less than 4181 * or equal to arc_c (we enforced this above), which means we 4182 * can use the simpler of the two equations below: 4183 * 4184 * mru + mfu + mru ghost + mfu ghost <= 2 * arc_c 4185 * mru ghost + mfu ghost <= arc_c 4186 */ 4187 target = refcount_count(&arc_mru_ghost->arcs_size) + 4188 refcount_count(&arc_mfu_ghost->arcs_size) - arc_c; 4189 4190 bytes = arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_DATA); 4191 total_evicted += bytes; 4192 4193 target -= bytes; 4194 4195 total_evicted += 4196 arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_METADATA); 4197 4198 return (total_evicted); 4199} 4200 4201void 4202arc_flush(spa_t *spa, boolean_t retry) 4203{ 4204 uint64_t guid = 0; 4205 4206 /* 4207 * If retry is B_TRUE, a spa must not be specified since we have 4208 * no good way to determine if all of a spa's buffers have been 4209 * evicted from an arc state. 4210 */ 4211 ASSERT(!retry || spa == 0); 4212 4213 if (spa != NULL) 4214 guid = spa_load_guid(spa); 4215 4216 (void) arc_flush_state(arc_mru, guid, ARC_BUFC_DATA, retry); 4217 (void) arc_flush_state(arc_mru, guid, ARC_BUFC_METADATA, retry); 4218 4219 (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_DATA, retry); 4220 (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_METADATA, retry); 4221 4222 (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_DATA, retry); 4223 (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_METADATA, retry); 4224 4225 (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_DATA, retry); 4226 (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_METADATA, retry); 4227} 4228 4229void 4230arc_shrink(int64_t to_free) 4231{ 4232 uint64_t asize = aggsum_value(&arc_size); 4233 if (arc_c > arc_c_min) { 4234 DTRACE_PROBE4(arc__shrink, uint64_t, arc_c, uint64_t, 4235 arc_c_min, uint64_t, arc_p, uint64_t, to_free); 4236 if (arc_c > arc_c_min + to_free) 4237 atomic_add_64(&arc_c, -to_free); 4238 else 4239 arc_c = arc_c_min; 4240 4241 atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift)); 4242 if (asize < arc_c) 4243 arc_c = MAX(asize, arc_c_min); 4244 if (arc_p > arc_c) 4245 arc_p = (arc_c >> 1); 4246 4247 DTRACE_PROBE2(arc__shrunk, uint64_t, arc_c, uint64_t, 4248 arc_p); 4249 4250 ASSERT(arc_c >= arc_c_min); 4251 ASSERT((int64_t)arc_p >= 0); 4252 } 4253 4254 if (asize > arc_c) { 4255 DTRACE_PROBE2(arc__shrink_adjust, uint64_t, asize, 4256 uint64_t, arc_c); 4257 (void) arc_adjust(); 4258 } 4259} 4260 4261typedef enum free_memory_reason_t { 4262 FMR_UNKNOWN, 4263 FMR_NEEDFREE, 4264 FMR_LOTSFREE, 4265 FMR_SWAPFS_MINFREE, 4266 FMR_PAGES_PP_MAXIMUM, 4267 FMR_HEAP_ARENA, 4268 FMR_ZIO_ARENA, 4269 FMR_ZIO_FRAG, 4270} free_memory_reason_t; 4271 4272int64_t last_free_memory; 4273free_memory_reason_t last_free_reason; 4274 4275/* 4276 * Additional reserve of pages for pp_reserve. 4277 */ 4278int64_t arc_pages_pp_reserve = 64; 4279 4280/* 4281 * Additional reserve of pages for swapfs. 4282 */ 4283int64_t arc_swapfs_reserve = 64; 4284 4285/* 4286 * Return the amount of memory that can be consumed before reclaim will be 4287 * needed. Positive if there is sufficient free memory, negative indicates 4288 * the amount of memory that needs to be freed up. 4289 */ 4290static int64_t 4291arc_available_memory(void) 4292{ 4293 int64_t lowest = INT64_MAX; 4294 int64_t n; 4295 free_memory_reason_t r = FMR_UNKNOWN; 4296 4297#ifdef _KERNEL 4298#ifdef __FreeBSD__ 4299 /* 4300 * Cooperate with pagedaemon when it's time for it to scan 4301 * and reclaim some pages. 4302 */ 4303 n = PAGESIZE * ((int64_t)freemem - zfs_arc_free_target); 4304 if (n < lowest) { 4305 lowest = n; 4306 r = FMR_LOTSFREE; 4307 } 4308 4309#else 4310 if (needfree > 0) { 4311 n = PAGESIZE * (-needfree); 4312 if (n < lowest) { 4313 lowest = n; 4314 r = FMR_NEEDFREE; 4315 } 4316 } 4317 4318 /* 4319 * check that we're out of range of the pageout scanner. It starts to 4320 * schedule paging if freemem is less than lotsfree and needfree. 4321 * lotsfree is the high-water mark for pageout, and needfree is the 4322 * number of needed free pages. We add extra pages here to make sure 4323 * the scanner doesn't start up while we're freeing memory. 4324 */ 4325 n = PAGESIZE * (freemem - lotsfree - needfree - desfree); 4326 if (n < lowest) { 4327 lowest = n; 4328 r = FMR_LOTSFREE; 4329 } 4330 4331 /* 4332 * check to make sure that swapfs has enough space so that anon 4333 * reservations can still succeed. anon_resvmem() checks that the 4334 * availrmem is greater than swapfs_minfree, and the number of reserved 4335 * swap pages. We also add a bit of extra here just to prevent 4336 * circumstances from getting really dire. 4337 */ 4338 n = PAGESIZE * (availrmem - swapfs_minfree - swapfs_reserve - 4339 desfree - arc_swapfs_reserve); 4340 if (n < lowest) { 4341 lowest = n; 4342 r = FMR_SWAPFS_MINFREE; 4343 } 4344 4345 4346 /* 4347 * Check that we have enough availrmem that memory locking (e.g., via 4348 * mlock(3C) or memcntl(2)) can still succeed. (pages_pp_maximum 4349 * stores the number of pages that cannot be locked; when availrmem 4350 * drops below pages_pp_maximum, page locking mechanisms such as 4351 * page_pp_lock() will fail.) 4352 */ 4353 n = PAGESIZE * (availrmem - pages_pp_maximum - 4354 arc_pages_pp_reserve); 4355 if (n < lowest) { 4356 lowest = n; 4357 r = FMR_PAGES_PP_MAXIMUM; 4358 } 4359 4360#endif /* __FreeBSD__ */ 4361#if defined(__i386) || !defined(UMA_MD_SMALL_ALLOC) 4362 /* 4363 * If we're on an i386 platform, it's possible that we'll exhaust the 4364 * kernel heap space before we ever run out of available physical 4365 * memory. Most checks of the size of the heap_area compare against 4366 * tune.t_minarmem, which is the minimum available real memory that we 4367 * can have in the system. However, this is generally fixed at 25 pages 4368 * which is so low that it's useless. In this comparison, we seek to 4369 * calculate the total heap-size, and reclaim if more than 3/4ths of the 4370 * heap is allocated. (Or, in the calculation, if less than 1/4th is 4371 * free) 4372 */ 4373 n = (int64_t)vmem_size(heap_arena, VMEM_FREE) - 4374 (vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC) >> 2); 4375 if (n < lowest) { 4376 lowest = n; 4377 r = FMR_HEAP_ARENA; 4378 } 4379#define zio_arena NULL 4380#else 4381#define zio_arena heap_arena 4382#endif 4383 4384 /* 4385 * If zio data pages are being allocated out of a separate heap segment, 4386 * then enforce that the size of available vmem for this arena remains 4387 * above about 1/4th (1/(2^arc_zio_arena_free_shift)) free. 4388 * 4389 * Note that reducing the arc_zio_arena_free_shift keeps more virtual 4390 * memory (in the zio_arena) free, which can avoid memory 4391 * fragmentation issues. 4392 */ 4393 if (zio_arena != NULL) { 4394 n = (int64_t)vmem_size(zio_arena, VMEM_FREE) - 4395 (vmem_size(zio_arena, VMEM_ALLOC) >> 4396 arc_zio_arena_free_shift); 4397 if (n < lowest) { 4398 lowest = n; 4399 r = FMR_ZIO_ARENA; 4400 } 4401 } 4402 4403 /* 4404 * Above limits know nothing about real level of KVA fragmentation. 4405 * Start aggressive reclamation if too little sequential KVA left. 4406 */ 4407 if (lowest > 0) { 4408 n = (vmem_size(heap_arena, VMEM_MAXFREE) < SPA_MAXBLOCKSIZE) ? 4409 -((int64_t)vmem_size(heap_arena, VMEM_ALLOC) >> 4) : 4410 INT64_MAX; 4411 if (n < lowest) { 4412 lowest = n; 4413 r = FMR_ZIO_FRAG; 4414 } 4415 } 4416 4417#else /* _KERNEL */ 4418 /* Every 100 calls, free a small amount */ 4419 if (spa_get_random(100) == 0) 4420 lowest = -1024; 4421#endif /* _KERNEL */ 4422 4423 last_free_memory = lowest; 4424 last_free_reason = r; 4425 DTRACE_PROBE2(arc__available_memory, int64_t, lowest, int, r); 4426 return (lowest); 4427} 4428 4429 4430/* 4431 * Determine if the system is under memory pressure and is asking 4432 * to reclaim memory. A return value of B_TRUE indicates that the system 4433 * is under memory pressure and that the arc should adjust accordingly. 4434 */ 4435static boolean_t 4436arc_reclaim_needed(void) 4437{ 4438 return (arc_available_memory() < 0); 4439} 4440 4441extern kmem_cache_t *zio_buf_cache[]; 4442extern kmem_cache_t *zio_data_buf_cache[]; 4443extern kmem_cache_t *range_seg_cache; 4444extern kmem_cache_t *abd_chunk_cache; 4445 4446static __noinline void 4447arc_kmem_reap_now(void) 4448{ 4449 size_t i; 4450 kmem_cache_t *prev_cache = NULL; 4451 kmem_cache_t *prev_data_cache = NULL; 4452 4453 DTRACE_PROBE(arc__kmem_reap_start); 4454#ifdef _KERNEL 4455 if (aggsum_compare(&arc_meta_used, arc_meta_limit) >= 0) { 4456 /* 4457 * We are exceeding our meta-data cache limit. 4458 * Purge some DNLC entries to release holds on meta-data. 4459 */ 4460 dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent); 4461 } 4462#if defined(__i386) 4463 /* 4464 * Reclaim unused memory from all kmem caches. 4465 */ 4466 kmem_reap(); 4467#endif 4468#endif 4469 4470 /* 4471 * If a kmem reap is already active, don't schedule more. We must 4472 * check for this because kmem_cache_reap_soon() won't actually 4473 * block on the cache being reaped (this is to prevent callers from 4474 * becoming implicitly blocked by a system-wide kmem reap -- which, 4475 * on a system with many, many full magazines, can take minutes). 4476 */ 4477 if (kmem_cache_reap_active()) 4478 return; 4479 4480 for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) { 4481 if (zio_buf_cache[i] != prev_cache) { 4482 prev_cache = zio_buf_cache[i]; 4483 kmem_cache_reap_soon(zio_buf_cache[i]); 4484 } 4485 if (zio_data_buf_cache[i] != prev_data_cache) { 4486 prev_data_cache = zio_data_buf_cache[i]; 4487 kmem_cache_reap_soon(zio_data_buf_cache[i]); 4488 } 4489 } 4490 kmem_cache_reap_soon(abd_chunk_cache); 4491 kmem_cache_reap_soon(buf_cache); 4492 kmem_cache_reap_soon(hdr_full_cache); 4493 kmem_cache_reap_soon(hdr_l2only_cache); 4494 kmem_cache_reap_soon(range_seg_cache); 4495 4496#ifdef illumos 4497 if (zio_arena != NULL) { 4498 /* 4499 * Ask the vmem arena to reclaim unused memory from its 4500 * quantum caches. 4501 */ 4502 vmem_qcache_reap(zio_arena); 4503 } 4504#endif 4505 DTRACE_PROBE(arc__kmem_reap_end); 4506} 4507 4508/* 4509 * Threads can block in arc_get_data_impl() waiting for this thread to evict 4510 * enough data and signal them to proceed. When this happens, the threads in 4511 * arc_get_data_impl() are sleeping while holding the hash lock for their 4512 * particular arc header. Thus, we must be careful to never sleep on a 4513 * hash lock in this thread. This is to prevent the following deadlock: 4514 * 4515 * - Thread A sleeps on CV in arc_get_data_impl() holding hash lock "L", 4516 * waiting for the reclaim thread to signal it. 4517 * 4518 * - arc_reclaim_thread() tries to acquire hash lock "L" using mutex_enter, 4519 * fails, and goes to sleep forever. 4520 * 4521 * This possible deadlock is avoided by always acquiring a hash lock 4522 * using mutex_tryenter() from arc_reclaim_thread(). 4523 */ 4524/* ARGSUSED */ 4525static void 4526arc_reclaim_thread(void *unused __unused) 4527{ 4528 hrtime_t growtime = 0; 4529 hrtime_t kmem_reap_time = 0; 4530 callb_cpr_t cpr; 4531 4532 CALLB_CPR_INIT(&cpr, &arc_reclaim_lock, callb_generic_cpr, FTAG); 4533 4534 mutex_enter(&arc_reclaim_lock); 4535 while (!arc_reclaim_thread_exit) { 4536 uint64_t evicted = 0; 4537 4538 /* 4539 * This is necessary in order for the mdb ::arc dcmd to 4540 * show up to date information. Since the ::arc command 4541 * does not call the kstat's update function, without 4542 * this call, the command may show stale stats for the 4543 * anon, mru, mru_ghost, mfu, and mfu_ghost lists. Even 4544 * with this change, the data might be up to 1 second 4545 * out of date; but that should suffice. The arc_state_t 4546 * structures can be queried directly if more accurate 4547 * information is needed. 4548 */ 4549 if (arc_ksp != NULL) 4550 arc_ksp->ks_update(arc_ksp, KSTAT_READ); 4551 4552 mutex_exit(&arc_reclaim_lock); 4553 4554 /* 4555 * We call arc_adjust() before (possibly) calling 4556 * arc_kmem_reap_now(), so that we can wake up 4557 * arc_get_data_impl() sooner. 4558 */ 4559 evicted = arc_adjust(); 4560 4561 int64_t free_memory = arc_available_memory(); 4562 if (free_memory < 0) { 4563 hrtime_t curtime = gethrtime(); 4564 arc_no_grow = B_TRUE; 4565 arc_warm = B_TRUE; 4566 4567 /* 4568 * Wait at least zfs_grow_retry (default 60) seconds 4569 * before considering growing. 4570 */ 4571 growtime = curtime + SEC2NSEC(arc_grow_retry); 4572 4573 /* 4574 * Wait at least arc_kmem_cache_reap_retry_ms 4575 * between arc_kmem_reap_now() calls. Without 4576 * this check it is possible to end up in a 4577 * situation where we spend lots of time 4578 * reaping caches, while we're near arc_c_min. 4579 */ 4580 if (curtime >= kmem_reap_time) { 4581 arc_kmem_reap_now(); 4582 kmem_reap_time = gethrtime() + 4583 MSEC2NSEC(arc_kmem_cache_reap_retry_ms); 4584 } 4585 4586 /* 4587 * If we are still low on memory, shrink the ARC 4588 * so that we have arc_shrink_min free space. 4589 */ 4590 free_memory = arc_available_memory(); 4591 4592 int64_t to_free = 4593 (arc_c >> arc_shrink_shift) - free_memory; 4594 if (to_free > 0) { 4595#ifdef _KERNEL 4596#ifdef illumos 4597 to_free = MAX(to_free, ptob(needfree)); 4598#endif 4599#endif 4600 arc_shrink(to_free); 4601 } 4602 } else if (free_memory < arc_c >> arc_no_grow_shift) { 4603 arc_no_grow = B_TRUE; 4604 } else if (gethrtime() >= growtime) { 4605 arc_no_grow = B_FALSE; 4606 } 4607 4608 mutex_enter(&arc_reclaim_lock); 4609 4610 /* 4611 * If evicted is zero, we couldn't evict anything via 4612 * arc_adjust(). This could be due to hash lock 4613 * collisions, but more likely due to the majority of 4614 * arc buffers being unevictable. Therefore, even if 4615 * arc_size is above arc_c, another pass is unlikely to 4616 * be helpful and could potentially cause us to enter an 4617 * infinite loop. 4618 */ 4619 if (aggsum_compare(&arc_size, arc_c) <= 0|| evicted == 0) { 4620 /* 4621 * We're either no longer overflowing, or we 4622 * can't evict anything more, so we should wake 4623 * up any threads before we go to sleep. 4624 */ 4625 cv_broadcast(&arc_reclaim_waiters_cv); 4626 4627 /* 4628 * Block until signaled, or after one second (we 4629 * might need to perform arc_kmem_reap_now() 4630 * even if we aren't being signalled) 4631 */ 4632 CALLB_CPR_SAFE_BEGIN(&cpr); 4633 (void) cv_timedwait_hires(&arc_reclaim_thread_cv, 4634 &arc_reclaim_lock, SEC2NSEC(1), MSEC2NSEC(1), 0); 4635 CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_lock); 4636 } 4637 } 4638 4639 arc_reclaim_thread_exit = B_FALSE; 4640 cv_broadcast(&arc_reclaim_thread_cv); 4641 CALLB_CPR_EXIT(&cpr); /* drops arc_reclaim_lock */ 4642 thread_exit(); 4643} 4644 4645static u_int arc_dnlc_evicts_arg; 4646extern struct vfsops zfs_vfsops; 4647 4648static void 4649arc_dnlc_evicts_thread(void *dummy __unused) 4650{ 4651 callb_cpr_t cpr; 4652 u_int percent; 4653 4654 CALLB_CPR_INIT(&cpr, &arc_dnlc_evicts_lock, callb_generic_cpr, FTAG); 4655 4656 mutex_enter(&arc_dnlc_evicts_lock); 4657 while (!arc_dnlc_evicts_thread_exit) { 4658 CALLB_CPR_SAFE_BEGIN(&cpr); 4659 (void) cv_wait(&arc_dnlc_evicts_cv, &arc_dnlc_evicts_lock); 4660 CALLB_CPR_SAFE_END(&cpr, &arc_dnlc_evicts_lock); 4661 if (arc_dnlc_evicts_arg != 0) { 4662 percent = arc_dnlc_evicts_arg; 4663 mutex_exit(&arc_dnlc_evicts_lock); 4664#ifdef _KERNEL 4665 vnlru_free(desiredvnodes * percent / 100, &zfs_vfsops); 4666#endif 4667 mutex_enter(&arc_dnlc_evicts_lock); 4668 /* 4669 * Clear our token only after vnlru_free() 4670 * pass is done, to avoid false queueing of 4671 * the requests. 4672 */ 4673 arc_dnlc_evicts_arg = 0; 4674 } 4675 } 4676 arc_dnlc_evicts_thread_exit = FALSE; 4677 cv_broadcast(&arc_dnlc_evicts_cv); 4678 CALLB_CPR_EXIT(&cpr); 4679 thread_exit(); 4680} 4681 4682void 4683dnlc_reduce_cache(void *arg) 4684{ 4685 u_int percent; 4686 4687 percent = (u_int)(uintptr_t)arg; 4688 mutex_enter(&arc_dnlc_evicts_lock); 4689 if (arc_dnlc_evicts_arg == 0) { 4690 arc_dnlc_evicts_arg = percent; 4691 cv_broadcast(&arc_dnlc_evicts_cv); 4692 } 4693 mutex_exit(&arc_dnlc_evicts_lock); 4694} 4695 4696/* 4697 * Adapt arc info given the number of bytes we are trying to add and 4698 * the state that we are comming from. This function is only called 4699 * when we are adding new content to the cache. 4700 */ 4701static void 4702arc_adapt(int bytes, arc_state_t *state) 4703{ 4704 int mult; 4705 uint64_t arc_p_min = (arc_c >> arc_p_min_shift); 4706 int64_t mrug_size = refcount_count(&arc_mru_ghost->arcs_size); 4707 int64_t mfug_size = refcount_count(&arc_mfu_ghost->arcs_size); 4708 4709 if (state == arc_l2c_only) 4710 return; 4711 4712 ASSERT(bytes > 0); 4713 /* 4714 * Adapt the target size of the MRU list: 4715 * - if we just hit in the MRU ghost list, then increase 4716 * the target size of the MRU list. 4717 * - if we just hit in the MFU ghost list, then increase 4718 * the target size of the MFU list by decreasing the 4719 * target size of the MRU list. 4720 */ 4721 if (state == arc_mru_ghost) { 4722 mult = (mrug_size >= mfug_size) ? 1 : (mfug_size / mrug_size); 4723 mult = MIN(mult, 10); /* avoid wild arc_p adjustment */ 4724 4725 arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult); 4726 } else if (state == arc_mfu_ghost) { 4727 uint64_t delta; 4728 4729 mult = (mfug_size >= mrug_size) ? 1 : (mrug_size / mfug_size); 4730 mult = MIN(mult, 10); 4731 4732 delta = MIN(bytes * mult, arc_p); 4733 arc_p = MAX(arc_p_min, arc_p - delta); 4734 } 4735 ASSERT((int64_t)arc_p >= 0); 4736 4737 if (arc_reclaim_needed()) { 4738 cv_signal(&arc_reclaim_thread_cv); 4739 return; 4740 } 4741 4742 if (arc_no_grow) 4743 return; 4744 4745 if (arc_c >= arc_c_max) 4746 return; 4747 4748 /* 4749 * If we're within (2 * maxblocksize) bytes of the target 4750 * cache size, increment the target cache size 4751 */ 4752 if (aggsum_compare(&arc_size, arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) > 4753 0) { 4754 DTRACE_PROBE1(arc__inc_adapt, int, bytes); 4755 atomic_add_64(&arc_c, (int64_t)bytes); 4756 if (arc_c > arc_c_max) 4757 arc_c = arc_c_max; 4758 else if (state == arc_anon) 4759 atomic_add_64(&arc_p, (int64_t)bytes); 4760 if (arc_p > arc_c) 4761 arc_p = arc_c; 4762 } 4763 ASSERT((int64_t)arc_p >= 0); 4764} 4765 4766/* 4767 * Check if arc_size has grown past our upper threshold, determined by 4768 * zfs_arc_overflow_shift. 4769 */ 4770static boolean_t 4771arc_is_overflowing(void) 4772{ 4773 /* Always allow at least one block of overflow */ 4774 uint64_t overflow = MAX(SPA_MAXBLOCKSIZE, 4775 arc_c >> zfs_arc_overflow_shift); 4776 4777 /* 4778 * We just compare the lower bound here for performance reasons. Our 4779 * primary goals are to make sure that the arc never grows without 4780 * bound, and that it can reach its maximum size. This check 4781 * accomplishes both goals. The maximum amount we could run over by is 4782 * 2 * aggsum_borrow_multiplier * NUM_CPUS * the average size of a block 4783 * in the ARC. In practice, that's in the tens of MB, which is low 4784 * enough to be safe. 4785 */ 4786 return (aggsum_lower_bound(&arc_size) >= arc_c + overflow); 4787} 4788 4789static abd_t * 4790arc_get_data_abd(arc_buf_hdr_t *hdr, uint64_t size, void *tag) 4791{ 4792 arc_buf_contents_t type = arc_buf_type(hdr); 4793 4794 arc_get_data_impl(hdr, size, tag); 4795 if (type == ARC_BUFC_METADATA) { 4796 return (abd_alloc(size, B_TRUE)); 4797 } else { 4798 ASSERT(type == ARC_BUFC_DATA); 4799 return (abd_alloc(size, B_FALSE)); 4800 } 4801} 4802 4803static void * 4804arc_get_data_buf(arc_buf_hdr_t *hdr, uint64_t size, void *tag) 4805{ 4806 arc_buf_contents_t type = arc_buf_type(hdr); 4807 4808 arc_get_data_impl(hdr, size, tag); 4809 if (type == ARC_BUFC_METADATA) { 4810 return (zio_buf_alloc(size)); 4811 } else { 4812 ASSERT(type == ARC_BUFC_DATA); 4813 return (zio_data_buf_alloc(size)); 4814 } 4815} 4816 4817/* 4818 * Allocate a block and return it to the caller. If we are hitting the 4819 * hard limit for the cache size, we must sleep, waiting for the eviction 4820 * thread to catch up. If we're past the target size but below the hard 4821 * limit, we'll only signal the reclaim thread and continue on. 4822 */ 4823static void 4824arc_get_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag) 4825{ 4826 arc_state_t *state = hdr->b_l1hdr.b_state; 4827 arc_buf_contents_t type = arc_buf_type(hdr); 4828 4829 arc_adapt(size, state); 4830 4831 /* 4832 * If arc_size is currently overflowing, and has grown past our 4833 * upper limit, we must be adding data faster than the evict 4834 * thread can evict. Thus, to ensure we don't compound the 4835 * problem by adding more data and forcing arc_size to grow even 4836 * further past it's target size, we halt and wait for the 4837 * eviction thread to catch up. 4838 * 4839 * It's also possible that the reclaim thread is unable to evict 4840 * enough buffers to get arc_size below the overflow limit (e.g. 4841 * due to buffers being un-evictable, or hash lock collisions). 4842 * In this case, we want to proceed regardless if we're 4843 * overflowing; thus we don't use a while loop here. 4844 */ 4845 if (arc_is_overflowing()) { 4846 mutex_enter(&arc_reclaim_lock); 4847 4848 /* 4849 * Now that we've acquired the lock, we may no longer be 4850 * over the overflow limit, lets check. 4851 * 4852 * We're ignoring the case of spurious wake ups. If that 4853 * were to happen, it'd let this thread consume an ARC 4854 * buffer before it should have (i.e. before we're under 4855 * the overflow limit and were signalled by the reclaim 4856 * thread). As long as that is a rare occurrence, it 4857 * shouldn't cause any harm. 4858 */ 4859 if (arc_is_overflowing()) { 4860 cv_signal(&arc_reclaim_thread_cv); 4861 cv_wait(&arc_reclaim_waiters_cv, &arc_reclaim_lock); 4862 } 4863 4864 mutex_exit(&arc_reclaim_lock); 4865 } 4866 4867 VERIFY3U(hdr->b_type, ==, type); 4868 if (type == ARC_BUFC_METADATA) { 4869 arc_space_consume(size, ARC_SPACE_META); 4870 } else { 4871 arc_space_consume(size, ARC_SPACE_DATA); 4872 } 4873 4874 /* 4875 * Update the state size. Note that ghost states have a 4876 * "ghost size" and so don't need to be updated. 4877 */ 4878 if (!GHOST_STATE(state)) { 4879 4880 (void) refcount_add_many(&state->arcs_size, size, tag); 4881 4882 /* 4883 * If this is reached via arc_read, the link is 4884 * protected by the hash lock. If reached via 4885 * arc_buf_alloc, the header should not be accessed by 4886 * any other thread. And, if reached via arc_read_done, 4887 * the hash lock will protect it if it's found in the 4888 * hash table; otherwise no other thread should be 4889 * trying to [add|remove]_reference it. 4890 */ 4891 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) { 4892 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 4893 (void) refcount_add_many(&state->arcs_esize[type], 4894 size, tag); 4895 } 4896 4897 /* 4898 * If we are growing the cache, and we are adding anonymous 4899 * data, and we have outgrown arc_p, update arc_p 4900 */ 4901 if (aggsum_compare(&arc_size, arc_c) < 0 && 4902 hdr->b_l1hdr.b_state == arc_anon && 4903 (refcount_count(&arc_anon->arcs_size) + 4904 refcount_count(&arc_mru->arcs_size) > arc_p)) 4905 arc_p = MIN(arc_c, arc_p + size); 4906 } 4907 ARCSTAT_BUMP(arcstat_allocated); 4908} 4909 4910static void 4911arc_free_data_abd(arc_buf_hdr_t *hdr, abd_t *abd, uint64_t size, void *tag) 4912{ 4913 arc_free_data_impl(hdr, size, tag); 4914 abd_free(abd); 4915} 4916 4917static void 4918arc_free_data_buf(arc_buf_hdr_t *hdr, void *buf, uint64_t size, void *tag) 4919{ 4920 arc_buf_contents_t type = arc_buf_type(hdr); 4921 4922 arc_free_data_impl(hdr, size, tag); 4923 if (type == ARC_BUFC_METADATA) { 4924 zio_buf_free(buf, size); 4925 } else { 4926 ASSERT(type == ARC_BUFC_DATA); 4927 zio_data_buf_free(buf, size); 4928 } 4929} 4930 4931/* 4932 * Free the arc data buffer. 4933 */ 4934static void 4935arc_free_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag) 4936{ 4937 arc_state_t *state = hdr->b_l1hdr.b_state; 4938 arc_buf_contents_t type = arc_buf_type(hdr); 4939 4940 /* protected by hash lock, if in the hash table */ 4941 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) { 4942 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 4943 ASSERT(state != arc_anon && state != arc_l2c_only); 4944 4945 (void) refcount_remove_many(&state->arcs_esize[type], 4946 size, tag); 4947 } 4948 (void) refcount_remove_many(&state->arcs_size, size, tag); 4949 4950 VERIFY3U(hdr->b_type, ==, type); 4951 if (type == ARC_BUFC_METADATA) { 4952 arc_space_return(size, ARC_SPACE_META); 4953 } else { 4954 ASSERT(type == ARC_BUFC_DATA); 4955 arc_space_return(size, ARC_SPACE_DATA); 4956 } 4957} 4958 4959/* 4960 * This routine is called whenever a buffer is accessed. 4961 * NOTE: the hash lock is dropped in this function. 4962 */ 4963static void 4964arc_access(arc_buf_hdr_t *hdr, kmutex_t *hash_lock) 4965{ 4966 clock_t now; 4967 4968 ASSERT(MUTEX_HELD(hash_lock)); 4969 ASSERT(HDR_HAS_L1HDR(hdr)); 4970 4971 if (hdr->b_l1hdr.b_state == arc_anon) { 4972 /* 4973 * This buffer is not in the cache, and does not 4974 * appear in our "ghost" list. Add the new buffer 4975 * to the MRU state. 4976 */ 4977 4978 ASSERT0(hdr->b_l1hdr.b_arc_access); 4979 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 4980 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr); 4981 arc_change_state(arc_mru, hdr, hash_lock); 4982 4983 } else if (hdr->b_l1hdr.b_state == arc_mru) { 4984 now = ddi_get_lbolt(); 4985 4986 /* 4987 * If this buffer is here because of a prefetch, then either: 4988 * - clear the flag if this is a "referencing" read 4989 * (any subsequent access will bump this into the MFU state). 4990 * or 4991 * - move the buffer to the head of the list if this is 4992 * another prefetch (to make it less likely to be evicted). 4993 */ 4994 if (HDR_PREFETCH(hdr)) { 4995 if (refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) { 4996 /* link protected by hash lock */ 4997 ASSERT(multilist_link_active( 4998 &hdr->b_l1hdr.b_arc_node)); 4999 } else { 5000 arc_hdr_clear_flags(hdr, ARC_FLAG_PREFETCH); 5001 ARCSTAT_BUMP(arcstat_mru_hits); 5002 } 5003 hdr->b_l1hdr.b_arc_access = now; 5004 return; 5005 } 5006 5007 /* 5008 * This buffer has been "accessed" only once so far, 5009 * but it is still in the cache. Move it to the MFU 5010 * state. 5011 */ 5012 if (now > hdr->b_l1hdr.b_arc_access + ARC_MINTIME) { 5013 /* 5014 * More than 125ms have passed since we 5015 * instantiated this buffer. Move it to the 5016 * most frequently used state. 5017 */ 5018 hdr->b_l1hdr.b_arc_access = now; 5019 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 5020 arc_change_state(arc_mfu, hdr, hash_lock); 5021 } 5022 ARCSTAT_BUMP(arcstat_mru_hits); 5023 } else if (hdr->b_l1hdr.b_state == arc_mru_ghost) { 5024 arc_state_t *new_state; 5025 /* 5026 * This buffer has been "accessed" recently, but 5027 * was evicted from the cache. Move it to the 5028 * MFU state. 5029 */ 5030 5031 if (HDR_PREFETCH(hdr)) { 5032 new_state = arc_mru; 5033 if (refcount_count(&hdr->b_l1hdr.b_refcnt) > 0) 5034 arc_hdr_clear_flags(hdr, ARC_FLAG_PREFETCH); 5035 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr); 5036 } else { 5037 new_state = arc_mfu; 5038 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 5039 } 5040 5041 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 5042 arc_change_state(new_state, hdr, hash_lock); 5043 5044 ARCSTAT_BUMP(arcstat_mru_ghost_hits); 5045 } else if (hdr->b_l1hdr.b_state == arc_mfu) { 5046 /* 5047 * This buffer has been accessed more than once and is 5048 * still in the cache. Keep it in the MFU state. 5049 * 5050 * NOTE: an add_reference() that occurred when we did 5051 * the arc_read() will have kicked this off the list. 5052 * If it was a prefetch, we will explicitly move it to 5053 * the head of the list now. 5054 */ 5055 if ((HDR_PREFETCH(hdr)) != 0) { 5056 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 5057 /* link protected by hash_lock */ 5058 ASSERT(multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 5059 } 5060 ARCSTAT_BUMP(arcstat_mfu_hits); 5061 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 5062 } else if (hdr->b_l1hdr.b_state == arc_mfu_ghost) { 5063 arc_state_t *new_state = arc_mfu; 5064 /* 5065 * This buffer has been accessed more than once but has 5066 * been evicted from the cache. Move it back to the 5067 * MFU state. 5068 */ 5069 5070 if (HDR_PREFETCH(hdr)) { 5071 /* 5072 * This is a prefetch access... 5073 * move this block back to the MRU state. 5074 */ 5075 ASSERT0(refcount_count(&hdr->b_l1hdr.b_refcnt)); 5076 new_state = arc_mru; 5077 } 5078 5079 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 5080 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 5081 arc_change_state(new_state, hdr, hash_lock); 5082 5083 ARCSTAT_BUMP(arcstat_mfu_ghost_hits); 5084 } else if (hdr->b_l1hdr.b_state == arc_l2c_only) { 5085 /* 5086 * This buffer is on the 2nd Level ARC. 5087 */ 5088 5089 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 5090 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 5091 arc_change_state(arc_mfu, hdr, hash_lock); 5092 } else { 5093 ASSERT(!"invalid arc state"); 5094 } 5095} 5096 5097/* 5098 * This routine is called by dbuf_hold() to update the arc_access() state 5099 * which otherwise would be skipped for entries in the dbuf cache. 5100 */ 5101void 5102arc_buf_access(arc_buf_t *buf) 5103{ 5104 mutex_enter(&buf->b_evict_lock); 5105 arc_buf_hdr_t *hdr = buf->b_hdr; 5106 5107 /* 5108 * Avoid taking the hash_lock when possible as an optimization. 5109 * The header must be checked again under the hash_lock in order 5110 * to handle the case where it is concurrently being released. 5111 */ 5112 if (hdr->b_l1hdr.b_state == arc_anon || HDR_EMPTY(hdr)) { 5113 mutex_exit(&buf->b_evict_lock); 5114 ARCSTAT_BUMP(arcstat_access_skip); 5115 return; 5116 } 5117 5118 kmutex_t *hash_lock = HDR_LOCK(hdr); 5119 mutex_enter(hash_lock); 5120 5121 if (hdr->b_l1hdr.b_state == arc_anon || HDR_EMPTY(hdr)) { 5122 mutex_exit(hash_lock); 5123 mutex_exit(&buf->b_evict_lock); 5124 ARCSTAT_BUMP(arcstat_access_skip); 5125 return; 5126 } 5127 5128 mutex_exit(&buf->b_evict_lock); 5129 5130 ASSERT(hdr->b_l1hdr.b_state == arc_mru || 5131 hdr->b_l1hdr.b_state == arc_mfu); 5132 5133 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr); 5134 arc_access(hdr, hash_lock); 5135 mutex_exit(hash_lock); 5136 5137 ARCSTAT_BUMP(arcstat_hits); 5138 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr), 5139 demand, prefetch, !HDR_ISTYPE_METADATA(hdr), data, metadata, hits); 5140} 5141 5142/* a generic arc_done_func_t which you can use */ 5143/* ARGSUSED */ 5144void 5145arc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg) 5146{ 5147 if (zio == NULL || zio->io_error == 0) 5148 bcopy(buf->b_data, arg, arc_buf_size(buf)); 5149 arc_buf_destroy(buf, arg); 5150} 5151 5152/* a generic arc_done_func_t */ 5153void 5154arc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg) 5155{ 5156 arc_buf_t **bufp = arg; 5157 if (zio && zio->io_error) { 5158 arc_buf_destroy(buf, arg); 5159 *bufp = NULL; 5160 } else { 5161 *bufp = buf; 5162 ASSERT(buf->b_data); 5163 } 5164} 5165 5166static void 5167arc_hdr_verify(arc_buf_hdr_t *hdr, blkptr_t *bp) 5168{ 5169 if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp)) { 5170 ASSERT3U(HDR_GET_PSIZE(hdr), ==, 0); 5171 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF); 5172 } else { 5173 if (HDR_COMPRESSION_ENABLED(hdr)) { 5174 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, 5175 BP_GET_COMPRESS(bp)); 5176 } 5177 ASSERT3U(HDR_GET_LSIZE(hdr), ==, BP_GET_LSIZE(bp)); 5178 ASSERT3U(HDR_GET_PSIZE(hdr), ==, BP_GET_PSIZE(bp)); 5179 } 5180} 5181 5182static void 5183arc_read_done(zio_t *zio) 5184{ 5185 arc_buf_hdr_t *hdr = zio->io_private; 5186 kmutex_t *hash_lock = NULL; 5187 arc_callback_t *callback_list; 5188 arc_callback_t *acb; 5189 boolean_t freeable = B_FALSE; 5190 boolean_t no_zio_error = (zio->io_error == 0); 5191 5192 /* 5193 * The hdr was inserted into hash-table and removed from lists 5194 * prior to starting I/O. We should find this header, since 5195 * it's in the hash table, and it should be legit since it's 5196 * not possible to evict it during the I/O. The only possible 5197 * reason for it not to be found is if we were freed during the 5198 * read. 5199 */ 5200 if (HDR_IN_HASH_TABLE(hdr)) { 5201 ASSERT3U(hdr->b_birth, ==, BP_PHYSICAL_BIRTH(zio->io_bp)); 5202 ASSERT3U(hdr->b_dva.dva_word[0], ==, 5203 BP_IDENTITY(zio->io_bp)->dva_word[0]); 5204 ASSERT3U(hdr->b_dva.dva_word[1], ==, 5205 BP_IDENTITY(zio->io_bp)->dva_word[1]); 5206 5207 arc_buf_hdr_t *found = buf_hash_find(hdr->b_spa, zio->io_bp, 5208 &hash_lock); 5209 5210 ASSERT((found == hdr && 5211 DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) || 5212 (found == hdr && HDR_L2_READING(hdr))); 5213 ASSERT3P(hash_lock, !=, NULL); 5214 } 5215 5216 if (no_zio_error) { 5217 /* byteswap if necessary */ 5218 if (BP_SHOULD_BYTESWAP(zio->io_bp)) { 5219 if (BP_GET_LEVEL(zio->io_bp) > 0) { 5220 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_UINT64; 5221 } else { 5222 hdr->b_l1hdr.b_byteswap = 5223 DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp)); 5224 } 5225 } else { 5226 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS; 5227 } 5228 } 5229 5230 arc_hdr_clear_flags(hdr, ARC_FLAG_L2_EVICTED); 5231 if (l2arc_noprefetch && HDR_PREFETCH(hdr)) 5232 arc_hdr_clear_flags(hdr, ARC_FLAG_L2CACHE); 5233 5234 callback_list = hdr->b_l1hdr.b_acb; 5235 ASSERT3P(callback_list, !=, NULL); 5236 5237 if (hash_lock && no_zio_error && hdr->b_l1hdr.b_state == arc_anon) { 5238 /* 5239 * Only call arc_access on anonymous buffers. This is because 5240 * if we've issued an I/O for an evicted buffer, we've already 5241 * called arc_access (to prevent any simultaneous readers from 5242 * getting confused). 5243 */ 5244 arc_access(hdr, hash_lock); 5245 } 5246 5247 /* 5248 * If a read request has a callback (i.e. acb_done is not NULL), then we 5249 * make a buf containing the data according to the parameters which were 5250 * passed in. The implementation of arc_buf_alloc_impl() ensures that we 5251 * aren't needlessly decompressing the data multiple times. 5252 */ 5253 int callback_cnt = 0; 5254 for (acb = callback_list; acb != NULL; acb = acb->acb_next) { 5255 if (!acb->acb_done) 5256 continue; 5257 5258 /* This is a demand read since prefetches don't use callbacks */ 5259 callback_cnt++; 5260 5261 int error = arc_buf_alloc_impl(hdr, acb->acb_private, 5262 acb->acb_compressed, no_zio_error, &acb->acb_buf); 5263 if (no_zio_error) { 5264 zio->io_error = error; 5265 } 5266 } 5267 hdr->b_l1hdr.b_acb = NULL; 5268 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS); 5269 if (callback_cnt == 0) { 5270 ASSERT(HDR_PREFETCH(hdr)); 5271 ASSERT0(hdr->b_l1hdr.b_bufcnt); 5272 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL); 5273 } 5274 5275 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt) || 5276 callback_list != NULL); 5277 5278 if (no_zio_error) { 5279 arc_hdr_verify(hdr, zio->io_bp); 5280 } else { 5281 arc_hdr_set_flags(hdr, ARC_FLAG_IO_ERROR); 5282 if (hdr->b_l1hdr.b_state != arc_anon) 5283 arc_change_state(arc_anon, hdr, hash_lock); 5284 if (HDR_IN_HASH_TABLE(hdr)) 5285 buf_hash_remove(hdr); 5286 freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt); 5287 } 5288 5289 /* 5290 * Broadcast before we drop the hash_lock to avoid the possibility 5291 * that the hdr (and hence the cv) might be freed before we get to 5292 * the cv_broadcast(). 5293 */ 5294 cv_broadcast(&hdr->b_l1hdr.b_cv); 5295 5296 if (hash_lock != NULL) { 5297 mutex_exit(hash_lock); 5298 } else { 5299 /* 5300 * This block was freed while we waited for the read to 5301 * complete. It has been removed from the hash table and 5302 * moved to the anonymous state (so that it won't show up 5303 * in the cache). 5304 */ 5305 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon); 5306 freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt); 5307 } 5308 5309 /* execute each callback and free its structure */ 5310 while ((acb = callback_list) != NULL) { 5311 if (acb->acb_done) 5312 acb->acb_done(zio, acb->acb_buf, acb->acb_private); 5313 5314 if (acb->acb_zio_dummy != NULL) { 5315 acb->acb_zio_dummy->io_error = zio->io_error; 5316 zio_nowait(acb->acb_zio_dummy); 5317 } 5318 5319 callback_list = acb->acb_next; 5320 kmem_free(acb, sizeof (arc_callback_t)); 5321 } 5322 5323 if (freeable) 5324 arc_hdr_destroy(hdr); 5325} 5326 5327/* 5328 * "Read" the block at the specified DVA (in bp) via the 5329 * cache. If the block is found in the cache, invoke the provided 5330 * callback immediately and return. Note that the `zio' parameter 5331 * in the callback will be NULL in this case, since no IO was 5332 * required. If the block is not in the cache pass the read request 5333 * on to the spa with a substitute callback function, so that the 5334 * requested block will be added to the cache. 5335 * 5336 * If a read request arrives for a block that has a read in-progress, 5337 * either wait for the in-progress read to complete (and return the 5338 * results); or, if this is a read with a "done" func, add a record 5339 * to the read to invoke the "done" func when the read completes, 5340 * and return; or just return. 5341 * 5342 * arc_read_done() will invoke all the requested "done" functions 5343 * for readers of this block. 5344 */ 5345int 5346arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_done_func_t *done, 5347 void *private, zio_priority_t priority, int zio_flags, 5348 arc_flags_t *arc_flags, const zbookmark_phys_t *zb) 5349{ 5350 arc_buf_hdr_t *hdr = NULL; 5351 kmutex_t *hash_lock = NULL; 5352 zio_t *rzio; 5353 uint64_t guid = spa_load_guid(spa); 5354 boolean_t compressed_read = (zio_flags & ZIO_FLAG_RAW) != 0; 5355 5356 ASSERT(!BP_IS_EMBEDDED(bp) || 5357 BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA); 5358 5359top: 5360 if (!BP_IS_EMBEDDED(bp)) { 5361 /* 5362 * Embedded BP's have no DVA and require no I/O to "read". 5363 * Create an anonymous arc buf to back it. 5364 */ 5365 hdr = buf_hash_find(guid, bp, &hash_lock); 5366 } 5367 5368 if (hdr != NULL && HDR_HAS_L1HDR(hdr) && hdr->b_l1hdr.b_pabd != NULL) { 5369 arc_buf_t *buf = NULL; 5370 *arc_flags |= ARC_FLAG_CACHED; 5371 5372 if (HDR_IO_IN_PROGRESS(hdr)) { 5373 5374 if ((hdr->b_flags & ARC_FLAG_PRIO_ASYNC_READ) && 5375 priority == ZIO_PRIORITY_SYNC_READ) { 5376 /* 5377 * This sync read must wait for an 5378 * in-progress async read (e.g. a predictive 5379 * prefetch). Async reads are queued 5380 * separately at the vdev_queue layer, so 5381 * this is a form of priority inversion. 5382 * Ideally, we would "inherit" the demand 5383 * i/o's priority by moving the i/o from 5384 * the async queue to the synchronous queue, 5385 * but there is currently no mechanism to do 5386 * so. Track this so that we can evaluate 5387 * the magnitude of this potential performance 5388 * problem. 5389 * 5390 * Note that if the prefetch i/o is already 5391 * active (has been issued to the device), 5392 * the prefetch improved performance, because 5393 * we issued it sooner than we would have 5394 * without the prefetch. 5395 */ 5396 DTRACE_PROBE1(arc__sync__wait__for__async, 5397 arc_buf_hdr_t *, hdr); 5398 ARCSTAT_BUMP(arcstat_sync_wait_for_async); 5399 } 5400 if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) { 5401 arc_hdr_clear_flags(hdr, 5402 ARC_FLAG_PREDICTIVE_PREFETCH); 5403 } 5404 5405 if (*arc_flags & ARC_FLAG_WAIT) { 5406 cv_wait(&hdr->b_l1hdr.b_cv, hash_lock); 5407 mutex_exit(hash_lock); 5408 goto top; 5409 } 5410 ASSERT(*arc_flags & ARC_FLAG_NOWAIT); 5411 5412 if (done) { 5413 arc_callback_t *acb = NULL; 5414 5415 acb = kmem_zalloc(sizeof (arc_callback_t), 5416 KM_SLEEP); 5417 acb->acb_done = done; 5418 acb->acb_private = private; 5419 acb->acb_compressed = compressed_read; 5420 if (pio != NULL) 5421 acb->acb_zio_dummy = zio_null(pio, 5422 spa, NULL, NULL, NULL, zio_flags); 5423 5424 ASSERT3P(acb->acb_done, !=, NULL); 5425 acb->acb_next = hdr->b_l1hdr.b_acb; 5426 hdr->b_l1hdr.b_acb = acb; 5427 mutex_exit(hash_lock); 5428 return (0); 5429 } 5430 mutex_exit(hash_lock); 5431 return (0); 5432 } 5433 5434 ASSERT(hdr->b_l1hdr.b_state == arc_mru || 5435 hdr->b_l1hdr.b_state == arc_mfu); 5436 5437 if (done) { 5438 if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) { 5439 /* 5440 * This is a demand read which does not have to 5441 * wait for i/o because we did a predictive 5442 * prefetch i/o for it, which has completed. 5443 */ 5444 DTRACE_PROBE1( 5445 arc__demand__hit__predictive__prefetch, 5446 arc_buf_hdr_t *, hdr); 5447 ARCSTAT_BUMP( 5448 arcstat_demand_hit_predictive_prefetch); 5449 arc_hdr_clear_flags(hdr, 5450 ARC_FLAG_PREDICTIVE_PREFETCH); 5451 } 5452 ASSERT(!BP_IS_EMBEDDED(bp) || !BP_IS_HOLE(bp)); 5453 5454 /* Get a buf with the desired data in it. */ 5455 VERIFY0(arc_buf_alloc_impl(hdr, private, 5456 compressed_read, B_TRUE, &buf)); 5457 } else if (*arc_flags & ARC_FLAG_PREFETCH && 5458 refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) { 5459 arc_hdr_set_flags(hdr, ARC_FLAG_PREFETCH); 5460 } 5461 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr); 5462 arc_access(hdr, hash_lock); 5463 if (*arc_flags & ARC_FLAG_L2CACHE) 5464 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE); 5465 mutex_exit(hash_lock); 5466 ARCSTAT_BUMP(arcstat_hits); 5467 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr), 5468 demand, prefetch, !HDR_ISTYPE_METADATA(hdr), 5469 data, metadata, hits); 5470 5471 if (done) 5472 done(NULL, buf, private); 5473 } else { 5474 uint64_t lsize = BP_GET_LSIZE(bp); 5475 uint64_t psize = BP_GET_PSIZE(bp); 5476 arc_callback_t *acb; 5477 vdev_t *vd = NULL; 5478 uint64_t addr = 0; 5479 boolean_t devw = B_FALSE; 5480 uint64_t size; 5481 5482 if (hdr == NULL) { 5483 /* this block is not in the cache */ 5484 arc_buf_hdr_t *exists = NULL; 5485 arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp); 5486 hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize, 5487 BP_GET_COMPRESS(bp), type); 5488 5489 if (!BP_IS_EMBEDDED(bp)) { 5490 hdr->b_dva = *BP_IDENTITY(bp); 5491 hdr->b_birth = BP_PHYSICAL_BIRTH(bp); 5492 exists = buf_hash_insert(hdr, &hash_lock); 5493 } 5494 if (exists != NULL) { 5495 /* somebody beat us to the hash insert */ 5496 mutex_exit(hash_lock); 5497 buf_discard_identity(hdr); 5498 arc_hdr_destroy(hdr); 5499 goto top; /* restart the IO request */ 5500 } 5501 } else { 5502 /* 5503 * This block is in the ghost cache. If it was L2-only 5504 * (and thus didn't have an L1 hdr), we realloc the 5505 * header to add an L1 hdr. 5506 */ 5507 if (!HDR_HAS_L1HDR(hdr)) { 5508 hdr = arc_hdr_realloc(hdr, hdr_l2only_cache, 5509 hdr_full_cache); 5510 } 5511 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL); 5512 ASSERT(GHOST_STATE(hdr->b_l1hdr.b_state)); 5513 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 5514 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 5515 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 5516 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL); 5517 5518 /* 5519 * This is a delicate dance that we play here. 5520 * This hdr is in the ghost list so we access it 5521 * to move it out of the ghost list before we 5522 * initiate the read. If it's a prefetch then 5523 * it won't have a callback so we'll remove the 5524 * reference that arc_buf_alloc_impl() created. We 5525 * do this after we've called arc_access() to 5526 * avoid hitting an assert in remove_reference(). 5527 */ 5528 arc_access(hdr, hash_lock); 5529 arc_hdr_alloc_pabd(hdr); 5530 } 5531 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL); 5532 size = arc_hdr_size(hdr); 5533 5534 /* 5535 * If compression is enabled on the hdr, then will do 5536 * RAW I/O and will store the compressed data in the hdr's 5537 * data block. Otherwise, the hdr's data block will contain 5538 * the uncompressed data. 5539 */ 5540 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF) { 5541 zio_flags |= ZIO_FLAG_RAW; 5542 } 5543 5544 if (*arc_flags & ARC_FLAG_PREFETCH) 5545 arc_hdr_set_flags(hdr, ARC_FLAG_PREFETCH); 5546 if (*arc_flags & ARC_FLAG_L2CACHE) 5547 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE); 5548 if (BP_GET_LEVEL(bp) > 0) 5549 arc_hdr_set_flags(hdr, ARC_FLAG_INDIRECT); 5550 if (*arc_flags & ARC_FLAG_PREDICTIVE_PREFETCH) 5551 arc_hdr_set_flags(hdr, ARC_FLAG_PREDICTIVE_PREFETCH); 5552 ASSERT(!GHOST_STATE(hdr->b_l1hdr.b_state)); 5553 5554 acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP); 5555 acb->acb_done = done; 5556 acb->acb_private = private; 5557 acb->acb_compressed = compressed_read; 5558 5559 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL); 5560 hdr->b_l1hdr.b_acb = acb; 5561 arc_hdr_set_flags(hdr, ARC_FLAG_IO_IN_PROGRESS); 5562 5563 if (HDR_HAS_L2HDR(hdr) && 5564 (vd = hdr->b_l2hdr.b_dev->l2ad_vdev) != NULL) { 5565 devw = hdr->b_l2hdr.b_dev->l2ad_writing; 5566 addr = hdr->b_l2hdr.b_daddr; 5567 /* 5568 * Lock out L2ARC device removal. 5569 */ 5570 if (vdev_is_dead(vd) || 5571 !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER)) 5572 vd = NULL; 5573 } 5574 5575 if (priority == ZIO_PRIORITY_ASYNC_READ) 5576 arc_hdr_set_flags(hdr, ARC_FLAG_PRIO_ASYNC_READ); 5577 else 5578 arc_hdr_clear_flags(hdr, ARC_FLAG_PRIO_ASYNC_READ); 5579 5580 if (hash_lock != NULL) 5581 mutex_exit(hash_lock); 5582 5583 /* 5584 * At this point, we have a level 1 cache miss. Try again in 5585 * L2ARC if possible. 5586 */ 5587 ASSERT3U(HDR_GET_LSIZE(hdr), ==, lsize); 5588 5589 DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp, 5590 uint64_t, lsize, zbookmark_phys_t *, zb); 5591 ARCSTAT_BUMP(arcstat_misses); 5592 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr), 5593 demand, prefetch, !HDR_ISTYPE_METADATA(hdr), 5594 data, metadata, misses); 5595#ifdef _KERNEL 5596#ifdef RACCT 5597 if (racct_enable) { 5598 PROC_LOCK(curproc); 5599 racct_add_force(curproc, RACCT_READBPS, size); 5600 racct_add_force(curproc, RACCT_READIOPS, 1); 5601 PROC_UNLOCK(curproc); 5602 } 5603#endif /* RACCT */ 5604 curthread->td_ru.ru_inblock++; 5605#endif 5606 5607 if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) { 5608 /* 5609 * Read from the L2ARC if the following are true: 5610 * 1. The L2ARC vdev was previously cached. 5611 * 2. This buffer still has L2ARC metadata. 5612 * 3. This buffer isn't currently writing to the L2ARC. 5613 * 4. The L2ARC entry wasn't evicted, which may 5614 * also have invalidated the vdev. 5615 * 5. This isn't prefetch and l2arc_noprefetch is set. 5616 */ 5617 if (HDR_HAS_L2HDR(hdr) && 5618 !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) && 5619 !(l2arc_noprefetch && HDR_PREFETCH(hdr))) { 5620 l2arc_read_callback_t *cb; 5621 abd_t *abd; 5622 uint64_t asize; 5623 5624 DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr); 5625 ARCSTAT_BUMP(arcstat_l2_hits); 5626 5627 cb = kmem_zalloc(sizeof (l2arc_read_callback_t), 5628 KM_SLEEP); 5629 cb->l2rcb_hdr = hdr; 5630 cb->l2rcb_bp = *bp; 5631 cb->l2rcb_zb = *zb; 5632 cb->l2rcb_flags = zio_flags; 5633 5634 asize = vdev_psize_to_asize(vd, size); 5635 if (asize != size) { 5636 abd = abd_alloc_for_io(asize, 5637 HDR_ISTYPE_METADATA(hdr)); 5638 cb->l2rcb_abd = abd; 5639 } else { 5640 abd = hdr->b_l1hdr.b_pabd; 5641 } 5642 5643 ASSERT(addr >= VDEV_LABEL_START_SIZE && 5644 addr + asize <= vd->vdev_psize - 5645 VDEV_LABEL_END_SIZE); 5646 5647 /* 5648 * l2arc read. The SCL_L2ARC lock will be 5649 * released by l2arc_read_done(). 5650 * Issue a null zio if the underlying buffer 5651 * was squashed to zero size by compression. 5652 */ 5653 ASSERT3U(HDR_GET_COMPRESS(hdr), !=, 5654 ZIO_COMPRESS_EMPTY); 5655 rzio = zio_read_phys(pio, vd, addr, 5656 asize, abd, 5657 ZIO_CHECKSUM_OFF, 5658 l2arc_read_done, cb, priority, 5659 zio_flags | ZIO_FLAG_DONT_CACHE | 5660 ZIO_FLAG_CANFAIL | 5661 ZIO_FLAG_DONT_PROPAGATE | 5662 ZIO_FLAG_DONT_RETRY, B_FALSE); 5663 DTRACE_PROBE2(l2arc__read, vdev_t *, vd, 5664 zio_t *, rzio); 5665 ARCSTAT_INCR(arcstat_l2_read_bytes, size); 5666 5667 if (*arc_flags & ARC_FLAG_NOWAIT) { 5668 zio_nowait(rzio); 5669 return (0); 5670 } 5671 5672 ASSERT(*arc_flags & ARC_FLAG_WAIT); 5673 if (zio_wait(rzio) == 0) 5674 return (0); 5675 5676 /* l2arc read error; goto zio_read() */ 5677 } else { 5678 DTRACE_PROBE1(l2arc__miss, 5679 arc_buf_hdr_t *, hdr); 5680 ARCSTAT_BUMP(arcstat_l2_misses); 5681 if (HDR_L2_WRITING(hdr)) 5682 ARCSTAT_BUMP(arcstat_l2_rw_clash); 5683 spa_config_exit(spa, SCL_L2ARC, vd); 5684 } 5685 } else { 5686 if (vd != NULL) 5687 spa_config_exit(spa, SCL_L2ARC, vd); 5688 if (l2arc_ndev != 0) { 5689 DTRACE_PROBE1(l2arc__miss, 5690 arc_buf_hdr_t *, hdr); 5691 ARCSTAT_BUMP(arcstat_l2_misses); 5692 } 5693 } 5694 5695 rzio = zio_read(pio, spa, bp, hdr->b_l1hdr.b_pabd, size, 5696 arc_read_done, hdr, priority, zio_flags, zb); 5697 5698 if (*arc_flags & ARC_FLAG_WAIT) 5699 return (zio_wait(rzio)); 5700 5701 ASSERT(*arc_flags & ARC_FLAG_NOWAIT); 5702 zio_nowait(rzio); 5703 } 5704 return (0); 5705} 5706 5707/* 5708 * Notify the arc that a block was freed, and thus will never be used again. 5709 */ 5710void 5711arc_freed(spa_t *spa, const blkptr_t *bp) 5712{ 5713 arc_buf_hdr_t *hdr; 5714 kmutex_t *hash_lock; 5715 uint64_t guid = spa_load_guid(spa); 5716 5717 ASSERT(!BP_IS_EMBEDDED(bp)); 5718 5719 hdr = buf_hash_find(guid, bp, &hash_lock); 5720 if (hdr == NULL) 5721 return; 5722 5723 /* 5724 * We might be trying to free a block that is still doing I/O 5725 * (i.e. prefetch) or has a reference (i.e. a dedup-ed, 5726 * dmu_sync-ed block). If this block is being prefetched, then it 5727 * would still have the ARC_FLAG_IO_IN_PROGRESS flag set on the hdr 5728 * until the I/O completes. A block may also have a reference if it is 5729 * part of a dedup-ed, dmu_synced write. The dmu_sync() function would 5730 * have written the new block to its final resting place on disk but 5731 * without the dedup flag set. This would have left the hdr in the MRU 5732 * state and discoverable. When the txg finally syncs it detects that 5733 * the block was overridden in open context and issues an override I/O. 5734 * Since this is a dedup block, the override I/O will determine if the 5735 * block is already in the DDT. If so, then it will replace the io_bp 5736 * with the bp from the DDT and allow the I/O to finish. When the I/O 5737 * reaches the done callback, dbuf_write_override_done, it will 5738 * check to see if the io_bp and io_bp_override are identical. 5739 * If they are not, then it indicates that the bp was replaced with 5740 * the bp in the DDT and the override bp is freed. This allows 5741 * us to arrive here with a reference on a block that is being 5742 * freed. So if we have an I/O in progress, or a reference to 5743 * this hdr, then we don't destroy the hdr. 5744 */ 5745 if (!HDR_HAS_L1HDR(hdr) || (!HDR_IO_IN_PROGRESS(hdr) && 5746 refcount_is_zero(&hdr->b_l1hdr.b_refcnt))) { 5747 arc_change_state(arc_anon, hdr, hash_lock); 5748 arc_hdr_destroy(hdr); 5749 mutex_exit(hash_lock); 5750 } else { 5751 mutex_exit(hash_lock); 5752 } 5753 5754} 5755 5756/* 5757 * Release this buffer from the cache, making it an anonymous buffer. This 5758 * must be done after a read and prior to modifying the buffer contents. 5759 * If the buffer has more than one reference, we must make 5760 * a new hdr for the buffer. 5761 */ 5762void 5763arc_release(arc_buf_t *buf, void *tag) 5764{ 5765 arc_buf_hdr_t *hdr = buf->b_hdr; 5766 5767 /* 5768 * It would be nice to assert that if it's DMU metadata (level > 5769 * 0 || it's the dnode file), then it must be syncing context. 5770 * But we don't know that information at this level. 5771 */ 5772 5773 mutex_enter(&buf->b_evict_lock); 5774 5775 ASSERT(HDR_HAS_L1HDR(hdr)); 5776 5777 /* 5778 * We don't grab the hash lock prior to this check, because if 5779 * the buffer's header is in the arc_anon state, it won't be 5780 * linked into the hash table. 5781 */ 5782 if (hdr->b_l1hdr.b_state == arc_anon) { 5783 mutex_exit(&buf->b_evict_lock); 5784 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 5785 ASSERT(!HDR_IN_HASH_TABLE(hdr)); 5786 ASSERT(!HDR_HAS_L2HDR(hdr)); 5787 ASSERT(HDR_EMPTY(hdr)); 5788 ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1); 5789 ASSERT3S(refcount_count(&hdr->b_l1hdr.b_refcnt), ==, 1); 5790 ASSERT(!list_link_active(&hdr->b_l1hdr.b_arc_node)); 5791 5792 hdr->b_l1hdr.b_arc_access = 0; 5793 5794 /* 5795 * If the buf is being overridden then it may already 5796 * have a hdr that is not empty. 5797 */ 5798 buf_discard_identity(hdr); 5799 arc_buf_thaw(buf); 5800 5801 return; 5802 } 5803 5804 kmutex_t *hash_lock = HDR_LOCK(hdr); 5805 mutex_enter(hash_lock); 5806 5807 /* 5808 * This assignment is only valid as long as the hash_lock is 5809 * held, we must be careful not to reference state or the 5810 * b_state field after dropping the lock. 5811 */ 5812 arc_state_t *state = hdr->b_l1hdr.b_state; 5813 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 5814 ASSERT3P(state, !=, arc_anon); 5815 5816 /* this buffer is not on any list */ 5817 ASSERT3S(refcount_count(&hdr->b_l1hdr.b_refcnt), >, 0); 5818 5819 if (HDR_HAS_L2HDR(hdr)) { 5820 mutex_enter(&hdr->b_l2hdr.b_dev->l2ad_mtx); 5821 5822 /* 5823 * We have to recheck this conditional again now that 5824 * we're holding the l2ad_mtx to prevent a race with 5825 * another thread which might be concurrently calling 5826 * l2arc_evict(). In that case, l2arc_evict() might have 5827 * destroyed the header's L2 portion as we were waiting 5828 * to acquire the l2ad_mtx. 5829 */ 5830 if (HDR_HAS_L2HDR(hdr)) { 5831 l2arc_trim(hdr); 5832 arc_hdr_l2hdr_destroy(hdr); 5833 } 5834 5835 mutex_exit(&hdr->b_l2hdr.b_dev->l2ad_mtx); 5836 } 5837 5838 /* 5839 * Do we have more than one buf? 5840 */ 5841 if (hdr->b_l1hdr.b_bufcnt > 1) { 5842 arc_buf_hdr_t *nhdr; 5843 uint64_t spa = hdr->b_spa; 5844 uint64_t psize = HDR_GET_PSIZE(hdr); 5845 uint64_t lsize = HDR_GET_LSIZE(hdr); 5846 enum zio_compress compress = HDR_GET_COMPRESS(hdr); 5847 arc_buf_contents_t type = arc_buf_type(hdr); 5848 VERIFY3U(hdr->b_type, ==, type); 5849 5850 ASSERT(hdr->b_l1hdr.b_buf != buf || buf->b_next != NULL); 5851 (void) remove_reference(hdr, hash_lock, tag); 5852 5853 if (arc_buf_is_shared(buf) && !ARC_BUF_COMPRESSED(buf)) { 5854 ASSERT3P(hdr->b_l1hdr.b_buf, !=, buf); 5855 ASSERT(ARC_BUF_LAST(buf)); 5856 } 5857 5858 /* 5859 * Pull the data off of this hdr and attach it to 5860 * a new anonymous hdr. Also find the last buffer 5861 * in the hdr's buffer list. 5862 */ 5863 arc_buf_t *lastbuf = arc_buf_remove(hdr, buf); 5864 ASSERT3P(lastbuf, !=, NULL); 5865 5866 /* 5867 * If the current arc_buf_t and the hdr are sharing their data 5868 * buffer, then we must stop sharing that block. 5869 */ 5870 if (arc_buf_is_shared(buf)) { 5871 VERIFY(!arc_buf_is_shared(lastbuf)); 5872 5873 /* 5874 * First, sever the block sharing relationship between 5875 * buf and the arc_buf_hdr_t. 5876 */ 5877 arc_unshare_buf(hdr, buf); 5878 5879 /* 5880 * Now we need to recreate the hdr's b_pabd. Since we 5881 * have lastbuf handy, we try to share with it, but if 5882 * we can't then we allocate a new b_pabd and copy the 5883 * data from buf into it. 5884 */ 5885 if (arc_can_share(hdr, lastbuf)) { 5886 arc_share_buf(hdr, lastbuf); 5887 } else { 5888 arc_hdr_alloc_pabd(hdr); 5889 abd_copy_from_buf(hdr->b_l1hdr.b_pabd, 5890 buf->b_data, psize); 5891 } 5892 VERIFY3P(lastbuf->b_data, !=, NULL); 5893 } else if (HDR_SHARED_DATA(hdr)) { 5894 /* 5895 * Uncompressed shared buffers are always at the end 5896 * of the list. Compressed buffers don't have the 5897 * same requirements. This makes it hard to 5898 * simply assert that the lastbuf is shared so 5899 * we rely on the hdr's compression flags to determine 5900 * if we have a compressed, shared buffer. 5901 */ 5902 ASSERT(arc_buf_is_shared(lastbuf) || 5903 HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF); 5904 ASSERT(!ARC_BUF_SHARED(buf)); 5905 } 5906 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL); 5907 ASSERT3P(state, !=, arc_l2c_only); 5908 5909 (void) refcount_remove_many(&state->arcs_size, 5910 arc_buf_size(buf), buf); 5911 5912 if (refcount_is_zero(&hdr->b_l1hdr.b_refcnt)) { 5913 ASSERT3P(state, !=, arc_l2c_only); 5914 (void) refcount_remove_many(&state->arcs_esize[type], 5915 arc_buf_size(buf), buf); 5916 } 5917 5918 hdr->b_l1hdr.b_bufcnt -= 1; 5919 arc_cksum_verify(buf); 5920#ifdef illumos 5921 arc_buf_unwatch(buf); 5922#endif 5923 5924 mutex_exit(hash_lock); 5925 5926 /* 5927 * Allocate a new hdr. The new hdr will contain a b_pabd 5928 * buffer which will be freed in arc_write(). 5929 */ 5930 nhdr = arc_hdr_alloc(spa, psize, lsize, compress, type); 5931 ASSERT3P(nhdr->b_l1hdr.b_buf, ==, NULL); 5932 ASSERT0(nhdr->b_l1hdr.b_bufcnt); 5933 ASSERT0(refcount_count(&nhdr->b_l1hdr.b_refcnt)); 5934 VERIFY3U(nhdr->b_type, ==, type); 5935 ASSERT(!HDR_SHARED_DATA(nhdr)); 5936 5937 nhdr->b_l1hdr.b_buf = buf; 5938 nhdr->b_l1hdr.b_bufcnt = 1; 5939 (void) refcount_add(&nhdr->b_l1hdr.b_refcnt, tag); 5940 buf->b_hdr = nhdr; 5941 5942 mutex_exit(&buf->b_evict_lock); 5943 (void) refcount_add_many(&arc_anon->arcs_size, 5944 arc_buf_size(buf), buf); 5945 } else { 5946 mutex_exit(&buf->b_evict_lock); 5947 ASSERT(refcount_count(&hdr->b_l1hdr.b_refcnt) == 1); 5948 /* protected by hash lock, or hdr is on arc_anon */ 5949 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 5950 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 5951 arc_change_state(arc_anon, hdr, hash_lock); 5952 hdr->b_l1hdr.b_arc_access = 0; 5953 mutex_exit(hash_lock); 5954 5955 buf_discard_identity(hdr); 5956 arc_buf_thaw(buf); 5957 } 5958} 5959 5960int 5961arc_released(arc_buf_t *buf) 5962{ 5963 int released; 5964 5965 mutex_enter(&buf->b_evict_lock); 5966 released = (buf->b_data != NULL && 5967 buf->b_hdr->b_l1hdr.b_state == arc_anon); 5968 mutex_exit(&buf->b_evict_lock); 5969 return (released); 5970} 5971 5972#ifdef ZFS_DEBUG 5973int 5974arc_referenced(arc_buf_t *buf) 5975{ 5976 int referenced; 5977 5978 mutex_enter(&buf->b_evict_lock); 5979 referenced = (refcount_count(&buf->b_hdr->b_l1hdr.b_refcnt)); 5980 mutex_exit(&buf->b_evict_lock); 5981 return (referenced); 5982} 5983#endif 5984 5985static void 5986arc_write_ready(zio_t *zio) 5987{ 5988 arc_write_callback_t *callback = zio->io_private; 5989 arc_buf_t *buf = callback->awcb_buf; 5990 arc_buf_hdr_t *hdr = buf->b_hdr; 5991 uint64_t psize = BP_IS_HOLE(zio->io_bp) ? 0 : BP_GET_PSIZE(zio->io_bp); 5992 5993 ASSERT(HDR_HAS_L1HDR(hdr)); 5994 ASSERT(!refcount_is_zero(&buf->b_hdr->b_l1hdr.b_refcnt)); 5995 ASSERT(hdr->b_l1hdr.b_bufcnt > 0); 5996 5997 /* 5998 * If we're reexecuting this zio because the pool suspended, then 5999 * cleanup any state that was previously set the first time the 6000 * callback was invoked. 6001 */ 6002 if (zio->io_flags & ZIO_FLAG_REEXECUTED) { 6003 arc_cksum_free(hdr); 6004#ifdef illumos 6005 arc_buf_unwatch(buf); 6006#endif 6007 if (hdr->b_l1hdr.b_pabd != NULL) { 6008 if (arc_buf_is_shared(buf)) { 6009 arc_unshare_buf(hdr, buf); 6010 } else { 6011 arc_hdr_free_pabd(hdr); 6012 } 6013 } 6014 } 6015 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL); 6016 ASSERT(!HDR_SHARED_DATA(hdr)); 6017 ASSERT(!arc_buf_is_shared(buf)); 6018 6019 callback->awcb_ready(zio, buf, callback->awcb_private); 6020 6021 if (HDR_IO_IN_PROGRESS(hdr)) 6022 ASSERT(zio->io_flags & ZIO_FLAG_REEXECUTED); 6023 6024 arc_cksum_compute(buf); 6025 arc_hdr_set_flags(hdr, ARC_FLAG_IO_IN_PROGRESS); 6026 6027 enum zio_compress compress; 6028 if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) { 6029 compress = ZIO_COMPRESS_OFF; 6030 } else { 6031 ASSERT3U(HDR_GET_LSIZE(hdr), ==, BP_GET_LSIZE(zio->io_bp)); 6032 compress = BP_GET_COMPRESS(zio->io_bp); 6033 } 6034 HDR_SET_PSIZE(hdr, psize); 6035 arc_hdr_set_compress(hdr, compress); 6036 6037 6038 /* 6039 * Fill the hdr with data. If the hdr is compressed, the data we want 6040 * is available from the zio, otherwise we can take it from the buf. 6041 * 6042 * We might be able to share the buf's data with the hdr here. However, 6043 * doing so would cause the ARC to be full of linear ABDs if we write a 6044 * lot of shareable data. As a compromise, we check whether scattered 6045 * ABDs are allowed, and assume that if they are then the user wants 6046 * the ARC to be primarily filled with them regardless of the data being 6047 * written. Therefore, if they're allowed then we allocate one and copy 6048 * the data into it; otherwise, we share the data directly if we can. 6049 */ 6050 if (zfs_abd_scatter_enabled || !arc_can_share(hdr, buf)) { 6051 arc_hdr_alloc_pabd(hdr); 6052 6053 /* 6054 * Ideally, we would always copy the io_abd into b_pabd, but the 6055 * user may have disabled compressed ARC, thus we must check the 6056 * hdr's compression setting rather than the io_bp's. 6057 */ 6058 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF) { 6059 ASSERT3U(BP_GET_COMPRESS(zio->io_bp), !=, 6060 ZIO_COMPRESS_OFF); 6061 ASSERT3U(psize, >, 0); 6062 6063 abd_copy(hdr->b_l1hdr.b_pabd, zio->io_abd, psize); 6064 } else { 6065 ASSERT3U(zio->io_orig_size, ==, arc_hdr_size(hdr)); 6066 6067 abd_copy_from_buf(hdr->b_l1hdr.b_pabd, buf->b_data, 6068 arc_buf_size(buf)); 6069 } 6070 } else { 6071 ASSERT3P(buf->b_data, ==, abd_to_buf(zio->io_orig_abd)); 6072 ASSERT3U(zio->io_orig_size, ==, arc_buf_size(buf)); 6073 ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1); 6074 6075 arc_share_buf(hdr, buf); 6076 } 6077 6078 arc_hdr_verify(hdr, zio->io_bp); 6079} 6080 6081static void 6082arc_write_children_ready(zio_t *zio) 6083{ 6084 arc_write_callback_t *callback = zio->io_private; 6085 arc_buf_t *buf = callback->awcb_buf; 6086 6087 callback->awcb_children_ready(zio, buf, callback->awcb_private); 6088} 6089 6090/* 6091 * The SPA calls this callback for each physical write that happens on behalf 6092 * of a logical write. See the comment in dbuf_write_physdone() for details. 6093 */ 6094static void 6095arc_write_physdone(zio_t *zio) 6096{ 6097 arc_write_callback_t *cb = zio->io_private; 6098 if (cb->awcb_physdone != NULL) 6099 cb->awcb_physdone(zio, cb->awcb_buf, cb->awcb_private); 6100} 6101 6102static void 6103arc_write_done(zio_t *zio) 6104{ 6105 arc_write_callback_t *callback = zio->io_private; 6106 arc_buf_t *buf = callback->awcb_buf; 6107 arc_buf_hdr_t *hdr = buf->b_hdr; 6108 6109 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL); 6110 6111 if (zio->io_error == 0) { 6112 arc_hdr_verify(hdr, zio->io_bp); 6113 6114 if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) { 6115 buf_discard_identity(hdr); 6116 } else { 6117 hdr->b_dva = *BP_IDENTITY(zio->io_bp); 6118 hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp); 6119 } 6120 } else { 6121 ASSERT(HDR_EMPTY(hdr)); 6122 } 6123 6124 /* 6125 * If the block to be written was all-zero or compressed enough to be 6126 * embedded in the BP, no write was performed so there will be no 6127 * dva/birth/checksum. The buffer must therefore remain anonymous 6128 * (and uncached). 6129 */ 6130 if (!HDR_EMPTY(hdr)) { 6131 arc_buf_hdr_t *exists; 6132 kmutex_t *hash_lock; 6133 6134 ASSERT3U(zio->io_error, ==, 0); 6135 6136 arc_cksum_verify(buf); 6137 6138 exists = buf_hash_insert(hdr, &hash_lock); 6139 if (exists != NULL) { 6140 /* 6141 * This can only happen if we overwrite for 6142 * sync-to-convergence, because we remove 6143 * buffers from the hash table when we arc_free(). 6144 */ 6145 if (zio->io_flags & ZIO_FLAG_IO_REWRITE) { 6146 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) 6147 panic("bad overwrite, hdr=%p exists=%p", 6148 (void *)hdr, (void *)exists); 6149 ASSERT(refcount_is_zero( 6150 &exists->b_l1hdr.b_refcnt)); 6151 arc_change_state(arc_anon, exists, hash_lock); 6152 mutex_exit(hash_lock); 6153 arc_hdr_destroy(exists); 6154 exists = buf_hash_insert(hdr, &hash_lock); 6155 ASSERT3P(exists, ==, NULL); 6156 } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) { 6157 /* nopwrite */ 6158 ASSERT(zio->io_prop.zp_nopwrite); 6159 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) 6160 panic("bad nopwrite, hdr=%p exists=%p", 6161 (void *)hdr, (void *)exists); 6162 } else { 6163 /* Dedup */ 6164 ASSERT(hdr->b_l1hdr.b_bufcnt == 1); 6165 ASSERT(hdr->b_l1hdr.b_state == arc_anon); 6166 ASSERT(BP_GET_DEDUP(zio->io_bp)); 6167 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0); 6168 } 6169 } 6170 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS); 6171 /* if it's not anon, we are doing a scrub */ 6172 if (exists == NULL && hdr->b_l1hdr.b_state == arc_anon) 6173 arc_access(hdr, hash_lock); 6174 mutex_exit(hash_lock); 6175 } else { 6176 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS); 6177 } 6178 6179 ASSERT(!refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 6180 callback->awcb_done(zio, buf, callback->awcb_private); 6181 6182 abd_put(zio->io_abd); 6183 kmem_free(callback, sizeof (arc_write_callback_t)); 6184} 6185 6186zio_t * 6187arc_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, arc_buf_t *buf, 6188 boolean_t l2arc, const zio_prop_t *zp, arc_done_func_t *ready, 6189 arc_done_func_t *children_ready, arc_done_func_t *physdone, 6190 arc_done_func_t *done, void *private, zio_priority_t priority, 6191 int zio_flags, const zbookmark_phys_t *zb) 6192{ 6193 arc_buf_hdr_t *hdr = buf->b_hdr; 6194 arc_write_callback_t *callback; 6195 zio_t *zio; 6196 zio_prop_t localprop = *zp; 6197 6198 ASSERT3P(ready, !=, NULL); 6199 ASSERT3P(done, !=, NULL); 6200 ASSERT(!HDR_IO_ERROR(hdr)); 6201 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 6202 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL); 6203 ASSERT3U(hdr->b_l1hdr.b_bufcnt, >, 0); 6204 if (l2arc) 6205 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE); 6206 if (ARC_BUF_COMPRESSED(buf)) { 6207 /* 6208 * We're writing a pre-compressed buffer. Make the 6209 * compression algorithm requested by the zio_prop_t match 6210 * the pre-compressed buffer's compression algorithm. 6211 */ 6212 localprop.zp_compress = HDR_GET_COMPRESS(hdr); 6213 6214 ASSERT3U(HDR_GET_LSIZE(hdr), !=, arc_buf_size(buf)); 6215 zio_flags |= ZIO_FLAG_RAW; 6216 } 6217 callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP); 6218 callback->awcb_ready = ready; 6219 callback->awcb_children_ready = children_ready; 6220 callback->awcb_physdone = physdone; 6221 callback->awcb_done = done; 6222 callback->awcb_private = private; 6223 callback->awcb_buf = buf; 6224 6225 /* 6226 * The hdr's b_pabd is now stale, free it now. A new data block 6227 * will be allocated when the zio pipeline calls arc_write_ready(). 6228 */ 6229 if (hdr->b_l1hdr.b_pabd != NULL) { 6230 /* 6231 * If the buf is currently sharing the data block with 6232 * the hdr then we need to break that relationship here. 6233 * The hdr will remain with a NULL data pointer and the 6234 * buf will take sole ownership of the block. 6235 */ 6236 if (arc_buf_is_shared(buf)) { 6237 arc_unshare_buf(hdr, buf); 6238 } else { 6239 arc_hdr_free_pabd(hdr); 6240 } 6241 VERIFY3P(buf->b_data, !=, NULL); 6242 arc_hdr_set_compress(hdr, ZIO_COMPRESS_OFF); 6243 } 6244 ASSERT(!arc_buf_is_shared(buf)); 6245 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL); 6246 6247 zio = zio_write(pio, spa, txg, bp, 6248 abd_get_from_buf(buf->b_data, HDR_GET_LSIZE(hdr)), 6249 HDR_GET_LSIZE(hdr), arc_buf_size(buf), &localprop, arc_write_ready, 6250 (children_ready != NULL) ? arc_write_children_ready : NULL, 6251 arc_write_physdone, arc_write_done, callback, 6252 priority, zio_flags, zb); 6253 6254 return (zio); 6255} 6256 6257static int 6258arc_memory_throttle(uint64_t reserve, uint64_t txg) 6259{ 6260#ifdef _KERNEL 6261 uint64_t available_memory = ptob(freemem); 6262 static uint64_t page_load = 0; 6263 static uint64_t last_txg = 0; 6264 6265#if defined(__i386) || !defined(UMA_MD_SMALL_ALLOC) 6266 available_memory = 6267 MIN(available_memory, ptob(vmem_size(heap_arena, VMEM_FREE))); 6268#endif 6269 6270 if (freemem > (uint64_t)physmem * arc_lotsfree_percent / 100) 6271 return (0); 6272 6273 if (txg > last_txg) { 6274 last_txg = txg; 6275 page_load = 0; 6276 } 6277 /* 6278 * If we are in pageout, we know that memory is already tight, 6279 * the arc is already going to be evicting, so we just want to 6280 * continue to let page writes occur as quickly as possible. 6281 */ 6282 if (curproc == pageproc) { 6283 if (page_load > MAX(ptob(minfree), available_memory) / 4) 6284 return (SET_ERROR(ERESTART)); 6285 /* Note: reserve is inflated, so we deflate */ 6286 page_load += reserve / 8; 6287 return (0); 6288 } else if (page_load > 0 && arc_reclaim_needed()) { 6289 /* memory is low, delay before restarting */ 6290 ARCSTAT_INCR(arcstat_memory_throttle_count, 1); 6291 return (SET_ERROR(EAGAIN)); 6292 } 6293 page_load = 0; 6294#endif 6295 return (0); 6296} 6297 6298void 6299arc_tempreserve_clear(uint64_t reserve) 6300{ 6301 atomic_add_64(&arc_tempreserve, -reserve); 6302 ASSERT((int64_t)arc_tempreserve >= 0); 6303} 6304 6305int 6306arc_tempreserve_space(uint64_t reserve, uint64_t txg) 6307{ 6308 int error; 6309 uint64_t anon_size; 6310 6311 if (reserve > arc_c/4 && !arc_no_grow) { 6312 arc_c = MIN(arc_c_max, reserve * 4); 6313 DTRACE_PROBE1(arc__set_reserve, uint64_t, arc_c); 6314 } 6315 if (reserve > arc_c) 6316 return (SET_ERROR(ENOMEM)); 6317 6318 /* 6319 * Don't count loaned bufs as in flight dirty data to prevent long 6320 * network delays from blocking transactions that are ready to be 6321 * assigned to a txg. 6322 */ 6323 6324 /* assert that it has not wrapped around */ 6325 ASSERT3S(atomic_add_64_nv(&arc_loaned_bytes, 0), >=, 0); 6326 6327 anon_size = MAX((int64_t)(refcount_count(&arc_anon->arcs_size) - 6328 arc_loaned_bytes), 0); 6329 6330 /* 6331 * Writes will, almost always, require additional memory allocations 6332 * in order to compress/encrypt/etc the data. We therefore need to 6333 * make sure that there is sufficient available memory for this. 6334 */ 6335 error = arc_memory_throttle(reserve, txg); 6336 if (error != 0) 6337 return (error); 6338 6339 /* 6340 * Throttle writes when the amount of dirty data in the cache 6341 * gets too large. We try to keep the cache less than half full 6342 * of dirty blocks so that our sync times don't grow too large. 6343 * Note: if two requests come in concurrently, we might let them 6344 * both succeed, when one of them should fail. Not a huge deal. 6345 */ 6346 6347 if (reserve + arc_tempreserve + anon_size > arc_c / 2 && 6348 anon_size > arc_c / 4) { 6349 uint64_t meta_esize = 6350 refcount_count(&arc_anon->arcs_esize[ARC_BUFC_METADATA]); 6351 uint64_t data_esize = 6352 refcount_count(&arc_anon->arcs_esize[ARC_BUFC_DATA]); 6353 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK " 6354 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n", 6355 arc_tempreserve >> 10, meta_esize >> 10, 6356 data_esize >> 10, reserve >> 10, arc_c >> 10); 6357 return (SET_ERROR(ERESTART)); 6358 } 6359 atomic_add_64(&arc_tempreserve, reserve); 6360 return (0); 6361} 6362 6363static void 6364arc_kstat_update_state(arc_state_t *state, kstat_named_t *size, 6365 kstat_named_t *evict_data, kstat_named_t *evict_metadata) 6366{ 6367 size->value.ui64 = refcount_count(&state->arcs_size); 6368 evict_data->value.ui64 = 6369 refcount_count(&state->arcs_esize[ARC_BUFC_DATA]); 6370 evict_metadata->value.ui64 = 6371 refcount_count(&state->arcs_esize[ARC_BUFC_METADATA]); 6372} 6373 6374static int 6375arc_kstat_update(kstat_t *ksp, int rw) 6376{ 6377 arc_stats_t *as = ksp->ks_data; 6378 6379 if (rw == KSTAT_WRITE) { 6380 return (EACCES); 6381 } else { 6382 arc_kstat_update_state(arc_anon, 6383 &as->arcstat_anon_size, 6384 &as->arcstat_anon_evictable_data, 6385 &as->arcstat_anon_evictable_metadata); 6386 arc_kstat_update_state(arc_mru, 6387 &as->arcstat_mru_size, 6388 &as->arcstat_mru_evictable_data, 6389 &as->arcstat_mru_evictable_metadata); 6390 arc_kstat_update_state(arc_mru_ghost, 6391 &as->arcstat_mru_ghost_size, 6392 &as->arcstat_mru_ghost_evictable_data, 6393 &as->arcstat_mru_ghost_evictable_metadata); 6394 arc_kstat_update_state(arc_mfu, 6395 &as->arcstat_mfu_size, 6396 &as->arcstat_mfu_evictable_data, 6397 &as->arcstat_mfu_evictable_metadata); 6398 arc_kstat_update_state(arc_mfu_ghost, 6399 &as->arcstat_mfu_ghost_size, 6400 &as->arcstat_mfu_ghost_evictable_data, 6401 &as->arcstat_mfu_ghost_evictable_metadata); 6402 6403 ARCSTAT(arcstat_size) = aggsum_value(&arc_size); 6404 ARCSTAT(arcstat_meta_used) = aggsum_value(&arc_meta_used); 6405 ARCSTAT(arcstat_data_size) = aggsum_value(&astat_data_size); 6406 ARCSTAT(arcstat_metadata_size) = 6407 aggsum_value(&astat_metadata_size); 6408 ARCSTAT(arcstat_hdr_size) = aggsum_value(&astat_hdr_size); 6409 ARCSTAT(arcstat_other_size) = aggsum_value(&astat_other_size); 6410 ARCSTAT(arcstat_l2_hdr_size) = aggsum_value(&astat_l2_hdr_size); 6411 } 6412 6413 return (0); 6414} 6415 6416/* 6417 * This function *must* return indices evenly distributed between all 6418 * sublists of the multilist. This is needed due to how the ARC eviction 6419 * code is laid out; arc_evict_state() assumes ARC buffers are evenly 6420 * distributed between all sublists and uses this assumption when 6421 * deciding which sublist to evict from and how much to evict from it. 6422 */ 6423unsigned int 6424arc_state_multilist_index_func(multilist_t *ml, void *obj) 6425{ 6426 arc_buf_hdr_t *hdr = obj; 6427 6428 /* 6429 * We rely on b_dva to generate evenly distributed index 6430 * numbers using buf_hash below. So, as an added precaution, 6431 * let's make sure we never add empty buffers to the arc lists. 6432 */ 6433 ASSERT(!HDR_EMPTY(hdr)); 6434 6435 /* 6436 * The assumption here, is the hash value for a given 6437 * arc_buf_hdr_t will remain constant throughout it's lifetime 6438 * (i.e. it's b_spa, b_dva, and b_birth fields don't change). 6439 * Thus, we don't need to store the header's sublist index 6440 * on insertion, as this index can be recalculated on removal. 6441 * 6442 * Also, the low order bits of the hash value are thought to be 6443 * distributed evenly. Otherwise, in the case that the multilist 6444 * has a power of two number of sublists, each sublists' usage 6445 * would not be evenly distributed. 6446 */ 6447 return (buf_hash(hdr->b_spa, &hdr->b_dva, hdr->b_birth) % 6448 multilist_get_num_sublists(ml)); 6449} 6450 6451#ifdef _KERNEL 6452static eventhandler_tag arc_event_lowmem = NULL; 6453 6454static void 6455arc_lowmem(void *arg __unused, int howto __unused) 6456{ 6457 6458 mutex_enter(&arc_reclaim_lock); 6459 DTRACE_PROBE1(arc__needfree, int64_t, ((int64_t)freemem - zfs_arc_free_target) * PAGESIZE); 6460 cv_signal(&arc_reclaim_thread_cv); 6461 6462 /* 6463 * It is unsafe to block here in arbitrary threads, because we can come 6464 * here from ARC itself and may hold ARC locks and thus risk a deadlock 6465 * with ARC reclaim thread. 6466 */ 6467 if (curproc == pageproc) 6468 (void) cv_wait(&arc_reclaim_waiters_cv, &arc_reclaim_lock); 6469 mutex_exit(&arc_reclaim_lock); 6470} 6471#endif 6472 6473static void 6474arc_state_init(void) 6475{ 6476 arc_anon = &ARC_anon; 6477 arc_mru = &ARC_mru; 6478 arc_mru_ghost = &ARC_mru_ghost; 6479 arc_mfu = &ARC_mfu; 6480 arc_mfu_ghost = &ARC_mfu_ghost; 6481 arc_l2c_only = &ARC_l2c_only; 6482 6483 arc_mru->arcs_list[ARC_BUFC_METADATA] = 6484 multilist_create(sizeof (arc_buf_hdr_t), 6485 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 6486 arc_state_multilist_index_func); 6487 arc_mru->arcs_list[ARC_BUFC_DATA] = 6488 multilist_create(sizeof (arc_buf_hdr_t), 6489 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 6490 arc_state_multilist_index_func); 6491 arc_mru_ghost->arcs_list[ARC_BUFC_METADATA] = 6492 multilist_create(sizeof (arc_buf_hdr_t), 6493 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 6494 arc_state_multilist_index_func); 6495 arc_mru_ghost->arcs_list[ARC_BUFC_DATA] = 6496 multilist_create(sizeof (arc_buf_hdr_t), 6497 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 6498 arc_state_multilist_index_func); 6499 arc_mfu->arcs_list[ARC_BUFC_METADATA] = 6500 multilist_create(sizeof (arc_buf_hdr_t), 6501 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 6502 arc_state_multilist_index_func); 6503 arc_mfu->arcs_list[ARC_BUFC_DATA] = 6504 multilist_create(sizeof (arc_buf_hdr_t), 6505 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 6506 arc_state_multilist_index_func); 6507 arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA] = 6508 multilist_create(sizeof (arc_buf_hdr_t), 6509 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 6510 arc_state_multilist_index_func); 6511 arc_mfu_ghost->arcs_list[ARC_BUFC_DATA] = 6512 multilist_create(sizeof (arc_buf_hdr_t), 6513 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 6514 arc_state_multilist_index_func); 6515 arc_l2c_only->arcs_list[ARC_BUFC_METADATA] = 6516 multilist_create(sizeof (arc_buf_hdr_t), 6517 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 6518 arc_state_multilist_index_func); 6519 arc_l2c_only->arcs_list[ARC_BUFC_DATA] = 6520 multilist_create(sizeof (arc_buf_hdr_t), 6521 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 6522 arc_state_multilist_index_func); 6523 6524 refcount_create(&arc_anon->arcs_esize[ARC_BUFC_METADATA]); 6525 refcount_create(&arc_anon->arcs_esize[ARC_BUFC_DATA]); 6526 refcount_create(&arc_mru->arcs_esize[ARC_BUFC_METADATA]); 6527 refcount_create(&arc_mru->arcs_esize[ARC_BUFC_DATA]); 6528 refcount_create(&arc_mru_ghost->arcs_esize[ARC_BUFC_METADATA]); 6529 refcount_create(&arc_mru_ghost->arcs_esize[ARC_BUFC_DATA]); 6530 refcount_create(&arc_mfu->arcs_esize[ARC_BUFC_METADATA]); 6531 refcount_create(&arc_mfu->arcs_esize[ARC_BUFC_DATA]); 6532 refcount_create(&arc_mfu_ghost->arcs_esize[ARC_BUFC_METADATA]); 6533 refcount_create(&arc_mfu_ghost->arcs_esize[ARC_BUFC_DATA]); 6534 refcount_create(&arc_l2c_only->arcs_esize[ARC_BUFC_METADATA]); 6535 refcount_create(&arc_l2c_only->arcs_esize[ARC_BUFC_DATA]); 6536 6537 refcount_create(&arc_anon->arcs_size); 6538 refcount_create(&arc_mru->arcs_size); 6539 refcount_create(&arc_mru_ghost->arcs_size); 6540 refcount_create(&arc_mfu->arcs_size); 6541 refcount_create(&arc_mfu_ghost->arcs_size); 6542 refcount_create(&arc_l2c_only->arcs_size); 6543 6544 aggsum_init(&arc_meta_used, 0); 6545 aggsum_init(&arc_size, 0); 6546 aggsum_init(&astat_data_size, 0); 6547 aggsum_init(&astat_metadata_size, 0); 6548 aggsum_init(&astat_hdr_size, 0); 6549 aggsum_init(&astat_other_size, 0); 6550 aggsum_init(&astat_l2_hdr_size, 0); 6551} 6552 6553static void 6554arc_state_fini(void) 6555{ 6556 refcount_destroy(&arc_anon->arcs_esize[ARC_BUFC_METADATA]); 6557 refcount_destroy(&arc_anon->arcs_esize[ARC_BUFC_DATA]); 6558 refcount_destroy(&arc_mru->arcs_esize[ARC_BUFC_METADATA]); 6559 refcount_destroy(&arc_mru->arcs_esize[ARC_BUFC_DATA]); 6560 refcount_destroy(&arc_mru_ghost->arcs_esize[ARC_BUFC_METADATA]); 6561 refcount_destroy(&arc_mru_ghost->arcs_esize[ARC_BUFC_DATA]); 6562 refcount_destroy(&arc_mfu->arcs_esize[ARC_BUFC_METADATA]); 6563 refcount_destroy(&arc_mfu->arcs_esize[ARC_BUFC_DATA]); 6564 refcount_destroy(&arc_mfu_ghost->arcs_esize[ARC_BUFC_METADATA]); 6565 refcount_destroy(&arc_mfu_ghost->arcs_esize[ARC_BUFC_DATA]); 6566 refcount_destroy(&arc_l2c_only->arcs_esize[ARC_BUFC_METADATA]); 6567 refcount_destroy(&arc_l2c_only->arcs_esize[ARC_BUFC_DATA]); 6568 6569 refcount_destroy(&arc_anon->arcs_size); 6570 refcount_destroy(&arc_mru->arcs_size); 6571 refcount_destroy(&arc_mru_ghost->arcs_size); 6572 refcount_destroy(&arc_mfu->arcs_size); 6573 refcount_destroy(&arc_mfu_ghost->arcs_size); 6574 refcount_destroy(&arc_l2c_only->arcs_size); 6575 6576 multilist_destroy(arc_mru->arcs_list[ARC_BUFC_METADATA]); 6577 multilist_destroy(arc_mru_ghost->arcs_list[ARC_BUFC_METADATA]); 6578 multilist_destroy(arc_mfu->arcs_list[ARC_BUFC_METADATA]); 6579 multilist_destroy(arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA]); 6580 multilist_destroy(arc_mru->arcs_list[ARC_BUFC_DATA]); 6581 multilist_destroy(arc_mru_ghost->arcs_list[ARC_BUFC_DATA]); 6582 multilist_destroy(arc_mfu->arcs_list[ARC_BUFC_DATA]); 6583 multilist_destroy(arc_mfu_ghost->arcs_list[ARC_BUFC_DATA]); 6584} 6585 6586uint64_t 6587arc_max_bytes(void) 6588{ 6589 return (arc_c_max); 6590} 6591 6592void 6593arc_init(void) 6594{ 6595 int i, prefetch_tunable_set = 0; 6596 6597 /* 6598 * allmem is "all memory that we could possibly use". 6599 */ 6600#ifdef illumos 6601#ifdef _KERNEL 6602 uint64_t allmem = ptob(physmem - swapfs_minfree); 6603#else 6604 uint64_t allmem = (physmem * PAGESIZE) / 2; 6605#endif 6606#else 6607 uint64_t allmem = kmem_size(); 6608#endif 6609 6610 6611 mutex_init(&arc_reclaim_lock, NULL, MUTEX_DEFAULT, NULL); 6612 cv_init(&arc_reclaim_thread_cv, NULL, CV_DEFAULT, NULL); 6613 cv_init(&arc_reclaim_waiters_cv, NULL, CV_DEFAULT, NULL); 6614 6615 mutex_init(&arc_dnlc_evicts_lock, NULL, MUTEX_DEFAULT, NULL); 6616 cv_init(&arc_dnlc_evicts_cv, NULL, CV_DEFAULT, NULL); 6617 6618 /* Convert seconds to clock ticks */ 6619 arc_min_prefetch_lifespan = 1 * hz; 6620 6621 /* set min cache to 1/32 of all memory, or arc_abs_min, whichever is more */ 6622 arc_c_min = MAX(allmem / 32, arc_abs_min); 6623 /* set max to 5/8 of all memory, or all but 1GB, whichever is more */ 6624 if (allmem >= 1 << 30) 6625 arc_c_max = allmem - (1 << 30); 6626 else 6627 arc_c_max = arc_c_min; 6628 arc_c_max = MAX(allmem * 5 / 8, arc_c_max); 6629 6630 /* 6631 * In userland, there's only the memory pressure that we artificially 6632 * create (see arc_available_memory()). Don't let arc_c get too 6633 * small, because it can cause transactions to be larger than 6634 * arc_c, causing arc_tempreserve_space() to fail. 6635 */ 6636#ifndef _KERNEL 6637 arc_c_min = arc_c_max / 2; 6638#endif 6639 6640#ifdef _KERNEL 6641 /* 6642 * Allow the tunables to override our calculations if they are 6643 * reasonable. 6644 */ 6645 if (zfs_arc_max > arc_abs_min && zfs_arc_max < allmem) { 6646 arc_c_max = zfs_arc_max; 6647 arc_c_min = MIN(arc_c_min, arc_c_max); 6648 } 6649 if (zfs_arc_min > arc_abs_min && zfs_arc_min <= arc_c_max) 6650 arc_c_min = zfs_arc_min; 6651#endif 6652 6653 arc_c = arc_c_max; 6654 arc_p = (arc_c >> 1); 6655 6656 /* limit meta-data to 1/4 of the arc capacity */ 6657 arc_meta_limit = arc_c_max / 4; 6658 6659#ifdef _KERNEL 6660 /* 6661 * Metadata is stored in the kernel's heap. Don't let us 6662 * use more than half the heap for the ARC. 6663 */ 6664 arc_meta_limit = MIN(arc_meta_limit, 6665 vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 2); 6666#endif 6667 6668 /* Allow the tunable to override if it is reasonable */ 6669 if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max) 6670 arc_meta_limit = zfs_arc_meta_limit; 6671 6672 if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0) 6673 arc_c_min = arc_meta_limit / 2; 6674 6675 if (zfs_arc_meta_min > 0) { 6676 arc_meta_min = zfs_arc_meta_min; 6677 } else { 6678 arc_meta_min = arc_c_min / 2; 6679 } 6680 6681 if (zfs_arc_grow_retry > 0) 6682 arc_grow_retry = zfs_arc_grow_retry; 6683 6684 if (zfs_arc_shrink_shift > 0) 6685 arc_shrink_shift = zfs_arc_shrink_shift; 6686 6687 if (zfs_arc_no_grow_shift > 0) 6688 arc_no_grow_shift = zfs_arc_no_grow_shift; 6689 /* 6690 * Ensure that arc_no_grow_shift is less than arc_shrink_shift. 6691 */ 6692 if (arc_no_grow_shift >= arc_shrink_shift) 6693 arc_no_grow_shift = arc_shrink_shift - 1; 6694 6695 if (zfs_arc_p_min_shift > 0) 6696 arc_p_min_shift = zfs_arc_p_min_shift; 6697 6698 /* if kmem_flags are set, lets try to use less memory */ 6699 if (kmem_debugging()) 6700 arc_c = arc_c / 2; 6701 if (arc_c < arc_c_min) 6702 arc_c = arc_c_min; 6703 6704 zfs_arc_min = arc_c_min; 6705 zfs_arc_max = arc_c_max; 6706 6707 arc_state_init(); 6708 buf_init(); 6709 6710 arc_reclaim_thread_exit = B_FALSE; 6711 arc_dnlc_evicts_thread_exit = FALSE; 6712 6713 arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED, 6714 sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL); 6715 6716 if (arc_ksp != NULL) { 6717 arc_ksp->ks_data = &arc_stats; 6718 arc_ksp->ks_update = arc_kstat_update; 6719 kstat_install(arc_ksp); 6720 } 6721 6722 (void) thread_create(NULL, 0, arc_reclaim_thread, NULL, 0, &p0, 6723 TS_RUN, minclsyspri); 6724 6725#ifdef _KERNEL 6726 arc_event_lowmem = EVENTHANDLER_REGISTER(vm_lowmem, arc_lowmem, NULL, 6727 EVENTHANDLER_PRI_FIRST); 6728#endif 6729 6730 (void) thread_create(NULL, 0, arc_dnlc_evicts_thread, NULL, 0, &p0, 6731 TS_RUN, minclsyspri); 6732 6733 arc_dead = B_FALSE; 6734 arc_warm = B_FALSE; 6735 6736 /* 6737 * Calculate maximum amount of dirty data per pool. 6738 * 6739 * If it has been set by /etc/system, take that. 6740 * Otherwise, use a percentage of physical memory defined by 6741 * zfs_dirty_data_max_percent (default 10%) with a cap at 6742 * zfs_dirty_data_max_max (default 4GB). 6743 */ 6744 if (zfs_dirty_data_max == 0) { 6745 zfs_dirty_data_max = ptob(physmem) * 6746 zfs_dirty_data_max_percent / 100; 6747 zfs_dirty_data_max = MIN(zfs_dirty_data_max, 6748 zfs_dirty_data_max_max); 6749 } 6750 6751#ifdef _KERNEL 6752 if (TUNABLE_INT_FETCH("vfs.zfs.prefetch_disable", &zfs_prefetch_disable)) 6753 prefetch_tunable_set = 1; 6754 6755#ifdef __i386__ 6756 if (prefetch_tunable_set == 0) { 6757 printf("ZFS NOTICE: Prefetch is disabled by default on i386 " 6758 "-- to enable,\n"); 6759 printf(" add \"vfs.zfs.prefetch_disable=0\" " 6760 "to /boot/loader.conf.\n"); 6761 zfs_prefetch_disable = 1; 6762 } 6763#else 6764 if ((((uint64_t)physmem * PAGESIZE) < (1ULL << 32)) && 6765 prefetch_tunable_set == 0) { 6766 printf("ZFS NOTICE: Prefetch is disabled by default if less " 6767 "than 4GB of RAM is present;\n" 6768 " to enable, add \"vfs.zfs.prefetch_disable=0\" " 6769 "to /boot/loader.conf.\n"); 6770 zfs_prefetch_disable = 1; 6771 } 6772#endif 6773 /* Warn about ZFS memory and address space requirements. */ 6774 if (((uint64_t)physmem * PAGESIZE) < (256 + 128 + 64) * (1 << 20)) { 6775 printf("ZFS WARNING: Recommended minimum RAM size is 512MB; " 6776 "expect unstable behavior.\n"); 6777 } 6778 if (allmem < 512 * (1 << 20)) { 6779 printf("ZFS WARNING: Recommended minimum kmem_size is 512MB; " 6780 "expect unstable behavior.\n"); 6781 printf(" Consider tuning vm.kmem_size and " 6782 "vm.kmem_size_max\n"); 6783 printf(" in /boot/loader.conf.\n"); 6784 } 6785#endif 6786} 6787 6788void 6789arc_fini(void) 6790{ 6791#ifdef _KERNEL 6792 if (arc_event_lowmem != NULL) 6793 EVENTHANDLER_DEREGISTER(vm_lowmem, arc_event_lowmem); 6794#endif 6795 6796 mutex_enter(&arc_reclaim_lock); 6797 arc_reclaim_thread_exit = B_TRUE; 6798 /* 6799 * The reclaim thread will set arc_reclaim_thread_exit back to 6800 * B_FALSE when it is finished exiting; we're waiting for that. 6801 */ 6802 while (arc_reclaim_thread_exit) { 6803 cv_signal(&arc_reclaim_thread_cv); 6804 cv_wait(&arc_reclaim_thread_cv, &arc_reclaim_lock); 6805 } 6806 mutex_exit(&arc_reclaim_lock); 6807 6808 /* Use B_TRUE to ensure *all* buffers are evicted */ 6809 arc_flush(NULL, B_TRUE); 6810 6811 mutex_enter(&arc_dnlc_evicts_lock); 6812 arc_dnlc_evicts_thread_exit = TRUE; 6813 /* 6814 * The user evicts thread will set arc_user_evicts_thread_exit 6815 * to FALSE when it is finished exiting; we're waiting for that. 6816 */ 6817 while (arc_dnlc_evicts_thread_exit) { 6818 cv_signal(&arc_dnlc_evicts_cv); 6819 cv_wait(&arc_dnlc_evicts_cv, &arc_dnlc_evicts_lock); 6820 } 6821 mutex_exit(&arc_dnlc_evicts_lock); 6822 6823 arc_dead = B_TRUE; 6824 6825 if (arc_ksp != NULL) { 6826 kstat_delete(arc_ksp); 6827 arc_ksp = NULL; 6828 } 6829 6830 mutex_destroy(&arc_reclaim_lock); 6831 cv_destroy(&arc_reclaim_thread_cv); 6832 cv_destroy(&arc_reclaim_waiters_cv); 6833 6834 mutex_destroy(&arc_dnlc_evicts_lock); 6835 cv_destroy(&arc_dnlc_evicts_cv); 6836 6837 arc_state_fini(); 6838 buf_fini(); 6839 6840 ASSERT0(arc_loaned_bytes); 6841} 6842 6843/* 6844 * Level 2 ARC 6845 * 6846 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk. 6847 * It uses dedicated storage devices to hold cached data, which are populated 6848 * using large infrequent writes. The main role of this cache is to boost 6849 * the performance of random read workloads. The intended L2ARC devices 6850 * include short-stroked disks, solid state disks, and other media with 6851 * substantially faster read latency than disk. 6852 * 6853 * +-----------------------+ 6854 * | ARC | 6855 * +-----------------------+ 6856 * | ^ ^ 6857 * | | | 6858 * l2arc_feed_thread() arc_read() 6859 * | | | 6860 * | l2arc read | 6861 * V | | 6862 * +---------------+ | 6863 * | L2ARC | | 6864 * +---------------+ | 6865 * | ^ | 6866 * l2arc_write() | | 6867 * | | | 6868 * V | | 6869 * +-------+ +-------+ 6870 * | vdev | | vdev | 6871 * | cache | | cache | 6872 * +-------+ +-------+ 6873 * +=========+ .-----. 6874 * : L2ARC : |-_____-| 6875 * : devices : | Disks | 6876 * +=========+ `-_____-' 6877 * 6878 * Read requests are satisfied from the following sources, in order: 6879 * 6880 * 1) ARC 6881 * 2) vdev cache of L2ARC devices 6882 * 3) L2ARC devices 6883 * 4) vdev cache of disks 6884 * 5) disks 6885 * 6886 * Some L2ARC device types exhibit extremely slow write performance. 6887 * To accommodate for this there are some significant differences between 6888 * the L2ARC and traditional cache design: 6889 * 6890 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from 6891 * the ARC behave as usual, freeing buffers and placing headers on ghost 6892 * lists. The ARC does not send buffers to the L2ARC during eviction as 6893 * this would add inflated write latencies for all ARC memory pressure. 6894 * 6895 * 2. The L2ARC attempts to cache data from the ARC before it is evicted. 6896 * It does this by periodically scanning buffers from the eviction-end of 6897 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are 6898 * not already there. It scans until a headroom of buffers is satisfied, 6899 * which itself is a buffer for ARC eviction. If a compressible buffer is 6900 * found during scanning and selected for writing to an L2ARC device, we 6901 * temporarily boost scanning headroom during the next scan cycle to make 6902 * sure we adapt to compression effects (which might significantly reduce 6903 * the data volume we write to L2ARC). The thread that does this is 6904 * l2arc_feed_thread(), illustrated below; example sizes are included to 6905 * provide a better sense of ratio than this diagram: 6906 * 6907 * head --> tail 6908 * +---------------------+----------+ 6909 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC 6910 * +---------------------+----------+ | o L2ARC eligible 6911 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer 6912 * +---------------------+----------+ | 6913 * 15.9 Gbytes ^ 32 Mbytes | 6914 * headroom | 6915 * l2arc_feed_thread() 6916 * | 6917 * l2arc write hand <--[oooo]--' 6918 * | 8 Mbyte 6919 * | write max 6920 * V 6921 * +==============================+ 6922 * L2ARC dev |####|#|###|###| |####| ... | 6923 * +==============================+ 6924 * 32 Gbytes 6925 * 6926 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of 6927 * evicted, then the L2ARC has cached a buffer much sooner than it probably 6928 * needed to, potentially wasting L2ARC device bandwidth and storage. It is 6929 * safe to say that this is an uncommon case, since buffers at the end of 6930 * the ARC lists have moved there due to inactivity. 6931 * 6932 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom, 6933 * then the L2ARC simply misses copying some buffers. This serves as a 6934 * pressure valve to prevent heavy read workloads from both stalling the ARC 6935 * with waits and clogging the L2ARC with writes. This also helps prevent 6936 * the potential for the L2ARC to churn if it attempts to cache content too 6937 * quickly, such as during backups of the entire pool. 6938 * 6939 * 5. After system boot and before the ARC has filled main memory, there are 6940 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru 6941 * lists can remain mostly static. Instead of searching from tail of these 6942 * lists as pictured, the l2arc_feed_thread() will search from the list heads 6943 * for eligible buffers, greatly increasing its chance of finding them. 6944 * 6945 * The L2ARC device write speed is also boosted during this time so that 6946 * the L2ARC warms up faster. Since there have been no ARC evictions yet, 6947 * there are no L2ARC reads, and no fear of degrading read performance 6948 * through increased writes. 6949 * 6950 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that 6951 * the vdev queue can aggregate them into larger and fewer writes. Each 6952 * device is written to in a rotor fashion, sweeping writes through 6953 * available space then repeating. 6954 * 6955 * 7. The L2ARC does not store dirty content. It never needs to flush 6956 * write buffers back to disk based storage. 6957 * 6958 * 8. If an ARC buffer is written (and dirtied) which also exists in the 6959 * L2ARC, the now stale L2ARC buffer is immediately dropped. 6960 * 6961 * The performance of the L2ARC can be tweaked by a number of tunables, which 6962 * may be necessary for different workloads: 6963 * 6964 * l2arc_write_max max write bytes per interval 6965 * l2arc_write_boost extra write bytes during device warmup 6966 * l2arc_noprefetch skip caching prefetched buffers 6967 * l2arc_headroom number of max device writes to precache 6968 * l2arc_headroom_boost when we find compressed buffers during ARC 6969 * scanning, we multiply headroom by this 6970 * percentage factor for the next scan cycle, 6971 * since more compressed buffers are likely to 6972 * be present 6973 * l2arc_feed_secs seconds between L2ARC writing 6974 * 6975 * Tunables may be removed or added as future performance improvements are 6976 * integrated, and also may become zpool properties. 6977 * 6978 * There are three key functions that control how the L2ARC warms up: 6979 * 6980 * l2arc_write_eligible() check if a buffer is eligible to cache 6981 * l2arc_write_size() calculate how much to write 6982 * l2arc_write_interval() calculate sleep delay between writes 6983 * 6984 * These three functions determine what to write, how much, and how quickly 6985 * to send writes. 6986 */ 6987 6988static boolean_t 6989l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *hdr) 6990{ 6991 /* 6992 * A buffer is *not* eligible for the L2ARC if it: 6993 * 1. belongs to a different spa. 6994 * 2. is already cached on the L2ARC. 6995 * 3. has an I/O in progress (it may be an incomplete read). 6996 * 4. is flagged not eligible (zfs property). 6997 */ 6998 if (hdr->b_spa != spa_guid) { 6999 ARCSTAT_BUMP(arcstat_l2_write_spa_mismatch); 7000 return (B_FALSE); 7001 } 7002 if (HDR_HAS_L2HDR(hdr)) { 7003 ARCSTAT_BUMP(arcstat_l2_write_in_l2); 7004 return (B_FALSE); 7005 } 7006 if (HDR_IO_IN_PROGRESS(hdr)) { 7007 ARCSTAT_BUMP(arcstat_l2_write_hdr_io_in_progress); 7008 return (B_FALSE); 7009 } 7010 if (!HDR_L2CACHE(hdr)) { 7011 ARCSTAT_BUMP(arcstat_l2_write_not_cacheable); 7012 return (B_FALSE); 7013 } 7014 7015 return (B_TRUE); 7016} 7017 7018static uint64_t 7019l2arc_write_size(void) 7020{ 7021 uint64_t size; 7022 7023 /* 7024 * Make sure our globals have meaningful values in case the user 7025 * altered them. 7026 */ 7027 size = l2arc_write_max; 7028 if (size == 0) { 7029 cmn_err(CE_NOTE, "Bad value for l2arc_write_max, value must " 7030 "be greater than zero, resetting it to the default (%d)", 7031 L2ARC_WRITE_SIZE); 7032 size = l2arc_write_max = L2ARC_WRITE_SIZE; 7033 } 7034 7035 if (arc_warm == B_FALSE) 7036 size += l2arc_write_boost; 7037 7038 return (size); 7039 7040} 7041 7042static clock_t 7043l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote) 7044{ 7045 clock_t interval, next, now; 7046 7047 /* 7048 * If the ARC lists are busy, increase our write rate; if the 7049 * lists are stale, idle back. This is achieved by checking 7050 * how much we previously wrote - if it was more than half of 7051 * what we wanted, schedule the next write much sooner. 7052 */ 7053 if (l2arc_feed_again && wrote > (wanted / 2)) 7054 interval = (hz * l2arc_feed_min_ms) / 1000; 7055 else 7056 interval = hz * l2arc_feed_secs; 7057 7058 now = ddi_get_lbolt(); 7059 next = MAX(now, MIN(now + interval, began + interval)); 7060 7061 return (next); 7062} 7063 7064/* 7065 * Cycle through L2ARC devices. This is how L2ARC load balances. 7066 * If a device is returned, this also returns holding the spa config lock. 7067 */ 7068static l2arc_dev_t * 7069l2arc_dev_get_next(void) 7070{ 7071 l2arc_dev_t *first, *next = NULL; 7072 7073 /* 7074 * Lock out the removal of spas (spa_namespace_lock), then removal 7075 * of cache devices (l2arc_dev_mtx). Once a device has been selected, 7076 * both locks will be dropped and a spa config lock held instead. 7077 */ 7078 mutex_enter(&spa_namespace_lock); 7079 mutex_enter(&l2arc_dev_mtx); 7080 7081 /* if there are no vdevs, there is nothing to do */ 7082 if (l2arc_ndev == 0) 7083 goto out; 7084 7085 first = NULL; 7086 next = l2arc_dev_last; 7087 do { 7088 /* loop around the list looking for a non-faulted vdev */ 7089 if (next == NULL) { 7090 next = list_head(l2arc_dev_list); 7091 } else { 7092 next = list_next(l2arc_dev_list, next); 7093 if (next == NULL) 7094 next = list_head(l2arc_dev_list); 7095 } 7096 7097 /* if we have come back to the start, bail out */ 7098 if (first == NULL) 7099 first = next; 7100 else if (next == first) 7101 break; 7102 7103 } while (vdev_is_dead(next->l2ad_vdev)); 7104 7105 /* if we were unable to find any usable vdevs, return NULL */ 7106 if (vdev_is_dead(next->l2ad_vdev)) 7107 next = NULL; 7108 7109 l2arc_dev_last = next; 7110 7111out: 7112 mutex_exit(&l2arc_dev_mtx); 7113 7114 /* 7115 * Grab the config lock to prevent the 'next' device from being 7116 * removed while we are writing to it. 7117 */ 7118 if (next != NULL) 7119 spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER); 7120 mutex_exit(&spa_namespace_lock); 7121 7122 return (next); 7123} 7124 7125/* 7126 * Free buffers that were tagged for destruction. 7127 */ 7128static void 7129l2arc_do_free_on_write() 7130{ 7131 list_t *buflist; 7132 l2arc_data_free_t *df, *df_prev; 7133 7134 mutex_enter(&l2arc_free_on_write_mtx); 7135 buflist = l2arc_free_on_write; 7136 7137 for (df = list_tail(buflist); df; df = df_prev) { 7138 df_prev = list_prev(buflist, df); 7139 ASSERT3P(df->l2df_abd, !=, NULL); 7140 abd_free(df->l2df_abd); 7141 list_remove(buflist, df); 7142 kmem_free(df, sizeof (l2arc_data_free_t)); 7143 } 7144 7145 mutex_exit(&l2arc_free_on_write_mtx); 7146} 7147 7148/* 7149 * A write to a cache device has completed. Update all headers to allow 7150 * reads from these buffers to begin. 7151 */ 7152static void 7153l2arc_write_done(zio_t *zio) 7154{ 7155 l2arc_write_callback_t *cb; 7156 l2arc_dev_t *dev; 7157 list_t *buflist; 7158 arc_buf_hdr_t *head, *hdr, *hdr_prev; 7159 kmutex_t *hash_lock; 7160 int64_t bytes_dropped = 0; 7161 7162 cb = zio->io_private; 7163 ASSERT3P(cb, !=, NULL); 7164 dev = cb->l2wcb_dev; 7165 ASSERT3P(dev, !=, NULL); 7166 head = cb->l2wcb_head; 7167 ASSERT3P(head, !=, NULL); 7168 buflist = &dev->l2ad_buflist; 7169 ASSERT3P(buflist, !=, NULL); 7170 DTRACE_PROBE2(l2arc__iodone, zio_t *, zio, 7171 l2arc_write_callback_t *, cb); 7172 7173 if (zio->io_error != 0) 7174 ARCSTAT_BUMP(arcstat_l2_writes_error); 7175 7176 /* 7177 * All writes completed, or an error was hit. 7178 */ 7179top: 7180 mutex_enter(&dev->l2ad_mtx); 7181 for (hdr = list_prev(buflist, head); hdr; hdr = hdr_prev) { 7182 hdr_prev = list_prev(buflist, hdr); 7183 7184 hash_lock = HDR_LOCK(hdr); 7185 7186 /* 7187 * We cannot use mutex_enter or else we can deadlock 7188 * with l2arc_write_buffers (due to swapping the order 7189 * the hash lock and l2ad_mtx are taken). 7190 */ 7191 if (!mutex_tryenter(hash_lock)) { 7192 /* 7193 * Missed the hash lock. We must retry so we 7194 * don't leave the ARC_FLAG_L2_WRITING bit set. 7195 */ 7196 ARCSTAT_BUMP(arcstat_l2_writes_lock_retry); 7197 7198 /* 7199 * We don't want to rescan the headers we've 7200 * already marked as having been written out, so 7201 * we reinsert the head node so we can pick up 7202 * where we left off. 7203 */ 7204 list_remove(buflist, head); 7205 list_insert_after(buflist, hdr, head); 7206 7207 mutex_exit(&dev->l2ad_mtx); 7208 7209 /* 7210 * We wait for the hash lock to become available 7211 * to try and prevent busy waiting, and increase 7212 * the chance we'll be able to acquire the lock 7213 * the next time around. 7214 */ 7215 mutex_enter(hash_lock); 7216 mutex_exit(hash_lock); 7217 goto top; 7218 } 7219 7220 /* 7221 * We could not have been moved into the arc_l2c_only 7222 * state while in-flight due to our ARC_FLAG_L2_WRITING 7223 * bit being set. Let's just ensure that's being enforced. 7224 */ 7225 ASSERT(HDR_HAS_L1HDR(hdr)); 7226 7227 if (zio->io_error != 0) { 7228 /* 7229 * Error - drop L2ARC entry. 7230 */ 7231 list_remove(buflist, hdr); 7232 l2arc_trim(hdr); 7233 arc_hdr_clear_flags(hdr, ARC_FLAG_HAS_L2HDR); 7234 7235 ARCSTAT_INCR(arcstat_l2_psize, -arc_hdr_size(hdr)); 7236 ARCSTAT_INCR(arcstat_l2_lsize, -HDR_GET_LSIZE(hdr)); 7237 7238 bytes_dropped += arc_hdr_size(hdr); 7239 (void) refcount_remove_many(&dev->l2ad_alloc, 7240 arc_hdr_size(hdr), hdr); 7241 } 7242 7243 /* 7244 * Allow ARC to begin reads and ghost list evictions to 7245 * this L2ARC entry. 7246 */ 7247 arc_hdr_clear_flags(hdr, ARC_FLAG_L2_WRITING); 7248 7249 mutex_exit(hash_lock); 7250 } 7251 7252 atomic_inc_64(&l2arc_writes_done); 7253 list_remove(buflist, head); 7254 ASSERT(!HDR_HAS_L1HDR(head)); 7255 kmem_cache_free(hdr_l2only_cache, head); 7256 mutex_exit(&dev->l2ad_mtx); 7257 7258 vdev_space_update(dev->l2ad_vdev, -bytes_dropped, 0, 0); 7259 7260 l2arc_do_free_on_write(); 7261 7262 kmem_free(cb, sizeof (l2arc_write_callback_t)); 7263} 7264 7265/* 7266 * A read to a cache device completed. Validate buffer contents before 7267 * handing over to the regular ARC routines. 7268 */ 7269static void 7270l2arc_read_done(zio_t *zio) 7271{ 7272 l2arc_read_callback_t *cb; 7273 arc_buf_hdr_t *hdr; 7274 kmutex_t *hash_lock; 7275 boolean_t valid_cksum; 7276 7277 ASSERT3P(zio->io_vd, !=, NULL); 7278 ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE); 7279 7280 spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd); 7281 7282 cb = zio->io_private; 7283 ASSERT3P(cb, !=, NULL); 7284 hdr = cb->l2rcb_hdr; 7285 ASSERT3P(hdr, !=, NULL); 7286 7287 hash_lock = HDR_LOCK(hdr); 7288 mutex_enter(hash_lock); 7289 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 7290 7291 /* 7292 * If the data was read into a temporary buffer, 7293 * move it and free the buffer. 7294 */ 7295 if (cb->l2rcb_abd != NULL) { 7296 ASSERT3U(arc_hdr_size(hdr), <, zio->io_size); 7297 if (zio->io_error == 0) { 7298 abd_copy(hdr->b_l1hdr.b_pabd, cb->l2rcb_abd, 7299 arc_hdr_size(hdr)); 7300 } 7301 7302 /* 7303 * The following must be done regardless of whether 7304 * there was an error: 7305 * - free the temporary buffer 7306 * - point zio to the real ARC buffer 7307 * - set zio size accordingly 7308 * These are required because zio is either re-used for 7309 * an I/O of the block in the case of the error 7310 * or the zio is passed to arc_read_done() and it 7311 * needs real data. 7312 */ 7313 abd_free(cb->l2rcb_abd); 7314 zio->io_size = zio->io_orig_size = arc_hdr_size(hdr); 7315 zio->io_abd = zio->io_orig_abd = hdr->b_l1hdr.b_pabd; 7316 } 7317 7318 ASSERT3P(zio->io_abd, !=, NULL); 7319 7320 /* 7321 * Check this survived the L2ARC journey. 7322 */ 7323 ASSERT3P(zio->io_abd, ==, hdr->b_l1hdr.b_pabd); 7324 zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */ 7325 zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */ 7326 7327 valid_cksum = arc_cksum_is_equal(hdr, zio); 7328 if (valid_cksum && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) { 7329 mutex_exit(hash_lock); 7330 zio->io_private = hdr; 7331 arc_read_done(zio); 7332 } else { 7333 mutex_exit(hash_lock); 7334 /* 7335 * Buffer didn't survive caching. Increment stats and 7336 * reissue to the original storage device. 7337 */ 7338 if (zio->io_error != 0) { 7339 ARCSTAT_BUMP(arcstat_l2_io_error); 7340 } else { 7341 zio->io_error = SET_ERROR(EIO); 7342 } 7343 if (!valid_cksum) 7344 ARCSTAT_BUMP(arcstat_l2_cksum_bad); 7345 7346 /* 7347 * If there's no waiter, issue an async i/o to the primary 7348 * storage now. If there *is* a waiter, the caller must 7349 * issue the i/o in a context where it's OK to block. 7350 */ 7351 if (zio->io_waiter == NULL) { 7352 zio_t *pio = zio_unique_parent(zio); 7353 7354 ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL); 7355 7356 zio_nowait(zio_read(pio, zio->io_spa, zio->io_bp, 7357 hdr->b_l1hdr.b_pabd, zio->io_size, arc_read_done, 7358 hdr, zio->io_priority, cb->l2rcb_flags, 7359 &cb->l2rcb_zb)); 7360 } 7361 } 7362 7363 kmem_free(cb, sizeof (l2arc_read_callback_t)); 7364} 7365 7366/* 7367 * This is the list priority from which the L2ARC will search for pages to 7368 * cache. This is used within loops (0..3) to cycle through lists in the 7369 * desired order. This order can have a significant effect on cache 7370 * performance. 7371 * 7372 * Currently the metadata lists are hit first, MFU then MRU, followed by 7373 * the data lists. This function returns a locked list, and also returns 7374 * the lock pointer. 7375 */ 7376static multilist_sublist_t * 7377l2arc_sublist_lock(int list_num) 7378{ 7379 multilist_t *ml = NULL; 7380 unsigned int idx; 7381 7382 ASSERT(list_num >= 0 && list_num <= 3); 7383 7384 switch (list_num) { 7385 case 0: 7386 ml = arc_mfu->arcs_list[ARC_BUFC_METADATA]; 7387 break; 7388 case 1: 7389 ml = arc_mru->arcs_list[ARC_BUFC_METADATA]; 7390 break; 7391 case 2: 7392 ml = arc_mfu->arcs_list[ARC_BUFC_DATA]; 7393 break; 7394 case 3: 7395 ml = arc_mru->arcs_list[ARC_BUFC_DATA]; 7396 break; 7397 } 7398 7399 /* 7400 * Return a randomly-selected sublist. This is acceptable 7401 * because the caller feeds only a little bit of data for each 7402 * call (8MB). Subsequent calls will result in different 7403 * sublists being selected. 7404 */ 7405 idx = multilist_get_random_index(ml); 7406 return (multilist_sublist_lock(ml, idx)); 7407} 7408 7409/* 7410 * Evict buffers from the device write hand to the distance specified in 7411 * bytes. This distance may span populated buffers, it may span nothing. 7412 * This is clearing a region on the L2ARC device ready for writing. 7413 * If the 'all' boolean is set, every buffer is evicted. 7414 */ 7415static void 7416l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all) 7417{ 7418 list_t *buflist; 7419 arc_buf_hdr_t *hdr, *hdr_prev; 7420 kmutex_t *hash_lock; 7421 uint64_t taddr; 7422 7423 buflist = &dev->l2ad_buflist; 7424 7425 if (!all && dev->l2ad_first) { 7426 /* 7427 * This is the first sweep through the device. There is 7428 * nothing to evict. 7429 */ 7430 return; 7431 } 7432 7433 if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) { 7434 /* 7435 * When nearing the end of the device, evict to the end 7436 * before the device write hand jumps to the start. 7437 */ 7438 taddr = dev->l2ad_end; 7439 } else { 7440 taddr = dev->l2ad_hand + distance; 7441 } 7442 DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist, 7443 uint64_t, taddr, boolean_t, all); 7444 7445top: 7446 mutex_enter(&dev->l2ad_mtx); 7447 for (hdr = list_tail(buflist); hdr; hdr = hdr_prev) { 7448 hdr_prev = list_prev(buflist, hdr); 7449 7450 hash_lock = HDR_LOCK(hdr); 7451 7452 /* 7453 * We cannot use mutex_enter or else we can deadlock 7454 * with l2arc_write_buffers (due to swapping the order 7455 * the hash lock and l2ad_mtx are taken). 7456 */ 7457 if (!mutex_tryenter(hash_lock)) { 7458 /* 7459 * Missed the hash lock. Retry. 7460 */ 7461 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry); 7462 mutex_exit(&dev->l2ad_mtx); 7463 mutex_enter(hash_lock); 7464 mutex_exit(hash_lock); 7465 goto top; 7466 } 7467 7468 /* 7469 * A header can't be on this list if it doesn't have L2 header. 7470 */ 7471 ASSERT(HDR_HAS_L2HDR(hdr)); 7472 7473 /* Ensure this header has finished being written. */ 7474 ASSERT(!HDR_L2_WRITING(hdr)); 7475 ASSERT(!HDR_L2_WRITE_HEAD(hdr)); 7476 7477 if (!all && (hdr->b_l2hdr.b_daddr >= taddr || 7478 hdr->b_l2hdr.b_daddr < dev->l2ad_hand)) { 7479 /* 7480 * We've evicted to the target address, 7481 * or the end of the device. 7482 */ 7483 mutex_exit(hash_lock); 7484 break; 7485 } 7486 7487 if (!HDR_HAS_L1HDR(hdr)) { 7488 ASSERT(!HDR_L2_READING(hdr)); 7489 /* 7490 * This doesn't exist in the ARC. Destroy. 7491 * arc_hdr_destroy() will call list_remove() 7492 * and decrement arcstat_l2_lsize. 7493 */ 7494 arc_change_state(arc_anon, hdr, hash_lock); 7495 arc_hdr_destroy(hdr); 7496 } else { 7497 ASSERT(hdr->b_l1hdr.b_state != arc_l2c_only); 7498 ARCSTAT_BUMP(arcstat_l2_evict_l1cached); 7499 /* 7500 * Invalidate issued or about to be issued 7501 * reads, since we may be about to write 7502 * over this location. 7503 */ 7504 if (HDR_L2_READING(hdr)) { 7505 ARCSTAT_BUMP(arcstat_l2_evict_reading); 7506 arc_hdr_set_flags(hdr, ARC_FLAG_L2_EVICTED); 7507 } 7508 7509 arc_hdr_l2hdr_destroy(hdr); 7510 } 7511 mutex_exit(hash_lock); 7512 } 7513 mutex_exit(&dev->l2ad_mtx); 7514} 7515 7516/* 7517 * Find and write ARC buffers to the L2ARC device. 7518 * 7519 * An ARC_FLAG_L2_WRITING flag is set so that the L2ARC buffers are not valid 7520 * for reading until they have completed writing. 7521 * The headroom_boost is an in-out parameter used to maintain headroom boost 7522 * state between calls to this function. 7523 * 7524 * Returns the number of bytes actually written (which may be smaller than 7525 * the delta by which the device hand has changed due to alignment). 7526 */ 7527static uint64_t 7528l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz) 7529{ 7530 arc_buf_hdr_t *hdr, *hdr_prev, *head; 7531 uint64_t write_asize, write_psize, write_lsize, headroom; 7532 boolean_t full; 7533 l2arc_write_callback_t *cb; 7534 zio_t *pio, *wzio; 7535 uint64_t guid = spa_load_guid(spa); 7536 int try; 7537 7538 ASSERT3P(dev->l2ad_vdev, !=, NULL); 7539 7540 pio = NULL; 7541 write_lsize = write_asize = write_psize = 0; 7542 full = B_FALSE; 7543 head = kmem_cache_alloc(hdr_l2only_cache, KM_PUSHPAGE); 7544 arc_hdr_set_flags(head, ARC_FLAG_L2_WRITE_HEAD | ARC_FLAG_HAS_L2HDR); 7545 7546 ARCSTAT_BUMP(arcstat_l2_write_buffer_iter); 7547 /* 7548 * Copy buffers for L2ARC writing. 7549 */ 7550 for (try = 0; try <= 3; try++) { 7551 multilist_sublist_t *mls = l2arc_sublist_lock(try); 7552 uint64_t passed_sz = 0; 7553 7554 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_iter); 7555 7556 /* 7557 * L2ARC fast warmup. 7558 * 7559 * Until the ARC is warm and starts to evict, read from the 7560 * head of the ARC lists rather than the tail. 7561 */ 7562 if (arc_warm == B_FALSE) 7563 hdr = multilist_sublist_head(mls); 7564 else 7565 hdr = multilist_sublist_tail(mls); 7566 if (hdr == NULL) 7567 ARCSTAT_BUMP(arcstat_l2_write_buffer_list_null_iter); 7568 7569 headroom = target_sz * l2arc_headroom; 7570 if (zfs_compressed_arc_enabled) 7571 headroom = (headroom * l2arc_headroom_boost) / 100; 7572 7573 for (; hdr; hdr = hdr_prev) { 7574 kmutex_t *hash_lock; 7575 7576 if (arc_warm == B_FALSE) 7577 hdr_prev = multilist_sublist_next(mls, hdr); 7578 else 7579 hdr_prev = multilist_sublist_prev(mls, hdr); 7580 ARCSTAT_INCR(arcstat_l2_write_buffer_bytes_scanned, 7581 HDR_GET_LSIZE(hdr)); 7582 7583 hash_lock = HDR_LOCK(hdr); 7584 if (!mutex_tryenter(hash_lock)) { 7585 ARCSTAT_BUMP(arcstat_l2_write_trylock_fail); 7586 /* 7587 * Skip this buffer rather than waiting. 7588 */ 7589 continue; 7590 } 7591 7592 passed_sz += HDR_GET_LSIZE(hdr); 7593 if (passed_sz > headroom) { 7594 /* 7595 * Searched too far. 7596 */ 7597 mutex_exit(hash_lock); 7598 ARCSTAT_BUMP(arcstat_l2_write_passed_headroom); 7599 break; 7600 } 7601 7602 if (!l2arc_write_eligible(guid, hdr)) { 7603 mutex_exit(hash_lock); 7604 continue; 7605 } 7606 7607 /* 7608 * We rely on the L1 portion of the header below, so 7609 * it's invalid for this header to have been evicted out 7610 * of the ghost cache, prior to being written out. The 7611 * ARC_FLAG_L2_WRITING bit ensures this won't happen. 7612 */ 7613 ASSERT(HDR_HAS_L1HDR(hdr)); 7614 7615 ASSERT3U(HDR_GET_PSIZE(hdr), >, 0); 7616 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL); 7617 ASSERT3U(arc_hdr_size(hdr), >, 0); 7618 uint64_t psize = arc_hdr_size(hdr); 7619 uint64_t asize = vdev_psize_to_asize(dev->l2ad_vdev, 7620 psize); 7621 7622 if ((write_asize + asize) > target_sz) { 7623 full = B_TRUE; 7624 mutex_exit(hash_lock); 7625 ARCSTAT_BUMP(arcstat_l2_write_full); 7626 break; 7627 } 7628 7629 if (pio == NULL) { 7630 /* 7631 * Insert a dummy header on the buflist so 7632 * l2arc_write_done() can find where the 7633 * write buffers begin without searching. 7634 */ 7635 mutex_enter(&dev->l2ad_mtx); 7636 list_insert_head(&dev->l2ad_buflist, head); 7637 mutex_exit(&dev->l2ad_mtx); 7638 7639 cb = kmem_alloc( 7640 sizeof (l2arc_write_callback_t), KM_SLEEP); 7641 cb->l2wcb_dev = dev; 7642 cb->l2wcb_head = head; 7643 pio = zio_root(spa, l2arc_write_done, cb, 7644 ZIO_FLAG_CANFAIL); 7645 ARCSTAT_BUMP(arcstat_l2_write_pios); 7646 } 7647 7648 hdr->b_l2hdr.b_dev = dev; 7649 hdr->b_l2hdr.b_daddr = dev->l2ad_hand; 7650 arc_hdr_set_flags(hdr, 7651 ARC_FLAG_L2_WRITING | ARC_FLAG_HAS_L2HDR); 7652 7653 mutex_enter(&dev->l2ad_mtx); 7654 list_insert_head(&dev->l2ad_buflist, hdr); 7655 mutex_exit(&dev->l2ad_mtx); 7656 7657 (void) refcount_add_many(&dev->l2ad_alloc, psize, hdr); 7658 7659 /* 7660 * Normally the L2ARC can use the hdr's data, but if 7661 * we're sharing data between the hdr and one of its 7662 * bufs, L2ARC needs its own copy of the data so that 7663 * the ZIO below can't race with the buf consumer. 7664 * Another case where we need to create a copy of the 7665 * data is when the buffer size is not device-aligned 7666 * and we need to pad the block to make it such. 7667 * That also keeps the clock hand suitably aligned. 7668 * 7669 * To ensure that the copy will be available for the 7670 * lifetime of the ZIO and be cleaned up afterwards, we 7671 * add it to the l2arc_free_on_write queue. 7672 */ 7673 abd_t *to_write; 7674 if (!HDR_SHARED_DATA(hdr) && psize == asize) { 7675 to_write = hdr->b_l1hdr.b_pabd; 7676 } else { 7677 to_write = abd_alloc_for_io(asize, 7678 HDR_ISTYPE_METADATA(hdr)); 7679 abd_copy(to_write, hdr->b_l1hdr.b_pabd, psize); 7680 if (asize != psize) { 7681 abd_zero_off(to_write, psize, 7682 asize - psize); 7683 } 7684 l2arc_free_abd_on_write(to_write, asize, 7685 arc_buf_type(hdr)); 7686 } 7687 wzio = zio_write_phys(pio, dev->l2ad_vdev, 7688 hdr->b_l2hdr.b_daddr, asize, to_write, 7689 ZIO_CHECKSUM_OFF, NULL, hdr, 7690 ZIO_PRIORITY_ASYNC_WRITE, 7691 ZIO_FLAG_CANFAIL, B_FALSE); 7692 7693 write_lsize += HDR_GET_LSIZE(hdr); 7694 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev, 7695 zio_t *, wzio); 7696 7697 write_psize += psize; 7698 write_asize += asize; 7699 dev->l2ad_hand += asize; 7700 7701 mutex_exit(hash_lock); 7702 7703 (void) zio_nowait(wzio); 7704 } 7705 7706 multilist_sublist_unlock(mls); 7707 7708 if (full == B_TRUE) 7709 break; 7710 } 7711 7712 /* No buffers selected for writing? */ 7713 if (pio == NULL) { 7714 ASSERT0(write_lsize); 7715 ASSERT(!HDR_HAS_L1HDR(head)); 7716 kmem_cache_free(hdr_l2only_cache, head); 7717 return (0); 7718 } 7719 7720 ASSERT3U(write_psize, <=, target_sz); 7721 ARCSTAT_BUMP(arcstat_l2_writes_sent); 7722 ARCSTAT_INCR(arcstat_l2_write_bytes, write_psize); 7723 ARCSTAT_INCR(arcstat_l2_lsize, write_lsize); 7724 ARCSTAT_INCR(arcstat_l2_psize, write_psize); 7725 vdev_space_update(dev->l2ad_vdev, write_psize, 0, 0); 7726 7727 /* 7728 * Bump device hand to the device start if it is approaching the end. 7729 * l2arc_evict() will already have evicted ahead for this case. 7730 */ 7731 if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) { 7732 dev->l2ad_hand = dev->l2ad_start; 7733 dev->l2ad_first = B_FALSE; 7734 } 7735 7736 dev->l2ad_writing = B_TRUE; 7737 (void) zio_wait(pio); 7738 dev->l2ad_writing = B_FALSE; 7739 7740 return (write_asize); 7741} 7742 7743/* 7744 * This thread feeds the L2ARC at regular intervals. This is the beating 7745 * heart of the L2ARC. 7746 */ 7747/* ARGSUSED */ 7748static void 7749l2arc_feed_thread(void *unused __unused) 7750{ 7751 callb_cpr_t cpr; 7752 l2arc_dev_t *dev; 7753 spa_t *spa; 7754 uint64_t size, wrote; 7755 clock_t begin, next = ddi_get_lbolt(); 7756 7757 CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG); 7758 7759 mutex_enter(&l2arc_feed_thr_lock); 7760 7761 while (l2arc_thread_exit == 0) { 7762 CALLB_CPR_SAFE_BEGIN(&cpr); 7763 (void) cv_timedwait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock, 7764 next - ddi_get_lbolt()); 7765 CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock); 7766 next = ddi_get_lbolt() + hz; 7767 7768 /* 7769 * Quick check for L2ARC devices. 7770 */ 7771 mutex_enter(&l2arc_dev_mtx); 7772 if (l2arc_ndev == 0) { 7773 mutex_exit(&l2arc_dev_mtx); 7774 continue; 7775 } 7776 mutex_exit(&l2arc_dev_mtx); 7777 begin = ddi_get_lbolt(); 7778 7779 /* 7780 * This selects the next l2arc device to write to, and in 7781 * doing so the next spa to feed from: dev->l2ad_spa. This 7782 * will return NULL if there are now no l2arc devices or if 7783 * they are all faulted. 7784 * 7785 * If a device is returned, its spa's config lock is also 7786 * held to prevent device removal. l2arc_dev_get_next() 7787 * will grab and release l2arc_dev_mtx. 7788 */ 7789 if ((dev = l2arc_dev_get_next()) == NULL) 7790 continue; 7791 7792 spa = dev->l2ad_spa; 7793 ASSERT3P(spa, !=, NULL); 7794 7795 /* 7796 * If the pool is read-only then force the feed thread to 7797 * sleep a little longer. 7798 */ 7799 if (!spa_writeable(spa)) { 7800 next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz; 7801 spa_config_exit(spa, SCL_L2ARC, dev); 7802 continue; 7803 } 7804 7805 /* 7806 * Avoid contributing to memory pressure. 7807 */ 7808 if (arc_reclaim_needed()) { 7809 ARCSTAT_BUMP(arcstat_l2_abort_lowmem); 7810 spa_config_exit(spa, SCL_L2ARC, dev); 7811 continue; 7812 } 7813 7814 ARCSTAT_BUMP(arcstat_l2_feeds); 7815 7816 size = l2arc_write_size(); 7817 7818 /* 7819 * Evict L2ARC buffers that will be overwritten. 7820 */ 7821 l2arc_evict(dev, size, B_FALSE); 7822 7823 /* 7824 * Write ARC buffers. 7825 */ 7826 wrote = l2arc_write_buffers(spa, dev, size); 7827 7828 /* 7829 * Calculate interval between writes. 7830 */ 7831 next = l2arc_write_interval(begin, size, wrote); 7832 spa_config_exit(spa, SCL_L2ARC, dev); 7833 } 7834 7835 l2arc_thread_exit = 0; 7836 cv_broadcast(&l2arc_feed_thr_cv); 7837 CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */ 7838 thread_exit(); 7839} 7840 7841boolean_t 7842l2arc_vdev_present(vdev_t *vd) 7843{ 7844 l2arc_dev_t *dev; 7845 7846 mutex_enter(&l2arc_dev_mtx); 7847 for (dev = list_head(l2arc_dev_list); dev != NULL; 7848 dev = list_next(l2arc_dev_list, dev)) { 7849 if (dev->l2ad_vdev == vd) 7850 break; 7851 } 7852 mutex_exit(&l2arc_dev_mtx); 7853 7854 return (dev != NULL); 7855} 7856 7857/* 7858 * Add a vdev for use by the L2ARC. By this point the spa has already 7859 * validated the vdev and opened it. 7860 */ 7861void 7862l2arc_add_vdev(spa_t *spa, vdev_t *vd) 7863{ 7864 l2arc_dev_t *adddev; 7865 7866 ASSERT(!l2arc_vdev_present(vd)); 7867 7868 vdev_ashift_optimize(vd); 7869 7870 /* 7871 * Create a new l2arc device entry. 7872 */ 7873 adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP); 7874 adddev->l2ad_spa = spa; 7875 adddev->l2ad_vdev = vd; 7876 adddev->l2ad_start = VDEV_LABEL_START_SIZE; 7877 adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd); 7878 adddev->l2ad_hand = adddev->l2ad_start; 7879 adddev->l2ad_first = B_TRUE; 7880 adddev->l2ad_writing = B_FALSE; 7881 7882 mutex_init(&adddev->l2ad_mtx, NULL, MUTEX_DEFAULT, NULL); 7883 /* 7884 * This is a list of all ARC buffers that are still valid on the 7885 * device. 7886 */ 7887 list_create(&adddev->l2ad_buflist, sizeof (arc_buf_hdr_t), 7888 offsetof(arc_buf_hdr_t, b_l2hdr.b_l2node)); 7889 7890 vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand); 7891 refcount_create(&adddev->l2ad_alloc); 7892 7893 /* 7894 * Add device to global list 7895 */ 7896 mutex_enter(&l2arc_dev_mtx); 7897 list_insert_head(l2arc_dev_list, adddev); 7898 atomic_inc_64(&l2arc_ndev); 7899 mutex_exit(&l2arc_dev_mtx); 7900} 7901 7902/* 7903 * Remove a vdev from the L2ARC. 7904 */ 7905void 7906l2arc_remove_vdev(vdev_t *vd) 7907{ 7908 l2arc_dev_t *dev, *nextdev, *remdev = NULL; 7909 7910 /* 7911 * Find the device by vdev 7912 */ 7913 mutex_enter(&l2arc_dev_mtx); 7914 for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) { 7915 nextdev = list_next(l2arc_dev_list, dev); 7916 if (vd == dev->l2ad_vdev) { 7917 remdev = dev; 7918 break; 7919 } 7920 } 7921 ASSERT3P(remdev, !=, NULL); 7922 7923 /* 7924 * Remove device from global list 7925 */ 7926 list_remove(l2arc_dev_list, remdev); 7927 l2arc_dev_last = NULL; /* may have been invalidated */ 7928 atomic_dec_64(&l2arc_ndev); 7929 mutex_exit(&l2arc_dev_mtx); 7930 7931 /* 7932 * Clear all buflists and ARC references. L2ARC device flush. 7933 */ 7934 l2arc_evict(remdev, 0, B_TRUE); 7935 list_destroy(&remdev->l2ad_buflist); 7936 mutex_destroy(&remdev->l2ad_mtx); 7937 refcount_destroy(&remdev->l2ad_alloc); 7938 kmem_free(remdev, sizeof (l2arc_dev_t)); 7939} 7940 7941void 7942l2arc_init(void) 7943{ 7944 l2arc_thread_exit = 0; 7945 l2arc_ndev = 0; 7946 l2arc_writes_sent = 0; 7947 l2arc_writes_done = 0; 7948 7949 mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL); 7950 cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL); 7951 mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL); 7952 mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL); 7953 7954 l2arc_dev_list = &L2ARC_dev_list; 7955 l2arc_free_on_write = &L2ARC_free_on_write; 7956 list_create(l2arc_dev_list, sizeof (l2arc_dev_t), 7957 offsetof(l2arc_dev_t, l2ad_node)); 7958 list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t), 7959 offsetof(l2arc_data_free_t, l2df_list_node)); 7960} 7961 7962void 7963l2arc_fini(void) 7964{ 7965 /* 7966 * This is called from dmu_fini(), which is called from spa_fini(); 7967 * Because of this, we can assume that all l2arc devices have 7968 * already been removed when the pools themselves were removed. 7969 */ 7970 7971 l2arc_do_free_on_write(); 7972 7973 mutex_destroy(&l2arc_feed_thr_lock); 7974 cv_destroy(&l2arc_feed_thr_cv); 7975 mutex_destroy(&l2arc_dev_mtx); 7976 mutex_destroy(&l2arc_free_on_write_mtx); 7977 7978 list_destroy(l2arc_dev_list); 7979 list_destroy(l2arc_free_on_write); 7980} 7981 7982void 7983l2arc_start(void) 7984{ 7985 if (!(spa_mode_global & FWRITE)) 7986 return; 7987 7988 (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0, 7989 TS_RUN, minclsyspri); 7990} 7991 7992void 7993l2arc_stop(void) 7994{ 7995 if (!(spa_mode_global & FWRITE)) 7996 return; 7997 7998 mutex_enter(&l2arc_feed_thr_lock); 7999 cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */ 8000 l2arc_thread_exit = 1; 8001 while (l2arc_thread_exit != 0) 8002 cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock); 8003 mutex_exit(&l2arc_feed_thr_lock); 8004} 8005