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 zfs_arc_min_prefetch_ms = 1; 352 static int zfs_arc_min_prescient_prefetch_ms = 6; 353 354 /* 355 * If this percent of memory is free, don't throttle. 356 */ 357 int arc_lotsfree_percent = 10; 358 359 static boolean_t arc_initialized; 360 361 /* 362 * The arc has filled available memory and has now warmed up. 363 */ 364 static boolean_t arc_warm; 365 366 /* 367 * log2 fraction of the zio arena to keep free. 368 */ 369 int arc_zio_arena_free_shift = 2; 370 371 /* 372 * These tunables are for performance analysis. 373 */ 374 uint64_t zfs_arc_max; 375 uint64_t zfs_arc_min; 376 uint64_t zfs_arc_meta_limit = 0; 377 uint64_t zfs_arc_meta_min = 0; 378 int zfs_arc_grow_retry = 0; 379 int zfs_arc_shrink_shift = 0; 380 int zfs_arc_p_min_shift = 0; 381 int zfs_arc_average_blocksize = 8 * 1024; /* 8KB */ 382 383 /* 384 * ARC dirty data constraints for arc_tempreserve_space() throttle 385 */ 386 uint_t zfs_arc_dirty_limit_percent = 50; /* total dirty data limit */ 387 uint_t zfs_arc_anon_limit_percent = 25; /* anon block dirty limit */ 388 uint_t zfs_arc_pool_dirty_percent = 20; /* each pool's anon allowance */ 389 390 boolean_t zfs_compressed_arc_enabled = B_TRUE; 391 392 /* 393 * Note that buffers can be in one of 6 states: 394 * ARC_anon - anonymous (discussed below) 395 * ARC_mru - recently used, currently cached 396 * ARC_mru_ghost - recentely used, no longer in cache 397 * ARC_mfu - frequently used, currently cached 398 * ARC_mfu_ghost - frequently used, no longer in cache 399 * ARC_l2c_only - exists in L2ARC but not other states 400 * When there are no active references to the buffer, they are 401 * are linked onto a list in one of these arc states. These are 402 * the only buffers that can be evicted or deleted. Within each 403 * state there are multiple lists, one for meta-data and one for 404 * non-meta-data. Meta-data (indirect blocks, blocks of dnodes, 405 * etc.) is tracked separately so that it can be managed more 406 * explicitly: favored over data, limited explicitly. 407 * 408 * Anonymous buffers are buffers that are not associated with 409 * a DVA. These are buffers that hold dirty block copies 410 * before they are written to stable storage. By definition, 411 * they are "ref'd" and are considered part of arc_mru 412 * that cannot be freed. Generally, they will aquire a DVA 413 * as they are written and migrate onto the arc_mru list. 414 * 415 * The ARC_l2c_only state is for buffers that are in the second 416 * level ARC but no longer in any of the ARC_m* lists. The second 417 * level ARC itself may also contain buffers that are in any of 418 * the ARC_m* states - meaning that a buffer can exist in two 419 * places. The reason for the ARC_l2c_only state is to keep the 420 * buffer header in the hash table, so that reads that hit the 421 * second level ARC benefit from these fast lookups. 422 */ 423 424 typedef struct arc_state { 425 /* 426 * list of evictable buffers 427 */ 428 multilist_t *arcs_list[ARC_BUFC_NUMTYPES]; 429 /* 430 * total amount of evictable data in this state 431 */ 432 zfs_refcount_t arcs_esize[ARC_BUFC_NUMTYPES]; 433 /* 434 * total amount of data in this state; this includes: evictable, 435 * non-evictable, ARC_BUFC_DATA, and ARC_BUFC_METADATA. 436 */ 437 zfs_refcount_t arcs_size; 438 } arc_state_t; 439 440 /* The 6 states: */ 441 static arc_state_t ARC_anon; 442 static arc_state_t ARC_mru; 443 static arc_state_t ARC_mru_ghost; 444 static arc_state_t ARC_mfu; 445 static arc_state_t ARC_mfu_ghost; 446 static arc_state_t ARC_l2c_only; 447 448 typedef struct arc_stats { 449 kstat_named_t arcstat_hits; 450 kstat_named_t arcstat_misses; 451 kstat_named_t arcstat_demand_data_hits; 452 kstat_named_t arcstat_demand_data_misses; 453 kstat_named_t arcstat_demand_metadata_hits; 454 kstat_named_t arcstat_demand_metadata_misses; 455 kstat_named_t arcstat_prefetch_data_hits; 456 kstat_named_t arcstat_prefetch_data_misses; 457 kstat_named_t arcstat_prefetch_metadata_hits; 458 kstat_named_t arcstat_prefetch_metadata_misses; 459 kstat_named_t arcstat_mru_hits; 460 kstat_named_t arcstat_mru_ghost_hits; 461 kstat_named_t arcstat_mfu_hits; 462 kstat_named_t arcstat_mfu_ghost_hits; 463 kstat_named_t arcstat_deleted; 464 /* 465 * Number of buffers that could not be evicted because the hash lock 466 * was held by another thread. The lock may not necessarily be held 467 * by something using the same buffer, since hash locks are shared 468 * by multiple buffers. 469 */ 470 kstat_named_t arcstat_mutex_miss; 471 /* 472 * Number of buffers skipped when updating the access state due to the 473 * header having already been released after acquiring the hash lock. 474 */ 475 kstat_named_t arcstat_access_skip; 476 /* 477 * Number of buffers skipped because they have I/O in progress, are 478 * indirect prefetch buffers that have not lived long enough, or are 479 * not from the spa we're trying to evict from. 480 */ 481 kstat_named_t arcstat_evict_skip; 482 /* 483 * Number of times arc_evict_state() was unable to evict enough 484 * buffers to reach it's target amount. 485 */ 486 kstat_named_t arcstat_evict_not_enough; 487 kstat_named_t arcstat_evict_l2_cached; 488 kstat_named_t arcstat_evict_l2_eligible; 489 kstat_named_t arcstat_evict_l2_ineligible; 490 kstat_named_t arcstat_evict_l2_skip; 491 kstat_named_t arcstat_hash_elements; 492 kstat_named_t arcstat_hash_elements_max; 493 kstat_named_t arcstat_hash_collisions; 494 kstat_named_t arcstat_hash_chains; 495 kstat_named_t arcstat_hash_chain_max; 496 kstat_named_t arcstat_p; 497 kstat_named_t arcstat_c; 498 kstat_named_t arcstat_c_min; 499 kstat_named_t arcstat_c_max; 500 /* Not updated directly; only synced in arc_kstat_update. */ 501 kstat_named_t arcstat_size; 502 /* 503 * Number of compressed bytes stored in the arc_buf_hdr_t's b_pabd. 504 * Note that the compressed bytes may match the uncompressed bytes 505 * if the block is either not compressed or compressed arc is disabled. 506 */ 507 kstat_named_t arcstat_compressed_size; 508 /* 509 * Uncompressed size of the data stored in b_pabd. If compressed 510 * arc is disabled then this value will be identical to the stat 511 * above. 512 */ 513 kstat_named_t arcstat_uncompressed_size; 514 /* 515 * Number of bytes stored in all the arc_buf_t's. This is classified 516 * as "overhead" since this data is typically short-lived and will 517 * be evicted from the arc when it becomes unreferenced unless the 518 * zfs_keep_uncompressed_metadata or zfs_keep_uncompressed_level 519 * values have been set (see comment in dbuf.c for more information). 520 */ 521 kstat_named_t arcstat_overhead_size; 522 /* 523 * Number of bytes consumed by internal ARC structures necessary 524 * for tracking purposes; these structures are not actually 525 * backed by ARC buffers. This includes arc_buf_hdr_t structures 526 * (allocated via arc_buf_hdr_t_full and arc_buf_hdr_t_l2only 527 * caches), and arc_buf_t structures (allocated via arc_buf_t 528 * cache). 529 * Not updated directly; only synced in arc_kstat_update. 530 */ 531 kstat_named_t arcstat_hdr_size; 532 /* 533 * Number of bytes consumed by ARC buffers of type equal to 534 * ARC_BUFC_DATA. This is generally consumed by buffers backing 535 * on disk user data (e.g. plain file contents). 536 * Not updated directly; only synced in arc_kstat_update. 537 */ 538 kstat_named_t arcstat_data_size; 539 /* 540 * Number of bytes consumed by ARC buffers of type equal to 541 * ARC_BUFC_METADATA. This is generally consumed by buffers 542 * backing on disk data that is used for internal ZFS 543 * structures (e.g. ZAP, dnode, indirect blocks, etc). 544 * Not updated directly; only synced in arc_kstat_update. 545 */ 546 kstat_named_t arcstat_metadata_size; 547 /* 548 * Number of bytes consumed by various buffers and structures 549 * not actually backed with ARC buffers. This includes bonus 550 * buffers (allocated directly via zio_buf_* functions), 551 * dmu_buf_impl_t structures (allocated via dmu_buf_impl_t 552 * cache), and dnode_t structures (allocated via dnode_t cache). 553 * Not updated directly; only synced in arc_kstat_update. 554 */ 555 kstat_named_t arcstat_other_size; 556 /* 557 * Total number of bytes consumed by ARC buffers residing in the 558 * arc_anon state. This includes *all* buffers in the arc_anon 559 * state; e.g. data, metadata, evictable, and unevictable buffers 560 * are all included in this value. 561 * Not updated directly; only synced in arc_kstat_update. 562 */ 563 kstat_named_t arcstat_anon_size; 564 /* 565 * Number of bytes consumed by ARC buffers that meet the 566 * following criteria: backing buffers of type ARC_BUFC_DATA, 567 * residing in the arc_anon state, and are eligible for eviction 568 * (e.g. have no outstanding holds on the buffer). 569 * Not updated directly; only synced in arc_kstat_update. 570 */ 571 kstat_named_t arcstat_anon_evictable_data; 572 /* 573 * Number of bytes consumed by ARC buffers that meet the 574 * following criteria: backing buffers of type ARC_BUFC_METADATA, 575 * residing in the arc_anon state, and are eligible for eviction 576 * (e.g. have no outstanding holds on the buffer). 577 * Not updated directly; only synced in arc_kstat_update. 578 */ 579 kstat_named_t arcstat_anon_evictable_metadata; 580 /* 581 * Total number of bytes consumed by ARC buffers residing in the 582 * arc_mru state. This includes *all* buffers in the arc_mru 583 * state; e.g. data, metadata, evictable, and unevictable buffers 584 * are all included in this value. 585 * Not updated directly; only synced in arc_kstat_update. 586 */ 587 kstat_named_t arcstat_mru_size; 588 /* 589 * Number of bytes consumed by ARC buffers that meet the 590 * following criteria: backing buffers of type ARC_BUFC_DATA, 591 * residing in the arc_mru state, and are eligible for eviction 592 * (e.g. have no outstanding holds on the buffer). 593 * Not updated directly; only synced in arc_kstat_update. 594 */ 595 kstat_named_t arcstat_mru_evictable_data; 596 /* 597 * Number of bytes consumed by ARC buffers that meet the 598 * following criteria: backing buffers of type ARC_BUFC_METADATA, 599 * residing in the arc_mru state, and are eligible for eviction 600 * (e.g. have no outstanding holds on the buffer). 601 * Not updated directly; only synced in arc_kstat_update. 602 */ 603 kstat_named_t arcstat_mru_evictable_metadata; 604 /* 605 * Total number of bytes that *would have been* consumed by ARC 606 * buffers in the arc_mru_ghost state. The key thing to note 607 * here, is the fact that this size doesn't actually indicate 608 * RAM consumption. The ghost lists only consist of headers and 609 * don't actually have ARC buffers linked off of these headers. 610 * Thus, *if* the headers had associated ARC buffers, these 611 * buffers *would have* consumed this number of bytes. 612 * Not updated directly; only synced in arc_kstat_update. 613 */ 614 kstat_named_t arcstat_mru_ghost_size; 615 /* 616 * Number of bytes that *would have been* consumed by ARC 617 * buffers that are eligible for eviction, of type 618 * ARC_BUFC_DATA, and linked off the arc_mru_ghost state. 619 * Not updated directly; only synced in arc_kstat_update. 620 */ 621 kstat_named_t arcstat_mru_ghost_evictable_data; 622 /* 623 * Number of bytes that *would have been* consumed by ARC 624 * buffers that are eligible for eviction, of type 625 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state. 626 * Not updated directly; only synced in arc_kstat_update. 627 */ 628 kstat_named_t arcstat_mru_ghost_evictable_metadata; 629 /* 630 * Total number of bytes consumed by ARC buffers residing in the 631 * arc_mfu state. This includes *all* buffers in the arc_mfu 632 * state; e.g. data, metadata, evictable, and unevictable buffers 633 * are all included in this value. 634 * Not updated directly; only synced in arc_kstat_update. 635 */ 636 kstat_named_t arcstat_mfu_size; 637 /* 638 * Number of bytes consumed by ARC buffers that are eligible for 639 * eviction, of type ARC_BUFC_DATA, and reside in the arc_mfu 640 * state. 641 * Not updated directly; only synced in arc_kstat_update. 642 */ 643 kstat_named_t arcstat_mfu_evictable_data; 644 /* 645 * Number of bytes consumed by ARC buffers that are eligible for 646 * eviction, of type ARC_BUFC_METADATA, and reside in the 647 * arc_mfu state. 648 * Not updated directly; only synced in arc_kstat_update. 649 */ 650 kstat_named_t arcstat_mfu_evictable_metadata; 651 /* 652 * Total number of bytes that *would have been* consumed by ARC 653 * buffers in the arc_mfu_ghost state. See the comment above 654 * arcstat_mru_ghost_size for more details. 655 * Not updated directly; only synced in arc_kstat_update. 656 */ 657 kstat_named_t arcstat_mfu_ghost_size; 658 /* 659 * Number of bytes that *would have been* consumed by ARC 660 * buffers that are eligible for eviction, of type 661 * ARC_BUFC_DATA, and linked off the arc_mfu_ghost state. 662 * Not updated directly; only synced in arc_kstat_update. 663 */ 664 kstat_named_t arcstat_mfu_ghost_evictable_data; 665 /* 666 * Number of bytes that *would have been* consumed by ARC 667 * buffers that are eligible for eviction, of type 668 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state. 669 * Not updated directly; only synced in arc_kstat_update. 670 */ 671 kstat_named_t arcstat_mfu_ghost_evictable_metadata; 672 kstat_named_t arcstat_l2_hits; 673 kstat_named_t arcstat_l2_misses; 674 kstat_named_t arcstat_l2_feeds; 675 kstat_named_t arcstat_l2_rw_clash; 676 kstat_named_t arcstat_l2_read_bytes; 677 kstat_named_t arcstat_l2_write_bytes; 678 kstat_named_t arcstat_l2_writes_sent; 679 kstat_named_t arcstat_l2_writes_done; 680 kstat_named_t arcstat_l2_writes_error; 681 kstat_named_t arcstat_l2_writes_lock_retry; 682 kstat_named_t arcstat_l2_evict_lock_retry; 683 kstat_named_t arcstat_l2_evict_reading; 684 kstat_named_t arcstat_l2_evict_l1cached; 685 kstat_named_t arcstat_l2_free_on_write; 686 kstat_named_t arcstat_l2_abort_lowmem; 687 kstat_named_t arcstat_l2_cksum_bad; 688 kstat_named_t arcstat_l2_io_error; 689 kstat_named_t arcstat_l2_lsize; 690 kstat_named_t arcstat_l2_psize; 691 /* Not updated directly; only synced in arc_kstat_update. */ 692 kstat_named_t arcstat_l2_hdr_size; 693 kstat_named_t arcstat_memory_throttle_count; 694 /* Not updated directly; only synced in arc_kstat_update. */ 695 kstat_named_t arcstat_meta_used; 696 kstat_named_t arcstat_meta_limit; 697 kstat_named_t arcstat_meta_max; 698 kstat_named_t arcstat_meta_min; 699 kstat_named_t arcstat_async_upgrade_sync; 700 kstat_named_t arcstat_demand_hit_predictive_prefetch; 701 kstat_named_t arcstat_demand_hit_prescient_prefetch; 702 } arc_stats_t; 703 704 static arc_stats_t arc_stats = { 705 { "hits", KSTAT_DATA_UINT64 }, 706 { "misses", KSTAT_DATA_UINT64 }, 707 { "demand_data_hits", KSTAT_DATA_UINT64 }, 708 { "demand_data_misses", KSTAT_DATA_UINT64 }, 709 { "demand_metadata_hits", KSTAT_DATA_UINT64 }, 710 { "demand_metadata_misses", KSTAT_DATA_UINT64 }, 711 { "prefetch_data_hits", KSTAT_DATA_UINT64 }, 712 { "prefetch_data_misses", KSTAT_DATA_UINT64 }, 713 { "prefetch_metadata_hits", KSTAT_DATA_UINT64 }, 714 { "prefetch_metadata_misses", KSTAT_DATA_UINT64 }, 715 { "mru_hits", KSTAT_DATA_UINT64 }, 716 { "mru_ghost_hits", KSTAT_DATA_UINT64 }, 717 { "mfu_hits", KSTAT_DATA_UINT64 }, 718 { "mfu_ghost_hits", KSTAT_DATA_UINT64 }, 719 { "deleted", KSTAT_DATA_UINT64 }, 720 { "mutex_miss", KSTAT_DATA_UINT64 }, 721 { "access_skip", KSTAT_DATA_UINT64 }, 722 { "evict_skip", KSTAT_DATA_UINT64 }, 723 { "evict_not_enough", KSTAT_DATA_UINT64 }, 724 { "evict_l2_cached", KSTAT_DATA_UINT64 }, 725 { "evict_l2_eligible", KSTAT_DATA_UINT64 }, 726 { "evict_l2_ineligible", KSTAT_DATA_UINT64 }, 727 { "evict_l2_skip", KSTAT_DATA_UINT64 }, 728 { "hash_elements", KSTAT_DATA_UINT64 }, 729 { "hash_elements_max", KSTAT_DATA_UINT64 }, 730 { "hash_collisions", KSTAT_DATA_UINT64 }, 731 { "hash_chains", KSTAT_DATA_UINT64 }, 732 { "hash_chain_max", KSTAT_DATA_UINT64 }, 733 { "p", KSTAT_DATA_UINT64 }, 734 { "c", KSTAT_DATA_UINT64 }, 735 { "c_min", KSTAT_DATA_UINT64 }, 736 { "c_max", KSTAT_DATA_UINT64 }, 737 { "size", KSTAT_DATA_UINT64 }, 738 { "compressed_size", KSTAT_DATA_UINT64 }, 739 { "uncompressed_size", KSTAT_DATA_UINT64 }, 740 { "overhead_size", KSTAT_DATA_UINT64 }, 741 { "hdr_size", KSTAT_DATA_UINT64 }, 742 { "data_size", KSTAT_DATA_UINT64 }, 743 { "metadata_size", KSTAT_DATA_UINT64 }, 744 { "other_size", KSTAT_DATA_UINT64 }, 745 { "anon_size", KSTAT_DATA_UINT64 }, 746 { "anon_evictable_data", KSTAT_DATA_UINT64 }, 747 { "anon_evictable_metadata", KSTAT_DATA_UINT64 }, 748 { "mru_size", KSTAT_DATA_UINT64 }, 749 { "mru_evictable_data", KSTAT_DATA_UINT64 }, 750 { "mru_evictable_metadata", KSTAT_DATA_UINT64 }, 751 { "mru_ghost_size", KSTAT_DATA_UINT64 }, 752 { "mru_ghost_evictable_data", KSTAT_DATA_UINT64 }, 753 { "mru_ghost_evictable_metadata", KSTAT_DATA_UINT64 }, 754 { "mfu_size", KSTAT_DATA_UINT64 }, 755 { "mfu_evictable_data", KSTAT_DATA_UINT64 }, 756 { "mfu_evictable_metadata", KSTAT_DATA_UINT64 }, 757 { "mfu_ghost_size", KSTAT_DATA_UINT64 }, 758 { "mfu_ghost_evictable_data", KSTAT_DATA_UINT64 }, 759 { "mfu_ghost_evictable_metadata", KSTAT_DATA_UINT64 }, 760 { "l2_hits", KSTAT_DATA_UINT64 }, 761 { "l2_misses", KSTAT_DATA_UINT64 }, 762 { "l2_feeds", KSTAT_DATA_UINT64 }, 763 { "l2_rw_clash", KSTAT_DATA_UINT64 }, 764 { "l2_read_bytes", KSTAT_DATA_UINT64 }, 765 { "l2_write_bytes", KSTAT_DATA_UINT64 }, 766 { "l2_writes_sent", KSTAT_DATA_UINT64 }, 767 { "l2_writes_done", KSTAT_DATA_UINT64 }, 768 { "l2_writes_error", KSTAT_DATA_UINT64 }, 769 { "l2_writes_lock_retry", KSTAT_DATA_UINT64 }, 770 { "l2_evict_lock_retry", KSTAT_DATA_UINT64 }, 771 { "l2_evict_reading", KSTAT_DATA_UINT64 }, 772 { "l2_evict_l1cached", KSTAT_DATA_UINT64 }, 773 { "l2_free_on_write", KSTAT_DATA_UINT64 }, 774 { "l2_abort_lowmem", KSTAT_DATA_UINT64 }, 775 { "l2_cksum_bad", KSTAT_DATA_UINT64 }, 776 { "l2_io_error", KSTAT_DATA_UINT64 }, 777 { "l2_size", KSTAT_DATA_UINT64 }, 778 { "l2_asize", KSTAT_DATA_UINT64 }, 779 { "l2_hdr_size", KSTAT_DATA_UINT64 }, 780 { "memory_throttle_count", KSTAT_DATA_UINT64 }, 781 { "arc_meta_used", KSTAT_DATA_UINT64 }, 782 { "arc_meta_limit", KSTAT_DATA_UINT64 }, 783 { "arc_meta_max", KSTAT_DATA_UINT64 }, 784 { "arc_meta_min", KSTAT_DATA_UINT64 }, 785 { "async_upgrade_sync", KSTAT_DATA_UINT64 }, 786 { "demand_hit_predictive_prefetch", KSTAT_DATA_UINT64 }, 787 { "demand_hit_prescient_prefetch", KSTAT_DATA_UINT64 }, 788 }; 789 790 #define ARCSTAT(stat) (arc_stats.stat.value.ui64) 791 792 #define ARCSTAT_INCR(stat, val) \ 793 atomic_add_64(&arc_stats.stat.value.ui64, (val)) 794 795 #define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1) 796 #define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1) 797 798 #define ARCSTAT_MAX(stat, val) { \ 799 uint64_t m; \ 800 while ((val) > (m = arc_stats.stat.value.ui64) && \ 801 (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \ 802 continue; \ 803 } 804 805 #define ARCSTAT_MAXSTAT(stat) \ 806 ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64) 807 808 /* 809 * We define a macro to allow ARC hits/misses to be easily broken down by 810 * two separate conditions, giving a total of four different subtypes for 811 * each of hits and misses (so eight statistics total). 812 */ 813 #define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \ 814 if (cond1) { \ 815 if (cond2) { \ 816 ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \ 817 } else { \ 818 ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \ 819 } \ 820 } else { \ 821 if (cond2) { \ 822 ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \ 823 } else { \ 824 ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\ 825 } \ 826 } 827 828 kstat_t *arc_ksp; 829 static arc_state_t *arc_anon; 830 static arc_state_t *arc_mru; 831 static arc_state_t *arc_mru_ghost; 832 static arc_state_t *arc_mfu; 833 static arc_state_t *arc_mfu_ghost; 834 static arc_state_t *arc_l2c_only; 835 836 /* 837 * There are several ARC variables that are critical to export as kstats -- 838 * but we don't want to have to grovel around in the kstat whenever we wish to 839 * manipulate them. For these variables, we therefore define them to be in 840 * terms of the statistic variable. This assures that we are not introducing 841 * the possibility of inconsistency by having shadow copies of the variables, 842 * while still allowing the code to be readable. 843 */ 844 #define arc_p ARCSTAT(arcstat_p) /* target size of MRU */ 845 #define arc_c ARCSTAT(arcstat_c) /* target size of cache */ 846 #define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */ 847 #define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */ 848 #define arc_meta_limit ARCSTAT(arcstat_meta_limit) /* max size for metadata */ 849 #define arc_meta_min ARCSTAT(arcstat_meta_min) /* min size for metadata */ 850 #define arc_meta_max ARCSTAT(arcstat_meta_max) /* max size of metadata */ 851 852 /* compressed size of entire arc */ 853 #define arc_compressed_size ARCSTAT(arcstat_compressed_size) 854 /* uncompressed size of entire arc */ 855 #define arc_uncompressed_size ARCSTAT(arcstat_uncompressed_size) 856 /* number of bytes in the arc from arc_buf_t's */ 857 #define arc_overhead_size ARCSTAT(arcstat_overhead_size) 858 859 /* 860 * There are also some ARC variables that we want to export, but that are 861 * updated so often that having the canonical representation be the statistic 862 * variable causes a performance bottleneck. We want to use aggsum_t's for these 863 * instead, but still be able to export the kstat in the same way as before. 864 * The solution is to always use the aggsum version, except in the kstat update 865 * callback. 866 */ 867 aggsum_t arc_size; 868 aggsum_t arc_meta_used; 869 aggsum_t astat_data_size; 870 aggsum_t astat_metadata_size; 871 aggsum_t astat_hdr_size; 872 aggsum_t astat_other_size; 873 aggsum_t astat_l2_hdr_size; 874 875 static int arc_no_grow; /* Don't try to grow cache size */ 876 static hrtime_t arc_growtime; 877 static uint64_t arc_tempreserve; 878 static uint64_t arc_loaned_bytes; 879 880 typedef struct arc_callback arc_callback_t; 881 882 struct arc_callback { 883 void *acb_private; 884 arc_read_done_func_t *acb_done; 885 arc_buf_t *acb_buf; 886 boolean_t acb_compressed; 887 zio_t *acb_zio_dummy; 888 zio_t *acb_zio_head; 889 arc_callback_t *acb_next; 890 }; 891 892 typedef struct arc_write_callback arc_write_callback_t; 893 894 struct arc_write_callback { 895 void *awcb_private; 896 arc_write_done_func_t *awcb_ready; 897 arc_write_done_func_t *awcb_children_ready; 898 arc_write_done_func_t *awcb_physdone; 899 arc_write_done_func_t *awcb_done; 900 arc_buf_t *awcb_buf; 901 }; 902 903 /* 904 * ARC buffers are separated into multiple structs as a memory saving measure: 905 * - Common fields struct, always defined, and embedded within it: 906 * - L2-only fields, always allocated but undefined when not in L2ARC 907 * - L1-only fields, only allocated when in L1ARC 908 * 909 * Buffer in L1 Buffer only in L2 910 * +------------------------+ +------------------------+ 911 * | arc_buf_hdr_t | | arc_buf_hdr_t | 912 * | | | | 913 * | | | | 914 * | | | | 915 * +------------------------+ +------------------------+ 916 * | l2arc_buf_hdr_t | | l2arc_buf_hdr_t | 917 * | (undefined if L1-only) | | | 918 * +------------------------+ +------------------------+ 919 * | l1arc_buf_hdr_t | 920 * | | 921 * | | 922 * | | 923 * | | 924 * +------------------------+ 925 * 926 * Because it's possible for the L2ARC to become extremely large, we can wind 927 * up eating a lot of memory in L2ARC buffer headers, so the size of a header 928 * is minimized by only allocating the fields necessary for an L1-cached buffer 929 * when a header is actually in the L1 cache. The sub-headers (l1arc_buf_hdr and 930 * l2arc_buf_hdr) are embedded rather than allocated separately to save a couple 931 * words in pointers. arc_hdr_realloc() is used to switch a header between 932 * these two allocation states. 933 */ 934 typedef struct l1arc_buf_hdr { 935 kmutex_t b_freeze_lock; 936 zio_cksum_t *b_freeze_cksum; 937 #ifdef ZFS_DEBUG 938 /* 939 * Used for debugging with kmem_flags - by allocating and freeing 940 * b_thawed when the buffer is thawed, we get a record of the stack 941 * trace that thawed it. 942 */ 943 void *b_thawed; 944 #endif 945 946 arc_buf_t *b_buf; 947 uint32_t b_bufcnt; 948 /* for waiting on writes to complete */ 949 kcondvar_t b_cv; 950 uint8_t b_byteswap; 951 952 /* protected by arc state mutex */ 953 arc_state_t *b_state; 954 multilist_node_t b_arc_node; 955 956 /* updated atomically */ 957 clock_t b_arc_access; 958 959 /* self protecting */ 960 zfs_refcount_t b_refcnt; 961 962 arc_callback_t *b_acb; 963 abd_t *b_pabd; 964 } l1arc_buf_hdr_t; 965 966 typedef struct l2arc_dev l2arc_dev_t; 967 968 typedef struct l2arc_buf_hdr { 969 /* protected by arc_buf_hdr mutex */ 970 l2arc_dev_t *b_dev; /* L2ARC device */ 971 uint64_t b_daddr; /* disk address, offset byte */ 972 973 list_node_t b_l2node; 974 } l2arc_buf_hdr_t; 975 976 struct arc_buf_hdr { 977 /* protected by hash lock */ 978 dva_t b_dva; 979 uint64_t b_birth; 980 981 arc_buf_contents_t b_type; 982 arc_buf_hdr_t *b_hash_next; 983 arc_flags_t b_flags; 984 985 /* 986 * This field stores the size of the data buffer after 987 * compression, and is set in the arc's zio completion handlers. 988 * It is in units of SPA_MINBLOCKSIZE (e.g. 1 == 512 bytes). 989 * 990 * While the block pointers can store up to 32MB in their psize 991 * field, we can only store up to 32MB minus 512B. This is due 992 * to the bp using a bias of 1, whereas we use a bias of 0 (i.e. 993 * a field of zeros represents 512B in the bp). We can't use a 994 * bias of 1 since we need to reserve a psize of zero, here, to 995 * represent holes and embedded blocks. 996 * 997 * This isn't a problem in practice, since the maximum size of a 998 * buffer is limited to 16MB, so we never need to store 32MB in 999 * this field. Even in the upstream illumos code base, the 1000 * maximum size of a buffer is limited to 16MB. 1001 */ 1002 uint16_t b_psize; 1003 1004 /* 1005 * This field stores the size of the data buffer before 1006 * compression, and cannot change once set. It is in units 1007 * of SPA_MINBLOCKSIZE (e.g. 2 == 1024 bytes) 1008 */ 1009 uint16_t b_lsize; /* immutable */ 1010 uint64_t b_spa; /* immutable */ 1011 1012 /* L2ARC fields. Undefined when not in L2ARC. */ 1013 l2arc_buf_hdr_t b_l2hdr; 1014 /* L1ARC fields. Undefined when in l2arc_only state */ 1015 l1arc_buf_hdr_t b_l1hdr; 1016 }; 1017 1018 #define GHOST_STATE(state) \ 1019 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \ 1020 (state) == arc_l2c_only) 1021 1022 #define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_FLAG_IN_HASH_TABLE) 1023 #define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) 1024 #define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_FLAG_IO_ERROR) 1025 #define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_FLAG_PREFETCH) 1026 #define HDR_PRESCIENT_PREFETCH(hdr) \ 1027 ((hdr)->b_flags & ARC_FLAG_PRESCIENT_PREFETCH) 1028 #define HDR_COMPRESSION_ENABLED(hdr) \ 1029 ((hdr)->b_flags & ARC_FLAG_COMPRESSED_ARC) 1030 1031 #define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_FLAG_L2CACHE) 1032 #define HDR_L2_READING(hdr) \ 1033 (((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) && \ 1034 ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR)) 1035 #define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITING) 1036 #define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_FLAG_L2_EVICTED) 1037 #define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITE_HEAD) 1038 #define HDR_SHARED_DATA(hdr) ((hdr)->b_flags & ARC_FLAG_SHARED_DATA) 1039 1040 #define HDR_ISTYPE_METADATA(hdr) \ 1041 ((hdr)->b_flags & ARC_FLAG_BUFC_METADATA) 1042 #define HDR_ISTYPE_DATA(hdr) (!HDR_ISTYPE_METADATA(hdr)) 1043 1044 #define HDR_HAS_L1HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L1HDR) 1045 #define HDR_HAS_L2HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR) 1046 1047 /* For storing compression mode in b_flags */ 1048 #define HDR_COMPRESS_OFFSET (highbit64(ARC_FLAG_COMPRESS_0) - 1) 1049 1050 #define HDR_GET_COMPRESS(hdr) ((enum zio_compress)BF32_GET((hdr)->b_flags, \ 1051 HDR_COMPRESS_OFFSET, SPA_COMPRESSBITS)) 1052 #define HDR_SET_COMPRESS(hdr, cmp) BF32_SET((hdr)->b_flags, \ 1053 HDR_COMPRESS_OFFSET, SPA_COMPRESSBITS, (cmp)); 1054 1055 #define ARC_BUF_LAST(buf) ((buf)->b_next == NULL) 1056 #define ARC_BUF_SHARED(buf) ((buf)->b_flags & ARC_BUF_FLAG_SHARED) 1057 #define ARC_BUF_COMPRESSED(buf) ((buf)->b_flags & ARC_BUF_FLAG_COMPRESSED) 1058 1059 /* 1060 * Other sizes 1061 */ 1062 1063 #define HDR_FULL_SIZE ((int64_t)sizeof (arc_buf_hdr_t)) 1064 #define HDR_L2ONLY_SIZE ((int64_t)offsetof(arc_buf_hdr_t, b_l1hdr)) 1065 1066 /* 1067 * Hash table routines 1068 */ 1069 1070 #define HT_LOCK_PAD 64 1071 1072 struct ht_lock { 1073 kmutex_t ht_lock; 1074 #ifdef _KERNEL 1075 unsigned char pad[(HT_LOCK_PAD - sizeof (kmutex_t))]; 1076 #endif 1077 }; 1078 1079 #define BUF_LOCKS 256 1080 typedef struct buf_hash_table { 1081 uint64_t ht_mask; 1082 arc_buf_hdr_t **ht_table; 1083 struct ht_lock ht_locks[BUF_LOCKS]; 1084 } buf_hash_table_t; 1085 1086 static buf_hash_table_t buf_hash_table; 1087 1088 #define BUF_HASH_INDEX(spa, dva, birth) \ 1089 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask) 1090 #define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)]) 1091 #define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock)) 1092 #define HDR_LOCK(hdr) \ 1093 (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth))) 1094 1095 uint64_t zfs_crc64_table[256]; 1096 1097 /* 1098 * Level 2 ARC 1099 */ 1100 1101 #define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */ 1102 #define L2ARC_HEADROOM 2 /* num of writes */ 1103 /* 1104 * If we discover during ARC scan any buffers to be compressed, we boost 1105 * our headroom for the next scanning cycle by this percentage multiple. 1106 */ 1107 #define L2ARC_HEADROOM_BOOST 200 1108 #define L2ARC_FEED_SECS 1 /* caching interval secs */ 1109 #define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */ 1110 1111 #define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent) 1112 #define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done) 1113 1114 /* L2ARC Performance Tunables */ 1115 uint64_t l2arc_write_max = L2ARC_WRITE_SIZE; /* default max write size */ 1116 uint64_t l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra write during warmup */ 1117 uint64_t l2arc_headroom = L2ARC_HEADROOM; /* number of dev writes */ 1118 uint64_t l2arc_headroom_boost = L2ARC_HEADROOM_BOOST; 1119 uint64_t l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */ 1120 uint64_t l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval milliseconds */ 1121 boolean_t l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */ 1122 boolean_t l2arc_feed_again = B_TRUE; /* turbo warmup */ 1123 boolean_t l2arc_norw = B_TRUE; /* no reads during writes */ 1124 1125 /* 1126 * L2ARC Internals 1127 */ 1128 struct l2arc_dev { 1129 vdev_t *l2ad_vdev; /* vdev */ 1130 spa_t *l2ad_spa; /* spa */ 1131 uint64_t l2ad_hand; /* next write location */ 1132 uint64_t l2ad_start; /* first addr on device */ 1133 uint64_t l2ad_end; /* last addr on device */ 1134 boolean_t l2ad_first; /* first sweep through */ 1135 boolean_t l2ad_writing; /* currently writing */ 1136 kmutex_t l2ad_mtx; /* lock for buffer list */ 1137 list_t l2ad_buflist; /* buffer list */ 1138 list_node_t l2ad_node; /* device list node */ 1139 zfs_refcount_t l2ad_alloc; /* allocated bytes */ 1140 }; 1141 1142 static list_t L2ARC_dev_list; /* device list */ 1143 static list_t *l2arc_dev_list; /* device list pointer */ 1144 static kmutex_t l2arc_dev_mtx; /* device list mutex */ 1145 static l2arc_dev_t *l2arc_dev_last; /* last device used */ 1146 static list_t L2ARC_free_on_write; /* free after write buf list */ 1147 static list_t *l2arc_free_on_write; /* free after write list ptr */ 1148 static kmutex_t l2arc_free_on_write_mtx; /* mutex for list */ 1149 static uint64_t l2arc_ndev; /* number of devices */ 1150 1151 typedef struct l2arc_read_callback { 1152 arc_buf_hdr_t *l2rcb_hdr; /* read header */ 1153 blkptr_t l2rcb_bp; /* original blkptr */ 1154 zbookmark_phys_t l2rcb_zb; /* original bookmark */ 1155 int l2rcb_flags; /* original flags */ 1156 abd_t *l2rcb_abd; /* temporary buffer */ 1157 } l2arc_read_callback_t; 1158 1159 typedef struct l2arc_write_callback { 1160 l2arc_dev_t *l2wcb_dev; /* device info */ 1161 arc_buf_hdr_t *l2wcb_head; /* head of write buflist */ 1162 } l2arc_write_callback_t; 1163 1164 typedef struct l2arc_data_free { 1165 /* protected by l2arc_free_on_write_mtx */ 1166 abd_t *l2df_abd; 1167 size_t l2df_size; 1168 arc_buf_contents_t l2df_type; 1169 list_node_t l2df_list_node; 1170 } l2arc_data_free_t; 1171 1172 static kmutex_t l2arc_feed_thr_lock; 1173 static kcondvar_t l2arc_feed_thr_cv; 1174 static uint8_t l2arc_thread_exit; 1175 1176 static abd_t *arc_get_data_abd(arc_buf_hdr_t *, uint64_t, void *); 1177 static void *arc_get_data_buf(arc_buf_hdr_t *, uint64_t, void *); 1178 static void arc_get_data_impl(arc_buf_hdr_t *, uint64_t, void *); 1179 static void arc_free_data_abd(arc_buf_hdr_t *, abd_t *, uint64_t, void *); 1180 static void arc_free_data_buf(arc_buf_hdr_t *, void *, uint64_t, void *); 1181 static void arc_free_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag); 1182 static void arc_hdr_free_pabd(arc_buf_hdr_t *); 1183 static void arc_hdr_alloc_pabd(arc_buf_hdr_t *); 1184 static void arc_access(arc_buf_hdr_t *, kmutex_t *); 1185 static boolean_t arc_is_overflowing(); 1186 static void arc_buf_watch(arc_buf_t *); 1187 1188 static arc_buf_contents_t arc_buf_type(arc_buf_hdr_t *); 1189 static uint32_t arc_bufc_to_flags(arc_buf_contents_t); 1190 static inline void arc_hdr_set_flags(arc_buf_hdr_t *hdr, arc_flags_t flags); 1191 static inline void arc_hdr_clear_flags(arc_buf_hdr_t *hdr, arc_flags_t flags); 1192 1193 static boolean_t l2arc_write_eligible(uint64_t, arc_buf_hdr_t *); 1194 static void l2arc_read_done(zio_t *); 1195 1196 1197 /* 1198 * We use Cityhash for this. It's fast, and has good hash properties without 1199 * requiring any large static buffers. 1200 */ 1201 static uint64_t 1202 buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth) 1203 { 1204 return (cityhash4(spa, dva->dva_word[0], dva->dva_word[1], birth)); 1205 } 1206 1207 #define HDR_EMPTY(hdr) \ 1208 ((hdr)->b_dva.dva_word[0] == 0 && \ 1209 (hdr)->b_dva.dva_word[1] == 0) 1210 1211 #define HDR_EQUAL(spa, dva, birth, hdr) \ 1212 ((hdr)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \ 1213 ((hdr)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \ 1214 ((hdr)->b_birth == birth) && ((hdr)->b_spa == spa) 1215 1216 static void 1217 buf_discard_identity(arc_buf_hdr_t *hdr) 1218 { 1219 hdr->b_dva.dva_word[0] = 0; 1220 hdr->b_dva.dva_word[1] = 0; 1221 hdr->b_birth = 0; 1222 } 1223 1224 static arc_buf_hdr_t * 1225 buf_hash_find(uint64_t spa, const blkptr_t *bp, kmutex_t **lockp) 1226 { 1227 const dva_t *dva = BP_IDENTITY(bp); 1228 uint64_t birth = BP_PHYSICAL_BIRTH(bp); 1229 uint64_t idx = BUF_HASH_INDEX(spa, dva, birth); 1230 kmutex_t *hash_lock = BUF_HASH_LOCK(idx); 1231 arc_buf_hdr_t *hdr; 1232 1233 mutex_enter(hash_lock); 1234 for (hdr = buf_hash_table.ht_table[idx]; hdr != NULL; 1235 hdr = hdr->b_hash_next) { 1236 if (HDR_EQUAL(spa, dva, birth, hdr)) { 1237 *lockp = hash_lock; 1238 return (hdr); 1239 } 1240 } 1241 mutex_exit(hash_lock); 1242 *lockp = NULL; 1243 return (NULL); 1244 } 1245 1246 /* 1247 * Insert an entry into the hash table. If there is already an element 1248 * equal to elem in the hash table, then the already existing element 1249 * will be returned and the new element will not be inserted. 1250 * Otherwise returns NULL. 1251 * If lockp == NULL, the caller is assumed to already hold the hash lock. 1252 */ 1253 static arc_buf_hdr_t * 1254 buf_hash_insert(arc_buf_hdr_t *hdr, kmutex_t **lockp) 1255 { 1256 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth); 1257 kmutex_t *hash_lock = BUF_HASH_LOCK(idx); 1258 arc_buf_hdr_t *fhdr; 1259 uint32_t i; 1260 1261 ASSERT(!DVA_IS_EMPTY(&hdr->b_dva)); 1262 ASSERT(hdr->b_birth != 0); 1263 ASSERT(!HDR_IN_HASH_TABLE(hdr)); 1264 1265 if (lockp != NULL) { 1266 *lockp = hash_lock; 1267 mutex_enter(hash_lock); 1268 } else { 1269 ASSERT(MUTEX_HELD(hash_lock)); 1270 } 1271 1272 for (fhdr = buf_hash_table.ht_table[idx], i = 0; fhdr != NULL; 1273 fhdr = fhdr->b_hash_next, i++) { 1274 if (HDR_EQUAL(hdr->b_spa, &hdr->b_dva, hdr->b_birth, fhdr)) 1275 return (fhdr); 1276 } 1277 1278 hdr->b_hash_next = buf_hash_table.ht_table[idx]; 1279 buf_hash_table.ht_table[idx] = hdr; 1280 arc_hdr_set_flags(hdr, ARC_FLAG_IN_HASH_TABLE); 1281 1282 /* collect some hash table performance data */ 1283 if (i > 0) { 1284 ARCSTAT_BUMP(arcstat_hash_collisions); 1285 if (i == 1) 1286 ARCSTAT_BUMP(arcstat_hash_chains); 1287 1288 ARCSTAT_MAX(arcstat_hash_chain_max, i); 1289 } 1290 1291 ARCSTAT_BUMP(arcstat_hash_elements); 1292 ARCSTAT_MAXSTAT(arcstat_hash_elements); 1293 1294 return (NULL); 1295 } 1296 1297 static void 1298 buf_hash_remove(arc_buf_hdr_t *hdr) 1299 { 1300 arc_buf_hdr_t *fhdr, **hdrp; 1301 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth); 1302 1303 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx))); 1304 ASSERT(HDR_IN_HASH_TABLE(hdr)); 1305 1306 hdrp = &buf_hash_table.ht_table[idx]; 1307 while ((fhdr = *hdrp) != hdr) { 1308 ASSERT3P(fhdr, !=, NULL); 1309 hdrp = &fhdr->b_hash_next; 1310 } 1311 *hdrp = hdr->b_hash_next; 1312 hdr->b_hash_next = NULL; 1313 arc_hdr_clear_flags(hdr, ARC_FLAG_IN_HASH_TABLE); 1314 1315 /* collect some hash table performance data */ 1316 ARCSTAT_BUMPDOWN(arcstat_hash_elements); 1317 1318 if (buf_hash_table.ht_table[idx] && 1319 buf_hash_table.ht_table[idx]->b_hash_next == NULL) 1320 ARCSTAT_BUMPDOWN(arcstat_hash_chains); 1321 } 1322 1323 /* 1324 * Global data structures and functions for the buf kmem cache. 1325 */ 1326 static kmem_cache_t *hdr_full_cache; 1327 static kmem_cache_t *hdr_l2only_cache; 1328 static kmem_cache_t *buf_cache; 1329 1330 static void 1331 buf_fini(void) 1332 { 1333 int i; 1334 1335 kmem_free(buf_hash_table.ht_table, 1336 (buf_hash_table.ht_mask + 1) * sizeof (void *)); 1337 for (i = 0; i < BUF_LOCKS; i++) 1338 mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock); 1339 kmem_cache_destroy(hdr_full_cache); 1340 kmem_cache_destroy(hdr_l2only_cache); 1341 kmem_cache_destroy(buf_cache); 1342 } 1343 1344 /* 1345 * Constructor callback - called when the cache is empty 1346 * and a new buf is requested. 1347 */ 1348 /* ARGSUSED */ 1349 static int 1350 hdr_full_cons(void *vbuf, void *unused, int kmflag) 1351 { 1352 arc_buf_hdr_t *hdr = vbuf; 1353 1354 bzero(hdr, HDR_FULL_SIZE); 1355 cv_init(&hdr->b_l1hdr.b_cv, NULL, CV_DEFAULT, NULL); 1356 zfs_refcount_create(&hdr->b_l1hdr.b_refcnt); 1357 mutex_init(&hdr->b_l1hdr.b_freeze_lock, NULL, MUTEX_DEFAULT, NULL); 1358 multilist_link_init(&hdr->b_l1hdr.b_arc_node); 1359 arc_space_consume(HDR_FULL_SIZE, ARC_SPACE_HDRS); 1360 1361 return (0); 1362 } 1363 1364 /* ARGSUSED */ 1365 static int 1366 hdr_l2only_cons(void *vbuf, void *unused, int kmflag) 1367 { 1368 arc_buf_hdr_t *hdr = vbuf; 1369 1370 bzero(hdr, HDR_L2ONLY_SIZE); 1371 arc_space_consume(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS); 1372 1373 return (0); 1374 } 1375 1376 /* ARGSUSED */ 1377 static int 1378 buf_cons(void *vbuf, void *unused, int kmflag) 1379 { 1380 arc_buf_t *buf = vbuf; 1381 1382 bzero(buf, sizeof (arc_buf_t)); 1383 mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL); 1384 arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS); 1385 1386 return (0); 1387 } 1388 1389 /* 1390 * Destructor callback - called when a cached buf is 1391 * no longer required. 1392 */ 1393 /* ARGSUSED */ 1394 static void 1395 hdr_full_dest(void *vbuf, void *unused) 1396 { 1397 arc_buf_hdr_t *hdr = vbuf; 1398 1399 ASSERT(HDR_EMPTY(hdr)); 1400 cv_destroy(&hdr->b_l1hdr.b_cv); 1401 zfs_refcount_destroy(&hdr->b_l1hdr.b_refcnt); 1402 mutex_destroy(&hdr->b_l1hdr.b_freeze_lock); 1403 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 1404 arc_space_return(HDR_FULL_SIZE, ARC_SPACE_HDRS); 1405 } 1406 1407 /* ARGSUSED */ 1408 static void 1409 hdr_l2only_dest(void *vbuf, void *unused) 1410 { 1411 arc_buf_hdr_t *hdr = vbuf; 1412 1413 ASSERT(HDR_EMPTY(hdr)); 1414 arc_space_return(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS); 1415 } 1416 1417 /* ARGSUSED */ 1418 static void 1419 buf_dest(void *vbuf, void *unused) 1420 { 1421 arc_buf_t *buf = vbuf; 1422 1423 mutex_destroy(&buf->b_evict_lock); 1424 arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS); 1425 } 1426 1427 /* 1428 * Reclaim callback -- invoked when memory is low. 1429 */ 1430 /* ARGSUSED */ 1431 static void 1432 hdr_recl(void *unused) 1433 { 1434 dprintf("hdr_recl called\n"); 1435 /* 1436 * umem calls the reclaim func when we destroy the buf cache, 1437 * which is after we do arc_fini(). 1438 */ 1439 if (arc_initialized) 1440 zthr_wakeup(arc_reap_zthr); 1441 } 1442 1443 static void 1444 buf_init(void) 1445 { 1446 uint64_t *ct; 1447 uint64_t hsize = 1ULL << 12; 1448 int i, j; 1449 1450 /* 1451 * The hash table is big enough to fill all of physical memory 1452 * with an average block size of zfs_arc_average_blocksize (default 8K). 1453 * By default, the table will take up 1454 * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers). 1455 */ 1456 while (hsize * zfs_arc_average_blocksize < physmem * PAGESIZE) 1457 hsize <<= 1; 1458 retry: 1459 buf_hash_table.ht_mask = hsize - 1; 1460 buf_hash_table.ht_table = 1461 kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP); 1462 if (buf_hash_table.ht_table == NULL) { 1463 ASSERT(hsize > (1ULL << 8)); 1464 hsize >>= 1; 1465 goto retry; 1466 } 1467 1468 hdr_full_cache = kmem_cache_create("arc_buf_hdr_t_full", HDR_FULL_SIZE, 1469 0, hdr_full_cons, hdr_full_dest, hdr_recl, NULL, NULL, 0); 1470 hdr_l2only_cache = kmem_cache_create("arc_buf_hdr_t_l2only", 1471 HDR_L2ONLY_SIZE, 0, hdr_l2only_cons, hdr_l2only_dest, hdr_recl, 1472 NULL, NULL, 0); 1473 buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t), 1474 0, buf_cons, buf_dest, NULL, NULL, NULL, 0); 1475 1476 for (i = 0; i < 256; i++) 1477 for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--) 1478 *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY); 1479 1480 for (i = 0; i < BUF_LOCKS; i++) { 1481 mutex_init(&buf_hash_table.ht_locks[i].ht_lock, 1482 NULL, MUTEX_DEFAULT, NULL); 1483 } 1484 } 1485 1486 /* 1487 * This is the size that the buf occupies in memory. If the buf is compressed, 1488 * it will correspond to the compressed size. You should use this method of 1489 * getting the buf size unless you explicitly need the logical size. 1490 */ 1491 int32_t 1492 arc_buf_size(arc_buf_t *buf) 1493 { 1494 return (ARC_BUF_COMPRESSED(buf) ? 1495 HDR_GET_PSIZE(buf->b_hdr) : HDR_GET_LSIZE(buf->b_hdr)); 1496 } 1497 1498 int32_t 1499 arc_buf_lsize(arc_buf_t *buf) 1500 { 1501 return (HDR_GET_LSIZE(buf->b_hdr)); 1502 } 1503 1504 enum zio_compress 1505 arc_get_compression(arc_buf_t *buf) 1506 { 1507 return (ARC_BUF_COMPRESSED(buf) ? 1508 HDR_GET_COMPRESS(buf->b_hdr) : ZIO_COMPRESS_OFF); 1509 } 1510 1511 #define ARC_MINTIME (hz>>4) /* 62 ms */ 1512 1513 static inline boolean_t 1514 arc_buf_is_shared(arc_buf_t *buf) 1515 { 1516 boolean_t shared = (buf->b_data != NULL && 1517 buf->b_hdr->b_l1hdr.b_pabd != NULL && 1518 abd_is_linear(buf->b_hdr->b_l1hdr.b_pabd) && 1519 buf->b_data == abd_to_buf(buf->b_hdr->b_l1hdr.b_pabd)); 1520 IMPLY(shared, HDR_SHARED_DATA(buf->b_hdr)); 1521 IMPLY(shared, ARC_BUF_SHARED(buf)); 1522 IMPLY(shared, ARC_BUF_COMPRESSED(buf) || ARC_BUF_LAST(buf)); 1523 1524 /* 1525 * It would be nice to assert arc_can_share() too, but the "hdr isn't 1526 * already being shared" requirement prevents us from doing that. 1527 */ 1528 1529 return (shared); 1530 } 1531 1532 /* 1533 * Free the checksum associated with this header. If there is no checksum, this 1534 * is a no-op. 1535 */ 1536 static inline void 1537 arc_cksum_free(arc_buf_hdr_t *hdr) 1538 { 1539 ASSERT(HDR_HAS_L1HDR(hdr)); 1540 mutex_enter(&hdr->b_l1hdr.b_freeze_lock); 1541 if (hdr->b_l1hdr.b_freeze_cksum != NULL) { 1542 kmem_free(hdr->b_l1hdr.b_freeze_cksum, sizeof (zio_cksum_t)); 1543 hdr->b_l1hdr.b_freeze_cksum = NULL; 1544 } 1545 mutex_exit(&hdr->b_l1hdr.b_freeze_lock); 1546 } 1547 1548 /* 1549 * Return true iff at least one of the bufs on hdr is not compressed. 1550 */ 1551 static boolean_t 1552 arc_hdr_has_uncompressed_buf(arc_buf_hdr_t *hdr) 1553 { 1554 for (arc_buf_t *b = hdr->b_l1hdr.b_buf; b != NULL; b = b->b_next) { 1555 if (!ARC_BUF_COMPRESSED(b)) { 1556 return (B_TRUE); 1557 } 1558 } 1559 return (B_FALSE); 1560 } 1561 1562 /* 1563 * If we've turned on the ZFS_DEBUG_MODIFY flag, verify that the buf's data 1564 * matches the checksum that is stored in the hdr. If there is no checksum, 1565 * or if the buf is compressed, this is a no-op. 1566 */ 1567 static void 1568 arc_cksum_verify(arc_buf_t *buf) 1569 { 1570 arc_buf_hdr_t *hdr = buf->b_hdr; 1571 zio_cksum_t zc; 1572 1573 if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 1574 return; 1575 1576 if (ARC_BUF_COMPRESSED(buf)) { 1577 ASSERT(hdr->b_l1hdr.b_freeze_cksum == NULL || 1578 arc_hdr_has_uncompressed_buf(hdr)); 1579 return; 1580 } 1581 1582 ASSERT(HDR_HAS_L1HDR(hdr)); 1583 1584 mutex_enter(&hdr->b_l1hdr.b_freeze_lock); 1585 if (hdr->b_l1hdr.b_freeze_cksum == NULL || HDR_IO_ERROR(hdr)) { 1586 mutex_exit(&hdr->b_l1hdr.b_freeze_lock); 1587 return; 1588 } 1589 1590 fletcher_2_native(buf->b_data, arc_buf_size(buf), NULL, &zc); 1591 if (!ZIO_CHECKSUM_EQUAL(*hdr->b_l1hdr.b_freeze_cksum, zc)) 1592 panic("buffer modified while frozen!"); 1593 mutex_exit(&hdr->b_l1hdr.b_freeze_lock); 1594 } 1595 1596 static boolean_t 1597 arc_cksum_is_equal(arc_buf_hdr_t *hdr, zio_t *zio) 1598 { 1599 enum zio_compress compress = BP_GET_COMPRESS(zio->io_bp); 1600 boolean_t valid_cksum; 1601 1602 ASSERT(!BP_IS_EMBEDDED(zio->io_bp)); 1603 VERIFY3U(BP_GET_PSIZE(zio->io_bp), ==, HDR_GET_PSIZE(hdr)); 1604 1605 /* 1606 * We rely on the blkptr's checksum to determine if the block 1607 * is valid or not. When compressed arc is enabled, the l2arc 1608 * writes the block to the l2arc just as it appears in the pool. 1609 * This allows us to use the blkptr's checksum to validate the 1610 * data that we just read off of the l2arc without having to store 1611 * a separate checksum in the arc_buf_hdr_t. However, if compressed 1612 * arc is disabled, then the data written to the l2arc is always 1613 * uncompressed and won't match the block as it exists in the main 1614 * pool. When this is the case, we must first compress it if it is 1615 * compressed on the main pool before we can validate the checksum. 1616 */ 1617 if (!HDR_COMPRESSION_ENABLED(hdr) && compress != ZIO_COMPRESS_OFF) { 1618 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF); 1619 uint64_t lsize = HDR_GET_LSIZE(hdr); 1620 uint64_t csize; 1621 1622 abd_t *cdata = abd_alloc_linear(HDR_GET_PSIZE(hdr), B_TRUE); 1623 csize = zio_compress_data(compress, zio->io_abd, 1624 abd_to_buf(cdata), lsize); 1625 1626 ASSERT3U(csize, <=, HDR_GET_PSIZE(hdr)); 1627 if (csize < HDR_GET_PSIZE(hdr)) { 1628 /* 1629 * Compressed blocks are always a multiple of the 1630 * smallest ashift in the pool. Ideally, we would 1631 * like to round up the csize to the next 1632 * spa_min_ashift but that value may have changed 1633 * since the block was last written. Instead, 1634 * we rely on the fact that the hdr's psize 1635 * was set to the psize of the block when it was 1636 * last written. We set the csize to that value 1637 * and zero out any part that should not contain 1638 * data. 1639 */ 1640 abd_zero_off(cdata, csize, HDR_GET_PSIZE(hdr) - csize); 1641 csize = HDR_GET_PSIZE(hdr); 1642 } 1643 zio_push_transform(zio, cdata, csize, HDR_GET_PSIZE(hdr), NULL); 1644 } 1645 1646 /* 1647 * Block pointers always store the checksum for the logical data. 1648 * If the block pointer has the gang bit set, then the checksum 1649 * it represents is for the reconstituted data and not for an 1650 * individual gang member. The zio pipeline, however, must be able to 1651 * determine the checksum of each of the gang constituents so it 1652 * treats the checksum comparison differently than what we need 1653 * for l2arc blocks. This prevents us from using the 1654 * zio_checksum_error() interface directly. Instead we must call the 1655 * zio_checksum_error_impl() so that we can ensure the checksum is 1656 * generated using the correct checksum algorithm and accounts for the 1657 * logical I/O size and not just a gang fragment. 1658 */ 1659 valid_cksum = (zio_checksum_error_impl(zio->io_spa, zio->io_bp, 1660 BP_GET_CHECKSUM(zio->io_bp), zio->io_abd, zio->io_size, 1661 zio->io_offset, NULL) == 0); 1662 zio_pop_transforms(zio); 1663 return (valid_cksum); 1664 } 1665 1666 /* 1667 * Given a buf full of data, if ZFS_DEBUG_MODIFY is enabled this computes a 1668 * checksum and attaches it to the buf's hdr so that we can ensure that the buf 1669 * isn't modified later on. If buf is compressed or there is already a checksum 1670 * on the hdr, this is a no-op (we only checksum uncompressed bufs). 1671 */ 1672 static void 1673 arc_cksum_compute(arc_buf_t *buf) 1674 { 1675 arc_buf_hdr_t *hdr = buf->b_hdr; 1676 1677 if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 1678 return; 1679 1680 ASSERT(HDR_HAS_L1HDR(hdr)); 1681 1682 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1683 if (hdr->b_l1hdr.b_freeze_cksum != NULL) { 1684 ASSERT(arc_hdr_has_uncompressed_buf(hdr)); 1685 mutex_exit(&hdr->b_l1hdr.b_freeze_lock); 1686 return; 1687 } else if (ARC_BUF_COMPRESSED(buf)) { 1688 mutex_exit(&hdr->b_l1hdr.b_freeze_lock); 1689 return; 1690 } 1691 1692 ASSERT(!ARC_BUF_COMPRESSED(buf)); 1693 hdr->b_l1hdr.b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t), 1694 KM_SLEEP); 1695 fletcher_2_native(buf->b_data, arc_buf_size(buf), NULL, 1696 hdr->b_l1hdr.b_freeze_cksum); 1697 mutex_exit(&hdr->b_l1hdr.b_freeze_lock); 1698 arc_buf_watch(buf); 1699 } 1700 1701 #ifndef _KERNEL 1702 typedef struct procctl { 1703 long cmd; 1704 prwatch_t prwatch; 1705 } procctl_t; 1706 #endif 1707 1708 /* ARGSUSED */ 1709 static void 1710 arc_buf_unwatch(arc_buf_t *buf) 1711 { 1712 #ifndef _KERNEL 1713 if (arc_watch) { 1714 int result; 1715 procctl_t ctl; 1716 ctl.cmd = PCWATCH; 1717 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data; 1718 ctl.prwatch.pr_size = 0; 1719 ctl.prwatch.pr_wflags = 0; 1720 result = write(arc_procfd, &ctl, sizeof (ctl)); 1721 ASSERT3U(result, ==, sizeof (ctl)); 1722 } 1723 #endif 1724 } 1725 1726 /* ARGSUSED */ 1727 static void 1728 arc_buf_watch(arc_buf_t *buf) 1729 { 1730 #ifndef _KERNEL 1731 if (arc_watch) { 1732 int result; 1733 procctl_t ctl; 1734 ctl.cmd = PCWATCH; 1735 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data; 1736 ctl.prwatch.pr_size = arc_buf_size(buf); 1737 ctl.prwatch.pr_wflags = WA_WRITE; 1738 result = write(arc_procfd, &ctl, sizeof (ctl)); 1739 ASSERT3U(result, ==, sizeof (ctl)); 1740 } 1741 #endif 1742 } 1743 1744 static arc_buf_contents_t 1745 arc_buf_type(arc_buf_hdr_t *hdr) 1746 { 1747 arc_buf_contents_t type; 1748 if (HDR_ISTYPE_METADATA(hdr)) { 1749 type = ARC_BUFC_METADATA; 1750 } else { 1751 type = ARC_BUFC_DATA; 1752 } 1753 VERIFY3U(hdr->b_type, ==, type); 1754 return (type); 1755 } 1756 1757 boolean_t 1758 arc_is_metadata(arc_buf_t *buf) 1759 { 1760 return (HDR_ISTYPE_METADATA(buf->b_hdr) != 0); 1761 } 1762 1763 static uint32_t 1764 arc_bufc_to_flags(arc_buf_contents_t type) 1765 { 1766 switch (type) { 1767 case ARC_BUFC_DATA: 1768 /* metadata field is 0 if buffer contains normal data */ 1769 return (0); 1770 case ARC_BUFC_METADATA: 1771 return (ARC_FLAG_BUFC_METADATA); 1772 default: 1773 break; 1774 } 1775 panic("undefined ARC buffer type!"); 1776 return ((uint32_t)-1); 1777 } 1778 1779 void 1780 arc_buf_thaw(arc_buf_t *buf) 1781 { 1782 arc_buf_hdr_t *hdr = buf->b_hdr; 1783 1784 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon); 1785 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 1786 1787 arc_cksum_verify(buf); 1788 1789 /* 1790 * Compressed buffers do not manipulate the b_freeze_cksum or 1791 * allocate b_thawed. 1792 */ 1793 if (ARC_BUF_COMPRESSED(buf)) { 1794 ASSERT(hdr->b_l1hdr.b_freeze_cksum == NULL || 1795 arc_hdr_has_uncompressed_buf(hdr)); 1796 return; 1797 } 1798 1799 ASSERT(HDR_HAS_L1HDR(hdr)); 1800 arc_cksum_free(hdr); 1801 1802 mutex_enter(&hdr->b_l1hdr.b_freeze_lock); 1803 #ifdef ZFS_DEBUG 1804 if (zfs_flags & ZFS_DEBUG_MODIFY) { 1805 if (hdr->b_l1hdr.b_thawed != NULL) 1806 kmem_free(hdr->b_l1hdr.b_thawed, 1); 1807 hdr->b_l1hdr.b_thawed = kmem_alloc(1, KM_SLEEP); 1808 } 1809 #endif 1810 1811 mutex_exit(&hdr->b_l1hdr.b_freeze_lock); 1812 1813 arc_buf_unwatch(buf); 1814 } 1815 1816 void 1817 arc_buf_freeze(arc_buf_t *buf) 1818 { 1819 arc_buf_hdr_t *hdr = buf->b_hdr; 1820 kmutex_t *hash_lock; 1821 1822 if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 1823 return; 1824 1825 if (ARC_BUF_COMPRESSED(buf)) { 1826 ASSERT(hdr->b_l1hdr.b_freeze_cksum == NULL || 1827 arc_hdr_has_uncompressed_buf(hdr)); 1828 return; 1829 } 1830 1831 hash_lock = HDR_LOCK(hdr); 1832 mutex_enter(hash_lock); 1833 1834 ASSERT(HDR_HAS_L1HDR(hdr)); 1835 ASSERT(hdr->b_l1hdr.b_freeze_cksum != NULL || 1836 hdr->b_l1hdr.b_state == arc_anon); 1837 arc_cksum_compute(buf); 1838 mutex_exit(hash_lock); 1839 } 1840 1841 /* 1842 * The arc_buf_hdr_t's b_flags should never be modified directly. Instead, 1843 * the following functions should be used to ensure that the flags are 1844 * updated in a thread-safe way. When manipulating the flags either 1845 * the hash_lock must be held or the hdr must be undiscoverable. This 1846 * ensures that we're not racing with any other threads when updating 1847 * the flags. 1848 */ 1849 static inline void 1850 arc_hdr_set_flags(arc_buf_hdr_t *hdr, arc_flags_t flags) 1851 { 1852 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); 1853 hdr->b_flags |= flags; 1854 } 1855 1856 static inline void 1857 arc_hdr_clear_flags(arc_buf_hdr_t *hdr, arc_flags_t flags) 1858 { 1859 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); 1860 hdr->b_flags &= ~flags; 1861 } 1862 1863 /* 1864 * Setting the compression bits in the arc_buf_hdr_t's b_flags is 1865 * done in a special way since we have to clear and set bits 1866 * at the same time. Consumers that wish to set the compression bits 1867 * must use this function to ensure that the flags are updated in 1868 * thread-safe manner. 1869 */ 1870 static void 1871 arc_hdr_set_compress(arc_buf_hdr_t *hdr, enum zio_compress cmp) 1872 { 1873 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); 1874 1875 /* 1876 * Holes and embedded blocks will always have a psize = 0 so 1877 * we ignore the compression of the blkptr and set the 1878 * arc_buf_hdr_t's compression to ZIO_COMPRESS_OFF. 1879 * Holes and embedded blocks remain anonymous so we don't 1880 * want to uncompress them. Mark them as uncompressed. 1881 */ 1882 if (!zfs_compressed_arc_enabled || HDR_GET_PSIZE(hdr) == 0) { 1883 arc_hdr_clear_flags(hdr, ARC_FLAG_COMPRESSED_ARC); 1884 HDR_SET_COMPRESS(hdr, ZIO_COMPRESS_OFF); 1885 ASSERT(!HDR_COMPRESSION_ENABLED(hdr)); 1886 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF); 1887 } else { 1888 arc_hdr_set_flags(hdr, ARC_FLAG_COMPRESSED_ARC); 1889 HDR_SET_COMPRESS(hdr, cmp); 1890 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, cmp); 1891 ASSERT(HDR_COMPRESSION_ENABLED(hdr)); 1892 } 1893 } 1894 1895 /* 1896 * Looks for another buf on the same hdr which has the data decompressed, copies 1897 * from it, and returns true. If no such buf exists, returns false. 1898 */ 1899 static boolean_t 1900 arc_buf_try_copy_decompressed_data(arc_buf_t *buf) 1901 { 1902 arc_buf_hdr_t *hdr = buf->b_hdr; 1903 boolean_t copied = B_FALSE; 1904 1905 ASSERT(HDR_HAS_L1HDR(hdr)); 1906 ASSERT3P(buf->b_data, !=, NULL); 1907 ASSERT(!ARC_BUF_COMPRESSED(buf)); 1908 1909 for (arc_buf_t *from = hdr->b_l1hdr.b_buf; from != NULL; 1910 from = from->b_next) { 1911 /* can't use our own data buffer */ 1912 if (from == buf) { 1913 continue; 1914 } 1915 1916 if (!ARC_BUF_COMPRESSED(from)) { 1917 bcopy(from->b_data, buf->b_data, arc_buf_size(buf)); 1918 copied = B_TRUE; 1919 break; 1920 } 1921 } 1922 1923 /* 1924 * There were no decompressed bufs, so there should not be a 1925 * checksum on the hdr either. 1926 */ 1927 EQUIV(!copied, hdr->b_l1hdr.b_freeze_cksum == NULL); 1928 1929 return (copied); 1930 } 1931 1932 /* 1933 * Given a buf that has a data buffer attached to it, this function will 1934 * efficiently fill the buf with data of the specified compression setting from 1935 * the hdr and update the hdr's b_freeze_cksum if necessary. If the buf and hdr 1936 * are already sharing a data buf, no copy is performed. 1937 * 1938 * If the buf is marked as compressed but uncompressed data was requested, this 1939 * will allocate a new data buffer for the buf, remove that flag, and fill the 1940 * buf with uncompressed data. You can't request a compressed buf on a hdr with 1941 * uncompressed data, and (since we haven't added support for it yet) if you 1942 * want compressed data your buf must already be marked as compressed and have 1943 * the correct-sized data buffer. 1944 */ 1945 static int 1946 arc_buf_fill(arc_buf_t *buf, boolean_t compressed) 1947 { 1948 arc_buf_hdr_t *hdr = buf->b_hdr; 1949 boolean_t hdr_compressed = (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF); 1950 dmu_object_byteswap_t bswap = hdr->b_l1hdr.b_byteswap; 1951 1952 ASSERT3P(buf->b_data, !=, NULL); 1953 IMPLY(compressed, hdr_compressed); 1954 IMPLY(compressed, ARC_BUF_COMPRESSED(buf)); 1955 1956 if (hdr_compressed == compressed) { 1957 if (!arc_buf_is_shared(buf)) { 1958 abd_copy_to_buf(buf->b_data, hdr->b_l1hdr.b_pabd, 1959 arc_buf_size(buf)); 1960 } 1961 } else { 1962 ASSERT(hdr_compressed); 1963 ASSERT(!compressed); 1964 ASSERT3U(HDR_GET_LSIZE(hdr), !=, HDR_GET_PSIZE(hdr)); 1965 1966 /* 1967 * If the buf is sharing its data with the hdr, unlink it and 1968 * allocate a new data buffer for the buf. 1969 */ 1970 if (arc_buf_is_shared(buf)) { 1971 ASSERT(ARC_BUF_COMPRESSED(buf)); 1972 1973 /* We need to give the buf it's own b_data */ 1974 buf->b_flags &= ~ARC_BUF_FLAG_SHARED; 1975 buf->b_data = 1976 arc_get_data_buf(hdr, HDR_GET_LSIZE(hdr), buf); 1977 arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA); 1978 1979 /* Previously overhead was 0; just add new overhead */ 1980 ARCSTAT_INCR(arcstat_overhead_size, HDR_GET_LSIZE(hdr)); 1981 } else if (ARC_BUF_COMPRESSED(buf)) { 1982 /* We need to reallocate the buf's b_data */ 1983 arc_free_data_buf(hdr, buf->b_data, HDR_GET_PSIZE(hdr), 1984 buf); 1985 buf->b_data = 1986 arc_get_data_buf(hdr, HDR_GET_LSIZE(hdr), buf); 1987 1988 /* We increased the size of b_data; update overhead */ 1989 ARCSTAT_INCR(arcstat_overhead_size, 1990 HDR_GET_LSIZE(hdr) - HDR_GET_PSIZE(hdr)); 1991 } 1992 1993 /* 1994 * Regardless of the buf's previous compression settings, it 1995 * should not be compressed at the end of this function. 1996 */ 1997 buf->b_flags &= ~ARC_BUF_FLAG_COMPRESSED; 1998 1999 /* 2000 * Try copying the data from another buf which already has a 2001 * decompressed version. If that's not possible, it's time to 2002 * bite the bullet and decompress the data from the hdr. 2003 */ 2004 if (arc_buf_try_copy_decompressed_data(buf)) { 2005 /* Skip byteswapping and checksumming (already done) */ 2006 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, !=, NULL); 2007 return (0); 2008 } else { 2009 int error = zio_decompress_data(HDR_GET_COMPRESS(hdr), 2010 hdr->b_l1hdr.b_pabd, buf->b_data, 2011 HDR_GET_PSIZE(hdr), HDR_GET_LSIZE(hdr)); 2012 2013 /* 2014 * Absent hardware errors or software bugs, this should 2015 * be impossible, but log it anyway so we can debug it. 2016 */ 2017 if (error != 0) { 2018 zfs_dbgmsg( 2019 "hdr %p, compress %d, psize %d, lsize %d", 2020 hdr, HDR_GET_COMPRESS(hdr), 2021 HDR_GET_PSIZE(hdr), HDR_GET_LSIZE(hdr)); 2022 return (SET_ERROR(EIO)); 2023 } 2024 } 2025 } 2026 2027 /* Byteswap the buf's data if necessary */ 2028 if (bswap != DMU_BSWAP_NUMFUNCS) { 2029 ASSERT(!HDR_SHARED_DATA(hdr)); 2030 ASSERT3U(bswap, <, DMU_BSWAP_NUMFUNCS); 2031 dmu_ot_byteswap[bswap].ob_func(buf->b_data, HDR_GET_LSIZE(hdr)); 2032 } 2033 2034 /* Compute the hdr's checksum if necessary */ 2035 arc_cksum_compute(buf); 2036 2037 return (0); 2038 } 2039 2040 int 2041 arc_decompress(arc_buf_t *buf) 2042 { 2043 return (arc_buf_fill(buf, B_FALSE)); 2044 } 2045 2046 /* 2047 * Return the size of the block, b_pabd, that is stored in the arc_buf_hdr_t. 2048 */ 2049 static uint64_t 2050 arc_hdr_size(arc_buf_hdr_t *hdr) 2051 { 2052 uint64_t size; 2053 2054 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF && 2055 HDR_GET_PSIZE(hdr) > 0) { 2056 size = HDR_GET_PSIZE(hdr); 2057 } else { 2058 ASSERT3U(HDR_GET_LSIZE(hdr), !=, 0); 2059 size = HDR_GET_LSIZE(hdr); 2060 } 2061 return (size); 2062 } 2063 2064 /* 2065 * Increment the amount of evictable space in the arc_state_t's refcount. 2066 * We account for the space used by the hdr and the arc buf individually 2067 * so that we can add and remove them from the refcount individually. 2068 */ 2069 static void 2070 arc_evictable_space_increment(arc_buf_hdr_t *hdr, arc_state_t *state) 2071 { 2072 arc_buf_contents_t type = arc_buf_type(hdr); 2073 2074 ASSERT(HDR_HAS_L1HDR(hdr)); 2075 2076 if (GHOST_STATE(state)) { 2077 ASSERT0(hdr->b_l1hdr.b_bufcnt); 2078 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 2079 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL); 2080 (void) zfs_refcount_add_many(&state->arcs_esize[type], 2081 HDR_GET_LSIZE(hdr), hdr); 2082 return; 2083 } 2084 2085 ASSERT(!GHOST_STATE(state)); 2086 if (hdr->b_l1hdr.b_pabd != NULL) { 2087 (void) zfs_refcount_add_many(&state->arcs_esize[type], 2088 arc_hdr_size(hdr), hdr); 2089 } 2090 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL; 2091 buf = buf->b_next) { 2092 if (arc_buf_is_shared(buf)) 2093 continue; 2094 (void) zfs_refcount_add_many(&state->arcs_esize[type], 2095 arc_buf_size(buf), buf); 2096 } 2097 } 2098 2099 /* 2100 * Decrement the amount of evictable space in the arc_state_t's refcount. 2101 * We account for the space used by the hdr and the arc buf individually 2102 * so that we can add and remove them from the refcount individually. 2103 */ 2104 static void 2105 arc_evictable_space_decrement(arc_buf_hdr_t *hdr, arc_state_t *state) 2106 { 2107 arc_buf_contents_t type = arc_buf_type(hdr); 2108 2109 ASSERT(HDR_HAS_L1HDR(hdr)); 2110 2111 if (GHOST_STATE(state)) { 2112 ASSERT0(hdr->b_l1hdr.b_bufcnt); 2113 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 2114 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL); 2115 (void) zfs_refcount_remove_many(&state->arcs_esize[type], 2116 HDR_GET_LSIZE(hdr), hdr); 2117 return; 2118 } 2119 2120 ASSERT(!GHOST_STATE(state)); 2121 if (hdr->b_l1hdr.b_pabd != NULL) { 2122 (void) zfs_refcount_remove_many(&state->arcs_esize[type], 2123 arc_hdr_size(hdr), hdr); 2124 } 2125 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL; 2126 buf = buf->b_next) { 2127 if (arc_buf_is_shared(buf)) 2128 continue; 2129 (void) zfs_refcount_remove_many(&state->arcs_esize[type], 2130 arc_buf_size(buf), buf); 2131 } 2132 } 2133 2134 /* 2135 * Add a reference to this hdr indicating that someone is actively 2136 * referencing that memory. When the refcount transitions from 0 to 1, 2137 * we remove it from the respective arc_state_t list to indicate that 2138 * it is not evictable. 2139 */ 2140 static void 2141 add_reference(arc_buf_hdr_t *hdr, void *tag) 2142 { 2143 ASSERT(HDR_HAS_L1HDR(hdr)); 2144 if (!MUTEX_HELD(HDR_LOCK(hdr))) { 2145 ASSERT(hdr->b_l1hdr.b_state == arc_anon); 2146 ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 2147 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 2148 } 2149 2150 arc_state_t *state = hdr->b_l1hdr.b_state; 2151 2152 if ((zfs_refcount_add(&hdr->b_l1hdr.b_refcnt, tag) == 1) && 2153 (state != arc_anon)) { 2154 /* We don't use the L2-only state list. */ 2155 if (state != arc_l2c_only) { 2156 multilist_remove(state->arcs_list[arc_buf_type(hdr)], 2157 hdr); 2158 arc_evictable_space_decrement(hdr, state); 2159 } 2160 /* remove the prefetch flag if we get a reference */ 2161 arc_hdr_clear_flags(hdr, ARC_FLAG_PREFETCH); 2162 } 2163 } 2164 2165 /* 2166 * Remove a reference from this hdr. When the reference transitions from 2167 * 1 to 0 and we're not anonymous, then we add this hdr to the arc_state_t's 2168 * list making it eligible for eviction. 2169 */ 2170 static int 2171 remove_reference(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, void *tag) 2172 { 2173 int cnt; 2174 arc_state_t *state = hdr->b_l1hdr.b_state; 2175 2176 ASSERT(HDR_HAS_L1HDR(hdr)); 2177 ASSERT(state == arc_anon || MUTEX_HELD(hash_lock)); 2178 ASSERT(!GHOST_STATE(state)); 2179 2180 /* 2181 * arc_l2c_only counts as a ghost state so we don't need to explicitly 2182 * check to prevent usage of the arc_l2c_only list. 2183 */ 2184 if (((cnt = zfs_refcount_remove(&hdr->b_l1hdr.b_refcnt, tag)) == 0) && 2185 (state != arc_anon)) { 2186 multilist_insert(state->arcs_list[arc_buf_type(hdr)], hdr); 2187 ASSERT3U(hdr->b_l1hdr.b_bufcnt, >, 0); 2188 arc_evictable_space_increment(hdr, state); 2189 } 2190 return (cnt); 2191 } 2192 2193 /* 2194 * Move the supplied buffer to the indicated state. The hash lock 2195 * for the buffer must be held by the caller. 2196 */ 2197 static void 2198 arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *hdr, 2199 kmutex_t *hash_lock) 2200 { 2201 arc_state_t *old_state; 2202 int64_t refcnt; 2203 uint32_t bufcnt; 2204 boolean_t update_old, update_new; 2205 arc_buf_contents_t buftype = arc_buf_type(hdr); 2206 2207 /* 2208 * We almost always have an L1 hdr here, since we call arc_hdr_realloc() 2209 * in arc_read() when bringing a buffer out of the L2ARC. However, the 2210 * L1 hdr doesn't always exist when we change state to arc_anon before 2211 * destroying a header, in which case reallocating to add the L1 hdr is 2212 * pointless. 2213 */ 2214 if (HDR_HAS_L1HDR(hdr)) { 2215 old_state = hdr->b_l1hdr.b_state; 2216 refcnt = zfs_refcount_count(&hdr->b_l1hdr.b_refcnt); 2217 bufcnt = hdr->b_l1hdr.b_bufcnt; 2218 update_old = (bufcnt > 0 || hdr->b_l1hdr.b_pabd != NULL); 2219 } else { 2220 old_state = arc_l2c_only; 2221 refcnt = 0; 2222 bufcnt = 0; 2223 update_old = B_FALSE; 2224 } 2225 update_new = update_old; 2226 2227 ASSERT(MUTEX_HELD(hash_lock)); 2228 ASSERT3P(new_state, !=, old_state); 2229 ASSERT(!GHOST_STATE(new_state) || bufcnt == 0); 2230 ASSERT(old_state != arc_anon || bufcnt <= 1); 2231 2232 /* 2233 * If this buffer is evictable, transfer it from the 2234 * old state list to the new state list. 2235 */ 2236 if (refcnt == 0) { 2237 if (old_state != arc_anon && old_state != arc_l2c_only) { 2238 ASSERT(HDR_HAS_L1HDR(hdr)); 2239 multilist_remove(old_state->arcs_list[buftype], hdr); 2240 2241 if (GHOST_STATE(old_state)) { 2242 ASSERT0(bufcnt); 2243 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 2244 update_old = B_TRUE; 2245 } 2246 arc_evictable_space_decrement(hdr, old_state); 2247 } 2248 if (new_state != arc_anon && new_state != arc_l2c_only) { 2249 2250 /* 2251 * An L1 header always exists here, since if we're 2252 * moving to some L1-cached state (i.e. not l2c_only or 2253 * anonymous), we realloc the header to add an L1hdr 2254 * beforehand. 2255 */ 2256 ASSERT(HDR_HAS_L1HDR(hdr)); 2257 multilist_insert(new_state->arcs_list[buftype], hdr); 2258 2259 if (GHOST_STATE(new_state)) { 2260 ASSERT0(bufcnt); 2261 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 2262 update_new = B_TRUE; 2263 } 2264 arc_evictable_space_increment(hdr, new_state); 2265 } 2266 } 2267 2268 ASSERT(!HDR_EMPTY(hdr)); 2269 if (new_state == arc_anon && HDR_IN_HASH_TABLE(hdr)) 2270 buf_hash_remove(hdr); 2271 2272 /* adjust state sizes (ignore arc_l2c_only) */ 2273 2274 if (update_new && new_state != arc_l2c_only) { 2275 ASSERT(HDR_HAS_L1HDR(hdr)); 2276 if (GHOST_STATE(new_state)) { 2277 ASSERT0(bufcnt); 2278 2279 /* 2280 * When moving a header to a ghost state, we first 2281 * remove all arc buffers. Thus, we'll have a 2282 * bufcnt of zero, and no arc buffer to use for 2283 * the reference. As a result, we use the arc 2284 * header pointer for the reference. 2285 */ 2286 (void) zfs_refcount_add_many(&new_state->arcs_size, 2287 HDR_GET_LSIZE(hdr), hdr); 2288 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL); 2289 } else { 2290 uint32_t buffers = 0; 2291 2292 /* 2293 * Each individual buffer holds a unique reference, 2294 * thus we must remove each of these references one 2295 * at a time. 2296 */ 2297 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL; 2298 buf = buf->b_next) { 2299 ASSERT3U(bufcnt, !=, 0); 2300 buffers++; 2301 2302 /* 2303 * When the arc_buf_t is sharing the data 2304 * block with the hdr, the owner of the 2305 * reference belongs to the hdr. Only 2306 * add to the refcount if the arc_buf_t is 2307 * not shared. 2308 */ 2309 if (arc_buf_is_shared(buf)) 2310 continue; 2311 2312 (void) zfs_refcount_add_many( 2313 &new_state->arcs_size, 2314 arc_buf_size(buf), buf); 2315 } 2316 ASSERT3U(bufcnt, ==, buffers); 2317 2318 if (hdr->b_l1hdr.b_pabd != NULL) { 2319 (void) zfs_refcount_add_many( 2320 &new_state->arcs_size, 2321 arc_hdr_size(hdr), hdr); 2322 } else { 2323 ASSERT(GHOST_STATE(old_state)); 2324 } 2325 } 2326 } 2327 2328 if (update_old && old_state != arc_l2c_only) { 2329 ASSERT(HDR_HAS_L1HDR(hdr)); 2330 if (GHOST_STATE(old_state)) { 2331 ASSERT0(bufcnt); 2332 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL); 2333 2334 /* 2335 * When moving a header off of a ghost state, 2336 * the header will not contain any arc buffers. 2337 * We use the arc header pointer for the reference 2338 * which is exactly what we did when we put the 2339 * header on the ghost state. 2340 */ 2341 2342 (void) zfs_refcount_remove_many(&old_state->arcs_size, 2343 HDR_GET_LSIZE(hdr), hdr); 2344 } else { 2345 uint32_t buffers = 0; 2346 2347 /* 2348 * Each individual buffer holds a unique reference, 2349 * thus we must remove each of these references one 2350 * at a time. 2351 */ 2352 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL; 2353 buf = buf->b_next) { 2354 ASSERT3U(bufcnt, !=, 0); 2355 buffers++; 2356 2357 /* 2358 * When the arc_buf_t is sharing the data 2359 * block with the hdr, the owner of the 2360 * reference belongs to the hdr. Only 2361 * add to the refcount if the arc_buf_t is 2362 * not shared. 2363 */ 2364 if (arc_buf_is_shared(buf)) 2365 continue; 2366 2367 (void) zfs_refcount_remove_many( 2368 &old_state->arcs_size, arc_buf_size(buf), 2369 buf); 2370 } 2371 ASSERT3U(bufcnt, ==, buffers); 2372 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL); 2373 (void) zfs_refcount_remove_many( 2374 &old_state->arcs_size, arc_hdr_size(hdr), hdr); 2375 } 2376 } 2377 2378 if (HDR_HAS_L1HDR(hdr)) 2379 hdr->b_l1hdr.b_state = new_state; 2380 2381 /* 2382 * L2 headers should never be on the L2 state list since they don't 2383 * have L1 headers allocated. 2384 */ 2385 ASSERT(multilist_is_empty(arc_l2c_only->arcs_list[ARC_BUFC_DATA]) && 2386 multilist_is_empty(arc_l2c_only->arcs_list[ARC_BUFC_METADATA])); 2387 } 2388 2389 void 2390 arc_space_consume(uint64_t space, arc_space_type_t type) 2391 { 2392 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES); 2393 2394 switch (type) { 2395 case ARC_SPACE_DATA: 2396 aggsum_add(&astat_data_size, space); 2397 break; 2398 case ARC_SPACE_META: 2399 aggsum_add(&astat_metadata_size, space); 2400 break; 2401 case ARC_SPACE_OTHER: 2402 aggsum_add(&astat_other_size, space); 2403 break; 2404 case ARC_SPACE_HDRS: 2405 aggsum_add(&astat_hdr_size, space); 2406 break; 2407 case ARC_SPACE_L2HDRS: 2408 aggsum_add(&astat_l2_hdr_size, space); 2409 break; 2410 } 2411 2412 if (type != ARC_SPACE_DATA) 2413 aggsum_add(&arc_meta_used, space); 2414 2415 aggsum_add(&arc_size, space); 2416 } 2417 2418 void 2419 arc_space_return(uint64_t space, arc_space_type_t type) 2420 { 2421 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES); 2422 2423 switch (type) { 2424 case ARC_SPACE_DATA: 2425 aggsum_add(&astat_data_size, -space); 2426 break; 2427 case ARC_SPACE_META: 2428 aggsum_add(&astat_metadata_size, -space); 2429 break; 2430 case ARC_SPACE_OTHER: 2431 aggsum_add(&astat_other_size, -space); 2432 break; 2433 case ARC_SPACE_HDRS: 2434 aggsum_add(&astat_hdr_size, -space); 2435 break; 2436 case ARC_SPACE_L2HDRS: 2437 aggsum_add(&astat_l2_hdr_size, -space); 2438 break; 2439 } 2440 2441 if (type != ARC_SPACE_DATA) { 2442 ASSERT(aggsum_compare(&arc_meta_used, space) >= 0); 2443 /* 2444 * We use the upper bound here rather than the precise value 2445 * because the arc_meta_max value doesn't need to be 2446 * precise. It's only consumed by humans via arcstats. 2447 */ 2448 if (arc_meta_max < aggsum_upper_bound(&arc_meta_used)) 2449 arc_meta_max = aggsum_upper_bound(&arc_meta_used); 2450 aggsum_add(&arc_meta_used, -space); 2451 } 2452 2453 ASSERT(aggsum_compare(&arc_size, space) >= 0); 2454 aggsum_add(&arc_size, -space); 2455 } 2456 2457 /* 2458 * Given a hdr and a buf, returns whether that buf can share its b_data buffer 2459 * with the hdr's b_pabd. 2460 */ 2461 static boolean_t 2462 arc_can_share(arc_buf_hdr_t *hdr, arc_buf_t *buf) 2463 { 2464 /* 2465 * The criteria for sharing a hdr's data are: 2466 * 1. the hdr's compression matches the buf's compression 2467 * 2. the hdr doesn't need to be byteswapped 2468 * 3. the hdr isn't already being shared 2469 * 4. the buf is either compressed or it is the last buf in the hdr list 2470 * 2471 * Criterion #4 maintains the invariant that shared uncompressed 2472 * bufs must be the final buf in the hdr's b_buf list. Reading this, you 2473 * might ask, "if a compressed buf is allocated first, won't that be the 2474 * last thing in the list?", but in that case it's impossible to create 2475 * a shared uncompressed buf anyway (because the hdr must be compressed 2476 * to have the compressed buf). You might also think that #3 is 2477 * sufficient to make this guarantee, however it's possible 2478 * (specifically in the rare L2ARC write race mentioned in 2479 * arc_buf_alloc_impl()) there will be an existing uncompressed buf that 2480 * is sharable, but wasn't at the time of its allocation. Rather than 2481 * allow a new shared uncompressed buf to be created and then shuffle 2482 * the list around to make it the last element, this simply disallows 2483 * sharing if the new buf isn't the first to be added. 2484 */ 2485 ASSERT3P(buf->b_hdr, ==, hdr); 2486 boolean_t hdr_compressed = HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF; 2487 boolean_t buf_compressed = ARC_BUF_COMPRESSED(buf) != 0; 2488 return (buf_compressed == hdr_compressed && 2489 hdr->b_l1hdr.b_byteswap == DMU_BSWAP_NUMFUNCS && 2490 !HDR_SHARED_DATA(hdr) && 2491 (ARC_BUF_LAST(buf) || ARC_BUF_COMPRESSED(buf))); 2492 } 2493 2494 /* 2495 * Allocate a buf for this hdr. If you care about the data that's in the hdr, 2496 * or if you want a compressed buffer, pass those flags in. Returns 0 if the 2497 * copy was made successfully, or an error code otherwise. 2498 */ 2499 static int 2500 arc_buf_alloc_impl(arc_buf_hdr_t *hdr, void *tag, boolean_t compressed, 2501 boolean_t fill, arc_buf_t **ret) 2502 { 2503 arc_buf_t *buf; 2504 2505 ASSERT(HDR_HAS_L1HDR(hdr)); 2506 ASSERT3U(HDR_GET_LSIZE(hdr), >, 0); 2507 VERIFY(hdr->b_type == ARC_BUFC_DATA || 2508 hdr->b_type == ARC_BUFC_METADATA); 2509 ASSERT3P(ret, !=, NULL); 2510 ASSERT3P(*ret, ==, NULL); 2511 2512 buf = *ret = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 2513 buf->b_hdr = hdr; 2514 buf->b_data = NULL; 2515 buf->b_next = hdr->b_l1hdr.b_buf; 2516 buf->b_flags = 0; 2517 2518 add_reference(hdr, tag); 2519 2520 /* 2521 * We're about to change the hdr's b_flags. We must either 2522 * hold the hash_lock or be undiscoverable. 2523 */ 2524 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); 2525 2526 /* 2527 * Only honor requests for compressed bufs if the hdr is actually 2528 * compressed. 2529 */ 2530 if (compressed && HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF) 2531 buf->b_flags |= ARC_BUF_FLAG_COMPRESSED; 2532 2533 /* 2534 * If the hdr's data can be shared then we share the data buffer and 2535 * set the appropriate bit in the hdr's b_flags to indicate the hdr is 2536 * sharing it's b_pabd with the arc_buf_t. Otherwise, we allocate a new 2537 * buffer to store the buf's data. 2538 * 2539 * There are two additional restrictions here because we're sharing 2540 * hdr -> buf instead of the usual buf -> hdr. First, the hdr can't be 2541 * actively involved in an L2ARC write, because if this buf is used by 2542 * an arc_write() then the hdr's data buffer will be released when the 2543 * write completes, even though the L2ARC write might still be using it. 2544 * Second, the hdr's ABD must be linear so that the buf's user doesn't 2545 * need to be ABD-aware. 2546 */ 2547 boolean_t can_share = arc_can_share(hdr, buf) && !HDR_L2_WRITING(hdr) && 2548 abd_is_linear(hdr->b_l1hdr.b_pabd); 2549 2550 /* Set up b_data and sharing */ 2551 if (can_share) { 2552 buf->b_data = abd_to_buf(hdr->b_l1hdr.b_pabd); 2553 buf->b_flags |= ARC_BUF_FLAG_SHARED; 2554 arc_hdr_set_flags(hdr, ARC_FLAG_SHARED_DATA); 2555 } else { 2556 buf->b_data = 2557 arc_get_data_buf(hdr, arc_buf_size(buf), buf); 2558 ARCSTAT_INCR(arcstat_overhead_size, arc_buf_size(buf)); 2559 } 2560 VERIFY3P(buf->b_data, !=, NULL); 2561 2562 hdr->b_l1hdr.b_buf = buf; 2563 hdr->b_l1hdr.b_bufcnt += 1; 2564 2565 /* 2566 * If the user wants the data from the hdr, we need to either copy or 2567 * decompress the data. 2568 */ 2569 if (fill) { 2570 return (arc_buf_fill(buf, ARC_BUF_COMPRESSED(buf) != 0)); 2571 } 2572 2573 return (0); 2574 } 2575 2576 static char *arc_onloan_tag = "onloan"; 2577 2578 static inline void 2579 arc_loaned_bytes_update(int64_t delta) 2580 { 2581 atomic_add_64(&arc_loaned_bytes, delta); 2582 2583 /* assert that it did not wrap around */ 2584 ASSERT3S(atomic_add_64_nv(&arc_loaned_bytes, 0), >=, 0); 2585 } 2586 2587 /* 2588 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in 2589 * flight data by arc_tempreserve_space() until they are "returned". Loaned 2590 * buffers must be returned to the arc before they can be used by the DMU or 2591 * freed. 2592 */ 2593 arc_buf_t * 2594 arc_loan_buf(spa_t *spa, boolean_t is_metadata, int size) 2595 { 2596 arc_buf_t *buf = arc_alloc_buf(spa, arc_onloan_tag, 2597 is_metadata ? ARC_BUFC_METADATA : ARC_BUFC_DATA, size); 2598 2599 arc_loaned_bytes_update(arc_buf_size(buf)); 2600 2601 return (buf); 2602 } 2603 2604 arc_buf_t * 2605 arc_loan_compressed_buf(spa_t *spa, uint64_t psize, uint64_t lsize, 2606 enum zio_compress compression_type) 2607 { 2608 arc_buf_t *buf = arc_alloc_compressed_buf(spa, arc_onloan_tag, 2609 psize, lsize, compression_type); 2610 2611 arc_loaned_bytes_update(arc_buf_size(buf)); 2612 2613 return (buf); 2614 } 2615 2616 2617 /* 2618 * Return a loaned arc buffer to the arc. 2619 */ 2620 void 2621 arc_return_buf(arc_buf_t *buf, void *tag) 2622 { 2623 arc_buf_hdr_t *hdr = buf->b_hdr; 2624 2625 ASSERT3P(buf->b_data, !=, NULL); 2626 ASSERT(HDR_HAS_L1HDR(hdr)); 2627 (void) zfs_refcount_add(&hdr->b_l1hdr.b_refcnt, tag); 2628 (void) zfs_refcount_remove(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag); 2629 2630 arc_loaned_bytes_update(-arc_buf_size(buf)); 2631 } 2632 2633 /* Detach an arc_buf from a dbuf (tag) */ 2634 void 2635 arc_loan_inuse_buf(arc_buf_t *buf, void *tag) 2636 { 2637 arc_buf_hdr_t *hdr = buf->b_hdr; 2638 2639 ASSERT3P(buf->b_data, !=, NULL); 2640 ASSERT(HDR_HAS_L1HDR(hdr)); 2641 (void) zfs_refcount_add(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag); 2642 (void) zfs_refcount_remove(&hdr->b_l1hdr.b_refcnt, tag); 2643 2644 arc_loaned_bytes_update(arc_buf_size(buf)); 2645 } 2646 2647 static void 2648 l2arc_free_abd_on_write(abd_t *abd, size_t size, arc_buf_contents_t type) 2649 { 2650 l2arc_data_free_t *df = kmem_alloc(sizeof (*df), KM_SLEEP); 2651 2652 df->l2df_abd = abd; 2653 df->l2df_size = size; 2654 df->l2df_type = type; 2655 mutex_enter(&l2arc_free_on_write_mtx); 2656 list_insert_head(l2arc_free_on_write, df); 2657 mutex_exit(&l2arc_free_on_write_mtx); 2658 } 2659 2660 static void 2661 arc_hdr_free_on_write(arc_buf_hdr_t *hdr) 2662 { 2663 arc_state_t *state = hdr->b_l1hdr.b_state; 2664 arc_buf_contents_t type = arc_buf_type(hdr); 2665 uint64_t size = arc_hdr_size(hdr); 2666 2667 /* protected by hash lock, if in the hash table */ 2668 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) { 2669 ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 2670 ASSERT(state != arc_anon && state != arc_l2c_only); 2671 2672 (void) zfs_refcount_remove_many(&state->arcs_esize[type], 2673 size, hdr); 2674 } 2675 (void) zfs_refcount_remove_many(&state->arcs_size, size, hdr); 2676 if (type == ARC_BUFC_METADATA) { 2677 arc_space_return(size, ARC_SPACE_META); 2678 } else { 2679 ASSERT(type == ARC_BUFC_DATA); 2680 arc_space_return(size, ARC_SPACE_DATA); 2681 } 2682 2683 l2arc_free_abd_on_write(hdr->b_l1hdr.b_pabd, size, type); 2684 } 2685 2686 /* 2687 * Share the arc_buf_t's data with the hdr. Whenever we are sharing the 2688 * data buffer, we transfer the refcount ownership to the hdr and update 2689 * the appropriate kstats. 2690 */ 2691 static void 2692 arc_share_buf(arc_buf_hdr_t *hdr, arc_buf_t *buf) 2693 { 2694 arc_state_t *state = hdr->b_l1hdr.b_state; 2695 2696 ASSERT(arc_can_share(hdr, buf)); 2697 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL); 2698 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); 2699 2700 /* 2701 * Start sharing the data buffer. We transfer the 2702 * refcount ownership to the hdr since it always owns 2703 * the refcount whenever an arc_buf_t is shared. 2704 */ 2705 zfs_refcount_transfer_ownership(&state->arcs_size, buf, hdr); 2706 hdr->b_l1hdr.b_pabd = abd_get_from_buf(buf->b_data, arc_buf_size(buf)); 2707 abd_take_ownership_of_buf(hdr->b_l1hdr.b_pabd, 2708 HDR_ISTYPE_METADATA(hdr)); 2709 arc_hdr_set_flags(hdr, ARC_FLAG_SHARED_DATA); 2710 buf->b_flags |= ARC_BUF_FLAG_SHARED; 2711 2712 /* 2713 * Since we've transferred ownership to the hdr we need 2714 * to increment its compressed and uncompressed kstats and 2715 * decrement the overhead size. 2716 */ 2717 ARCSTAT_INCR(arcstat_compressed_size, arc_hdr_size(hdr)); 2718 ARCSTAT_INCR(arcstat_uncompressed_size, HDR_GET_LSIZE(hdr)); 2719 ARCSTAT_INCR(arcstat_overhead_size, -arc_buf_size(buf)); 2720 } 2721 2722 static void 2723 arc_unshare_buf(arc_buf_hdr_t *hdr, arc_buf_t *buf) 2724 { 2725 arc_state_t *state = hdr->b_l1hdr.b_state; 2726 2727 ASSERT(arc_buf_is_shared(buf)); 2728 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL); 2729 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); 2730 2731 /* 2732 * We are no longer sharing this buffer so we need 2733 * to transfer its ownership to the rightful owner. 2734 */ 2735 zfs_refcount_transfer_ownership(&state->arcs_size, hdr, buf); 2736 arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA); 2737 abd_release_ownership_of_buf(hdr->b_l1hdr.b_pabd); 2738 abd_put(hdr->b_l1hdr.b_pabd); 2739 hdr->b_l1hdr.b_pabd = NULL; 2740 buf->b_flags &= ~ARC_BUF_FLAG_SHARED; 2741 2742 /* 2743 * Since the buffer is no longer shared between 2744 * the arc buf and the hdr, count it as overhead. 2745 */ 2746 ARCSTAT_INCR(arcstat_compressed_size, -arc_hdr_size(hdr)); 2747 ARCSTAT_INCR(arcstat_uncompressed_size, -HDR_GET_LSIZE(hdr)); 2748 ARCSTAT_INCR(arcstat_overhead_size, arc_buf_size(buf)); 2749 } 2750 2751 /* 2752 * Remove an arc_buf_t from the hdr's buf list and return the last 2753 * arc_buf_t on the list. If no buffers remain on the list then return 2754 * NULL. 2755 */ 2756 static arc_buf_t * 2757 arc_buf_remove(arc_buf_hdr_t *hdr, arc_buf_t *buf) 2758 { 2759 ASSERT(HDR_HAS_L1HDR(hdr)); 2760 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); 2761 2762 arc_buf_t **bufp = &hdr->b_l1hdr.b_buf; 2763 arc_buf_t *lastbuf = NULL; 2764 2765 /* 2766 * Remove the buf from the hdr list and locate the last 2767 * remaining buffer on the list. 2768 */ 2769 while (*bufp != NULL) { 2770 if (*bufp == buf) 2771 *bufp = buf->b_next; 2772 2773 /* 2774 * If we've removed a buffer in the middle of 2775 * the list then update the lastbuf and update 2776 * bufp. 2777 */ 2778 if (*bufp != NULL) { 2779 lastbuf = *bufp; 2780 bufp = &(*bufp)->b_next; 2781 } 2782 } 2783 buf->b_next = NULL; 2784 ASSERT3P(lastbuf, !=, buf); 2785 IMPLY(hdr->b_l1hdr.b_bufcnt > 0, lastbuf != NULL); 2786 IMPLY(hdr->b_l1hdr.b_bufcnt > 0, hdr->b_l1hdr.b_buf != NULL); 2787 IMPLY(lastbuf != NULL, ARC_BUF_LAST(lastbuf)); 2788 2789 return (lastbuf); 2790 } 2791 2792 /* 2793 * Free up buf->b_data and pull the arc_buf_t off of the the arc_buf_hdr_t's 2794 * list and free it. 2795 */ 2796 static void 2797 arc_buf_destroy_impl(arc_buf_t *buf) 2798 { 2799 arc_buf_hdr_t *hdr = buf->b_hdr; 2800 2801 /* 2802 * Free up the data associated with the buf but only if we're not 2803 * sharing this with the hdr. If we are sharing it with the hdr, the 2804 * hdr is responsible for doing the free. 2805 */ 2806 if (buf->b_data != NULL) { 2807 /* 2808 * We're about to change the hdr's b_flags. We must either 2809 * hold the hash_lock or be undiscoverable. 2810 */ 2811 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); 2812 2813 arc_cksum_verify(buf); 2814 arc_buf_unwatch(buf); 2815 2816 if (arc_buf_is_shared(buf)) { 2817 arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA); 2818 } else { 2819 uint64_t size = arc_buf_size(buf); 2820 arc_free_data_buf(hdr, buf->b_data, size, buf); 2821 ARCSTAT_INCR(arcstat_overhead_size, -size); 2822 } 2823 buf->b_data = NULL; 2824 2825 ASSERT(hdr->b_l1hdr.b_bufcnt > 0); 2826 hdr->b_l1hdr.b_bufcnt -= 1; 2827 } 2828 2829 arc_buf_t *lastbuf = arc_buf_remove(hdr, buf); 2830 2831 if (ARC_BUF_SHARED(buf) && !ARC_BUF_COMPRESSED(buf)) { 2832 /* 2833 * If the current arc_buf_t is sharing its data buffer with the 2834 * hdr, then reassign the hdr's b_pabd to share it with the new 2835 * buffer at the end of the list. The shared buffer is always 2836 * the last one on the hdr's buffer list. 2837 * 2838 * There is an equivalent case for compressed bufs, but since 2839 * they aren't guaranteed to be the last buf in the list and 2840 * that is an exceedingly rare case, we just allow that space be 2841 * wasted temporarily. 2842 */ 2843 if (lastbuf != NULL) { 2844 /* Only one buf can be shared at once */ 2845 VERIFY(!arc_buf_is_shared(lastbuf)); 2846 /* hdr is uncompressed so can't have compressed buf */ 2847 VERIFY(!ARC_BUF_COMPRESSED(lastbuf)); 2848 2849 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL); 2850 arc_hdr_free_pabd(hdr); 2851 2852 /* 2853 * We must setup a new shared block between the 2854 * last buffer and the hdr. The data would have 2855 * been allocated by the arc buf so we need to transfer 2856 * ownership to the hdr since it's now being shared. 2857 */ 2858 arc_share_buf(hdr, lastbuf); 2859 } 2860 } else if (HDR_SHARED_DATA(hdr)) { 2861 /* 2862 * Uncompressed shared buffers are always at the end 2863 * of the list. Compressed buffers don't have the 2864 * same requirements. This makes it hard to 2865 * simply assert that the lastbuf is shared so 2866 * we rely on the hdr's compression flags to determine 2867 * if we have a compressed, shared buffer. 2868 */ 2869 ASSERT3P(lastbuf, !=, NULL); 2870 ASSERT(arc_buf_is_shared(lastbuf) || 2871 HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF); 2872 } 2873 2874 /* 2875 * Free the checksum if we're removing the last uncompressed buf from 2876 * this hdr. 2877 */ 2878 if (!arc_hdr_has_uncompressed_buf(hdr)) { 2879 arc_cksum_free(hdr); 2880 } 2881 2882 /* clean up the buf */ 2883 buf->b_hdr = NULL; 2884 kmem_cache_free(buf_cache, buf); 2885 } 2886 2887 static void 2888 arc_hdr_alloc_pabd(arc_buf_hdr_t *hdr) 2889 { 2890 ASSERT3U(HDR_GET_LSIZE(hdr), >, 0); 2891 ASSERT(HDR_HAS_L1HDR(hdr)); 2892 ASSERT(!HDR_SHARED_DATA(hdr)); 2893 2894 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL); 2895 hdr->b_l1hdr.b_pabd = arc_get_data_abd(hdr, arc_hdr_size(hdr), hdr); 2896 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS; 2897 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL); 2898 2899 ARCSTAT_INCR(arcstat_compressed_size, arc_hdr_size(hdr)); 2900 ARCSTAT_INCR(arcstat_uncompressed_size, HDR_GET_LSIZE(hdr)); 2901 } 2902 2903 static void 2904 arc_hdr_free_pabd(arc_buf_hdr_t *hdr) 2905 { 2906 ASSERT(HDR_HAS_L1HDR(hdr)); 2907 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL); 2908 2909 /* 2910 * If the hdr is currently being written to the l2arc then 2911 * we defer freeing the data by adding it to the l2arc_free_on_write 2912 * list. The l2arc will free the data once it's finished 2913 * writing it to the l2arc device. 2914 */ 2915 if (HDR_L2_WRITING(hdr)) { 2916 arc_hdr_free_on_write(hdr); 2917 ARCSTAT_BUMP(arcstat_l2_free_on_write); 2918 } else { 2919 arc_free_data_abd(hdr, hdr->b_l1hdr.b_pabd, 2920 arc_hdr_size(hdr), hdr); 2921 } 2922 hdr->b_l1hdr.b_pabd = NULL; 2923 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS; 2924 2925 ARCSTAT_INCR(arcstat_compressed_size, -arc_hdr_size(hdr)); 2926 ARCSTAT_INCR(arcstat_uncompressed_size, -HDR_GET_LSIZE(hdr)); 2927 } 2928 2929 static arc_buf_hdr_t * 2930 arc_hdr_alloc(uint64_t spa, int32_t psize, int32_t lsize, 2931 enum zio_compress compression_type, arc_buf_contents_t type) 2932 { 2933 arc_buf_hdr_t *hdr; 2934 2935 VERIFY(type == ARC_BUFC_DATA || type == ARC_BUFC_METADATA); 2936 2937 hdr = kmem_cache_alloc(hdr_full_cache, KM_PUSHPAGE); 2938 ASSERT(HDR_EMPTY(hdr)); 2939 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL); 2940 ASSERT3P(hdr->b_l1hdr.b_thawed, ==, NULL); 2941 HDR_SET_PSIZE(hdr, psize); 2942 HDR_SET_LSIZE(hdr, lsize); 2943 hdr->b_spa = spa; 2944 hdr->b_type = type; 2945 hdr->b_flags = 0; 2946 arc_hdr_set_flags(hdr, arc_bufc_to_flags(type) | ARC_FLAG_HAS_L1HDR); 2947 arc_hdr_set_compress(hdr, compression_type); 2948 2949 hdr->b_l1hdr.b_state = arc_anon; 2950 hdr->b_l1hdr.b_arc_access = 0; 2951 hdr->b_l1hdr.b_bufcnt = 0; 2952 hdr->b_l1hdr.b_buf = NULL; 2953 2954 /* 2955 * Allocate the hdr's buffer. This will contain either 2956 * the compressed or uncompressed data depending on the block 2957 * it references and compressed arc enablement. 2958 */ 2959 arc_hdr_alloc_pabd(hdr); 2960 ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 2961 2962 return (hdr); 2963 } 2964 2965 /* 2966 * Transition between the two allocation states for the arc_buf_hdr struct. 2967 * The arc_buf_hdr struct can be allocated with (hdr_full_cache) or without 2968 * (hdr_l2only_cache) the fields necessary for the L1 cache - the smaller 2969 * version is used when a cache buffer is only in the L2ARC in order to reduce 2970 * memory usage. 2971 */ 2972 static arc_buf_hdr_t * 2973 arc_hdr_realloc(arc_buf_hdr_t *hdr, kmem_cache_t *old, kmem_cache_t *new) 2974 { 2975 ASSERT(HDR_HAS_L2HDR(hdr)); 2976 2977 arc_buf_hdr_t *nhdr; 2978 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev; 2979 2980 ASSERT((old == hdr_full_cache && new == hdr_l2only_cache) || 2981 (old == hdr_l2only_cache && new == hdr_full_cache)); 2982 2983 nhdr = kmem_cache_alloc(new, KM_PUSHPAGE); 2984 2985 ASSERT(MUTEX_HELD(HDR_LOCK(hdr))); 2986 buf_hash_remove(hdr); 2987 2988 bcopy(hdr, nhdr, HDR_L2ONLY_SIZE); 2989 2990 if (new == hdr_full_cache) { 2991 arc_hdr_set_flags(nhdr, ARC_FLAG_HAS_L1HDR); 2992 /* 2993 * arc_access and arc_change_state need to be aware that a 2994 * header has just come out of L2ARC, so we set its state to 2995 * l2c_only even though it's about to change. 2996 */ 2997 nhdr->b_l1hdr.b_state = arc_l2c_only; 2998 2999 /* Verify previous threads set to NULL before freeing */ 3000 ASSERT3P(nhdr->b_l1hdr.b_pabd, ==, NULL); 3001 } else { 3002 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 3003 ASSERT0(hdr->b_l1hdr.b_bufcnt); 3004 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL); 3005 3006 /* 3007 * If we've reached here, We must have been called from 3008 * arc_evict_hdr(), as such we should have already been 3009 * removed from any ghost list we were previously on 3010 * (which protects us from racing with arc_evict_state), 3011 * thus no locking is needed during this check. 3012 */ 3013 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 3014 3015 /* 3016 * A buffer must not be moved into the arc_l2c_only 3017 * state if it's not finished being written out to the 3018 * l2arc device. Otherwise, the b_l1hdr.b_pabd field 3019 * might try to be accessed, even though it was removed. 3020 */ 3021 VERIFY(!HDR_L2_WRITING(hdr)); 3022 VERIFY3P(hdr->b_l1hdr.b_pabd, ==, NULL); 3023 3024 #ifdef ZFS_DEBUG 3025 if (hdr->b_l1hdr.b_thawed != NULL) { 3026 kmem_free(hdr->b_l1hdr.b_thawed, 1); 3027 hdr->b_l1hdr.b_thawed = NULL; 3028 } 3029 #endif 3030 3031 arc_hdr_clear_flags(nhdr, ARC_FLAG_HAS_L1HDR); 3032 } 3033 /* 3034 * The header has been reallocated so we need to re-insert it into any 3035 * lists it was on. 3036 */ 3037 (void) buf_hash_insert(nhdr, NULL); 3038 3039 ASSERT(list_link_active(&hdr->b_l2hdr.b_l2node)); 3040 3041 mutex_enter(&dev->l2ad_mtx); 3042 3043 /* 3044 * We must place the realloc'ed header back into the list at 3045 * the same spot. Otherwise, if it's placed earlier in the list, 3046 * l2arc_write_buffers() could find it during the function's 3047 * write phase, and try to write it out to the l2arc. 3048 */ 3049 list_insert_after(&dev->l2ad_buflist, hdr, nhdr); 3050 list_remove(&dev->l2ad_buflist, hdr); 3051 3052 mutex_exit(&dev->l2ad_mtx); 3053 3054 /* 3055 * Since we're using the pointer address as the tag when 3056 * incrementing and decrementing the l2ad_alloc refcount, we 3057 * must remove the old pointer (that we're about to destroy) and 3058 * add the new pointer to the refcount. Otherwise we'd remove 3059 * the wrong pointer address when calling arc_hdr_destroy() later. 3060 */ 3061 3062 (void) zfs_refcount_remove_many(&dev->l2ad_alloc, arc_hdr_size(hdr), 3063 hdr); 3064 (void) zfs_refcount_add_many(&dev->l2ad_alloc, arc_hdr_size(nhdr), 3065 nhdr); 3066 3067 buf_discard_identity(hdr); 3068 kmem_cache_free(old, hdr); 3069 3070 return (nhdr); 3071 } 3072 3073 /* 3074 * Allocate a new arc_buf_hdr_t and arc_buf_t and return the buf to the caller. 3075 * The buf is returned thawed since we expect the consumer to modify it. 3076 */ 3077 arc_buf_t * 3078 arc_alloc_buf(spa_t *spa, void *tag, arc_buf_contents_t type, int32_t size) 3079 { 3080 arc_buf_hdr_t *hdr = arc_hdr_alloc(spa_load_guid(spa), size, size, 3081 ZIO_COMPRESS_OFF, type); 3082 ASSERT(!MUTEX_HELD(HDR_LOCK(hdr))); 3083 3084 arc_buf_t *buf = NULL; 3085 VERIFY0(arc_buf_alloc_impl(hdr, tag, B_FALSE, B_FALSE, &buf)); 3086 arc_buf_thaw(buf); 3087 3088 return (buf); 3089 } 3090 3091 /* 3092 * Allocate a compressed buf in the same manner as arc_alloc_buf. Don't use this 3093 * for bufs containing metadata. 3094 */ 3095 arc_buf_t * 3096 arc_alloc_compressed_buf(spa_t *spa, void *tag, uint64_t psize, uint64_t lsize, 3097 enum zio_compress compression_type) 3098 { 3099 ASSERT3U(lsize, >, 0); 3100 ASSERT3U(lsize, >=, psize); 3101 ASSERT(compression_type > ZIO_COMPRESS_OFF); 3102 ASSERT(compression_type < ZIO_COMPRESS_FUNCTIONS); 3103 3104 arc_buf_hdr_t *hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize, 3105 compression_type, ARC_BUFC_DATA); 3106 ASSERT(!MUTEX_HELD(HDR_LOCK(hdr))); 3107 3108 arc_buf_t *buf = NULL; 3109 VERIFY0(arc_buf_alloc_impl(hdr, tag, B_TRUE, B_FALSE, &buf)); 3110 arc_buf_thaw(buf); 3111 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL); 3112 3113 if (!arc_buf_is_shared(buf)) { 3114 /* 3115 * To ensure that the hdr has the correct data in it if we call 3116 * arc_decompress() on this buf before it's been written to 3117 * disk, it's easiest if we just set up sharing between the 3118 * buf and the hdr. 3119 */ 3120 ASSERT(!abd_is_linear(hdr->b_l1hdr.b_pabd)); 3121 arc_hdr_free_pabd(hdr); 3122 arc_share_buf(hdr, buf); 3123 } 3124 3125 return (buf); 3126 } 3127 3128 static void 3129 arc_hdr_l2hdr_destroy(arc_buf_hdr_t *hdr) 3130 { 3131 l2arc_buf_hdr_t *l2hdr = &hdr->b_l2hdr; 3132 l2arc_dev_t *dev = l2hdr->b_dev; 3133 uint64_t psize = arc_hdr_size(hdr); 3134 3135 ASSERT(MUTEX_HELD(&dev->l2ad_mtx)); 3136 ASSERT(HDR_HAS_L2HDR(hdr)); 3137 3138 list_remove(&dev->l2ad_buflist, hdr); 3139 3140 ARCSTAT_INCR(arcstat_l2_psize, -psize); 3141 ARCSTAT_INCR(arcstat_l2_lsize, -HDR_GET_LSIZE(hdr)); 3142 3143 vdev_space_update(dev->l2ad_vdev, -psize, 0, 0); 3144 3145 (void) zfs_refcount_remove_many(&dev->l2ad_alloc, psize, hdr); 3146 arc_hdr_clear_flags(hdr, ARC_FLAG_HAS_L2HDR); 3147 } 3148 3149 static void 3150 arc_hdr_destroy(arc_buf_hdr_t *hdr) 3151 { 3152 if (HDR_HAS_L1HDR(hdr)) { 3153 ASSERT(hdr->b_l1hdr.b_buf == NULL || 3154 hdr->b_l1hdr.b_bufcnt > 0); 3155 ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 3156 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon); 3157 } 3158 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 3159 ASSERT(!HDR_IN_HASH_TABLE(hdr)); 3160 3161 if (!HDR_EMPTY(hdr)) 3162 buf_discard_identity(hdr); 3163 3164 if (HDR_HAS_L2HDR(hdr)) { 3165 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev; 3166 boolean_t buflist_held = MUTEX_HELD(&dev->l2ad_mtx); 3167 3168 if (!buflist_held) 3169 mutex_enter(&dev->l2ad_mtx); 3170 3171 /* 3172 * Even though we checked this conditional above, we 3173 * need to check this again now that we have the 3174 * l2ad_mtx. This is because we could be racing with 3175 * another thread calling l2arc_evict() which might have 3176 * destroyed this header's L2 portion as we were waiting 3177 * to acquire the l2ad_mtx. If that happens, we don't 3178 * want to re-destroy the header's L2 portion. 3179 */ 3180 if (HDR_HAS_L2HDR(hdr)) 3181 arc_hdr_l2hdr_destroy(hdr); 3182 3183 if (!buflist_held) 3184 mutex_exit(&dev->l2ad_mtx); 3185 } 3186 3187 if (HDR_HAS_L1HDR(hdr)) { 3188 arc_cksum_free(hdr); 3189 3190 while (hdr->b_l1hdr.b_buf != NULL) 3191 arc_buf_destroy_impl(hdr->b_l1hdr.b_buf); 3192 3193 #ifdef ZFS_DEBUG 3194 if (hdr->b_l1hdr.b_thawed != NULL) { 3195 kmem_free(hdr->b_l1hdr.b_thawed, 1); 3196 hdr->b_l1hdr.b_thawed = NULL; 3197 } 3198 #endif 3199 3200 if (hdr->b_l1hdr.b_pabd != NULL) { 3201 arc_hdr_free_pabd(hdr); 3202 } 3203 } 3204 3205 ASSERT3P(hdr->b_hash_next, ==, NULL); 3206 if (HDR_HAS_L1HDR(hdr)) { 3207 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 3208 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL); 3209 kmem_cache_free(hdr_full_cache, hdr); 3210 } else { 3211 kmem_cache_free(hdr_l2only_cache, hdr); 3212 } 3213 } 3214 3215 void 3216 arc_buf_destroy(arc_buf_t *buf, void* tag) 3217 { 3218 arc_buf_hdr_t *hdr = buf->b_hdr; 3219 kmutex_t *hash_lock = HDR_LOCK(hdr); 3220 3221 if (hdr->b_l1hdr.b_state == arc_anon) { 3222 ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1); 3223 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 3224 VERIFY0(remove_reference(hdr, NULL, tag)); 3225 arc_hdr_destroy(hdr); 3226 return; 3227 } 3228 3229 mutex_enter(hash_lock); 3230 ASSERT3P(hdr, ==, buf->b_hdr); 3231 ASSERT(hdr->b_l1hdr.b_bufcnt > 0); 3232 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 3233 ASSERT3P(hdr->b_l1hdr.b_state, !=, arc_anon); 3234 ASSERT3P(buf->b_data, !=, NULL); 3235 3236 (void) remove_reference(hdr, hash_lock, tag); 3237 arc_buf_destroy_impl(buf); 3238 mutex_exit(hash_lock); 3239 } 3240 3241 /* 3242 * Evict the arc_buf_hdr that is provided as a parameter. The resultant 3243 * state of the header is dependent on it's state prior to entering this 3244 * function. The following transitions are possible: 3245 * 3246 * - arc_mru -> arc_mru_ghost 3247 * - arc_mfu -> arc_mfu_ghost 3248 * - arc_mru_ghost -> arc_l2c_only 3249 * - arc_mru_ghost -> deleted 3250 * - arc_mfu_ghost -> arc_l2c_only 3251 * - arc_mfu_ghost -> deleted 3252 */ 3253 static int64_t 3254 arc_evict_hdr(arc_buf_hdr_t *hdr, kmutex_t *hash_lock) 3255 { 3256 arc_state_t *evicted_state, *state; 3257 int64_t bytes_evicted = 0; 3258 int min_lifetime = HDR_PRESCIENT_PREFETCH(hdr) ? 3259 zfs_arc_min_prescient_prefetch_ms : zfs_arc_min_prefetch_ms; 3260 3261 ASSERT(MUTEX_HELD(hash_lock)); 3262 ASSERT(HDR_HAS_L1HDR(hdr)); 3263 3264 state = hdr->b_l1hdr.b_state; 3265 if (GHOST_STATE(state)) { 3266 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 3267 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 3268 3269 /* 3270 * l2arc_write_buffers() relies on a header's L1 portion 3271 * (i.e. its b_pabd field) during it's write phase. 3272 * Thus, we cannot push a header onto the arc_l2c_only 3273 * state (removing it's L1 piece) until the header is 3274 * done being written to the l2arc. 3275 */ 3276 if (HDR_HAS_L2HDR(hdr) && HDR_L2_WRITING(hdr)) { 3277 ARCSTAT_BUMP(arcstat_evict_l2_skip); 3278 return (bytes_evicted); 3279 } 3280 3281 ARCSTAT_BUMP(arcstat_deleted); 3282 bytes_evicted += HDR_GET_LSIZE(hdr); 3283 3284 DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, hdr); 3285 3286 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL); 3287 if (HDR_HAS_L2HDR(hdr)) { 3288 /* 3289 * This buffer is cached on the 2nd Level ARC; 3290 * don't destroy the header. 3291 */ 3292 arc_change_state(arc_l2c_only, hdr, hash_lock); 3293 /* 3294 * dropping from L1+L2 cached to L2-only, 3295 * realloc to remove the L1 header. 3296 */ 3297 hdr = arc_hdr_realloc(hdr, hdr_full_cache, 3298 hdr_l2only_cache); 3299 } else { 3300 arc_change_state(arc_anon, hdr, hash_lock); 3301 arc_hdr_destroy(hdr); 3302 } 3303 return (bytes_evicted); 3304 } 3305 3306 ASSERT(state == arc_mru || state == arc_mfu); 3307 evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost; 3308 3309 /* prefetch buffers have a minimum lifespan */ 3310 if (HDR_IO_IN_PROGRESS(hdr) || 3311 ((hdr->b_flags & (ARC_FLAG_PREFETCH | ARC_FLAG_INDIRECT)) && 3312 ddi_get_lbolt() - hdr->b_l1hdr.b_arc_access < min_lifetime * hz)) { 3313 ARCSTAT_BUMP(arcstat_evict_skip); 3314 return (bytes_evicted); 3315 } 3316 3317 ASSERT0(zfs_refcount_count(&hdr->b_l1hdr.b_refcnt)); 3318 while (hdr->b_l1hdr.b_buf) { 3319 arc_buf_t *buf = hdr->b_l1hdr.b_buf; 3320 if (!mutex_tryenter(&buf->b_evict_lock)) { 3321 ARCSTAT_BUMP(arcstat_mutex_miss); 3322 break; 3323 } 3324 if (buf->b_data != NULL) 3325 bytes_evicted += HDR_GET_LSIZE(hdr); 3326 mutex_exit(&buf->b_evict_lock); 3327 arc_buf_destroy_impl(buf); 3328 } 3329 3330 if (HDR_HAS_L2HDR(hdr)) { 3331 ARCSTAT_INCR(arcstat_evict_l2_cached, HDR_GET_LSIZE(hdr)); 3332 } else { 3333 if (l2arc_write_eligible(hdr->b_spa, hdr)) { 3334 ARCSTAT_INCR(arcstat_evict_l2_eligible, 3335 HDR_GET_LSIZE(hdr)); 3336 } else { 3337 ARCSTAT_INCR(arcstat_evict_l2_ineligible, 3338 HDR_GET_LSIZE(hdr)); 3339 } 3340 } 3341 3342 if (hdr->b_l1hdr.b_bufcnt == 0) { 3343 arc_cksum_free(hdr); 3344 3345 bytes_evicted += arc_hdr_size(hdr); 3346 3347 /* 3348 * If this hdr is being evicted and has a compressed 3349 * buffer then we discard it here before we change states. 3350 * This ensures that the accounting is updated correctly 3351 * in arc_free_data_impl(). 3352 */ 3353 arc_hdr_free_pabd(hdr); 3354 3355 arc_change_state(evicted_state, hdr, hash_lock); 3356 ASSERT(HDR_IN_HASH_TABLE(hdr)); 3357 arc_hdr_set_flags(hdr, ARC_FLAG_IN_HASH_TABLE); 3358 DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, hdr); 3359 } 3360 3361 return (bytes_evicted); 3362 } 3363 3364 static uint64_t 3365 arc_evict_state_impl(multilist_t *ml, int idx, arc_buf_hdr_t *marker, 3366 uint64_t spa, int64_t bytes) 3367 { 3368 multilist_sublist_t *mls; 3369 uint64_t bytes_evicted = 0; 3370 arc_buf_hdr_t *hdr; 3371 kmutex_t *hash_lock; 3372 int evict_count = 0; 3373 3374 ASSERT3P(marker, !=, NULL); 3375 IMPLY(bytes < 0, bytes == ARC_EVICT_ALL); 3376 3377 mls = multilist_sublist_lock(ml, idx); 3378 3379 for (hdr = multilist_sublist_prev(mls, marker); hdr != NULL; 3380 hdr = multilist_sublist_prev(mls, marker)) { 3381 if ((bytes != ARC_EVICT_ALL && bytes_evicted >= bytes) || 3382 (evict_count >= zfs_arc_evict_batch_limit)) 3383 break; 3384 3385 /* 3386 * To keep our iteration location, move the marker 3387 * forward. Since we're not holding hdr's hash lock, we 3388 * must be very careful and not remove 'hdr' from the 3389 * sublist. Otherwise, other consumers might mistake the 3390 * 'hdr' as not being on a sublist when they call the 3391 * multilist_link_active() function (they all rely on 3392 * the hash lock protecting concurrent insertions and 3393 * removals). multilist_sublist_move_forward() was 3394 * specifically implemented to ensure this is the case 3395 * (only 'marker' will be removed and re-inserted). 3396 */ 3397 multilist_sublist_move_forward(mls, marker); 3398 3399 /* 3400 * The only case where the b_spa field should ever be 3401 * zero, is the marker headers inserted by 3402 * arc_evict_state(). It's possible for multiple threads 3403 * to be calling arc_evict_state() concurrently (e.g. 3404 * dsl_pool_close() and zio_inject_fault()), so we must 3405 * skip any markers we see from these other threads. 3406 */ 3407 if (hdr->b_spa == 0) 3408 continue; 3409 3410 /* we're only interested in evicting buffers of a certain spa */ 3411 if (spa != 0 && hdr->b_spa != spa) { 3412 ARCSTAT_BUMP(arcstat_evict_skip); 3413 continue; 3414 } 3415 3416 hash_lock = HDR_LOCK(hdr); 3417 3418 /* 3419 * We aren't calling this function from any code path 3420 * that would already be holding a hash lock, so we're 3421 * asserting on this assumption to be defensive in case 3422 * this ever changes. Without this check, it would be 3423 * possible to incorrectly increment arcstat_mutex_miss 3424 * below (e.g. if the code changed such that we called 3425 * this function with a hash lock held). 3426 */ 3427 ASSERT(!MUTEX_HELD(hash_lock)); 3428 3429 if (mutex_tryenter(hash_lock)) { 3430 uint64_t evicted = arc_evict_hdr(hdr, hash_lock); 3431 mutex_exit(hash_lock); 3432 3433 bytes_evicted += evicted; 3434 3435 /* 3436 * If evicted is zero, arc_evict_hdr() must have 3437 * decided to skip this header, don't increment 3438 * evict_count in this case. 3439 */ 3440 if (evicted != 0) 3441 evict_count++; 3442 3443 /* 3444 * If arc_size isn't overflowing, signal any 3445 * threads that might happen to be waiting. 3446 * 3447 * For each header evicted, we wake up a single 3448 * thread. If we used cv_broadcast, we could 3449 * wake up "too many" threads causing arc_size 3450 * to significantly overflow arc_c; since 3451 * arc_get_data_impl() doesn't check for overflow 3452 * when it's woken up (it doesn't because it's 3453 * possible for the ARC to be overflowing while 3454 * full of un-evictable buffers, and the 3455 * function should proceed in this case). 3456 * 3457 * If threads are left sleeping, due to not 3458 * using cv_broadcast here, they will be woken 3459 * up via cv_broadcast in arc_adjust_cb() just 3460 * before arc_adjust_zthr sleeps. 3461 */ 3462 mutex_enter(&arc_adjust_lock); 3463 if (!arc_is_overflowing()) 3464 cv_signal(&arc_adjust_waiters_cv); 3465 mutex_exit(&arc_adjust_lock); 3466 } else { 3467 ARCSTAT_BUMP(arcstat_mutex_miss); 3468 } 3469 } 3470 3471 multilist_sublist_unlock(mls); 3472 3473 return (bytes_evicted); 3474 } 3475 3476 /* 3477 * Evict buffers from the given arc state, until we've removed the 3478 * specified number of bytes. Move the removed buffers to the 3479 * appropriate evict state. 3480 * 3481 * This function makes a "best effort". It skips over any buffers 3482 * it can't get a hash_lock on, and so, may not catch all candidates. 3483 * It may also return without evicting as much space as requested. 3484 * 3485 * If bytes is specified using the special value ARC_EVICT_ALL, this 3486 * will evict all available (i.e. unlocked and evictable) buffers from 3487 * the given arc state; which is used by arc_flush(). 3488 */ 3489 static uint64_t 3490 arc_evict_state(arc_state_t *state, uint64_t spa, int64_t bytes, 3491 arc_buf_contents_t type) 3492 { 3493 uint64_t total_evicted = 0; 3494 multilist_t *ml = state->arcs_list[type]; 3495 int num_sublists; 3496 arc_buf_hdr_t **markers; 3497 3498 IMPLY(bytes < 0, bytes == ARC_EVICT_ALL); 3499 3500 num_sublists = multilist_get_num_sublists(ml); 3501 3502 /* 3503 * If we've tried to evict from each sublist, made some 3504 * progress, but still have not hit the target number of bytes 3505 * to evict, we want to keep trying. The markers allow us to 3506 * pick up where we left off for each individual sublist, rather 3507 * than starting from the tail each time. 3508 */ 3509 markers = kmem_zalloc(sizeof (*markers) * num_sublists, KM_SLEEP); 3510 for (int i = 0; i < num_sublists; i++) { 3511 markers[i] = kmem_cache_alloc(hdr_full_cache, KM_SLEEP); 3512 3513 /* 3514 * A b_spa of 0 is used to indicate that this header is 3515 * a marker. This fact is used in arc_adjust_type() and 3516 * arc_evict_state_impl(). 3517 */ 3518 markers[i]->b_spa = 0; 3519 3520 multilist_sublist_t *mls = multilist_sublist_lock(ml, i); 3521 multilist_sublist_insert_tail(mls, markers[i]); 3522 multilist_sublist_unlock(mls); 3523 } 3524 3525 /* 3526 * While we haven't hit our target number of bytes to evict, or 3527 * we're evicting all available buffers. 3528 */ 3529 while (total_evicted < bytes || bytes == ARC_EVICT_ALL) { 3530 /* 3531 * Start eviction using a randomly selected sublist, 3532 * this is to try and evenly balance eviction across all 3533 * sublists. Always starting at the same sublist 3534 * (e.g. index 0) would cause evictions to favor certain 3535 * sublists over others. 3536 */ 3537 int sublist_idx = multilist_get_random_index(ml); 3538 uint64_t scan_evicted = 0; 3539 3540 for (int i = 0; i < num_sublists; i++) { 3541 uint64_t bytes_remaining; 3542 uint64_t bytes_evicted; 3543 3544 if (bytes == ARC_EVICT_ALL) 3545 bytes_remaining = ARC_EVICT_ALL; 3546 else if (total_evicted < bytes) 3547 bytes_remaining = bytes - total_evicted; 3548 else 3549 break; 3550 3551 bytes_evicted = arc_evict_state_impl(ml, sublist_idx, 3552 markers[sublist_idx], spa, bytes_remaining); 3553 3554 scan_evicted += bytes_evicted; 3555 total_evicted += bytes_evicted; 3556 3557 /* we've reached the end, wrap to the beginning */ 3558 if (++sublist_idx >= num_sublists) 3559 sublist_idx = 0; 3560 } 3561 3562 /* 3563 * If we didn't evict anything during this scan, we have 3564 * no reason to believe we'll evict more during another 3565 * scan, so break the loop. 3566 */ 3567 if (scan_evicted == 0) { 3568 /* This isn't possible, let's make that obvious */ 3569 ASSERT3S(bytes, !=, 0); 3570 3571 /* 3572 * When bytes is ARC_EVICT_ALL, the only way to 3573 * break the loop is when scan_evicted is zero. 3574 * In that case, we actually have evicted enough, 3575 * so we don't want to increment the kstat. 3576 */ 3577 if (bytes != ARC_EVICT_ALL) { 3578 ASSERT3S(total_evicted, <, bytes); 3579 ARCSTAT_BUMP(arcstat_evict_not_enough); 3580 } 3581 3582 break; 3583 } 3584 } 3585 3586 for (int i = 0; i < num_sublists; i++) { 3587 multilist_sublist_t *mls = multilist_sublist_lock(ml, i); 3588 multilist_sublist_remove(mls, markers[i]); 3589 multilist_sublist_unlock(mls); 3590 3591 kmem_cache_free(hdr_full_cache, markers[i]); 3592 } 3593 kmem_free(markers, sizeof (*markers) * num_sublists); 3594 3595 return (total_evicted); 3596 } 3597 3598 /* 3599 * Flush all "evictable" data of the given type from the arc state 3600 * specified. This will not evict any "active" buffers (i.e. referenced). 3601 * 3602 * When 'retry' is set to B_FALSE, the function will make a single pass 3603 * over the state and evict any buffers that it can. Since it doesn't 3604 * continually retry the eviction, it might end up leaving some buffers 3605 * in the ARC due to lock misses. 3606 * 3607 * When 'retry' is set to B_TRUE, the function will continually retry the 3608 * eviction until *all* evictable buffers have been removed from the 3609 * state. As a result, if concurrent insertions into the state are 3610 * allowed (e.g. if the ARC isn't shutting down), this function might 3611 * wind up in an infinite loop, continually trying to evict buffers. 3612 */ 3613 static uint64_t 3614 arc_flush_state(arc_state_t *state, uint64_t spa, arc_buf_contents_t type, 3615 boolean_t retry) 3616 { 3617 uint64_t evicted = 0; 3618 3619 while (zfs_refcount_count(&state->arcs_esize[type]) != 0) { 3620 evicted += arc_evict_state(state, spa, ARC_EVICT_ALL, type); 3621 3622 if (!retry) 3623 break; 3624 } 3625 3626 return (evicted); 3627 } 3628 3629 /* 3630 * Evict the specified number of bytes from the state specified, 3631 * restricting eviction to the spa and type given. This function 3632 * prevents us from trying to evict more from a state's list than 3633 * is "evictable", and to skip evicting altogether when passed a 3634 * negative value for "bytes". In contrast, arc_evict_state() will 3635 * evict everything it can, when passed a negative value for "bytes". 3636 */ 3637 static uint64_t 3638 arc_adjust_impl(arc_state_t *state, uint64_t spa, int64_t bytes, 3639 arc_buf_contents_t type) 3640 { 3641 int64_t delta; 3642 3643 if (bytes > 0 && zfs_refcount_count(&state->arcs_esize[type]) > 0) { 3644 delta = MIN(zfs_refcount_count(&state->arcs_esize[type]), 3645 bytes); 3646 return (arc_evict_state(state, spa, delta, type)); 3647 } 3648 3649 return (0); 3650 } 3651 3652 /* 3653 * Evict metadata buffers from the cache, such that arc_meta_used is 3654 * capped by the arc_meta_limit tunable. 3655 */ 3656 static uint64_t 3657 arc_adjust_meta(uint64_t meta_used) 3658 { 3659 uint64_t total_evicted = 0; 3660 int64_t target; 3661 3662 /* 3663 * If we're over the meta limit, we want to evict enough 3664 * metadata to get back under the meta limit. We don't want to 3665 * evict so much that we drop the MRU below arc_p, though. If 3666 * we're over the meta limit more than we're over arc_p, we 3667 * evict some from the MRU here, and some from the MFU below. 3668 */ 3669 target = MIN((int64_t)(meta_used - arc_meta_limit), 3670 (int64_t)(zfs_refcount_count(&arc_anon->arcs_size) + 3671 zfs_refcount_count(&arc_mru->arcs_size) - arc_p)); 3672 3673 total_evicted += arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA); 3674 3675 /* 3676 * Similar to the above, we want to evict enough bytes to get us 3677 * below the meta limit, but not so much as to drop us below the 3678 * space allotted to the MFU (which is defined as arc_c - arc_p). 3679 */ 3680 target = MIN((int64_t)(meta_used - arc_meta_limit), 3681 (int64_t)(zfs_refcount_count(&arc_mfu->arcs_size) - 3682 (arc_c - arc_p))); 3683 3684 total_evicted += arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA); 3685 3686 return (total_evicted); 3687 } 3688 3689 /* 3690 * Return the type of the oldest buffer in the given arc state 3691 * 3692 * This function will select a random sublist of type ARC_BUFC_DATA and 3693 * a random sublist of type ARC_BUFC_METADATA. The tail of each sublist 3694 * is compared, and the type which contains the "older" buffer will be 3695 * returned. 3696 */ 3697 static arc_buf_contents_t 3698 arc_adjust_type(arc_state_t *state) 3699 { 3700 multilist_t *data_ml = state->arcs_list[ARC_BUFC_DATA]; 3701 multilist_t *meta_ml = state->arcs_list[ARC_BUFC_METADATA]; 3702 int data_idx = multilist_get_random_index(data_ml); 3703 int meta_idx = multilist_get_random_index(meta_ml); 3704 multilist_sublist_t *data_mls; 3705 multilist_sublist_t *meta_mls; 3706 arc_buf_contents_t type; 3707 arc_buf_hdr_t *data_hdr; 3708 arc_buf_hdr_t *meta_hdr; 3709 3710 /* 3711 * We keep the sublist lock until we're finished, to prevent 3712 * the headers from being destroyed via arc_evict_state(). 3713 */ 3714 data_mls = multilist_sublist_lock(data_ml, data_idx); 3715 meta_mls = multilist_sublist_lock(meta_ml, meta_idx); 3716 3717 /* 3718 * These two loops are to ensure we skip any markers that 3719 * might be at the tail of the lists due to arc_evict_state(). 3720 */ 3721 3722 for (data_hdr = multilist_sublist_tail(data_mls); data_hdr != NULL; 3723 data_hdr = multilist_sublist_prev(data_mls, data_hdr)) { 3724 if (data_hdr->b_spa != 0) 3725 break; 3726 } 3727 3728 for (meta_hdr = multilist_sublist_tail(meta_mls); meta_hdr != NULL; 3729 meta_hdr = multilist_sublist_prev(meta_mls, meta_hdr)) { 3730 if (meta_hdr->b_spa != 0) 3731 break; 3732 } 3733 3734 if (data_hdr == NULL && meta_hdr == NULL) { 3735 type = ARC_BUFC_DATA; 3736 } else if (data_hdr == NULL) { 3737 ASSERT3P(meta_hdr, !=, NULL); 3738 type = ARC_BUFC_METADATA; 3739 } else if (meta_hdr == NULL) { 3740 ASSERT3P(data_hdr, !=, NULL); 3741 type = ARC_BUFC_DATA; 3742 } else { 3743 ASSERT3P(data_hdr, !=, NULL); 3744 ASSERT3P(meta_hdr, !=, NULL); 3745 3746 /* The headers can't be on the sublist without an L1 header */ 3747 ASSERT(HDR_HAS_L1HDR(data_hdr)); 3748 ASSERT(HDR_HAS_L1HDR(meta_hdr)); 3749 3750 if (data_hdr->b_l1hdr.b_arc_access < 3751 meta_hdr->b_l1hdr.b_arc_access) { 3752 type = ARC_BUFC_DATA; 3753 } else { 3754 type = ARC_BUFC_METADATA; 3755 } 3756 } 3757 3758 multilist_sublist_unlock(meta_mls); 3759 multilist_sublist_unlock(data_mls); 3760 3761 return (type); 3762 } 3763 3764 /* 3765 * Evict buffers from the cache, such that arc_size is capped by arc_c. 3766 */ 3767 static uint64_t 3768 arc_adjust(void) 3769 { 3770 uint64_t total_evicted = 0; 3771 uint64_t bytes; 3772 int64_t target; 3773 uint64_t asize = aggsum_value(&arc_size); 3774 uint64_t ameta = aggsum_value(&arc_meta_used); 3775 3776 /* 3777 * If we're over arc_meta_limit, we want to correct that before 3778 * potentially evicting data buffers below. 3779 */ 3780 total_evicted += arc_adjust_meta(ameta); 3781 3782 /* 3783 * Adjust MRU size 3784 * 3785 * If we're over the target cache size, we want to evict enough 3786 * from the list to get back to our target size. We don't want 3787 * to evict too much from the MRU, such that it drops below 3788 * arc_p. So, if we're over our target cache size more than 3789 * the MRU is over arc_p, we'll evict enough to get back to 3790 * arc_p here, and then evict more from the MFU below. 3791 */ 3792 target = MIN((int64_t)(asize - arc_c), 3793 (int64_t)(zfs_refcount_count(&arc_anon->arcs_size) + 3794 zfs_refcount_count(&arc_mru->arcs_size) + ameta - arc_p)); 3795 3796 /* 3797 * If we're below arc_meta_min, always prefer to evict data. 3798 * Otherwise, try to satisfy the requested number of bytes to 3799 * evict from the type which contains older buffers; in an 3800 * effort to keep newer buffers in the cache regardless of their 3801 * type. If we cannot satisfy the number of bytes from this 3802 * type, spill over into the next type. 3803 */ 3804 if (arc_adjust_type(arc_mru) == ARC_BUFC_METADATA && 3805 ameta > arc_meta_min) { 3806 bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA); 3807 total_evicted += bytes; 3808 3809 /* 3810 * If we couldn't evict our target number of bytes from 3811 * metadata, we try to get the rest from data. 3812 */ 3813 target -= bytes; 3814 3815 total_evicted += 3816 arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA); 3817 } else { 3818 bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA); 3819 total_evicted += bytes; 3820 3821 /* 3822 * If we couldn't evict our target number of bytes from 3823 * data, we try to get the rest from metadata. 3824 */ 3825 target -= bytes; 3826 3827 total_evicted += 3828 arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA); 3829 } 3830 3831 /* 3832 * Adjust MFU size 3833 * 3834 * Now that we've tried to evict enough from the MRU to get its 3835 * size back to arc_p, if we're still above the target cache 3836 * size, we evict the rest from the MFU. 3837 */ 3838 target = asize - arc_c; 3839 3840 if (arc_adjust_type(arc_mfu) == ARC_BUFC_METADATA && 3841 ameta > arc_meta_min) { 3842 bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA); 3843 total_evicted += bytes; 3844 3845 /* 3846 * If we couldn't evict our target number of bytes from 3847 * metadata, we try to get the rest from data. 3848 */ 3849 target -= bytes; 3850 3851 total_evicted += 3852 arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA); 3853 } else { 3854 bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA); 3855 total_evicted += bytes; 3856 3857 /* 3858 * If we couldn't evict our target number of bytes from 3859 * data, we try to get the rest from data. 3860 */ 3861 target -= bytes; 3862 3863 total_evicted += 3864 arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA); 3865 } 3866 3867 /* 3868 * Adjust ghost lists 3869 * 3870 * In addition to the above, the ARC also defines target values 3871 * for the ghost lists. The sum of the mru list and mru ghost 3872 * list should never exceed the target size of the cache, and 3873 * the sum of the mru list, mfu list, mru ghost list, and mfu 3874 * ghost list should never exceed twice the target size of the 3875 * cache. The following logic enforces these limits on the ghost 3876 * caches, and evicts from them as needed. 3877 */ 3878 target = zfs_refcount_count(&arc_mru->arcs_size) + 3879 zfs_refcount_count(&arc_mru_ghost->arcs_size) - arc_c; 3880 3881 bytes = arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_DATA); 3882 total_evicted += bytes; 3883 3884 target -= bytes; 3885 3886 total_evicted += 3887 arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_METADATA); 3888 3889 /* 3890 * We assume the sum of the mru list and mfu list is less than 3891 * or equal to arc_c (we enforced this above), which means we 3892 * can use the simpler of the two equations below: 3893 * 3894 * mru + mfu + mru ghost + mfu ghost <= 2 * arc_c 3895 * mru ghost + mfu ghost <= arc_c 3896 */ 3897 target = zfs_refcount_count(&arc_mru_ghost->arcs_size) + 3898 zfs_refcount_count(&arc_mfu_ghost->arcs_size) - arc_c; 3899 3900 bytes = arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_DATA); 3901 total_evicted += bytes; 3902 3903 target -= bytes; 3904 3905 total_evicted += 3906 arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_METADATA); 3907 3908 return (total_evicted); 3909 } 3910 3911 void 3912 arc_flush(spa_t *spa, boolean_t retry) 3913 { 3914 uint64_t guid = 0; 3915 3916 /* 3917 * If retry is B_TRUE, a spa must not be specified since we have 3918 * no good way to determine if all of a spa's buffers have been 3919 * evicted from an arc state. 3920 */ 3921 ASSERT(!retry || spa == 0); 3922 3923 if (spa != NULL) 3924 guid = spa_load_guid(spa); 3925 3926 (void) arc_flush_state(arc_mru, guid, ARC_BUFC_DATA, retry); 3927 (void) arc_flush_state(arc_mru, guid, ARC_BUFC_METADATA, retry); 3928 3929 (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_DATA, retry); 3930 (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_METADATA, retry); 3931 3932 (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_DATA, retry); 3933 (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_METADATA, retry); 3934 3935 (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_DATA, retry); 3936 (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_METADATA, retry); 3937 } 3938 3939 static void 3940 arc_reduce_target_size(int64_t to_free) 3941 { 3942 uint64_t asize = aggsum_value(&arc_size); 3943 if (arc_c > arc_c_min) { 3944 3945 if (arc_c > arc_c_min + to_free) 3946 atomic_add_64(&arc_c, -to_free); 3947 else 3948 arc_c = arc_c_min; 3949 3950 atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift)); 3951 if (asize < arc_c) 3952 arc_c = MAX(asize, arc_c_min); 3953 if (arc_p > arc_c) 3954 arc_p = (arc_c >> 1); 3955 ASSERT(arc_c >= arc_c_min); 3956 ASSERT((int64_t)arc_p >= 0); 3957 } 3958 3959 if (asize > arc_c) { 3960 /* See comment in arc_adjust_cb_check() on why lock+flag */ 3961 mutex_enter(&arc_adjust_lock); 3962 arc_adjust_needed = B_TRUE; 3963 mutex_exit(&arc_adjust_lock); 3964 zthr_wakeup(arc_adjust_zthr); 3965 } 3966 } 3967 3968 typedef enum free_memory_reason_t { 3969 FMR_UNKNOWN, 3970 FMR_NEEDFREE, 3971 FMR_LOTSFREE, 3972 FMR_SWAPFS_MINFREE, 3973 FMR_PAGES_PP_MAXIMUM, 3974 FMR_HEAP_ARENA, 3975 FMR_ZIO_ARENA, 3976 } free_memory_reason_t; 3977 3978 int64_t last_free_memory; 3979 free_memory_reason_t last_free_reason; 3980 3981 /* 3982 * Additional reserve of pages for pp_reserve. 3983 */ 3984 int64_t arc_pages_pp_reserve = 64; 3985 3986 /* 3987 * Additional reserve of pages for swapfs. 3988 */ 3989 int64_t arc_swapfs_reserve = 64; 3990 3991 /* 3992 * Return the amount of memory that can be consumed before reclaim will be 3993 * needed. Positive if there is sufficient free memory, negative indicates 3994 * the amount of memory that needs to be freed up. 3995 */ 3996 static int64_t 3997 arc_available_memory(void) 3998 { 3999 int64_t lowest = INT64_MAX; 4000 int64_t n; 4001 free_memory_reason_t r = FMR_UNKNOWN; 4002 4003 #ifdef _KERNEL 4004 if (needfree > 0) { 4005 n = PAGESIZE * (-needfree); 4006 if (n < lowest) { 4007 lowest = n; 4008 r = FMR_NEEDFREE; 4009 } 4010 } 4011 4012 /* 4013 * check that we're out of range of the pageout scanner. It starts to 4014 * schedule paging if freemem is less than lotsfree and needfree. 4015 * lotsfree is the high-water mark for pageout, and needfree is the 4016 * number of needed free pages. We add extra pages here to make sure 4017 * the scanner doesn't start up while we're freeing memory. 4018 */ 4019 n = PAGESIZE * (freemem - lotsfree - needfree - desfree); 4020 if (n < lowest) { 4021 lowest = n; 4022 r = FMR_LOTSFREE; 4023 } 4024 4025 /* 4026 * check to make sure that swapfs has enough space so that anon 4027 * reservations can still succeed. anon_resvmem() checks that the 4028 * availrmem is greater than swapfs_minfree, and the number of reserved 4029 * swap pages. We also add a bit of extra here just to prevent 4030 * circumstances from getting really dire. 4031 */ 4032 n = PAGESIZE * (availrmem - swapfs_minfree - swapfs_reserve - 4033 desfree - arc_swapfs_reserve); 4034 if (n < lowest) { 4035 lowest = n; 4036 r = FMR_SWAPFS_MINFREE; 4037 } 4038 4039 4040 /* 4041 * Check that we have enough availrmem that memory locking (e.g., via 4042 * mlock(3C) or memcntl(2)) can still succeed. (pages_pp_maximum 4043 * stores the number of pages that cannot be locked; when availrmem 4044 * drops below pages_pp_maximum, page locking mechanisms such as 4045 * page_pp_lock() will fail.) 4046 */ 4047 n = PAGESIZE * (availrmem - pages_pp_maximum - 4048 arc_pages_pp_reserve); 4049 if (n < lowest) { 4050 lowest = n; 4051 r = FMR_PAGES_PP_MAXIMUM; 4052 } 4053 4054 #if defined(__i386) 4055 /* 4056 * If we're on an i386 platform, it's possible that we'll exhaust the 4057 * kernel heap space before we ever run out of available physical 4058 * memory. Most checks of the size of the heap_area compare against 4059 * tune.t_minarmem, which is the minimum available real memory that we 4060 * can have in the system. However, this is generally fixed at 25 pages 4061 * which is so low that it's useless. In this comparison, we seek to 4062 * calculate the total heap-size, and reclaim if more than 3/4ths of the 4063 * heap is allocated. (Or, in the calculation, if less than 1/4th is 4064 * free) 4065 */ 4066 n = (int64_t)vmem_size(heap_arena, VMEM_FREE) - 4067 (vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC) >> 2); 4068 if (n < lowest) { 4069 lowest = n; 4070 r = FMR_HEAP_ARENA; 4071 } 4072 #endif 4073 4074 /* 4075 * If zio data pages are being allocated out of a separate heap segment, 4076 * then enforce that the size of available vmem for this arena remains 4077 * above about 1/4th (1/(2^arc_zio_arena_free_shift)) free. 4078 * 4079 * Note that reducing the arc_zio_arena_free_shift keeps more virtual 4080 * memory (in the zio_arena) free, which can avoid memory 4081 * fragmentation issues. 4082 */ 4083 if (zio_arena != NULL) { 4084 n = (int64_t)vmem_size(zio_arena, VMEM_FREE) - 4085 (vmem_size(zio_arena, VMEM_ALLOC) >> 4086 arc_zio_arena_free_shift); 4087 if (n < lowest) { 4088 lowest = n; 4089 r = FMR_ZIO_ARENA; 4090 } 4091 } 4092 #else 4093 /* Every 100 calls, free a small amount */ 4094 if (spa_get_random(100) == 0) 4095 lowest = -1024; 4096 #endif 4097 4098 last_free_memory = lowest; 4099 last_free_reason = r; 4100 4101 return (lowest); 4102 } 4103 4104 4105 /* 4106 * Determine if the system is under memory pressure and is asking 4107 * to reclaim memory. A return value of B_TRUE indicates that the system 4108 * is under memory pressure and that the arc should adjust accordingly. 4109 */ 4110 static boolean_t 4111 arc_reclaim_needed(void) 4112 { 4113 return (arc_available_memory() < 0); 4114 } 4115 4116 static void 4117 arc_kmem_reap_soon(void) 4118 { 4119 size_t i; 4120 kmem_cache_t *prev_cache = NULL; 4121 kmem_cache_t *prev_data_cache = NULL; 4122 extern kmem_cache_t *zio_buf_cache[]; 4123 extern kmem_cache_t *zio_data_buf_cache[]; 4124 extern kmem_cache_t *range_seg_cache; 4125 extern kmem_cache_t *abd_chunk_cache; 4126 4127 #ifdef _KERNEL 4128 if (aggsum_compare(&arc_meta_used, arc_meta_limit) >= 0) { 4129 /* 4130 * We are exceeding our meta-data cache limit. 4131 * Purge some DNLC entries to release holds on meta-data. 4132 */ 4133 dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent); 4134 } 4135 #if defined(__i386) 4136 /* 4137 * Reclaim unused memory from all kmem caches. 4138 */ 4139 kmem_reap(); 4140 #endif 4141 #endif 4142 4143 for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) { 4144 if (zio_buf_cache[i] != prev_cache) { 4145 prev_cache = zio_buf_cache[i]; 4146 kmem_cache_reap_soon(zio_buf_cache[i]); 4147 } 4148 if (zio_data_buf_cache[i] != prev_data_cache) { 4149 prev_data_cache = zio_data_buf_cache[i]; 4150 kmem_cache_reap_soon(zio_data_buf_cache[i]); 4151 } 4152 } 4153 kmem_cache_reap_soon(abd_chunk_cache); 4154 kmem_cache_reap_soon(buf_cache); 4155 kmem_cache_reap_soon(hdr_full_cache); 4156 kmem_cache_reap_soon(hdr_l2only_cache); 4157 kmem_cache_reap_soon(range_seg_cache); 4158 4159 if (zio_arena != NULL) { 4160 /* 4161 * Ask the vmem arena to reclaim unused memory from its 4162 * quantum caches. 4163 */ 4164 vmem_qcache_reap(zio_arena); 4165 } 4166 } 4167 4168 /* ARGSUSED */ 4169 static boolean_t 4170 arc_adjust_cb_check(void *arg, zthr_t *zthr) 4171 { 4172 /* 4173 * This is necessary in order for the mdb ::arc dcmd to 4174 * show up to date information. Since the ::arc command 4175 * does not call the kstat's update function, without 4176 * this call, the command may show stale stats for the 4177 * anon, mru, mru_ghost, mfu, and mfu_ghost lists. Even 4178 * with this change, the data might be up to 1 second 4179 * out of date(the arc_adjust_zthr has a maximum sleep 4180 * time of 1 second); but that should suffice. The 4181 * arc_state_t structures can be queried directly if more 4182 * accurate information is needed. 4183 */ 4184 if (arc_ksp != NULL) 4185 arc_ksp->ks_update(arc_ksp, KSTAT_READ); 4186 4187 /* 4188 * We have to rely on arc_get_data_impl() to tell us when to adjust, 4189 * rather than checking if we are overflowing here, so that we are 4190 * sure to not leave arc_get_data_impl() waiting on 4191 * arc_adjust_waiters_cv. If we have become "not overflowing" since 4192 * arc_get_data_impl() checked, we need to wake it up. We could 4193 * broadcast the CV here, but arc_get_data_impl() may have not yet 4194 * gone to sleep. We would need to use a mutex to ensure that this 4195 * function doesn't broadcast until arc_get_data_impl() has gone to 4196 * sleep (e.g. the arc_adjust_lock). However, the lock ordering of 4197 * such a lock would necessarily be incorrect with respect to the 4198 * zthr_lock, which is held before this function is called, and is 4199 * held by arc_get_data_impl() when it calls zthr_wakeup(). 4200 */ 4201 return (arc_adjust_needed); 4202 } 4203 4204 /* 4205 * Keep arc_size under arc_c by running arc_adjust which evicts data 4206 * from the ARC. 4207 */ 4208 /* ARGSUSED */ 4209 static void 4210 arc_adjust_cb(void *arg, zthr_t *zthr) 4211 { 4212 uint64_t evicted = 0; 4213 4214 /* Evict from cache */ 4215 evicted = arc_adjust(); 4216 4217 /* 4218 * If evicted is zero, we couldn't evict anything 4219 * via arc_adjust(). This could be due to hash lock 4220 * collisions, but more likely due to the majority of 4221 * arc buffers being unevictable. Therefore, even if 4222 * arc_size is above arc_c, another pass is unlikely to 4223 * be helpful and could potentially cause us to enter an 4224 * infinite loop. Additionally, zthr_iscancelled() is 4225 * checked here so that if the arc is shutting down, the 4226 * broadcast will wake any remaining arc adjust waiters. 4227 */ 4228 mutex_enter(&arc_adjust_lock); 4229 arc_adjust_needed = !zthr_iscancelled(arc_adjust_zthr) && 4230 evicted > 0 && aggsum_compare(&arc_size, arc_c) > 0; 4231 if (!arc_adjust_needed) { 4232 /* 4233 * We're either no longer overflowing, or we 4234 * can't evict anything more, so we should wake 4235 * up any waiters. 4236 */ 4237 cv_broadcast(&arc_adjust_waiters_cv); 4238 } 4239 mutex_exit(&arc_adjust_lock); 4240 } 4241 4242 /* ARGSUSED */ 4243 static boolean_t 4244 arc_reap_cb_check(void *arg, zthr_t *zthr) 4245 { 4246 int64_t free_memory = arc_available_memory(); 4247 4248 /* 4249 * If a kmem reap is already active, don't schedule more. We must 4250 * check for this because kmem_cache_reap_soon() won't actually 4251 * block on the cache being reaped (this is to prevent callers from 4252 * becoming implicitly blocked by a system-wide kmem reap -- which, 4253 * on a system with many, many full magazines, can take minutes). 4254 */ 4255 if (!kmem_cache_reap_active() && 4256 free_memory < 0) { 4257 arc_no_grow = B_TRUE; 4258 arc_warm = B_TRUE; 4259 /* 4260 * Wait at least zfs_grow_retry (default 60) seconds 4261 * before considering growing. 4262 */ 4263 arc_growtime = gethrtime() + SEC2NSEC(arc_grow_retry); 4264 return (B_TRUE); 4265 } else if (free_memory < arc_c >> arc_no_grow_shift) { 4266 arc_no_grow = B_TRUE; 4267 } else if (gethrtime() >= arc_growtime) { 4268 arc_no_grow = B_FALSE; 4269 } 4270 4271 return (B_FALSE); 4272 } 4273 4274 /* 4275 * Keep enough free memory in the system by reaping the ARC's kmem 4276 * caches. To cause more slabs to be reapable, we may reduce the 4277 * target size of the cache (arc_c), causing the arc_adjust_cb() 4278 * to free more buffers. 4279 */ 4280 /* ARGSUSED */ 4281 static void 4282 arc_reap_cb(void *arg, zthr_t *zthr) 4283 { 4284 int64_t free_memory; 4285 4286 /* 4287 * Kick off asynchronous kmem_reap()'s of all our caches. 4288 */ 4289 arc_kmem_reap_soon(); 4290 4291 /* 4292 * Wait at least arc_kmem_cache_reap_retry_ms between 4293 * arc_kmem_reap_soon() calls. Without this check it is possible to 4294 * end up in a situation where we spend lots of time reaping 4295 * caches, while we're near arc_c_min. Waiting here also gives the 4296 * subsequent free memory check a chance of finding that the 4297 * asynchronous reap has already freed enough memory, and we don't 4298 * need to call arc_reduce_target_size(). 4299 */ 4300 delay((hz * arc_kmem_cache_reap_retry_ms + 999) / 1000); 4301 4302 /* 4303 * Reduce the target size as needed to maintain the amount of free 4304 * memory in the system at a fraction of the arc_size (1/128th by 4305 * default). If oversubscribed (free_memory < 0) then reduce the 4306 * target arc_size by the deficit amount plus the fractional 4307 * amount. If free memory is positive but less then the fractional 4308 * amount, reduce by what is needed to hit the fractional amount. 4309 */ 4310 free_memory = arc_available_memory(); 4311 4312 int64_t to_free = 4313 (arc_c >> arc_shrink_shift) - free_memory; 4314 if (to_free > 0) { 4315 #ifdef _KERNEL 4316 to_free = MAX(to_free, ptob(needfree)); 4317 #endif 4318 arc_reduce_target_size(to_free); 4319 } 4320 } 4321 4322 /* 4323 * Adapt arc info given the number of bytes we are trying to add and 4324 * the state that we are comming from. This function is only called 4325 * when we are adding new content to the cache. 4326 */ 4327 static void 4328 arc_adapt(int bytes, arc_state_t *state) 4329 { 4330 int mult; 4331 uint64_t arc_p_min = (arc_c >> arc_p_min_shift); 4332 int64_t mrug_size = zfs_refcount_count(&arc_mru_ghost->arcs_size); 4333 int64_t mfug_size = zfs_refcount_count(&arc_mfu_ghost->arcs_size); 4334 4335 if (state == arc_l2c_only) 4336 return; 4337 4338 ASSERT(bytes > 0); 4339 /* 4340 * Adapt the target size of the MRU list: 4341 * - if we just hit in the MRU ghost list, then increase 4342 * the target size of the MRU list. 4343 * - if we just hit in the MFU ghost list, then increase 4344 * the target size of the MFU list by decreasing the 4345 * target size of the MRU list. 4346 */ 4347 if (state == arc_mru_ghost) { 4348 mult = (mrug_size >= mfug_size) ? 1 : (mfug_size / mrug_size); 4349 mult = MIN(mult, 10); /* avoid wild arc_p adjustment */ 4350 4351 arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult); 4352 } else if (state == arc_mfu_ghost) { 4353 uint64_t delta; 4354 4355 mult = (mfug_size >= mrug_size) ? 1 : (mrug_size / mfug_size); 4356 mult = MIN(mult, 10); 4357 4358 delta = MIN(bytes * mult, arc_p); 4359 arc_p = MAX(arc_p_min, arc_p - delta); 4360 } 4361 ASSERT((int64_t)arc_p >= 0); 4362 4363 /* 4364 * Wake reap thread if we do not have any available memory 4365 */ 4366 if (arc_reclaim_needed()) { 4367 zthr_wakeup(arc_reap_zthr); 4368 return; 4369 } 4370 4371 4372 if (arc_no_grow) 4373 return; 4374 4375 if (arc_c >= arc_c_max) 4376 return; 4377 4378 /* 4379 * If we're within (2 * maxblocksize) bytes of the target 4380 * cache size, increment the target cache size 4381 */ 4382 if (aggsum_compare(&arc_size, arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) > 4383 0) { 4384 atomic_add_64(&arc_c, (int64_t)bytes); 4385 if (arc_c > arc_c_max) 4386 arc_c = arc_c_max; 4387 else if (state == arc_anon) 4388 atomic_add_64(&arc_p, (int64_t)bytes); 4389 if (arc_p > arc_c) 4390 arc_p = arc_c; 4391 } 4392 ASSERT((int64_t)arc_p >= 0); 4393 } 4394 4395 /* 4396 * Check if arc_size has grown past our upper threshold, determined by 4397 * zfs_arc_overflow_shift. 4398 */ 4399 static boolean_t 4400 arc_is_overflowing(void) 4401 { 4402 /* Always allow at least one block of overflow */ 4403 uint64_t overflow = MAX(SPA_MAXBLOCKSIZE, 4404 arc_c >> zfs_arc_overflow_shift); 4405 4406 /* 4407 * We just compare the lower bound here for performance reasons. Our 4408 * primary goals are to make sure that the arc never grows without 4409 * bound, and that it can reach its maximum size. This check 4410 * accomplishes both goals. The maximum amount we could run over by is 4411 * 2 * aggsum_borrow_multiplier * NUM_CPUS * the average size of a block 4412 * in the ARC. In practice, that's in the tens of MB, which is low 4413 * enough to be safe. 4414 */ 4415 return (aggsum_lower_bound(&arc_size) >= arc_c + overflow); 4416 } 4417 4418 static abd_t * 4419 arc_get_data_abd(arc_buf_hdr_t *hdr, uint64_t size, void *tag) 4420 { 4421 arc_buf_contents_t type = arc_buf_type(hdr); 4422 4423 arc_get_data_impl(hdr, size, tag); 4424 if (type == ARC_BUFC_METADATA) { 4425 return (abd_alloc(size, B_TRUE)); 4426 } else { 4427 ASSERT(type == ARC_BUFC_DATA); 4428 return (abd_alloc(size, B_FALSE)); 4429 } 4430 } 4431 4432 static void * 4433 arc_get_data_buf(arc_buf_hdr_t *hdr, uint64_t size, void *tag) 4434 { 4435 arc_buf_contents_t type = arc_buf_type(hdr); 4436 4437 arc_get_data_impl(hdr, size, tag); 4438 if (type == ARC_BUFC_METADATA) { 4439 return (zio_buf_alloc(size)); 4440 } else { 4441 ASSERT(type == ARC_BUFC_DATA); 4442 return (zio_data_buf_alloc(size)); 4443 } 4444 } 4445 4446 /* 4447 * Allocate a block and return it to the caller. If we are hitting the 4448 * hard limit for the cache size, we must sleep, waiting for the eviction 4449 * thread to catch up. If we're past the target size but below the hard 4450 * limit, we'll only signal the reclaim thread and continue on. 4451 */ 4452 static void 4453 arc_get_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag) 4454 { 4455 arc_state_t *state = hdr->b_l1hdr.b_state; 4456 arc_buf_contents_t type = arc_buf_type(hdr); 4457 4458 arc_adapt(size, state); 4459 4460 /* 4461 * If arc_size is currently overflowing, and has grown past our 4462 * upper limit, we must be adding data faster than the evict 4463 * thread can evict. Thus, to ensure we don't compound the 4464 * problem by adding more data and forcing arc_size to grow even 4465 * further past it's target size, we halt and wait for the 4466 * eviction thread to catch up. 4467 * 4468 * It's also possible that the reclaim thread is unable to evict 4469 * enough buffers to get arc_size below the overflow limit (e.g. 4470 * due to buffers being un-evictable, or hash lock collisions). 4471 * In this case, we want to proceed regardless if we're 4472 * overflowing; thus we don't use a while loop here. 4473 */ 4474 if (arc_is_overflowing()) { 4475 mutex_enter(&arc_adjust_lock); 4476 4477 /* 4478 * Now that we've acquired the lock, we may no longer be 4479 * over the overflow limit, lets check. 4480 * 4481 * We're ignoring the case of spurious wake ups. If that 4482 * were to happen, it'd let this thread consume an ARC 4483 * buffer before it should have (i.e. before we're under 4484 * the overflow limit and were signalled by the reclaim 4485 * thread). As long as that is a rare occurrence, it 4486 * shouldn't cause any harm. 4487 */ 4488 if (arc_is_overflowing()) { 4489 arc_adjust_needed = B_TRUE; 4490 zthr_wakeup(arc_adjust_zthr); 4491 (void) cv_wait(&arc_adjust_waiters_cv, 4492 &arc_adjust_lock); 4493 } 4494 mutex_exit(&arc_adjust_lock); 4495 } 4496 4497 VERIFY3U(hdr->b_type, ==, type); 4498 if (type == ARC_BUFC_METADATA) { 4499 arc_space_consume(size, ARC_SPACE_META); 4500 } else { 4501 arc_space_consume(size, ARC_SPACE_DATA); 4502 } 4503 4504 /* 4505 * Update the state size. Note that ghost states have a 4506 * "ghost size" and so don't need to be updated. 4507 */ 4508 if (!GHOST_STATE(state)) { 4509 4510 (void) zfs_refcount_add_many(&state->arcs_size, size, tag); 4511 4512 /* 4513 * If this is reached via arc_read, the link is 4514 * protected by the hash lock. If reached via 4515 * arc_buf_alloc, the header should not be accessed by 4516 * any other thread. And, if reached via arc_read_done, 4517 * the hash lock will protect it if it's found in the 4518 * hash table; otherwise no other thread should be 4519 * trying to [add|remove]_reference it. 4520 */ 4521 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) { 4522 ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 4523 (void) zfs_refcount_add_many(&state->arcs_esize[type], 4524 size, tag); 4525 } 4526 4527 /* 4528 * If we are growing the cache, and we are adding anonymous 4529 * data, and we have outgrown arc_p, update arc_p 4530 */ 4531 if (aggsum_compare(&arc_size, arc_c) < 0 && 4532 hdr->b_l1hdr.b_state == arc_anon && 4533 (zfs_refcount_count(&arc_anon->arcs_size) + 4534 zfs_refcount_count(&arc_mru->arcs_size) > arc_p)) 4535 arc_p = MIN(arc_c, arc_p + size); 4536 } 4537 } 4538 4539 static void 4540 arc_free_data_abd(arc_buf_hdr_t *hdr, abd_t *abd, uint64_t size, void *tag) 4541 { 4542 arc_free_data_impl(hdr, size, tag); 4543 abd_free(abd); 4544 } 4545 4546 static void 4547 arc_free_data_buf(arc_buf_hdr_t *hdr, void *buf, uint64_t size, void *tag) 4548 { 4549 arc_buf_contents_t type = arc_buf_type(hdr); 4550 4551 arc_free_data_impl(hdr, size, tag); 4552 if (type == ARC_BUFC_METADATA) { 4553 zio_buf_free(buf, size); 4554 } else { 4555 ASSERT(type == ARC_BUFC_DATA); 4556 zio_data_buf_free(buf, size); 4557 } 4558 } 4559 4560 /* 4561 * Free the arc data buffer. 4562 */ 4563 static void 4564 arc_free_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag) 4565 { 4566 arc_state_t *state = hdr->b_l1hdr.b_state; 4567 arc_buf_contents_t type = arc_buf_type(hdr); 4568 4569 /* protected by hash lock, if in the hash table */ 4570 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) { 4571 ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 4572 ASSERT(state != arc_anon && state != arc_l2c_only); 4573 4574 (void) zfs_refcount_remove_many(&state->arcs_esize[type], 4575 size, tag); 4576 } 4577 (void) zfs_refcount_remove_many(&state->arcs_size, size, tag); 4578 4579 VERIFY3U(hdr->b_type, ==, type); 4580 if (type == ARC_BUFC_METADATA) { 4581 arc_space_return(size, ARC_SPACE_META); 4582 } else { 4583 ASSERT(type == ARC_BUFC_DATA); 4584 arc_space_return(size, ARC_SPACE_DATA); 4585 } 4586 } 4587 4588 /* 4589 * This routine is called whenever a buffer is accessed. 4590 * NOTE: the hash lock is dropped in this function. 4591 */ 4592 static void 4593 arc_access(arc_buf_hdr_t *hdr, kmutex_t *hash_lock) 4594 { 4595 clock_t now; 4596 4597 ASSERT(MUTEX_HELD(hash_lock)); 4598 ASSERT(HDR_HAS_L1HDR(hdr)); 4599 4600 if (hdr->b_l1hdr.b_state == arc_anon) { 4601 /* 4602 * This buffer is not in the cache, and does not 4603 * appear in our "ghost" list. Add the new buffer 4604 * to the MRU state. 4605 */ 4606 4607 ASSERT0(hdr->b_l1hdr.b_arc_access); 4608 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 4609 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr); 4610 arc_change_state(arc_mru, hdr, hash_lock); 4611 4612 } else if (hdr->b_l1hdr.b_state == arc_mru) { 4613 now = ddi_get_lbolt(); 4614 4615 /* 4616 * If this buffer is here because of a prefetch, then either: 4617 * - clear the flag if this is a "referencing" read 4618 * (any subsequent access will bump this into the MFU state). 4619 * or 4620 * - move the buffer to the head of the list if this is 4621 * another prefetch (to make it less likely to be evicted). 4622 */ 4623 if (HDR_PREFETCH(hdr) || HDR_PRESCIENT_PREFETCH(hdr)) { 4624 if (zfs_refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) { 4625 /* link protected by hash lock */ 4626 ASSERT(multilist_link_active( 4627 &hdr->b_l1hdr.b_arc_node)); 4628 } else { 4629 arc_hdr_clear_flags(hdr, 4630 ARC_FLAG_PREFETCH | 4631 ARC_FLAG_PRESCIENT_PREFETCH); 4632 ARCSTAT_BUMP(arcstat_mru_hits); 4633 } 4634 hdr->b_l1hdr.b_arc_access = now; 4635 return; 4636 } 4637 4638 /* 4639 * This buffer has been "accessed" only once so far, 4640 * but it is still in the cache. Move it to the MFU 4641 * state. 4642 */ 4643 if (now > hdr->b_l1hdr.b_arc_access + ARC_MINTIME) { 4644 /* 4645 * More than 125ms have passed since we 4646 * instantiated this buffer. Move it to the 4647 * most frequently used state. 4648 */ 4649 hdr->b_l1hdr.b_arc_access = now; 4650 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 4651 arc_change_state(arc_mfu, hdr, hash_lock); 4652 } 4653 ARCSTAT_BUMP(arcstat_mru_hits); 4654 } else if (hdr->b_l1hdr.b_state == arc_mru_ghost) { 4655 arc_state_t *new_state; 4656 /* 4657 * This buffer has been "accessed" recently, but 4658 * was evicted from the cache. Move it to the 4659 * MFU state. 4660 */ 4661 4662 if (HDR_PREFETCH(hdr) || HDR_PRESCIENT_PREFETCH(hdr)) { 4663 new_state = arc_mru; 4664 if (zfs_refcount_count(&hdr->b_l1hdr.b_refcnt) > 0) { 4665 arc_hdr_clear_flags(hdr, 4666 ARC_FLAG_PREFETCH | 4667 ARC_FLAG_PRESCIENT_PREFETCH); 4668 } 4669 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr); 4670 } else { 4671 new_state = arc_mfu; 4672 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 4673 } 4674 4675 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 4676 arc_change_state(new_state, hdr, hash_lock); 4677 4678 ARCSTAT_BUMP(arcstat_mru_ghost_hits); 4679 } else if (hdr->b_l1hdr.b_state == arc_mfu) { 4680 /* 4681 * This buffer has been accessed more than once and is 4682 * still in the cache. Keep it in the MFU state. 4683 * 4684 * NOTE: an add_reference() that occurred when we did 4685 * the arc_read() will have kicked this off the list. 4686 * If it was a prefetch, we will explicitly move it to 4687 * the head of the list now. 4688 */ 4689 ARCSTAT_BUMP(arcstat_mfu_hits); 4690 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 4691 } else if (hdr->b_l1hdr.b_state == arc_mfu_ghost) { 4692 arc_state_t *new_state = arc_mfu; 4693 /* 4694 * This buffer has been accessed more than once but has 4695 * been evicted from the cache. Move it back to the 4696 * MFU state. 4697 */ 4698 4699 if (HDR_PREFETCH(hdr) || HDR_PRESCIENT_PREFETCH(hdr)) { 4700 /* 4701 * This is a prefetch access... 4702 * move this block back to the MRU state. 4703 */ 4704 new_state = arc_mru; 4705 } 4706 4707 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 4708 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 4709 arc_change_state(new_state, hdr, hash_lock); 4710 4711 ARCSTAT_BUMP(arcstat_mfu_ghost_hits); 4712 } else if (hdr->b_l1hdr.b_state == arc_l2c_only) { 4713 /* 4714 * This buffer is on the 2nd Level ARC. 4715 */ 4716 4717 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 4718 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 4719 arc_change_state(arc_mfu, hdr, hash_lock); 4720 } else { 4721 ASSERT(!"invalid arc state"); 4722 } 4723 } 4724 4725 /* 4726 * This routine is called by dbuf_hold() to update the arc_access() state 4727 * which otherwise would be skipped for entries in the dbuf cache. 4728 */ 4729 void 4730 arc_buf_access(arc_buf_t *buf) 4731 { 4732 mutex_enter(&buf->b_evict_lock); 4733 arc_buf_hdr_t *hdr = buf->b_hdr; 4734 4735 /* 4736 * Avoid taking the hash_lock when possible as an optimization. 4737 * The header must be checked again under the hash_lock in order 4738 * to handle the case where it is concurrently being released. 4739 */ 4740 if (hdr->b_l1hdr.b_state == arc_anon || HDR_EMPTY(hdr)) { 4741 mutex_exit(&buf->b_evict_lock); 4742 return; 4743 } 4744 4745 kmutex_t *hash_lock = HDR_LOCK(hdr); 4746 mutex_enter(hash_lock); 4747 4748 if (hdr->b_l1hdr.b_state == arc_anon || HDR_EMPTY(hdr)) { 4749 mutex_exit(hash_lock); 4750 mutex_exit(&buf->b_evict_lock); 4751 ARCSTAT_BUMP(arcstat_access_skip); 4752 return; 4753 } 4754 4755 mutex_exit(&buf->b_evict_lock); 4756 4757 ASSERT(hdr->b_l1hdr.b_state == arc_mru || 4758 hdr->b_l1hdr.b_state == arc_mfu); 4759 4760 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr); 4761 arc_access(hdr, hash_lock); 4762 mutex_exit(hash_lock); 4763 4764 ARCSTAT_BUMP(arcstat_hits); 4765 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr), 4766 demand, prefetch, !HDR_ISTYPE_METADATA(hdr), data, metadata, hits); 4767 } 4768 4769 /* a generic arc_read_done_func_t which you can use */ 4770 /* ARGSUSED */ 4771 void 4772 arc_bcopy_func(zio_t *zio, const zbookmark_phys_t *zb, const blkptr_t *bp, 4773 arc_buf_t *buf, void *arg) 4774 { 4775 if (buf == NULL) 4776 return; 4777 4778 bcopy(buf->b_data, arg, arc_buf_size(buf)); 4779 arc_buf_destroy(buf, arg); 4780 } 4781 4782 /* a generic arc_read_done_func_t */ 4783 void 4784 arc_getbuf_func(zio_t *zio, const zbookmark_phys_t *zb, const blkptr_t *bp, 4785 arc_buf_t *buf, void *arg) 4786 { 4787 arc_buf_t **bufp = arg; 4788 4789 if (buf == NULL) { 4790 ASSERT(zio == NULL || zio->io_error != 0); 4791 *bufp = NULL; 4792 } else { 4793 ASSERT(zio == NULL || zio->io_error == 0); 4794 *bufp = buf; 4795 ASSERT(buf->b_data != NULL); 4796 } 4797 } 4798 4799 static void 4800 arc_hdr_verify(arc_buf_hdr_t *hdr, blkptr_t *bp) 4801 { 4802 if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp)) { 4803 ASSERT3U(HDR_GET_PSIZE(hdr), ==, 0); 4804 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF); 4805 } else { 4806 if (HDR_COMPRESSION_ENABLED(hdr)) { 4807 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, 4808 BP_GET_COMPRESS(bp)); 4809 } 4810 ASSERT3U(HDR_GET_LSIZE(hdr), ==, BP_GET_LSIZE(bp)); 4811 ASSERT3U(HDR_GET_PSIZE(hdr), ==, BP_GET_PSIZE(bp)); 4812 } 4813 } 4814 4815 static void 4816 arc_read_done(zio_t *zio) 4817 { 4818 arc_buf_hdr_t *hdr = zio->io_private; 4819 kmutex_t *hash_lock = NULL; 4820 arc_callback_t *callback_list; 4821 arc_callback_t *acb; 4822 boolean_t freeable = B_FALSE; 4823 4824 /* 4825 * The hdr was inserted into hash-table and removed from lists 4826 * prior to starting I/O. We should find this header, since 4827 * it's in the hash table, and it should be legit since it's 4828 * not possible to evict it during the I/O. The only possible 4829 * reason for it not to be found is if we were freed during the 4830 * read. 4831 */ 4832 if (HDR_IN_HASH_TABLE(hdr)) { 4833 ASSERT3U(hdr->b_birth, ==, BP_PHYSICAL_BIRTH(zio->io_bp)); 4834 ASSERT3U(hdr->b_dva.dva_word[0], ==, 4835 BP_IDENTITY(zio->io_bp)->dva_word[0]); 4836 ASSERT3U(hdr->b_dva.dva_word[1], ==, 4837 BP_IDENTITY(zio->io_bp)->dva_word[1]); 4838 4839 arc_buf_hdr_t *found = buf_hash_find(hdr->b_spa, zio->io_bp, 4840 &hash_lock); 4841 4842 ASSERT((found == hdr && 4843 DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) || 4844 (found == hdr && HDR_L2_READING(hdr))); 4845 ASSERT3P(hash_lock, !=, NULL); 4846 } 4847 4848 if (zio->io_error == 0) { 4849 /* byteswap if necessary */ 4850 if (BP_SHOULD_BYTESWAP(zio->io_bp)) { 4851 if (BP_GET_LEVEL(zio->io_bp) > 0) { 4852 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_UINT64; 4853 } else { 4854 hdr->b_l1hdr.b_byteswap = 4855 DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp)); 4856 } 4857 } else { 4858 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS; 4859 } 4860 } 4861 4862 arc_hdr_clear_flags(hdr, ARC_FLAG_L2_EVICTED); 4863 if (l2arc_noprefetch && HDR_PREFETCH(hdr)) 4864 arc_hdr_clear_flags(hdr, ARC_FLAG_L2CACHE); 4865 4866 callback_list = hdr->b_l1hdr.b_acb; 4867 ASSERT3P(callback_list, !=, NULL); 4868 4869 if (hash_lock && zio->io_error == 0 && 4870 hdr->b_l1hdr.b_state == arc_anon) { 4871 /* 4872 * Only call arc_access on anonymous buffers. This is because 4873 * if we've issued an I/O for an evicted buffer, we've already 4874 * called arc_access (to prevent any simultaneous readers from 4875 * getting confused). 4876 */ 4877 arc_access(hdr, hash_lock); 4878 } 4879 4880 /* 4881 * If a read request has a callback (i.e. acb_done is not NULL), then we 4882 * make a buf containing the data according to the parameters which were 4883 * passed in. The implementation of arc_buf_alloc_impl() ensures that we 4884 * aren't needlessly decompressing the data multiple times. 4885 */ 4886 int callback_cnt = 0; 4887 for (acb = callback_list; acb != NULL; acb = acb->acb_next) { 4888 if (!acb->acb_done) 4889 continue; 4890 4891 callback_cnt++; 4892 4893 if (zio->io_error != 0) 4894 continue; 4895 4896 int error = arc_buf_alloc_impl(hdr, acb->acb_private, 4897 acb->acb_compressed, B_TRUE, &acb->acb_buf); 4898 if (error != 0) { 4899 /* 4900 * Decompression failed. Set io_error 4901 * so that when we call acb_done (below), 4902 * we will indicate that the read failed. 4903 * Note that in the unusual case where one 4904 * callback is compressed and another 4905 * uncompressed, we will mark all of them 4906 * as failed, even though the uncompressed 4907 * one can't actually fail. In this case, 4908 * the hdr will not be anonymous, because 4909 * if there are multiple callbacks, it's 4910 * because multiple threads found the same 4911 * arc buf in the hash table. 4912 */ 4913 zio->io_error = error; 4914 } 4915 } 4916 /* 4917 * If there are multiple callbacks, we must have the hash lock, 4918 * because the only way for multiple threads to find this hdr is 4919 * in the hash table. This ensures that if there are multiple 4920 * callbacks, the hdr is not anonymous. If it were anonymous, 4921 * we couldn't use arc_buf_destroy() in the error case below. 4922 */ 4923 ASSERT(callback_cnt < 2 || hash_lock != NULL); 4924 4925 hdr->b_l1hdr.b_acb = NULL; 4926 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS); 4927 if (callback_cnt == 0) { 4928 ASSERT(HDR_PREFETCH(hdr)); 4929 ASSERT0(hdr->b_l1hdr.b_bufcnt); 4930 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL); 4931 } 4932 4933 ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt) || 4934 callback_list != NULL); 4935 4936 if (zio->io_error == 0) { 4937 arc_hdr_verify(hdr, zio->io_bp); 4938 } else { 4939 arc_hdr_set_flags(hdr, ARC_FLAG_IO_ERROR); 4940 if (hdr->b_l1hdr.b_state != arc_anon) 4941 arc_change_state(arc_anon, hdr, hash_lock); 4942 if (HDR_IN_HASH_TABLE(hdr)) 4943 buf_hash_remove(hdr); 4944 freeable = zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt); 4945 } 4946 4947 /* 4948 * Broadcast before we drop the hash_lock to avoid the possibility 4949 * that the hdr (and hence the cv) might be freed before we get to 4950 * the cv_broadcast(). 4951 */ 4952 cv_broadcast(&hdr->b_l1hdr.b_cv); 4953 4954 if (hash_lock != NULL) { 4955 mutex_exit(hash_lock); 4956 } else { 4957 /* 4958 * This block was freed while we waited for the read to 4959 * complete. It has been removed from the hash table and 4960 * moved to the anonymous state (so that it won't show up 4961 * in the cache). 4962 */ 4963 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon); 4964 freeable = zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt); 4965 } 4966 4967 /* execute each callback and free its structure */ 4968 while ((acb = callback_list) != NULL) { 4969 if (acb->acb_done != NULL) { 4970 if (zio->io_error != 0 && acb->acb_buf != NULL) { 4971 /* 4972 * If arc_buf_alloc_impl() fails during 4973 * decompression, the buf will still be 4974 * allocated, and needs to be freed here. 4975 */ 4976 arc_buf_destroy(acb->acb_buf, acb->acb_private); 4977 acb->acb_buf = NULL; 4978 } 4979 acb->acb_done(zio, &zio->io_bookmark, zio->io_bp, 4980 acb->acb_buf, acb->acb_private); 4981 } 4982 4983 if (acb->acb_zio_dummy != NULL) { 4984 acb->acb_zio_dummy->io_error = zio->io_error; 4985 zio_nowait(acb->acb_zio_dummy); 4986 } 4987 4988 callback_list = acb->acb_next; 4989 kmem_free(acb, sizeof (arc_callback_t)); 4990 } 4991 4992 if (freeable) 4993 arc_hdr_destroy(hdr); 4994 } 4995 4996 /* 4997 * "Read" the block at the specified DVA (in bp) via the 4998 * cache. If the block is found in the cache, invoke the provided 4999 * callback immediately and return. Note that the `zio' parameter 5000 * in the callback will be NULL in this case, since no IO was 5001 * required. If the block is not in the cache pass the read request 5002 * on to the spa with a substitute callback function, so that the 5003 * requested block will be added to the cache. 5004 * 5005 * If a read request arrives for a block that has a read in-progress, 5006 * either wait for the in-progress read to complete (and return the 5007 * results); or, if this is a read with a "done" func, add a record 5008 * to the read to invoke the "done" func when the read completes, 5009 * and return; or just return. 5010 * 5011 * arc_read_done() will invoke all the requested "done" functions 5012 * for readers of this block. 5013 */ 5014 int 5015 arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_read_done_func_t *done, 5016 void *private, zio_priority_t priority, int zio_flags, 5017 arc_flags_t *arc_flags, const zbookmark_phys_t *zb) 5018 { 5019 arc_buf_hdr_t *hdr = NULL; 5020 kmutex_t *hash_lock = NULL; 5021 zio_t *rzio; 5022 uint64_t guid = spa_load_guid(spa); 5023 boolean_t compressed_read = (zio_flags & ZIO_FLAG_RAW) != 0; 5024 int rc = 0; 5025 5026 ASSERT(!BP_IS_EMBEDDED(bp) || 5027 BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA); 5028 5029 top: 5030 if (!BP_IS_EMBEDDED(bp)) { 5031 /* 5032 * Embedded BP's have no DVA and require no I/O to "read". 5033 * Create an anonymous arc buf to back it. 5034 */ 5035 hdr = buf_hash_find(guid, bp, &hash_lock); 5036 } 5037 5038 if (hdr != NULL && HDR_HAS_L1HDR(hdr) && hdr->b_l1hdr.b_pabd != NULL) { 5039 arc_buf_t *buf = NULL; 5040 *arc_flags |= ARC_FLAG_CACHED; 5041 5042 if (HDR_IO_IN_PROGRESS(hdr)) { 5043 zio_t *head_zio = hdr->b_l1hdr.b_acb->acb_zio_head; 5044 5045 ASSERT3P(head_zio, !=, NULL); 5046 if ((hdr->b_flags & ARC_FLAG_PRIO_ASYNC_READ) && 5047 priority == ZIO_PRIORITY_SYNC_READ) { 5048 /* 5049 * This is a sync read that needs to wait for 5050 * an in-flight async read. Request that the 5051 * zio have its priority upgraded. 5052 */ 5053 zio_change_priority(head_zio, priority); 5054 DTRACE_PROBE1(arc__async__upgrade__sync, 5055 arc_buf_hdr_t *, hdr); 5056 ARCSTAT_BUMP(arcstat_async_upgrade_sync); 5057 } 5058 if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) { 5059 arc_hdr_clear_flags(hdr, 5060 ARC_FLAG_PREDICTIVE_PREFETCH); 5061 } 5062 5063 if (*arc_flags & ARC_FLAG_WAIT) { 5064 cv_wait(&hdr->b_l1hdr.b_cv, hash_lock); 5065 mutex_exit(hash_lock); 5066 goto top; 5067 } 5068 ASSERT(*arc_flags & ARC_FLAG_NOWAIT); 5069 5070 if (done) { 5071 arc_callback_t *acb = NULL; 5072 5073 acb = kmem_zalloc(sizeof (arc_callback_t), 5074 KM_SLEEP); 5075 acb->acb_done = done; 5076 acb->acb_private = private; 5077 acb->acb_compressed = compressed_read; 5078 if (pio != NULL) 5079 acb->acb_zio_dummy = zio_null(pio, 5080 spa, NULL, NULL, NULL, zio_flags); 5081 5082 ASSERT3P(acb->acb_done, !=, NULL); 5083 acb->acb_zio_head = head_zio; 5084 acb->acb_next = hdr->b_l1hdr.b_acb; 5085 hdr->b_l1hdr.b_acb = acb; 5086 mutex_exit(hash_lock); 5087 return (0); 5088 } 5089 mutex_exit(hash_lock); 5090 return (0); 5091 } 5092 5093 ASSERT(hdr->b_l1hdr.b_state == arc_mru || 5094 hdr->b_l1hdr.b_state == arc_mfu); 5095 5096 if (done) { 5097 if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) { 5098 /* 5099 * This is a demand read which does not have to 5100 * wait for i/o because we did a predictive 5101 * prefetch i/o for it, which has completed. 5102 */ 5103 DTRACE_PROBE1( 5104 arc__demand__hit__predictive__prefetch, 5105 arc_buf_hdr_t *, hdr); 5106 ARCSTAT_BUMP( 5107 arcstat_demand_hit_predictive_prefetch); 5108 arc_hdr_clear_flags(hdr, 5109 ARC_FLAG_PREDICTIVE_PREFETCH); 5110 } 5111 5112 if (hdr->b_flags & ARC_FLAG_PRESCIENT_PREFETCH) { 5113 ARCSTAT_BUMP( 5114 arcstat_demand_hit_prescient_prefetch); 5115 arc_hdr_clear_flags(hdr, 5116 ARC_FLAG_PRESCIENT_PREFETCH); 5117 } 5118 5119 ASSERT(!BP_IS_EMBEDDED(bp) || !BP_IS_HOLE(bp)); 5120 5121 /* Get a buf with the desired data in it. */ 5122 rc = arc_buf_alloc_impl(hdr, private, 5123 compressed_read, B_TRUE, &buf); 5124 if (rc != 0) { 5125 arc_buf_destroy(buf, private); 5126 buf = NULL; 5127 } 5128 ASSERT((zio_flags & ZIO_FLAG_SPECULATIVE) || 5129 rc == 0 || rc != ENOENT); 5130 } else if (*arc_flags & ARC_FLAG_PREFETCH && 5131 zfs_refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) { 5132 arc_hdr_set_flags(hdr, ARC_FLAG_PREFETCH); 5133 } 5134 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr); 5135 arc_access(hdr, hash_lock); 5136 if (*arc_flags & ARC_FLAG_PRESCIENT_PREFETCH) 5137 arc_hdr_set_flags(hdr, ARC_FLAG_PRESCIENT_PREFETCH); 5138 if (*arc_flags & ARC_FLAG_L2CACHE) 5139 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE); 5140 mutex_exit(hash_lock); 5141 ARCSTAT_BUMP(arcstat_hits); 5142 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr), 5143 demand, prefetch, !HDR_ISTYPE_METADATA(hdr), 5144 data, metadata, hits); 5145 5146 if (done) 5147 done(NULL, zb, bp, buf, private); 5148 } else { 5149 uint64_t lsize = BP_GET_LSIZE(bp); 5150 uint64_t psize = BP_GET_PSIZE(bp); 5151 arc_callback_t *acb; 5152 vdev_t *vd = NULL; 5153 uint64_t addr = 0; 5154 boolean_t devw = B_FALSE; 5155 uint64_t size; 5156 5157 if (hdr == NULL) { 5158 /* this block is not in the cache */ 5159 arc_buf_hdr_t *exists = NULL; 5160 arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp); 5161 hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize, 5162 BP_GET_COMPRESS(bp), type); 5163 5164 if (!BP_IS_EMBEDDED(bp)) { 5165 hdr->b_dva = *BP_IDENTITY(bp); 5166 hdr->b_birth = BP_PHYSICAL_BIRTH(bp); 5167 exists = buf_hash_insert(hdr, &hash_lock); 5168 } 5169 if (exists != NULL) { 5170 /* somebody beat us to the hash insert */ 5171 mutex_exit(hash_lock); 5172 buf_discard_identity(hdr); 5173 arc_hdr_destroy(hdr); 5174 goto top; /* restart the IO request */ 5175 } 5176 } else { 5177 /* 5178 * This block is in the ghost cache. If it was L2-only 5179 * (and thus didn't have an L1 hdr), we realloc the 5180 * header to add an L1 hdr. 5181 */ 5182 if (!HDR_HAS_L1HDR(hdr)) { 5183 hdr = arc_hdr_realloc(hdr, hdr_l2only_cache, 5184 hdr_full_cache); 5185 } 5186 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL); 5187 ASSERT(GHOST_STATE(hdr->b_l1hdr.b_state)); 5188 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 5189 ASSERT(zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 5190 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 5191 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL); 5192 5193 /* 5194 * This is a delicate dance that we play here. 5195 * This hdr is in the ghost list so we access it 5196 * to move it out of the ghost list before we 5197 * initiate the read. If it's a prefetch then 5198 * it won't have a callback so we'll remove the 5199 * reference that arc_buf_alloc_impl() created. We 5200 * do this after we've called arc_access() to 5201 * avoid hitting an assert in remove_reference(). 5202 */ 5203 arc_access(hdr, hash_lock); 5204 arc_hdr_alloc_pabd(hdr); 5205 } 5206 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL); 5207 size = arc_hdr_size(hdr); 5208 5209 /* 5210 * If compression is enabled on the hdr, then will do 5211 * RAW I/O and will store the compressed data in the hdr's 5212 * data block. Otherwise, the hdr's data block will contain 5213 * the uncompressed data. 5214 */ 5215 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF) { 5216 zio_flags |= ZIO_FLAG_RAW; 5217 } 5218 5219 if (*arc_flags & ARC_FLAG_PREFETCH) 5220 arc_hdr_set_flags(hdr, ARC_FLAG_PREFETCH); 5221 if (*arc_flags & ARC_FLAG_PRESCIENT_PREFETCH) 5222 arc_hdr_set_flags(hdr, ARC_FLAG_PRESCIENT_PREFETCH); 5223 5224 if (*arc_flags & ARC_FLAG_L2CACHE) 5225 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE); 5226 if (BP_GET_LEVEL(bp) > 0) 5227 arc_hdr_set_flags(hdr, ARC_FLAG_INDIRECT); 5228 if (*arc_flags & ARC_FLAG_PREDICTIVE_PREFETCH) 5229 arc_hdr_set_flags(hdr, ARC_FLAG_PREDICTIVE_PREFETCH); 5230 ASSERT(!GHOST_STATE(hdr->b_l1hdr.b_state)); 5231 5232 acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP); 5233 acb->acb_done = done; 5234 acb->acb_private = private; 5235 acb->acb_compressed = compressed_read; 5236 5237 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL); 5238 hdr->b_l1hdr.b_acb = acb; 5239 arc_hdr_set_flags(hdr, ARC_FLAG_IO_IN_PROGRESS); 5240 5241 if (HDR_HAS_L2HDR(hdr) && 5242 (vd = hdr->b_l2hdr.b_dev->l2ad_vdev) != NULL) { 5243 devw = hdr->b_l2hdr.b_dev->l2ad_writing; 5244 addr = hdr->b_l2hdr.b_daddr; 5245 /* 5246 * Lock out L2ARC device removal. 5247 */ 5248 if (vdev_is_dead(vd) || 5249 !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER)) 5250 vd = NULL; 5251 } 5252 5253 /* 5254 * We count both async reads and scrub IOs as asynchronous so 5255 * that both can be upgraded in the event of a cache hit while 5256 * the read IO is still in-flight. 5257 */ 5258 if (priority == ZIO_PRIORITY_ASYNC_READ || 5259 priority == ZIO_PRIORITY_SCRUB) 5260 arc_hdr_set_flags(hdr, ARC_FLAG_PRIO_ASYNC_READ); 5261 else 5262 arc_hdr_clear_flags(hdr, ARC_FLAG_PRIO_ASYNC_READ); 5263 5264 /* 5265 * At this point, we have a level 1 cache miss. Try again in 5266 * L2ARC if possible. 5267 */ 5268 ASSERT3U(HDR_GET_LSIZE(hdr), ==, lsize); 5269 5270 DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp, 5271 uint64_t, lsize, zbookmark_phys_t *, zb); 5272 ARCSTAT_BUMP(arcstat_misses); 5273 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr), 5274 demand, prefetch, !HDR_ISTYPE_METADATA(hdr), 5275 data, metadata, misses); 5276 5277 if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) { 5278 /* 5279 * Read from the L2ARC if the following are true: 5280 * 1. The L2ARC vdev was previously cached. 5281 * 2. This buffer still has L2ARC metadata. 5282 * 3. This buffer isn't currently writing to the L2ARC. 5283 * 4. The L2ARC entry wasn't evicted, which may 5284 * also have invalidated the vdev. 5285 * 5. This isn't prefetch and l2arc_noprefetch is set. 5286 */ 5287 if (HDR_HAS_L2HDR(hdr) && 5288 !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) && 5289 !(l2arc_noprefetch && HDR_PREFETCH(hdr))) { 5290 l2arc_read_callback_t *cb; 5291 abd_t *abd; 5292 uint64_t asize; 5293 5294 DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr); 5295 ARCSTAT_BUMP(arcstat_l2_hits); 5296 5297 cb = kmem_zalloc(sizeof (l2arc_read_callback_t), 5298 KM_SLEEP); 5299 cb->l2rcb_hdr = hdr; 5300 cb->l2rcb_bp = *bp; 5301 cb->l2rcb_zb = *zb; 5302 cb->l2rcb_flags = zio_flags; 5303 5304 asize = vdev_psize_to_asize(vd, size); 5305 if (asize != size) { 5306 abd = abd_alloc_for_io(asize, 5307 HDR_ISTYPE_METADATA(hdr)); 5308 cb->l2rcb_abd = abd; 5309 } else { 5310 abd = hdr->b_l1hdr.b_pabd; 5311 } 5312 5313 ASSERT(addr >= VDEV_LABEL_START_SIZE && 5314 addr + asize <= vd->vdev_psize - 5315 VDEV_LABEL_END_SIZE); 5316 5317 /* 5318 * l2arc read. The SCL_L2ARC lock will be 5319 * released by l2arc_read_done(). 5320 * Issue a null zio if the underlying buffer 5321 * was squashed to zero size by compression. 5322 */ 5323 ASSERT3U(HDR_GET_COMPRESS(hdr), !=, 5324 ZIO_COMPRESS_EMPTY); 5325 rzio = zio_read_phys(pio, vd, addr, 5326 asize, abd, 5327 ZIO_CHECKSUM_OFF, 5328 l2arc_read_done, cb, priority, 5329 zio_flags | ZIO_FLAG_DONT_CACHE | 5330 ZIO_FLAG_CANFAIL | 5331 ZIO_FLAG_DONT_PROPAGATE | 5332 ZIO_FLAG_DONT_RETRY, B_FALSE); 5333 acb->acb_zio_head = rzio; 5334 5335 if (hash_lock != NULL) 5336 mutex_exit(hash_lock); 5337 5338 DTRACE_PROBE2(l2arc__read, vdev_t *, vd, 5339 zio_t *, rzio); 5340 ARCSTAT_INCR(arcstat_l2_read_bytes, size); 5341 5342 if (*arc_flags & ARC_FLAG_NOWAIT) { 5343 zio_nowait(rzio); 5344 return (0); 5345 } 5346 5347 ASSERT(*arc_flags & ARC_FLAG_WAIT); 5348 if (zio_wait(rzio) == 0) 5349 return (0); 5350 5351 /* l2arc read error; goto zio_read() */ 5352 if (hash_lock != NULL) 5353 mutex_enter(hash_lock); 5354 } else { 5355 DTRACE_PROBE1(l2arc__miss, 5356 arc_buf_hdr_t *, hdr); 5357 ARCSTAT_BUMP(arcstat_l2_misses); 5358 if (HDR_L2_WRITING(hdr)) 5359 ARCSTAT_BUMP(arcstat_l2_rw_clash); 5360 spa_config_exit(spa, SCL_L2ARC, vd); 5361 } 5362 } else { 5363 if (vd != NULL) 5364 spa_config_exit(spa, SCL_L2ARC, vd); 5365 if (l2arc_ndev != 0) { 5366 DTRACE_PROBE1(l2arc__miss, 5367 arc_buf_hdr_t *, hdr); 5368 ARCSTAT_BUMP(arcstat_l2_misses); 5369 } 5370 } 5371 5372 rzio = zio_read(pio, spa, bp, hdr->b_l1hdr.b_pabd, size, 5373 arc_read_done, hdr, priority, zio_flags, zb); 5374 acb->acb_zio_head = rzio; 5375 5376 if (hash_lock != NULL) 5377 mutex_exit(hash_lock); 5378 5379 if (*arc_flags & ARC_FLAG_WAIT) 5380 return (zio_wait(rzio)); 5381 5382 ASSERT(*arc_flags & ARC_FLAG_NOWAIT); 5383 zio_nowait(rzio); 5384 } 5385 return (0); 5386 } 5387 5388 /* 5389 * Notify the arc that a block was freed, and thus will never be used again. 5390 */ 5391 void 5392 arc_freed(spa_t *spa, const blkptr_t *bp) 5393 { 5394 arc_buf_hdr_t *hdr; 5395 kmutex_t *hash_lock; 5396 uint64_t guid = spa_load_guid(spa); 5397 5398 ASSERT(!BP_IS_EMBEDDED(bp)); 5399 5400 hdr = buf_hash_find(guid, bp, &hash_lock); 5401 if (hdr == NULL) 5402 return; 5403 5404 /* 5405 * We might be trying to free a block that is still doing I/O 5406 * (i.e. prefetch) or has a reference (i.e. a dedup-ed, 5407 * dmu_sync-ed block). If this block is being prefetched, then it 5408 * would still have the ARC_FLAG_IO_IN_PROGRESS flag set on the hdr 5409 * until the I/O completes. A block may also have a reference if it is 5410 * part of a dedup-ed, dmu_synced write. The dmu_sync() function would 5411 * have written the new block to its final resting place on disk but 5412 * without the dedup flag set. This would have left the hdr in the MRU 5413 * state and discoverable. When the txg finally syncs it detects that 5414 * the block was overridden in open context and issues an override I/O. 5415 * Since this is a dedup block, the override I/O will determine if the 5416 * block is already in the DDT. If so, then it will replace the io_bp 5417 * with the bp from the DDT and allow the I/O to finish. When the I/O 5418 * reaches the done callback, dbuf_write_override_done, it will 5419 * check to see if the io_bp and io_bp_override are identical. 5420 * If they are not, then it indicates that the bp was replaced with 5421 * the bp in the DDT and the override bp is freed. This allows 5422 * us to arrive here with a reference on a block that is being 5423 * freed. So if we have an I/O in progress, or a reference to 5424 * this hdr, then we don't destroy the hdr. 5425 */ 5426 if (!HDR_HAS_L1HDR(hdr) || (!HDR_IO_IN_PROGRESS(hdr) && 5427 zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt))) { 5428 arc_change_state(arc_anon, hdr, hash_lock); 5429 arc_hdr_destroy(hdr); 5430 mutex_exit(hash_lock); 5431 } else { 5432 mutex_exit(hash_lock); 5433 } 5434 5435 } 5436 5437 /* 5438 * Release this buffer from the cache, making it an anonymous buffer. This 5439 * must be done after a read and prior to modifying the buffer contents. 5440 * If the buffer has more than one reference, we must make 5441 * a new hdr for the buffer. 5442 */ 5443 void 5444 arc_release(arc_buf_t *buf, void *tag) 5445 { 5446 arc_buf_hdr_t *hdr = buf->b_hdr; 5447 5448 /* 5449 * It would be nice to assert that if it's DMU metadata (level > 5450 * 0 || it's the dnode file), then it must be syncing context. 5451 * But we don't know that information at this level. 5452 */ 5453 5454 mutex_enter(&buf->b_evict_lock); 5455 5456 ASSERT(HDR_HAS_L1HDR(hdr)); 5457 5458 /* 5459 * We don't grab the hash lock prior to this check, because if 5460 * the buffer's header is in the arc_anon state, it won't be 5461 * linked into the hash table. 5462 */ 5463 if (hdr->b_l1hdr.b_state == arc_anon) { 5464 mutex_exit(&buf->b_evict_lock); 5465 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 5466 ASSERT(!HDR_IN_HASH_TABLE(hdr)); 5467 ASSERT(!HDR_HAS_L2HDR(hdr)); 5468 ASSERT(HDR_EMPTY(hdr)); 5469 5470 ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1); 5471 ASSERT3S(zfs_refcount_count(&hdr->b_l1hdr.b_refcnt), ==, 1); 5472 ASSERT(!list_link_active(&hdr->b_l1hdr.b_arc_node)); 5473 5474 hdr->b_l1hdr.b_arc_access = 0; 5475 5476 /* 5477 * If the buf is being overridden then it may already 5478 * have a hdr that is not empty. 5479 */ 5480 buf_discard_identity(hdr); 5481 arc_buf_thaw(buf); 5482 5483 return; 5484 } 5485 5486 kmutex_t *hash_lock = HDR_LOCK(hdr); 5487 mutex_enter(hash_lock); 5488 5489 /* 5490 * This assignment is only valid as long as the hash_lock is 5491 * held, we must be careful not to reference state or the 5492 * b_state field after dropping the lock. 5493 */ 5494 arc_state_t *state = hdr->b_l1hdr.b_state; 5495 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 5496 ASSERT3P(state, !=, arc_anon); 5497 5498 /* this buffer is not on any list */ 5499 ASSERT3S(zfs_refcount_count(&hdr->b_l1hdr.b_refcnt), >, 0); 5500 5501 if (HDR_HAS_L2HDR(hdr)) { 5502 mutex_enter(&hdr->b_l2hdr.b_dev->l2ad_mtx); 5503 5504 /* 5505 * We have to recheck this conditional again now that 5506 * we're holding the l2ad_mtx to prevent a race with 5507 * another thread which might be concurrently calling 5508 * l2arc_evict(). In that case, l2arc_evict() might have 5509 * destroyed the header's L2 portion as we were waiting 5510 * to acquire the l2ad_mtx. 5511 */ 5512 if (HDR_HAS_L2HDR(hdr)) 5513 arc_hdr_l2hdr_destroy(hdr); 5514 5515 mutex_exit(&hdr->b_l2hdr.b_dev->l2ad_mtx); 5516 } 5517 5518 /* 5519 * Do we have more than one buf? 5520 */ 5521 if (hdr->b_l1hdr.b_bufcnt > 1) { 5522 arc_buf_hdr_t *nhdr; 5523 uint64_t spa = hdr->b_spa; 5524 uint64_t psize = HDR_GET_PSIZE(hdr); 5525 uint64_t lsize = HDR_GET_LSIZE(hdr); 5526 enum zio_compress compress = HDR_GET_COMPRESS(hdr); 5527 arc_buf_contents_t type = arc_buf_type(hdr); 5528 VERIFY3U(hdr->b_type, ==, type); 5529 5530 ASSERT(hdr->b_l1hdr.b_buf != buf || buf->b_next != NULL); 5531 (void) remove_reference(hdr, hash_lock, tag); 5532 5533 if (arc_buf_is_shared(buf) && !ARC_BUF_COMPRESSED(buf)) { 5534 ASSERT3P(hdr->b_l1hdr.b_buf, !=, buf); 5535 ASSERT(ARC_BUF_LAST(buf)); 5536 } 5537 5538 /* 5539 * Pull the data off of this hdr and attach it to 5540 * a new anonymous hdr. Also find the last buffer 5541 * in the hdr's buffer list. 5542 */ 5543 arc_buf_t *lastbuf = arc_buf_remove(hdr, buf); 5544 ASSERT3P(lastbuf, !=, NULL); 5545 5546 /* 5547 * If the current arc_buf_t and the hdr are sharing their data 5548 * buffer, then we must stop sharing that block. 5549 */ 5550 if (arc_buf_is_shared(buf)) { 5551 VERIFY(!arc_buf_is_shared(lastbuf)); 5552 5553 /* 5554 * First, sever the block sharing relationship between 5555 * buf and the arc_buf_hdr_t. 5556 */ 5557 arc_unshare_buf(hdr, buf); 5558 5559 /* 5560 * Now we need to recreate the hdr's b_pabd. Since we 5561 * have lastbuf handy, we try to share with it, but if 5562 * we can't then we allocate a new b_pabd and copy the 5563 * data from buf into it. 5564 */ 5565 if (arc_can_share(hdr, lastbuf)) { 5566 arc_share_buf(hdr, lastbuf); 5567 } else { 5568 arc_hdr_alloc_pabd(hdr); 5569 abd_copy_from_buf(hdr->b_l1hdr.b_pabd, 5570 buf->b_data, psize); 5571 } 5572 VERIFY3P(lastbuf->b_data, !=, NULL); 5573 } else if (HDR_SHARED_DATA(hdr)) { 5574 /* 5575 * Uncompressed shared buffers are always at the end 5576 * of the list. Compressed buffers don't have the 5577 * same requirements. This makes it hard to 5578 * simply assert that the lastbuf is shared so 5579 * we rely on the hdr's compression flags to determine 5580 * if we have a compressed, shared buffer. 5581 */ 5582 ASSERT(arc_buf_is_shared(lastbuf) || 5583 HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF); 5584 ASSERT(!ARC_BUF_SHARED(buf)); 5585 } 5586 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL); 5587 ASSERT3P(state, !=, arc_l2c_only); 5588 5589 (void) zfs_refcount_remove_many(&state->arcs_size, 5590 arc_buf_size(buf), buf); 5591 5592 if (zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt)) { 5593 ASSERT3P(state, !=, arc_l2c_only); 5594 (void) zfs_refcount_remove_many( 5595 &state->arcs_esize[type], 5596 arc_buf_size(buf), buf); 5597 } 5598 5599 hdr->b_l1hdr.b_bufcnt -= 1; 5600 arc_cksum_verify(buf); 5601 arc_buf_unwatch(buf); 5602 5603 mutex_exit(hash_lock); 5604 5605 /* 5606 * Allocate a new hdr. The new hdr will contain a b_pabd 5607 * buffer which will be freed in arc_write(). 5608 */ 5609 nhdr = arc_hdr_alloc(spa, psize, lsize, compress, type); 5610 ASSERT3P(nhdr->b_l1hdr.b_buf, ==, NULL); 5611 ASSERT0(nhdr->b_l1hdr.b_bufcnt); 5612 ASSERT0(zfs_refcount_count(&nhdr->b_l1hdr.b_refcnt)); 5613 VERIFY3U(nhdr->b_type, ==, type); 5614 ASSERT(!HDR_SHARED_DATA(nhdr)); 5615 5616 nhdr->b_l1hdr.b_buf = buf; 5617 nhdr->b_l1hdr.b_bufcnt = 1; 5618 (void) zfs_refcount_add(&nhdr->b_l1hdr.b_refcnt, tag); 5619 buf->b_hdr = nhdr; 5620 5621 mutex_exit(&buf->b_evict_lock); 5622 (void) zfs_refcount_add_many(&arc_anon->arcs_size, 5623 arc_buf_size(buf), buf); 5624 } else { 5625 mutex_exit(&buf->b_evict_lock); 5626 ASSERT(zfs_refcount_count(&hdr->b_l1hdr.b_refcnt) == 1); 5627 /* protected by hash lock, or hdr is on arc_anon */ 5628 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 5629 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 5630 arc_change_state(arc_anon, hdr, hash_lock); 5631 hdr->b_l1hdr.b_arc_access = 0; 5632 mutex_exit(hash_lock); 5633 5634 buf_discard_identity(hdr); 5635 arc_buf_thaw(buf); 5636 } 5637 } 5638 5639 int 5640 arc_released(arc_buf_t *buf) 5641 { 5642 int released; 5643 5644 mutex_enter(&buf->b_evict_lock); 5645 released = (buf->b_data != NULL && 5646 buf->b_hdr->b_l1hdr.b_state == arc_anon); 5647 mutex_exit(&buf->b_evict_lock); 5648 return (released); 5649 } 5650 5651 #ifdef ZFS_DEBUG 5652 int 5653 arc_referenced(arc_buf_t *buf) 5654 { 5655 int referenced; 5656 5657 mutex_enter(&buf->b_evict_lock); 5658 referenced = (zfs_refcount_count(&buf->b_hdr->b_l1hdr.b_refcnt)); 5659 mutex_exit(&buf->b_evict_lock); 5660 return (referenced); 5661 } 5662 #endif 5663 5664 static void 5665 arc_write_ready(zio_t *zio) 5666 { 5667 arc_write_callback_t *callback = zio->io_private; 5668 arc_buf_t *buf = callback->awcb_buf; 5669 arc_buf_hdr_t *hdr = buf->b_hdr; 5670 uint64_t psize = BP_IS_HOLE(zio->io_bp) ? 0 : BP_GET_PSIZE(zio->io_bp); 5671 5672 ASSERT(HDR_HAS_L1HDR(hdr)); 5673 ASSERT(!zfs_refcount_is_zero(&buf->b_hdr->b_l1hdr.b_refcnt)); 5674 ASSERT(hdr->b_l1hdr.b_bufcnt > 0); 5675 5676 /* 5677 * If we're reexecuting this zio because the pool suspended, then 5678 * cleanup any state that was previously set the first time the 5679 * callback was invoked. 5680 */ 5681 if (zio->io_flags & ZIO_FLAG_REEXECUTED) { 5682 arc_cksum_free(hdr); 5683 arc_buf_unwatch(buf); 5684 if (hdr->b_l1hdr.b_pabd != NULL) { 5685 if (arc_buf_is_shared(buf)) { 5686 arc_unshare_buf(hdr, buf); 5687 } else { 5688 arc_hdr_free_pabd(hdr); 5689 } 5690 } 5691 } 5692 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL); 5693 ASSERT(!HDR_SHARED_DATA(hdr)); 5694 ASSERT(!arc_buf_is_shared(buf)); 5695 5696 callback->awcb_ready(zio, buf, callback->awcb_private); 5697 5698 if (HDR_IO_IN_PROGRESS(hdr)) 5699 ASSERT(zio->io_flags & ZIO_FLAG_REEXECUTED); 5700 5701 arc_cksum_compute(buf); 5702 arc_hdr_set_flags(hdr, ARC_FLAG_IO_IN_PROGRESS); 5703 5704 enum zio_compress compress; 5705 if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) { 5706 compress = ZIO_COMPRESS_OFF; 5707 } else { 5708 ASSERT3U(HDR_GET_LSIZE(hdr), ==, BP_GET_LSIZE(zio->io_bp)); 5709 compress = BP_GET_COMPRESS(zio->io_bp); 5710 } 5711 HDR_SET_PSIZE(hdr, psize); 5712 arc_hdr_set_compress(hdr, compress); 5713 5714 5715 /* 5716 * Fill the hdr with data. If the hdr is compressed, the data we want 5717 * is available from the zio, otherwise we can take it from the buf. 5718 * 5719 * We might be able to share the buf's data with the hdr here. However, 5720 * doing so would cause the ARC to be full of linear ABDs if we write a 5721 * lot of shareable data. As a compromise, we check whether scattered 5722 * ABDs are allowed, and assume that if they are then the user wants 5723 * the ARC to be primarily filled with them regardless of the data being 5724 * written. Therefore, if they're allowed then we allocate one and copy 5725 * the data into it; otherwise, we share the data directly if we can. 5726 */ 5727 if (zfs_abd_scatter_enabled || !arc_can_share(hdr, buf)) { 5728 arc_hdr_alloc_pabd(hdr); 5729 5730 /* 5731 * Ideally, we would always copy the io_abd into b_pabd, but the 5732 * user may have disabled compressed ARC, thus we must check the 5733 * hdr's compression setting rather than the io_bp's. 5734 */ 5735 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF) { 5736 ASSERT3U(BP_GET_COMPRESS(zio->io_bp), !=, 5737 ZIO_COMPRESS_OFF); 5738 ASSERT3U(psize, >, 0); 5739 5740 abd_copy(hdr->b_l1hdr.b_pabd, zio->io_abd, psize); 5741 } else { 5742 ASSERT3U(zio->io_orig_size, ==, arc_hdr_size(hdr)); 5743 5744 abd_copy_from_buf(hdr->b_l1hdr.b_pabd, buf->b_data, 5745 arc_buf_size(buf)); 5746 } 5747 } else { 5748 ASSERT3P(buf->b_data, ==, abd_to_buf(zio->io_orig_abd)); 5749 ASSERT3U(zio->io_orig_size, ==, arc_buf_size(buf)); 5750 ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1); 5751 5752 arc_share_buf(hdr, buf); 5753 } 5754 5755 arc_hdr_verify(hdr, zio->io_bp); 5756 } 5757 5758 static void 5759 arc_write_children_ready(zio_t *zio) 5760 { 5761 arc_write_callback_t *callback = zio->io_private; 5762 arc_buf_t *buf = callback->awcb_buf; 5763 5764 callback->awcb_children_ready(zio, buf, callback->awcb_private); 5765 } 5766 5767 /* 5768 * The SPA calls this callback for each physical write that happens on behalf 5769 * of a logical write. See the comment in dbuf_write_physdone() for details. 5770 */ 5771 static void 5772 arc_write_physdone(zio_t *zio) 5773 { 5774 arc_write_callback_t *cb = zio->io_private; 5775 if (cb->awcb_physdone != NULL) 5776 cb->awcb_physdone(zio, cb->awcb_buf, cb->awcb_private); 5777 } 5778 5779 static void 5780 arc_write_done(zio_t *zio) 5781 { 5782 arc_write_callback_t *callback = zio->io_private; 5783 arc_buf_t *buf = callback->awcb_buf; 5784 arc_buf_hdr_t *hdr = buf->b_hdr; 5785 5786 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL); 5787 5788 if (zio->io_error == 0) { 5789 arc_hdr_verify(hdr, zio->io_bp); 5790 5791 if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) { 5792 buf_discard_identity(hdr); 5793 } else { 5794 hdr->b_dva = *BP_IDENTITY(zio->io_bp); 5795 hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp); 5796 } 5797 } else { 5798 ASSERT(HDR_EMPTY(hdr)); 5799 } 5800 5801 /* 5802 * If the block to be written was all-zero or compressed enough to be 5803 * embedded in the BP, no write was performed so there will be no 5804 * dva/birth/checksum. The buffer must therefore remain anonymous 5805 * (and uncached). 5806 */ 5807 if (!HDR_EMPTY(hdr)) { 5808 arc_buf_hdr_t *exists; 5809 kmutex_t *hash_lock; 5810 5811 ASSERT3U(zio->io_error, ==, 0); 5812 5813 arc_cksum_verify(buf); 5814 5815 exists = buf_hash_insert(hdr, &hash_lock); 5816 if (exists != NULL) { 5817 /* 5818 * This can only happen if we overwrite for 5819 * sync-to-convergence, because we remove 5820 * buffers from the hash table when we arc_free(). 5821 */ 5822 if (zio->io_flags & ZIO_FLAG_IO_REWRITE) { 5823 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) 5824 panic("bad overwrite, hdr=%p exists=%p", 5825 (void *)hdr, (void *)exists); 5826 ASSERT(zfs_refcount_is_zero( 5827 &exists->b_l1hdr.b_refcnt)); 5828 arc_change_state(arc_anon, exists, hash_lock); 5829 mutex_exit(hash_lock); 5830 arc_hdr_destroy(exists); 5831 exists = buf_hash_insert(hdr, &hash_lock); 5832 ASSERT3P(exists, ==, NULL); 5833 } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) { 5834 /* nopwrite */ 5835 ASSERT(zio->io_prop.zp_nopwrite); 5836 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) 5837 panic("bad nopwrite, hdr=%p exists=%p", 5838 (void *)hdr, (void *)exists); 5839 } else { 5840 /* Dedup */ 5841 ASSERT(hdr->b_l1hdr.b_bufcnt == 1); 5842 ASSERT(hdr->b_l1hdr.b_state == arc_anon); 5843 ASSERT(BP_GET_DEDUP(zio->io_bp)); 5844 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0); 5845 } 5846 } 5847 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS); 5848 /* if it's not anon, we are doing a scrub */ 5849 if (exists == NULL && hdr->b_l1hdr.b_state == arc_anon) 5850 arc_access(hdr, hash_lock); 5851 mutex_exit(hash_lock); 5852 } else { 5853 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS); 5854 } 5855 5856 ASSERT(!zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 5857 callback->awcb_done(zio, buf, callback->awcb_private); 5858 5859 abd_put(zio->io_abd); 5860 kmem_free(callback, sizeof (arc_write_callback_t)); 5861 } 5862 5863 zio_t * 5864 arc_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, arc_buf_t *buf, 5865 boolean_t l2arc, const zio_prop_t *zp, arc_write_done_func_t *ready, 5866 arc_write_done_func_t *children_ready, arc_write_done_func_t *physdone, 5867 arc_write_done_func_t *done, void *private, zio_priority_t priority, 5868 int zio_flags, const zbookmark_phys_t *zb) 5869 { 5870 arc_buf_hdr_t *hdr = buf->b_hdr; 5871 arc_write_callback_t *callback; 5872 zio_t *zio; 5873 zio_prop_t localprop = *zp; 5874 5875 ASSERT3P(ready, !=, NULL); 5876 ASSERT3P(done, !=, NULL); 5877 ASSERT(!HDR_IO_ERROR(hdr)); 5878 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 5879 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL); 5880 ASSERT3U(hdr->b_l1hdr.b_bufcnt, >, 0); 5881 if (l2arc) 5882 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE); 5883 if (ARC_BUF_COMPRESSED(buf)) { 5884 /* 5885 * We're writing a pre-compressed buffer. Make the 5886 * compression algorithm requested by the zio_prop_t match 5887 * the pre-compressed buffer's compression algorithm. 5888 */ 5889 localprop.zp_compress = HDR_GET_COMPRESS(hdr); 5890 5891 ASSERT3U(HDR_GET_LSIZE(hdr), !=, arc_buf_size(buf)); 5892 zio_flags |= ZIO_FLAG_RAW; 5893 } 5894 callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP); 5895 callback->awcb_ready = ready; 5896 callback->awcb_children_ready = children_ready; 5897 callback->awcb_physdone = physdone; 5898 callback->awcb_done = done; 5899 callback->awcb_private = private; 5900 callback->awcb_buf = buf; 5901 5902 /* 5903 * The hdr's b_pabd is now stale, free it now. A new data block 5904 * will be allocated when the zio pipeline calls arc_write_ready(). 5905 */ 5906 if (hdr->b_l1hdr.b_pabd != NULL) { 5907 /* 5908 * If the buf is currently sharing the data block with 5909 * the hdr then we need to break that relationship here. 5910 * The hdr will remain with a NULL data pointer and the 5911 * buf will take sole ownership of the block. 5912 */ 5913 if (arc_buf_is_shared(buf)) { 5914 arc_unshare_buf(hdr, buf); 5915 } else { 5916 arc_hdr_free_pabd(hdr); 5917 } 5918 VERIFY3P(buf->b_data, !=, NULL); 5919 arc_hdr_set_compress(hdr, ZIO_COMPRESS_OFF); 5920 } 5921 ASSERT(!arc_buf_is_shared(buf)); 5922 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL); 5923 5924 zio = zio_write(pio, spa, txg, bp, 5925 abd_get_from_buf(buf->b_data, HDR_GET_LSIZE(hdr)), 5926 HDR_GET_LSIZE(hdr), arc_buf_size(buf), &localprop, arc_write_ready, 5927 (children_ready != NULL) ? arc_write_children_ready : NULL, 5928 arc_write_physdone, arc_write_done, callback, 5929 priority, zio_flags, zb); 5930 5931 return (zio); 5932 } 5933 5934 static int 5935 arc_memory_throttle(spa_t *spa, uint64_t reserve, uint64_t txg) 5936 { 5937 #ifdef _KERNEL 5938 uint64_t available_memory = ptob(freemem); 5939 5940 #if defined(__i386) 5941 available_memory = 5942 MIN(available_memory, vmem_size(heap_arena, VMEM_FREE)); 5943 #endif 5944 5945 if (freemem > physmem * arc_lotsfree_percent / 100) 5946 return (0); 5947 5948 if (txg > spa->spa_lowmem_last_txg) { 5949 spa->spa_lowmem_last_txg = txg; 5950 spa->spa_lowmem_page_load = 0; 5951 } 5952 /* 5953 * If we are in pageout, we know that memory is already tight, 5954 * the arc is already going to be evicting, so we just want to 5955 * continue to let page writes occur as quickly as possible. 5956 */ 5957 if (curproc == proc_pageout) { 5958 if (spa->spa_lowmem_page_load > 5959 MAX(ptob(minfree), available_memory) / 4) 5960 return (SET_ERROR(ERESTART)); 5961 /* Note: reserve is inflated, so we deflate */ 5962 atomic_add_64(&spa->spa_lowmem_page_load, reserve / 8); 5963 return (0); 5964 } else if (spa->spa_lowmem_page_load > 0 && arc_reclaim_needed()) { 5965 /* memory is low, delay before restarting */ 5966 ARCSTAT_INCR(arcstat_memory_throttle_count, 1); 5967 return (SET_ERROR(EAGAIN)); 5968 } 5969 spa->spa_lowmem_page_load = 0; 5970 #endif /* _KERNEL */ 5971 return (0); 5972 } 5973 5974 void 5975 arc_tempreserve_clear(uint64_t reserve) 5976 { 5977 atomic_add_64(&arc_tempreserve, -reserve); 5978 ASSERT((int64_t)arc_tempreserve >= 0); 5979 } 5980 5981 int 5982 arc_tempreserve_space(spa_t *spa, uint64_t reserve, uint64_t txg) 5983 { 5984 int error; 5985 uint64_t anon_size; 5986 5987 if (reserve > arc_c/4 && !arc_no_grow) 5988 arc_c = MIN(arc_c_max, reserve * 4); 5989 if (reserve > arc_c) 5990 return (SET_ERROR(ENOMEM)); 5991 5992 /* 5993 * Don't count loaned bufs as in flight dirty data to prevent long 5994 * network delays from blocking transactions that are ready to be 5995 * assigned to a txg. 5996 */ 5997 5998 /* assert that it has not wrapped around */ 5999 ASSERT3S(atomic_add_64_nv(&arc_loaned_bytes, 0), >=, 0); 6000 6001 anon_size = MAX((int64_t)(zfs_refcount_count(&arc_anon->arcs_size) - 6002 arc_loaned_bytes), 0); 6003 6004 /* 6005 * Writes will, almost always, require additional memory allocations 6006 * in order to compress/encrypt/etc the data. We therefore need to 6007 * make sure that there is sufficient available memory for this. 6008 */ 6009 error = arc_memory_throttle(spa, reserve, txg); 6010 if (error != 0) 6011 return (error); 6012 6013 /* 6014 * Throttle writes when the amount of dirty data in the cache 6015 * gets too large. We try to keep the cache less than half full 6016 * of dirty blocks so that our sync times don't grow too large. 6017 * 6018 * In the case of one pool being built on another pool, we want 6019 * to make sure we don't end up throttling the lower (backing) 6020 * pool when the upper pool is the majority contributor to dirty 6021 * data. To insure we make forward progress during throttling, we 6022 * also check the current pool's net dirty data and only throttle 6023 * if it exceeds zfs_arc_pool_dirty_percent of the anonymous dirty 6024 * data in the cache. 6025 * 6026 * Note: if two requests come in concurrently, we might let them 6027 * both succeed, when one of them should fail. Not a huge deal. 6028 */ 6029 uint64_t total_dirty = reserve + arc_tempreserve + anon_size; 6030 uint64_t spa_dirty_anon = spa_dirty_data(spa); 6031 6032 if (total_dirty > arc_c * zfs_arc_dirty_limit_percent / 100 && 6033 anon_size > arc_c * zfs_arc_anon_limit_percent / 100 && 6034 spa_dirty_anon > anon_size * zfs_arc_pool_dirty_percent / 100) { 6035 uint64_t meta_esize = 6036 zfs_refcount_count( 6037 &arc_anon->arcs_esize[ARC_BUFC_METADATA]); 6038 uint64_t data_esize = 6039 zfs_refcount_count(&arc_anon->arcs_esize[ARC_BUFC_DATA]); 6040 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK " 6041 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n", 6042 arc_tempreserve >> 10, meta_esize >> 10, 6043 data_esize >> 10, reserve >> 10, arc_c >> 10); 6044 return (SET_ERROR(ERESTART)); 6045 } 6046 atomic_add_64(&arc_tempreserve, reserve); 6047 return (0); 6048 } 6049 6050 static void 6051 arc_kstat_update_state(arc_state_t *state, kstat_named_t *size, 6052 kstat_named_t *evict_data, kstat_named_t *evict_metadata) 6053 { 6054 size->value.ui64 = zfs_refcount_count(&state->arcs_size); 6055 evict_data->value.ui64 = 6056 zfs_refcount_count(&state->arcs_esize[ARC_BUFC_DATA]); 6057 evict_metadata->value.ui64 = 6058 zfs_refcount_count(&state->arcs_esize[ARC_BUFC_METADATA]); 6059 } 6060 6061 static int 6062 arc_kstat_update(kstat_t *ksp, int rw) 6063 { 6064 arc_stats_t *as = ksp->ks_data; 6065 6066 if (rw == KSTAT_WRITE) { 6067 return (EACCES); 6068 } else { 6069 arc_kstat_update_state(arc_anon, 6070 &as->arcstat_anon_size, 6071 &as->arcstat_anon_evictable_data, 6072 &as->arcstat_anon_evictable_metadata); 6073 arc_kstat_update_state(arc_mru, 6074 &as->arcstat_mru_size, 6075 &as->arcstat_mru_evictable_data, 6076 &as->arcstat_mru_evictable_metadata); 6077 arc_kstat_update_state(arc_mru_ghost, 6078 &as->arcstat_mru_ghost_size, 6079 &as->arcstat_mru_ghost_evictable_data, 6080 &as->arcstat_mru_ghost_evictable_metadata); 6081 arc_kstat_update_state(arc_mfu, 6082 &as->arcstat_mfu_size, 6083 &as->arcstat_mfu_evictable_data, 6084 &as->arcstat_mfu_evictable_metadata); 6085 arc_kstat_update_state(arc_mfu_ghost, 6086 &as->arcstat_mfu_ghost_size, 6087 &as->arcstat_mfu_ghost_evictable_data, 6088 &as->arcstat_mfu_ghost_evictable_metadata); 6089 6090 ARCSTAT(arcstat_size) = aggsum_value(&arc_size); 6091 ARCSTAT(arcstat_meta_used) = aggsum_value(&arc_meta_used); 6092 ARCSTAT(arcstat_data_size) = aggsum_value(&astat_data_size); 6093 ARCSTAT(arcstat_metadata_size) = 6094 aggsum_value(&astat_metadata_size); 6095 ARCSTAT(arcstat_hdr_size) = aggsum_value(&astat_hdr_size); 6096 ARCSTAT(arcstat_other_size) = aggsum_value(&astat_other_size); 6097 ARCSTAT(arcstat_l2_hdr_size) = aggsum_value(&astat_l2_hdr_size); 6098 } 6099 6100 return (0); 6101 } 6102 6103 /* 6104 * This function *must* return indices evenly distributed between all 6105 * sublists of the multilist. This is needed due to how the ARC eviction 6106 * code is laid out; arc_evict_state() assumes ARC buffers are evenly 6107 * distributed between all sublists and uses this assumption when 6108 * deciding which sublist to evict from and how much to evict from it. 6109 */ 6110 unsigned int 6111 arc_state_multilist_index_func(multilist_t *ml, void *obj) 6112 { 6113 arc_buf_hdr_t *hdr = obj; 6114 6115 /* 6116 * We rely on b_dva to generate evenly distributed index 6117 * numbers using buf_hash below. So, as an added precaution, 6118 * let's make sure we never add empty buffers to the arc lists. 6119 */ 6120 ASSERT(!HDR_EMPTY(hdr)); 6121 6122 /* 6123 * The assumption here, is the hash value for a given 6124 * arc_buf_hdr_t will remain constant throughout it's lifetime 6125 * (i.e. it's b_spa, b_dva, and b_birth fields don't change). 6126 * Thus, we don't need to store the header's sublist index 6127 * on insertion, as this index can be recalculated on removal. 6128 * 6129 * Also, the low order bits of the hash value are thought to be 6130 * distributed evenly. Otherwise, in the case that the multilist 6131 * has a power of two number of sublists, each sublists' usage 6132 * would not be evenly distributed. 6133 */ 6134 return (buf_hash(hdr->b_spa, &hdr->b_dva, hdr->b_birth) % 6135 multilist_get_num_sublists(ml)); 6136 } 6137 6138 static void 6139 arc_state_init(void) 6140 { 6141 arc_anon = &ARC_anon; 6142 arc_mru = &ARC_mru; 6143 arc_mru_ghost = &ARC_mru_ghost; 6144 arc_mfu = &ARC_mfu; 6145 arc_mfu_ghost = &ARC_mfu_ghost; 6146 arc_l2c_only = &ARC_l2c_only; 6147 6148 arc_mru->arcs_list[ARC_BUFC_METADATA] = 6149 multilist_create(sizeof (arc_buf_hdr_t), 6150 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 6151 arc_state_multilist_index_func); 6152 arc_mru->arcs_list[ARC_BUFC_DATA] = 6153 multilist_create(sizeof (arc_buf_hdr_t), 6154 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 6155 arc_state_multilist_index_func); 6156 arc_mru_ghost->arcs_list[ARC_BUFC_METADATA] = 6157 multilist_create(sizeof (arc_buf_hdr_t), 6158 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 6159 arc_state_multilist_index_func); 6160 arc_mru_ghost->arcs_list[ARC_BUFC_DATA] = 6161 multilist_create(sizeof (arc_buf_hdr_t), 6162 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 6163 arc_state_multilist_index_func); 6164 arc_mfu->arcs_list[ARC_BUFC_METADATA] = 6165 multilist_create(sizeof (arc_buf_hdr_t), 6166 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 6167 arc_state_multilist_index_func); 6168 arc_mfu->arcs_list[ARC_BUFC_DATA] = 6169 multilist_create(sizeof (arc_buf_hdr_t), 6170 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 6171 arc_state_multilist_index_func); 6172 arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA] = 6173 multilist_create(sizeof (arc_buf_hdr_t), 6174 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 6175 arc_state_multilist_index_func); 6176 arc_mfu_ghost->arcs_list[ARC_BUFC_DATA] = 6177 multilist_create(sizeof (arc_buf_hdr_t), 6178 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 6179 arc_state_multilist_index_func); 6180 arc_l2c_only->arcs_list[ARC_BUFC_METADATA] = 6181 multilist_create(sizeof (arc_buf_hdr_t), 6182 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 6183 arc_state_multilist_index_func); 6184 arc_l2c_only->arcs_list[ARC_BUFC_DATA] = 6185 multilist_create(sizeof (arc_buf_hdr_t), 6186 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 6187 arc_state_multilist_index_func); 6188 6189 zfs_refcount_create(&arc_anon->arcs_esize[ARC_BUFC_METADATA]); 6190 zfs_refcount_create(&arc_anon->arcs_esize[ARC_BUFC_DATA]); 6191 zfs_refcount_create(&arc_mru->arcs_esize[ARC_BUFC_METADATA]); 6192 zfs_refcount_create(&arc_mru->arcs_esize[ARC_BUFC_DATA]); 6193 zfs_refcount_create(&arc_mru_ghost->arcs_esize[ARC_BUFC_METADATA]); 6194 zfs_refcount_create(&arc_mru_ghost->arcs_esize[ARC_BUFC_DATA]); 6195 zfs_refcount_create(&arc_mfu->arcs_esize[ARC_BUFC_METADATA]); 6196 zfs_refcount_create(&arc_mfu->arcs_esize[ARC_BUFC_DATA]); 6197 zfs_refcount_create(&arc_mfu_ghost->arcs_esize[ARC_BUFC_METADATA]); 6198 zfs_refcount_create(&arc_mfu_ghost->arcs_esize[ARC_BUFC_DATA]); 6199 zfs_refcount_create(&arc_l2c_only->arcs_esize[ARC_BUFC_METADATA]); 6200 zfs_refcount_create(&arc_l2c_only->arcs_esize[ARC_BUFC_DATA]); 6201 6202 zfs_refcount_create(&arc_anon->arcs_size); 6203 zfs_refcount_create(&arc_mru->arcs_size); 6204 zfs_refcount_create(&arc_mru_ghost->arcs_size); 6205 zfs_refcount_create(&arc_mfu->arcs_size); 6206 zfs_refcount_create(&arc_mfu_ghost->arcs_size); 6207 zfs_refcount_create(&arc_l2c_only->arcs_size); 6208 6209 aggsum_init(&arc_meta_used, 0); 6210 aggsum_init(&arc_size, 0); 6211 aggsum_init(&astat_data_size, 0); 6212 aggsum_init(&astat_metadata_size, 0); 6213 aggsum_init(&astat_hdr_size, 0); 6214 aggsum_init(&astat_other_size, 0); 6215 aggsum_init(&astat_l2_hdr_size, 0); 6216 } 6217 6218 static void 6219 arc_state_fini(void) 6220 { 6221 zfs_refcount_destroy(&arc_anon->arcs_esize[ARC_BUFC_METADATA]); 6222 zfs_refcount_destroy(&arc_anon->arcs_esize[ARC_BUFC_DATA]); 6223 zfs_refcount_destroy(&arc_mru->arcs_esize[ARC_BUFC_METADATA]); 6224 zfs_refcount_destroy(&arc_mru->arcs_esize[ARC_BUFC_DATA]); 6225 zfs_refcount_destroy(&arc_mru_ghost->arcs_esize[ARC_BUFC_METADATA]); 6226 zfs_refcount_destroy(&arc_mru_ghost->arcs_esize[ARC_BUFC_DATA]); 6227 zfs_refcount_destroy(&arc_mfu->arcs_esize[ARC_BUFC_METADATA]); 6228 zfs_refcount_destroy(&arc_mfu->arcs_esize[ARC_BUFC_DATA]); 6229 zfs_refcount_destroy(&arc_mfu_ghost->arcs_esize[ARC_BUFC_METADATA]); 6230 zfs_refcount_destroy(&arc_mfu_ghost->arcs_esize[ARC_BUFC_DATA]); 6231 zfs_refcount_destroy(&arc_l2c_only->arcs_esize[ARC_BUFC_METADATA]); 6232 zfs_refcount_destroy(&arc_l2c_only->arcs_esize[ARC_BUFC_DATA]); 6233 6234 zfs_refcount_destroy(&arc_anon->arcs_size); 6235 zfs_refcount_destroy(&arc_mru->arcs_size); 6236 zfs_refcount_destroy(&arc_mru_ghost->arcs_size); 6237 zfs_refcount_destroy(&arc_mfu->arcs_size); 6238 zfs_refcount_destroy(&arc_mfu_ghost->arcs_size); 6239 zfs_refcount_destroy(&arc_l2c_only->arcs_size); 6240 6241 multilist_destroy(arc_mru->arcs_list[ARC_BUFC_METADATA]); 6242 multilist_destroy(arc_mru_ghost->arcs_list[ARC_BUFC_METADATA]); 6243 multilist_destroy(arc_mfu->arcs_list[ARC_BUFC_METADATA]); 6244 multilist_destroy(arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA]); 6245 multilist_destroy(arc_mru->arcs_list[ARC_BUFC_DATA]); 6246 multilist_destroy(arc_mru_ghost->arcs_list[ARC_BUFC_DATA]); 6247 multilist_destroy(arc_mfu->arcs_list[ARC_BUFC_DATA]); 6248 multilist_destroy(arc_mfu_ghost->arcs_list[ARC_BUFC_DATA]); 6249 6250 aggsum_fini(&arc_meta_used); 6251 aggsum_fini(&arc_size); 6252 aggsum_fini(&astat_data_size); 6253 aggsum_fini(&astat_metadata_size); 6254 aggsum_fini(&astat_hdr_size); 6255 aggsum_fini(&astat_other_size); 6256 aggsum_fini(&astat_l2_hdr_size); 6257 } 6258 6259 uint64_t 6260 arc_max_bytes(void) 6261 { 6262 return (arc_c_max); 6263 } 6264 6265 void 6266 arc_init(void) 6267 { 6268 /* 6269 * allmem is "all memory that we could possibly use". 6270 */ 6271 #ifdef _KERNEL 6272 uint64_t allmem = ptob(physmem - swapfs_minfree); 6273 #else 6274 uint64_t allmem = (physmem * PAGESIZE) / 2; 6275 #endif 6276 mutex_init(&arc_adjust_lock, NULL, MUTEX_DEFAULT, NULL); 6277 cv_init(&arc_adjust_waiters_cv, NULL, CV_DEFAULT, NULL); 6278 6279 /* set min cache to 1/32 of all memory, or 64MB, whichever is more */ 6280 arc_c_min = MAX(allmem / 32, 64 << 20); 6281 /* set max to 3/4 of all memory, or all but 1GB, whichever is more */ 6282 if (allmem >= 1 << 30) 6283 arc_c_max = allmem - (1 << 30); 6284 else 6285 arc_c_max = arc_c_min; 6286 arc_c_max = MAX(allmem * 3 / 4, arc_c_max); 6287 6288 /* 6289 * In userland, there's only the memory pressure that we artificially 6290 * create (see arc_available_memory()). Don't let arc_c get too 6291 * small, because it can cause transactions to be larger than 6292 * arc_c, causing arc_tempreserve_space() to fail. 6293 */ 6294 #ifndef _KERNEL 6295 arc_c_min = arc_c_max / 2; 6296 #endif 6297 6298 /* 6299 * Allow the tunables to override our calculations if they are 6300 * reasonable (ie. over 64MB) 6301 */ 6302 if (zfs_arc_max > 64 << 20 && zfs_arc_max < allmem) { 6303 arc_c_max = zfs_arc_max; 6304 arc_c_min = MIN(arc_c_min, arc_c_max); 6305 } 6306 if (zfs_arc_min > 64 << 20 && zfs_arc_min <= arc_c_max) 6307 arc_c_min = zfs_arc_min; 6308 6309 arc_c = arc_c_max; 6310 arc_p = (arc_c >> 1); 6311 6312 /* limit meta-data to 1/4 of the arc capacity */ 6313 arc_meta_limit = arc_c_max / 4; 6314 6315 #ifdef _KERNEL 6316 /* 6317 * Metadata is stored in the kernel's heap. Don't let us 6318 * use more than half the heap for the ARC. 6319 */ 6320 arc_meta_limit = MIN(arc_meta_limit, 6321 vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 2); 6322 #endif 6323 6324 /* Allow the tunable to override if it is reasonable */ 6325 if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max) 6326 arc_meta_limit = zfs_arc_meta_limit; 6327 6328 if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0) 6329 arc_c_min = arc_meta_limit / 2; 6330 6331 if (zfs_arc_meta_min > 0) { 6332 arc_meta_min = zfs_arc_meta_min; 6333 } else { 6334 arc_meta_min = arc_c_min / 2; 6335 } 6336 6337 if (zfs_arc_grow_retry > 0) 6338 arc_grow_retry = zfs_arc_grow_retry; 6339 6340 if (zfs_arc_shrink_shift > 0) 6341 arc_shrink_shift = zfs_arc_shrink_shift; 6342 6343 /* 6344 * Ensure that arc_no_grow_shift is less than arc_shrink_shift. 6345 */ 6346 if (arc_no_grow_shift >= arc_shrink_shift) 6347 arc_no_grow_shift = arc_shrink_shift - 1; 6348 6349 if (zfs_arc_p_min_shift > 0) 6350 arc_p_min_shift = zfs_arc_p_min_shift; 6351 6352 /* if kmem_flags are set, lets try to use less memory */ 6353 if (kmem_debugging()) 6354 arc_c = arc_c / 2; 6355 if (arc_c < arc_c_min) 6356 arc_c = arc_c_min; 6357 6358 arc_state_init(); 6359 6360 /* 6361 * The arc must be "uninitialized", so that hdr_recl() (which is 6362 * registered by buf_init()) will not access arc_reap_zthr before 6363 * it is created. 6364 */ 6365 ASSERT(!arc_initialized); 6366 buf_init(); 6367 6368 arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED, 6369 sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL); 6370 6371 if (arc_ksp != NULL) { 6372 arc_ksp->ks_data = &arc_stats; 6373 arc_ksp->ks_update = arc_kstat_update; 6374 kstat_install(arc_ksp); 6375 } 6376 6377 arc_adjust_zthr = zthr_create(arc_adjust_cb_check, 6378 arc_adjust_cb, NULL); 6379 arc_reap_zthr = zthr_create_timer(arc_reap_cb_check, 6380 arc_reap_cb, NULL, SEC2NSEC(1)); 6381 6382 arc_initialized = B_TRUE; 6383 arc_warm = B_FALSE; 6384 6385 /* 6386 * Calculate maximum amount of dirty data per pool. 6387 * 6388 * If it has been set by /etc/system, take that. 6389 * Otherwise, use a percentage of physical memory defined by 6390 * zfs_dirty_data_max_percent (default 10%) with a cap at 6391 * zfs_dirty_data_max_max (default 4GB). 6392 */ 6393 if (zfs_dirty_data_max == 0) { 6394 zfs_dirty_data_max = physmem * PAGESIZE * 6395 zfs_dirty_data_max_percent / 100; 6396 zfs_dirty_data_max = MIN(zfs_dirty_data_max, 6397 zfs_dirty_data_max_max); 6398 } 6399 } 6400 6401 void 6402 arc_fini(void) 6403 { 6404 /* Use B_TRUE to ensure *all* buffers are evicted */ 6405 arc_flush(NULL, B_TRUE); 6406 6407 arc_initialized = B_FALSE; 6408 6409 if (arc_ksp != NULL) { 6410 kstat_delete(arc_ksp); 6411 arc_ksp = NULL; 6412 } 6413 6414 (void) zthr_cancel(arc_adjust_zthr); 6415 zthr_destroy(arc_adjust_zthr); 6416 6417 (void) zthr_cancel(arc_reap_zthr); 6418 zthr_destroy(arc_reap_zthr); 6419 6420 mutex_destroy(&arc_adjust_lock); 6421 cv_destroy(&arc_adjust_waiters_cv); 6422 6423 /* 6424 * buf_fini() must proceed arc_state_fini() because buf_fin() may 6425 * trigger the release of kmem magazines, which can callback to 6426 * arc_space_return() which accesses aggsums freed in act_state_fini(). 6427 */ 6428 buf_fini(); 6429 arc_state_fini(); 6430 6431 ASSERT0(arc_loaned_bytes); 6432 } 6433 6434 /* 6435 * Level 2 ARC 6436 * 6437 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk. 6438 * It uses dedicated storage devices to hold cached data, which are populated 6439 * using large infrequent writes. The main role of this cache is to boost 6440 * the performance of random read workloads. The intended L2ARC devices 6441 * include short-stroked disks, solid state disks, and other media with 6442 * substantially faster read latency than disk. 6443 * 6444 * +-----------------------+ 6445 * | ARC | 6446 * +-----------------------+ 6447 * | ^ ^ 6448 * | | | 6449 * l2arc_feed_thread() arc_read() 6450 * | | | 6451 * | l2arc read | 6452 * V | | 6453 * +---------------+ | 6454 * | L2ARC | | 6455 * +---------------+ | 6456 * | ^ | 6457 * l2arc_write() | | 6458 * | | | 6459 * V | | 6460 * +-------+ +-------+ 6461 * | vdev | | vdev | 6462 * | cache | | cache | 6463 * +-------+ +-------+ 6464 * +=========+ .-----. 6465 * : L2ARC : |-_____-| 6466 * : devices : | Disks | 6467 * +=========+ `-_____-' 6468 * 6469 * Read requests are satisfied from the following sources, in order: 6470 * 6471 * 1) ARC 6472 * 2) vdev cache of L2ARC devices 6473 * 3) L2ARC devices 6474 * 4) vdev cache of disks 6475 * 5) disks 6476 * 6477 * Some L2ARC device types exhibit extremely slow write performance. 6478 * To accommodate for this there are some significant differences between 6479 * the L2ARC and traditional cache design: 6480 * 6481 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from 6482 * the ARC behave as usual, freeing buffers and placing headers on ghost 6483 * lists. The ARC does not send buffers to the L2ARC during eviction as 6484 * this would add inflated write latencies for all ARC memory pressure. 6485 * 6486 * 2. The L2ARC attempts to cache data from the ARC before it is evicted. 6487 * It does this by periodically scanning buffers from the eviction-end of 6488 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are 6489 * not already there. It scans until a headroom of buffers is satisfied, 6490 * which itself is a buffer for ARC eviction. If a compressible buffer is 6491 * found during scanning and selected for writing to an L2ARC device, we 6492 * temporarily boost scanning headroom during the next scan cycle to make 6493 * sure we adapt to compression effects (which might significantly reduce 6494 * the data volume we write to L2ARC). The thread that does this is 6495 * l2arc_feed_thread(), illustrated below; example sizes are included to 6496 * provide a better sense of ratio than this diagram: 6497 * 6498 * head --> tail 6499 * +---------------------+----------+ 6500 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC 6501 * +---------------------+----------+ | o L2ARC eligible 6502 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer 6503 * +---------------------+----------+ | 6504 * 15.9 Gbytes ^ 32 Mbytes | 6505 * headroom | 6506 * l2arc_feed_thread() 6507 * | 6508 * l2arc write hand <--[oooo]--' 6509 * | 8 Mbyte 6510 * | write max 6511 * V 6512 * +==============================+ 6513 * L2ARC dev |####|#|###|###| |####| ... | 6514 * +==============================+ 6515 * 32 Gbytes 6516 * 6517 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of 6518 * evicted, then the L2ARC has cached a buffer much sooner than it probably 6519 * needed to, potentially wasting L2ARC device bandwidth and storage. It is 6520 * safe to say that this is an uncommon case, since buffers at the end of 6521 * the ARC lists have moved there due to inactivity. 6522 * 6523 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom, 6524 * then the L2ARC simply misses copying some buffers. This serves as a 6525 * pressure valve to prevent heavy read workloads from both stalling the ARC 6526 * with waits and clogging the L2ARC with writes. This also helps prevent 6527 * the potential for the L2ARC to churn if it attempts to cache content too 6528 * quickly, such as during backups of the entire pool. 6529 * 6530 * 5. After system boot and before the ARC has filled main memory, there are 6531 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru 6532 * lists can remain mostly static. Instead of searching from tail of these 6533 * lists as pictured, the l2arc_feed_thread() will search from the list heads 6534 * for eligible buffers, greatly increasing its chance of finding them. 6535 * 6536 * The L2ARC device write speed is also boosted during this time so that 6537 * the L2ARC warms up faster. Since there have been no ARC evictions yet, 6538 * there are no L2ARC reads, and no fear of degrading read performance 6539 * through increased writes. 6540 * 6541 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that 6542 * the vdev queue can aggregate them into larger and fewer writes. Each 6543 * device is written to in a rotor fashion, sweeping writes through 6544 * available space then repeating. 6545 * 6546 * 7. The L2ARC does not store dirty content. It never needs to flush 6547 * write buffers back to disk based storage. 6548 * 6549 * 8. If an ARC buffer is written (and dirtied) which also exists in the 6550 * L2ARC, the now stale L2ARC buffer is immediately dropped. 6551 * 6552 * The performance of the L2ARC can be tweaked by a number of tunables, which 6553 * may be necessary for different workloads: 6554 * 6555 * l2arc_write_max max write bytes per interval 6556 * l2arc_write_boost extra write bytes during device warmup 6557 * l2arc_noprefetch skip caching prefetched buffers 6558 * l2arc_headroom number of max device writes to precache 6559 * l2arc_headroom_boost when we find compressed buffers during ARC 6560 * scanning, we multiply headroom by this 6561 * percentage factor for the next scan cycle, 6562 * since more compressed buffers are likely to 6563 * be present 6564 * l2arc_feed_secs seconds between L2ARC writing 6565 * 6566 * Tunables may be removed or added as future performance improvements are 6567 * integrated, and also may become zpool properties. 6568 * 6569 * There are three key functions that control how the L2ARC warms up: 6570 * 6571 * l2arc_write_eligible() check if a buffer is eligible to cache 6572 * l2arc_write_size() calculate how much to write 6573 * l2arc_write_interval() calculate sleep delay between writes 6574 * 6575 * These three functions determine what to write, how much, and how quickly 6576 * to send writes. 6577 */ 6578 6579 static boolean_t 6580 l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *hdr) 6581 { 6582 /* 6583 * A buffer is *not* eligible for the L2ARC if it: 6584 * 1. belongs to a different spa. 6585 * 2. is already cached on the L2ARC. 6586 * 3. has an I/O in progress (it may be an incomplete read). 6587 * 4. is flagged not eligible (zfs property). 6588 */ 6589 if (hdr->b_spa != spa_guid || HDR_HAS_L2HDR(hdr) || 6590 HDR_IO_IN_PROGRESS(hdr) || !HDR_L2CACHE(hdr)) 6591 return (B_FALSE); 6592 6593 return (B_TRUE); 6594 } 6595 6596 static uint64_t 6597 l2arc_write_size(void) 6598 { 6599 uint64_t size; 6600 6601 /* 6602 * Make sure our globals have meaningful values in case the user 6603 * altered them. 6604 */ 6605 size = l2arc_write_max; 6606 if (size == 0) { 6607 cmn_err(CE_NOTE, "Bad value for l2arc_write_max, value must " 6608 "be greater than zero, resetting it to the default (%d)", 6609 L2ARC_WRITE_SIZE); 6610 size = l2arc_write_max = L2ARC_WRITE_SIZE; 6611 } 6612 6613 if (arc_warm == B_FALSE) 6614 size += l2arc_write_boost; 6615 6616 return (size); 6617 6618 } 6619 6620 static clock_t 6621 l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote) 6622 { 6623 clock_t interval, next, now; 6624 6625 /* 6626 * If the ARC lists are busy, increase our write rate; if the 6627 * lists are stale, idle back. This is achieved by checking 6628 * how much we previously wrote - if it was more than half of 6629 * what we wanted, schedule the next write much sooner. 6630 */ 6631 if (l2arc_feed_again && wrote > (wanted / 2)) 6632 interval = (hz * l2arc_feed_min_ms) / 1000; 6633 else 6634 interval = hz * l2arc_feed_secs; 6635 6636 now = ddi_get_lbolt(); 6637 next = MAX(now, MIN(now + interval, began + interval)); 6638 6639 return (next); 6640 } 6641 6642 /* 6643 * Cycle through L2ARC devices. This is how L2ARC load balances. 6644 * If a device is returned, this also returns holding the spa config lock. 6645 */ 6646 static l2arc_dev_t * 6647 l2arc_dev_get_next(void) 6648 { 6649 l2arc_dev_t *first, *next = NULL; 6650 6651 /* 6652 * Lock out the removal of spas (spa_namespace_lock), then removal 6653 * of cache devices (l2arc_dev_mtx). Once a device has been selected, 6654 * both locks will be dropped and a spa config lock held instead. 6655 */ 6656 mutex_enter(&spa_namespace_lock); 6657 mutex_enter(&l2arc_dev_mtx); 6658 6659 /* if there are no vdevs, there is nothing to do */ 6660 if (l2arc_ndev == 0) 6661 goto out; 6662 6663 first = NULL; 6664 next = l2arc_dev_last; 6665 do { 6666 /* loop around the list looking for a non-faulted vdev */ 6667 if (next == NULL) { 6668 next = list_head(l2arc_dev_list); 6669 } else { 6670 next = list_next(l2arc_dev_list, next); 6671 if (next == NULL) 6672 next = list_head(l2arc_dev_list); 6673 } 6674 6675 /* if we have come back to the start, bail out */ 6676 if (first == NULL) 6677 first = next; 6678 else if (next == first) 6679 break; 6680 6681 } while (vdev_is_dead(next->l2ad_vdev)); 6682 6683 /* if we were unable to find any usable vdevs, return NULL */ 6684 if (vdev_is_dead(next->l2ad_vdev)) 6685 next = NULL; 6686 6687 l2arc_dev_last = next; 6688 6689 out: 6690 mutex_exit(&l2arc_dev_mtx); 6691 6692 /* 6693 * Grab the config lock to prevent the 'next' device from being 6694 * removed while we are writing to it. 6695 */ 6696 if (next != NULL) 6697 spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER); 6698 mutex_exit(&spa_namespace_lock); 6699 6700 return (next); 6701 } 6702 6703 /* 6704 * Free buffers that were tagged for destruction. 6705 */ 6706 static void 6707 l2arc_do_free_on_write() 6708 { 6709 list_t *buflist; 6710 l2arc_data_free_t *df, *df_prev; 6711 6712 mutex_enter(&l2arc_free_on_write_mtx); 6713 buflist = l2arc_free_on_write; 6714 6715 for (df = list_tail(buflist); df; df = df_prev) { 6716 df_prev = list_prev(buflist, df); 6717 ASSERT3P(df->l2df_abd, !=, NULL); 6718 abd_free(df->l2df_abd); 6719 list_remove(buflist, df); 6720 kmem_free(df, sizeof (l2arc_data_free_t)); 6721 } 6722 6723 mutex_exit(&l2arc_free_on_write_mtx); 6724 } 6725 6726 /* 6727 * A write to a cache device has completed. Update all headers to allow 6728 * reads from these buffers to begin. 6729 */ 6730 static void 6731 l2arc_write_done(zio_t *zio) 6732 { 6733 l2arc_write_callback_t *cb; 6734 l2arc_dev_t *dev; 6735 list_t *buflist; 6736 arc_buf_hdr_t *head, *hdr, *hdr_prev; 6737 kmutex_t *hash_lock; 6738 int64_t bytes_dropped = 0; 6739 6740 cb = zio->io_private; 6741 ASSERT3P(cb, !=, NULL); 6742 dev = cb->l2wcb_dev; 6743 ASSERT3P(dev, !=, NULL); 6744 head = cb->l2wcb_head; 6745 ASSERT3P(head, !=, NULL); 6746 buflist = &dev->l2ad_buflist; 6747 ASSERT3P(buflist, !=, NULL); 6748 DTRACE_PROBE2(l2arc__iodone, zio_t *, zio, 6749 l2arc_write_callback_t *, cb); 6750 6751 if (zio->io_error != 0) 6752 ARCSTAT_BUMP(arcstat_l2_writes_error); 6753 6754 /* 6755 * All writes completed, or an error was hit. 6756 */ 6757 top: 6758 mutex_enter(&dev->l2ad_mtx); 6759 for (hdr = list_prev(buflist, head); hdr; hdr = hdr_prev) { 6760 hdr_prev = list_prev(buflist, hdr); 6761 6762 hash_lock = HDR_LOCK(hdr); 6763 6764 /* 6765 * We cannot use mutex_enter or else we can deadlock 6766 * with l2arc_write_buffers (due to swapping the order 6767 * the hash lock and l2ad_mtx are taken). 6768 */ 6769 if (!mutex_tryenter(hash_lock)) { 6770 /* 6771 * Missed the hash lock. We must retry so we 6772 * don't leave the ARC_FLAG_L2_WRITING bit set. 6773 */ 6774 ARCSTAT_BUMP(arcstat_l2_writes_lock_retry); 6775 6776 /* 6777 * We don't want to rescan the headers we've 6778 * already marked as having been written out, so 6779 * we reinsert the head node so we can pick up 6780 * where we left off. 6781 */ 6782 list_remove(buflist, head); 6783 list_insert_after(buflist, hdr, head); 6784 6785 mutex_exit(&dev->l2ad_mtx); 6786 6787 /* 6788 * We wait for the hash lock to become available 6789 * to try and prevent busy waiting, and increase 6790 * the chance we'll be able to acquire the lock 6791 * the next time around. 6792 */ 6793 mutex_enter(hash_lock); 6794 mutex_exit(hash_lock); 6795 goto top; 6796 } 6797 6798 /* 6799 * We could not have been moved into the arc_l2c_only 6800 * state while in-flight due to our ARC_FLAG_L2_WRITING 6801 * bit being set. Let's just ensure that's being enforced. 6802 */ 6803 ASSERT(HDR_HAS_L1HDR(hdr)); 6804 6805 if (zio->io_error != 0) { 6806 /* 6807 * Error - drop L2ARC entry. 6808 */ 6809 list_remove(buflist, hdr); 6810 arc_hdr_clear_flags(hdr, ARC_FLAG_HAS_L2HDR); 6811 6812 ARCSTAT_INCR(arcstat_l2_psize, -arc_hdr_size(hdr)); 6813 ARCSTAT_INCR(arcstat_l2_lsize, -HDR_GET_LSIZE(hdr)); 6814 6815 bytes_dropped += arc_hdr_size(hdr); 6816 (void) zfs_refcount_remove_many(&dev->l2ad_alloc, 6817 arc_hdr_size(hdr), hdr); 6818 } 6819 6820 /* 6821 * Allow ARC to begin reads and ghost list evictions to 6822 * this L2ARC entry. 6823 */ 6824 arc_hdr_clear_flags(hdr, ARC_FLAG_L2_WRITING); 6825 6826 mutex_exit(hash_lock); 6827 } 6828 6829 atomic_inc_64(&l2arc_writes_done); 6830 list_remove(buflist, head); 6831 ASSERT(!HDR_HAS_L1HDR(head)); 6832 kmem_cache_free(hdr_l2only_cache, head); 6833 mutex_exit(&dev->l2ad_mtx); 6834 6835 vdev_space_update(dev->l2ad_vdev, -bytes_dropped, 0, 0); 6836 6837 l2arc_do_free_on_write(); 6838 6839 kmem_free(cb, sizeof (l2arc_write_callback_t)); 6840 } 6841 6842 /* 6843 * A read to a cache device completed. Validate buffer contents before 6844 * handing over to the regular ARC routines. 6845 */ 6846 static void 6847 l2arc_read_done(zio_t *zio) 6848 { 6849 l2arc_read_callback_t *cb; 6850 arc_buf_hdr_t *hdr; 6851 kmutex_t *hash_lock; 6852 boolean_t valid_cksum; 6853 6854 ASSERT3P(zio->io_vd, !=, NULL); 6855 ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE); 6856 6857 spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd); 6858 6859 cb = zio->io_private; 6860 ASSERT3P(cb, !=, NULL); 6861 hdr = cb->l2rcb_hdr; 6862 ASSERT3P(hdr, !=, NULL); 6863 6864 hash_lock = HDR_LOCK(hdr); 6865 mutex_enter(hash_lock); 6866 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 6867 6868 /* 6869 * If the data was read into a temporary buffer, 6870 * move it and free the buffer. 6871 */ 6872 if (cb->l2rcb_abd != NULL) { 6873 ASSERT3U(arc_hdr_size(hdr), <, zio->io_size); 6874 if (zio->io_error == 0) { 6875 abd_copy(hdr->b_l1hdr.b_pabd, cb->l2rcb_abd, 6876 arc_hdr_size(hdr)); 6877 } 6878 6879 /* 6880 * The following must be done regardless of whether 6881 * there was an error: 6882 * - free the temporary buffer 6883 * - point zio to the real ARC buffer 6884 * - set zio size accordingly 6885 * These are required because zio is either re-used for 6886 * an I/O of the block in the case of the error 6887 * or the zio is passed to arc_read_done() and it 6888 * needs real data. 6889 */ 6890 abd_free(cb->l2rcb_abd); 6891 zio->io_size = zio->io_orig_size = arc_hdr_size(hdr); 6892 zio->io_abd = zio->io_orig_abd = hdr->b_l1hdr.b_pabd; 6893 } 6894 6895 ASSERT3P(zio->io_abd, !=, NULL); 6896 6897 /* 6898 * Check this survived the L2ARC journey. 6899 */ 6900 ASSERT3P(zio->io_abd, ==, hdr->b_l1hdr.b_pabd); 6901 zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */ 6902 zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */ 6903 6904 valid_cksum = arc_cksum_is_equal(hdr, zio); 6905 if (valid_cksum && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) { 6906 mutex_exit(hash_lock); 6907 zio->io_private = hdr; 6908 arc_read_done(zio); 6909 } else { 6910 mutex_exit(hash_lock); 6911 /* 6912 * Buffer didn't survive caching. Increment stats and 6913 * reissue to the original storage device. 6914 */ 6915 if (zio->io_error != 0) { 6916 ARCSTAT_BUMP(arcstat_l2_io_error); 6917 } else { 6918 zio->io_error = SET_ERROR(EIO); 6919 } 6920 if (!valid_cksum) 6921 ARCSTAT_BUMP(arcstat_l2_cksum_bad); 6922 6923 /* 6924 * If there's no waiter, issue an async i/o to the primary 6925 * storage now. If there *is* a waiter, the caller must 6926 * issue the i/o in a context where it's OK to block. 6927 */ 6928 if (zio->io_waiter == NULL) { 6929 zio_t *pio = zio_unique_parent(zio); 6930 6931 ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL); 6932 6933 zio_nowait(zio_read(pio, zio->io_spa, zio->io_bp, 6934 hdr->b_l1hdr.b_pabd, zio->io_size, arc_read_done, 6935 hdr, zio->io_priority, cb->l2rcb_flags, 6936 &cb->l2rcb_zb)); 6937 } 6938 } 6939 6940 kmem_free(cb, sizeof (l2arc_read_callback_t)); 6941 } 6942 6943 /* 6944 * This is the list priority from which the L2ARC will search for pages to 6945 * cache. This is used within loops (0..3) to cycle through lists in the 6946 * desired order. This order can have a significant effect on cache 6947 * performance. 6948 * 6949 * Currently the metadata lists are hit first, MFU then MRU, followed by 6950 * the data lists. This function returns a locked list, and also returns 6951 * the lock pointer. 6952 */ 6953 static multilist_sublist_t * 6954 l2arc_sublist_lock(int list_num) 6955 { 6956 multilist_t *ml = NULL; 6957 unsigned int idx; 6958 6959 ASSERT(list_num >= 0 && list_num <= 3); 6960 6961 switch (list_num) { 6962 case 0: 6963 ml = arc_mfu->arcs_list[ARC_BUFC_METADATA]; 6964 break; 6965 case 1: 6966 ml = arc_mru->arcs_list[ARC_BUFC_METADATA]; 6967 break; 6968 case 2: 6969 ml = arc_mfu->arcs_list[ARC_BUFC_DATA]; 6970 break; 6971 case 3: 6972 ml = arc_mru->arcs_list[ARC_BUFC_DATA]; 6973 break; 6974 } 6975 6976 /* 6977 * Return a randomly-selected sublist. This is acceptable 6978 * because the caller feeds only a little bit of data for each 6979 * call (8MB). Subsequent calls will result in different 6980 * sublists being selected. 6981 */ 6982 idx = multilist_get_random_index(ml); 6983 return (multilist_sublist_lock(ml, idx)); 6984 } 6985 6986 /* 6987 * Evict buffers from the device write hand to the distance specified in 6988 * bytes. This distance may span populated buffers, it may span nothing. 6989 * This is clearing a region on the L2ARC device ready for writing. 6990 * If the 'all' boolean is set, every buffer is evicted. 6991 */ 6992 static void 6993 l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all) 6994 { 6995 list_t *buflist; 6996 arc_buf_hdr_t *hdr, *hdr_prev; 6997 kmutex_t *hash_lock; 6998 uint64_t taddr; 6999 7000 buflist = &dev->l2ad_buflist; 7001 7002 if (!all && dev->l2ad_first) { 7003 /* 7004 * This is the first sweep through the device. There is 7005 * nothing to evict. 7006 */ 7007 return; 7008 } 7009 7010 if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) { 7011 /* 7012 * When nearing the end of the device, evict to the end 7013 * before the device write hand jumps to the start. 7014 */ 7015 taddr = dev->l2ad_end; 7016 } else { 7017 taddr = dev->l2ad_hand + distance; 7018 } 7019 DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist, 7020 uint64_t, taddr, boolean_t, all); 7021 7022 top: 7023 mutex_enter(&dev->l2ad_mtx); 7024 for (hdr = list_tail(buflist); hdr; hdr = hdr_prev) { 7025 hdr_prev = list_prev(buflist, hdr); 7026 7027 hash_lock = HDR_LOCK(hdr); 7028 7029 /* 7030 * We cannot use mutex_enter or else we can deadlock 7031 * with l2arc_write_buffers (due to swapping the order 7032 * the hash lock and l2ad_mtx are taken). 7033 */ 7034 if (!mutex_tryenter(hash_lock)) { 7035 /* 7036 * Missed the hash lock. Retry. 7037 */ 7038 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry); 7039 mutex_exit(&dev->l2ad_mtx); 7040 mutex_enter(hash_lock); 7041 mutex_exit(hash_lock); 7042 goto top; 7043 } 7044 7045 /* 7046 * A header can't be on this list if it doesn't have L2 header. 7047 */ 7048 ASSERT(HDR_HAS_L2HDR(hdr)); 7049 7050 /* Ensure this header has finished being written. */ 7051 ASSERT(!HDR_L2_WRITING(hdr)); 7052 ASSERT(!HDR_L2_WRITE_HEAD(hdr)); 7053 7054 if (!all && (hdr->b_l2hdr.b_daddr >= taddr || 7055 hdr->b_l2hdr.b_daddr < dev->l2ad_hand)) { 7056 /* 7057 * We've evicted to the target address, 7058 * or the end of the device. 7059 */ 7060 mutex_exit(hash_lock); 7061 break; 7062 } 7063 7064 if (!HDR_HAS_L1HDR(hdr)) { 7065 ASSERT(!HDR_L2_READING(hdr)); 7066 /* 7067 * This doesn't exist in the ARC. Destroy. 7068 * arc_hdr_destroy() will call list_remove() 7069 * and decrement arcstat_l2_lsize. 7070 */ 7071 arc_change_state(arc_anon, hdr, hash_lock); 7072 arc_hdr_destroy(hdr); 7073 } else { 7074 ASSERT(hdr->b_l1hdr.b_state != arc_l2c_only); 7075 ARCSTAT_BUMP(arcstat_l2_evict_l1cached); 7076 /* 7077 * Invalidate issued or about to be issued 7078 * reads, since we may be about to write 7079 * over this location. 7080 */ 7081 if (HDR_L2_READING(hdr)) { 7082 ARCSTAT_BUMP(arcstat_l2_evict_reading); 7083 arc_hdr_set_flags(hdr, ARC_FLAG_L2_EVICTED); 7084 } 7085 7086 arc_hdr_l2hdr_destroy(hdr); 7087 } 7088 mutex_exit(hash_lock); 7089 } 7090 mutex_exit(&dev->l2ad_mtx); 7091 } 7092 7093 /* 7094 * Find and write ARC buffers to the L2ARC device. 7095 * 7096 * An ARC_FLAG_L2_WRITING flag is set so that the L2ARC buffers are not valid 7097 * for reading until they have completed writing. 7098 * The headroom_boost is an in-out parameter used to maintain headroom boost 7099 * state between calls to this function. 7100 * 7101 * Returns the number of bytes actually written (which may be smaller than 7102 * the delta by which the device hand has changed due to alignment). 7103 */ 7104 static uint64_t 7105 l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz) 7106 { 7107 arc_buf_hdr_t *hdr, *hdr_prev, *head; 7108 uint64_t write_asize, write_psize, write_lsize, headroom; 7109 boolean_t full; 7110 l2arc_write_callback_t *cb; 7111 zio_t *pio, *wzio; 7112 uint64_t guid = spa_load_guid(spa); 7113 7114 ASSERT3P(dev->l2ad_vdev, !=, NULL); 7115 7116 pio = NULL; 7117 write_lsize = write_asize = write_psize = 0; 7118 full = B_FALSE; 7119 head = kmem_cache_alloc(hdr_l2only_cache, KM_PUSHPAGE); 7120 arc_hdr_set_flags(head, ARC_FLAG_L2_WRITE_HEAD | ARC_FLAG_HAS_L2HDR); 7121 7122 /* 7123 * Copy buffers for L2ARC writing. 7124 */ 7125 for (int try = 0; try <= 3; try++) { 7126 multilist_sublist_t *mls = l2arc_sublist_lock(try); 7127 uint64_t passed_sz = 0; 7128 7129 /* 7130 * L2ARC fast warmup. 7131 * 7132 * Until the ARC is warm and starts to evict, read from the 7133 * head of the ARC lists rather than the tail. 7134 */ 7135 if (arc_warm == B_FALSE) 7136 hdr = multilist_sublist_head(mls); 7137 else 7138 hdr = multilist_sublist_tail(mls); 7139 7140 headroom = target_sz * l2arc_headroom; 7141 if (zfs_compressed_arc_enabled) 7142 headroom = (headroom * l2arc_headroom_boost) / 100; 7143 7144 for (; hdr; hdr = hdr_prev) { 7145 kmutex_t *hash_lock; 7146 7147 if (arc_warm == B_FALSE) 7148 hdr_prev = multilist_sublist_next(mls, hdr); 7149 else 7150 hdr_prev = multilist_sublist_prev(mls, hdr); 7151 7152 hash_lock = HDR_LOCK(hdr); 7153 if (!mutex_tryenter(hash_lock)) { 7154 /* 7155 * Skip this buffer rather than waiting. 7156 */ 7157 continue; 7158 } 7159 7160 passed_sz += HDR_GET_LSIZE(hdr); 7161 if (passed_sz > headroom) { 7162 /* 7163 * Searched too far. 7164 */ 7165 mutex_exit(hash_lock); 7166 break; 7167 } 7168 7169 if (!l2arc_write_eligible(guid, hdr)) { 7170 mutex_exit(hash_lock); 7171 continue; 7172 } 7173 7174 /* 7175 * We rely on the L1 portion of the header below, so 7176 * it's invalid for this header to have been evicted out 7177 * of the ghost cache, prior to being written out. The 7178 * ARC_FLAG_L2_WRITING bit ensures this won't happen. 7179 */ 7180 ASSERT(HDR_HAS_L1HDR(hdr)); 7181 7182 ASSERT3U(HDR_GET_PSIZE(hdr), >, 0); 7183 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL); 7184 ASSERT3U(arc_hdr_size(hdr), >, 0); 7185 uint64_t psize = arc_hdr_size(hdr); 7186 uint64_t asize = vdev_psize_to_asize(dev->l2ad_vdev, 7187 psize); 7188 7189 if ((write_asize + asize) > target_sz) { 7190 full = B_TRUE; 7191 mutex_exit(hash_lock); 7192 break; 7193 } 7194 7195 if (pio == NULL) { 7196 /* 7197 * Insert a dummy header on the buflist so 7198 * l2arc_write_done() can find where the 7199 * write buffers begin without searching. 7200 */ 7201 mutex_enter(&dev->l2ad_mtx); 7202 list_insert_head(&dev->l2ad_buflist, head); 7203 mutex_exit(&dev->l2ad_mtx); 7204 7205 cb = kmem_alloc( 7206 sizeof (l2arc_write_callback_t), KM_SLEEP); 7207 cb->l2wcb_dev = dev; 7208 cb->l2wcb_head = head; 7209 pio = zio_root(spa, l2arc_write_done, cb, 7210 ZIO_FLAG_CANFAIL); 7211 } 7212 7213 hdr->b_l2hdr.b_dev = dev; 7214 hdr->b_l2hdr.b_daddr = dev->l2ad_hand; 7215 arc_hdr_set_flags(hdr, 7216 ARC_FLAG_L2_WRITING | ARC_FLAG_HAS_L2HDR); 7217 7218 mutex_enter(&dev->l2ad_mtx); 7219 list_insert_head(&dev->l2ad_buflist, hdr); 7220 mutex_exit(&dev->l2ad_mtx); 7221 7222 (void) zfs_refcount_add_many(&dev->l2ad_alloc, psize, 7223 hdr); 7224 7225 /* 7226 * Normally the L2ARC can use the hdr's data, but if 7227 * we're sharing data between the hdr and one of its 7228 * bufs, L2ARC needs its own copy of the data so that 7229 * the ZIO below can't race with the buf consumer. 7230 * Another case where we need to create a copy of the 7231 * data is when the buffer size is not device-aligned 7232 * and we need to pad the block to make it such. 7233 * That also keeps the clock hand suitably aligned. 7234 * 7235 * To ensure that the copy will be available for the 7236 * lifetime of the ZIO and be cleaned up afterwards, we 7237 * add it to the l2arc_free_on_write queue. 7238 */ 7239 abd_t *to_write; 7240 if (!HDR_SHARED_DATA(hdr) && psize == asize) { 7241 to_write = hdr->b_l1hdr.b_pabd; 7242 } else { 7243 to_write = abd_alloc_for_io(asize, 7244 HDR_ISTYPE_METADATA(hdr)); 7245 abd_copy(to_write, hdr->b_l1hdr.b_pabd, psize); 7246 if (asize != psize) { 7247 abd_zero_off(to_write, psize, 7248 asize - psize); 7249 } 7250 l2arc_free_abd_on_write(to_write, asize, 7251 arc_buf_type(hdr)); 7252 } 7253 wzio = zio_write_phys(pio, dev->l2ad_vdev, 7254 hdr->b_l2hdr.b_daddr, asize, to_write, 7255 ZIO_CHECKSUM_OFF, NULL, hdr, 7256 ZIO_PRIORITY_ASYNC_WRITE, 7257 ZIO_FLAG_CANFAIL, B_FALSE); 7258 7259 write_lsize += HDR_GET_LSIZE(hdr); 7260 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev, 7261 zio_t *, wzio); 7262 7263 write_psize += psize; 7264 write_asize += asize; 7265 dev->l2ad_hand += asize; 7266 7267 mutex_exit(hash_lock); 7268 7269 (void) zio_nowait(wzio); 7270 } 7271 7272 multilist_sublist_unlock(mls); 7273 7274 if (full == B_TRUE) 7275 break; 7276 } 7277 7278 /* No buffers selected for writing? */ 7279 if (pio == NULL) { 7280 ASSERT0(write_lsize); 7281 ASSERT(!HDR_HAS_L1HDR(head)); 7282 kmem_cache_free(hdr_l2only_cache, head); 7283 return (0); 7284 } 7285 7286 ASSERT3U(write_asize, <=, target_sz); 7287 ARCSTAT_BUMP(arcstat_l2_writes_sent); 7288 ARCSTAT_INCR(arcstat_l2_write_bytes, write_psize); 7289 ARCSTAT_INCR(arcstat_l2_lsize, write_lsize); 7290 ARCSTAT_INCR(arcstat_l2_psize, write_psize); 7291 vdev_space_update(dev->l2ad_vdev, write_psize, 0, 0); 7292 7293 /* 7294 * Bump device hand to the device start if it is approaching the end. 7295 * l2arc_evict() will already have evicted ahead for this case. 7296 */ 7297 if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) { 7298 dev->l2ad_hand = dev->l2ad_start; 7299 dev->l2ad_first = B_FALSE; 7300 } 7301 7302 dev->l2ad_writing = B_TRUE; 7303 (void) zio_wait(pio); 7304 dev->l2ad_writing = B_FALSE; 7305 7306 return (write_asize); 7307 } 7308 7309 /* 7310 * This thread feeds the L2ARC at regular intervals. This is the beating 7311 * heart of the L2ARC. 7312 */ 7313 /* ARGSUSED */ 7314 static void 7315 l2arc_feed_thread(void *unused) 7316 { 7317 callb_cpr_t cpr; 7318 l2arc_dev_t *dev; 7319 spa_t *spa; 7320 uint64_t size, wrote; 7321 clock_t begin, next = ddi_get_lbolt(); 7322 7323 CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG); 7324 7325 mutex_enter(&l2arc_feed_thr_lock); 7326 7327 while (l2arc_thread_exit == 0) { 7328 CALLB_CPR_SAFE_BEGIN(&cpr); 7329 (void) cv_timedwait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock, 7330 next); 7331 CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock); 7332 next = ddi_get_lbolt() + hz; 7333 7334 /* 7335 * Quick check for L2ARC devices. 7336 */ 7337 mutex_enter(&l2arc_dev_mtx); 7338 if (l2arc_ndev == 0) { 7339 mutex_exit(&l2arc_dev_mtx); 7340 continue; 7341 } 7342 mutex_exit(&l2arc_dev_mtx); 7343 begin = ddi_get_lbolt(); 7344 7345 /* 7346 * This selects the next l2arc device to write to, and in 7347 * doing so the next spa to feed from: dev->l2ad_spa. This 7348 * will return NULL if there are now no l2arc devices or if 7349 * they are all faulted. 7350 * 7351 * If a device is returned, its spa's config lock is also 7352 * held to prevent device removal. l2arc_dev_get_next() 7353 * will grab and release l2arc_dev_mtx. 7354 */ 7355 if ((dev = l2arc_dev_get_next()) == NULL) 7356 continue; 7357 7358 spa = dev->l2ad_spa; 7359 ASSERT3P(spa, !=, NULL); 7360 7361 /* 7362 * If the pool is read-only then force the feed thread to 7363 * sleep a little longer. 7364 */ 7365 if (!spa_writeable(spa)) { 7366 next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz; 7367 spa_config_exit(spa, SCL_L2ARC, dev); 7368 continue; 7369 } 7370 7371 /* 7372 * Avoid contributing to memory pressure. 7373 */ 7374 if (arc_reclaim_needed()) { 7375 ARCSTAT_BUMP(arcstat_l2_abort_lowmem); 7376 spa_config_exit(spa, SCL_L2ARC, dev); 7377 continue; 7378 } 7379 7380 ARCSTAT_BUMP(arcstat_l2_feeds); 7381 7382 size = l2arc_write_size(); 7383 7384 /* 7385 * Evict L2ARC buffers that will be overwritten. 7386 */ 7387 l2arc_evict(dev, size, B_FALSE); 7388 7389 /* 7390 * Write ARC buffers. 7391 */ 7392 wrote = l2arc_write_buffers(spa, dev, size); 7393 7394 /* 7395 * Calculate interval between writes. 7396 */ 7397 next = l2arc_write_interval(begin, size, wrote); 7398 spa_config_exit(spa, SCL_L2ARC, dev); 7399 } 7400 7401 l2arc_thread_exit = 0; 7402 cv_broadcast(&l2arc_feed_thr_cv); 7403 CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */ 7404 thread_exit(); 7405 } 7406 7407 boolean_t 7408 l2arc_vdev_present(vdev_t *vd) 7409 { 7410 l2arc_dev_t *dev; 7411 7412 mutex_enter(&l2arc_dev_mtx); 7413 for (dev = list_head(l2arc_dev_list); dev != NULL; 7414 dev = list_next(l2arc_dev_list, dev)) { 7415 if (dev->l2ad_vdev == vd) 7416 break; 7417 } 7418 mutex_exit(&l2arc_dev_mtx); 7419 7420 return (dev != NULL); 7421 } 7422 7423 /* 7424 * Add a vdev for use by the L2ARC. By this point the spa has already 7425 * validated the vdev and opened it. 7426 */ 7427 void 7428 l2arc_add_vdev(spa_t *spa, vdev_t *vd) 7429 { 7430 l2arc_dev_t *adddev; 7431 7432 ASSERT(!l2arc_vdev_present(vd)); 7433 7434 /* 7435 * Create a new l2arc device entry. 7436 */ 7437 adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP); 7438 adddev->l2ad_spa = spa; 7439 adddev->l2ad_vdev = vd; 7440 adddev->l2ad_start = VDEV_LABEL_START_SIZE; 7441 adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd); 7442 adddev->l2ad_hand = adddev->l2ad_start; 7443 adddev->l2ad_first = B_TRUE; 7444 adddev->l2ad_writing = B_FALSE; 7445 7446 mutex_init(&adddev->l2ad_mtx, NULL, MUTEX_DEFAULT, NULL); 7447 /* 7448 * This is a list of all ARC buffers that are still valid on the 7449 * device. 7450 */ 7451 list_create(&adddev->l2ad_buflist, sizeof (arc_buf_hdr_t), 7452 offsetof(arc_buf_hdr_t, b_l2hdr.b_l2node)); 7453 7454 vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand); 7455 zfs_refcount_create(&adddev->l2ad_alloc); 7456 7457 /* 7458 * Add device to global list 7459 */ 7460 mutex_enter(&l2arc_dev_mtx); 7461 list_insert_head(l2arc_dev_list, adddev); 7462 atomic_inc_64(&l2arc_ndev); 7463 mutex_exit(&l2arc_dev_mtx); 7464 } 7465 7466 /* 7467 * Remove a vdev from the L2ARC. 7468 */ 7469 void 7470 l2arc_remove_vdev(vdev_t *vd) 7471 { 7472 l2arc_dev_t *dev, *nextdev, *remdev = NULL; 7473 7474 /* 7475 * Find the device by vdev 7476 */ 7477 mutex_enter(&l2arc_dev_mtx); 7478 for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) { 7479 nextdev = list_next(l2arc_dev_list, dev); 7480 if (vd == dev->l2ad_vdev) { 7481 remdev = dev; 7482 break; 7483 } 7484 } 7485 ASSERT3P(remdev, !=, NULL); 7486 7487 /* 7488 * Remove device from global list 7489 */ 7490 list_remove(l2arc_dev_list, remdev); 7491 l2arc_dev_last = NULL; /* may have been invalidated */ 7492 atomic_dec_64(&l2arc_ndev); 7493 mutex_exit(&l2arc_dev_mtx); 7494 7495 /* 7496 * Clear all buflists and ARC references. L2ARC device flush. 7497 */ 7498 l2arc_evict(remdev, 0, B_TRUE); 7499 list_destroy(&remdev->l2ad_buflist); 7500 mutex_destroy(&remdev->l2ad_mtx); 7501 zfs_refcount_destroy(&remdev->l2ad_alloc); 7502 kmem_free(remdev, sizeof (l2arc_dev_t)); 7503 } 7504 7505 void 7506 l2arc_init(void) 7507 { 7508 l2arc_thread_exit = 0; 7509 l2arc_ndev = 0; 7510 l2arc_writes_sent = 0; 7511 l2arc_writes_done = 0; 7512 7513 mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL); 7514 cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL); 7515 mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL); 7516 mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL); 7517 7518 l2arc_dev_list = &L2ARC_dev_list; 7519 l2arc_free_on_write = &L2ARC_free_on_write; 7520 list_create(l2arc_dev_list, sizeof (l2arc_dev_t), 7521 offsetof(l2arc_dev_t, l2ad_node)); 7522 list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t), 7523 offsetof(l2arc_data_free_t, l2df_list_node)); 7524 } 7525 7526 void 7527 l2arc_fini(void) 7528 { 7529 /* 7530 * This is called from dmu_fini(), which is called from spa_fini(); 7531 * Because of this, we can assume that all l2arc devices have 7532 * already been removed when the pools themselves were removed. 7533 */ 7534 7535 l2arc_do_free_on_write(); 7536 7537 mutex_destroy(&l2arc_feed_thr_lock); 7538 cv_destroy(&l2arc_feed_thr_cv); 7539 mutex_destroy(&l2arc_dev_mtx); 7540 mutex_destroy(&l2arc_free_on_write_mtx); 7541 7542 list_destroy(l2arc_dev_list); 7543 list_destroy(l2arc_free_on_write); 7544 } 7545 7546 void 7547 l2arc_start(void) 7548 { 7549 if (!(spa_mode_global & FWRITE)) 7550 return; 7551 7552 (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0, 7553 TS_RUN, minclsyspri); 7554 } 7555 7556 void 7557 l2arc_stop(void) 7558 { 7559 if (!(spa_mode_global & FWRITE)) 7560 return; 7561 7562 mutex_enter(&l2arc_feed_thr_lock); 7563 cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */ 7564 l2arc_thread_exit = 1; 7565 while (l2arc_thread_exit != 0) 7566 cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock); 7567 mutex_exit(&l2arc_feed_thr_lock); 7568 } 7569