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, 2018 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/vmsystm.h> 271 #include <vm/anon.h> 272 #include <sys/fs/swapnode.h> 273 #include <sys/dnlc.h> 274 #endif 275 #include <sys/callb.h> 276 #include <sys/kstat.h> 277 #include <sys/zthr.h> 278 #include <zfs_fletcher.h> 279 #include <sys/aggsum.h> 280 #include <sys/cityhash.h> 281 282 #ifndef _KERNEL 283 /* set with ZFS_DEBUG=watch, to enable watchpoints on frozen buffers */ 284 boolean_t arc_watch = B_FALSE; 285 int arc_procfd; 286 #endif 287 288 /* 289 * This thread's job is to keep enough free memory in the system, by 290 * calling arc_kmem_reap_now() plus arc_shrink(), which improves 291 * arc_available_memory(). 292 */ 293 static zthr_t *arc_reap_zthr; 294 295 /* 296 * This thread's job is to keep arc_size under arc_c, by calling 297 * arc_adjust(), which improves arc_is_overflowing(). 298 */ 299 static zthr_t *arc_adjust_zthr; 300 301 static kmutex_t arc_adjust_lock; 302 static kcondvar_t arc_adjust_waiters_cv; 303 static boolean_t arc_adjust_needed = B_FALSE; 304 305 uint_t arc_reduce_dnlc_percent = 3; 306 307 /* 308 * The number of headers to evict in arc_evict_state_impl() before 309 * dropping the sublist lock and evicting from another sublist. A lower 310 * value means we're more likely to evict the "correct" header (i.e. the 311 * oldest header in the arc state), but comes with higher overhead 312 * (i.e. more invocations of arc_evict_state_impl()). 313 */ 314 int zfs_arc_evict_batch_limit = 10; 315 316 /* number of seconds before growing cache again */ 317 int arc_grow_retry = 60; 318 319 /* 320 * Minimum time between calls to arc_kmem_reap_soon(). Note that this will 321 * be converted to ticks, so with the default hz=100, a setting of 15 ms 322 * will actually wait 2 ticks, or 20ms. 323 */ 324 int arc_kmem_cache_reap_retry_ms = 1000; 325 326 /* shift of arc_c for calculating overflow limit in arc_get_data_impl */ 327 int zfs_arc_overflow_shift = 8; 328 329 /* shift of arc_c for calculating both min and max arc_p */ 330 int arc_p_min_shift = 4; 331 332 /* log2(fraction of arc to reclaim) */ 333 int arc_shrink_shift = 7; 334 335 /* 336 * log2(fraction of ARC which must be free to allow growing). 337 * I.e. If there is less than arc_c >> arc_no_grow_shift free memory, 338 * when reading a new block into the ARC, we will evict an equal-sized block 339 * from the ARC. 340 * 341 * This must be less than arc_shrink_shift, so that when we shrink the ARC, 342 * we will still not allow it to grow. 343 */ 344 int arc_no_grow_shift = 5; 345 346 347 /* 348 * minimum lifespan of a prefetch block in clock ticks 349 * (initialized in arc_init()) 350 */ 351 static int arc_min_prefetch_lifespan; 352 353 /* 354 * If this percent of memory is free, don't throttle. 355 */ 356 int arc_lotsfree_percent = 10; 357 358 static boolean_t arc_initialized; 359 360 /* 361 * The arc has filled available memory and has now warmed up. 362 */ 363 static boolean_t arc_warm; 364 365 /* 366 * log2 fraction of the zio arena to keep free. 367 */ 368 int arc_zio_arena_free_shift = 2; 369 370 /* 371 * These tunables are for performance analysis. 372 */ 373 uint64_t zfs_arc_max; 374 uint64_t zfs_arc_min; 375 uint64_t zfs_arc_meta_limit = 0; 376 uint64_t zfs_arc_meta_min = 0; 377 int zfs_arc_grow_retry = 0; 378 int zfs_arc_shrink_shift = 0; 379 int zfs_arc_p_min_shift = 0; 380 int zfs_arc_average_blocksize = 8 * 1024; /* 8KB */ 381 382 /* 383 * ARC dirty data constraints for arc_tempreserve_space() throttle 384 */ 385 uint_t zfs_arc_dirty_limit_percent = 50; /* total dirty data limit */ 386 uint_t zfs_arc_anon_limit_percent = 25; /* anon block dirty limit */ 387 uint_t zfs_arc_pool_dirty_percent = 20; /* each pool's anon allowance */ 388 389 boolean_t zfs_compressed_arc_enabled = B_TRUE; 390 391 /* 392 * Note that buffers can be in one of 6 states: 393 * ARC_anon - anonymous (discussed below) 394 * ARC_mru - recently used, currently cached 395 * ARC_mru_ghost - recentely used, no longer in cache 396 * ARC_mfu - frequently used, currently cached 397 * ARC_mfu_ghost - frequently used, no longer in cache 398 * ARC_l2c_only - exists in L2ARC but not other states 399 * When there are no active references to the buffer, they are 400 * are linked onto a list in one of these arc states. These are 401 * the only buffers that can be evicted or deleted. Within each 402 * state there are multiple lists, one for meta-data and one for 403 * non-meta-data. Meta-data (indirect blocks, blocks of dnodes, 404 * etc.) is tracked separately so that it can be managed more 405 * explicitly: favored over data, limited explicitly. 406 * 407 * Anonymous buffers are buffers that are not associated with 408 * a DVA. These are buffers that hold dirty block copies 409 * before they are written to stable storage. By definition, 410 * they are "ref'd" and are considered part of arc_mru 411 * that cannot be freed. Generally, they will aquire a DVA 412 * as they are written and migrate onto the arc_mru list. 413 * 414 * The ARC_l2c_only state is for buffers that are in the second 415 * level ARC but no longer in any of the ARC_m* lists. The second 416 * level ARC itself may also contain buffers that are in any of 417 * the ARC_m* states - meaning that a buffer can exist in two 418 * places. The reason for the ARC_l2c_only state is to keep the 419 * buffer header in the hash table, so that reads that hit the 420 * second level ARC benefit from these fast lookups. 421 */ 422 423 typedef struct arc_state { 424 /* 425 * list of evictable buffers 426 */ 427 multilist_t *arcs_list[ARC_BUFC_NUMTYPES]; 428 /* 429 * total amount of evictable data in this state 430 */ 431 zfs_refcount_t arcs_esize[ARC_BUFC_NUMTYPES]; 432 /* 433 * total amount of data in this state; this includes: evictable, 434 * non-evictable, ARC_BUFC_DATA, and ARC_BUFC_METADATA. 435 */ 436 zfs_refcount_t arcs_size; 437 } arc_state_t; 438 439 /* The 6 states: */ 440 static arc_state_t ARC_anon; 441 static arc_state_t ARC_mru; 442 static arc_state_t ARC_mru_ghost; 443 static arc_state_t ARC_mfu; 444 static arc_state_t ARC_mfu_ghost; 445 static arc_state_t ARC_l2c_only; 446 447 typedef struct arc_stats { 448 kstat_named_t arcstat_hits; 449 kstat_named_t arcstat_misses; 450 kstat_named_t arcstat_demand_data_hits; 451 kstat_named_t arcstat_demand_data_misses; 452 kstat_named_t arcstat_demand_metadata_hits; 453 kstat_named_t arcstat_demand_metadata_misses; 454 kstat_named_t arcstat_prefetch_data_hits; 455 kstat_named_t arcstat_prefetch_data_misses; 456 kstat_named_t arcstat_prefetch_metadata_hits; 457 kstat_named_t arcstat_prefetch_metadata_misses; 458 kstat_named_t arcstat_mru_hits; 459 kstat_named_t arcstat_mru_ghost_hits; 460 kstat_named_t arcstat_mfu_hits; 461 kstat_named_t arcstat_mfu_ghost_hits; 462 kstat_named_t arcstat_deleted; 463 /* 464 * Number of buffers that could not be evicted because the hash lock 465 * was held by another thread. The lock may not necessarily be held 466 * by something using the same buffer, since hash locks are shared 467 * by multiple buffers. 468 */ 469 kstat_named_t arcstat_mutex_miss; 470 /* 471 * Number of buffers skipped because they have I/O in progress, are 472 * indrect prefetch buffers that have not lived long enough, or are 473 * not from the spa we're trying to evict from. 474 */ 475 kstat_named_t arcstat_evict_skip; 476 /* 477 * Number of times arc_evict_state() was unable to evict enough 478 * buffers to reach it's target amount. 479 */ 480 kstat_named_t arcstat_evict_not_enough; 481 kstat_named_t arcstat_evict_l2_cached; 482 kstat_named_t arcstat_evict_l2_eligible; 483 kstat_named_t arcstat_evict_l2_ineligible; 484 kstat_named_t arcstat_evict_l2_skip; 485 kstat_named_t arcstat_hash_elements; 486 kstat_named_t arcstat_hash_elements_max; 487 kstat_named_t arcstat_hash_collisions; 488 kstat_named_t arcstat_hash_chains; 489 kstat_named_t arcstat_hash_chain_max; 490 kstat_named_t arcstat_p; 491 kstat_named_t arcstat_c; 492 kstat_named_t arcstat_c_min; 493 kstat_named_t arcstat_c_max; 494 /* Not updated directly; only synced in arc_kstat_update. */ 495 kstat_named_t arcstat_size; 496 /* 497 * Number of compressed bytes stored in the arc_buf_hdr_t's b_pabd. 498 * Note that the compressed bytes may match the uncompressed bytes 499 * if the block is either not compressed or compressed arc is disabled. 500 */ 501 kstat_named_t arcstat_compressed_size; 502 /* 503 * Uncompressed size of the data stored in b_pabd. If compressed 504 * arc is disabled then this value will be identical to the stat 505 * above. 506 */ 507 kstat_named_t arcstat_uncompressed_size; 508 /* 509 * Number of bytes stored in all the arc_buf_t's. This is classified 510 * as "overhead" since this data is typically short-lived and will 511 * be evicted from the arc when it becomes unreferenced unless the 512 * zfs_keep_uncompressed_metadata or zfs_keep_uncompressed_level 513 * values have been set (see comment in dbuf.c for more information). 514 */ 515 kstat_named_t arcstat_overhead_size; 516 /* 517 * Number of bytes consumed by internal ARC structures necessary 518 * for tracking purposes; these structures are not actually 519 * backed by ARC buffers. This includes arc_buf_hdr_t structures 520 * (allocated via arc_buf_hdr_t_full and arc_buf_hdr_t_l2only 521 * caches), and arc_buf_t structures (allocated via arc_buf_t 522 * cache). 523 * Not updated directly; only synced in arc_kstat_update. 524 */ 525 kstat_named_t arcstat_hdr_size; 526 /* 527 * Number of bytes consumed by ARC buffers of type equal to 528 * ARC_BUFC_DATA. This is generally consumed by buffers backing 529 * on disk user data (e.g. plain file contents). 530 * Not updated directly; only synced in arc_kstat_update. 531 */ 532 kstat_named_t arcstat_data_size; 533 /* 534 * Number of bytes consumed by ARC buffers of type equal to 535 * ARC_BUFC_METADATA. This is generally consumed by buffers 536 * backing on disk data that is used for internal ZFS 537 * structures (e.g. ZAP, dnode, indirect blocks, etc). 538 * Not updated directly; only synced in arc_kstat_update. 539 */ 540 kstat_named_t arcstat_metadata_size; 541 /* 542 * Number of bytes consumed by various buffers and structures 543 * not actually backed with ARC buffers. This includes bonus 544 * buffers (allocated directly via zio_buf_* functions), 545 * dmu_buf_impl_t structures (allocated via dmu_buf_impl_t 546 * cache), and dnode_t structures (allocated via dnode_t cache). 547 * Not updated directly; only synced in arc_kstat_update. 548 */ 549 kstat_named_t arcstat_other_size; 550 /* 551 * Total number of bytes consumed by ARC buffers residing in the 552 * arc_anon state. This includes *all* buffers in the arc_anon 553 * state; e.g. data, metadata, evictable, and unevictable buffers 554 * are all included in this value. 555 * Not updated directly; only synced in arc_kstat_update. 556 */ 557 kstat_named_t arcstat_anon_size; 558 /* 559 * Number of bytes consumed by ARC buffers that meet the 560 * following criteria: backing buffers of type ARC_BUFC_DATA, 561 * residing in the arc_anon state, and are eligible for eviction 562 * (e.g. have no outstanding holds on the buffer). 563 * Not updated directly; only synced in arc_kstat_update. 564 */ 565 kstat_named_t arcstat_anon_evictable_data; 566 /* 567 * Number of bytes consumed by ARC buffers that meet the 568 * following criteria: backing buffers of type ARC_BUFC_METADATA, 569 * residing in the arc_anon state, and are eligible for eviction 570 * (e.g. have no outstanding holds on the buffer). 571 * Not updated directly; only synced in arc_kstat_update. 572 */ 573 kstat_named_t arcstat_anon_evictable_metadata; 574 /* 575 * Total number of bytes consumed by ARC buffers residing in the 576 * arc_mru state. This includes *all* buffers in the arc_mru 577 * state; e.g. data, metadata, evictable, and unevictable buffers 578 * are all included in this value. 579 * Not updated directly; only synced in arc_kstat_update. 580 */ 581 kstat_named_t arcstat_mru_size; 582 /* 583 * Number of bytes consumed by ARC buffers that meet the 584 * following criteria: backing buffers of type ARC_BUFC_DATA, 585 * residing in the arc_mru state, and are eligible for eviction 586 * (e.g. have no outstanding holds on the buffer). 587 * Not updated directly; only synced in arc_kstat_update. 588 */ 589 kstat_named_t arcstat_mru_evictable_data; 590 /* 591 * Number of bytes consumed by ARC buffers that meet the 592 * following criteria: backing buffers of type ARC_BUFC_METADATA, 593 * residing in the arc_mru state, and are eligible for eviction 594 * (e.g. have no outstanding holds on the buffer). 595 * Not updated directly; only synced in arc_kstat_update. 596 */ 597 kstat_named_t arcstat_mru_evictable_metadata; 598 /* 599 * Total number of bytes that *would have been* consumed by ARC 600 * buffers in the arc_mru_ghost state. The key thing to note 601 * here, is the fact that this size doesn't actually indicate 602 * RAM consumption. The ghost lists only consist of headers and 603 * don't actually have ARC buffers linked off of these headers. 604 * Thus, *if* the headers had associated ARC buffers, these 605 * buffers *would have* consumed this number of bytes. 606 * Not updated directly; only synced in arc_kstat_update. 607 */ 608 kstat_named_t arcstat_mru_ghost_size; 609 /* 610 * Number of bytes that *would have been* consumed by ARC 611 * buffers that are eligible for eviction, of type 612 * ARC_BUFC_DATA, and linked off the arc_mru_ghost state. 613 * Not updated directly; only synced in arc_kstat_update. 614 */ 615 kstat_named_t arcstat_mru_ghost_evictable_data; 616 /* 617 * Number of bytes that *would have been* consumed by ARC 618 * buffers that are eligible for eviction, of type 619 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state. 620 * Not updated directly; only synced in arc_kstat_update. 621 */ 622 kstat_named_t arcstat_mru_ghost_evictable_metadata; 623 /* 624 * Total number of bytes consumed by ARC buffers residing in the 625 * arc_mfu state. This includes *all* buffers in the arc_mfu 626 * state; e.g. data, metadata, evictable, and unevictable buffers 627 * are all included in this value. 628 * Not updated directly; only synced in arc_kstat_update. 629 */ 630 kstat_named_t arcstat_mfu_size; 631 /* 632 * Number of bytes consumed by ARC buffers that are eligible for 633 * eviction, of type ARC_BUFC_DATA, and reside in the arc_mfu 634 * state. 635 * Not updated directly; only synced in arc_kstat_update. 636 */ 637 kstat_named_t arcstat_mfu_evictable_data; 638 /* 639 * Number of bytes consumed by ARC buffers that are eligible for 640 * eviction, of type ARC_BUFC_METADATA, and reside in the 641 * arc_mfu state. 642 * Not updated directly; only synced in arc_kstat_update. 643 */ 644 kstat_named_t arcstat_mfu_evictable_metadata; 645 /* 646 * Total number of bytes that *would have been* consumed by ARC 647 * buffers in the arc_mfu_ghost state. See the comment above 648 * arcstat_mru_ghost_size for more details. 649 * Not updated directly; only synced in arc_kstat_update. 650 */ 651 kstat_named_t arcstat_mfu_ghost_size; 652 /* 653 * Number of bytes that *would have been* consumed by ARC 654 * buffers that are eligible for eviction, of type 655 * ARC_BUFC_DATA, and linked off the arc_mfu_ghost state. 656 * Not updated directly; only synced in arc_kstat_update. 657 */ 658 kstat_named_t arcstat_mfu_ghost_evictable_data; 659 /* 660 * Number of bytes that *would have been* consumed by ARC 661 * buffers that are eligible for eviction, of type 662 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state. 663 * Not updated directly; only synced in arc_kstat_update. 664 */ 665 kstat_named_t arcstat_mfu_ghost_evictable_metadata; 666 kstat_named_t arcstat_l2_hits; 667 kstat_named_t arcstat_l2_misses; 668 kstat_named_t arcstat_l2_feeds; 669 kstat_named_t arcstat_l2_rw_clash; 670 kstat_named_t arcstat_l2_read_bytes; 671 kstat_named_t arcstat_l2_write_bytes; 672 kstat_named_t arcstat_l2_writes_sent; 673 kstat_named_t arcstat_l2_writes_done; 674 kstat_named_t arcstat_l2_writes_error; 675 kstat_named_t arcstat_l2_writes_lock_retry; 676 kstat_named_t arcstat_l2_evict_lock_retry; 677 kstat_named_t arcstat_l2_evict_reading; 678 kstat_named_t arcstat_l2_evict_l1cached; 679 kstat_named_t arcstat_l2_free_on_write; 680 kstat_named_t arcstat_l2_abort_lowmem; 681 kstat_named_t arcstat_l2_cksum_bad; 682 kstat_named_t arcstat_l2_io_error; 683 kstat_named_t arcstat_l2_lsize; 684 kstat_named_t arcstat_l2_psize; 685 /* Not updated directly; only synced in arc_kstat_update. */ 686 kstat_named_t arcstat_l2_hdr_size; 687 kstat_named_t arcstat_memory_throttle_count; 688 /* Not updated directly; only synced in arc_kstat_update. */ 689 kstat_named_t arcstat_meta_used; 690 kstat_named_t arcstat_meta_limit; 691 kstat_named_t arcstat_meta_max; 692 kstat_named_t arcstat_meta_min; 693 kstat_named_t arcstat_sync_wait_for_async; 694 kstat_named_t arcstat_demand_hit_predictive_prefetch; 695 } arc_stats_t; 696 697 static arc_stats_t arc_stats = { 698 { "hits", KSTAT_DATA_UINT64 }, 699 { "misses", KSTAT_DATA_UINT64 }, 700 { "demand_data_hits", KSTAT_DATA_UINT64 }, 701 { "demand_data_misses", KSTAT_DATA_UINT64 }, 702 { "demand_metadata_hits", KSTAT_DATA_UINT64 }, 703 { "demand_metadata_misses", KSTAT_DATA_UINT64 }, 704 { "prefetch_data_hits", KSTAT_DATA_UINT64 }, 705 { "prefetch_data_misses", KSTAT_DATA_UINT64 }, 706 { "prefetch_metadata_hits", KSTAT_DATA_UINT64 }, 707 { "prefetch_metadata_misses", KSTAT_DATA_UINT64 }, 708 { "mru_hits", KSTAT_DATA_UINT64 }, 709 { "mru_ghost_hits", KSTAT_DATA_UINT64 }, 710 { "mfu_hits", KSTAT_DATA_UINT64 }, 711 { "mfu_ghost_hits", KSTAT_DATA_UINT64 }, 712 { "deleted", KSTAT_DATA_UINT64 }, 713 { "mutex_miss", KSTAT_DATA_UINT64 }, 714 { "evict_skip", KSTAT_DATA_UINT64 }, 715 { "evict_not_enough", KSTAT_DATA_UINT64 }, 716 { "evict_l2_cached", KSTAT_DATA_UINT64 }, 717 { "evict_l2_eligible", KSTAT_DATA_UINT64 }, 718 { "evict_l2_ineligible", KSTAT_DATA_UINT64 }, 719 { "evict_l2_skip", KSTAT_DATA_UINT64 }, 720 { "hash_elements", KSTAT_DATA_UINT64 }, 721 { "hash_elements_max", KSTAT_DATA_UINT64 }, 722 { "hash_collisions", KSTAT_DATA_UINT64 }, 723 { "hash_chains", KSTAT_DATA_UINT64 }, 724 { "hash_chain_max", KSTAT_DATA_UINT64 }, 725 { "p", KSTAT_DATA_UINT64 }, 726 { "c", KSTAT_DATA_UINT64 }, 727 { "c_min", KSTAT_DATA_UINT64 }, 728 { "c_max", KSTAT_DATA_UINT64 }, 729 { "size", KSTAT_DATA_UINT64 }, 730 { "compressed_size", KSTAT_DATA_UINT64 }, 731 { "uncompressed_size", KSTAT_DATA_UINT64 }, 732 { "overhead_size", KSTAT_DATA_UINT64 }, 733 { "hdr_size", KSTAT_DATA_UINT64 }, 734 { "data_size", KSTAT_DATA_UINT64 }, 735 { "metadata_size", KSTAT_DATA_UINT64 }, 736 { "other_size", KSTAT_DATA_UINT64 }, 737 { "anon_size", KSTAT_DATA_UINT64 }, 738 { "anon_evictable_data", KSTAT_DATA_UINT64 }, 739 { "anon_evictable_metadata", KSTAT_DATA_UINT64 }, 740 { "mru_size", KSTAT_DATA_UINT64 }, 741 { "mru_evictable_data", KSTAT_DATA_UINT64 }, 742 { "mru_evictable_metadata", KSTAT_DATA_UINT64 }, 743 { "mru_ghost_size", KSTAT_DATA_UINT64 }, 744 { "mru_ghost_evictable_data", KSTAT_DATA_UINT64 }, 745 { "mru_ghost_evictable_metadata", KSTAT_DATA_UINT64 }, 746 { "mfu_size", KSTAT_DATA_UINT64 }, 747 { "mfu_evictable_data", KSTAT_DATA_UINT64 }, 748 { "mfu_evictable_metadata", KSTAT_DATA_UINT64 }, 749 { "mfu_ghost_size", KSTAT_DATA_UINT64 }, 750 { "mfu_ghost_evictable_data", KSTAT_DATA_UINT64 }, 751 { "mfu_ghost_evictable_metadata", KSTAT_DATA_UINT64 }, 752 { "l2_hits", KSTAT_DATA_UINT64 }, 753 { "l2_misses", KSTAT_DATA_UINT64 }, 754 { "l2_feeds", KSTAT_DATA_UINT64 }, 755 { "l2_rw_clash", KSTAT_DATA_UINT64 }, 756 { "l2_read_bytes", KSTAT_DATA_UINT64 }, 757 { "l2_write_bytes", KSTAT_DATA_UINT64 }, 758 { "l2_writes_sent", KSTAT_DATA_UINT64 }, 759 { "l2_writes_done", KSTAT_DATA_UINT64 }, 760 { "l2_writes_error", KSTAT_DATA_UINT64 }, 761 { "l2_writes_lock_retry", KSTAT_DATA_UINT64 }, 762 { "l2_evict_lock_retry", KSTAT_DATA_UINT64 }, 763 { "l2_evict_reading", KSTAT_DATA_UINT64 }, 764 { "l2_evict_l1cached", KSTAT_DATA_UINT64 }, 765 { "l2_free_on_write", KSTAT_DATA_UINT64 }, 766 { "l2_abort_lowmem", KSTAT_DATA_UINT64 }, 767 { "l2_cksum_bad", KSTAT_DATA_UINT64 }, 768 { "l2_io_error", KSTAT_DATA_UINT64 }, 769 { "l2_size", KSTAT_DATA_UINT64 }, 770 { "l2_asize", KSTAT_DATA_UINT64 }, 771 { "l2_hdr_size", KSTAT_DATA_UINT64 }, 772 { "memory_throttle_count", KSTAT_DATA_UINT64 }, 773 { "arc_meta_used", KSTAT_DATA_UINT64 }, 774 { "arc_meta_limit", KSTAT_DATA_UINT64 }, 775 { "arc_meta_max", KSTAT_DATA_UINT64 }, 776 { "arc_meta_min", KSTAT_DATA_UINT64 }, 777 { "sync_wait_for_async", KSTAT_DATA_UINT64 }, 778 { "demand_hit_predictive_prefetch", KSTAT_DATA_UINT64 }, 779 }; 780 781 #define ARCSTAT(stat) (arc_stats.stat.value.ui64) 782 783 #define ARCSTAT_INCR(stat, val) \ 784 atomic_add_64(&arc_stats.stat.value.ui64, (val)) 785 786 #define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1) 787 #define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1) 788 789 #define ARCSTAT_MAX(stat, val) { \ 790 uint64_t m; \ 791 while ((val) > (m = arc_stats.stat.value.ui64) && \ 792 (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \ 793 continue; \ 794 } 795 796 #define ARCSTAT_MAXSTAT(stat) \ 797 ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64) 798 799 /* 800 * We define a macro to allow ARC hits/misses to be easily broken down by 801 * two separate conditions, giving a total of four different subtypes for 802 * each of hits and misses (so eight statistics total). 803 */ 804 #define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \ 805 if (cond1) { \ 806 if (cond2) { \ 807 ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \ 808 } else { \ 809 ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \ 810 } \ 811 } else { \ 812 if (cond2) { \ 813 ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \ 814 } else { \ 815 ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\ 816 } \ 817 } 818 819 kstat_t *arc_ksp; 820 static arc_state_t *arc_anon; 821 static arc_state_t *arc_mru; 822 static arc_state_t *arc_mru_ghost; 823 static arc_state_t *arc_mfu; 824 static arc_state_t *arc_mfu_ghost; 825 static arc_state_t *arc_l2c_only; 826 827 /* 828 * There are several ARC variables that are critical to export as kstats -- 829 * but we don't want to have to grovel around in the kstat whenever we wish to 830 * manipulate them. For these variables, we therefore define them to be in 831 * terms of the statistic variable. This assures that we are not introducing 832 * the possibility of inconsistency by having shadow copies of the variables, 833 * while still allowing the code to be readable. 834 */ 835 #define arc_p ARCSTAT(arcstat_p) /* target size of MRU */ 836 #define arc_c ARCSTAT(arcstat_c) /* target size of cache */ 837 #define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */ 838 #define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */ 839 #define arc_meta_limit ARCSTAT(arcstat_meta_limit) /* max size for metadata */ 840 #define arc_meta_min ARCSTAT(arcstat_meta_min) /* min size for metadata */ 841 #define arc_meta_max ARCSTAT(arcstat_meta_max) /* max size of metadata */ 842 843 /* compressed size of entire arc */ 844 #define arc_compressed_size ARCSTAT(arcstat_compressed_size) 845 /* uncompressed size of entire arc */ 846 #define arc_uncompressed_size ARCSTAT(arcstat_uncompressed_size) 847 /* number of bytes in the arc from arc_buf_t's */ 848 #define arc_overhead_size ARCSTAT(arcstat_overhead_size) 849 850 /* 851 * There are also some ARC variables that we want to export, but that are 852 * updated so often that having the canonical representation be the statistic 853 * variable causes a performance bottleneck. We want to use aggsum_t's for these 854 * instead, but still be able to export the kstat in the same way as before. 855 * The solution is to always use the aggsum version, except in the kstat update 856 * callback. 857 */ 858 aggsum_t arc_size; 859 aggsum_t arc_meta_used; 860 aggsum_t astat_data_size; 861 aggsum_t astat_metadata_size; 862 aggsum_t astat_hdr_size; 863 aggsum_t astat_other_size; 864 aggsum_t astat_l2_hdr_size; 865 866 static int arc_no_grow; /* Don't try to grow cache size */ 867 static hrtime_t arc_growtime; 868 static uint64_t arc_tempreserve; 869 static uint64_t arc_loaned_bytes; 870 871 typedef struct arc_callback arc_callback_t; 872 873 struct arc_callback { 874 void *acb_private; 875 arc_done_func_t *acb_done; 876 arc_buf_t *acb_buf; 877 boolean_t acb_compressed; 878 zio_t *acb_zio_dummy; 879 arc_callback_t *acb_next; 880 }; 881 882 typedef struct arc_write_callback arc_write_callback_t; 883 884 struct arc_write_callback { 885 void *awcb_private; 886 arc_done_func_t *awcb_ready; 887 arc_done_func_t *awcb_children_ready; 888 arc_done_func_t *awcb_physdone; 889 arc_done_func_t *awcb_done; 890 arc_buf_t *awcb_buf; 891 }; 892 893 /* 894 * ARC buffers are separated into multiple structs as a memory saving measure: 895 * - Common fields struct, always defined, and embedded within it: 896 * - L2-only fields, always allocated but undefined when not in L2ARC 897 * - L1-only fields, only allocated when in L1ARC 898 * 899 * Buffer in L1 Buffer only in L2 900 * +------------------------+ +------------------------+ 901 * | arc_buf_hdr_t | | arc_buf_hdr_t | 902 * | | | | 903 * | | | | 904 * | | | | 905 * +------------------------+ +------------------------+ 906 * | l2arc_buf_hdr_t | | l2arc_buf_hdr_t | 907 * | (undefined if L1-only) | | | 908 * +------------------------+ +------------------------+ 909 * | l1arc_buf_hdr_t | 910 * | | 911 * | | 912 * | | 913 * | | 914 * +------------------------+ 915 * 916 * Because it's possible for the L2ARC to become extremely large, we can wind 917 * up eating a lot of memory in L2ARC buffer headers, so the size of a header 918 * is minimized by only allocating the fields necessary for an L1-cached buffer 919 * when a header is actually in the L1 cache. The sub-headers (l1arc_buf_hdr and 920 * l2arc_buf_hdr) are embedded rather than allocated separately to save a couple 921 * words in pointers. arc_hdr_realloc() is used to switch a header between 922 * these two allocation states. 923 */ 924 typedef struct l1arc_buf_hdr { 925 kmutex_t b_freeze_lock; 926 zio_cksum_t *b_freeze_cksum; 927 #ifdef ZFS_DEBUG 928 /* 929 * Used for debugging with kmem_flags - by allocating and freeing 930 * b_thawed when the buffer is thawed, we get a record of the stack 931 * trace that thawed it. 932 */ 933 void *b_thawed; 934 #endif 935 936 arc_buf_t *b_buf; 937 uint32_t b_bufcnt; 938 /* for waiting on writes to complete */ 939 kcondvar_t b_cv; 940 uint8_t b_byteswap; 941 942 /* protected by arc state mutex */ 943 arc_state_t *b_state; 944 multilist_node_t b_arc_node; 945 946 /* updated atomically */ 947 clock_t b_arc_access; 948 949 /* self protecting */ 950 zfs_refcount_t b_refcnt; 951 952 arc_callback_t *b_acb; 953 abd_t *b_pabd; 954 } l1arc_buf_hdr_t; 955 956 typedef struct l2arc_dev l2arc_dev_t; 957 958 typedef struct l2arc_buf_hdr { 959 /* protected by arc_buf_hdr mutex */ 960 l2arc_dev_t *b_dev; /* L2ARC device */ 961 uint64_t b_daddr; /* disk address, offset byte */ 962 963 list_node_t b_l2node; 964 } l2arc_buf_hdr_t; 965 966 struct arc_buf_hdr { 967 /* protected by hash lock */ 968 dva_t b_dva; 969 uint64_t b_birth; 970 971 arc_buf_contents_t b_type; 972 arc_buf_hdr_t *b_hash_next; 973 arc_flags_t b_flags; 974 975 /* 976 * This field stores the size of the data buffer after 977 * compression, and is set in the arc's zio completion handlers. 978 * It is in units of SPA_MINBLOCKSIZE (e.g. 1 == 512 bytes). 979 * 980 * While the block pointers can store up to 32MB in their psize 981 * field, we can only store up to 32MB minus 512B. This is due 982 * to the bp using a bias of 1, whereas we use a bias of 0 (i.e. 983 * a field of zeros represents 512B in the bp). We can't use a 984 * bias of 1 since we need to reserve a psize of zero, here, to 985 * represent holes and embedded blocks. 986 * 987 * This isn't a problem in practice, since the maximum size of a 988 * buffer is limited to 16MB, so we never need to store 32MB in 989 * this field. Even in the upstream illumos code base, the 990 * maximum size of a buffer is limited to 16MB. 991 */ 992 uint16_t b_psize; 993 994 /* 995 * This field stores the size of the data buffer before 996 * compression, and cannot change once set. It is in units 997 * of SPA_MINBLOCKSIZE (e.g. 2 == 1024 bytes) 998 */ 999 uint16_t b_lsize; /* immutable */ 1000 uint64_t b_spa; /* immutable */ 1001 1002 /* L2ARC fields. Undefined when not in L2ARC. */ 1003 l2arc_buf_hdr_t b_l2hdr; 1004 /* L1ARC fields. Undefined when in l2arc_only state */ 1005 l1arc_buf_hdr_t b_l1hdr; 1006 }; 1007 1008 #define GHOST_STATE(state) \ 1009 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \ 1010 (state) == arc_l2c_only) 1011 1012 #define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_FLAG_IN_HASH_TABLE) 1013 #define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) 1014 #define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_FLAG_IO_ERROR) 1015 #define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_FLAG_PREFETCH) 1016 #define HDR_COMPRESSION_ENABLED(hdr) \ 1017 ((hdr)->b_flags & ARC_FLAG_COMPRESSED_ARC) 1018 1019 #define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_FLAG_L2CACHE) 1020 #define HDR_L2_READING(hdr) \ 1021 (((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) && \ 1022 ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR)) 1023 #define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITING) 1024 #define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_FLAG_L2_EVICTED) 1025 #define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITE_HEAD) 1026 #define HDR_SHARED_DATA(hdr) ((hdr)->b_flags & ARC_FLAG_SHARED_DATA) 1027 1028 #define HDR_ISTYPE_METADATA(hdr) \ 1029 ((hdr)->b_flags & ARC_FLAG_BUFC_METADATA) 1030 #define HDR_ISTYPE_DATA(hdr) (!HDR_ISTYPE_METADATA(hdr)) 1031 1032 #define HDR_HAS_L1HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L1HDR) 1033 #define HDR_HAS_L2HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR) 1034 1035 /* For storing compression mode in b_flags */ 1036 #define HDR_COMPRESS_OFFSET (highbit64(ARC_FLAG_COMPRESS_0) - 1) 1037 1038 #define HDR_GET_COMPRESS(hdr) ((enum zio_compress)BF32_GET((hdr)->b_flags, \ 1039 HDR_COMPRESS_OFFSET, SPA_COMPRESSBITS)) 1040 #define HDR_SET_COMPRESS(hdr, cmp) BF32_SET((hdr)->b_flags, \ 1041 HDR_COMPRESS_OFFSET, SPA_COMPRESSBITS, (cmp)); 1042 1043 #define ARC_BUF_LAST(buf) ((buf)->b_next == NULL) 1044 #define ARC_BUF_SHARED(buf) ((buf)->b_flags & ARC_BUF_FLAG_SHARED) 1045 #define ARC_BUF_COMPRESSED(buf) ((buf)->b_flags & ARC_BUF_FLAG_COMPRESSED) 1046 1047 /* 1048 * Other sizes 1049 */ 1050 1051 #define HDR_FULL_SIZE ((int64_t)sizeof (arc_buf_hdr_t)) 1052 #define HDR_L2ONLY_SIZE ((int64_t)offsetof(arc_buf_hdr_t, b_l1hdr)) 1053 1054 /* 1055 * Hash table routines 1056 */ 1057 1058 #define HT_LOCK_PAD 64 1059 1060 struct ht_lock { 1061 kmutex_t ht_lock; 1062 #ifdef _KERNEL 1063 unsigned char pad[(HT_LOCK_PAD - sizeof (kmutex_t))]; 1064 #endif 1065 }; 1066 1067 #define BUF_LOCKS 256 1068 typedef struct buf_hash_table { 1069 uint64_t ht_mask; 1070 arc_buf_hdr_t **ht_table; 1071 struct ht_lock ht_locks[BUF_LOCKS]; 1072 } buf_hash_table_t; 1073 1074 static buf_hash_table_t buf_hash_table; 1075 1076 #define BUF_HASH_INDEX(spa, dva, birth) \ 1077 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask) 1078 #define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)]) 1079 #define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock)) 1080 #define HDR_LOCK(hdr) \ 1081 (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth))) 1082 1083 uint64_t zfs_crc64_table[256]; 1084 1085 /* 1086 * Level 2 ARC 1087 */ 1088 1089 #define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */ 1090 #define L2ARC_HEADROOM 2 /* num of writes */ 1091 /* 1092 * If we discover during ARC scan any buffers to be compressed, we boost 1093 * our headroom for the next scanning cycle by this percentage multiple. 1094 */ 1095 #define L2ARC_HEADROOM_BOOST 200 1096 #define L2ARC_FEED_SECS 1 /* caching interval secs */ 1097 #define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */ 1098 1099 #define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent) 1100 #define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done) 1101 1102 /* L2ARC Performance Tunables */ 1103 uint64_t l2arc_write_max = L2ARC_WRITE_SIZE; /* default max write size */ 1104 uint64_t l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra write during warmup */ 1105 uint64_t l2arc_headroom = L2ARC_HEADROOM; /* number of dev writes */ 1106 uint64_t l2arc_headroom_boost = L2ARC_HEADROOM_BOOST; 1107 uint64_t l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */ 1108 uint64_t l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval milliseconds */ 1109 boolean_t l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */ 1110 boolean_t l2arc_feed_again = B_TRUE; /* turbo warmup */ 1111 boolean_t l2arc_norw = B_TRUE; /* no reads during writes */ 1112 1113 /* 1114 * L2ARC Internals 1115 */ 1116 struct l2arc_dev { 1117 vdev_t *l2ad_vdev; /* vdev */ 1118 spa_t *l2ad_spa; /* spa */ 1119 uint64_t l2ad_hand; /* next write location */ 1120 uint64_t l2ad_start; /* first addr on device */ 1121 uint64_t l2ad_end; /* last addr on device */ 1122 boolean_t l2ad_first; /* first sweep through */ 1123 boolean_t l2ad_writing; /* currently writing */ 1124 kmutex_t l2ad_mtx; /* lock for buffer list */ 1125 list_t l2ad_buflist; /* buffer list */ 1126 list_node_t l2ad_node; /* device list node */ 1127 zfs_refcount_t l2ad_alloc; /* allocated bytes */ 1128 }; 1129 1130 static list_t L2ARC_dev_list; /* device list */ 1131 static list_t *l2arc_dev_list; /* device list pointer */ 1132 static kmutex_t l2arc_dev_mtx; /* device list mutex */ 1133 static l2arc_dev_t *l2arc_dev_last; /* last device used */ 1134 static list_t L2ARC_free_on_write; /* free after write buf list */ 1135 static list_t *l2arc_free_on_write; /* free after write list ptr */ 1136 static kmutex_t l2arc_free_on_write_mtx; /* mutex for list */ 1137 static uint64_t l2arc_ndev; /* number of devices */ 1138 1139 typedef struct l2arc_read_callback { 1140 arc_buf_hdr_t *l2rcb_hdr; /* read header */ 1141 blkptr_t l2rcb_bp; /* original blkptr */ 1142 zbookmark_phys_t l2rcb_zb; /* original bookmark */ 1143 int l2rcb_flags; /* original flags */ 1144 abd_t *l2rcb_abd; /* temporary buffer */ 1145 } l2arc_read_callback_t; 1146 1147 typedef struct l2arc_write_callback { 1148 l2arc_dev_t *l2wcb_dev; /* device info */ 1149 arc_buf_hdr_t *l2wcb_head; /* head of write buflist */ 1150 } l2arc_write_callback_t; 1151 1152 typedef struct l2arc_data_free { 1153 /* protected by l2arc_free_on_write_mtx */ 1154 abd_t *l2df_abd; 1155 size_t l2df_size; 1156 arc_buf_contents_t l2df_type; 1157 list_node_t l2df_list_node; 1158 } l2arc_data_free_t; 1159 1160 static kmutex_t l2arc_feed_thr_lock; 1161 static kcondvar_t l2arc_feed_thr_cv; 1162 static uint8_t l2arc_thread_exit; 1163 1164 static abd_t *arc_get_data_abd(arc_buf_hdr_t *, uint64_t, void *); 1165 static void *arc_get_data_buf(arc_buf_hdr_t *, uint64_t, void *); 1166 static void arc_get_data_impl(arc_buf_hdr_t *, uint64_t, void *); 1167 static void arc_free_data_abd(arc_buf_hdr_t *, abd_t *, uint64_t, void *); 1168 static void arc_free_data_buf(arc_buf_hdr_t *, void *, uint64_t, void *); 1169 static void arc_free_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag); 1170 static void arc_hdr_free_pabd(arc_buf_hdr_t *); 1171 static void arc_hdr_alloc_pabd(arc_buf_hdr_t *); 1172 static void arc_access(arc_buf_hdr_t *, kmutex_t *); 1173 static boolean_t arc_is_overflowing(); 1174 static void arc_buf_watch(arc_buf_t *); 1175 1176 static arc_buf_contents_t arc_buf_type(arc_buf_hdr_t *); 1177 static uint32_t arc_bufc_to_flags(arc_buf_contents_t); 1178 static inline void arc_hdr_set_flags(arc_buf_hdr_t *hdr, arc_flags_t flags); 1179 static inline void arc_hdr_clear_flags(arc_buf_hdr_t *hdr, arc_flags_t flags); 1180 1181 static boolean_t l2arc_write_eligible(uint64_t, arc_buf_hdr_t *); 1182 static void l2arc_read_done(zio_t *); 1183 1184 1185 /* 1186 * We use Cityhash for this. It's fast, and has good hash properties without 1187 * requiring any large static buffers. 1188 */ 1189 static uint64_t 1190 buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth) 1191 { 1192 return (cityhash4(spa, dva->dva_word[0], dva->dva_word[1], birth)); 1193 } 1194 1195 #define HDR_EMPTY(hdr) \ 1196 ((hdr)->b_dva.dva_word[0] == 0 && \ 1197 (hdr)->b_dva.dva_word[1] == 0) 1198 1199 #define HDR_EQUAL(spa, dva, birth, hdr) \ 1200 ((hdr)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \ 1201 ((hdr)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \ 1202 ((hdr)->b_birth == birth) && ((hdr)->b_spa == spa) 1203 1204 static void 1205 buf_discard_identity(arc_buf_hdr_t *hdr) 1206 { 1207 hdr->b_dva.dva_word[0] = 0; 1208 hdr->b_dva.dva_word[1] = 0; 1209 hdr->b_birth = 0; 1210 } 1211 1212 static arc_buf_hdr_t * 1213 buf_hash_find(uint64_t spa, const blkptr_t *bp, kmutex_t **lockp) 1214 { 1215 const dva_t *dva = BP_IDENTITY(bp); 1216 uint64_t birth = BP_PHYSICAL_BIRTH(bp); 1217 uint64_t idx = BUF_HASH_INDEX(spa, dva, birth); 1218 kmutex_t *hash_lock = BUF_HASH_LOCK(idx); 1219 arc_buf_hdr_t *hdr; 1220 1221 mutex_enter(hash_lock); 1222 for (hdr = buf_hash_table.ht_table[idx]; hdr != NULL; 1223 hdr = hdr->b_hash_next) { 1224 if (HDR_EQUAL(spa, dva, birth, hdr)) { 1225 *lockp = hash_lock; 1226 return (hdr); 1227 } 1228 } 1229 mutex_exit(hash_lock); 1230 *lockp = NULL; 1231 return (NULL); 1232 } 1233 1234 /* 1235 * Insert an entry into the hash table. If there is already an element 1236 * equal to elem in the hash table, then the already existing element 1237 * will be returned and the new element will not be inserted. 1238 * Otherwise returns NULL. 1239 * If lockp == NULL, the caller is assumed to already hold the hash lock. 1240 */ 1241 static arc_buf_hdr_t * 1242 buf_hash_insert(arc_buf_hdr_t *hdr, kmutex_t **lockp) 1243 { 1244 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth); 1245 kmutex_t *hash_lock = BUF_HASH_LOCK(idx); 1246 arc_buf_hdr_t *fhdr; 1247 uint32_t i; 1248 1249 ASSERT(!DVA_IS_EMPTY(&hdr->b_dva)); 1250 ASSERT(hdr->b_birth != 0); 1251 ASSERT(!HDR_IN_HASH_TABLE(hdr)); 1252 1253 if (lockp != NULL) { 1254 *lockp = hash_lock; 1255 mutex_enter(hash_lock); 1256 } else { 1257 ASSERT(MUTEX_HELD(hash_lock)); 1258 } 1259 1260 for (fhdr = buf_hash_table.ht_table[idx], i = 0; fhdr != NULL; 1261 fhdr = fhdr->b_hash_next, i++) { 1262 if (HDR_EQUAL(hdr->b_spa, &hdr->b_dva, hdr->b_birth, fhdr)) 1263 return (fhdr); 1264 } 1265 1266 hdr->b_hash_next = buf_hash_table.ht_table[idx]; 1267 buf_hash_table.ht_table[idx] = hdr; 1268 arc_hdr_set_flags(hdr, ARC_FLAG_IN_HASH_TABLE); 1269 1270 /* collect some hash table performance data */ 1271 if (i > 0) { 1272 ARCSTAT_BUMP(arcstat_hash_collisions); 1273 if (i == 1) 1274 ARCSTAT_BUMP(arcstat_hash_chains); 1275 1276 ARCSTAT_MAX(arcstat_hash_chain_max, i); 1277 } 1278 1279 ARCSTAT_BUMP(arcstat_hash_elements); 1280 ARCSTAT_MAXSTAT(arcstat_hash_elements); 1281 1282 return (NULL); 1283 } 1284 1285 static void 1286 buf_hash_remove(arc_buf_hdr_t *hdr) 1287 { 1288 arc_buf_hdr_t *fhdr, **hdrp; 1289 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth); 1290 1291 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx))); 1292 ASSERT(HDR_IN_HASH_TABLE(hdr)); 1293 1294 hdrp = &buf_hash_table.ht_table[idx]; 1295 while ((fhdr = *hdrp) != hdr) { 1296 ASSERT3P(fhdr, !=, NULL); 1297 hdrp = &fhdr->b_hash_next; 1298 } 1299 *hdrp = hdr->b_hash_next; 1300 hdr->b_hash_next = NULL; 1301 arc_hdr_clear_flags(hdr, ARC_FLAG_IN_HASH_TABLE); 1302 1303 /* collect some hash table performance data */ 1304 ARCSTAT_BUMPDOWN(arcstat_hash_elements); 1305 1306 if (buf_hash_table.ht_table[idx] && 1307 buf_hash_table.ht_table[idx]->b_hash_next == NULL) 1308 ARCSTAT_BUMPDOWN(arcstat_hash_chains); 1309 } 1310 1311 /* 1312 * Global data structures and functions for the buf kmem cache. 1313 */ 1314 static kmem_cache_t *hdr_full_cache; 1315 static kmem_cache_t *hdr_l2only_cache; 1316 static kmem_cache_t *buf_cache; 1317 1318 static void 1319 buf_fini(void) 1320 { 1321 int i; 1322 1323 kmem_free(buf_hash_table.ht_table, 1324 (buf_hash_table.ht_mask + 1) * sizeof (void *)); 1325 for (i = 0; i < BUF_LOCKS; i++) 1326 mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock); 1327 kmem_cache_destroy(hdr_full_cache); 1328 kmem_cache_destroy(hdr_l2only_cache); 1329 kmem_cache_destroy(buf_cache); 1330 } 1331 1332 /* 1333 * Constructor callback - called when the cache is empty 1334 * and a new buf is requested. 1335 */ 1336 /* ARGSUSED */ 1337 static int 1338 hdr_full_cons(void *vbuf, void *unused, int kmflag) 1339 { 1340 arc_buf_hdr_t *hdr = vbuf; 1341 1342 bzero(hdr, HDR_FULL_SIZE); 1343 cv_init(&hdr->b_l1hdr.b_cv, NULL, CV_DEFAULT, NULL); 1344 zfs_refcount_create(&hdr->b_l1hdr.b_refcnt); 1345 mutex_init(&hdr->b_l1hdr.b_freeze_lock, NULL, MUTEX_DEFAULT, NULL); 1346 multilist_link_init(&hdr->b_l1hdr.b_arc_node); 1347 arc_space_consume(HDR_FULL_SIZE, ARC_SPACE_HDRS); 1348 1349 return (0); 1350 } 1351 1352 /* ARGSUSED */ 1353 static int 1354 hdr_l2only_cons(void *vbuf, void *unused, int kmflag) 1355 { 1356 arc_buf_hdr_t *hdr = vbuf; 1357 1358 bzero(hdr, HDR_L2ONLY_SIZE); 1359 arc_space_consume(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS); 1360 1361 return (0); 1362 } 1363 1364 /* ARGSUSED */ 1365 static int 1366 buf_cons(void *vbuf, void *unused, int kmflag) 1367 { 1368 arc_buf_t *buf = vbuf; 1369 1370 bzero(buf, sizeof (arc_buf_t)); 1371 mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL); 1372 arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS); 1373 1374 return (0); 1375 } 1376 1377 /* 1378 * Destructor callback - called when a cached buf is 1379 * no longer required. 1380 */ 1381 /* ARGSUSED */ 1382 static void 1383 hdr_full_dest(void *vbuf, void *unused) 1384 { 1385 arc_buf_hdr_t *hdr = vbuf; 1386 1387 ASSERT(HDR_EMPTY(hdr)); 1388 cv_destroy(&hdr->b_l1hdr.b_cv); 1389 zfs_refcount_destroy(&hdr->b_l1hdr.b_refcnt); 1390 mutex_destroy(&hdr->b_l1hdr.b_freeze_lock); 1391 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 1392 arc_space_return(HDR_FULL_SIZE, ARC_SPACE_HDRS); 1393 } 1394 1395 /* ARGSUSED */ 1396 static void 1397 hdr_l2only_dest(void *vbuf, void *unused) 1398 { 1399 arc_buf_hdr_t *hdr = vbuf; 1400 1401 ASSERT(HDR_EMPTY(hdr)); 1402 arc_space_return(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS); 1403 } 1404 1405 /* ARGSUSED */ 1406 static void 1407 buf_dest(void *vbuf, void *unused) 1408 { 1409 arc_buf_t *buf = vbuf; 1410 1411 mutex_destroy(&buf->b_evict_lock); 1412 arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS); 1413 } 1414 1415 /* 1416 * Reclaim callback -- invoked when memory is low. 1417 */ 1418 /* ARGSUSED */ 1419 static void 1420 hdr_recl(void *unused) 1421 { 1422 dprintf("hdr_recl called\n"); 1423 /* 1424 * umem calls the reclaim func when we destroy the buf cache, 1425 * which is after we do arc_fini(). 1426 */ 1427 if (arc_initialized) 1428 zthr_wakeup(arc_reap_zthr); 1429 } 1430 1431 static void 1432 buf_init(void) 1433 { 1434 uint64_t *ct; 1435 uint64_t hsize = 1ULL << 12; 1436 int i, j; 1437 1438 /* 1439 * The hash table is big enough to fill all of physical memory 1440 * with an average block size of zfs_arc_average_blocksize (default 8K). 1441 * By default, the table will take up 1442 * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers). 1443 */ 1444 while (hsize * zfs_arc_average_blocksize < physmem * PAGESIZE) 1445 hsize <<= 1; 1446 retry: 1447 buf_hash_table.ht_mask = hsize - 1; 1448 buf_hash_table.ht_table = 1449 kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP); 1450 if (buf_hash_table.ht_table == NULL) { 1451 ASSERT(hsize > (1ULL << 8)); 1452 hsize >>= 1; 1453 goto retry; 1454 } 1455 1456 hdr_full_cache = kmem_cache_create("arc_buf_hdr_t_full", HDR_FULL_SIZE, 1457 0, hdr_full_cons, hdr_full_dest, hdr_recl, NULL, NULL, 0); 1458 hdr_l2only_cache = kmem_cache_create("arc_buf_hdr_t_l2only", 1459 HDR_L2ONLY_SIZE, 0, hdr_l2only_cons, hdr_l2only_dest, hdr_recl, 1460 NULL, NULL, 0); 1461 buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t), 1462 0, buf_cons, buf_dest, NULL, NULL, NULL, 0); 1463 1464 for (i = 0; i < 256; i++) 1465 for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--) 1466 *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY); 1467 1468 for (i = 0; i < BUF_LOCKS; i++) { 1469 mutex_init(&buf_hash_table.ht_locks[i].ht_lock, 1470 NULL, MUTEX_DEFAULT, NULL); 1471 } 1472 } 1473 1474 /* 1475 * This is the size that the buf occupies in memory. If the buf is compressed, 1476 * it will correspond to the compressed size. You should use this method of 1477 * getting the buf size unless you explicitly need the logical size. 1478 */ 1479 int32_t 1480 arc_buf_size(arc_buf_t *buf) 1481 { 1482 return (ARC_BUF_COMPRESSED(buf) ? 1483 HDR_GET_PSIZE(buf->b_hdr) : HDR_GET_LSIZE(buf->b_hdr)); 1484 } 1485 1486 int32_t 1487 arc_buf_lsize(arc_buf_t *buf) 1488 { 1489 return (HDR_GET_LSIZE(buf->b_hdr)); 1490 } 1491 1492 enum zio_compress 1493 arc_get_compression(arc_buf_t *buf) 1494 { 1495 return (ARC_BUF_COMPRESSED(buf) ? 1496 HDR_GET_COMPRESS(buf->b_hdr) : ZIO_COMPRESS_OFF); 1497 } 1498 1499 #define ARC_MINTIME (hz>>4) /* 62 ms */ 1500 1501 static inline boolean_t 1502 arc_buf_is_shared(arc_buf_t *buf) 1503 { 1504 boolean_t shared = (buf->b_data != NULL && 1505 buf->b_hdr->b_l1hdr.b_pabd != NULL && 1506 abd_is_linear(buf->b_hdr->b_l1hdr.b_pabd) && 1507 buf->b_data == abd_to_buf(buf->b_hdr->b_l1hdr.b_pabd)); 1508 IMPLY(shared, HDR_SHARED_DATA(buf->b_hdr)); 1509 IMPLY(shared, ARC_BUF_SHARED(buf)); 1510 IMPLY(shared, ARC_BUF_COMPRESSED(buf) || ARC_BUF_LAST(buf)); 1511 1512 /* 1513 * It would be nice to assert arc_can_share() too, but the "hdr isn't 1514 * already being shared" requirement prevents us from doing that. 1515 */ 1516 1517 return (shared); 1518 } 1519 1520 /* 1521 * Free the checksum associated with this header. If there is no checksum, this 1522 * is a no-op. 1523 */ 1524 static inline void 1525 arc_cksum_free(arc_buf_hdr_t *hdr) 1526 { 1527 ASSERT(HDR_HAS_L1HDR(hdr)); 1528 mutex_enter(&hdr->b_l1hdr.b_freeze_lock); 1529 if (hdr->b_l1hdr.b_freeze_cksum != NULL) { 1530 kmem_free(hdr->b_l1hdr.b_freeze_cksum, sizeof (zio_cksum_t)); 1531 hdr->b_l1hdr.b_freeze_cksum = NULL; 1532 } 1533 mutex_exit(&hdr->b_l1hdr.b_freeze_lock); 1534 } 1535 1536 /* 1537 * Return true iff at least one of the bufs on hdr is not compressed. 1538 */ 1539 static boolean_t 1540 arc_hdr_has_uncompressed_buf(arc_buf_hdr_t *hdr) 1541 { 1542 for (arc_buf_t *b = hdr->b_l1hdr.b_buf; b != NULL; b = b->b_next) { 1543 if (!ARC_BUF_COMPRESSED(b)) { 1544 return (B_TRUE); 1545 } 1546 } 1547 return (B_FALSE); 1548 } 1549 1550 /* 1551 * If we've turned on the ZFS_DEBUG_MODIFY flag, verify that the buf's data 1552 * matches the checksum that is stored in the hdr. If there is no checksum, 1553 * or if the buf is compressed, this is a no-op. 1554 */ 1555 static void 1556 arc_cksum_verify(arc_buf_t *buf) 1557 { 1558 arc_buf_hdr_t *hdr = buf->b_hdr; 1559 zio_cksum_t zc; 1560 1561 if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 1562 return; 1563 1564 if (ARC_BUF_COMPRESSED(buf)) { 1565 ASSERT(hdr->b_l1hdr.b_freeze_cksum == NULL || 1566 arc_hdr_has_uncompressed_buf(hdr)); 1567 return; 1568 } 1569 1570 ASSERT(HDR_HAS_L1HDR(hdr)); 1571 1572 mutex_enter(&hdr->b_l1hdr.b_freeze_lock); 1573 if (hdr->b_l1hdr.b_freeze_cksum == NULL || HDR_IO_ERROR(hdr)) { 1574 mutex_exit(&hdr->b_l1hdr.b_freeze_lock); 1575 return; 1576 } 1577 1578 fletcher_2_native(buf->b_data, arc_buf_size(buf), NULL, &zc); 1579 if (!ZIO_CHECKSUM_EQUAL(*hdr->b_l1hdr.b_freeze_cksum, zc)) 1580 panic("buffer modified while frozen!"); 1581 mutex_exit(&hdr->b_l1hdr.b_freeze_lock); 1582 } 1583 1584 static boolean_t 1585 arc_cksum_is_equal(arc_buf_hdr_t *hdr, zio_t *zio) 1586 { 1587 enum zio_compress compress = BP_GET_COMPRESS(zio->io_bp); 1588 boolean_t valid_cksum; 1589 1590 ASSERT(!BP_IS_EMBEDDED(zio->io_bp)); 1591 VERIFY3U(BP_GET_PSIZE(zio->io_bp), ==, HDR_GET_PSIZE(hdr)); 1592 1593 /* 1594 * We rely on the blkptr's checksum to determine if the block 1595 * is valid or not. When compressed arc is enabled, the l2arc 1596 * writes the block to the l2arc just as it appears in the pool. 1597 * This allows us to use the blkptr's checksum to validate the 1598 * data that we just read off of the l2arc without having to store 1599 * a separate checksum in the arc_buf_hdr_t. However, if compressed 1600 * arc is disabled, then the data written to the l2arc is always 1601 * uncompressed and won't match the block as it exists in the main 1602 * pool. When this is the case, we must first compress it if it is 1603 * compressed on the main pool before we can validate the checksum. 1604 */ 1605 if (!HDR_COMPRESSION_ENABLED(hdr) && compress != ZIO_COMPRESS_OFF) { 1606 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF); 1607 uint64_t lsize = HDR_GET_LSIZE(hdr); 1608 uint64_t csize; 1609 1610 abd_t *cdata = abd_alloc_linear(HDR_GET_PSIZE(hdr), B_TRUE); 1611 csize = zio_compress_data(compress, zio->io_abd, 1612 abd_to_buf(cdata), lsize); 1613 1614 ASSERT3U(csize, <=, HDR_GET_PSIZE(hdr)); 1615 if (csize < HDR_GET_PSIZE(hdr)) { 1616 /* 1617 * Compressed blocks are always a multiple of the 1618 * smallest ashift in the pool. Ideally, we would 1619 * like to round up the csize to the next 1620 * spa_min_ashift but that value may have changed 1621 * since the block was last written. Instead, 1622 * we rely on the fact that the hdr's psize 1623 * was set to the psize of the block when it was 1624 * last written. We set the csize to that value 1625 * and zero out any part that should not contain 1626 * data. 1627 */ 1628 abd_zero_off(cdata, csize, HDR_GET_PSIZE(hdr) - csize); 1629 csize = HDR_GET_PSIZE(hdr); 1630 } 1631 zio_push_transform(zio, cdata, csize, HDR_GET_PSIZE(hdr), NULL); 1632 } 1633 1634 /* 1635 * Block pointers always store the checksum for the logical data. 1636 * If the block pointer has the gang bit set, then the checksum 1637 * it represents is for the reconstituted data and not for an 1638 * individual gang member. The zio pipeline, however, must be able to 1639 * determine the checksum of each of the gang constituents so it 1640 * treats the checksum comparison differently than what we need 1641 * for l2arc blocks. This prevents us from using the 1642 * zio_checksum_error() interface directly. Instead we must call the 1643 * zio_checksum_error_impl() so that we can ensure the checksum is 1644 * generated using the correct checksum algorithm and accounts for the 1645 * logical I/O size and not just a gang fragment. 1646 */ 1647 valid_cksum = (zio_checksum_error_impl(zio->io_spa, zio->io_bp, 1648 BP_GET_CHECKSUM(zio->io_bp), zio->io_abd, zio->io_size, 1649 zio->io_offset, NULL) == 0); 1650 zio_pop_transforms(zio); 1651 return (valid_cksum); 1652 } 1653 1654 /* 1655 * Given a buf full of data, if ZFS_DEBUG_MODIFY is enabled this computes a 1656 * checksum and attaches it to the buf's hdr so that we can ensure that the buf 1657 * isn't modified later on. If buf is compressed or there is already a checksum 1658 * on the hdr, this is a no-op (we only checksum uncompressed bufs). 1659 */ 1660 static void 1661 arc_cksum_compute(arc_buf_t *buf) 1662 { 1663 arc_buf_hdr_t *hdr = buf->b_hdr; 1664 1665 if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 1666 return; 1667 1668 ASSERT(HDR_HAS_L1HDR(hdr)); 1669 1670 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1671 if (hdr->b_l1hdr.b_freeze_cksum != NULL) { 1672 ASSERT(arc_hdr_has_uncompressed_buf(hdr)); 1673 mutex_exit(&hdr->b_l1hdr.b_freeze_lock); 1674 return; 1675 } else if (ARC_BUF_COMPRESSED(buf)) { 1676 mutex_exit(&hdr->b_l1hdr.b_freeze_lock); 1677 return; 1678 } 1679 1680 ASSERT(!ARC_BUF_COMPRESSED(buf)); 1681 hdr->b_l1hdr.b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t), 1682 KM_SLEEP); 1683 fletcher_2_native(buf->b_data, arc_buf_size(buf), NULL, 1684 hdr->b_l1hdr.b_freeze_cksum); 1685 mutex_exit(&hdr->b_l1hdr.b_freeze_lock); 1686 arc_buf_watch(buf); 1687 } 1688 1689 #ifndef _KERNEL 1690 typedef struct procctl { 1691 long cmd; 1692 prwatch_t prwatch; 1693 } procctl_t; 1694 #endif 1695 1696 /* ARGSUSED */ 1697 static void 1698 arc_buf_unwatch(arc_buf_t *buf) 1699 { 1700 #ifndef _KERNEL 1701 if (arc_watch) { 1702 int result; 1703 procctl_t ctl; 1704 ctl.cmd = PCWATCH; 1705 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data; 1706 ctl.prwatch.pr_size = 0; 1707 ctl.prwatch.pr_wflags = 0; 1708 result = write(arc_procfd, &ctl, sizeof (ctl)); 1709 ASSERT3U(result, ==, sizeof (ctl)); 1710 } 1711 #endif 1712 } 1713 1714 /* ARGSUSED */ 1715 static void 1716 arc_buf_watch(arc_buf_t *buf) 1717 { 1718 #ifndef _KERNEL 1719 if (arc_watch) { 1720 int result; 1721 procctl_t ctl; 1722 ctl.cmd = PCWATCH; 1723 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data; 1724 ctl.prwatch.pr_size = arc_buf_size(buf); 1725 ctl.prwatch.pr_wflags = WA_WRITE; 1726 result = write(arc_procfd, &ctl, sizeof (ctl)); 1727 ASSERT3U(result, ==, sizeof (ctl)); 1728 } 1729 #endif 1730 } 1731 1732 static arc_buf_contents_t 1733 arc_buf_type(arc_buf_hdr_t *hdr) 1734 { 1735 arc_buf_contents_t type; 1736 if (HDR_ISTYPE_METADATA(hdr)) { 1737 type = ARC_BUFC_METADATA; 1738 } else { 1739 type = ARC_BUFC_DATA; 1740 } 1741 VERIFY3U(hdr->b_type, ==, type); 1742 return (type); 1743 } 1744 1745 boolean_t 1746 arc_is_metadata(arc_buf_t *buf) 1747 { 1748 return (HDR_ISTYPE_METADATA(buf->b_hdr) != 0); 1749 } 1750 1751 static uint32_t 1752 arc_bufc_to_flags(arc_buf_contents_t type) 1753 { 1754 switch (type) { 1755 case ARC_BUFC_DATA: 1756 /* metadata field is 0 if buffer contains normal data */ 1757 return (0); 1758 case ARC_BUFC_METADATA: 1759 return (ARC_FLAG_BUFC_METADATA); 1760 default: 1761 break; 1762 } 1763 panic("undefined ARC buffer type!"); 1764 return ((uint32_t)-1); 1765 } 1766 1767 void 1768 arc_buf_thaw(arc_buf_t *buf) 1769 { 1770 arc_buf_hdr_t *hdr = buf->b_hdr; 1771 1772 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon); 1773 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 1774 1775 arc_cksum_verify(buf); 1776 1777 /* 1778 * Compressed buffers do not manipulate the b_freeze_cksum or 1779 * allocate b_thawed. 1780 */ 1781 if (ARC_BUF_COMPRESSED(buf)) { 1782 ASSERT(hdr->b_l1hdr.b_freeze_cksum == NULL || 1783 arc_hdr_has_uncompressed_buf(hdr)); 1784 return; 1785 } 1786 1787 ASSERT(HDR_HAS_L1HDR(hdr)); 1788 arc_cksum_free(hdr); 1789 1790 mutex_enter(&hdr->b_l1hdr.b_freeze_lock); 1791 #ifdef ZFS_DEBUG 1792 if (zfs_flags & ZFS_DEBUG_MODIFY) { 1793 if (hdr->b_l1hdr.b_thawed != NULL) 1794 kmem_free(hdr->b_l1hdr.b_thawed, 1); 1795 hdr->b_l1hdr.b_thawed = kmem_alloc(1, KM_SLEEP); 1796 } 1797 #endif 1798 1799 mutex_exit(&hdr->b_l1hdr.b_freeze_lock); 1800 1801 arc_buf_unwatch(buf); 1802 } 1803 1804 void 1805 arc_buf_freeze(arc_buf_t *buf) 1806 { 1807 arc_buf_hdr_t *hdr = buf->b_hdr; 1808 kmutex_t *hash_lock; 1809 1810 if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 1811 return; 1812 1813 if (ARC_BUF_COMPRESSED(buf)) { 1814 ASSERT(hdr->b_l1hdr.b_freeze_cksum == NULL || 1815 arc_hdr_has_uncompressed_buf(hdr)); 1816 return; 1817 } 1818 1819 hash_lock = HDR_LOCK(hdr); 1820 mutex_enter(hash_lock); 1821 1822 ASSERT(HDR_HAS_L1HDR(hdr)); 1823 ASSERT(hdr->b_l1hdr.b_freeze_cksum != NULL || 1824 hdr->b_l1hdr.b_state == arc_anon); 1825 arc_cksum_compute(buf); 1826 mutex_exit(hash_lock); 1827 } 1828 1829 /* 1830 * The arc_buf_hdr_t's b_flags should never be modified directly. Instead, 1831 * the following functions should be used to ensure that the flags are 1832 * updated in a thread-safe way. When manipulating the flags either 1833 * the hash_lock must be held or the hdr must be undiscoverable. This 1834 * ensures that we're not racing with any other threads when updating 1835 * the flags. 1836 */ 1837 static inline void 1838 arc_hdr_set_flags(arc_buf_hdr_t *hdr, arc_flags_t flags) 1839 { 1840 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); 1841 hdr->b_flags |= flags; 1842 } 1843 1844 static inline void 1845 arc_hdr_clear_flags(arc_buf_hdr_t *hdr, arc_flags_t flags) 1846 { 1847 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); 1848 hdr->b_flags &= ~flags; 1849 } 1850 1851 /* 1852 * Setting the compression bits in the arc_buf_hdr_t's b_flags is 1853 * done in a special way since we have to clear and set bits 1854 * at the same time. Consumers that wish to set the compression bits 1855 * must use this function to ensure that the flags are updated in 1856 * thread-safe manner. 1857 */ 1858 static void 1859 arc_hdr_set_compress(arc_buf_hdr_t *hdr, enum zio_compress cmp) 1860 { 1861 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); 1862 1863 /* 1864 * Holes and embedded blocks will always have a psize = 0 so 1865 * we ignore the compression of the blkptr and set the 1866 * arc_buf_hdr_t's compression to ZIO_COMPRESS_OFF. 1867 * Holes and embedded blocks remain anonymous so we don't 1868 * want to uncompress them. Mark them as uncompressed. 1869 */ 1870 if (!zfs_compressed_arc_enabled || HDR_GET_PSIZE(hdr) == 0) { 1871 arc_hdr_clear_flags(hdr, ARC_FLAG_COMPRESSED_ARC); 1872 HDR_SET_COMPRESS(hdr, ZIO_COMPRESS_OFF); 1873 ASSERT(!HDR_COMPRESSION_ENABLED(hdr)); 1874 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF); 1875 } else { 1876 arc_hdr_set_flags(hdr, ARC_FLAG_COMPRESSED_ARC); 1877 HDR_SET_COMPRESS(hdr, cmp); 1878 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, cmp); 1879 ASSERT(HDR_COMPRESSION_ENABLED(hdr)); 1880 } 1881 } 1882 1883 /* 1884 * Looks for another buf on the same hdr which has the data decompressed, copies 1885 * from it, and returns true. If no such buf exists, returns false. 1886 */ 1887 static boolean_t 1888 arc_buf_try_copy_decompressed_data(arc_buf_t *buf) 1889 { 1890 arc_buf_hdr_t *hdr = buf->b_hdr; 1891 boolean_t copied = B_FALSE; 1892 1893 ASSERT(HDR_HAS_L1HDR(hdr)); 1894 ASSERT3P(buf->b_data, !=, NULL); 1895 ASSERT(!ARC_BUF_COMPRESSED(buf)); 1896 1897 for (arc_buf_t *from = hdr->b_l1hdr.b_buf; from != NULL; 1898 from = from->b_next) { 1899 /* can't use our own data buffer */ 1900 if (from == buf) { 1901 continue; 1902 } 1903 1904 if (!ARC_BUF_COMPRESSED(from)) { 1905 bcopy(from->b_data, buf->b_data, arc_buf_size(buf)); 1906 copied = B_TRUE; 1907 break; 1908 } 1909 } 1910 1911 /* 1912 * There were no decompressed bufs, so there should not be a 1913 * checksum on the hdr either. 1914 */ 1915 EQUIV(!copied, hdr->b_l1hdr.b_freeze_cksum == NULL); 1916 1917 return (copied); 1918 } 1919 1920 /* 1921 * Given a buf that has a data buffer attached to it, this function will 1922 * efficiently fill the buf with data of the specified compression setting from 1923 * the hdr and update the hdr's b_freeze_cksum if necessary. If the buf and hdr 1924 * are already sharing a data buf, no copy is performed. 1925 * 1926 * If the buf is marked as compressed but uncompressed data was requested, this 1927 * will allocate a new data buffer for the buf, remove that flag, and fill the 1928 * buf with uncompressed data. You can't request a compressed buf on a hdr with 1929 * uncompressed data, and (since we haven't added support for it yet) if you 1930 * want compressed data your buf must already be marked as compressed and have 1931 * the correct-sized data buffer. 1932 */ 1933 static int 1934 arc_buf_fill(arc_buf_t *buf, boolean_t compressed) 1935 { 1936 arc_buf_hdr_t *hdr = buf->b_hdr; 1937 boolean_t hdr_compressed = (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF); 1938 dmu_object_byteswap_t bswap = hdr->b_l1hdr.b_byteswap; 1939 1940 ASSERT3P(buf->b_data, !=, NULL); 1941 IMPLY(compressed, hdr_compressed); 1942 IMPLY(compressed, ARC_BUF_COMPRESSED(buf)); 1943 1944 if (hdr_compressed == compressed) { 1945 if (!arc_buf_is_shared(buf)) { 1946 abd_copy_to_buf(buf->b_data, hdr->b_l1hdr.b_pabd, 1947 arc_buf_size(buf)); 1948 } 1949 } else { 1950 ASSERT(hdr_compressed); 1951 ASSERT(!compressed); 1952 ASSERT3U(HDR_GET_LSIZE(hdr), !=, HDR_GET_PSIZE(hdr)); 1953 1954 /* 1955 * If the buf is sharing its data with the hdr, unlink it and 1956 * allocate a new data buffer for the buf. 1957 */ 1958 if (arc_buf_is_shared(buf)) { 1959 ASSERT(ARC_BUF_COMPRESSED(buf)); 1960 1961 /* We need to give the buf it's own b_data */ 1962 buf->b_flags &= ~ARC_BUF_FLAG_SHARED; 1963 buf->b_data = 1964 arc_get_data_buf(hdr, HDR_GET_LSIZE(hdr), buf); 1965 arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA); 1966 1967 /* Previously overhead was 0; just add new overhead */ 1968 ARCSTAT_INCR(arcstat_overhead_size, HDR_GET_LSIZE(hdr)); 1969 } else if (ARC_BUF_COMPRESSED(buf)) { 1970 /* We need to reallocate the buf's b_data */ 1971 arc_free_data_buf(hdr, buf->b_data, HDR_GET_PSIZE(hdr), 1972 buf); 1973 buf->b_data = 1974 arc_get_data_buf(hdr, HDR_GET_LSIZE(hdr), buf); 1975 1976 /* We increased the size of b_data; update overhead */ 1977 ARCSTAT_INCR(arcstat_overhead_size, 1978 HDR_GET_LSIZE(hdr) - HDR_GET_PSIZE(hdr)); 1979 } 1980 1981 /* 1982 * Regardless of the buf's previous compression settings, it 1983 * should not be compressed at the end of this function. 1984 */ 1985 buf->b_flags &= ~ARC_BUF_FLAG_COMPRESSED; 1986 1987 /* 1988 * Try copying the data from another buf which already has a 1989 * decompressed version. If that's not possible, it's time to 1990 * bite the bullet and decompress the data from the hdr. 1991 */ 1992 if (arc_buf_try_copy_decompressed_data(buf)) { 1993 /* Skip byteswapping and checksumming (already done) */ 1994 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, !=, NULL); 1995 return (0); 1996 } else { 1997 int error = zio_decompress_data(HDR_GET_COMPRESS(hdr), 1998 hdr->b_l1hdr.b_pabd, buf->b_data, 1999 HDR_GET_PSIZE(hdr), HDR_GET_LSIZE(hdr)); 2000 2001 /* 2002 * Absent hardware errors or software bugs, this should 2003 * be impossible, but log it anyway so we can debug it. 2004 */ 2005 if (error != 0) { 2006 zfs_dbgmsg( 2007 "hdr %p, compress %d, psize %d, lsize %d", 2008 hdr, HDR_GET_COMPRESS(hdr), 2009 HDR_GET_PSIZE(hdr), HDR_GET_LSIZE(hdr)); 2010 return (SET_ERROR(EIO)); 2011 } 2012 } 2013 } 2014 2015 /* Byteswap the buf's data if necessary */ 2016 if (bswap != DMU_BSWAP_NUMFUNCS) { 2017 ASSERT(!HDR_SHARED_DATA(hdr)); 2018 ASSERT3U(bswap, <, DMU_BSWAP_NUMFUNCS); 2019 dmu_ot_byteswap[bswap].ob_func(buf->b_data, HDR_GET_LSIZE(hdr)); 2020 } 2021 2022 /* Compute the hdr's checksum if necessary */ 2023 arc_cksum_compute(buf); 2024 2025 return (0); 2026 } 2027 2028 int 2029 arc_decompress(arc_buf_t *buf) 2030 { 2031 return (arc_buf_fill(buf, B_FALSE)); 2032 } 2033 2034 /* 2035 * Return the size of the block, b_pabd, that is stored in the arc_buf_hdr_t. 2036 */ 2037 static uint64_t 2038 arc_hdr_size(arc_buf_hdr_t *hdr) 2039 { 2040 uint64_t size; 2041 2042 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF && 2043 HDR_GET_PSIZE(hdr) > 0) { 2044 size = HDR_GET_PSIZE(hdr); 2045 } else { 2046 ASSERT3U(HDR_GET_LSIZE(hdr), !=, 0); 2047 size = HDR_GET_LSIZE(hdr); 2048 } 2049 return (size); 2050 } 2051 2052 /* 2053 * Increment the amount of evictable space in the arc_state_t's refcount. 2054 * We account for the space used by the hdr and the arc buf individually 2055 * so that we can add and remove them from the refcount individually. 2056 */ 2057 static void 2058 arc_evictable_space_increment(arc_buf_hdr_t *hdr, arc_state_t *state) 2059 { 2060 arc_buf_contents_t type = arc_buf_type(hdr); 2061 2062 ASSERT(HDR_HAS_L1HDR(hdr)); 2063 2064 if (GHOST_STATE(state)) { 2065 ASSERT0(hdr->b_l1hdr.b_bufcnt); 2066 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 2067 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL); 2068 (void) zfs_refcount_add_many(&state->arcs_esize[type], 2069 HDR_GET_LSIZE(hdr), hdr); 2070 return; 2071 } 2072 2073 ASSERT(!GHOST_STATE(state)); 2074 if (hdr->b_l1hdr.b_pabd != NULL) { 2075 (void) zfs_refcount_add_many(&state->arcs_esize[type], 2076 arc_hdr_size(hdr), hdr); 2077 } 2078 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL; 2079 buf = buf->b_next) { 2080 if (arc_buf_is_shared(buf)) 2081 continue; 2082 (void) zfs_refcount_add_many(&state->arcs_esize[type], 2083 arc_buf_size(buf), buf); 2084 } 2085 } 2086 2087 /* 2088 * Decrement the amount of evictable space in the arc_state_t's refcount. 2089 * We account for the space used by the hdr and the arc buf individually 2090 * so that we can add and remove them from the refcount individually. 2091 */ 2092 static void 2093 arc_evictable_space_decrement(arc_buf_hdr_t *hdr, arc_state_t *state) 2094 { 2095 arc_buf_contents_t type = arc_buf_type(hdr); 2096 2097 ASSERT(HDR_HAS_L1HDR(hdr)); 2098 2099 if (GHOST_STATE(state)) { 2100 ASSERT0(hdr->b_l1hdr.b_bufcnt); 2101 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 2102 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL); 2103 (void) zfs_refcount_remove_many(&state->arcs_esize[type], 2104 HDR_GET_LSIZE(hdr), hdr); 2105 return; 2106 } 2107 2108 ASSERT(!GHOST_STATE(state)); 2109 if (hdr->b_l1hdr.b_pabd != NULL) { 2110 (void) zfs_refcount_remove_many(&state->arcs_esize[type], 2111 arc_hdr_size(hdr), hdr); 2112 } 2113 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL; 2114 buf = buf->b_next) { 2115 if (arc_buf_is_shared(buf)) 2116 continue; 2117 (void) zfs_refcount_remove_many(&state->arcs_esize[type], 2118 arc_buf_size(buf), buf); 2119 } 2120 } 2121 2122 /* 2123 * Add a reference to this hdr indicating that someone is actively 2124 * referencing that memory. When the refcount transitions from 0 to 1, 2125 * we remove it from the respective arc_state_t list to indicate that 2126 * it is not evictable. 2127 */ 2128 static void 2129 add_reference(arc_buf_hdr_t *hdr, void *tag) 2130 { 2131 ASSERT(HDR_HAS_L1HDR(hdr)); 2132 if (!MUTEX_HELD(HDR_LOCK(hdr))) { 2133 ASSERT(hdr->b_l1hdr.b_state == arc_anon); 2134 ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 2135 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 2136 } 2137 2138 arc_state_t *state = hdr->b_l1hdr.b_state; 2139 2140 if ((zfs_refcount_add(&hdr->b_l1hdr.b_refcnt, tag) == 1) && 2141 (state != arc_anon)) { 2142 /* We don't use the L2-only state list. */ 2143 if (state != arc_l2c_only) { 2144 multilist_remove(state->arcs_list[arc_buf_type(hdr)], 2145 hdr); 2146 arc_evictable_space_decrement(hdr, state); 2147 } 2148 /* remove the prefetch flag if we get a reference */ 2149 arc_hdr_clear_flags(hdr, ARC_FLAG_PREFETCH); 2150 } 2151 } 2152 2153 /* 2154 * Remove a reference from this hdr. When the reference transitions from 2155 * 1 to 0 and we're not anonymous, then we add this hdr to the arc_state_t's 2156 * list making it eligible for eviction. 2157 */ 2158 static int 2159 remove_reference(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, void *tag) 2160 { 2161 int cnt; 2162 arc_state_t *state = hdr->b_l1hdr.b_state; 2163 2164 ASSERT(HDR_HAS_L1HDR(hdr)); 2165 ASSERT(state == arc_anon || MUTEX_HELD(hash_lock)); 2166 ASSERT(!GHOST_STATE(state)); 2167 2168 /* 2169 * arc_l2c_only counts as a ghost state so we don't need to explicitly 2170 * check to prevent usage of the arc_l2c_only list. 2171 */ 2172 if (((cnt = zfs_refcount_remove(&hdr->b_l1hdr.b_refcnt, tag)) == 0) && 2173 (state != arc_anon)) { 2174 multilist_insert(state->arcs_list[arc_buf_type(hdr)], hdr); 2175 ASSERT3U(hdr->b_l1hdr.b_bufcnt, >, 0); 2176 arc_evictable_space_increment(hdr, state); 2177 } 2178 return (cnt); 2179 } 2180 2181 /* 2182 * Move the supplied buffer to the indicated state. The hash lock 2183 * for the buffer must be held by the caller. 2184 */ 2185 static void 2186 arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *hdr, 2187 kmutex_t *hash_lock) 2188 { 2189 arc_state_t *old_state; 2190 int64_t refcnt; 2191 uint32_t bufcnt; 2192 boolean_t update_old, update_new; 2193 arc_buf_contents_t buftype = arc_buf_type(hdr); 2194 2195 /* 2196 * We almost always have an L1 hdr here, since we call arc_hdr_realloc() 2197 * in arc_read() when bringing a buffer out of the L2ARC. However, the 2198 * L1 hdr doesn't always exist when we change state to arc_anon before 2199 * destroying a header, in which case reallocating to add the L1 hdr is 2200 * pointless. 2201 */ 2202 if (HDR_HAS_L1HDR(hdr)) { 2203 old_state = hdr->b_l1hdr.b_state; 2204 refcnt = zfs_refcount_count(&hdr->b_l1hdr.b_refcnt); 2205 bufcnt = hdr->b_l1hdr.b_bufcnt; 2206 update_old = (bufcnt > 0 || hdr->b_l1hdr.b_pabd != NULL); 2207 } else { 2208 old_state = arc_l2c_only; 2209 refcnt = 0; 2210 bufcnt = 0; 2211 update_old = B_FALSE; 2212 } 2213 update_new = update_old; 2214 2215 ASSERT(MUTEX_HELD(hash_lock)); 2216 ASSERT3P(new_state, !=, old_state); 2217 ASSERT(!GHOST_STATE(new_state) || bufcnt == 0); 2218 ASSERT(old_state != arc_anon || bufcnt <= 1); 2219 2220 /* 2221 * If this buffer is evictable, transfer it from the 2222 * old state list to the new state list. 2223 */ 2224 if (refcnt == 0) { 2225 if (old_state != arc_anon && old_state != arc_l2c_only) { 2226 ASSERT(HDR_HAS_L1HDR(hdr)); 2227 multilist_remove(old_state->arcs_list[buftype], hdr); 2228 2229 if (GHOST_STATE(old_state)) { 2230 ASSERT0(bufcnt); 2231 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 2232 update_old = B_TRUE; 2233 } 2234 arc_evictable_space_decrement(hdr, old_state); 2235 } 2236 if (new_state != arc_anon && new_state != arc_l2c_only) { 2237 2238 /* 2239 * An L1 header always exists here, since if we're 2240 * moving to some L1-cached state (i.e. not l2c_only or 2241 * anonymous), we realloc the header to add an L1hdr 2242 * beforehand. 2243 */ 2244 ASSERT(HDR_HAS_L1HDR(hdr)); 2245 multilist_insert(new_state->arcs_list[buftype], hdr); 2246 2247 if (GHOST_STATE(new_state)) { 2248 ASSERT0(bufcnt); 2249 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 2250 update_new = B_TRUE; 2251 } 2252 arc_evictable_space_increment(hdr, new_state); 2253 } 2254 } 2255 2256 ASSERT(!HDR_EMPTY(hdr)); 2257 if (new_state == arc_anon && HDR_IN_HASH_TABLE(hdr)) 2258 buf_hash_remove(hdr); 2259 2260 /* adjust state sizes (ignore arc_l2c_only) */ 2261 2262 if (update_new && new_state != arc_l2c_only) { 2263 ASSERT(HDR_HAS_L1HDR(hdr)); 2264 if (GHOST_STATE(new_state)) { 2265 ASSERT0(bufcnt); 2266 2267 /* 2268 * When moving a header to a ghost state, we first 2269 * remove all arc buffers. Thus, we'll have a 2270 * bufcnt of zero, and no arc buffer to use for 2271 * the reference. As a result, we use the arc 2272 * header pointer for the reference. 2273 */ 2274 (void) zfs_refcount_add_many(&new_state->arcs_size, 2275 HDR_GET_LSIZE(hdr), hdr); 2276 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL); 2277 } else { 2278 uint32_t buffers = 0; 2279 2280 /* 2281 * Each individual buffer holds a unique reference, 2282 * thus we must remove each of these references one 2283 * at a time. 2284 */ 2285 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL; 2286 buf = buf->b_next) { 2287 ASSERT3U(bufcnt, !=, 0); 2288 buffers++; 2289 2290 /* 2291 * When the arc_buf_t is sharing the data 2292 * block with the hdr, the owner of the 2293 * reference belongs to the hdr. Only 2294 * add to the refcount if the arc_buf_t is 2295 * not shared. 2296 */ 2297 if (arc_buf_is_shared(buf)) 2298 continue; 2299 2300 (void) zfs_refcount_add_many( 2301 &new_state->arcs_size, 2302 arc_buf_size(buf), buf); 2303 } 2304 ASSERT3U(bufcnt, ==, buffers); 2305 2306 if (hdr->b_l1hdr.b_pabd != NULL) { 2307 (void) zfs_refcount_add_many( 2308 &new_state->arcs_size, 2309 arc_hdr_size(hdr), hdr); 2310 } else { 2311 ASSERT(GHOST_STATE(old_state)); 2312 } 2313 } 2314 } 2315 2316 if (update_old && old_state != arc_l2c_only) { 2317 ASSERT(HDR_HAS_L1HDR(hdr)); 2318 if (GHOST_STATE(old_state)) { 2319 ASSERT0(bufcnt); 2320 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL); 2321 2322 /* 2323 * When moving a header off of a ghost state, 2324 * the header will not contain any arc buffers. 2325 * We use the arc header pointer for the reference 2326 * which is exactly what we did when we put the 2327 * header on the ghost state. 2328 */ 2329 2330 (void) zfs_refcount_remove_many(&old_state->arcs_size, 2331 HDR_GET_LSIZE(hdr), hdr); 2332 } else { 2333 uint32_t buffers = 0; 2334 2335 /* 2336 * Each individual buffer holds a unique reference, 2337 * thus we must remove each of these references one 2338 * at a time. 2339 */ 2340 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL; 2341 buf = buf->b_next) { 2342 ASSERT3U(bufcnt, !=, 0); 2343 buffers++; 2344 2345 /* 2346 * When the arc_buf_t is sharing the data 2347 * block with the hdr, the owner of the 2348 * reference belongs to the hdr. Only 2349 * add to the refcount if the arc_buf_t is 2350 * not shared. 2351 */ 2352 if (arc_buf_is_shared(buf)) 2353 continue; 2354 2355 (void) zfs_refcount_remove_many( 2356 &old_state->arcs_size, arc_buf_size(buf), 2357 buf); 2358 } 2359 ASSERT3U(bufcnt, ==, buffers); 2360 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL); 2361 (void) zfs_refcount_remove_many( 2362 &old_state->arcs_size, arc_hdr_size(hdr), hdr); 2363 } 2364 } 2365 2366 if (HDR_HAS_L1HDR(hdr)) 2367 hdr->b_l1hdr.b_state = new_state; 2368 2369 /* 2370 * L2 headers should never be on the L2 state list since they don't 2371 * have L1 headers allocated. 2372 */ 2373 ASSERT(multilist_is_empty(arc_l2c_only->arcs_list[ARC_BUFC_DATA]) && 2374 multilist_is_empty(arc_l2c_only->arcs_list[ARC_BUFC_METADATA])); 2375 } 2376 2377 void 2378 arc_space_consume(uint64_t space, arc_space_type_t type) 2379 { 2380 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES); 2381 2382 switch (type) { 2383 case ARC_SPACE_DATA: 2384 aggsum_add(&astat_data_size, space); 2385 break; 2386 case ARC_SPACE_META: 2387 aggsum_add(&astat_metadata_size, space); 2388 break; 2389 case ARC_SPACE_OTHER: 2390 aggsum_add(&astat_other_size, space); 2391 break; 2392 case ARC_SPACE_HDRS: 2393 aggsum_add(&astat_hdr_size, space); 2394 break; 2395 case ARC_SPACE_L2HDRS: 2396 aggsum_add(&astat_l2_hdr_size, space); 2397 break; 2398 } 2399 2400 if (type != ARC_SPACE_DATA) 2401 aggsum_add(&arc_meta_used, space); 2402 2403 aggsum_add(&arc_size, space); 2404 } 2405 2406 void 2407 arc_space_return(uint64_t space, arc_space_type_t type) 2408 { 2409 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES); 2410 2411 switch (type) { 2412 case ARC_SPACE_DATA: 2413 aggsum_add(&astat_data_size, -space); 2414 break; 2415 case ARC_SPACE_META: 2416 aggsum_add(&astat_metadata_size, -space); 2417 break; 2418 case ARC_SPACE_OTHER: 2419 aggsum_add(&astat_other_size, -space); 2420 break; 2421 case ARC_SPACE_HDRS: 2422 aggsum_add(&astat_hdr_size, -space); 2423 break; 2424 case ARC_SPACE_L2HDRS: 2425 aggsum_add(&astat_l2_hdr_size, -space); 2426 break; 2427 } 2428 2429 if (type != ARC_SPACE_DATA) { 2430 ASSERT(aggsum_compare(&arc_meta_used, space) >= 0); 2431 /* 2432 * We use the upper bound here rather than the precise value 2433 * because the arc_meta_max value doesn't need to be 2434 * precise. It's only consumed by humans via arcstats. 2435 */ 2436 if (arc_meta_max < aggsum_upper_bound(&arc_meta_used)) 2437 arc_meta_max = aggsum_upper_bound(&arc_meta_used); 2438 aggsum_add(&arc_meta_used, -space); 2439 } 2440 2441 ASSERT(aggsum_compare(&arc_size, space) >= 0); 2442 aggsum_add(&arc_size, -space); 2443 } 2444 2445 /* 2446 * Given a hdr and a buf, returns whether that buf can share its b_data buffer 2447 * with the hdr's b_pabd. 2448 */ 2449 static boolean_t 2450 arc_can_share(arc_buf_hdr_t *hdr, arc_buf_t *buf) 2451 { 2452 /* 2453 * The criteria for sharing a hdr's data are: 2454 * 1. the hdr's compression matches the buf's compression 2455 * 2. the hdr doesn't need to be byteswapped 2456 * 3. the hdr isn't already being shared 2457 * 4. the buf is either compressed or it is the last buf in the hdr list 2458 * 2459 * Criterion #4 maintains the invariant that shared uncompressed 2460 * bufs must be the final buf in the hdr's b_buf list. Reading this, you 2461 * might ask, "if a compressed buf is allocated first, won't that be the 2462 * last thing in the list?", but in that case it's impossible to create 2463 * a shared uncompressed buf anyway (because the hdr must be compressed 2464 * to have the compressed buf). You might also think that #3 is 2465 * sufficient to make this guarantee, however it's possible 2466 * (specifically in the rare L2ARC write race mentioned in 2467 * arc_buf_alloc_impl()) there will be an existing uncompressed buf that 2468 * is sharable, but wasn't at the time of its allocation. Rather than 2469 * allow a new shared uncompressed buf to be created and then shuffle 2470 * the list around to make it the last element, this simply disallows 2471 * sharing if the new buf isn't the first to be added. 2472 */ 2473 ASSERT3P(buf->b_hdr, ==, hdr); 2474 boolean_t hdr_compressed = HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF; 2475 boolean_t buf_compressed = ARC_BUF_COMPRESSED(buf) != 0; 2476 return (buf_compressed == hdr_compressed && 2477 hdr->b_l1hdr.b_byteswap == DMU_BSWAP_NUMFUNCS && 2478 !HDR_SHARED_DATA(hdr) && 2479 (ARC_BUF_LAST(buf) || ARC_BUF_COMPRESSED(buf))); 2480 } 2481 2482 /* 2483 * Allocate a buf for this hdr. If you care about the data that's in the hdr, 2484 * or if you want a compressed buffer, pass those flags in. Returns 0 if the 2485 * copy was made successfully, or an error code otherwise. 2486 */ 2487 static int 2488 arc_buf_alloc_impl(arc_buf_hdr_t *hdr, void *tag, boolean_t compressed, 2489 boolean_t fill, arc_buf_t **ret) 2490 { 2491 arc_buf_t *buf; 2492 2493 ASSERT(HDR_HAS_L1HDR(hdr)); 2494 ASSERT3U(HDR_GET_LSIZE(hdr), >, 0); 2495 VERIFY(hdr->b_type == ARC_BUFC_DATA || 2496 hdr->b_type == ARC_BUFC_METADATA); 2497 ASSERT3P(ret, !=, NULL); 2498 ASSERT3P(*ret, ==, NULL); 2499 2500 buf = *ret = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 2501 buf->b_hdr = hdr; 2502 buf->b_data = NULL; 2503 buf->b_next = hdr->b_l1hdr.b_buf; 2504 buf->b_flags = 0; 2505 2506 add_reference(hdr, tag); 2507 2508 /* 2509 * We're about to change the hdr's b_flags. We must either 2510 * hold the hash_lock or be undiscoverable. 2511 */ 2512 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); 2513 2514 /* 2515 * Only honor requests for compressed bufs if the hdr is actually 2516 * compressed. 2517 */ 2518 if (compressed && HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF) 2519 buf->b_flags |= ARC_BUF_FLAG_COMPRESSED; 2520 2521 /* 2522 * If the hdr's data can be shared then we share the data buffer and 2523 * set the appropriate bit in the hdr's b_flags to indicate the hdr is 2524 * sharing it's b_pabd with the arc_buf_t. Otherwise, we allocate a new 2525 * buffer to store the buf's data. 2526 * 2527 * There are two additional restrictions here because we're sharing 2528 * hdr -> buf instead of the usual buf -> hdr. First, the hdr can't be 2529 * actively involved in an L2ARC write, because if this buf is used by 2530 * an arc_write() then the hdr's data buffer will be released when the 2531 * write completes, even though the L2ARC write might still be using it. 2532 * Second, the hdr's ABD must be linear so that the buf's user doesn't 2533 * need to be ABD-aware. 2534 */ 2535 boolean_t can_share = arc_can_share(hdr, buf) && !HDR_L2_WRITING(hdr) && 2536 abd_is_linear(hdr->b_l1hdr.b_pabd); 2537 2538 /* Set up b_data and sharing */ 2539 if (can_share) { 2540 buf->b_data = abd_to_buf(hdr->b_l1hdr.b_pabd); 2541 buf->b_flags |= ARC_BUF_FLAG_SHARED; 2542 arc_hdr_set_flags(hdr, ARC_FLAG_SHARED_DATA); 2543 } else { 2544 buf->b_data = 2545 arc_get_data_buf(hdr, arc_buf_size(buf), buf); 2546 ARCSTAT_INCR(arcstat_overhead_size, arc_buf_size(buf)); 2547 } 2548 VERIFY3P(buf->b_data, !=, NULL); 2549 2550 hdr->b_l1hdr.b_buf = buf; 2551 hdr->b_l1hdr.b_bufcnt += 1; 2552 2553 /* 2554 * If the user wants the data from the hdr, we need to either copy or 2555 * decompress the data. 2556 */ 2557 if (fill) { 2558 return (arc_buf_fill(buf, ARC_BUF_COMPRESSED(buf) != 0)); 2559 } 2560 2561 return (0); 2562 } 2563 2564 static char *arc_onloan_tag = "onloan"; 2565 2566 static inline void 2567 arc_loaned_bytes_update(int64_t delta) 2568 { 2569 atomic_add_64(&arc_loaned_bytes, delta); 2570 2571 /* assert that it did not wrap around */ 2572 ASSERT3S(atomic_add_64_nv(&arc_loaned_bytes, 0), >=, 0); 2573 } 2574 2575 /* 2576 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in 2577 * flight data by arc_tempreserve_space() until they are "returned". Loaned 2578 * buffers must be returned to the arc before they can be used by the DMU or 2579 * freed. 2580 */ 2581 arc_buf_t * 2582 arc_loan_buf(spa_t *spa, boolean_t is_metadata, int size) 2583 { 2584 arc_buf_t *buf = arc_alloc_buf(spa, arc_onloan_tag, 2585 is_metadata ? ARC_BUFC_METADATA : ARC_BUFC_DATA, size); 2586 2587 arc_loaned_bytes_update(arc_buf_size(buf)); 2588 2589 return (buf); 2590 } 2591 2592 arc_buf_t * 2593 arc_loan_compressed_buf(spa_t *spa, uint64_t psize, uint64_t lsize, 2594 enum zio_compress compression_type) 2595 { 2596 arc_buf_t *buf = arc_alloc_compressed_buf(spa, arc_onloan_tag, 2597 psize, lsize, compression_type); 2598 2599 arc_loaned_bytes_update(arc_buf_size(buf)); 2600 2601 return (buf); 2602 } 2603 2604 2605 /* 2606 * Return a loaned arc buffer to the arc. 2607 */ 2608 void 2609 arc_return_buf(arc_buf_t *buf, void *tag) 2610 { 2611 arc_buf_hdr_t *hdr = buf->b_hdr; 2612 2613 ASSERT3P(buf->b_data, !=, NULL); 2614 ASSERT(HDR_HAS_L1HDR(hdr)); 2615 (void) zfs_refcount_add(&hdr->b_l1hdr.b_refcnt, tag); 2616 (void) zfs_refcount_remove(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag); 2617 2618 arc_loaned_bytes_update(-arc_buf_size(buf)); 2619 } 2620 2621 /* Detach an arc_buf from a dbuf (tag) */ 2622 void 2623 arc_loan_inuse_buf(arc_buf_t *buf, void *tag) 2624 { 2625 arc_buf_hdr_t *hdr = buf->b_hdr; 2626 2627 ASSERT3P(buf->b_data, !=, NULL); 2628 ASSERT(HDR_HAS_L1HDR(hdr)); 2629 (void) zfs_refcount_add(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag); 2630 (void) zfs_refcount_remove(&hdr->b_l1hdr.b_refcnt, tag); 2631 2632 arc_loaned_bytes_update(arc_buf_size(buf)); 2633 } 2634 2635 static void 2636 l2arc_free_abd_on_write(abd_t *abd, size_t size, arc_buf_contents_t type) 2637 { 2638 l2arc_data_free_t *df = kmem_alloc(sizeof (*df), KM_SLEEP); 2639 2640 df->l2df_abd = abd; 2641 df->l2df_size = size; 2642 df->l2df_type = type; 2643 mutex_enter(&l2arc_free_on_write_mtx); 2644 list_insert_head(l2arc_free_on_write, df); 2645 mutex_exit(&l2arc_free_on_write_mtx); 2646 } 2647 2648 static void 2649 arc_hdr_free_on_write(arc_buf_hdr_t *hdr) 2650 { 2651 arc_state_t *state = hdr->b_l1hdr.b_state; 2652 arc_buf_contents_t type = arc_buf_type(hdr); 2653 uint64_t size = arc_hdr_size(hdr); 2654 2655 /* protected by hash lock, if in the hash table */ 2656 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) { 2657 ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 2658 ASSERT(state != arc_anon && state != arc_l2c_only); 2659 2660 (void) zfs_refcount_remove_many(&state->arcs_esize[type], 2661 size, hdr); 2662 } 2663 (void) zfs_refcount_remove_many(&state->arcs_size, size, hdr); 2664 if (type == ARC_BUFC_METADATA) { 2665 arc_space_return(size, ARC_SPACE_META); 2666 } else { 2667 ASSERT(type == ARC_BUFC_DATA); 2668 arc_space_return(size, ARC_SPACE_DATA); 2669 } 2670 2671 l2arc_free_abd_on_write(hdr->b_l1hdr.b_pabd, size, type); 2672 } 2673 2674 /* 2675 * Share the arc_buf_t's data with the hdr. Whenever we are sharing the 2676 * data buffer, we transfer the refcount ownership to the hdr and update 2677 * the appropriate kstats. 2678 */ 2679 static void 2680 arc_share_buf(arc_buf_hdr_t *hdr, arc_buf_t *buf) 2681 { 2682 arc_state_t *state = hdr->b_l1hdr.b_state; 2683 2684 ASSERT(arc_can_share(hdr, buf)); 2685 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL); 2686 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); 2687 2688 /* 2689 * Start sharing the data buffer. We transfer the 2690 * refcount ownership to the hdr since it always owns 2691 * the refcount whenever an arc_buf_t is shared. 2692 */ 2693 zfs_refcount_transfer_ownership(&state->arcs_size, buf, hdr); 2694 hdr->b_l1hdr.b_pabd = abd_get_from_buf(buf->b_data, arc_buf_size(buf)); 2695 abd_take_ownership_of_buf(hdr->b_l1hdr.b_pabd, 2696 HDR_ISTYPE_METADATA(hdr)); 2697 arc_hdr_set_flags(hdr, ARC_FLAG_SHARED_DATA); 2698 buf->b_flags |= ARC_BUF_FLAG_SHARED; 2699 2700 /* 2701 * Since we've transferred ownership to the hdr we need 2702 * to increment its compressed and uncompressed kstats and 2703 * decrement the overhead size. 2704 */ 2705 ARCSTAT_INCR(arcstat_compressed_size, arc_hdr_size(hdr)); 2706 ARCSTAT_INCR(arcstat_uncompressed_size, HDR_GET_LSIZE(hdr)); 2707 ARCSTAT_INCR(arcstat_overhead_size, -arc_buf_size(buf)); 2708 } 2709 2710 static void 2711 arc_unshare_buf(arc_buf_hdr_t *hdr, arc_buf_t *buf) 2712 { 2713 arc_state_t *state = hdr->b_l1hdr.b_state; 2714 2715 ASSERT(arc_buf_is_shared(buf)); 2716 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL); 2717 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); 2718 2719 /* 2720 * We are no longer sharing this buffer so we need 2721 * to transfer its ownership to the rightful owner. 2722 */ 2723 zfs_refcount_transfer_ownership(&state->arcs_size, hdr, buf); 2724 arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA); 2725 abd_release_ownership_of_buf(hdr->b_l1hdr.b_pabd); 2726 abd_put(hdr->b_l1hdr.b_pabd); 2727 hdr->b_l1hdr.b_pabd = NULL; 2728 buf->b_flags &= ~ARC_BUF_FLAG_SHARED; 2729 2730 /* 2731 * Since the buffer is no longer shared between 2732 * the arc buf and the hdr, count it as overhead. 2733 */ 2734 ARCSTAT_INCR(arcstat_compressed_size, -arc_hdr_size(hdr)); 2735 ARCSTAT_INCR(arcstat_uncompressed_size, -HDR_GET_LSIZE(hdr)); 2736 ARCSTAT_INCR(arcstat_overhead_size, arc_buf_size(buf)); 2737 } 2738 2739 /* 2740 * Remove an arc_buf_t from the hdr's buf list and return the last 2741 * arc_buf_t on the list. If no buffers remain on the list then return 2742 * NULL. 2743 */ 2744 static arc_buf_t * 2745 arc_buf_remove(arc_buf_hdr_t *hdr, arc_buf_t *buf) 2746 { 2747 ASSERT(HDR_HAS_L1HDR(hdr)); 2748 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); 2749 2750 arc_buf_t **bufp = &hdr->b_l1hdr.b_buf; 2751 arc_buf_t *lastbuf = NULL; 2752 2753 /* 2754 * Remove the buf from the hdr list and locate the last 2755 * remaining buffer on the list. 2756 */ 2757 while (*bufp != NULL) { 2758 if (*bufp == buf) 2759 *bufp = buf->b_next; 2760 2761 /* 2762 * If we've removed a buffer in the middle of 2763 * the list then update the lastbuf and update 2764 * bufp. 2765 */ 2766 if (*bufp != NULL) { 2767 lastbuf = *bufp; 2768 bufp = &(*bufp)->b_next; 2769 } 2770 } 2771 buf->b_next = NULL; 2772 ASSERT3P(lastbuf, !=, buf); 2773 IMPLY(hdr->b_l1hdr.b_bufcnt > 0, lastbuf != NULL); 2774 IMPLY(hdr->b_l1hdr.b_bufcnt > 0, hdr->b_l1hdr.b_buf != NULL); 2775 IMPLY(lastbuf != NULL, ARC_BUF_LAST(lastbuf)); 2776 2777 return (lastbuf); 2778 } 2779 2780 /* 2781 * Free up buf->b_data and pull the arc_buf_t off of the the arc_buf_hdr_t's 2782 * list and free it. 2783 */ 2784 static void 2785 arc_buf_destroy_impl(arc_buf_t *buf) 2786 { 2787 arc_buf_hdr_t *hdr = buf->b_hdr; 2788 2789 /* 2790 * Free up the data associated with the buf but only if we're not 2791 * sharing this with the hdr. If we are sharing it with the hdr, the 2792 * hdr is responsible for doing the free. 2793 */ 2794 if (buf->b_data != NULL) { 2795 /* 2796 * We're about to change the hdr's b_flags. We must either 2797 * hold the hash_lock or be undiscoverable. 2798 */ 2799 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); 2800 2801 arc_cksum_verify(buf); 2802 arc_buf_unwatch(buf); 2803 2804 if (arc_buf_is_shared(buf)) { 2805 arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA); 2806 } else { 2807 uint64_t size = arc_buf_size(buf); 2808 arc_free_data_buf(hdr, buf->b_data, size, buf); 2809 ARCSTAT_INCR(arcstat_overhead_size, -size); 2810 } 2811 buf->b_data = NULL; 2812 2813 ASSERT(hdr->b_l1hdr.b_bufcnt > 0); 2814 hdr->b_l1hdr.b_bufcnt -= 1; 2815 } 2816 2817 arc_buf_t *lastbuf = arc_buf_remove(hdr, buf); 2818 2819 if (ARC_BUF_SHARED(buf) && !ARC_BUF_COMPRESSED(buf)) { 2820 /* 2821 * If the current arc_buf_t is sharing its data buffer with the 2822 * hdr, then reassign the hdr's b_pabd to share it with the new 2823 * buffer at the end of the list. The shared buffer is always 2824 * the last one on the hdr's buffer list. 2825 * 2826 * There is an equivalent case for compressed bufs, but since 2827 * they aren't guaranteed to be the last buf in the list and 2828 * that is an exceedingly rare case, we just allow that space be 2829 * wasted temporarily. 2830 */ 2831 if (lastbuf != NULL) { 2832 /* Only one buf can be shared at once */ 2833 VERIFY(!arc_buf_is_shared(lastbuf)); 2834 /* hdr is uncompressed so can't have compressed buf */ 2835 VERIFY(!ARC_BUF_COMPRESSED(lastbuf)); 2836 2837 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL); 2838 arc_hdr_free_pabd(hdr); 2839 2840 /* 2841 * We must setup a new shared block between the 2842 * last buffer and the hdr. The data would have 2843 * been allocated by the arc buf so we need to transfer 2844 * ownership to the hdr since it's now being shared. 2845 */ 2846 arc_share_buf(hdr, lastbuf); 2847 } 2848 } else if (HDR_SHARED_DATA(hdr)) { 2849 /* 2850 * Uncompressed shared buffers are always at the end 2851 * of the list. Compressed buffers don't have the 2852 * same requirements. This makes it hard to 2853 * simply assert that the lastbuf is shared so 2854 * we rely on the hdr's compression flags to determine 2855 * if we have a compressed, shared buffer. 2856 */ 2857 ASSERT3P(lastbuf, !=, NULL); 2858 ASSERT(arc_buf_is_shared(lastbuf) || 2859 HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF); 2860 } 2861 2862 /* 2863 * Free the checksum if we're removing the last uncompressed buf from 2864 * this hdr. 2865 */ 2866 if (!arc_hdr_has_uncompressed_buf(hdr)) { 2867 arc_cksum_free(hdr); 2868 } 2869 2870 /* clean up the buf */ 2871 buf->b_hdr = NULL; 2872 kmem_cache_free(buf_cache, buf); 2873 } 2874 2875 static void 2876 arc_hdr_alloc_pabd(arc_buf_hdr_t *hdr) 2877 { 2878 ASSERT3U(HDR_GET_LSIZE(hdr), >, 0); 2879 ASSERT(HDR_HAS_L1HDR(hdr)); 2880 ASSERT(!HDR_SHARED_DATA(hdr)); 2881 2882 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL); 2883 hdr->b_l1hdr.b_pabd = arc_get_data_abd(hdr, arc_hdr_size(hdr), hdr); 2884 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS; 2885 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL); 2886 2887 ARCSTAT_INCR(arcstat_compressed_size, arc_hdr_size(hdr)); 2888 ARCSTAT_INCR(arcstat_uncompressed_size, HDR_GET_LSIZE(hdr)); 2889 } 2890 2891 static void 2892 arc_hdr_free_pabd(arc_buf_hdr_t *hdr) 2893 { 2894 ASSERT(HDR_HAS_L1HDR(hdr)); 2895 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL); 2896 2897 /* 2898 * If the hdr is currently being written to the l2arc then 2899 * we defer freeing the data by adding it to the l2arc_free_on_write 2900 * list. The l2arc will free the data once it's finished 2901 * writing it to the l2arc device. 2902 */ 2903 if (HDR_L2_WRITING(hdr)) { 2904 arc_hdr_free_on_write(hdr); 2905 ARCSTAT_BUMP(arcstat_l2_free_on_write); 2906 } else { 2907 arc_free_data_abd(hdr, hdr->b_l1hdr.b_pabd, 2908 arc_hdr_size(hdr), hdr); 2909 } 2910 hdr->b_l1hdr.b_pabd = NULL; 2911 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS; 2912 2913 ARCSTAT_INCR(arcstat_compressed_size, -arc_hdr_size(hdr)); 2914 ARCSTAT_INCR(arcstat_uncompressed_size, -HDR_GET_LSIZE(hdr)); 2915 } 2916 2917 static arc_buf_hdr_t * 2918 arc_hdr_alloc(uint64_t spa, int32_t psize, int32_t lsize, 2919 enum zio_compress compression_type, arc_buf_contents_t type) 2920 { 2921 arc_buf_hdr_t *hdr; 2922 2923 VERIFY(type == ARC_BUFC_DATA || type == ARC_BUFC_METADATA); 2924 2925 hdr = kmem_cache_alloc(hdr_full_cache, KM_PUSHPAGE); 2926 ASSERT(HDR_EMPTY(hdr)); 2927 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL); 2928 ASSERT3P(hdr->b_l1hdr.b_thawed, ==, NULL); 2929 HDR_SET_PSIZE(hdr, psize); 2930 HDR_SET_LSIZE(hdr, lsize); 2931 hdr->b_spa = spa; 2932 hdr->b_type = type; 2933 hdr->b_flags = 0; 2934 arc_hdr_set_flags(hdr, arc_bufc_to_flags(type) | ARC_FLAG_HAS_L1HDR); 2935 arc_hdr_set_compress(hdr, compression_type); 2936 2937 hdr->b_l1hdr.b_state = arc_anon; 2938 hdr->b_l1hdr.b_arc_access = 0; 2939 hdr->b_l1hdr.b_bufcnt = 0; 2940 hdr->b_l1hdr.b_buf = NULL; 2941 2942 /* 2943 * Allocate the hdr's buffer. This will contain either 2944 * the compressed or uncompressed data depending on the block 2945 * it references and compressed arc enablement. 2946 */ 2947 arc_hdr_alloc_pabd(hdr); 2948 ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 2949 2950 return (hdr); 2951 } 2952 2953 /* 2954 * Transition between the two allocation states for the arc_buf_hdr struct. 2955 * The arc_buf_hdr struct can be allocated with (hdr_full_cache) or without 2956 * (hdr_l2only_cache) the fields necessary for the L1 cache - the smaller 2957 * version is used when a cache buffer is only in the L2ARC in order to reduce 2958 * memory usage. 2959 */ 2960 static arc_buf_hdr_t * 2961 arc_hdr_realloc(arc_buf_hdr_t *hdr, kmem_cache_t *old, kmem_cache_t *new) 2962 { 2963 ASSERT(HDR_HAS_L2HDR(hdr)); 2964 2965 arc_buf_hdr_t *nhdr; 2966 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev; 2967 2968 ASSERT((old == hdr_full_cache && new == hdr_l2only_cache) || 2969 (old == hdr_l2only_cache && new == hdr_full_cache)); 2970 2971 nhdr = kmem_cache_alloc(new, KM_PUSHPAGE); 2972 2973 ASSERT(MUTEX_HELD(HDR_LOCK(hdr))); 2974 buf_hash_remove(hdr); 2975 2976 bcopy(hdr, nhdr, HDR_L2ONLY_SIZE); 2977 2978 if (new == hdr_full_cache) { 2979 arc_hdr_set_flags(nhdr, ARC_FLAG_HAS_L1HDR); 2980 /* 2981 * arc_access and arc_change_state need to be aware that a 2982 * header has just come out of L2ARC, so we set its state to 2983 * l2c_only even though it's about to change. 2984 */ 2985 nhdr->b_l1hdr.b_state = arc_l2c_only; 2986 2987 /* Verify previous threads set to NULL before freeing */ 2988 ASSERT3P(nhdr->b_l1hdr.b_pabd, ==, NULL); 2989 } else { 2990 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 2991 ASSERT0(hdr->b_l1hdr.b_bufcnt); 2992 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL); 2993 2994 /* 2995 * If we've reached here, We must have been called from 2996 * arc_evict_hdr(), as such we should have already been 2997 * removed from any ghost list we were previously on 2998 * (which protects us from racing with arc_evict_state), 2999 * thus no locking is needed during this check. 3000 */ 3001 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 3002 3003 /* 3004 * A buffer must not be moved into the arc_l2c_only 3005 * state if it's not finished being written out to the 3006 * l2arc device. Otherwise, the b_l1hdr.b_pabd field 3007 * might try to be accessed, even though it was removed. 3008 */ 3009 VERIFY(!HDR_L2_WRITING(hdr)); 3010 VERIFY3P(hdr->b_l1hdr.b_pabd, ==, NULL); 3011 3012 #ifdef ZFS_DEBUG 3013 if (hdr->b_l1hdr.b_thawed != NULL) { 3014 kmem_free(hdr->b_l1hdr.b_thawed, 1); 3015 hdr->b_l1hdr.b_thawed = NULL; 3016 } 3017 #endif 3018 3019 arc_hdr_clear_flags(nhdr, ARC_FLAG_HAS_L1HDR); 3020 } 3021 /* 3022 * The header has been reallocated so we need to re-insert it into any 3023 * lists it was on. 3024 */ 3025 (void) buf_hash_insert(nhdr, NULL); 3026 3027 ASSERT(list_link_active(&hdr->b_l2hdr.b_l2node)); 3028 3029 mutex_enter(&dev->l2ad_mtx); 3030 3031 /* 3032 * We must place the realloc'ed header back into the list at 3033 * the same spot. Otherwise, if it's placed earlier in the list, 3034 * l2arc_write_buffers() could find it during the function's 3035 * write phase, and try to write it out to the l2arc. 3036 */ 3037 list_insert_after(&dev->l2ad_buflist, hdr, nhdr); 3038 list_remove(&dev->l2ad_buflist, hdr); 3039 3040 mutex_exit(&dev->l2ad_mtx); 3041 3042 /* 3043 * Since we're using the pointer address as the tag when 3044 * incrementing and decrementing the l2ad_alloc refcount, we 3045 * must remove the old pointer (that we're about to destroy) and 3046 * add the new pointer to the refcount. Otherwise we'd remove 3047 * the wrong pointer address when calling arc_hdr_destroy() later. 3048 */ 3049 3050 (void) zfs_refcount_remove_many(&dev->l2ad_alloc, arc_hdr_size(hdr), 3051 hdr); 3052 (void) zfs_refcount_add_many(&dev->l2ad_alloc, arc_hdr_size(nhdr), 3053 nhdr); 3054 3055 buf_discard_identity(hdr); 3056 kmem_cache_free(old, hdr); 3057 3058 return (nhdr); 3059 } 3060 3061 /* 3062 * Allocate a new arc_buf_hdr_t and arc_buf_t and return the buf to the caller. 3063 * The buf is returned thawed since we expect the consumer to modify it. 3064 */ 3065 arc_buf_t * 3066 arc_alloc_buf(spa_t *spa, void *tag, arc_buf_contents_t type, int32_t size) 3067 { 3068 arc_buf_hdr_t *hdr = arc_hdr_alloc(spa_load_guid(spa), size, size, 3069 ZIO_COMPRESS_OFF, type); 3070 ASSERT(!MUTEX_HELD(HDR_LOCK(hdr))); 3071 3072 arc_buf_t *buf = NULL; 3073 VERIFY0(arc_buf_alloc_impl(hdr, tag, B_FALSE, B_FALSE, &buf)); 3074 arc_buf_thaw(buf); 3075 3076 return (buf); 3077 } 3078 3079 /* 3080 * Allocate a compressed buf in the same manner as arc_alloc_buf. Don't use this 3081 * for bufs containing metadata. 3082 */ 3083 arc_buf_t * 3084 arc_alloc_compressed_buf(spa_t *spa, void *tag, uint64_t psize, uint64_t lsize, 3085 enum zio_compress compression_type) 3086 { 3087 ASSERT3U(lsize, >, 0); 3088 ASSERT3U(lsize, >=, psize); 3089 ASSERT(compression_type > ZIO_COMPRESS_OFF); 3090 ASSERT(compression_type < ZIO_COMPRESS_FUNCTIONS); 3091 3092 arc_buf_hdr_t *hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize, 3093 compression_type, ARC_BUFC_DATA); 3094 ASSERT(!MUTEX_HELD(HDR_LOCK(hdr))); 3095 3096 arc_buf_t *buf = NULL; 3097 VERIFY0(arc_buf_alloc_impl(hdr, tag, B_TRUE, B_FALSE, &buf)); 3098 arc_buf_thaw(buf); 3099 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL); 3100 3101 if (!arc_buf_is_shared(buf)) { 3102 /* 3103 * To ensure that the hdr has the correct data in it if we call 3104 * arc_decompress() on this buf before it's been written to 3105 * disk, it's easiest if we just set up sharing between the 3106 * buf and the hdr. 3107 */ 3108 ASSERT(!abd_is_linear(hdr->b_l1hdr.b_pabd)); 3109 arc_hdr_free_pabd(hdr); 3110 arc_share_buf(hdr, buf); 3111 } 3112 3113 return (buf); 3114 } 3115 3116 static void 3117 arc_hdr_l2hdr_destroy(arc_buf_hdr_t *hdr) 3118 { 3119 l2arc_buf_hdr_t *l2hdr = &hdr->b_l2hdr; 3120 l2arc_dev_t *dev = l2hdr->b_dev; 3121 uint64_t psize = arc_hdr_size(hdr); 3122 3123 ASSERT(MUTEX_HELD(&dev->l2ad_mtx)); 3124 ASSERT(HDR_HAS_L2HDR(hdr)); 3125 3126 list_remove(&dev->l2ad_buflist, hdr); 3127 3128 ARCSTAT_INCR(arcstat_l2_psize, -psize); 3129 ARCSTAT_INCR(arcstat_l2_lsize, -HDR_GET_LSIZE(hdr)); 3130 3131 vdev_space_update(dev->l2ad_vdev, -psize, 0, 0); 3132 3133 (void) zfs_refcount_remove_many(&dev->l2ad_alloc, psize, hdr); 3134 arc_hdr_clear_flags(hdr, ARC_FLAG_HAS_L2HDR); 3135 } 3136 3137 static void 3138 arc_hdr_destroy(arc_buf_hdr_t *hdr) 3139 { 3140 if (HDR_HAS_L1HDR(hdr)) { 3141 ASSERT(hdr->b_l1hdr.b_buf == NULL || 3142 hdr->b_l1hdr.b_bufcnt > 0); 3143 ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 3144 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon); 3145 } 3146 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 3147 ASSERT(!HDR_IN_HASH_TABLE(hdr)); 3148 3149 if (!HDR_EMPTY(hdr)) 3150 buf_discard_identity(hdr); 3151 3152 if (HDR_HAS_L2HDR(hdr)) { 3153 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev; 3154 boolean_t buflist_held = MUTEX_HELD(&dev->l2ad_mtx); 3155 3156 if (!buflist_held) 3157 mutex_enter(&dev->l2ad_mtx); 3158 3159 /* 3160 * Even though we checked this conditional above, we 3161 * need to check this again now that we have the 3162 * l2ad_mtx. This is because we could be racing with 3163 * another thread calling l2arc_evict() which might have 3164 * destroyed this header's L2 portion as we were waiting 3165 * to acquire the l2ad_mtx. If that happens, we don't 3166 * want to re-destroy the header's L2 portion. 3167 */ 3168 if (HDR_HAS_L2HDR(hdr)) 3169 arc_hdr_l2hdr_destroy(hdr); 3170 3171 if (!buflist_held) 3172 mutex_exit(&dev->l2ad_mtx); 3173 } 3174 3175 if (HDR_HAS_L1HDR(hdr)) { 3176 arc_cksum_free(hdr); 3177 3178 while (hdr->b_l1hdr.b_buf != NULL) 3179 arc_buf_destroy_impl(hdr->b_l1hdr.b_buf); 3180 3181 #ifdef ZFS_DEBUG 3182 if (hdr->b_l1hdr.b_thawed != NULL) { 3183 kmem_free(hdr->b_l1hdr.b_thawed, 1); 3184 hdr->b_l1hdr.b_thawed = NULL; 3185 } 3186 #endif 3187 3188 if (hdr->b_l1hdr.b_pabd != NULL) { 3189 arc_hdr_free_pabd(hdr); 3190 } 3191 } 3192 3193 ASSERT3P(hdr->b_hash_next, ==, NULL); 3194 if (HDR_HAS_L1HDR(hdr)) { 3195 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 3196 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL); 3197 kmem_cache_free(hdr_full_cache, hdr); 3198 } else { 3199 kmem_cache_free(hdr_l2only_cache, hdr); 3200 } 3201 } 3202 3203 void 3204 arc_buf_destroy(arc_buf_t *buf, void* tag) 3205 { 3206 arc_buf_hdr_t *hdr = buf->b_hdr; 3207 kmutex_t *hash_lock = HDR_LOCK(hdr); 3208 3209 if (hdr->b_l1hdr.b_state == arc_anon) { 3210 ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1); 3211 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 3212 VERIFY0(remove_reference(hdr, NULL, tag)); 3213 arc_hdr_destroy(hdr); 3214 return; 3215 } 3216 3217 mutex_enter(hash_lock); 3218 ASSERT3P(hdr, ==, buf->b_hdr); 3219 ASSERT(hdr->b_l1hdr.b_bufcnt > 0); 3220 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 3221 ASSERT3P(hdr->b_l1hdr.b_state, !=, arc_anon); 3222 ASSERT3P(buf->b_data, !=, NULL); 3223 3224 (void) remove_reference(hdr, hash_lock, tag); 3225 arc_buf_destroy_impl(buf); 3226 mutex_exit(hash_lock); 3227 } 3228 3229 /* 3230 * Evict the arc_buf_hdr that is provided as a parameter. The resultant 3231 * state of the header is dependent on it's state prior to entering this 3232 * function. The following transitions are possible: 3233 * 3234 * - arc_mru -> arc_mru_ghost 3235 * - arc_mfu -> arc_mfu_ghost 3236 * - arc_mru_ghost -> arc_l2c_only 3237 * - arc_mru_ghost -> deleted 3238 * - arc_mfu_ghost -> arc_l2c_only 3239 * - arc_mfu_ghost -> deleted 3240 */ 3241 static int64_t 3242 arc_evict_hdr(arc_buf_hdr_t *hdr, kmutex_t *hash_lock) 3243 { 3244 arc_state_t *evicted_state, *state; 3245 int64_t bytes_evicted = 0; 3246 3247 ASSERT(MUTEX_HELD(hash_lock)); 3248 ASSERT(HDR_HAS_L1HDR(hdr)); 3249 3250 state = hdr->b_l1hdr.b_state; 3251 if (GHOST_STATE(state)) { 3252 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 3253 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 3254 3255 /* 3256 * l2arc_write_buffers() relies on a header's L1 portion 3257 * (i.e. its b_pabd field) during it's write phase. 3258 * Thus, we cannot push a header onto the arc_l2c_only 3259 * state (removing it's L1 piece) until the header is 3260 * done being written to the l2arc. 3261 */ 3262 if (HDR_HAS_L2HDR(hdr) && HDR_L2_WRITING(hdr)) { 3263 ARCSTAT_BUMP(arcstat_evict_l2_skip); 3264 return (bytes_evicted); 3265 } 3266 3267 ARCSTAT_BUMP(arcstat_deleted); 3268 bytes_evicted += HDR_GET_LSIZE(hdr); 3269 3270 DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, hdr); 3271 3272 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL); 3273 if (HDR_HAS_L2HDR(hdr)) { 3274 /* 3275 * This buffer is cached on the 2nd Level ARC; 3276 * don't destroy the header. 3277 */ 3278 arc_change_state(arc_l2c_only, hdr, hash_lock); 3279 /* 3280 * dropping from L1+L2 cached to L2-only, 3281 * realloc to remove the L1 header. 3282 */ 3283 hdr = arc_hdr_realloc(hdr, hdr_full_cache, 3284 hdr_l2only_cache); 3285 } else { 3286 arc_change_state(arc_anon, hdr, hash_lock); 3287 arc_hdr_destroy(hdr); 3288 } 3289 return (bytes_evicted); 3290 } 3291 3292 ASSERT(state == arc_mru || state == arc_mfu); 3293 evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost; 3294 3295 /* prefetch buffers have a minimum lifespan */ 3296 if (HDR_IO_IN_PROGRESS(hdr) || 3297 ((hdr->b_flags & (ARC_FLAG_PREFETCH | ARC_FLAG_INDIRECT)) && 3298 ddi_get_lbolt() - hdr->b_l1hdr.b_arc_access < 3299 arc_min_prefetch_lifespan)) { 3300 ARCSTAT_BUMP(arcstat_evict_skip); 3301 return (bytes_evicted); 3302 } 3303 3304 ASSERT0(zfs_refcount_count(&hdr->b_l1hdr.b_refcnt)); 3305 while (hdr->b_l1hdr.b_buf) { 3306 arc_buf_t *buf = hdr->b_l1hdr.b_buf; 3307 if (!mutex_tryenter(&buf->b_evict_lock)) { 3308 ARCSTAT_BUMP(arcstat_mutex_miss); 3309 break; 3310 } 3311 if (buf->b_data != NULL) 3312 bytes_evicted += HDR_GET_LSIZE(hdr); 3313 mutex_exit(&buf->b_evict_lock); 3314 arc_buf_destroy_impl(buf); 3315 } 3316 3317 if (HDR_HAS_L2HDR(hdr)) { 3318 ARCSTAT_INCR(arcstat_evict_l2_cached, HDR_GET_LSIZE(hdr)); 3319 } else { 3320 if (l2arc_write_eligible(hdr->b_spa, hdr)) { 3321 ARCSTAT_INCR(arcstat_evict_l2_eligible, 3322 HDR_GET_LSIZE(hdr)); 3323 } else { 3324 ARCSTAT_INCR(arcstat_evict_l2_ineligible, 3325 HDR_GET_LSIZE(hdr)); 3326 } 3327 } 3328 3329 if (hdr->b_l1hdr.b_bufcnt == 0) { 3330 arc_cksum_free(hdr); 3331 3332 bytes_evicted += arc_hdr_size(hdr); 3333 3334 /* 3335 * If this hdr is being evicted and has a compressed 3336 * buffer then we discard it here before we change states. 3337 * This ensures that the accounting is updated correctly 3338 * in arc_free_data_impl(). 3339 */ 3340 arc_hdr_free_pabd(hdr); 3341 3342 arc_change_state(evicted_state, hdr, hash_lock); 3343 ASSERT(HDR_IN_HASH_TABLE(hdr)); 3344 arc_hdr_set_flags(hdr, ARC_FLAG_IN_HASH_TABLE); 3345 DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, hdr); 3346 } 3347 3348 return (bytes_evicted); 3349 } 3350 3351 static uint64_t 3352 arc_evict_state_impl(multilist_t *ml, int idx, arc_buf_hdr_t *marker, 3353 uint64_t spa, int64_t bytes) 3354 { 3355 multilist_sublist_t *mls; 3356 uint64_t bytes_evicted = 0; 3357 arc_buf_hdr_t *hdr; 3358 kmutex_t *hash_lock; 3359 int evict_count = 0; 3360 3361 ASSERT3P(marker, !=, NULL); 3362 IMPLY(bytes < 0, bytes == ARC_EVICT_ALL); 3363 3364 mls = multilist_sublist_lock(ml, idx); 3365 3366 for (hdr = multilist_sublist_prev(mls, marker); hdr != NULL; 3367 hdr = multilist_sublist_prev(mls, marker)) { 3368 if ((bytes != ARC_EVICT_ALL && bytes_evicted >= bytes) || 3369 (evict_count >= zfs_arc_evict_batch_limit)) 3370 break; 3371 3372 /* 3373 * To keep our iteration location, move the marker 3374 * forward. Since we're not holding hdr's hash lock, we 3375 * must be very careful and not remove 'hdr' from the 3376 * sublist. Otherwise, other consumers might mistake the 3377 * 'hdr' as not being on a sublist when they call the 3378 * multilist_link_active() function (they all rely on 3379 * the hash lock protecting concurrent insertions and 3380 * removals). multilist_sublist_move_forward() was 3381 * specifically implemented to ensure this is the case 3382 * (only 'marker' will be removed and re-inserted). 3383 */ 3384 multilist_sublist_move_forward(mls, marker); 3385 3386 /* 3387 * The only case where the b_spa field should ever be 3388 * zero, is the marker headers inserted by 3389 * arc_evict_state(). It's possible for multiple threads 3390 * to be calling arc_evict_state() concurrently (e.g. 3391 * dsl_pool_close() and zio_inject_fault()), so we must 3392 * skip any markers we see from these other threads. 3393 */ 3394 if (hdr->b_spa == 0) 3395 continue; 3396 3397 /* we're only interested in evicting buffers of a certain spa */ 3398 if (spa != 0 && hdr->b_spa != spa) { 3399 ARCSTAT_BUMP(arcstat_evict_skip); 3400 continue; 3401 } 3402 3403 hash_lock = HDR_LOCK(hdr); 3404 3405 /* 3406 * We aren't calling this function from any code path 3407 * that would already be holding a hash lock, so we're 3408 * asserting on this assumption to be defensive in case 3409 * this ever changes. Without this check, it would be 3410 * possible to incorrectly increment arcstat_mutex_miss 3411 * below (e.g. if the code changed such that we called 3412 * this function with a hash lock held). 3413 */ 3414 ASSERT(!MUTEX_HELD(hash_lock)); 3415 3416 if (mutex_tryenter(hash_lock)) { 3417 uint64_t evicted = arc_evict_hdr(hdr, hash_lock); 3418 mutex_exit(hash_lock); 3419 3420 bytes_evicted += evicted; 3421 3422 /* 3423 * If evicted is zero, arc_evict_hdr() must have 3424 * decided to skip this header, don't increment 3425 * evict_count in this case. 3426 */ 3427 if (evicted != 0) 3428 evict_count++; 3429 3430 /* 3431 * If arc_size isn't overflowing, signal any 3432 * threads that might happen to be waiting. 3433 * 3434 * For each header evicted, we wake up a single 3435 * thread. If we used cv_broadcast, we could 3436 * wake up "too many" threads causing arc_size 3437 * to significantly overflow arc_c; since 3438 * arc_get_data_impl() doesn't check for overflow 3439 * when it's woken up (it doesn't because it's 3440 * possible for the ARC to be overflowing while 3441 * full of un-evictable buffers, and the 3442 * function should proceed in this case). 3443 * 3444 * If threads are left sleeping, due to not 3445 * using cv_broadcast here, they will be woken 3446 * up via cv_broadcast in arc_adjust_cb() just 3447 * before arc_adjust_zthr sleeps. 3448 */ 3449 mutex_enter(&arc_adjust_lock); 3450 if (!arc_is_overflowing()) 3451 cv_signal(&arc_adjust_waiters_cv); 3452 mutex_exit(&arc_adjust_lock); 3453 } else { 3454 ARCSTAT_BUMP(arcstat_mutex_miss); 3455 } 3456 } 3457 3458 multilist_sublist_unlock(mls); 3459 3460 return (bytes_evicted); 3461 } 3462 3463 /* 3464 * Evict buffers from the given arc state, until we've removed the 3465 * specified number of bytes. Move the removed buffers to the 3466 * appropriate evict state. 3467 * 3468 * This function makes a "best effort". It skips over any buffers 3469 * it can't get a hash_lock on, and so, may not catch all candidates. 3470 * It may also return without evicting as much space as requested. 3471 * 3472 * If bytes is specified using the special value ARC_EVICT_ALL, this 3473 * will evict all available (i.e. unlocked and evictable) buffers from 3474 * the given arc state; which is used by arc_flush(). 3475 */ 3476 static uint64_t 3477 arc_evict_state(arc_state_t *state, uint64_t spa, int64_t bytes, 3478 arc_buf_contents_t type) 3479 { 3480 uint64_t total_evicted = 0; 3481 multilist_t *ml = state->arcs_list[type]; 3482 int num_sublists; 3483 arc_buf_hdr_t **markers; 3484 3485 IMPLY(bytes < 0, bytes == ARC_EVICT_ALL); 3486 3487 num_sublists = multilist_get_num_sublists(ml); 3488 3489 /* 3490 * If we've tried to evict from each sublist, made some 3491 * progress, but still have not hit the target number of bytes 3492 * to evict, we want to keep trying. The markers allow us to 3493 * pick up where we left off for each individual sublist, rather 3494 * than starting from the tail each time. 3495 */ 3496 markers = kmem_zalloc(sizeof (*markers) * num_sublists, KM_SLEEP); 3497 for (int i = 0; i < num_sublists; i++) { 3498 markers[i] = kmem_cache_alloc(hdr_full_cache, KM_SLEEP); 3499 3500 /* 3501 * A b_spa of 0 is used to indicate that this header is 3502 * a marker. This fact is used in arc_adjust_type() and 3503 * arc_evict_state_impl(). 3504 */ 3505 markers[i]->b_spa = 0; 3506 3507 multilist_sublist_t *mls = multilist_sublist_lock(ml, i); 3508 multilist_sublist_insert_tail(mls, markers[i]); 3509 multilist_sublist_unlock(mls); 3510 } 3511 3512 /* 3513 * While we haven't hit our target number of bytes to evict, or 3514 * we're evicting all available buffers. 3515 */ 3516 while (total_evicted < bytes || bytes == ARC_EVICT_ALL) { 3517 /* 3518 * Start eviction using a randomly selected sublist, 3519 * this is to try and evenly balance eviction across all 3520 * sublists. Always starting at the same sublist 3521 * (e.g. index 0) would cause evictions to favor certain 3522 * sublists over others. 3523 */ 3524 int sublist_idx = multilist_get_random_index(ml); 3525 uint64_t scan_evicted = 0; 3526 3527 for (int i = 0; i < num_sublists; i++) { 3528 uint64_t bytes_remaining; 3529 uint64_t bytes_evicted; 3530 3531 if (bytes == ARC_EVICT_ALL) 3532 bytes_remaining = ARC_EVICT_ALL; 3533 else if (total_evicted < bytes) 3534 bytes_remaining = bytes - total_evicted; 3535 else 3536 break; 3537 3538 bytes_evicted = arc_evict_state_impl(ml, sublist_idx, 3539 markers[sublist_idx], spa, bytes_remaining); 3540 3541 scan_evicted += bytes_evicted; 3542 total_evicted += bytes_evicted; 3543 3544 /* we've reached the end, wrap to the beginning */ 3545 if (++sublist_idx >= num_sublists) 3546 sublist_idx = 0; 3547 } 3548 3549 /* 3550 * If we didn't evict anything during this scan, we have 3551 * no reason to believe we'll evict more during another 3552 * scan, so break the loop. 3553 */ 3554 if (scan_evicted == 0) { 3555 /* This isn't possible, let's make that obvious */ 3556 ASSERT3S(bytes, !=, 0); 3557 3558 /* 3559 * When bytes is ARC_EVICT_ALL, the only way to 3560 * break the loop is when scan_evicted is zero. 3561 * In that case, we actually have evicted enough, 3562 * so we don't want to increment the kstat. 3563 */ 3564 if (bytes != ARC_EVICT_ALL) { 3565 ASSERT3S(total_evicted, <, bytes); 3566 ARCSTAT_BUMP(arcstat_evict_not_enough); 3567 } 3568 3569 break; 3570 } 3571 } 3572 3573 for (int i = 0; i < num_sublists; i++) { 3574 multilist_sublist_t *mls = multilist_sublist_lock(ml, i); 3575 multilist_sublist_remove(mls, markers[i]); 3576 multilist_sublist_unlock(mls); 3577 3578 kmem_cache_free(hdr_full_cache, markers[i]); 3579 } 3580 kmem_free(markers, sizeof (*markers) * num_sublists); 3581 3582 return (total_evicted); 3583 } 3584 3585 /* 3586 * Flush all "evictable" data of the given type from the arc state 3587 * specified. This will not evict any "active" buffers (i.e. referenced). 3588 * 3589 * When 'retry' is set to B_FALSE, the function will make a single pass 3590 * over the state and evict any buffers that it can. Since it doesn't 3591 * continually retry the eviction, it might end up leaving some buffers 3592 * in the ARC due to lock misses. 3593 * 3594 * When 'retry' is set to B_TRUE, the function will continually retry the 3595 * eviction until *all* evictable buffers have been removed from the 3596 * state. As a result, if concurrent insertions into the state are 3597 * allowed (e.g. if the ARC isn't shutting down), this function might 3598 * wind up in an infinite loop, continually trying to evict buffers. 3599 */ 3600 static uint64_t 3601 arc_flush_state(arc_state_t *state, uint64_t spa, arc_buf_contents_t type, 3602 boolean_t retry) 3603 { 3604 uint64_t evicted = 0; 3605 3606 while (zfs_refcount_count(&state->arcs_esize[type]) != 0) { 3607 evicted += arc_evict_state(state, spa, ARC_EVICT_ALL, type); 3608 3609 if (!retry) 3610 break; 3611 } 3612 3613 return (evicted); 3614 } 3615 3616 /* 3617 * Evict the specified number of bytes from the state specified, 3618 * restricting eviction to the spa and type given. This function 3619 * prevents us from trying to evict more from a state's list than 3620 * is "evictable", and to skip evicting altogether when passed a 3621 * negative value for "bytes". In contrast, arc_evict_state() will 3622 * evict everything it can, when passed a negative value for "bytes". 3623 */ 3624 static uint64_t 3625 arc_adjust_impl(arc_state_t *state, uint64_t spa, int64_t bytes, 3626 arc_buf_contents_t type) 3627 { 3628 int64_t delta; 3629 3630 if (bytes > 0 && zfs_refcount_count(&state->arcs_esize[type]) > 0) { 3631 delta = MIN(zfs_refcount_count(&state->arcs_esize[type]), 3632 bytes); 3633 return (arc_evict_state(state, spa, delta, type)); 3634 } 3635 3636 return (0); 3637 } 3638 3639 /* 3640 * Evict metadata buffers from the cache, such that arc_meta_used is 3641 * capped by the arc_meta_limit tunable. 3642 */ 3643 static uint64_t 3644 arc_adjust_meta(uint64_t meta_used) 3645 { 3646 uint64_t total_evicted = 0; 3647 int64_t target; 3648 3649 /* 3650 * If we're over the meta limit, we want to evict enough 3651 * metadata to get back under the meta limit. We don't want to 3652 * evict so much that we drop the MRU below arc_p, though. If 3653 * we're over the meta limit more than we're over arc_p, we 3654 * evict some from the MRU here, and some from the MFU below. 3655 */ 3656 target = MIN((int64_t)(meta_used - arc_meta_limit), 3657 (int64_t)(zfs_refcount_count(&arc_anon->arcs_size) + 3658 zfs_refcount_count(&arc_mru->arcs_size) - arc_p)); 3659 3660 total_evicted += arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA); 3661 3662 /* 3663 * Similar to the above, we want to evict enough bytes to get us 3664 * below the meta limit, but not so much as to drop us below the 3665 * space allotted to the MFU (which is defined as arc_c - arc_p). 3666 */ 3667 target = MIN((int64_t)(meta_used - arc_meta_limit), 3668 (int64_t)(zfs_refcount_count(&arc_mfu->arcs_size) - 3669 (arc_c - arc_p))); 3670 3671 total_evicted += arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA); 3672 3673 return (total_evicted); 3674 } 3675 3676 /* 3677 * Return the type of the oldest buffer in the given arc state 3678 * 3679 * This function will select a random sublist of type ARC_BUFC_DATA and 3680 * a random sublist of type ARC_BUFC_METADATA. The tail of each sublist 3681 * is compared, and the type which contains the "older" buffer will be 3682 * returned. 3683 */ 3684 static arc_buf_contents_t 3685 arc_adjust_type(arc_state_t *state) 3686 { 3687 multilist_t *data_ml = state->arcs_list[ARC_BUFC_DATA]; 3688 multilist_t *meta_ml = state->arcs_list[ARC_BUFC_METADATA]; 3689 int data_idx = multilist_get_random_index(data_ml); 3690 int meta_idx = multilist_get_random_index(meta_ml); 3691 multilist_sublist_t *data_mls; 3692 multilist_sublist_t *meta_mls; 3693 arc_buf_contents_t type; 3694 arc_buf_hdr_t *data_hdr; 3695 arc_buf_hdr_t *meta_hdr; 3696 3697 /* 3698 * We keep the sublist lock until we're finished, to prevent 3699 * the headers from being destroyed via arc_evict_state(). 3700 */ 3701 data_mls = multilist_sublist_lock(data_ml, data_idx); 3702 meta_mls = multilist_sublist_lock(meta_ml, meta_idx); 3703 3704 /* 3705 * These two loops are to ensure we skip any markers that 3706 * might be at the tail of the lists due to arc_evict_state(). 3707 */ 3708 3709 for (data_hdr = multilist_sublist_tail(data_mls); data_hdr != NULL; 3710 data_hdr = multilist_sublist_prev(data_mls, data_hdr)) { 3711 if (data_hdr->b_spa != 0) 3712 break; 3713 } 3714 3715 for (meta_hdr = multilist_sublist_tail(meta_mls); meta_hdr != NULL; 3716 meta_hdr = multilist_sublist_prev(meta_mls, meta_hdr)) { 3717 if (meta_hdr->b_spa != 0) 3718 break; 3719 } 3720 3721 if (data_hdr == NULL && meta_hdr == NULL) { 3722 type = ARC_BUFC_DATA; 3723 } else if (data_hdr == NULL) { 3724 ASSERT3P(meta_hdr, !=, NULL); 3725 type = ARC_BUFC_METADATA; 3726 } else if (meta_hdr == NULL) { 3727 ASSERT3P(data_hdr, !=, NULL); 3728 type = ARC_BUFC_DATA; 3729 } else { 3730 ASSERT3P(data_hdr, !=, NULL); 3731 ASSERT3P(meta_hdr, !=, NULL); 3732 3733 /* The headers can't be on the sublist without an L1 header */ 3734 ASSERT(HDR_HAS_L1HDR(data_hdr)); 3735 ASSERT(HDR_HAS_L1HDR(meta_hdr)); 3736 3737 if (data_hdr->b_l1hdr.b_arc_access < 3738 meta_hdr->b_l1hdr.b_arc_access) { 3739 type = ARC_BUFC_DATA; 3740 } else { 3741 type = ARC_BUFC_METADATA; 3742 } 3743 } 3744 3745 multilist_sublist_unlock(meta_mls); 3746 multilist_sublist_unlock(data_mls); 3747 3748 return (type); 3749 } 3750 3751 /* 3752 * Evict buffers from the cache, such that arc_size is capped by arc_c. 3753 */ 3754 static uint64_t 3755 arc_adjust(void) 3756 { 3757 uint64_t total_evicted = 0; 3758 uint64_t bytes; 3759 int64_t target; 3760 uint64_t asize = aggsum_value(&arc_size); 3761 uint64_t ameta = aggsum_value(&arc_meta_used); 3762 3763 /* 3764 * If we're over arc_meta_limit, we want to correct that before 3765 * potentially evicting data buffers below. 3766 */ 3767 total_evicted += arc_adjust_meta(ameta); 3768 3769 /* 3770 * Adjust MRU size 3771 * 3772 * If we're over the target cache size, we want to evict enough 3773 * from the list to get back to our target size. We don't want 3774 * to evict too much from the MRU, such that it drops below 3775 * arc_p. So, if we're over our target cache size more than 3776 * the MRU is over arc_p, we'll evict enough to get back to 3777 * arc_p here, and then evict more from the MFU below. 3778 */ 3779 target = MIN((int64_t)(asize - arc_c), 3780 (int64_t)(zfs_refcount_count(&arc_anon->arcs_size) + 3781 zfs_refcount_count(&arc_mru->arcs_size) + ameta - arc_p)); 3782 3783 /* 3784 * If we're below arc_meta_min, always prefer to evict data. 3785 * Otherwise, try to satisfy the requested number of bytes to 3786 * evict from the type which contains older buffers; in an 3787 * effort to keep newer buffers in the cache regardless of their 3788 * type. If we cannot satisfy the number of bytes from this 3789 * type, spill over into the next type. 3790 */ 3791 if (arc_adjust_type(arc_mru) == ARC_BUFC_METADATA && 3792 ameta > arc_meta_min) { 3793 bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA); 3794 total_evicted += bytes; 3795 3796 /* 3797 * If we couldn't evict our target number of bytes from 3798 * metadata, we try to get the rest from data. 3799 */ 3800 target -= bytes; 3801 3802 total_evicted += 3803 arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA); 3804 } else { 3805 bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA); 3806 total_evicted += bytes; 3807 3808 /* 3809 * If we couldn't evict our target number of bytes from 3810 * data, we try to get the rest from metadata. 3811 */ 3812 target -= bytes; 3813 3814 total_evicted += 3815 arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA); 3816 } 3817 3818 /* 3819 * Adjust MFU size 3820 * 3821 * Now that we've tried to evict enough from the MRU to get its 3822 * size back to arc_p, if we're still above the target cache 3823 * size, we evict the rest from the MFU. 3824 */ 3825 target = asize - arc_c; 3826 3827 if (arc_adjust_type(arc_mfu) == ARC_BUFC_METADATA && 3828 ameta > arc_meta_min) { 3829 bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA); 3830 total_evicted += bytes; 3831 3832 /* 3833 * If we couldn't evict our target number of bytes from 3834 * metadata, we try to get the rest from data. 3835 */ 3836 target -= bytes; 3837 3838 total_evicted += 3839 arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA); 3840 } else { 3841 bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA); 3842 total_evicted += bytes; 3843 3844 /* 3845 * If we couldn't evict our target number of bytes from 3846 * data, we try to get the rest from data. 3847 */ 3848 target -= bytes; 3849 3850 total_evicted += 3851 arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA); 3852 } 3853 3854 /* 3855 * Adjust ghost lists 3856 * 3857 * In addition to the above, the ARC also defines target values 3858 * for the ghost lists. The sum of the mru list and mru ghost 3859 * list should never exceed the target size of the cache, and 3860 * the sum of the mru list, mfu list, mru ghost list, and mfu 3861 * ghost list should never exceed twice the target size of the 3862 * cache. The following logic enforces these limits on the ghost 3863 * caches, and evicts from them as needed. 3864 */ 3865 target = zfs_refcount_count(&arc_mru->arcs_size) + 3866 zfs_refcount_count(&arc_mru_ghost->arcs_size) - arc_c; 3867 3868 bytes = arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_DATA); 3869 total_evicted += bytes; 3870 3871 target -= bytes; 3872 3873 total_evicted += 3874 arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_METADATA); 3875 3876 /* 3877 * We assume the sum of the mru list and mfu list is less than 3878 * or equal to arc_c (we enforced this above), which means we 3879 * can use the simpler of the two equations below: 3880 * 3881 * mru + mfu + mru ghost + mfu ghost <= 2 * arc_c 3882 * mru ghost + mfu ghost <= arc_c 3883 */ 3884 target = zfs_refcount_count(&arc_mru_ghost->arcs_size) + 3885 zfs_refcount_count(&arc_mfu_ghost->arcs_size) - arc_c; 3886 3887 bytes = arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_DATA); 3888 total_evicted += bytes; 3889 3890 target -= bytes; 3891 3892 total_evicted += 3893 arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_METADATA); 3894 3895 return (total_evicted); 3896 } 3897 3898 void 3899 arc_flush(spa_t *spa, boolean_t retry) 3900 { 3901 uint64_t guid = 0; 3902 3903 /* 3904 * If retry is B_TRUE, a spa must not be specified since we have 3905 * no good way to determine if all of a spa's buffers have been 3906 * evicted from an arc state. 3907 */ 3908 ASSERT(!retry || spa == 0); 3909 3910 if (spa != NULL) 3911 guid = spa_load_guid(spa); 3912 3913 (void) arc_flush_state(arc_mru, guid, ARC_BUFC_DATA, retry); 3914 (void) arc_flush_state(arc_mru, guid, ARC_BUFC_METADATA, retry); 3915 3916 (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_DATA, retry); 3917 (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_METADATA, retry); 3918 3919 (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_DATA, retry); 3920 (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_METADATA, retry); 3921 3922 (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_DATA, retry); 3923 (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_METADATA, retry); 3924 } 3925 3926 static void 3927 arc_reduce_target_size(int64_t to_free) 3928 { 3929 uint64_t asize = aggsum_value(&arc_size); 3930 if (arc_c > arc_c_min) { 3931 3932 if (arc_c > arc_c_min + to_free) 3933 atomic_add_64(&arc_c, -to_free); 3934 else 3935 arc_c = arc_c_min; 3936 3937 atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift)); 3938 if (asize < arc_c) 3939 arc_c = MAX(asize, arc_c_min); 3940 if (arc_p > arc_c) 3941 arc_p = (arc_c >> 1); 3942 ASSERT(arc_c >= arc_c_min); 3943 ASSERT((int64_t)arc_p >= 0); 3944 } 3945 3946 if (asize > arc_c) { 3947 /* See comment in arc_adjust_cb_check() on why lock+flag */ 3948 mutex_enter(&arc_adjust_lock); 3949 arc_adjust_needed = B_TRUE; 3950 mutex_exit(&arc_adjust_lock); 3951 zthr_wakeup(arc_adjust_zthr); 3952 } 3953 } 3954 3955 typedef enum free_memory_reason_t { 3956 FMR_UNKNOWN, 3957 FMR_NEEDFREE, 3958 FMR_LOTSFREE, 3959 FMR_SWAPFS_MINFREE, 3960 FMR_PAGES_PP_MAXIMUM, 3961 FMR_HEAP_ARENA, 3962 FMR_ZIO_ARENA, 3963 } free_memory_reason_t; 3964 3965 int64_t last_free_memory; 3966 free_memory_reason_t last_free_reason; 3967 3968 /* 3969 * Additional reserve of pages for pp_reserve. 3970 */ 3971 int64_t arc_pages_pp_reserve = 64; 3972 3973 /* 3974 * Additional reserve of pages for swapfs. 3975 */ 3976 int64_t arc_swapfs_reserve = 64; 3977 3978 /* 3979 * Return the amount of memory that can be consumed before reclaim will be 3980 * needed. Positive if there is sufficient free memory, negative indicates 3981 * the amount of memory that needs to be freed up. 3982 */ 3983 static int64_t 3984 arc_available_memory(void) 3985 { 3986 int64_t lowest = INT64_MAX; 3987 int64_t n; 3988 free_memory_reason_t r = FMR_UNKNOWN; 3989 3990 #ifdef _KERNEL 3991 if (needfree > 0) { 3992 n = PAGESIZE * (-needfree); 3993 if (n < lowest) { 3994 lowest = n; 3995 r = FMR_NEEDFREE; 3996 } 3997 } 3998 3999 /* 4000 * check that we're out of range of the pageout scanner. It starts to 4001 * schedule paging if freemem is less than lotsfree and needfree. 4002 * lotsfree is the high-water mark for pageout, and needfree is the 4003 * number of needed free pages. We add extra pages here to make sure 4004 * the scanner doesn't start up while we're freeing memory. 4005 */ 4006 n = PAGESIZE * (freemem - lotsfree - needfree - desfree); 4007 if (n < lowest) { 4008 lowest = n; 4009 r = FMR_LOTSFREE; 4010 } 4011 4012 /* 4013 * check to make sure that swapfs has enough space so that anon 4014 * reservations can still succeed. anon_resvmem() checks that the 4015 * availrmem is greater than swapfs_minfree, and the number of reserved 4016 * swap pages. We also add a bit of extra here just to prevent 4017 * circumstances from getting really dire. 4018 */ 4019 n = PAGESIZE * (availrmem - swapfs_minfree - swapfs_reserve - 4020 desfree - arc_swapfs_reserve); 4021 if (n < lowest) { 4022 lowest = n; 4023 r = FMR_SWAPFS_MINFREE; 4024 } 4025 4026 4027 /* 4028 * Check that we have enough availrmem that memory locking (e.g., via 4029 * mlock(3C) or memcntl(2)) can still succeed. (pages_pp_maximum 4030 * stores the number of pages that cannot be locked; when availrmem 4031 * drops below pages_pp_maximum, page locking mechanisms such as 4032 * page_pp_lock() will fail.) 4033 */ 4034 n = PAGESIZE * (availrmem - pages_pp_maximum - 4035 arc_pages_pp_reserve); 4036 if (n < lowest) { 4037 lowest = n; 4038 r = FMR_PAGES_PP_MAXIMUM; 4039 } 4040 4041 #if defined(__i386) 4042 /* 4043 * If we're on an i386 platform, it's possible that we'll exhaust the 4044 * kernel heap space before we ever run out of available physical 4045 * memory. Most checks of the size of the heap_area compare against 4046 * tune.t_minarmem, which is the minimum available real memory that we 4047 * can have in the system. However, this is generally fixed at 25 pages 4048 * which is so low that it's useless. In this comparison, we seek to 4049 * calculate the total heap-size, and reclaim if more than 3/4ths of the 4050 * heap is allocated. (Or, in the calculation, if less than 1/4th is 4051 * free) 4052 */ 4053 n = (int64_t)vmem_size(heap_arena, VMEM_FREE) - 4054 (vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC) >> 2); 4055 if (n < lowest) { 4056 lowest = n; 4057 r = FMR_HEAP_ARENA; 4058 } 4059 #endif 4060 4061 /* 4062 * If zio data pages are being allocated out of a separate heap segment, 4063 * then enforce that the size of available vmem for this arena remains 4064 * above about 1/4th (1/(2^arc_zio_arena_free_shift)) free. 4065 * 4066 * Note that reducing the arc_zio_arena_free_shift keeps more virtual 4067 * memory (in the zio_arena) free, which can avoid memory 4068 * fragmentation issues. 4069 */ 4070 if (zio_arena != NULL) { 4071 n = (int64_t)vmem_size(zio_arena, VMEM_FREE) - 4072 (vmem_size(zio_arena, VMEM_ALLOC) >> 4073 arc_zio_arena_free_shift); 4074 if (n < lowest) { 4075 lowest = n; 4076 r = FMR_ZIO_ARENA; 4077 } 4078 } 4079 #else 4080 /* Every 100 calls, free a small amount */ 4081 if (spa_get_random(100) == 0) 4082 lowest = -1024; 4083 #endif 4084 4085 last_free_memory = lowest; 4086 last_free_reason = r; 4087 4088 return (lowest); 4089 } 4090 4091 4092 /* 4093 * Determine if the system is under memory pressure and is asking 4094 * to reclaim memory. A return value of B_TRUE indicates that the system 4095 * is under memory pressure and that the arc should adjust accordingly. 4096 */ 4097 static boolean_t 4098 arc_reclaim_needed(void) 4099 { 4100 return (arc_available_memory() < 0); 4101 } 4102 4103 static void 4104 arc_kmem_reap_soon(void) 4105 { 4106 size_t i; 4107 kmem_cache_t *prev_cache = NULL; 4108 kmem_cache_t *prev_data_cache = NULL; 4109 extern kmem_cache_t *zio_buf_cache[]; 4110 extern kmem_cache_t *zio_data_buf_cache[]; 4111 extern kmem_cache_t *range_seg_cache; 4112 extern kmem_cache_t *abd_chunk_cache; 4113 4114 #ifdef _KERNEL 4115 if (aggsum_compare(&arc_meta_used, arc_meta_limit) >= 0) { 4116 /* 4117 * We are exceeding our meta-data cache limit. 4118 * Purge some DNLC entries to release holds on meta-data. 4119 */ 4120 dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent); 4121 } 4122 #if defined(__i386) 4123 /* 4124 * Reclaim unused memory from all kmem caches. 4125 */ 4126 kmem_reap(); 4127 #endif 4128 #endif 4129 4130 for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) { 4131 if (zio_buf_cache[i] != prev_cache) { 4132 prev_cache = zio_buf_cache[i]; 4133 kmem_cache_reap_soon(zio_buf_cache[i]); 4134 } 4135 if (zio_data_buf_cache[i] != prev_data_cache) { 4136 prev_data_cache = zio_data_buf_cache[i]; 4137 kmem_cache_reap_soon(zio_data_buf_cache[i]); 4138 } 4139 } 4140 kmem_cache_reap_soon(abd_chunk_cache); 4141 kmem_cache_reap_soon(buf_cache); 4142 kmem_cache_reap_soon(hdr_full_cache); 4143 kmem_cache_reap_soon(hdr_l2only_cache); 4144 kmem_cache_reap_soon(range_seg_cache); 4145 4146 if (zio_arena != NULL) { 4147 /* 4148 * Ask the vmem arena to reclaim unused memory from its 4149 * quantum caches. 4150 */ 4151 vmem_qcache_reap(zio_arena); 4152 } 4153 } 4154 4155 /* ARGSUSED */ 4156 static boolean_t 4157 arc_adjust_cb_check(void *arg, zthr_t *zthr) 4158 { 4159 /* 4160 * This is necessary in order for the mdb ::arc dcmd to 4161 * show up to date information. Since the ::arc command 4162 * does not call the kstat's update function, without 4163 * this call, the command may show stale stats for the 4164 * anon, mru, mru_ghost, mfu, and mfu_ghost lists. Even 4165 * with this change, the data might be up to 1 second 4166 * out of date(the arc_adjust_zthr has a maximum sleep 4167 * time of 1 second); but that should suffice. The 4168 * arc_state_t structures can be queried directly if more 4169 * accurate information is needed. 4170 */ 4171 if (arc_ksp != NULL) 4172 arc_ksp->ks_update(arc_ksp, KSTAT_READ); 4173 4174 /* 4175 * We have to rely on arc_get_data_impl() to tell us when to adjust, 4176 * rather than checking if we are overflowing here, so that we are 4177 * sure to not leave arc_get_data_impl() waiting on 4178 * arc_adjust_waiters_cv. If we have become "not overflowing" since 4179 * arc_get_data_impl() checked, we need to wake it up. We could 4180 * broadcast the CV here, but arc_get_data_impl() may have not yet 4181 * gone to sleep. We would need to use a mutex to ensure that this 4182 * function doesn't broadcast until arc_get_data_impl() has gone to 4183 * sleep (e.g. the arc_adjust_lock). However, the lock ordering of 4184 * such a lock would necessarily be incorrect with respect to the 4185 * zthr_lock, which is held before this function is called, and is 4186 * held by arc_get_data_impl() when it calls zthr_wakeup(). 4187 */ 4188 return (arc_adjust_needed); 4189 } 4190 4191 /* 4192 * Keep arc_size under arc_c by running arc_adjust which evicts data 4193 * from the ARC. 4194 */ 4195 /* ARGSUSED */ 4196 static int 4197 arc_adjust_cb(void *arg, zthr_t *zthr) 4198 { 4199 uint64_t evicted = 0; 4200 4201 /* Evict from cache */ 4202 evicted = arc_adjust(); 4203 4204 /* 4205 * If evicted is zero, we couldn't evict anything 4206 * via arc_adjust(). This could be due to hash lock 4207 * collisions, but more likely due to the majority of 4208 * arc buffers being unevictable. Therefore, even if 4209 * arc_size is above arc_c, another pass is unlikely to 4210 * be helpful and could potentially cause us to enter an 4211 * infinite loop. Additionally, zthr_iscancelled() is 4212 * checked here so that if the arc is shutting down, the 4213 * broadcast will wake any remaining arc adjust waiters. 4214 */ 4215 mutex_enter(&arc_adjust_lock); 4216 arc_adjust_needed = !zthr_iscancelled(arc_adjust_zthr) && 4217 evicted > 0 && aggsum_compare(&arc_size, arc_c) > 0; 4218 if (!arc_adjust_needed) { 4219 /* 4220 * We're either no longer overflowing, or we 4221 * can't evict anything more, so we should wake 4222 * up any waiters. 4223 */ 4224 cv_broadcast(&arc_adjust_waiters_cv); 4225 } 4226 mutex_exit(&arc_adjust_lock); 4227 4228 return (0); 4229 } 4230 4231 /* ARGSUSED */ 4232 static boolean_t 4233 arc_reap_cb_check(void *arg, zthr_t *zthr) 4234 { 4235 int64_t free_memory = arc_available_memory(); 4236 4237 /* 4238 * If a kmem reap is already active, don't schedule more. We must 4239 * check for this because kmem_cache_reap_soon() won't actually 4240 * block on the cache being reaped (this is to prevent callers from 4241 * becoming implicitly blocked by a system-wide kmem reap -- which, 4242 * on a system with many, many full magazines, can take minutes). 4243 */ 4244 if (!kmem_cache_reap_active() && 4245 free_memory < 0) { 4246 arc_no_grow = B_TRUE; 4247 arc_warm = B_TRUE; 4248 /* 4249 * Wait at least zfs_grow_retry (default 60) seconds 4250 * before considering growing. 4251 */ 4252 arc_growtime = gethrtime() + SEC2NSEC(arc_grow_retry); 4253 return (B_TRUE); 4254 } else if (free_memory < arc_c >> arc_no_grow_shift) { 4255 arc_no_grow = B_TRUE; 4256 } else if (gethrtime() >= arc_growtime) { 4257 arc_no_grow = B_FALSE; 4258 } 4259 4260 return (B_FALSE); 4261 } 4262 4263 /* 4264 * Keep enough free memory in the system by reaping the ARC's kmem 4265 * caches. To cause more slabs to be reapable, we may reduce the 4266 * target size of the cache (arc_c), causing the arc_adjust_cb() 4267 * to free more buffers. 4268 */ 4269 /* ARGSUSED */ 4270 static int 4271 arc_reap_cb(void *arg, zthr_t *zthr) 4272 { 4273 int64_t free_memory; 4274 4275 /* 4276 * Kick off asynchronous kmem_reap()'s of all our caches. 4277 */ 4278 arc_kmem_reap_soon(); 4279 4280 /* 4281 * Wait at least arc_kmem_cache_reap_retry_ms between 4282 * arc_kmem_reap_soon() calls. Without this check it is possible to 4283 * end up in a situation where we spend lots of time reaping 4284 * caches, while we're near arc_c_min. Waiting here also gives the 4285 * subsequent free memory check a chance of finding that the 4286 * asynchronous reap has already freed enough memory, and we don't 4287 * need to call arc_reduce_target_size(). 4288 */ 4289 delay((hz * arc_kmem_cache_reap_retry_ms + 999) / 1000); 4290 4291 /* 4292 * Reduce the target size as needed to maintain the amount of free 4293 * memory in the system at a fraction of the arc_size (1/128th by 4294 * default). If oversubscribed (free_memory < 0) then reduce the 4295 * target arc_size by the deficit amount plus the fractional 4296 * amount. If free memory is positive but less then the fractional 4297 * amount, reduce by what is needed to hit the fractional amount. 4298 */ 4299 free_memory = arc_available_memory(); 4300 4301 int64_t to_free = 4302 (arc_c >> arc_shrink_shift) - free_memory; 4303 if (to_free > 0) { 4304 #ifdef _KERNEL 4305 to_free = MAX(to_free, ptob(needfree)); 4306 #endif 4307 arc_reduce_target_size(to_free); 4308 } 4309 4310 return (0); 4311 } 4312 4313 /* 4314 * Adapt arc info given the number of bytes we are trying to add and 4315 * the state that we are comming from. This function is only called 4316 * when we are adding new content to the cache. 4317 */ 4318 static void 4319 arc_adapt(int bytes, arc_state_t *state) 4320 { 4321 int mult; 4322 uint64_t arc_p_min = (arc_c >> arc_p_min_shift); 4323 int64_t mrug_size = zfs_refcount_count(&arc_mru_ghost->arcs_size); 4324 int64_t mfug_size = zfs_refcount_count(&arc_mfu_ghost->arcs_size); 4325 4326 if (state == arc_l2c_only) 4327 return; 4328 4329 ASSERT(bytes > 0); 4330 /* 4331 * Adapt the target size of the MRU list: 4332 * - if we just hit in the MRU ghost list, then increase 4333 * the target size of the MRU list. 4334 * - if we just hit in the MFU ghost list, then increase 4335 * the target size of the MFU list by decreasing the 4336 * target size of the MRU list. 4337 */ 4338 if (state == arc_mru_ghost) { 4339 mult = (mrug_size >= mfug_size) ? 1 : (mfug_size / mrug_size); 4340 mult = MIN(mult, 10); /* avoid wild arc_p adjustment */ 4341 4342 arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult); 4343 } else if (state == arc_mfu_ghost) { 4344 uint64_t delta; 4345 4346 mult = (mfug_size >= mrug_size) ? 1 : (mrug_size / mfug_size); 4347 mult = MIN(mult, 10); 4348 4349 delta = MIN(bytes * mult, arc_p); 4350 arc_p = MAX(arc_p_min, arc_p - delta); 4351 } 4352 ASSERT((int64_t)arc_p >= 0); 4353 4354 /* 4355 * Wake reap thread if we do not have any available memory 4356 */ 4357 if (arc_reclaim_needed()) { 4358 zthr_wakeup(arc_reap_zthr); 4359 return; 4360 } 4361 4362 4363 if (arc_no_grow) 4364 return; 4365 4366 if (arc_c >= arc_c_max) 4367 return; 4368 4369 /* 4370 * If we're within (2 * maxblocksize) bytes of the target 4371 * cache size, increment the target cache size 4372 */ 4373 if (aggsum_compare(&arc_size, arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) > 4374 0) { 4375 atomic_add_64(&arc_c, (int64_t)bytes); 4376 if (arc_c > arc_c_max) 4377 arc_c = arc_c_max; 4378 else if (state == arc_anon) 4379 atomic_add_64(&arc_p, (int64_t)bytes); 4380 if (arc_p > arc_c) 4381 arc_p = arc_c; 4382 } 4383 ASSERT((int64_t)arc_p >= 0); 4384 } 4385 4386 /* 4387 * Check if arc_size has grown past our upper threshold, determined by 4388 * zfs_arc_overflow_shift. 4389 */ 4390 static boolean_t 4391 arc_is_overflowing(void) 4392 { 4393 /* Always allow at least one block of overflow */ 4394 uint64_t overflow = MAX(SPA_MAXBLOCKSIZE, 4395 arc_c >> zfs_arc_overflow_shift); 4396 4397 /* 4398 * We just compare the lower bound here for performance reasons. Our 4399 * primary goals are to make sure that the arc never grows without 4400 * bound, and that it can reach its maximum size. This check 4401 * accomplishes both goals. The maximum amount we could run over by is 4402 * 2 * aggsum_borrow_multiplier * NUM_CPUS * the average size of a block 4403 * in the ARC. In practice, that's in the tens of MB, which is low 4404 * enough to be safe. 4405 */ 4406 return (aggsum_lower_bound(&arc_size) >= arc_c + overflow); 4407 } 4408 4409 static abd_t * 4410 arc_get_data_abd(arc_buf_hdr_t *hdr, uint64_t size, void *tag) 4411 { 4412 arc_buf_contents_t type = arc_buf_type(hdr); 4413 4414 arc_get_data_impl(hdr, size, tag); 4415 if (type == ARC_BUFC_METADATA) { 4416 return (abd_alloc(size, B_TRUE)); 4417 } else { 4418 ASSERT(type == ARC_BUFC_DATA); 4419 return (abd_alloc(size, B_FALSE)); 4420 } 4421 } 4422 4423 static void * 4424 arc_get_data_buf(arc_buf_hdr_t *hdr, uint64_t size, void *tag) 4425 { 4426 arc_buf_contents_t type = arc_buf_type(hdr); 4427 4428 arc_get_data_impl(hdr, size, tag); 4429 if (type == ARC_BUFC_METADATA) { 4430 return (zio_buf_alloc(size)); 4431 } else { 4432 ASSERT(type == ARC_BUFC_DATA); 4433 return (zio_data_buf_alloc(size)); 4434 } 4435 } 4436 4437 /* 4438 * Allocate a block and return it to the caller. If we are hitting the 4439 * hard limit for the cache size, we must sleep, waiting for the eviction 4440 * thread to catch up. If we're past the target size but below the hard 4441 * limit, we'll only signal the reclaim thread and continue on. 4442 */ 4443 static void 4444 arc_get_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag) 4445 { 4446 arc_state_t *state = hdr->b_l1hdr.b_state; 4447 arc_buf_contents_t type = arc_buf_type(hdr); 4448 4449 arc_adapt(size, state); 4450 4451 /* 4452 * If arc_size is currently overflowing, and has grown past our 4453 * upper limit, we must be adding data faster than the evict 4454 * thread can evict. Thus, to ensure we don't compound the 4455 * problem by adding more data and forcing arc_size to grow even 4456 * further past it's target size, we halt and wait for the 4457 * eviction thread to catch up. 4458 * 4459 * It's also possible that the reclaim thread is unable to evict 4460 * enough buffers to get arc_size below the overflow limit (e.g. 4461 * due to buffers being un-evictable, or hash lock collisions). 4462 * In this case, we want to proceed regardless if we're 4463 * overflowing; thus we don't use a while loop here. 4464 */ 4465 if (arc_is_overflowing()) { 4466 mutex_enter(&arc_adjust_lock); 4467 4468 /* 4469 * Now that we've acquired the lock, we may no longer be 4470 * over the overflow limit, lets check. 4471 * 4472 * We're ignoring the case of spurious wake ups. If that 4473 * were to happen, it'd let this thread consume an ARC 4474 * buffer before it should have (i.e. before we're under 4475 * the overflow limit and were signalled by the reclaim 4476 * thread). As long as that is a rare occurrence, it 4477 * shouldn't cause any harm. 4478 */ 4479 if (arc_is_overflowing()) { 4480 arc_adjust_needed = B_TRUE; 4481 zthr_wakeup(arc_adjust_zthr); 4482 (void) cv_wait(&arc_adjust_waiters_cv, 4483 &arc_adjust_lock); 4484 } 4485 mutex_exit(&arc_adjust_lock); 4486 } 4487 4488 VERIFY3U(hdr->b_type, ==, type); 4489 if (type == ARC_BUFC_METADATA) { 4490 arc_space_consume(size, ARC_SPACE_META); 4491 } else { 4492 arc_space_consume(size, ARC_SPACE_DATA); 4493 } 4494 4495 /* 4496 * Update the state size. Note that ghost states have a 4497 * "ghost size" and so don't need to be updated. 4498 */ 4499 if (!GHOST_STATE(state)) { 4500 4501 (void) zfs_refcount_add_many(&state->arcs_size, size, tag); 4502 4503 /* 4504 * If this is reached via arc_read, the link is 4505 * protected by the hash lock. If reached via 4506 * arc_buf_alloc, the header should not be accessed by 4507 * any other thread. And, if reached via arc_read_done, 4508 * the hash lock will protect it if it's found in the 4509 * hash table; otherwise no other thread should be 4510 * trying to [add|remove]_reference it. 4511 */ 4512 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) { 4513 ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 4514 (void) zfs_refcount_add_many(&state->arcs_esize[type], 4515 size, tag); 4516 } 4517 4518 /* 4519 * If we are growing the cache, and we are adding anonymous 4520 * data, and we have outgrown arc_p, update arc_p 4521 */ 4522 if (aggsum_compare(&arc_size, arc_c) < 0 && 4523 hdr->b_l1hdr.b_state == arc_anon && 4524 (zfs_refcount_count(&arc_anon->arcs_size) + 4525 zfs_refcount_count(&arc_mru->arcs_size) > arc_p)) 4526 arc_p = MIN(arc_c, arc_p + size); 4527 } 4528 } 4529 4530 static void 4531 arc_free_data_abd(arc_buf_hdr_t *hdr, abd_t *abd, uint64_t size, void *tag) 4532 { 4533 arc_free_data_impl(hdr, size, tag); 4534 abd_free(abd); 4535 } 4536 4537 static void 4538 arc_free_data_buf(arc_buf_hdr_t *hdr, void *buf, uint64_t size, void *tag) 4539 { 4540 arc_buf_contents_t type = arc_buf_type(hdr); 4541 4542 arc_free_data_impl(hdr, size, tag); 4543 if (type == ARC_BUFC_METADATA) { 4544 zio_buf_free(buf, size); 4545 } else { 4546 ASSERT(type == ARC_BUFC_DATA); 4547 zio_data_buf_free(buf, size); 4548 } 4549 } 4550 4551 /* 4552 * Free the arc data buffer. 4553 */ 4554 static void 4555 arc_free_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag) 4556 { 4557 arc_state_t *state = hdr->b_l1hdr.b_state; 4558 arc_buf_contents_t type = arc_buf_type(hdr); 4559 4560 /* protected by hash lock, if in the hash table */ 4561 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) { 4562 ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 4563 ASSERT(state != arc_anon && state != arc_l2c_only); 4564 4565 (void) zfs_refcount_remove_many(&state->arcs_esize[type], 4566 size, tag); 4567 } 4568 (void) zfs_refcount_remove_many(&state->arcs_size, size, tag); 4569 4570 VERIFY3U(hdr->b_type, ==, type); 4571 if (type == ARC_BUFC_METADATA) { 4572 arc_space_return(size, ARC_SPACE_META); 4573 } else { 4574 ASSERT(type == ARC_BUFC_DATA); 4575 arc_space_return(size, ARC_SPACE_DATA); 4576 } 4577 } 4578 4579 /* 4580 * This routine is called whenever a buffer is accessed. 4581 * NOTE: the hash lock is dropped in this function. 4582 */ 4583 static void 4584 arc_access(arc_buf_hdr_t *hdr, kmutex_t *hash_lock) 4585 { 4586 clock_t now; 4587 4588 ASSERT(MUTEX_HELD(hash_lock)); 4589 ASSERT(HDR_HAS_L1HDR(hdr)); 4590 4591 if (hdr->b_l1hdr.b_state == arc_anon) { 4592 /* 4593 * This buffer is not in the cache, and does not 4594 * appear in our "ghost" list. Add the new buffer 4595 * to the MRU state. 4596 */ 4597 4598 ASSERT0(hdr->b_l1hdr.b_arc_access); 4599 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 4600 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr); 4601 arc_change_state(arc_mru, hdr, hash_lock); 4602 4603 } else if (hdr->b_l1hdr.b_state == arc_mru) { 4604 now = ddi_get_lbolt(); 4605 4606 /* 4607 * If this buffer is here because of a prefetch, then either: 4608 * - clear the flag if this is a "referencing" read 4609 * (any subsequent access will bump this into the MFU state). 4610 * or 4611 * - move the buffer to the head of the list if this is 4612 * another prefetch (to make it less likely to be evicted). 4613 */ 4614 if (HDR_PREFETCH(hdr)) { 4615 if (zfs_refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) { 4616 /* link protected by hash lock */ 4617 ASSERT(multilist_link_active( 4618 &hdr->b_l1hdr.b_arc_node)); 4619 } else { 4620 arc_hdr_clear_flags(hdr, ARC_FLAG_PREFETCH); 4621 ARCSTAT_BUMP(arcstat_mru_hits); 4622 } 4623 hdr->b_l1hdr.b_arc_access = now; 4624 return; 4625 } 4626 4627 /* 4628 * This buffer has been "accessed" only once so far, 4629 * but it is still in the cache. Move it to the MFU 4630 * state. 4631 */ 4632 if (now > hdr->b_l1hdr.b_arc_access + ARC_MINTIME) { 4633 /* 4634 * More than 125ms have passed since we 4635 * instantiated this buffer. Move it to the 4636 * most frequently used state. 4637 */ 4638 hdr->b_l1hdr.b_arc_access = now; 4639 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 4640 arc_change_state(arc_mfu, hdr, hash_lock); 4641 } 4642 ARCSTAT_BUMP(arcstat_mru_hits); 4643 } else if (hdr->b_l1hdr.b_state == arc_mru_ghost) { 4644 arc_state_t *new_state; 4645 /* 4646 * This buffer has been "accessed" recently, but 4647 * was evicted from the cache. Move it to the 4648 * MFU state. 4649 */ 4650 4651 if (HDR_PREFETCH(hdr)) { 4652 new_state = arc_mru; 4653 if (zfs_refcount_count(&hdr->b_l1hdr.b_refcnt) > 0) 4654 arc_hdr_clear_flags(hdr, ARC_FLAG_PREFETCH); 4655 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr); 4656 } else { 4657 new_state = arc_mfu; 4658 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 4659 } 4660 4661 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 4662 arc_change_state(new_state, hdr, hash_lock); 4663 4664 ARCSTAT_BUMP(arcstat_mru_ghost_hits); 4665 } else if (hdr->b_l1hdr.b_state == arc_mfu) { 4666 /* 4667 * This buffer has been accessed more than once and is 4668 * still in the cache. Keep it in the MFU state. 4669 * 4670 * NOTE: an add_reference() that occurred when we did 4671 * the arc_read() will have kicked this off the list. 4672 * If it was a prefetch, we will explicitly move it to 4673 * the head of the list now. 4674 */ 4675 if ((HDR_PREFETCH(hdr)) != 0) { 4676 ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 4677 /* link protected by hash_lock */ 4678 ASSERT(multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 4679 } 4680 ARCSTAT_BUMP(arcstat_mfu_hits); 4681 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 4682 } else if (hdr->b_l1hdr.b_state == arc_mfu_ghost) { 4683 arc_state_t *new_state = arc_mfu; 4684 /* 4685 * This buffer has been accessed more than once but has 4686 * been evicted from the cache. Move it back to the 4687 * MFU state. 4688 */ 4689 4690 if (HDR_PREFETCH(hdr)) { 4691 /* 4692 * This is a prefetch access... 4693 * move this block back to the MRU state. 4694 */ 4695 ASSERT0(zfs_refcount_count(&hdr->b_l1hdr.b_refcnt)); 4696 new_state = arc_mru; 4697 } 4698 4699 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 4700 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 4701 arc_change_state(new_state, hdr, hash_lock); 4702 4703 ARCSTAT_BUMP(arcstat_mfu_ghost_hits); 4704 } else if (hdr->b_l1hdr.b_state == arc_l2c_only) { 4705 /* 4706 * This buffer is on the 2nd Level ARC. 4707 */ 4708 4709 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 4710 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 4711 arc_change_state(arc_mfu, hdr, hash_lock); 4712 } else { 4713 ASSERT(!"invalid arc state"); 4714 } 4715 } 4716 4717 /* a generic arc_done_func_t which you can use */ 4718 /* ARGSUSED */ 4719 void 4720 arc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg) 4721 { 4722 if (zio == NULL || zio->io_error == 0) 4723 bcopy(buf->b_data, arg, arc_buf_size(buf)); 4724 arc_buf_destroy(buf, arg); 4725 } 4726 4727 /* a generic arc_done_func_t */ 4728 void 4729 arc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg) 4730 { 4731 arc_buf_t **bufp = arg; 4732 if (buf == NULL) { 4733 ASSERT(zio == NULL || zio->io_error != 0); 4734 *bufp = NULL; 4735 } else { 4736 ASSERT(zio == NULL || zio->io_error == 0); 4737 *bufp = buf; 4738 ASSERT(buf->b_data != NULL); 4739 } 4740 } 4741 4742 static void 4743 arc_hdr_verify(arc_buf_hdr_t *hdr, blkptr_t *bp) 4744 { 4745 if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp)) { 4746 ASSERT3U(HDR_GET_PSIZE(hdr), ==, 0); 4747 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF); 4748 } else { 4749 if (HDR_COMPRESSION_ENABLED(hdr)) { 4750 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, 4751 BP_GET_COMPRESS(bp)); 4752 } 4753 ASSERT3U(HDR_GET_LSIZE(hdr), ==, BP_GET_LSIZE(bp)); 4754 ASSERT3U(HDR_GET_PSIZE(hdr), ==, BP_GET_PSIZE(bp)); 4755 } 4756 } 4757 4758 static void 4759 arc_read_done(zio_t *zio) 4760 { 4761 arc_buf_hdr_t *hdr = zio->io_private; 4762 kmutex_t *hash_lock = NULL; 4763 arc_callback_t *callback_list; 4764 arc_callback_t *acb; 4765 boolean_t freeable = B_FALSE; 4766 boolean_t no_zio_error = (zio->io_error == 0); 4767 4768 /* 4769 * The hdr was inserted into hash-table and removed from lists 4770 * prior to starting I/O. We should find this header, since 4771 * it's in the hash table, and it should be legit since it's 4772 * not possible to evict it during the I/O. The only possible 4773 * reason for it not to be found is if we were freed during the 4774 * read. 4775 */ 4776 if (HDR_IN_HASH_TABLE(hdr)) { 4777 ASSERT3U(hdr->b_birth, ==, BP_PHYSICAL_BIRTH(zio->io_bp)); 4778 ASSERT3U(hdr->b_dva.dva_word[0], ==, 4779 BP_IDENTITY(zio->io_bp)->dva_word[0]); 4780 ASSERT3U(hdr->b_dva.dva_word[1], ==, 4781 BP_IDENTITY(zio->io_bp)->dva_word[1]); 4782 4783 arc_buf_hdr_t *found = buf_hash_find(hdr->b_spa, zio->io_bp, 4784 &hash_lock); 4785 4786 ASSERT((found == hdr && 4787 DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) || 4788 (found == hdr && HDR_L2_READING(hdr))); 4789 ASSERT3P(hash_lock, !=, NULL); 4790 } 4791 4792 if (no_zio_error) { 4793 /* byteswap if necessary */ 4794 if (BP_SHOULD_BYTESWAP(zio->io_bp)) { 4795 if (BP_GET_LEVEL(zio->io_bp) > 0) { 4796 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_UINT64; 4797 } else { 4798 hdr->b_l1hdr.b_byteswap = 4799 DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp)); 4800 } 4801 } else { 4802 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS; 4803 } 4804 } 4805 4806 arc_hdr_clear_flags(hdr, ARC_FLAG_L2_EVICTED); 4807 if (l2arc_noprefetch && HDR_PREFETCH(hdr)) 4808 arc_hdr_clear_flags(hdr, ARC_FLAG_L2CACHE); 4809 4810 callback_list = hdr->b_l1hdr.b_acb; 4811 ASSERT3P(callback_list, !=, NULL); 4812 4813 if (hash_lock && no_zio_error && hdr->b_l1hdr.b_state == arc_anon) { 4814 /* 4815 * Only call arc_access on anonymous buffers. This is because 4816 * if we've issued an I/O for an evicted buffer, we've already 4817 * called arc_access (to prevent any simultaneous readers from 4818 * getting confused). 4819 */ 4820 arc_access(hdr, hash_lock); 4821 } 4822 4823 /* 4824 * If a read request has a callback (i.e. acb_done is not NULL), then we 4825 * make a buf containing the data according to the parameters which were 4826 * passed in. The implementation of arc_buf_alloc_impl() ensures that we 4827 * aren't needlessly decompressing the data multiple times. 4828 */ 4829 int callback_cnt = 0; 4830 for (acb = callback_list; acb != NULL; acb = acb->acb_next) { 4831 if (!acb->acb_done) 4832 continue; 4833 4834 /* This is a demand read since prefetches don't use callbacks */ 4835 callback_cnt++; 4836 4837 if (no_zio_error) { 4838 int error = arc_buf_alloc_impl(hdr, acb->acb_private, 4839 acb->acb_compressed, zio->io_error == 0, 4840 &acb->acb_buf); 4841 if (error != 0) { 4842 /* 4843 * Decompression failed. Set io_error 4844 * so that when we call acb_done (below), 4845 * we will indicate that the read failed. 4846 * Note that in the unusual case where one 4847 * callback is compressed and another 4848 * uncompressed, we will mark all of them 4849 * as failed, even though the uncompressed 4850 * one can't actually fail. In this case, 4851 * the hdr will not be anonymous, because 4852 * if there are multiple callbacks, it's 4853 * because multiple threads found the same 4854 * arc buf in the hash table. 4855 */ 4856 zio->io_error = error; 4857 } 4858 } 4859 } 4860 /* 4861 * If there are multiple callbacks, we must have the hash lock, 4862 * because the only way for multiple threads to find this hdr is 4863 * in the hash table. This ensures that if there are multiple 4864 * callbacks, the hdr is not anonymous. If it were anonymous, 4865 * we couldn't use arc_buf_destroy() in the error case below. 4866 */ 4867 ASSERT(callback_cnt < 2 || hash_lock != NULL); 4868 4869 hdr->b_l1hdr.b_acb = NULL; 4870 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS); 4871 if (callback_cnt == 0) { 4872 ASSERT(HDR_PREFETCH(hdr)); 4873 ASSERT0(hdr->b_l1hdr.b_bufcnt); 4874 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL); 4875 } 4876 4877 ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt) || 4878 callback_list != NULL); 4879 4880 if (no_zio_error) { 4881 arc_hdr_verify(hdr, zio->io_bp); 4882 } else { 4883 arc_hdr_set_flags(hdr, ARC_FLAG_IO_ERROR); 4884 if (hdr->b_l1hdr.b_state != arc_anon) 4885 arc_change_state(arc_anon, hdr, hash_lock); 4886 if (HDR_IN_HASH_TABLE(hdr)) 4887 buf_hash_remove(hdr); 4888 freeable = zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt); 4889 } 4890 4891 /* 4892 * Broadcast before we drop the hash_lock to avoid the possibility 4893 * that the hdr (and hence the cv) might be freed before we get to 4894 * the cv_broadcast(). 4895 */ 4896 cv_broadcast(&hdr->b_l1hdr.b_cv); 4897 4898 if (hash_lock != NULL) { 4899 mutex_exit(hash_lock); 4900 } else { 4901 /* 4902 * This block was freed while we waited for the read to 4903 * complete. It has been removed from the hash table and 4904 * moved to the anonymous state (so that it won't show up 4905 * in the cache). 4906 */ 4907 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon); 4908 freeable = zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt); 4909 } 4910 4911 /* execute each callback and free its structure */ 4912 while ((acb = callback_list) != NULL) { 4913 if (acb->acb_done != NULL) { 4914 if (zio->io_error != 0 && acb->acb_buf != NULL) { 4915 /* 4916 * If arc_buf_alloc_impl() fails during 4917 * decompression, the buf will still be 4918 * allocated, and needs to be freed here. 4919 */ 4920 arc_buf_destroy(acb->acb_buf, acb->acb_private); 4921 acb->acb_buf = NULL; 4922 } 4923 acb->acb_done(zio, acb->acb_buf, acb->acb_private); 4924 } 4925 4926 if (acb->acb_zio_dummy != NULL) { 4927 acb->acb_zio_dummy->io_error = zio->io_error; 4928 zio_nowait(acb->acb_zio_dummy); 4929 } 4930 4931 callback_list = acb->acb_next; 4932 kmem_free(acb, sizeof (arc_callback_t)); 4933 } 4934 4935 if (freeable) 4936 arc_hdr_destroy(hdr); 4937 } 4938 4939 /* 4940 * "Read" the block at the specified DVA (in bp) via the 4941 * cache. If the block is found in the cache, invoke the provided 4942 * callback immediately and return. Note that the `zio' parameter 4943 * in the callback will be NULL in this case, since no IO was 4944 * required. If the block is not in the cache pass the read request 4945 * on to the spa with a substitute callback function, so that the 4946 * requested block will be added to the cache. 4947 * 4948 * If a read request arrives for a block that has a read in-progress, 4949 * either wait for the in-progress read to complete (and return the 4950 * results); or, if this is a read with a "done" func, add a record 4951 * to the read to invoke the "done" func when the read completes, 4952 * and return; or just return. 4953 * 4954 * arc_read_done() will invoke all the requested "done" functions 4955 * for readers of this block. 4956 */ 4957 int 4958 arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_done_func_t *done, 4959 void *private, zio_priority_t priority, int zio_flags, 4960 arc_flags_t *arc_flags, const zbookmark_phys_t *zb) 4961 { 4962 arc_buf_hdr_t *hdr = NULL; 4963 kmutex_t *hash_lock = NULL; 4964 zio_t *rzio; 4965 uint64_t guid = spa_load_guid(spa); 4966 boolean_t compressed_read = (zio_flags & ZIO_FLAG_RAW) != 0; 4967 4968 ASSERT(!BP_IS_EMBEDDED(bp) || 4969 BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA); 4970 4971 top: 4972 if (!BP_IS_EMBEDDED(bp)) { 4973 /* 4974 * Embedded BP's have no DVA and require no I/O to "read". 4975 * Create an anonymous arc buf to back it. 4976 */ 4977 hdr = buf_hash_find(guid, bp, &hash_lock); 4978 } 4979 4980 if (hdr != NULL && HDR_HAS_L1HDR(hdr) && hdr->b_l1hdr.b_pabd != NULL) { 4981 arc_buf_t *buf = NULL; 4982 *arc_flags |= ARC_FLAG_CACHED; 4983 4984 if (HDR_IO_IN_PROGRESS(hdr)) { 4985 4986 if ((hdr->b_flags & ARC_FLAG_PRIO_ASYNC_READ) && 4987 priority == ZIO_PRIORITY_SYNC_READ) { 4988 /* 4989 * This sync read must wait for an 4990 * in-progress async read (e.g. a predictive 4991 * prefetch). Async reads are queued 4992 * separately at the vdev_queue layer, so 4993 * this is a form of priority inversion. 4994 * Ideally, we would "inherit" the demand 4995 * i/o's priority by moving the i/o from 4996 * the async queue to the synchronous queue, 4997 * but there is currently no mechanism to do 4998 * so. Track this so that we can evaluate 4999 * the magnitude of this potential performance 5000 * problem. 5001 * 5002 * Note that if the prefetch i/o is already 5003 * active (has been issued to the device), 5004 * the prefetch improved performance, because 5005 * we issued it sooner than we would have 5006 * without the prefetch. 5007 */ 5008 DTRACE_PROBE1(arc__sync__wait__for__async, 5009 arc_buf_hdr_t *, hdr); 5010 ARCSTAT_BUMP(arcstat_sync_wait_for_async); 5011 } 5012 if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) { 5013 arc_hdr_clear_flags(hdr, 5014 ARC_FLAG_PREDICTIVE_PREFETCH); 5015 } 5016 5017 if (*arc_flags & ARC_FLAG_WAIT) { 5018 cv_wait(&hdr->b_l1hdr.b_cv, hash_lock); 5019 mutex_exit(hash_lock); 5020 goto top; 5021 } 5022 ASSERT(*arc_flags & ARC_FLAG_NOWAIT); 5023 5024 if (done) { 5025 arc_callback_t *acb = NULL; 5026 5027 acb = kmem_zalloc(sizeof (arc_callback_t), 5028 KM_SLEEP); 5029 acb->acb_done = done; 5030 acb->acb_private = private; 5031 acb->acb_compressed = compressed_read; 5032 if (pio != NULL) 5033 acb->acb_zio_dummy = zio_null(pio, 5034 spa, NULL, NULL, NULL, zio_flags); 5035 5036 ASSERT3P(acb->acb_done, !=, NULL); 5037 acb->acb_next = hdr->b_l1hdr.b_acb; 5038 hdr->b_l1hdr.b_acb = acb; 5039 mutex_exit(hash_lock); 5040 return (0); 5041 } 5042 mutex_exit(hash_lock); 5043 return (0); 5044 } 5045 5046 ASSERT(hdr->b_l1hdr.b_state == arc_mru || 5047 hdr->b_l1hdr.b_state == arc_mfu); 5048 5049 if (done) { 5050 if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) { 5051 /* 5052 * This is a demand read which does not have to 5053 * wait for i/o because we did a predictive 5054 * prefetch i/o for it, which has completed. 5055 */ 5056 DTRACE_PROBE1( 5057 arc__demand__hit__predictive__prefetch, 5058 arc_buf_hdr_t *, hdr); 5059 ARCSTAT_BUMP( 5060 arcstat_demand_hit_predictive_prefetch); 5061 arc_hdr_clear_flags(hdr, 5062 ARC_FLAG_PREDICTIVE_PREFETCH); 5063 } 5064 ASSERT(!BP_IS_EMBEDDED(bp) || !BP_IS_HOLE(bp)); 5065 5066 /* Get a buf with the desired data in it. */ 5067 VERIFY0(arc_buf_alloc_impl(hdr, private, 5068 compressed_read, B_TRUE, &buf)); 5069 } else if (*arc_flags & ARC_FLAG_PREFETCH && 5070 zfs_refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) { 5071 arc_hdr_set_flags(hdr, ARC_FLAG_PREFETCH); 5072 } 5073 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr); 5074 arc_access(hdr, hash_lock); 5075 if (*arc_flags & ARC_FLAG_L2CACHE) 5076 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE); 5077 mutex_exit(hash_lock); 5078 ARCSTAT_BUMP(arcstat_hits); 5079 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr), 5080 demand, prefetch, !HDR_ISTYPE_METADATA(hdr), 5081 data, metadata, hits); 5082 5083 if (done) 5084 done(NULL, buf, private); 5085 } else { 5086 uint64_t lsize = BP_GET_LSIZE(bp); 5087 uint64_t psize = BP_GET_PSIZE(bp); 5088 arc_callback_t *acb; 5089 vdev_t *vd = NULL; 5090 uint64_t addr = 0; 5091 boolean_t devw = B_FALSE; 5092 uint64_t size; 5093 5094 if (hdr == NULL) { 5095 /* this block is not in the cache */ 5096 arc_buf_hdr_t *exists = NULL; 5097 arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp); 5098 hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize, 5099 BP_GET_COMPRESS(bp), type); 5100 5101 if (!BP_IS_EMBEDDED(bp)) { 5102 hdr->b_dva = *BP_IDENTITY(bp); 5103 hdr->b_birth = BP_PHYSICAL_BIRTH(bp); 5104 exists = buf_hash_insert(hdr, &hash_lock); 5105 } 5106 if (exists != NULL) { 5107 /* somebody beat us to the hash insert */ 5108 mutex_exit(hash_lock); 5109 buf_discard_identity(hdr); 5110 arc_hdr_destroy(hdr); 5111 goto top; /* restart the IO request */ 5112 } 5113 } else { 5114 /* 5115 * This block is in the ghost cache. If it was L2-only 5116 * (and thus didn't have an L1 hdr), we realloc the 5117 * header to add an L1 hdr. 5118 */ 5119 if (!HDR_HAS_L1HDR(hdr)) { 5120 hdr = arc_hdr_realloc(hdr, hdr_l2only_cache, 5121 hdr_full_cache); 5122 } 5123 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL); 5124 ASSERT(GHOST_STATE(hdr->b_l1hdr.b_state)); 5125 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 5126 ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 5127 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 5128 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL); 5129 5130 /* 5131 * This is a delicate dance that we play here. 5132 * This hdr is in the ghost list so we access it 5133 * to move it out of the ghost list before we 5134 * initiate the read. If it's a prefetch then 5135 * it won't have a callback so we'll remove the 5136 * reference that arc_buf_alloc_impl() created. We 5137 * do this after we've called arc_access() to 5138 * avoid hitting an assert in remove_reference(). 5139 */ 5140 arc_access(hdr, hash_lock); 5141 arc_hdr_alloc_pabd(hdr); 5142 } 5143 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL); 5144 size = arc_hdr_size(hdr); 5145 5146 /* 5147 * If compression is enabled on the hdr, then will do 5148 * RAW I/O and will store the compressed data in the hdr's 5149 * data block. Otherwise, the hdr's data block will contain 5150 * the uncompressed data. 5151 */ 5152 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF) { 5153 zio_flags |= ZIO_FLAG_RAW; 5154 } 5155 5156 if (*arc_flags & ARC_FLAG_PREFETCH) 5157 arc_hdr_set_flags(hdr, ARC_FLAG_PREFETCH); 5158 if (*arc_flags & ARC_FLAG_L2CACHE) 5159 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE); 5160 if (BP_GET_LEVEL(bp) > 0) 5161 arc_hdr_set_flags(hdr, ARC_FLAG_INDIRECT); 5162 if (*arc_flags & ARC_FLAG_PREDICTIVE_PREFETCH) 5163 arc_hdr_set_flags(hdr, ARC_FLAG_PREDICTIVE_PREFETCH); 5164 ASSERT(!GHOST_STATE(hdr->b_l1hdr.b_state)); 5165 5166 acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP); 5167 acb->acb_done = done; 5168 acb->acb_private = private; 5169 acb->acb_compressed = compressed_read; 5170 5171 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL); 5172 hdr->b_l1hdr.b_acb = acb; 5173 arc_hdr_set_flags(hdr, ARC_FLAG_IO_IN_PROGRESS); 5174 5175 if (HDR_HAS_L2HDR(hdr) && 5176 (vd = hdr->b_l2hdr.b_dev->l2ad_vdev) != NULL) { 5177 devw = hdr->b_l2hdr.b_dev->l2ad_writing; 5178 addr = hdr->b_l2hdr.b_daddr; 5179 /* 5180 * Lock out L2ARC device removal. 5181 */ 5182 if (vdev_is_dead(vd) || 5183 !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER)) 5184 vd = NULL; 5185 } 5186 5187 if (priority == ZIO_PRIORITY_ASYNC_READ) 5188 arc_hdr_set_flags(hdr, ARC_FLAG_PRIO_ASYNC_READ); 5189 else 5190 arc_hdr_clear_flags(hdr, ARC_FLAG_PRIO_ASYNC_READ); 5191 5192 if (hash_lock != NULL) 5193 mutex_exit(hash_lock); 5194 5195 /* 5196 * At this point, we have a level 1 cache miss. Try again in 5197 * L2ARC if possible. 5198 */ 5199 ASSERT3U(HDR_GET_LSIZE(hdr), ==, lsize); 5200 5201 DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp, 5202 uint64_t, lsize, zbookmark_phys_t *, zb); 5203 ARCSTAT_BUMP(arcstat_misses); 5204 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr), 5205 demand, prefetch, !HDR_ISTYPE_METADATA(hdr), 5206 data, metadata, misses); 5207 5208 if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) { 5209 /* 5210 * Read from the L2ARC if the following are true: 5211 * 1. The L2ARC vdev was previously cached. 5212 * 2. This buffer still has L2ARC metadata. 5213 * 3. This buffer isn't currently writing to the L2ARC. 5214 * 4. The L2ARC entry wasn't evicted, which may 5215 * also have invalidated the vdev. 5216 * 5. This isn't prefetch and l2arc_noprefetch is set. 5217 */ 5218 if (HDR_HAS_L2HDR(hdr) && 5219 !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) && 5220 !(l2arc_noprefetch && HDR_PREFETCH(hdr))) { 5221 l2arc_read_callback_t *cb; 5222 abd_t *abd; 5223 uint64_t asize; 5224 5225 DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr); 5226 ARCSTAT_BUMP(arcstat_l2_hits); 5227 5228 cb = kmem_zalloc(sizeof (l2arc_read_callback_t), 5229 KM_SLEEP); 5230 cb->l2rcb_hdr = hdr; 5231 cb->l2rcb_bp = *bp; 5232 cb->l2rcb_zb = *zb; 5233 cb->l2rcb_flags = zio_flags; 5234 5235 asize = vdev_psize_to_asize(vd, size); 5236 if (asize != size) { 5237 abd = abd_alloc_for_io(asize, 5238 HDR_ISTYPE_METADATA(hdr)); 5239 cb->l2rcb_abd = abd; 5240 } else { 5241 abd = hdr->b_l1hdr.b_pabd; 5242 } 5243 5244 ASSERT(addr >= VDEV_LABEL_START_SIZE && 5245 addr + asize <= vd->vdev_psize - 5246 VDEV_LABEL_END_SIZE); 5247 5248 /* 5249 * l2arc read. The SCL_L2ARC lock will be 5250 * released by l2arc_read_done(). 5251 * Issue a null zio if the underlying buffer 5252 * was squashed to zero size by compression. 5253 */ 5254 ASSERT3U(HDR_GET_COMPRESS(hdr), !=, 5255 ZIO_COMPRESS_EMPTY); 5256 rzio = zio_read_phys(pio, vd, addr, 5257 asize, abd, 5258 ZIO_CHECKSUM_OFF, 5259 l2arc_read_done, cb, priority, 5260 zio_flags | ZIO_FLAG_DONT_CACHE | 5261 ZIO_FLAG_CANFAIL | 5262 ZIO_FLAG_DONT_PROPAGATE | 5263 ZIO_FLAG_DONT_RETRY, B_FALSE); 5264 DTRACE_PROBE2(l2arc__read, vdev_t *, vd, 5265 zio_t *, rzio); 5266 ARCSTAT_INCR(arcstat_l2_read_bytes, size); 5267 5268 if (*arc_flags & ARC_FLAG_NOWAIT) { 5269 zio_nowait(rzio); 5270 return (0); 5271 } 5272 5273 ASSERT(*arc_flags & ARC_FLAG_WAIT); 5274 if (zio_wait(rzio) == 0) 5275 return (0); 5276 5277 /* l2arc read error; goto zio_read() */ 5278 } else { 5279 DTRACE_PROBE1(l2arc__miss, 5280 arc_buf_hdr_t *, hdr); 5281 ARCSTAT_BUMP(arcstat_l2_misses); 5282 if (HDR_L2_WRITING(hdr)) 5283 ARCSTAT_BUMP(arcstat_l2_rw_clash); 5284 spa_config_exit(spa, SCL_L2ARC, vd); 5285 } 5286 } else { 5287 if (vd != NULL) 5288 spa_config_exit(spa, SCL_L2ARC, vd); 5289 if (l2arc_ndev != 0) { 5290 DTRACE_PROBE1(l2arc__miss, 5291 arc_buf_hdr_t *, hdr); 5292 ARCSTAT_BUMP(arcstat_l2_misses); 5293 } 5294 } 5295 5296 rzio = zio_read(pio, spa, bp, hdr->b_l1hdr.b_pabd, size, 5297 arc_read_done, hdr, priority, zio_flags, zb); 5298 5299 if (*arc_flags & ARC_FLAG_WAIT) 5300 return (zio_wait(rzio)); 5301 5302 ASSERT(*arc_flags & ARC_FLAG_NOWAIT); 5303 zio_nowait(rzio); 5304 } 5305 return (0); 5306 } 5307 5308 /* 5309 * Notify the arc that a block was freed, and thus will never be used again. 5310 */ 5311 void 5312 arc_freed(spa_t *spa, const blkptr_t *bp) 5313 { 5314 arc_buf_hdr_t *hdr; 5315 kmutex_t *hash_lock; 5316 uint64_t guid = spa_load_guid(spa); 5317 5318 ASSERT(!BP_IS_EMBEDDED(bp)); 5319 5320 hdr = buf_hash_find(guid, bp, &hash_lock); 5321 if (hdr == NULL) 5322 return; 5323 5324 /* 5325 * We might be trying to free a block that is still doing I/O 5326 * (i.e. prefetch) or has a reference (i.e. a dedup-ed, 5327 * dmu_sync-ed block). If this block is being prefetched, then it 5328 * would still have the ARC_FLAG_IO_IN_PROGRESS flag set on the hdr 5329 * until the I/O completes. A block may also have a reference if it is 5330 * part of a dedup-ed, dmu_synced write. The dmu_sync() function would 5331 * have written the new block to its final resting place on disk but 5332 * without the dedup flag set. This would have left the hdr in the MRU 5333 * state and discoverable. When the txg finally syncs it detects that 5334 * the block was overridden in open context and issues an override I/O. 5335 * Since this is a dedup block, the override I/O will determine if the 5336 * block is already in the DDT. If so, then it will replace the io_bp 5337 * with the bp from the DDT and allow the I/O to finish. When the I/O 5338 * reaches the done callback, dbuf_write_override_done, it will 5339 * check to see if the io_bp and io_bp_override are identical. 5340 * If they are not, then it indicates that the bp was replaced with 5341 * the bp in the DDT and the override bp is freed. This allows 5342 * us to arrive here with a reference on a block that is being 5343 * freed. So if we have an I/O in progress, or a reference to 5344 * this hdr, then we don't destroy the hdr. 5345 */ 5346 if (!HDR_HAS_L1HDR(hdr) || (!HDR_IO_IN_PROGRESS(hdr) && 5347 zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt))) { 5348 arc_change_state(arc_anon, hdr, hash_lock); 5349 arc_hdr_destroy(hdr); 5350 mutex_exit(hash_lock); 5351 } else { 5352 mutex_exit(hash_lock); 5353 } 5354 5355 } 5356 5357 /* 5358 * Release this buffer from the cache, making it an anonymous buffer. This 5359 * must be done after a read and prior to modifying the buffer contents. 5360 * If the buffer has more than one reference, we must make 5361 * a new hdr for the buffer. 5362 */ 5363 void 5364 arc_release(arc_buf_t *buf, void *tag) 5365 { 5366 arc_buf_hdr_t *hdr = buf->b_hdr; 5367 5368 /* 5369 * It would be nice to assert that if it's DMU metadata (level > 5370 * 0 || it's the dnode file), then it must be syncing context. 5371 * But we don't know that information at this level. 5372 */ 5373 5374 mutex_enter(&buf->b_evict_lock); 5375 5376 ASSERT(HDR_HAS_L1HDR(hdr)); 5377 5378 /* 5379 * We don't grab the hash lock prior to this check, because if 5380 * the buffer's header is in the arc_anon state, it won't be 5381 * linked into the hash table. 5382 */ 5383 if (hdr->b_l1hdr.b_state == arc_anon) { 5384 mutex_exit(&buf->b_evict_lock); 5385 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 5386 ASSERT(!HDR_IN_HASH_TABLE(hdr)); 5387 ASSERT(!HDR_HAS_L2HDR(hdr)); 5388 ASSERT(HDR_EMPTY(hdr)); 5389 5390 ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1); 5391 ASSERT3S(zfs_refcount_count(&hdr->b_l1hdr.b_refcnt), ==, 1); 5392 ASSERT(!list_link_active(&hdr->b_l1hdr.b_arc_node)); 5393 5394 hdr->b_l1hdr.b_arc_access = 0; 5395 5396 /* 5397 * If the buf is being overridden then it may already 5398 * have a hdr that is not empty. 5399 */ 5400 buf_discard_identity(hdr); 5401 arc_buf_thaw(buf); 5402 5403 return; 5404 } 5405 5406 kmutex_t *hash_lock = HDR_LOCK(hdr); 5407 mutex_enter(hash_lock); 5408 5409 /* 5410 * This assignment is only valid as long as the hash_lock is 5411 * held, we must be careful not to reference state or the 5412 * b_state field after dropping the lock. 5413 */ 5414 arc_state_t *state = hdr->b_l1hdr.b_state; 5415 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 5416 ASSERT3P(state, !=, arc_anon); 5417 5418 /* this buffer is not on any list */ 5419 ASSERT3S(zfs_refcount_count(&hdr->b_l1hdr.b_refcnt), >, 0); 5420 5421 if (HDR_HAS_L2HDR(hdr)) { 5422 mutex_enter(&hdr->b_l2hdr.b_dev->l2ad_mtx); 5423 5424 /* 5425 * We have to recheck this conditional again now that 5426 * we're holding the l2ad_mtx to prevent a race with 5427 * another thread which might be concurrently calling 5428 * l2arc_evict(). In that case, l2arc_evict() might have 5429 * destroyed the header's L2 portion as we were waiting 5430 * to acquire the l2ad_mtx. 5431 */ 5432 if (HDR_HAS_L2HDR(hdr)) 5433 arc_hdr_l2hdr_destroy(hdr); 5434 5435 mutex_exit(&hdr->b_l2hdr.b_dev->l2ad_mtx); 5436 } 5437 5438 /* 5439 * Do we have more than one buf? 5440 */ 5441 if (hdr->b_l1hdr.b_bufcnt > 1) { 5442 arc_buf_hdr_t *nhdr; 5443 uint64_t spa = hdr->b_spa; 5444 uint64_t psize = HDR_GET_PSIZE(hdr); 5445 uint64_t lsize = HDR_GET_LSIZE(hdr); 5446 enum zio_compress compress = HDR_GET_COMPRESS(hdr); 5447 arc_buf_contents_t type = arc_buf_type(hdr); 5448 VERIFY3U(hdr->b_type, ==, type); 5449 5450 ASSERT(hdr->b_l1hdr.b_buf != buf || buf->b_next != NULL); 5451 (void) remove_reference(hdr, hash_lock, tag); 5452 5453 if (arc_buf_is_shared(buf) && !ARC_BUF_COMPRESSED(buf)) { 5454 ASSERT3P(hdr->b_l1hdr.b_buf, !=, buf); 5455 ASSERT(ARC_BUF_LAST(buf)); 5456 } 5457 5458 /* 5459 * Pull the data off of this hdr and attach it to 5460 * a new anonymous hdr. Also find the last buffer 5461 * in the hdr's buffer list. 5462 */ 5463 arc_buf_t *lastbuf = arc_buf_remove(hdr, buf); 5464 ASSERT3P(lastbuf, !=, NULL); 5465 5466 /* 5467 * If the current arc_buf_t and the hdr are sharing their data 5468 * buffer, then we must stop sharing that block. 5469 */ 5470 if (arc_buf_is_shared(buf)) { 5471 VERIFY(!arc_buf_is_shared(lastbuf)); 5472 5473 /* 5474 * First, sever the block sharing relationship between 5475 * buf and the arc_buf_hdr_t. 5476 */ 5477 arc_unshare_buf(hdr, buf); 5478 5479 /* 5480 * Now we need to recreate the hdr's b_pabd. Since we 5481 * have lastbuf handy, we try to share with it, but if 5482 * we can't then we allocate a new b_pabd and copy the 5483 * data from buf into it. 5484 */ 5485 if (arc_can_share(hdr, lastbuf)) { 5486 arc_share_buf(hdr, lastbuf); 5487 } else { 5488 arc_hdr_alloc_pabd(hdr); 5489 abd_copy_from_buf(hdr->b_l1hdr.b_pabd, 5490 buf->b_data, psize); 5491 } 5492 VERIFY3P(lastbuf->b_data, !=, NULL); 5493 } else if (HDR_SHARED_DATA(hdr)) { 5494 /* 5495 * Uncompressed shared buffers are always at the end 5496 * of the list. Compressed buffers don't have the 5497 * same requirements. This makes it hard to 5498 * simply assert that the lastbuf is shared so 5499 * we rely on the hdr's compression flags to determine 5500 * if we have a compressed, shared buffer. 5501 */ 5502 ASSERT(arc_buf_is_shared(lastbuf) || 5503 HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF); 5504 ASSERT(!ARC_BUF_SHARED(buf)); 5505 } 5506 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL); 5507 ASSERT3P(state, !=, arc_l2c_only); 5508 5509 (void) zfs_refcount_remove_many(&state->arcs_size, 5510 arc_buf_size(buf), buf); 5511 5512 if (zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt)) { 5513 ASSERT3P(state, !=, arc_l2c_only); 5514 (void) zfs_refcount_remove_many( 5515 &state->arcs_esize[type], 5516 arc_buf_size(buf), buf); 5517 } 5518 5519 hdr->b_l1hdr.b_bufcnt -= 1; 5520 arc_cksum_verify(buf); 5521 arc_buf_unwatch(buf); 5522 5523 mutex_exit(hash_lock); 5524 5525 /* 5526 * Allocate a new hdr. The new hdr will contain a b_pabd 5527 * buffer which will be freed in arc_write(). 5528 */ 5529 nhdr = arc_hdr_alloc(spa, psize, lsize, compress, type); 5530 ASSERT3P(nhdr->b_l1hdr.b_buf, ==, NULL); 5531 ASSERT0(nhdr->b_l1hdr.b_bufcnt); 5532 ASSERT0(zfs_refcount_count(&nhdr->b_l1hdr.b_refcnt)); 5533 VERIFY3U(nhdr->b_type, ==, type); 5534 ASSERT(!HDR_SHARED_DATA(nhdr)); 5535 5536 nhdr->b_l1hdr.b_buf = buf; 5537 nhdr->b_l1hdr.b_bufcnt = 1; 5538 (void) zfs_refcount_add(&nhdr->b_l1hdr.b_refcnt, tag); 5539 buf->b_hdr = nhdr; 5540 5541 mutex_exit(&buf->b_evict_lock); 5542 (void) zfs_refcount_add_many(&arc_anon->arcs_size, 5543 arc_buf_size(buf), buf); 5544 } else { 5545 mutex_exit(&buf->b_evict_lock); 5546 ASSERT(zfs_refcount_count(&hdr->b_l1hdr.b_refcnt) == 1); 5547 /* protected by hash lock, or hdr is on arc_anon */ 5548 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 5549 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 5550 arc_change_state(arc_anon, hdr, hash_lock); 5551 hdr->b_l1hdr.b_arc_access = 0; 5552 mutex_exit(hash_lock); 5553 5554 buf_discard_identity(hdr); 5555 arc_buf_thaw(buf); 5556 } 5557 } 5558 5559 int 5560 arc_released(arc_buf_t *buf) 5561 { 5562 int released; 5563 5564 mutex_enter(&buf->b_evict_lock); 5565 released = (buf->b_data != NULL && 5566 buf->b_hdr->b_l1hdr.b_state == arc_anon); 5567 mutex_exit(&buf->b_evict_lock); 5568 return (released); 5569 } 5570 5571 #ifdef ZFS_DEBUG 5572 int 5573 arc_referenced(arc_buf_t *buf) 5574 { 5575 int referenced; 5576 5577 mutex_enter(&buf->b_evict_lock); 5578 referenced = (zfs_refcount_count(&buf->b_hdr->b_l1hdr.b_refcnt)); 5579 mutex_exit(&buf->b_evict_lock); 5580 return (referenced); 5581 } 5582 #endif 5583 5584 static void 5585 arc_write_ready(zio_t *zio) 5586 { 5587 arc_write_callback_t *callback = zio->io_private; 5588 arc_buf_t *buf = callback->awcb_buf; 5589 arc_buf_hdr_t *hdr = buf->b_hdr; 5590 uint64_t psize = BP_IS_HOLE(zio->io_bp) ? 0 : BP_GET_PSIZE(zio->io_bp); 5591 5592 ASSERT(HDR_HAS_L1HDR(hdr)); 5593 ASSERT(!zfs_refcount_is_zero(&buf->b_hdr->b_l1hdr.b_refcnt)); 5594 ASSERT(hdr->b_l1hdr.b_bufcnt > 0); 5595 5596 /* 5597 * If we're reexecuting this zio because the pool suspended, then 5598 * cleanup any state that was previously set the first time the 5599 * callback was invoked. 5600 */ 5601 if (zio->io_flags & ZIO_FLAG_REEXECUTED) { 5602 arc_cksum_free(hdr); 5603 arc_buf_unwatch(buf); 5604 if (hdr->b_l1hdr.b_pabd != NULL) { 5605 if (arc_buf_is_shared(buf)) { 5606 arc_unshare_buf(hdr, buf); 5607 } else { 5608 arc_hdr_free_pabd(hdr); 5609 } 5610 } 5611 } 5612 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL); 5613 ASSERT(!HDR_SHARED_DATA(hdr)); 5614 ASSERT(!arc_buf_is_shared(buf)); 5615 5616 callback->awcb_ready(zio, buf, callback->awcb_private); 5617 5618 if (HDR_IO_IN_PROGRESS(hdr)) 5619 ASSERT(zio->io_flags & ZIO_FLAG_REEXECUTED); 5620 5621 arc_cksum_compute(buf); 5622 arc_hdr_set_flags(hdr, ARC_FLAG_IO_IN_PROGRESS); 5623 5624 enum zio_compress compress; 5625 if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) { 5626 compress = ZIO_COMPRESS_OFF; 5627 } else { 5628 ASSERT3U(HDR_GET_LSIZE(hdr), ==, BP_GET_LSIZE(zio->io_bp)); 5629 compress = BP_GET_COMPRESS(zio->io_bp); 5630 } 5631 HDR_SET_PSIZE(hdr, psize); 5632 arc_hdr_set_compress(hdr, compress); 5633 5634 5635 /* 5636 * Fill the hdr with data. If the hdr is compressed, the data we want 5637 * is available from the zio, otherwise we can take it from the buf. 5638 * 5639 * We might be able to share the buf's data with the hdr here. However, 5640 * doing so would cause the ARC to be full of linear ABDs if we write a 5641 * lot of shareable data. As a compromise, we check whether scattered 5642 * ABDs are allowed, and assume that if they are then the user wants 5643 * the ARC to be primarily filled with them regardless of the data being 5644 * written. Therefore, if they're allowed then we allocate one and copy 5645 * the data into it; otherwise, we share the data directly if we can. 5646 */ 5647 if (zfs_abd_scatter_enabled || !arc_can_share(hdr, buf)) { 5648 arc_hdr_alloc_pabd(hdr); 5649 5650 /* 5651 * Ideally, we would always copy the io_abd into b_pabd, but the 5652 * user may have disabled compressed ARC, thus we must check the 5653 * hdr's compression setting rather than the io_bp's. 5654 */ 5655 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF) { 5656 ASSERT3U(BP_GET_COMPRESS(zio->io_bp), !=, 5657 ZIO_COMPRESS_OFF); 5658 ASSERT3U(psize, >, 0); 5659 5660 abd_copy(hdr->b_l1hdr.b_pabd, zio->io_abd, psize); 5661 } else { 5662 ASSERT3U(zio->io_orig_size, ==, arc_hdr_size(hdr)); 5663 5664 abd_copy_from_buf(hdr->b_l1hdr.b_pabd, buf->b_data, 5665 arc_buf_size(buf)); 5666 } 5667 } else { 5668 ASSERT3P(buf->b_data, ==, abd_to_buf(zio->io_orig_abd)); 5669 ASSERT3U(zio->io_orig_size, ==, arc_buf_size(buf)); 5670 ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1); 5671 5672 arc_share_buf(hdr, buf); 5673 } 5674 5675 arc_hdr_verify(hdr, zio->io_bp); 5676 } 5677 5678 static void 5679 arc_write_children_ready(zio_t *zio) 5680 { 5681 arc_write_callback_t *callback = zio->io_private; 5682 arc_buf_t *buf = callback->awcb_buf; 5683 5684 callback->awcb_children_ready(zio, buf, callback->awcb_private); 5685 } 5686 5687 /* 5688 * The SPA calls this callback for each physical write that happens on behalf 5689 * of a logical write. See the comment in dbuf_write_physdone() for details. 5690 */ 5691 static void 5692 arc_write_physdone(zio_t *zio) 5693 { 5694 arc_write_callback_t *cb = zio->io_private; 5695 if (cb->awcb_physdone != NULL) 5696 cb->awcb_physdone(zio, cb->awcb_buf, cb->awcb_private); 5697 } 5698 5699 static void 5700 arc_write_done(zio_t *zio) 5701 { 5702 arc_write_callback_t *callback = zio->io_private; 5703 arc_buf_t *buf = callback->awcb_buf; 5704 arc_buf_hdr_t *hdr = buf->b_hdr; 5705 5706 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL); 5707 5708 if (zio->io_error == 0) { 5709 arc_hdr_verify(hdr, zio->io_bp); 5710 5711 if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) { 5712 buf_discard_identity(hdr); 5713 } else { 5714 hdr->b_dva = *BP_IDENTITY(zio->io_bp); 5715 hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp); 5716 } 5717 } else { 5718 ASSERT(HDR_EMPTY(hdr)); 5719 } 5720 5721 /* 5722 * If the block to be written was all-zero or compressed enough to be 5723 * embedded in the BP, no write was performed so there will be no 5724 * dva/birth/checksum. The buffer must therefore remain anonymous 5725 * (and uncached). 5726 */ 5727 if (!HDR_EMPTY(hdr)) { 5728 arc_buf_hdr_t *exists; 5729 kmutex_t *hash_lock; 5730 5731 ASSERT3U(zio->io_error, ==, 0); 5732 5733 arc_cksum_verify(buf); 5734 5735 exists = buf_hash_insert(hdr, &hash_lock); 5736 if (exists != NULL) { 5737 /* 5738 * This can only happen if we overwrite for 5739 * sync-to-convergence, because we remove 5740 * buffers from the hash table when we arc_free(). 5741 */ 5742 if (zio->io_flags & ZIO_FLAG_IO_REWRITE) { 5743 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) 5744 panic("bad overwrite, hdr=%p exists=%p", 5745 (void *)hdr, (void *)exists); 5746 ASSERT(zfs_refcount_is_zero( 5747 &exists->b_l1hdr.b_refcnt)); 5748 arc_change_state(arc_anon, exists, hash_lock); 5749 mutex_exit(hash_lock); 5750 arc_hdr_destroy(exists); 5751 exists = buf_hash_insert(hdr, &hash_lock); 5752 ASSERT3P(exists, ==, NULL); 5753 } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) { 5754 /* nopwrite */ 5755 ASSERT(zio->io_prop.zp_nopwrite); 5756 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) 5757 panic("bad nopwrite, hdr=%p exists=%p", 5758 (void *)hdr, (void *)exists); 5759 } else { 5760 /* Dedup */ 5761 ASSERT(hdr->b_l1hdr.b_bufcnt == 1); 5762 ASSERT(hdr->b_l1hdr.b_state == arc_anon); 5763 ASSERT(BP_GET_DEDUP(zio->io_bp)); 5764 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0); 5765 } 5766 } 5767 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS); 5768 /* if it's not anon, we are doing a scrub */ 5769 if (exists == NULL && hdr->b_l1hdr.b_state == arc_anon) 5770 arc_access(hdr, hash_lock); 5771 mutex_exit(hash_lock); 5772 } else { 5773 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS); 5774 } 5775 5776 ASSERT(!zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 5777 callback->awcb_done(zio, buf, callback->awcb_private); 5778 5779 abd_put(zio->io_abd); 5780 kmem_free(callback, sizeof (arc_write_callback_t)); 5781 } 5782 5783 zio_t * 5784 arc_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, arc_buf_t *buf, 5785 boolean_t l2arc, const zio_prop_t *zp, arc_done_func_t *ready, 5786 arc_done_func_t *children_ready, arc_done_func_t *physdone, 5787 arc_done_func_t *done, void *private, zio_priority_t priority, 5788 int zio_flags, const zbookmark_phys_t *zb) 5789 { 5790 arc_buf_hdr_t *hdr = buf->b_hdr; 5791 arc_write_callback_t *callback; 5792 zio_t *zio; 5793 zio_prop_t localprop = *zp; 5794 5795 ASSERT3P(ready, !=, NULL); 5796 ASSERT3P(done, !=, NULL); 5797 ASSERT(!HDR_IO_ERROR(hdr)); 5798 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 5799 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL); 5800 ASSERT3U(hdr->b_l1hdr.b_bufcnt, >, 0); 5801 if (l2arc) 5802 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE); 5803 if (ARC_BUF_COMPRESSED(buf)) { 5804 /* 5805 * We're writing a pre-compressed buffer. Make the 5806 * compression algorithm requested by the zio_prop_t match 5807 * the pre-compressed buffer's compression algorithm. 5808 */ 5809 localprop.zp_compress = HDR_GET_COMPRESS(hdr); 5810 5811 ASSERT3U(HDR_GET_LSIZE(hdr), !=, arc_buf_size(buf)); 5812 zio_flags |= ZIO_FLAG_RAW; 5813 } 5814 callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP); 5815 callback->awcb_ready = ready; 5816 callback->awcb_children_ready = children_ready; 5817 callback->awcb_physdone = physdone; 5818 callback->awcb_done = done; 5819 callback->awcb_private = private; 5820 callback->awcb_buf = buf; 5821 5822 /* 5823 * The hdr's b_pabd is now stale, free it now. A new data block 5824 * will be allocated when the zio pipeline calls arc_write_ready(). 5825 */ 5826 if (hdr->b_l1hdr.b_pabd != NULL) { 5827 /* 5828 * If the buf is currently sharing the data block with 5829 * the hdr then we need to break that relationship here. 5830 * The hdr will remain with a NULL data pointer and the 5831 * buf will take sole ownership of the block. 5832 */ 5833 if (arc_buf_is_shared(buf)) { 5834 arc_unshare_buf(hdr, buf); 5835 } else { 5836 arc_hdr_free_pabd(hdr); 5837 } 5838 VERIFY3P(buf->b_data, !=, NULL); 5839 arc_hdr_set_compress(hdr, ZIO_COMPRESS_OFF); 5840 } 5841 ASSERT(!arc_buf_is_shared(buf)); 5842 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL); 5843 5844 zio = zio_write(pio, spa, txg, bp, 5845 abd_get_from_buf(buf->b_data, HDR_GET_LSIZE(hdr)), 5846 HDR_GET_LSIZE(hdr), arc_buf_size(buf), &localprop, arc_write_ready, 5847 (children_ready != NULL) ? arc_write_children_ready : NULL, 5848 arc_write_physdone, arc_write_done, callback, 5849 priority, zio_flags, zb); 5850 5851 return (zio); 5852 } 5853 5854 static int 5855 arc_memory_throttle(spa_t *spa, uint64_t reserve, uint64_t txg) 5856 { 5857 #ifdef _KERNEL 5858 uint64_t available_memory = ptob(freemem); 5859 5860 #if defined(__i386) 5861 available_memory = 5862 MIN(available_memory, vmem_size(heap_arena, VMEM_FREE)); 5863 #endif 5864 5865 if (freemem > physmem * arc_lotsfree_percent / 100) 5866 return (0); 5867 5868 if (txg > spa->spa_lowmem_last_txg) { 5869 spa->spa_lowmem_last_txg = txg; 5870 spa->spa_lowmem_page_load = 0; 5871 } 5872 /* 5873 * If we are in pageout, we know that memory is already tight, 5874 * the arc is already going to be evicting, so we just want to 5875 * continue to let page writes occur as quickly as possible. 5876 */ 5877 if (curproc == proc_pageout) { 5878 if (spa->spa_lowmem_page_load > 5879 MAX(ptob(minfree), available_memory) / 4) 5880 return (SET_ERROR(ERESTART)); 5881 /* Note: reserve is inflated, so we deflate */ 5882 atomic_add_64(&spa->spa_lowmem_page_load, reserve / 8); 5883 return (0); 5884 } else if (spa->spa_lowmem_page_load > 0 && arc_reclaim_needed()) { 5885 /* memory is low, delay before restarting */ 5886 ARCSTAT_INCR(arcstat_memory_throttle_count, 1); 5887 return (SET_ERROR(EAGAIN)); 5888 } 5889 spa->spa_lowmem_page_load = 0; 5890 #endif /* _KERNEL */ 5891 return (0); 5892 } 5893 5894 void 5895 arc_tempreserve_clear(uint64_t reserve) 5896 { 5897 atomic_add_64(&arc_tempreserve, -reserve); 5898 ASSERT((int64_t)arc_tempreserve >= 0); 5899 } 5900 5901 int 5902 arc_tempreserve_space(spa_t *spa, uint64_t reserve, uint64_t txg) 5903 { 5904 int error; 5905 uint64_t anon_size; 5906 5907 if (reserve > arc_c/4 && !arc_no_grow) 5908 arc_c = MIN(arc_c_max, reserve * 4); 5909 if (reserve > arc_c) 5910 return (SET_ERROR(ENOMEM)); 5911 5912 /* 5913 * Don't count loaned bufs as in flight dirty data to prevent long 5914 * network delays from blocking transactions that are ready to be 5915 * assigned to a txg. 5916 */ 5917 5918 /* assert that it has not wrapped around */ 5919 ASSERT3S(atomic_add_64_nv(&arc_loaned_bytes, 0), >=, 0); 5920 5921 anon_size = MAX((int64_t)(zfs_refcount_count(&arc_anon->arcs_size) - 5922 arc_loaned_bytes), 0); 5923 5924 /* 5925 * Writes will, almost always, require additional memory allocations 5926 * in order to compress/encrypt/etc the data. We therefore need to 5927 * make sure that there is sufficient available memory for this. 5928 */ 5929 error = arc_memory_throttle(spa, reserve, txg); 5930 if (error != 0) 5931 return (error); 5932 5933 /* 5934 * Throttle writes when the amount of dirty data in the cache 5935 * gets too large. We try to keep the cache less than half full 5936 * of dirty blocks so that our sync times don't grow too large. 5937 * 5938 * In the case of one pool being built on another pool, we want 5939 * to make sure we don't end up throttling the lower (backing) 5940 * pool when the upper pool is the majority contributor to dirty 5941 * data. To insure we make forward progress during throttling, we 5942 * also check the current pool's net dirty data and only throttle 5943 * if it exceeds zfs_arc_pool_dirty_percent of the anonymous dirty 5944 * data in the cache. 5945 * 5946 * Note: if two requests come in concurrently, we might let them 5947 * both succeed, when one of them should fail. Not a huge deal. 5948 */ 5949 uint64_t total_dirty = reserve + arc_tempreserve + anon_size; 5950 uint64_t spa_dirty_anon = spa_dirty_data(spa); 5951 5952 if (total_dirty > arc_c * zfs_arc_dirty_limit_percent / 100 && 5953 anon_size > arc_c * zfs_arc_anon_limit_percent / 100 && 5954 spa_dirty_anon > anon_size * zfs_arc_pool_dirty_percent / 100) { 5955 uint64_t meta_esize = 5956 zfs_refcount_count( 5957 &arc_anon->arcs_esize[ARC_BUFC_METADATA]); 5958 uint64_t data_esize = 5959 zfs_refcount_count(&arc_anon->arcs_esize[ARC_BUFC_DATA]); 5960 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK " 5961 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n", 5962 arc_tempreserve >> 10, meta_esize >> 10, 5963 data_esize >> 10, reserve >> 10, arc_c >> 10); 5964 return (SET_ERROR(ERESTART)); 5965 } 5966 atomic_add_64(&arc_tempreserve, reserve); 5967 return (0); 5968 } 5969 5970 static void 5971 arc_kstat_update_state(arc_state_t *state, kstat_named_t *size, 5972 kstat_named_t *evict_data, kstat_named_t *evict_metadata) 5973 { 5974 size->value.ui64 = zfs_refcount_count(&state->arcs_size); 5975 evict_data->value.ui64 = 5976 zfs_refcount_count(&state->arcs_esize[ARC_BUFC_DATA]); 5977 evict_metadata->value.ui64 = 5978 zfs_refcount_count(&state->arcs_esize[ARC_BUFC_METADATA]); 5979 } 5980 5981 static int 5982 arc_kstat_update(kstat_t *ksp, int rw) 5983 { 5984 arc_stats_t *as = ksp->ks_data; 5985 5986 if (rw == KSTAT_WRITE) { 5987 return (EACCES); 5988 } else { 5989 arc_kstat_update_state(arc_anon, 5990 &as->arcstat_anon_size, 5991 &as->arcstat_anon_evictable_data, 5992 &as->arcstat_anon_evictable_metadata); 5993 arc_kstat_update_state(arc_mru, 5994 &as->arcstat_mru_size, 5995 &as->arcstat_mru_evictable_data, 5996 &as->arcstat_mru_evictable_metadata); 5997 arc_kstat_update_state(arc_mru_ghost, 5998 &as->arcstat_mru_ghost_size, 5999 &as->arcstat_mru_ghost_evictable_data, 6000 &as->arcstat_mru_ghost_evictable_metadata); 6001 arc_kstat_update_state(arc_mfu, 6002 &as->arcstat_mfu_size, 6003 &as->arcstat_mfu_evictable_data, 6004 &as->arcstat_mfu_evictable_metadata); 6005 arc_kstat_update_state(arc_mfu_ghost, 6006 &as->arcstat_mfu_ghost_size, 6007 &as->arcstat_mfu_ghost_evictable_data, 6008 &as->arcstat_mfu_ghost_evictable_metadata); 6009 6010 ARCSTAT(arcstat_size) = aggsum_value(&arc_size); 6011 ARCSTAT(arcstat_meta_used) = aggsum_value(&arc_meta_used); 6012 ARCSTAT(arcstat_data_size) = aggsum_value(&astat_data_size); 6013 ARCSTAT(arcstat_metadata_size) = 6014 aggsum_value(&astat_metadata_size); 6015 ARCSTAT(arcstat_hdr_size) = aggsum_value(&astat_hdr_size); 6016 ARCSTAT(arcstat_other_size) = aggsum_value(&astat_other_size); 6017 ARCSTAT(arcstat_l2_hdr_size) = aggsum_value(&astat_l2_hdr_size); 6018 } 6019 6020 return (0); 6021 } 6022 6023 /* 6024 * This function *must* return indices evenly distributed between all 6025 * sublists of the multilist. This is needed due to how the ARC eviction 6026 * code is laid out; arc_evict_state() assumes ARC buffers are evenly 6027 * distributed between all sublists and uses this assumption when 6028 * deciding which sublist to evict from and how much to evict from it. 6029 */ 6030 unsigned int 6031 arc_state_multilist_index_func(multilist_t *ml, void *obj) 6032 { 6033 arc_buf_hdr_t *hdr = obj; 6034 6035 /* 6036 * We rely on b_dva to generate evenly distributed index 6037 * numbers using buf_hash below. So, as an added precaution, 6038 * let's make sure we never add empty buffers to the arc lists. 6039 */ 6040 ASSERT(!HDR_EMPTY(hdr)); 6041 6042 /* 6043 * The assumption here, is the hash value for a given 6044 * arc_buf_hdr_t will remain constant throughout it's lifetime 6045 * (i.e. it's b_spa, b_dva, and b_birth fields don't change). 6046 * Thus, we don't need to store the header's sublist index 6047 * on insertion, as this index can be recalculated on removal. 6048 * 6049 * Also, the low order bits of the hash value are thought to be 6050 * distributed evenly. Otherwise, in the case that the multilist 6051 * has a power of two number of sublists, each sublists' usage 6052 * would not be evenly distributed. 6053 */ 6054 return (buf_hash(hdr->b_spa, &hdr->b_dva, hdr->b_birth) % 6055 multilist_get_num_sublists(ml)); 6056 } 6057 6058 static void 6059 arc_state_init(void) 6060 { 6061 arc_anon = &ARC_anon; 6062 arc_mru = &ARC_mru; 6063 arc_mru_ghost = &ARC_mru_ghost; 6064 arc_mfu = &ARC_mfu; 6065 arc_mfu_ghost = &ARC_mfu_ghost; 6066 arc_l2c_only = &ARC_l2c_only; 6067 6068 arc_mru->arcs_list[ARC_BUFC_METADATA] = 6069 multilist_create(sizeof (arc_buf_hdr_t), 6070 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 6071 arc_state_multilist_index_func); 6072 arc_mru->arcs_list[ARC_BUFC_DATA] = 6073 multilist_create(sizeof (arc_buf_hdr_t), 6074 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 6075 arc_state_multilist_index_func); 6076 arc_mru_ghost->arcs_list[ARC_BUFC_METADATA] = 6077 multilist_create(sizeof (arc_buf_hdr_t), 6078 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 6079 arc_state_multilist_index_func); 6080 arc_mru_ghost->arcs_list[ARC_BUFC_DATA] = 6081 multilist_create(sizeof (arc_buf_hdr_t), 6082 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 6083 arc_state_multilist_index_func); 6084 arc_mfu->arcs_list[ARC_BUFC_METADATA] = 6085 multilist_create(sizeof (arc_buf_hdr_t), 6086 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 6087 arc_state_multilist_index_func); 6088 arc_mfu->arcs_list[ARC_BUFC_DATA] = 6089 multilist_create(sizeof (arc_buf_hdr_t), 6090 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 6091 arc_state_multilist_index_func); 6092 arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA] = 6093 multilist_create(sizeof (arc_buf_hdr_t), 6094 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 6095 arc_state_multilist_index_func); 6096 arc_mfu_ghost->arcs_list[ARC_BUFC_DATA] = 6097 multilist_create(sizeof (arc_buf_hdr_t), 6098 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 6099 arc_state_multilist_index_func); 6100 arc_l2c_only->arcs_list[ARC_BUFC_METADATA] = 6101 multilist_create(sizeof (arc_buf_hdr_t), 6102 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 6103 arc_state_multilist_index_func); 6104 arc_l2c_only->arcs_list[ARC_BUFC_DATA] = 6105 multilist_create(sizeof (arc_buf_hdr_t), 6106 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 6107 arc_state_multilist_index_func); 6108 6109 zfs_refcount_create(&arc_anon->arcs_esize[ARC_BUFC_METADATA]); 6110 zfs_refcount_create(&arc_anon->arcs_esize[ARC_BUFC_DATA]); 6111 zfs_refcount_create(&arc_mru->arcs_esize[ARC_BUFC_METADATA]); 6112 zfs_refcount_create(&arc_mru->arcs_esize[ARC_BUFC_DATA]); 6113 zfs_refcount_create(&arc_mru_ghost->arcs_esize[ARC_BUFC_METADATA]); 6114 zfs_refcount_create(&arc_mru_ghost->arcs_esize[ARC_BUFC_DATA]); 6115 zfs_refcount_create(&arc_mfu->arcs_esize[ARC_BUFC_METADATA]); 6116 zfs_refcount_create(&arc_mfu->arcs_esize[ARC_BUFC_DATA]); 6117 zfs_refcount_create(&arc_mfu_ghost->arcs_esize[ARC_BUFC_METADATA]); 6118 zfs_refcount_create(&arc_mfu_ghost->arcs_esize[ARC_BUFC_DATA]); 6119 zfs_refcount_create(&arc_l2c_only->arcs_esize[ARC_BUFC_METADATA]); 6120 zfs_refcount_create(&arc_l2c_only->arcs_esize[ARC_BUFC_DATA]); 6121 6122 zfs_refcount_create(&arc_anon->arcs_size); 6123 zfs_refcount_create(&arc_mru->arcs_size); 6124 zfs_refcount_create(&arc_mru_ghost->arcs_size); 6125 zfs_refcount_create(&arc_mfu->arcs_size); 6126 zfs_refcount_create(&arc_mfu_ghost->arcs_size); 6127 zfs_refcount_create(&arc_l2c_only->arcs_size); 6128 6129 aggsum_init(&arc_meta_used, 0); 6130 aggsum_init(&arc_size, 0); 6131 aggsum_init(&astat_data_size, 0); 6132 aggsum_init(&astat_metadata_size, 0); 6133 aggsum_init(&astat_hdr_size, 0); 6134 aggsum_init(&astat_other_size, 0); 6135 aggsum_init(&astat_l2_hdr_size, 0); 6136 } 6137 6138 static void 6139 arc_state_fini(void) 6140 { 6141 zfs_refcount_destroy(&arc_anon->arcs_esize[ARC_BUFC_METADATA]); 6142 zfs_refcount_destroy(&arc_anon->arcs_esize[ARC_BUFC_DATA]); 6143 zfs_refcount_destroy(&arc_mru->arcs_esize[ARC_BUFC_METADATA]); 6144 zfs_refcount_destroy(&arc_mru->arcs_esize[ARC_BUFC_DATA]); 6145 zfs_refcount_destroy(&arc_mru_ghost->arcs_esize[ARC_BUFC_METADATA]); 6146 zfs_refcount_destroy(&arc_mru_ghost->arcs_esize[ARC_BUFC_DATA]); 6147 zfs_refcount_destroy(&arc_mfu->arcs_esize[ARC_BUFC_METADATA]); 6148 zfs_refcount_destroy(&arc_mfu->arcs_esize[ARC_BUFC_DATA]); 6149 zfs_refcount_destroy(&arc_mfu_ghost->arcs_esize[ARC_BUFC_METADATA]); 6150 zfs_refcount_destroy(&arc_mfu_ghost->arcs_esize[ARC_BUFC_DATA]); 6151 zfs_refcount_destroy(&arc_l2c_only->arcs_esize[ARC_BUFC_METADATA]); 6152 zfs_refcount_destroy(&arc_l2c_only->arcs_esize[ARC_BUFC_DATA]); 6153 6154 zfs_refcount_destroy(&arc_anon->arcs_size); 6155 zfs_refcount_destroy(&arc_mru->arcs_size); 6156 zfs_refcount_destroy(&arc_mru_ghost->arcs_size); 6157 zfs_refcount_destroy(&arc_mfu->arcs_size); 6158 zfs_refcount_destroy(&arc_mfu_ghost->arcs_size); 6159 zfs_refcount_destroy(&arc_l2c_only->arcs_size); 6160 6161 multilist_destroy(arc_mru->arcs_list[ARC_BUFC_METADATA]); 6162 multilist_destroy(arc_mru_ghost->arcs_list[ARC_BUFC_METADATA]); 6163 multilist_destroy(arc_mfu->arcs_list[ARC_BUFC_METADATA]); 6164 multilist_destroy(arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA]); 6165 multilist_destroy(arc_mru->arcs_list[ARC_BUFC_DATA]); 6166 multilist_destroy(arc_mru_ghost->arcs_list[ARC_BUFC_DATA]); 6167 multilist_destroy(arc_mfu->arcs_list[ARC_BUFC_DATA]); 6168 multilist_destroy(arc_mfu_ghost->arcs_list[ARC_BUFC_DATA]); 6169 6170 aggsum_fini(&arc_meta_used); 6171 aggsum_fini(&arc_size); 6172 aggsum_fini(&astat_data_size); 6173 aggsum_fini(&astat_metadata_size); 6174 aggsum_fini(&astat_hdr_size); 6175 aggsum_fini(&astat_other_size); 6176 aggsum_fini(&astat_l2_hdr_size); 6177 } 6178 6179 uint64_t 6180 arc_max_bytes(void) 6181 { 6182 return (arc_c_max); 6183 } 6184 6185 void 6186 arc_init(void) 6187 { 6188 /* 6189 * allmem is "all memory that we could possibly use". 6190 */ 6191 #ifdef _KERNEL 6192 uint64_t allmem = ptob(physmem - swapfs_minfree); 6193 #else 6194 uint64_t allmem = (physmem * PAGESIZE) / 2; 6195 #endif 6196 mutex_init(&arc_adjust_lock, NULL, MUTEX_DEFAULT, NULL); 6197 cv_init(&arc_adjust_waiters_cv, NULL, CV_DEFAULT, NULL); 6198 6199 /* Convert seconds to clock ticks */ 6200 arc_min_prefetch_lifespan = 1 * hz; 6201 6202 /* set min cache to 1/32 of all memory, or 64MB, whichever is more */ 6203 arc_c_min = MAX(allmem / 32, 64 << 20); 6204 /* set max to 3/4 of all memory, or all but 1GB, whichever is more */ 6205 if (allmem >= 1 << 30) 6206 arc_c_max = allmem - (1 << 30); 6207 else 6208 arc_c_max = arc_c_min; 6209 arc_c_max = MAX(allmem * 3 / 4, arc_c_max); 6210 6211 /* 6212 * In userland, there's only the memory pressure that we artificially 6213 * create (see arc_available_memory()). Don't let arc_c get too 6214 * small, because it can cause transactions to be larger than 6215 * arc_c, causing arc_tempreserve_space() to fail. 6216 */ 6217 #ifndef _KERNEL 6218 arc_c_min = arc_c_max / 2; 6219 #endif 6220 6221 /* 6222 * Allow the tunables to override our calculations if they are 6223 * reasonable (ie. over 64MB) 6224 */ 6225 if (zfs_arc_max > 64 << 20 && zfs_arc_max < allmem) { 6226 arc_c_max = zfs_arc_max; 6227 arc_c_min = MIN(arc_c_min, arc_c_max); 6228 } 6229 if (zfs_arc_min > 64 << 20 && zfs_arc_min <= arc_c_max) 6230 arc_c_min = zfs_arc_min; 6231 6232 arc_c = arc_c_max; 6233 arc_p = (arc_c >> 1); 6234 6235 /* limit meta-data to 1/4 of the arc capacity */ 6236 arc_meta_limit = arc_c_max / 4; 6237 6238 #ifdef _KERNEL 6239 /* 6240 * Metadata is stored in the kernel's heap. Don't let us 6241 * use more than half the heap for the ARC. 6242 */ 6243 arc_meta_limit = MIN(arc_meta_limit, 6244 vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 2); 6245 #endif 6246 6247 /* Allow the tunable to override if it is reasonable */ 6248 if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max) 6249 arc_meta_limit = zfs_arc_meta_limit; 6250 6251 if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0) 6252 arc_c_min = arc_meta_limit / 2; 6253 6254 if (zfs_arc_meta_min > 0) { 6255 arc_meta_min = zfs_arc_meta_min; 6256 } else { 6257 arc_meta_min = arc_c_min / 2; 6258 } 6259 6260 if (zfs_arc_grow_retry > 0) 6261 arc_grow_retry = zfs_arc_grow_retry; 6262 6263 if (zfs_arc_shrink_shift > 0) 6264 arc_shrink_shift = zfs_arc_shrink_shift; 6265 6266 /* 6267 * Ensure that arc_no_grow_shift is less than arc_shrink_shift. 6268 */ 6269 if (arc_no_grow_shift >= arc_shrink_shift) 6270 arc_no_grow_shift = arc_shrink_shift - 1; 6271 6272 if (zfs_arc_p_min_shift > 0) 6273 arc_p_min_shift = zfs_arc_p_min_shift; 6274 6275 /* if kmem_flags are set, lets try to use less memory */ 6276 if (kmem_debugging()) 6277 arc_c = arc_c / 2; 6278 if (arc_c < arc_c_min) 6279 arc_c = arc_c_min; 6280 6281 arc_state_init(); 6282 6283 /* 6284 * The arc must be "uninitialized", so that hdr_recl() (which is 6285 * registered by buf_init()) will not access arc_reap_zthr before 6286 * it is created. 6287 */ 6288 ASSERT(!arc_initialized); 6289 buf_init(); 6290 6291 arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED, 6292 sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL); 6293 6294 if (arc_ksp != NULL) { 6295 arc_ksp->ks_data = &arc_stats; 6296 arc_ksp->ks_update = arc_kstat_update; 6297 kstat_install(arc_ksp); 6298 } 6299 6300 arc_adjust_zthr = zthr_create(arc_adjust_cb_check, 6301 arc_adjust_cb, NULL); 6302 arc_reap_zthr = zthr_create_timer(arc_reap_cb_check, 6303 arc_reap_cb, NULL, SEC2NSEC(1)); 6304 6305 arc_initialized = B_TRUE; 6306 arc_warm = B_FALSE; 6307 6308 /* 6309 * Calculate maximum amount of dirty data per pool. 6310 * 6311 * If it has been set by /etc/system, take that. 6312 * Otherwise, use a percentage of physical memory defined by 6313 * zfs_dirty_data_max_percent (default 10%) with a cap at 6314 * zfs_dirty_data_max_max (default 4GB). 6315 */ 6316 if (zfs_dirty_data_max == 0) { 6317 zfs_dirty_data_max = physmem * PAGESIZE * 6318 zfs_dirty_data_max_percent / 100; 6319 zfs_dirty_data_max = MIN(zfs_dirty_data_max, 6320 zfs_dirty_data_max_max); 6321 } 6322 } 6323 6324 void 6325 arc_fini(void) 6326 { 6327 /* Use B_TRUE to ensure *all* buffers are evicted */ 6328 arc_flush(NULL, B_TRUE); 6329 6330 arc_initialized = B_FALSE; 6331 6332 if (arc_ksp != NULL) { 6333 kstat_delete(arc_ksp); 6334 arc_ksp = NULL; 6335 } 6336 6337 (void) zthr_cancel(arc_adjust_zthr); 6338 zthr_destroy(arc_adjust_zthr); 6339 6340 (void) zthr_cancel(arc_reap_zthr); 6341 zthr_destroy(arc_reap_zthr); 6342 6343 mutex_destroy(&arc_adjust_lock); 6344 cv_destroy(&arc_adjust_waiters_cv); 6345 6346 /* 6347 * buf_fini() must proceed arc_state_fini() because buf_fin() may 6348 * trigger the release of kmem magazines, which can callback to 6349 * arc_space_return() which accesses aggsums freed in act_state_fini(). 6350 */ 6351 buf_fini(); 6352 arc_state_fini(); 6353 6354 ASSERT0(arc_loaned_bytes); 6355 } 6356 6357 /* 6358 * Level 2 ARC 6359 * 6360 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk. 6361 * It uses dedicated storage devices to hold cached data, which are populated 6362 * using large infrequent writes. The main role of this cache is to boost 6363 * the performance of random read workloads. The intended L2ARC devices 6364 * include short-stroked disks, solid state disks, and other media with 6365 * substantially faster read latency than disk. 6366 * 6367 * +-----------------------+ 6368 * | ARC | 6369 * +-----------------------+ 6370 * | ^ ^ 6371 * | | | 6372 * l2arc_feed_thread() arc_read() 6373 * | | | 6374 * | l2arc read | 6375 * V | | 6376 * +---------------+ | 6377 * | L2ARC | | 6378 * +---------------+ | 6379 * | ^ | 6380 * l2arc_write() | | 6381 * | | | 6382 * V | | 6383 * +-------+ +-------+ 6384 * | vdev | | vdev | 6385 * | cache | | cache | 6386 * +-------+ +-------+ 6387 * +=========+ .-----. 6388 * : L2ARC : |-_____-| 6389 * : devices : | Disks | 6390 * +=========+ `-_____-' 6391 * 6392 * Read requests are satisfied from the following sources, in order: 6393 * 6394 * 1) ARC 6395 * 2) vdev cache of L2ARC devices 6396 * 3) L2ARC devices 6397 * 4) vdev cache of disks 6398 * 5) disks 6399 * 6400 * Some L2ARC device types exhibit extremely slow write performance. 6401 * To accommodate for this there are some significant differences between 6402 * the L2ARC and traditional cache design: 6403 * 6404 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from 6405 * the ARC behave as usual, freeing buffers and placing headers on ghost 6406 * lists. The ARC does not send buffers to the L2ARC during eviction as 6407 * this would add inflated write latencies for all ARC memory pressure. 6408 * 6409 * 2. The L2ARC attempts to cache data from the ARC before it is evicted. 6410 * It does this by periodically scanning buffers from the eviction-end of 6411 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are 6412 * not already there. It scans until a headroom of buffers is satisfied, 6413 * which itself is a buffer for ARC eviction. If a compressible buffer is 6414 * found during scanning and selected for writing to an L2ARC device, we 6415 * temporarily boost scanning headroom during the next scan cycle to make 6416 * sure we adapt to compression effects (which might significantly reduce 6417 * the data volume we write to L2ARC). The thread that does this is 6418 * l2arc_feed_thread(), illustrated below; example sizes are included to 6419 * provide a better sense of ratio than this diagram: 6420 * 6421 * head --> tail 6422 * +---------------------+----------+ 6423 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC 6424 * +---------------------+----------+ | o L2ARC eligible 6425 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer 6426 * +---------------------+----------+ | 6427 * 15.9 Gbytes ^ 32 Mbytes | 6428 * headroom | 6429 * l2arc_feed_thread() 6430 * | 6431 * l2arc write hand <--[oooo]--' 6432 * | 8 Mbyte 6433 * | write max 6434 * V 6435 * +==============================+ 6436 * L2ARC dev |####|#|###|###| |####| ... | 6437 * +==============================+ 6438 * 32 Gbytes 6439 * 6440 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of 6441 * evicted, then the L2ARC has cached a buffer much sooner than it probably 6442 * needed to, potentially wasting L2ARC device bandwidth and storage. It is 6443 * safe to say that this is an uncommon case, since buffers at the end of 6444 * the ARC lists have moved there due to inactivity. 6445 * 6446 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom, 6447 * then the L2ARC simply misses copying some buffers. This serves as a 6448 * pressure valve to prevent heavy read workloads from both stalling the ARC 6449 * with waits and clogging the L2ARC with writes. This also helps prevent 6450 * the potential for the L2ARC to churn if it attempts to cache content too 6451 * quickly, such as during backups of the entire pool. 6452 * 6453 * 5. After system boot and before the ARC has filled main memory, there are 6454 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru 6455 * lists can remain mostly static. Instead of searching from tail of these 6456 * lists as pictured, the l2arc_feed_thread() will search from the list heads 6457 * for eligible buffers, greatly increasing its chance of finding them. 6458 * 6459 * The L2ARC device write speed is also boosted during this time so that 6460 * the L2ARC warms up faster. Since there have been no ARC evictions yet, 6461 * there are no L2ARC reads, and no fear of degrading read performance 6462 * through increased writes. 6463 * 6464 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that 6465 * the vdev queue can aggregate them into larger and fewer writes. Each 6466 * device is written to in a rotor fashion, sweeping writes through 6467 * available space then repeating. 6468 * 6469 * 7. The L2ARC does not store dirty content. It never needs to flush 6470 * write buffers back to disk based storage. 6471 * 6472 * 8. If an ARC buffer is written (and dirtied) which also exists in the 6473 * L2ARC, the now stale L2ARC buffer is immediately dropped. 6474 * 6475 * The performance of the L2ARC can be tweaked by a number of tunables, which 6476 * may be necessary for different workloads: 6477 * 6478 * l2arc_write_max max write bytes per interval 6479 * l2arc_write_boost extra write bytes during device warmup 6480 * l2arc_noprefetch skip caching prefetched buffers 6481 * l2arc_headroom number of max device writes to precache 6482 * l2arc_headroom_boost when we find compressed buffers during ARC 6483 * scanning, we multiply headroom by this 6484 * percentage factor for the next scan cycle, 6485 * since more compressed buffers are likely to 6486 * be present 6487 * l2arc_feed_secs seconds between L2ARC writing 6488 * 6489 * Tunables may be removed or added as future performance improvements are 6490 * integrated, and also may become zpool properties. 6491 * 6492 * There are three key functions that control how the L2ARC warms up: 6493 * 6494 * l2arc_write_eligible() check if a buffer is eligible to cache 6495 * l2arc_write_size() calculate how much to write 6496 * l2arc_write_interval() calculate sleep delay between writes 6497 * 6498 * These three functions determine what to write, how much, and how quickly 6499 * to send writes. 6500 */ 6501 6502 static boolean_t 6503 l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *hdr) 6504 { 6505 /* 6506 * A buffer is *not* eligible for the L2ARC if it: 6507 * 1. belongs to a different spa. 6508 * 2. is already cached on the L2ARC. 6509 * 3. has an I/O in progress (it may be an incomplete read). 6510 * 4. is flagged not eligible (zfs property). 6511 */ 6512 if (hdr->b_spa != spa_guid || HDR_HAS_L2HDR(hdr) || 6513 HDR_IO_IN_PROGRESS(hdr) || !HDR_L2CACHE(hdr)) 6514 return (B_FALSE); 6515 6516 return (B_TRUE); 6517 } 6518 6519 static uint64_t 6520 l2arc_write_size(void) 6521 { 6522 uint64_t size; 6523 6524 /* 6525 * Make sure our globals have meaningful values in case the user 6526 * altered them. 6527 */ 6528 size = l2arc_write_max; 6529 if (size == 0) { 6530 cmn_err(CE_NOTE, "Bad value for l2arc_write_max, value must " 6531 "be greater than zero, resetting it to the default (%d)", 6532 L2ARC_WRITE_SIZE); 6533 size = l2arc_write_max = L2ARC_WRITE_SIZE; 6534 } 6535 6536 if (arc_warm == B_FALSE) 6537 size += l2arc_write_boost; 6538 6539 return (size); 6540 6541 } 6542 6543 static clock_t 6544 l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote) 6545 { 6546 clock_t interval, next, now; 6547 6548 /* 6549 * If the ARC lists are busy, increase our write rate; if the 6550 * lists are stale, idle back. This is achieved by checking 6551 * how much we previously wrote - if it was more than half of 6552 * what we wanted, schedule the next write much sooner. 6553 */ 6554 if (l2arc_feed_again && wrote > (wanted / 2)) 6555 interval = (hz * l2arc_feed_min_ms) / 1000; 6556 else 6557 interval = hz * l2arc_feed_secs; 6558 6559 now = ddi_get_lbolt(); 6560 next = MAX(now, MIN(now + interval, began + interval)); 6561 6562 return (next); 6563 } 6564 6565 /* 6566 * Cycle through L2ARC devices. This is how L2ARC load balances. 6567 * If a device is returned, this also returns holding the spa config lock. 6568 */ 6569 static l2arc_dev_t * 6570 l2arc_dev_get_next(void) 6571 { 6572 l2arc_dev_t *first, *next = NULL; 6573 6574 /* 6575 * Lock out the removal of spas (spa_namespace_lock), then removal 6576 * of cache devices (l2arc_dev_mtx). Once a device has been selected, 6577 * both locks will be dropped and a spa config lock held instead. 6578 */ 6579 mutex_enter(&spa_namespace_lock); 6580 mutex_enter(&l2arc_dev_mtx); 6581 6582 /* if there are no vdevs, there is nothing to do */ 6583 if (l2arc_ndev == 0) 6584 goto out; 6585 6586 first = NULL; 6587 next = l2arc_dev_last; 6588 do { 6589 /* loop around the list looking for a non-faulted vdev */ 6590 if (next == NULL) { 6591 next = list_head(l2arc_dev_list); 6592 } else { 6593 next = list_next(l2arc_dev_list, next); 6594 if (next == NULL) 6595 next = list_head(l2arc_dev_list); 6596 } 6597 6598 /* if we have come back to the start, bail out */ 6599 if (first == NULL) 6600 first = next; 6601 else if (next == first) 6602 break; 6603 6604 } while (vdev_is_dead(next->l2ad_vdev)); 6605 6606 /* if we were unable to find any usable vdevs, return NULL */ 6607 if (vdev_is_dead(next->l2ad_vdev)) 6608 next = NULL; 6609 6610 l2arc_dev_last = next; 6611 6612 out: 6613 mutex_exit(&l2arc_dev_mtx); 6614 6615 /* 6616 * Grab the config lock to prevent the 'next' device from being 6617 * removed while we are writing to it. 6618 */ 6619 if (next != NULL) 6620 spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER); 6621 mutex_exit(&spa_namespace_lock); 6622 6623 return (next); 6624 } 6625 6626 /* 6627 * Free buffers that were tagged for destruction. 6628 */ 6629 static void 6630 l2arc_do_free_on_write() 6631 { 6632 list_t *buflist; 6633 l2arc_data_free_t *df, *df_prev; 6634 6635 mutex_enter(&l2arc_free_on_write_mtx); 6636 buflist = l2arc_free_on_write; 6637 6638 for (df = list_tail(buflist); df; df = df_prev) { 6639 df_prev = list_prev(buflist, df); 6640 ASSERT3P(df->l2df_abd, !=, NULL); 6641 abd_free(df->l2df_abd); 6642 list_remove(buflist, df); 6643 kmem_free(df, sizeof (l2arc_data_free_t)); 6644 } 6645 6646 mutex_exit(&l2arc_free_on_write_mtx); 6647 } 6648 6649 /* 6650 * A write to a cache device has completed. Update all headers to allow 6651 * reads from these buffers to begin. 6652 */ 6653 static void 6654 l2arc_write_done(zio_t *zio) 6655 { 6656 l2arc_write_callback_t *cb; 6657 l2arc_dev_t *dev; 6658 list_t *buflist; 6659 arc_buf_hdr_t *head, *hdr, *hdr_prev; 6660 kmutex_t *hash_lock; 6661 int64_t bytes_dropped = 0; 6662 6663 cb = zio->io_private; 6664 ASSERT3P(cb, !=, NULL); 6665 dev = cb->l2wcb_dev; 6666 ASSERT3P(dev, !=, NULL); 6667 head = cb->l2wcb_head; 6668 ASSERT3P(head, !=, NULL); 6669 buflist = &dev->l2ad_buflist; 6670 ASSERT3P(buflist, !=, NULL); 6671 DTRACE_PROBE2(l2arc__iodone, zio_t *, zio, 6672 l2arc_write_callback_t *, cb); 6673 6674 if (zio->io_error != 0) 6675 ARCSTAT_BUMP(arcstat_l2_writes_error); 6676 6677 /* 6678 * All writes completed, or an error was hit. 6679 */ 6680 top: 6681 mutex_enter(&dev->l2ad_mtx); 6682 for (hdr = list_prev(buflist, head); hdr; hdr = hdr_prev) { 6683 hdr_prev = list_prev(buflist, hdr); 6684 6685 hash_lock = HDR_LOCK(hdr); 6686 6687 /* 6688 * We cannot use mutex_enter or else we can deadlock 6689 * with l2arc_write_buffers (due to swapping the order 6690 * the hash lock and l2ad_mtx are taken). 6691 */ 6692 if (!mutex_tryenter(hash_lock)) { 6693 /* 6694 * Missed the hash lock. We must retry so we 6695 * don't leave the ARC_FLAG_L2_WRITING bit set. 6696 */ 6697 ARCSTAT_BUMP(arcstat_l2_writes_lock_retry); 6698 6699 /* 6700 * We don't want to rescan the headers we've 6701 * already marked as having been written out, so 6702 * we reinsert the head node so we can pick up 6703 * where we left off. 6704 */ 6705 list_remove(buflist, head); 6706 list_insert_after(buflist, hdr, head); 6707 6708 mutex_exit(&dev->l2ad_mtx); 6709 6710 /* 6711 * We wait for the hash lock to become available 6712 * to try and prevent busy waiting, and increase 6713 * the chance we'll be able to acquire the lock 6714 * the next time around. 6715 */ 6716 mutex_enter(hash_lock); 6717 mutex_exit(hash_lock); 6718 goto top; 6719 } 6720 6721 /* 6722 * We could not have been moved into the arc_l2c_only 6723 * state while in-flight due to our ARC_FLAG_L2_WRITING 6724 * bit being set. Let's just ensure that's being enforced. 6725 */ 6726 ASSERT(HDR_HAS_L1HDR(hdr)); 6727 6728 if (zio->io_error != 0) { 6729 /* 6730 * Error - drop L2ARC entry. 6731 */ 6732 list_remove(buflist, hdr); 6733 arc_hdr_clear_flags(hdr, ARC_FLAG_HAS_L2HDR); 6734 6735 ARCSTAT_INCR(arcstat_l2_psize, -arc_hdr_size(hdr)); 6736 ARCSTAT_INCR(arcstat_l2_lsize, -HDR_GET_LSIZE(hdr)); 6737 6738 bytes_dropped += arc_hdr_size(hdr); 6739 (void) zfs_refcount_remove_many(&dev->l2ad_alloc, 6740 arc_hdr_size(hdr), hdr); 6741 } 6742 6743 /* 6744 * Allow ARC to begin reads and ghost list evictions to 6745 * this L2ARC entry. 6746 */ 6747 arc_hdr_clear_flags(hdr, ARC_FLAG_L2_WRITING); 6748 6749 mutex_exit(hash_lock); 6750 } 6751 6752 atomic_inc_64(&l2arc_writes_done); 6753 list_remove(buflist, head); 6754 ASSERT(!HDR_HAS_L1HDR(head)); 6755 kmem_cache_free(hdr_l2only_cache, head); 6756 mutex_exit(&dev->l2ad_mtx); 6757 6758 vdev_space_update(dev->l2ad_vdev, -bytes_dropped, 0, 0); 6759 6760 l2arc_do_free_on_write(); 6761 6762 kmem_free(cb, sizeof (l2arc_write_callback_t)); 6763 } 6764 6765 /* 6766 * A read to a cache device completed. Validate buffer contents before 6767 * handing over to the regular ARC routines. 6768 */ 6769 static void 6770 l2arc_read_done(zio_t *zio) 6771 { 6772 l2arc_read_callback_t *cb; 6773 arc_buf_hdr_t *hdr; 6774 kmutex_t *hash_lock; 6775 boolean_t valid_cksum; 6776 6777 ASSERT3P(zio->io_vd, !=, NULL); 6778 ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE); 6779 6780 spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd); 6781 6782 cb = zio->io_private; 6783 ASSERT3P(cb, !=, NULL); 6784 hdr = cb->l2rcb_hdr; 6785 ASSERT3P(hdr, !=, NULL); 6786 6787 hash_lock = HDR_LOCK(hdr); 6788 mutex_enter(hash_lock); 6789 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 6790 6791 /* 6792 * If the data was read into a temporary buffer, 6793 * move it and free the buffer. 6794 */ 6795 if (cb->l2rcb_abd != NULL) { 6796 ASSERT3U(arc_hdr_size(hdr), <, zio->io_size); 6797 if (zio->io_error == 0) { 6798 abd_copy(hdr->b_l1hdr.b_pabd, cb->l2rcb_abd, 6799 arc_hdr_size(hdr)); 6800 } 6801 6802 /* 6803 * The following must be done regardless of whether 6804 * there was an error: 6805 * - free the temporary buffer 6806 * - point zio to the real ARC buffer 6807 * - set zio size accordingly 6808 * These are required because zio is either re-used for 6809 * an I/O of the block in the case of the error 6810 * or the zio is passed to arc_read_done() and it 6811 * needs real data. 6812 */ 6813 abd_free(cb->l2rcb_abd); 6814 zio->io_size = zio->io_orig_size = arc_hdr_size(hdr); 6815 zio->io_abd = zio->io_orig_abd = hdr->b_l1hdr.b_pabd; 6816 } 6817 6818 ASSERT3P(zio->io_abd, !=, NULL); 6819 6820 /* 6821 * Check this survived the L2ARC journey. 6822 */ 6823 ASSERT3P(zio->io_abd, ==, hdr->b_l1hdr.b_pabd); 6824 zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */ 6825 zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */ 6826 6827 valid_cksum = arc_cksum_is_equal(hdr, zio); 6828 if (valid_cksum && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) { 6829 mutex_exit(hash_lock); 6830 zio->io_private = hdr; 6831 arc_read_done(zio); 6832 } else { 6833 mutex_exit(hash_lock); 6834 /* 6835 * Buffer didn't survive caching. Increment stats and 6836 * reissue to the original storage device. 6837 */ 6838 if (zio->io_error != 0) { 6839 ARCSTAT_BUMP(arcstat_l2_io_error); 6840 } else { 6841 zio->io_error = SET_ERROR(EIO); 6842 } 6843 if (!valid_cksum) 6844 ARCSTAT_BUMP(arcstat_l2_cksum_bad); 6845 6846 /* 6847 * If there's no waiter, issue an async i/o to the primary 6848 * storage now. If there *is* a waiter, the caller must 6849 * issue the i/o in a context where it's OK to block. 6850 */ 6851 if (zio->io_waiter == NULL) { 6852 zio_t *pio = zio_unique_parent(zio); 6853 6854 ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL); 6855 6856 zio_nowait(zio_read(pio, zio->io_spa, zio->io_bp, 6857 hdr->b_l1hdr.b_pabd, zio->io_size, arc_read_done, 6858 hdr, zio->io_priority, cb->l2rcb_flags, 6859 &cb->l2rcb_zb)); 6860 } 6861 } 6862 6863 kmem_free(cb, sizeof (l2arc_read_callback_t)); 6864 } 6865 6866 /* 6867 * This is the list priority from which the L2ARC will search for pages to 6868 * cache. This is used within loops (0..3) to cycle through lists in the 6869 * desired order. This order can have a significant effect on cache 6870 * performance. 6871 * 6872 * Currently the metadata lists are hit first, MFU then MRU, followed by 6873 * the data lists. This function returns a locked list, and also returns 6874 * the lock pointer. 6875 */ 6876 static multilist_sublist_t * 6877 l2arc_sublist_lock(int list_num) 6878 { 6879 multilist_t *ml = NULL; 6880 unsigned int idx; 6881 6882 ASSERT(list_num >= 0 && list_num <= 3); 6883 6884 switch (list_num) { 6885 case 0: 6886 ml = arc_mfu->arcs_list[ARC_BUFC_METADATA]; 6887 break; 6888 case 1: 6889 ml = arc_mru->arcs_list[ARC_BUFC_METADATA]; 6890 break; 6891 case 2: 6892 ml = arc_mfu->arcs_list[ARC_BUFC_DATA]; 6893 break; 6894 case 3: 6895 ml = arc_mru->arcs_list[ARC_BUFC_DATA]; 6896 break; 6897 } 6898 6899 /* 6900 * Return a randomly-selected sublist. This is acceptable 6901 * because the caller feeds only a little bit of data for each 6902 * call (8MB). Subsequent calls will result in different 6903 * sublists being selected. 6904 */ 6905 idx = multilist_get_random_index(ml); 6906 return (multilist_sublist_lock(ml, idx)); 6907 } 6908 6909 /* 6910 * Evict buffers from the device write hand to the distance specified in 6911 * bytes. This distance may span populated buffers, it may span nothing. 6912 * This is clearing a region on the L2ARC device ready for writing. 6913 * If the 'all' boolean is set, every buffer is evicted. 6914 */ 6915 static void 6916 l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all) 6917 { 6918 list_t *buflist; 6919 arc_buf_hdr_t *hdr, *hdr_prev; 6920 kmutex_t *hash_lock; 6921 uint64_t taddr; 6922 6923 buflist = &dev->l2ad_buflist; 6924 6925 if (!all && dev->l2ad_first) { 6926 /* 6927 * This is the first sweep through the device. There is 6928 * nothing to evict. 6929 */ 6930 return; 6931 } 6932 6933 if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) { 6934 /* 6935 * When nearing the end of the device, evict to the end 6936 * before the device write hand jumps to the start. 6937 */ 6938 taddr = dev->l2ad_end; 6939 } else { 6940 taddr = dev->l2ad_hand + distance; 6941 } 6942 DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist, 6943 uint64_t, taddr, boolean_t, all); 6944 6945 top: 6946 mutex_enter(&dev->l2ad_mtx); 6947 for (hdr = list_tail(buflist); hdr; hdr = hdr_prev) { 6948 hdr_prev = list_prev(buflist, hdr); 6949 6950 hash_lock = HDR_LOCK(hdr); 6951 6952 /* 6953 * We cannot use mutex_enter or else we can deadlock 6954 * with l2arc_write_buffers (due to swapping the order 6955 * the hash lock and l2ad_mtx are taken). 6956 */ 6957 if (!mutex_tryenter(hash_lock)) { 6958 /* 6959 * Missed the hash lock. Retry. 6960 */ 6961 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry); 6962 mutex_exit(&dev->l2ad_mtx); 6963 mutex_enter(hash_lock); 6964 mutex_exit(hash_lock); 6965 goto top; 6966 } 6967 6968 /* 6969 * A header can't be on this list if it doesn't have L2 header. 6970 */ 6971 ASSERT(HDR_HAS_L2HDR(hdr)); 6972 6973 /* Ensure this header has finished being written. */ 6974 ASSERT(!HDR_L2_WRITING(hdr)); 6975 ASSERT(!HDR_L2_WRITE_HEAD(hdr)); 6976 6977 if (!all && (hdr->b_l2hdr.b_daddr >= taddr || 6978 hdr->b_l2hdr.b_daddr < dev->l2ad_hand)) { 6979 /* 6980 * We've evicted to the target address, 6981 * or the end of the device. 6982 */ 6983 mutex_exit(hash_lock); 6984 break; 6985 } 6986 6987 if (!HDR_HAS_L1HDR(hdr)) { 6988 ASSERT(!HDR_L2_READING(hdr)); 6989 /* 6990 * This doesn't exist in the ARC. Destroy. 6991 * arc_hdr_destroy() will call list_remove() 6992 * and decrement arcstat_l2_lsize. 6993 */ 6994 arc_change_state(arc_anon, hdr, hash_lock); 6995 arc_hdr_destroy(hdr); 6996 } else { 6997 ASSERT(hdr->b_l1hdr.b_state != arc_l2c_only); 6998 ARCSTAT_BUMP(arcstat_l2_evict_l1cached); 6999 /* 7000 * Invalidate issued or about to be issued 7001 * reads, since we may be about to write 7002 * over this location. 7003 */ 7004 if (HDR_L2_READING(hdr)) { 7005 ARCSTAT_BUMP(arcstat_l2_evict_reading); 7006 arc_hdr_set_flags(hdr, ARC_FLAG_L2_EVICTED); 7007 } 7008 7009 arc_hdr_l2hdr_destroy(hdr); 7010 } 7011 mutex_exit(hash_lock); 7012 } 7013 mutex_exit(&dev->l2ad_mtx); 7014 } 7015 7016 /* 7017 * Find and write ARC buffers to the L2ARC device. 7018 * 7019 * An ARC_FLAG_L2_WRITING flag is set so that the L2ARC buffers are not valid 7020 * for reading until they have completed writing. 7021 * The headroom_boost is an in-out parameter used to maintain headroom boost 7022 * state between calls to this function. 7023 * 7024 * Returns the number of bytes actually written (which may be smaller than 7025 * the delta by which the device hand has changed due to alignment). 7026 */ 7027 static uint64_t 7028 l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz) 7029 { 7030 arc_buf_hdr_t *hdr, *hdr_prev, *head; 7031 uint64_t write_asize, write_psize, write_lsize, headroom; 7032 boolean_t full; 7033 l2arc_write_callback_t *cb; 7034 zio_t *pio, *wzio; 7035 uint64_t guid = spa_load_guid(spa); 7036 7037 ASSERT3P(dev->l2ad_vdev, !=, NULL); 7038 7039 pio = NULL; 7040 write_lsize = write_asize = write_psize = 0; 7041 full = B_FALSE; 7042 head = kmem_cache_alloc(hdr_l2only_cache, KM_PUSHPAGE); 7043 arc_hdr_set_flags(head, ARC_FLAG_L2_WRITE_HEAD | ARC_FLAG_HAS_L2HDR); 7044 7045 /* 7046 * Copy buffers for L2ARC writing. 7047 */ 7048 for (int try = 0; try <= 3; try++) { 7049 multilist_sublist_t *mls = l2arc_sublist_lock(try); 7050 uint64_t passed_sz = 0; 7051 7052 /* 7053 * L2ARC fast warmup. 7054 * 7055 * Until the ARC is warm and starts to evict, read from the 7056 * head of the ARC lists rather than the tail. 7057 */ 7058 if (arc_warm == B_FALSE) 7059 hdr = multilist_sublist_head(mls); 7060 else 7061 hdr = multilist_sublist_tail(mls); 7062 7063 headroom = target_sz * l2arc_headroom; 7064 if (zfs_compressed_arc_enabled) 7065 headroom = (headroom * l2arc_headroom_boost) / 100; 7066 7067 for (; hdr; hdr = hdr_prev) { 7068 kmutex_t *hash_lock; 7069 7070 if (arc_warm == B_FALSE) 7071 hdr_prev = multilist_sublist_next(mls, hdr); 7072 else 7073 hdr_prev = multilist_sublist_prev(mls, hdr); 7074 7075 hash_lock = HDR_LOCK(hdr); 7076 if (!mutex_tryenter(hash_lock)) { 7077 /* 7078 * Skip this buffer rather than waiting. 7079 */ 7080 continue; 7081 } 7082 7083 passed_sz += HDR_GET_LSIZE(hdr); 7084 if (passed_sz > headroom) { 7085 /* 7086 * Searched too far. 7087 */ 7088 mutex_exit(hash_lock); 7089 break; 7090 } 7091 7092 if (!l2arc_write_eligible(guid, hdr)) { 7093 mutex_exit(hash_lock); 7094 continue; 7095 } 7096 7097 /* 7098 * We rely on the L1 portion of the header below, so 7099 * it's invalid for this header to have been evicted out 7100 * of the ghost cache, prior to being written out. The 7101 * ARC_FLAG_L2_WRITING bit ensures this won't happen. 7102 */ 7103 ASSERT(HDR_HAS_L1HDR(hdr)); 7104 7105 ASSERT3U(HDR_GET_PSIZE(hdr), >, 0); 7106 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL); 7107 ASSERT3U(arc_hdr_size(hdr), >, 0); 7108 uint64_t psize = arc_hdr_size(hdr); 7109 uint64_t asize = vdev_psize_to_asize(dev->l2ad_vdev, 7110 psize); 7111 7112 if ((write_asize + asize) > target_sz) { 7113 full = B_TRUE; 7114 mutex_exit(hash_lock); 7115 break; 7116 } 7117 7118 if (pio == NULL) { 7119 /* 7120 * Insert a dummy header on the buflist so 7121 * l2arc_write_done() can find where the 7122 * write buffers begin without searching. 7123 */ 7124 mutex_enter(&dev->l2ad_mtx); 7125 list_insert_head(&dev->l2ad_buflist, head); 7126 mutex_exit(&dev->l2ad_mtx); 7127 7128 cb = kmem_alloc( 7129 sizeof (l2arc_write_callback_t), KM_SLEEP); 7130 cb->l2wcb_dev = dev; 7131 cb->l2wcb_head = head; 7132 pio = zio_root(spa, l2arc_write_done, cb, 7133 ZIO_FLAG_CANFAIL); 7134 } 7135 7136 hdr->b_l2hdr.b_dev = dev; 7137 hdr->b_l2hdr.b_daddr = dev->l2ad_hand; 7138 arc_hdr_set_flags(hdr, 7139 ARC_FLAG_L2_WRITING | ARC_FLAG_HAS_L2HDR); 7140 7141 mutex_enter(&dev->l2ad_mtx); 7142 list_insert_head(&dev->l2ad_buflist, hdr); 7143 mutex_exit(&dev->l2ad_mtx); 7144 7145 (void) zfs_refcount_add_many(&dev->l2ad_alloc, psize, 7146 hdr); 7147 7148 /* 7149 * Normally the L2ARC can use the hdr's data, but if 7150 * we're sharing data between the hdr and one of its 7151 * bufs, L2ARC needs its own copy of the data so that 7152 * the ZIO below can't race with the buf consumer. 7153 * Another case where we need to create a copy of the 7154 * data is when the buffer size is not device-aligned 7155 * and we need to pad the block to make it such. 7156 * That also keeps the clock hand suitably aligned. 7157 * 7158 * To ensure that the copy will be available for the 7159 * lifetime of the ZIO and be cleaned up afterwards, we 7160 * add it to the l2arc_free_on_write queue. 7161 */ 7162 abd_t *to_write; 7163 if (!HDR_SHARED_DATA(hdr) && psize == asize) { 7164 to_write = hdr->b_l1hdr.b_pabd; 7165 } else { 7166 to_write = abd_alloc_for_io(asize, 7167 HDR_ISTYPE_METADATA(hdr)); 7168 abd_copy(to_write, hdr->b_l1hdr.b_pabd, psize); 7169 if (asize != psize) { 7170 abd_zero_off(to_write, psize, 7171 asize - psize); 7172 } 7173 l2arc_free_abd_on_write(to_write, asize, 7174 arc_buf_type(hdr)); 7175 } 7176 wzio = zio_write_phys(pio, dev->l2ad_vdev, 7177 hdr->b_l2hdr.b_daddr, asize, to_write, 7178 ZIO_CHECKSUM_OFF, NULL, hdr, 7179 ZIO_PRIORITY_ASYNC_WRITE, 7180 ZIO_FLAG_CANFAIL, B_FALSE); 7181 7182 write_lsize += HDR_GET_LSIZE(hdr); 7183 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev, 7184 zio_t *, wzio); 7185 7186 write_psize += psize; 7187 write_asize += asize; 7188 dev->l2ad_hand += asize; 7189 7190 mutex_exit(hash_lock); 7191 7192 (void) zio_nowait(wzio); 7193 } 7194 7195 multilist_sublist_unlock(mls); 7196 7197 if (full == B_TRUE) 7198 break; 7199 } 7200 7201 /* No buffers selected for writing? */ 7202 if (pio == NULL) { 7203 ASSERT0(write_lsize); 7204 ASSERT(!HDR_HAS_L1HDR(head)); 7205 kmem_cache_free(hdr_l2only_cache, head); 7206 return (0); 7207 } 7208 7209 ASSERT3U(write_asize, <=, target_sz); 7210 ARCSTAT_BUMP(arcstat_l2_writes_sent); 7211 ARCSTAT_INCR(arcstat_l2_write_bytes, write_psize); 7212 ARCSTAT_INCR(arcstat_l2_lsize, write_lsize); 7213 ARCSTAT_INCR(arcstat_l2_psize, write_psize); 7214 vdev_space_update(dev->l2ad_vdev, write_psize, 0, 0); 7215 7216 /* 7217 * Bump device hand to the device start if it is approaching the end. 7218 * l2arc_evict() will already have evicted ahead for this case. 7219 */ 7220 if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) { 7221 dev->l2ad_hand = dev->l2ad_start; 7222 dev->l2ad_first = B_FALSE; 7223 } 7224 7225 dev->l2ad_writing = B_TRUE; 7226 (void) zio_wait(pio); 7227 dev->l2ad_writing = B_FALSE; 7228 7229 return (write_asize); 7230 } 7231 7232 /* 7233 * This thread feeds the L2ARC at regular intervals. This is the beating 7234 * heart of the L2ARC. 7235 */ 7236 /* ARGSUSED */ 7237 static void 7238 l2arc_feed_thread(void *unused) 7239 { 7240 callb_cpr_t cpr; 7241 l2arc_dev_t *dev; 7242 spa_t *spa; 7243 uint64_t size, wrote; 7244 clock_t begin, next = ddi_get_lbolt(); 7245 7246 CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG); 7247 7248 mutex_enter(&l2arc_feed_thr_lock); 7249 7250 while (l2arc_thread_exit == 0) { 7251 CALLB_CPR_SAFE_BEGIN(&cpr); 7252 (void) cv_timedwait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock, 7253 next); 7254 CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock); 7255 next = ddi_get_lbolt() + hz; 7256 7257 /* 7258 * Quick check for L2ARC devices. 7259 */ 7260 mutex_enter(&l2arc_dev_mtx); 7261 if (l2arc_ndev == 0) { 7262 mutex_exit(&l2arc_dev_mtx); 7263 continue; 7264 } 7265 mutex_exit(&l2arc_dev_mtx); 7266 begin = ddi_get_lbolt(); 7267 7268 /* 7269 * This selects the next l2arc device to write to, and in 7270 * doing so the next spa to feed from: dev->l2ad_spa. This 7271 * will return NULL if there are now no l2arc devices or if 7272 * they are all faulted. 7273 * 7274 * If a device is returned, its spa's config lock is also 7275 * held to prevent device removal. l2arc_dev_get_next() 7276 * will grab and release l2arc_dev_mtx. 7277 */ 7278 if ((dev = l2arc_dev_get_next()) == NULL) 7279 continue; 7280 7281 spa = dev->l2ad_spa; 7282 ASSERT3P(spa, !=, NULL); 7283 7284 /* 7285 * If the pool is read-only then force the feed thread to 7286 * sleep a little longer. 7287 */ 7288 if (!spa_writeable(spa)) { 7289 next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz; 7290 spa_config_exit(spa, SCL_L2ARC, dev); 7291 continue; 7292 } 7293 7294 /* 7295 * Avoid contributing to memory pressure. 7296 */ 7297 if (arc_reclaim_needed()) { 7298 ARCSTAT_BUMP(arcstat_l2_abort_lowmem); 7299 spa_config_exit(spa, SCL_L2ARC, dev); 7300 continue; 7301 } 7302 7303 ARCSTAT_BUMP(arcstat_l2_feeds); 7304 7305 size = l2arc_write_size(); 7306 7307 /* 7308 * Evict L2ARC buffers that will be overwritten. 7309 */ 7310 l2arc_evict(dev, size, B_FALSE); 7311 7312 /* 7313 * Write ARC buffers. 7314 */ 7315 wrote = l2arc_write_buffers(spa, dev, size); 7316 7317 /* 7318 * Calculate interval between writes. 7319 */ 7320 next = l2arc_write_interval(begin, size, wrote); 7321 spa_config_exit(spa, SCL_L2ARC, dev); 7322 } 7323 7324 l2arc_thread_exit = 0; 7325 cv_broadcast(&l2arc_feed_thr_cv); 7326 CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */ 7327 thread_exit(); 7328 } 7329 7330 boolean_t 7331 l2arc_vdev_present(vdev_t *vd) 7332 { 7333 l2arc_dev_t *dev; 7334 7335 mutex_enter(&l2arc_dev_mtx); 7336 for (dev = list_head(l2arc_dev_list); dev != NULL; 7337 dev = list_next(l2arc_dev_list, dev)) { 7338 if (dev->l2ad_vdev == vd) 7339 break; 7340 } 7341 mutex_exit(&l2arc_dev_mtx); 7342 7343 return (dev != NULL); 7344 } 7345 7346 /* 7347 * Add a vdev for use by the L2ARC. By this point the spa has already 7348 * validated the vdev and opened it. 7349 */ 7350 void 7351 l2arc_add_vdev(spa_t *spa, vdev_t *vd) 7352 { 7353 l2arc_dev_t *adddev; 7354 7355 ASSERT(!l2arc_vdev_present(vd)); 7356 7357 /* 7358 * Create a new l2arc device entry. 7359 */ 7360 adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP); 7361 adddev->l2ad_spa = spa; 7362 adddev->l2ad_vdev = vd; 7363 adddev->l2ad_start = VDEV_LABEL_START_SIZE; 7364 adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd); 7365 adddev->l2ad_hand = adddev->l2ad_start; 7366 adddev->l2ad_first = B_TRUE; 7367 adddev->l2ad_writing = B_FALSE; 7368 7369 mutex_init(&adddev->l2ad_mtx, NULL, MUTEX_DEFAULT, NULL); 7370 /* 7371 * This is a list of all ARC buffers that are still valid on the 7372 * device. 7373 */ 7374 list_create(&adddev->l2ad_buflist, sizeof (arc_buf_hdr_t), 7375 offsetof(arc_buf_hdr_t, b_l2hdr.b_l2node)); 7376 7377 vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand); 7378 zfs_refcount_create(&adddev->l2ad_alloc); 7379 7380 /* 7381 * Add device to global list 7382 */ 7383 mutex_enter(&l2arc_dev_mtx); 7384 list_insert_head(l2arc_dev_list, adddev); 7385 atomic_inc_64(&l2arc_ndev); 7386 mutex_exit(&l2arc_dev_mtx); 7387 } 7388 7389 /* 7390 * Remove a vdev from the L2ARC. 7391 */ 7392 void 7393 l2arc_remove_vdev(vdev_t *vd) 7394 { 7395 l2arc_dev_t *dev, *nextdev, *remdev = NULL; 7396 7397 /* 7398 * Find the device by vdev 7399 */ 7400 mutex_enter(&l2arc_dev_mtx); 7401 for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) { 7402 nextdev = list_next(l2arc_dev_list, dev); 7403 if (vd == dev->l2ad_vdev) { 7404 remdev = dev; 7405 break; 7406 } 7407 } 7408 ASSERT3P(remdev, !=, NULL); 7409 7410 /* 7411 * Remove device from global list 7412 */ 7413 list_remove(l2arc_dev_list, remdev); 7414 l2arc_dev_last = NULL; /* may have been invalidated */ 7415 atomic_dec_64(&l2arc_ndev); 7416 mutex_exit(&l2arc_dev_mtx); 7417 7418 /* 7419 * Clear all buflists and ARC references. L2ARC device flush. 7420 */ 7421 l2arc_evict(remdev, 0, B_TRUE); 7422 list_destroy(&remdev->l2ad_buflist); 7423 mutex_destroy(&remdev->l2ad_mtx); 7424 zfs_refcount_destroy(&remdev->l2ad_alloc); 7425 kmem_free(remdev, sizeof (l2arc_dev_t)); 7426 } 7427 7428 void 7429 l2arc_init(void) 7430 { 7431 l2arc_thread_exit = 0; 7432 l2arc_ndev = 0; 7433 l2arc_writes_sent = 0; 7434 l2arc_writes_done = 0; 7435 7436 mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL); 7437 cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL); 7438 mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL); 7439 mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL); 7440 7441 l2arc_dev_list = &L2ARC_dev_list; 7442 l2arc_free_on_write = &L2ARC_free_on_write; 7443 list_create(l2arc_dev_list, sizeof (l2arc_dev_t), 7444 offsetof(l2arc_dev_t, l2ad_node)); 7445 list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t), 7446 offsetof(l2arc_data_free_t, l2df_list_node)); 7447 } 7448 7449 void 7450 l2arc_fini(void) 7451 { 7452 /* 7453 * This is called from dmu_fini(), which is called from spa_fini(); 7454 * Because of this, we can assume that all l2arc devices have 7455 * already been removed when the pools themselves were removed. 7456 */ 7457 7458 l2arc_do_free_on_write(); 7459 7460 mutex_destroy(&l2arc_feed_thr_lock); 7461 cv_destroy(&l2arc_feed_thr_cv); 7462 mutex_destroy(&l2arc_dev_mtx); 7463 mutex_destroy(&l2arc_free_on_write_mtx); 7464 7465 list_destroy(l2arc_dev_list); 7466 list_destroy(l2arc_free_on_write); 7467 } 7468 7469 void 7470 l2arc_start(void) 7471 { 7472 if (!(spa_mode_global & FWRITE)) 7473 return; 7474 7475 (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0, 7476 TS_RUN, minclsyspri); 7477 } 7478 7479 void 7480 l2arc_stop(void) 7481 { 7482 if (!(spa_mode_global & FWRITE)) 7483 return; 7484 7485 mutex_enter(&l2arc_feed_thr_lock); 7486 cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */ 7487 l2arc_thread_exit = 1; 7488 while (l2arc_thread_exit != 0) 7489 cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock); 7490 mutex_exit(&l2arc_feed_thr_lock); 7491 } 7492