1 /* 2 * CDDL HEADER START 3 * 4 * The contents of this file are subject to the terms of the 5 * Common Development and Distribution License (the "License"). 6 * You may not use this file except in compliance with the License. 7 * 8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE 9 * or http://www.opensolaris.org/os/licensing. 10 * See the License for the specific language governing permissions 11 * and limitations under the License. 12 * 13 * When distributing Covered Code, include this CDDL HEADER in each 14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE. 15 * If applicable, add the following below this CDDL HEADER, with the 16 * fields enclosed by brackets "[]" replaced with your own identifying 17 * information: Portions Copyright [yyyy] [name of copyright owner] 18 * 19 * CDDL HEADER END 20 */ 21 /* 22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved. 23 * Copyright (c) 2012, Joyent, Inc. All rights reserved. 24 * Copyright (c) 2011, 2016 by Delphix. All rights reserved. 25 * Copyright (c) 2014 by Saso Kiselkov. All rights reserved. 26 * Copyright 2015 Nexenta Systems, Inc. All rights reserved. 27 */ 28 29 /* 30 * DVA-based Adjustable Replacement Cache 31 * 32 * While much of the theory of operation used here is 33 * based on the self-tuning, low overhead replacement cache 34 * presented by Megiddo and Modha at FAST 2003, there are some 35 * significant differences: 36 * 37 * 1. The Megiddo and Modha model assumes any page is evictable. 38 * Pages in its cache cannot be "locked" into memory. This makes 39 * the eviction algorithm simple: evict the last page in the list. 40 * This also make the performance characteristics easy to reason 41 * about. Our cache is not so simple. At any given moment, some 42 * subset of the blocks in the cache are un-evictable because we 43 * have handed out a reference to them. Blocks are only evictable 44 * when there are no external references active. This makes 45 * eviction far more problematic: we choose to evict the evictable 46 * blocks that are the "lowest" in the list. 47 * 48 * There are times when it is not possible to evict the requested 49 * space. In these circumstances we are unable to adjust the cache 50 * size. To prevent the cache growing unbounded at these times we 51 * implement a "cache throttle" that slows the flow of new data 52 * into the cache until we can make space available. 53 * 54 * 2. The Megiddo and Modha model assumes a fixed cache size. 55 * Pages are evicted when the cache is full and there is a cache 56 * miss. Our model has a variable sized cache. It grows with 57 * high use, but also tries to react to memory pressure from the 58 * operating system: decreasing its size when system memory is 59 * tight. 60 * 61 * 3. The Megiddo and Modha model assumes a fixed page size. All 62 * elements of the cache are therefore exactly the same size. So 63 * when adjusting the cache size following a cache miss, its simply 64 * a matter of choosing a single page to evict. In our model, we 65 * have variable sized cache blocks (rangeing from 512 bytes to 66 * 128K bytes). We therefore choose a set of blocks to evict to make 67 * space for a cache miss that approximates as closely as possible 68 * the space used by the new block. 69 * 70 * See also: "ARC: A Self-Tuning, Low Overhead Replacement Cache" 71 * by N. Megiddo & D. Modha, FAST 2003 72 */ 73 74 /* 75 * The locking model: 76 * 77 * A new reference to a cache buffer can be obtained in two 78 * ways: 1) via a hash table lookup using the DVA as a key, 79 * or 2) via one of the ARC lists. The arc_read() interface 80 * uses method 1, while the internal ARC algorithms for 81 * adjusting the cache use method 2. We therefore provide two 82 * types of locks: 1) the hash table lock array, and 2) the 83 * ARC list locks. 84 * 85 * Buffers do not have their own mutexes, rather they rely on the 86 * hash table mutexes for the bulk of their protection (i.e. most 87 * fields in the arc_buf_hdr_t are protected by these mutexes). 88 * 89 * buf_hash_find() returns the appropriate mutex (held) when it 90 * locates the requested buffer in the hash table. It returns 91 * NULL for the mutex if the buffer was not in the table. 92 * 93 * buf_hash_remove() expects the appropriate hash mutex to be 94 * already held before it is invoked. 95 * 96 * Each ARC state also has a mutex which is used to protect the 97 * buffer list associated with the state. When attempting to 98 * obtain a hash table lock while holding an ARC list lock you 99 * must use: mutex_tryenter() to avoid deadlock. Also note that 100 * the active state mutex must be held before the ghost state mutex. 101 * 102 * Note that the majority of the performance stats are manipulated 103 * with atomic operations. 104 * 105 * The L2ARC uses the l2ad_mtx on each vdev for the following: 106 * 107 * - L2ARC buflist creation 108 * - L2ARC buflist eviction 109 * - L2ARC write completion, which walks L2ARC buflists 110 * - ARC header destruction, as it removes from L2ARC buflists 111 * - ARC header release, as it removes from L2ARC buflists 112 */ 113 114 /* 115 * ARC operation: 116 * 117 * Every block that is in the ARC is tracked by an arc_buf_hdr_t structure. 118 * This structure can point either to a block that is still in the cache or to 119 * one that is only accessible in an L2 ARC device, or it can provide 120 * information about a block that was recently evicted. If a block is 121 * only accessible in the L2ARC, then the arc_buf_hdr_t only has enough 122 * information to retrieve it from the L2ARC device. This information is 123 * stored in the l2arc_buf_hdr_t sub-structure of the arc_buf_hdr_t. A block 124 * that is in this state cannot access the data directly. 125 * 126 * Blocks that are actively being referenced or have not been evicted 127 * are cached in the L1ARC. The L1ARC (l1arc_buf_hdr_t) is a structure within 128 * the arc_buf_hdr_t that will point to the data block in memory. A block can 129 * only be read by a consumer if it has an l1arc_buf_hdr_t. The L1ARC 130 * caches data in two ways -- in a list of ARC buffers (arc_buf_t) and 131 * also in the arc_buf_hdr_t's private physical data block pointer (b_pdata). 132 * 133 * The L1ARC's data pointer may or may not be uncompressed. The ARC has the 134 * ability to store the physical data (b_pdata) associated with the DVA of the 135 * arc_buf_hdr_t. Since the b_pdata is a copy of the on-disk physical block, 136 * it will match its on-disk compression characteristics. This behavior can be 137 * disabled by setting 'zfs_compressed_arc_enabled' to B_FALSE. When the 138 * compressed ARC functionality is disabled, the b_pdata will point to an 139 * uncompressed version of the on-disk data. 140 * 141 * Data in the L1ARC is not accessed by consumers of the ARC directly. Each 142 * arc_buf_hdr_t can have multiple ARC buffers (arc_buf_t) which reference it. 143 * Each ARC buffer (arc_buf_t) is being actively accessed by a specific ARC 144 * consumer. The ARC will provide references to this data and will keep it 145 * cached until it is no longer in use. The ARC caches only the L1ARC's physical 146 * data block and will evict any arc_buf_t that is no longer referenced. The 147 * amount of memory consumed by the arc_buf_ts' data buffers can be seen via the 148 * "overhead_size" kstat. 149 * 150 * Depending on the consumer, an arc_buf_t can be requested in uncompressed or 151 * compressed form. The typical case is that consumers will want uncompressed 152 * data, and when that happens a new data buffer is allocated where the data is 153 * decompressed for them to use. Currently the only consumer who wants 154 * compressed arc_buf_t's is "zfs send", when it streams data exactly as it 155 * exists on disk. When this happens, the arc_buf_t's data buffer is shared 156 * with the arc_buf_hdr_t. 157 * 158 * Here is a diagram showing an arc_buf_hdr_t referenced by two arc_buf_t's. The 159 * first one is owned by a compressed send consumer (and therefore references 160 * the same compressed data buffer as the arc_buf_hdr_t) and the second could be 161 * used by any other consumer (and has its own uncompressed copy of the data 162 * buffer). 163 * 164 * arc_buf_hdr_t 165 * +-----------+ 166 * | fields | 167 * | common to | 168 * | L1- and | 169 * | L2ARC | 170 * +-----------+ 171 * | l2arc_buf_hdr_t 172 * | | 173 * +-----------+ 174 * | l1arc_buf_hdr_t 175 * | | arc_buf_t 176 * | b_buf +------------>+-----------+ arc_buf_t 177 * | b_pdata +-+ |b_next +---->+-----------+ 178 * +-----------+ | |-----------| |b_next +-->NULL 179 * | |b_comp = T | +-----------+ 180 * | |b_data +-+ |b_comp = F | 181 * | +-----------+ | |b_data +-+ 182 * +->+------+ | +-----------+ | 183 * compressed | | | | 184 * data | |<--------------+ | uncompressed 185 * +------+ compressed, | data 186 * shared +-->+------+ 187 * data | | 188 * | | 189 * +------+ 190 * 191 * When a consumer reads a block, the ARC must first look to see if the 192 * arc_buf_hdr_t is cached. If the hdr is cached then the ARC allocates a new 193 * arc_buf_t and either copies uncompressed data into a new data buffer from an 194 * existing uncompressed arc_buf_t, decompresses the hdr's b_pdata buffer into a 195 * new data buffer, or shares the hdr's b_pdata buffer, depending on whether the 196 * hdr is compressed and the desired compression characteristics of the 197 * arc_buf_t consumer. If the arc_buf_t ends up sharing data with the 198 * arc_buf_hdr_t and both of them are uncompressed then the arc_buf_t must be 199 * the last buffer in the hdr's b_buf list, however a shared compressed buf can 200 * be anywhere in the hdr's list. 201 * 202 * The diagram below shows an example of an uncompressed ARC hdr that is 203 * sharing its data with an arc_buf_t (note that the shared uncompressed buf is 204 * the last element in the buf list): 205 * 206 * arc_buf_hdr_t 207 * +-----------+ 208 * | | 209 * | | 210 * | | 211 * +-----------+ 212 * l2arc_buf_hdr_t| | 213 * | | 214 * +-----------+ 215 * l1arc_buf_hdr_t| | 216 * | | arc_buf_t (shared) 217 * | b_buf +------------>+---------+ arc_buf_t 218 * | | |b_next +---->+---------+ 219 * | b_pdata +-+ |---------| |b_next +-->NULL 220 * +-----------+ | | | +---------+ 221 * | |b_data +-+ | | 222 * | +---------+ | |b_data +-+ 223 * +->+------+ | +---------+ | 224 * | | | | 225 * uncompressed | | | | 226 * data +------+ | | 227 * ^ +->+------+ | 228 * | uncompressed | | | 229 * | data | | | 230 * | +------+ | 231 * +---------------------------------+ 232 * 233 * Writing to the ARC requires that the ARC first discard the hdr's b_pdata 234 * since the physical block is about to be rewritten. The new data contents 235 * will be contained in the arc_buf_t. As the I/O pipeline performs the write, 236 * it may compress the data before writing it to disk. The ARC will be called 237 * with the transformed data and will bcopy the transformed on-disk block into 238 * a newly allocated b_pdata. Writes are always done into buffers which have 239 * either been loaned (and hence are new and don't have other readers) or 240 * buffers which have been released (and hence have their own hdr, if there 241 * were originally other readers of the buf's original hdr). This ensures that 242 * the ARC only needs to update a single buf and its hdr after a write occurs. 243 * 244 * When the L2ARC is in use, it will also take advantage of the b_pdata. The 245 * L2ARC will always write the contents of b_pdata to the L2ARC. This means 246 * that when compressed ARC is enabled that the L2ARC blocks are identical 247 * to the on-disk block in the main data pool. This provides a significant 248 * advantage since the ARC can leverage the bp's checksum when reading from the 249 * L2ARC to determine if the contents are valid. However, if the compressed 250 * ARC is disabled, then the L2ARC's block must be transformed to look 251 * like the physical block in the main data pool before comparing the 252 * checksum and determining its validity. 253 */ 254 255 #include <sys/spa.h> 256 #include <sys/zio.h> 257 #include <sys/spa_impl.h> 258 #include <sys/zio_compress.h> 259 #include <sys/zio_checksum.h> 260 #include <sys/zfs_context.h> 261 #include <sys/arc.h> 262 #include <sys/refcount.h> 263 #include <sys/vdev.h> 264 #include <sys/vdev_impl.h> 265 #include <sys/dsl_pool.h> 266 #include <sys/multilist.h> 267 #ifdef _KERNEL 268 #include <sys/vmsystm.h> 269 #include <vm/anon.h> 270 #include <sys/fs/swapnode.h> 271 #include <sys/dnlc.h> 272 #endif 273 #include <sys/callb.h> 274 #include <sys/kstat.h> 275 #include <zfs_fletcher.h> 276 277 #ifndef _KERNEL 278 /* set with ZFS_DEBUG=watch, to enable watchpoints on frozen buffers */ 279 boolean_t arc_watch = B_FALSE; 280 int arc_procfd; 281 #endif 282 283 static kmutex_t arc_reclaim_lock; 284 static kcondvar_t arc_reclaim_thread_cv; 285 static boolean_t arc_reclaim_thread_exit; 286 static kcondvar_t arc_reclaim_waiters_cv; 287 288 uint_t arc_reduce_dnlc_percent = 3; 289 290 /* 291 * The number of headers to evict in arc_evict_state_impl() before 292 * dropping the sublist lock and evicting from another sublist. A lower 293 * value means we're more likely to evict the "correct" header (i.e. the 294 * oldest header in the arc state), but comes with higher overhead 295 * (i.e. more invocations of arc_evict_state_impl()). 296 */ 297 int zfs_arc_evict_batch_limit = 10; 298 299 /* 300 * The number of sublists used for each of the arc state lists. If this 301 * is not set to a suitable value by the user, it will be configured to 302 * the number of CPUs on the system in arc_init(). 303 */ 304 int zfs_arc_num_sublists_per_state = 0; 305 306 /* number of seconds before growing cache again */ 307 static int arc_grow_retry = 60; 308 309 /* shift of arc_c for calculating overflow limit in arc_get_data_buf */ 310 int zfs_arc_overflow_shift = 8; 311 312 /* shift of arc_c for calculating both min and max arc_p */ 313 static int arc_p_min_shift = 4; 314 315 /* log2(fraction of arc to reclaim) */ 316 static int arc_shrink_shift = 7; 317 318 /* 319 * log2(fraction of ARC which must be free to allow growing). 320 * I.e. If there is less than arc_c >> arc_no_grow_shift free memory, 321 * when reading a new block into the ARC, we will evict an equal-sized block 322 * from the ARC. 323 * 324 * This must be less than arc_shrink_shift, so that when we shrink the ARC, 325 * we will still not allow it to grow. 326 */ 327 int arc_no_grow_shift = 5; 328 329 330 /* 331 * minimum lifespan of a prefetch block in clock ticks 332 * (initialized in arc_init()) 333 */ 334 static int arc_min_prefetch_lifespan; 335 336 /* 337 * If this percent of memory is free, don't throttle. 338 */ 339 int arc_lotsfree_percent = 10; 340 341 static int arc_dead; 342 343 /* 344 * The arc has filled available memory and has now warmed up. 345 */ 346 static boolean_t arc_warm; 347 348 /* 349 * log2 fraction of the zio arena to keep free. 350 */ 351 int arc_zio_arena_free_shift = 2; 352 353 /* 354 * These tunables are for performance analysis. 355 */ 356 uint64_t zfs_arc_max; 357 uint64_t zfs_arc_min; 358 uint64_t zfs_arc_meta_limit = 0; 359 uint64_t zfs_arc_meta_min = 0; 360 int zfs_arc_grow_retry = 0; 361 int zfs_arc_shrink_shift = 0; 362 int zfs_arc_p_min_shift = 0; 363 int zfs_arc_average_blocksize = 8 * 1024; /* 8KB */ 364 365 boolean_t zfs_compressed_arc_enabled = B_TRUE; 366 367 /* 368 * Note that buffers can be in one of 6 states: 369 * ARC_anon - anonymous (discussed below) 370 * ARC_mru - recently used, currently cached 371 * ARC_mru_ghost - recentely used, no longer in cache 372 * ARC_mfu - frequently used, currently cached 373 * ARC_mfu_ghost - frequently used, no longer in cache 374 * ARC_l2c_only - exists in L2ARC but not other states 375 * When there are no active references to the buffer, they are 376 * are linked onto a list in one of these arc states. These are 377 * the only buffers that can be evicted or deleted. Within each 378 * state there are multiple lists, one for meta-data and one for 379 * non-meta-data. Meta-data (indirect blocks, blocks of dnodes, 380 * etc.) is tracked separately so that it can be managed more 381 * explicitly: favored over data, limited explicitly. 382 * 383 * Anonymous buffers are buffers that are not associated with 384 * a DVA. These are buffers that hold dirty block copies 385 * before they are written to stable storage. By definition, 386 * they are "ref'd" and are considered part of arc_mru 387 * that cannot be freed. Generally, they will aquire a DVA 388 * as they are written and migrate onto the arc_mru list. 389 * 390 * The ARC_l2c_only state is for buffers that are in the second 391 * level ARC but no longer in any of the ARC_m* lists. The second 392 * level ARC itself may also contain buffers that are in any of 393 * the ARC_m* states - meaning that a buffer can exist in two 394 * places. The reason for the ARC_l2c_only state is to keep the 395 * buffer header in the hash table, so that reads that hit the 396 * second level ARC benefit from these fast lookups. 397 */ 398 399 typedef struct arc_state { 400 /* 401 * list of evictable buffers 402 */ 403 multilist_t arcs_list[ARC_BUFC_NUMTYPES]; 404 /* 405 * total amount of evictable data in this state 406 */ 407 refcount_t arcs_esize[ARC_BUFC_NUMTYPES]; 408 /* 409 * total amount of data in this state; this includes: evictable, 410 * non-evictable, ARC_BUFC_DATA, and ARC_BUFC_METADATA. 411 */ 412 refcount_t arcs_size; 413 } arc_state_t; 414 415 /* The 6 states: */ 416 static arc_state_t ARC_anon; 417 static arc_state_t ARC_mru; 418 static arc_state_t ARC_mru_ghost; 419 static arc_state_t ARC_mfu; 420 static arc_state_t ARC_mfu_ghost; 421 static arc_state_t ARC_l2c_only; 422 423 typedef struct arc_stats { 424 kstat_named_t arcstat_hits; 425 kstat_named_t arcstat_misses; 426 kstat_named_t arcstat_demand_data_hits; 427 kstat_named_t arcstat_demand_data_misses; 428 kstat_named_t arcstat_demand_metadata_hits; 429 kstat_named_t arcstat_demand_metadata_misses; 430 kstat_named_t arcstat_prefetch_data_hits; 431 kstat_named_t arcstat_prefetch_data_misses; 432 kstat_named_t arcstat_prefetch_metadata_hits; 433 kstat_named_t arcstat_prefetch_metadata_misses; 434 kstat_named_t arcstat_mru_hits; 435 kstat_named_t arcstat_mru_ghost_hits; 436 kstat_named_t arcstat_mfu_hits; 437 kstat_named_t arcstat_mfu_ghost_hits; 438 kstat_named_t arcstat_deleted; 439 /* 440 * Number of buffers that could not be evicted because the hash lock 441 * was held by another thread. The lock may not necessarily be held 442 * by something using the same buffer, since hash locks are shared 443 * by multiple buffers. 444 */ 445 kstat_named_t arcstat_mutex_miss; 446 /* 447 * Number of buffers skipped because they have I/O in progress, are 448 * indrect prefetch buffers that have not lived long enough, or are 449 * not from the spa we're trying to evict from. 450 */ 451 kstat_named_t arcstat_evict_skip; 452 /* 453 * Number of times arc_evict_state() was unable to evict enough 454 * buffers to reach it's target amount. 455 */ 456 kstat_named_t arcstat_evict_not_enough; 457 kstat_named_t arcstat_evict_l2_cached; 458 kstat_named_t arcstat_evict_l2_eligible; 459 kstat_named_t arcstat_evict_l2_ineligible; 460 kstat_named_t arcstat_evict_l2_skip; 461 kstat_named_t arcstat_hash_elements; 462 kstat_named_t arcstat_hash_elements_max; 463 kstat_named_t arcstat_hash_collisions; 464 kstat_named_t arcstat_hash_chains; 465 kstat_named_t arcstat_hash_chain_max; 466 kstat_named_t arcstat_p; 467 kstat_named_t arcstat_c; 468 kstat_named_t arcstat_c_min; 469 kstat_named_t arcstat_c_max; 470 kstat_named_t arcstat_size; 471 /* 472 * Number of compressed bytes stored in the arc_buf_hdr_t's b_pdata. 473 * Note that the compressed bytes may match the uncompressed bytes 474 * if the block is either not compressed or compressed arc is disabled. 475 */ 476 kstat_named_t arcstat_compressed_size; 477 /* 478 * Uncompressed size of the data stored in b_pdata. If compressed 479 * arc is disabled then this value will be identical to the stat 480 * above. 481 */ 482 kstat_named_t arcstat_uncompressed_size; 483 /* 484 * Number of bytes stored in all the arc_buf_t's. This is classified 485 * as "overhead" since this data is typically short-lived and will 486 * be evicted from the arc when it becomes unreferenced unless the 487 * zfs_keep_uncompressed_metadata or zfs_keep_uncompressed_level 488 * values have been set (see comment in dbuf.c for more information). 489 */ 490 kstat_named_t arcstat_overhead_size; 491 /* 492 * Number of bytes consumed by internal ARC structures necessary 493 * for tracking purposes; these structures are not actually 494 * backed by ARC buffers. This includes arc_buf_hdr_t structures 495 * (allocated via arc_buf_hdr_t_full and arc_buf_hdr_t_l2only 496 * caches), and arc_buf_t structures (allocated via arc_buf_t 497 * cache). 498 */ 499 kstat_named_t arcstat_hdr_size; 500 /* 501 * Number of bytes consumed by ARC buffers of type equal to 502 * ARC_BUFC_DATA. This is generally consumed by buffers backing 503 * on disk user data (e.g. plain file contents). 504 */ 505 kstat_named_t arcstat_data_size; 506 /* 507 * Number of bytes consumed by ARC buffers of type equal to 508 * ARC_BUFC_METADATA. This is generally consumed by buffers 509 * backing on disk data that is used for internal ZFS 510 * structures (e.g. ZAP, dnode, indirect blocks, etc). 511 */ 512 kstat_named_t arcstat_metadata_size; 513 /* 514 * Number of bytes consumed by various buffers and structures 515 * not actually backed with ARC buffers. This includes bonus 516 * buffers (allocated directly via zio_buf_* functions), 517 * dmu_buf_impl_t structures (allocated via dmu_buf_impl_t 518 * cache), and dnode_t structures (allocated via dnode_t cache). 519 */ 520 kstat_named_t arcstat_other_size; 521 /* 522 * Total number of bytes consumed by ARC buffers residing in the 523 * arc_anon state. This includes *all* buffers in the arc_anon 524 * state; e.g. data, metadata, evictable, and unevictable buffers 525 * are all included in this value. 526 */ 527 kstat_named_t arcstat_anon_size; 528 /* 529 * Number of bytes consumed by ARC buffers that meet the 530 * following criteria: backing buffers of type ARC_BUFC_DATA, 531 * residing in the arc_anon state, and are eligible for eviction 532 * (e.g. have no outstanding holds on the buffer). 533 */ 534 kstat_named_t arcstat_anon_evictable_data; 535 /* 536 * Number of bytes consumed by ARC buffers that meet the 537 * following criteria: backing buffers of type ARC_BUFC_METADATA, 538 * residing in the arc_anon state, and are eligible for eviction 539 * (e.g. have no outstanding holds on the buffer). 540 */ 541 kstat_named_t arcstat_anon_evictable_metadata; 542 /* 543 * Total number of bytes consumed by ARC buffers residing in the 544 * arc_mru state. This includes *all* buffers in the arc_mru 545 * state; e.g. data, metadata, evictable, and unevictable buffers 546 * are all included in this value. 547 */ 548 kstat_named_t arcstat_mru_size; 549 /* 550 * Number of bytes consumed by ARC buffers that meet the 551 * following criteria: backing buffers of type ARC_BUFC_DATA, 552 * residing in the arc_mru state, and are eligible for eviction 553 * (e.g. have no outstanding holds on the buffer). 554 */ 555 kstat_named_t arcstat_mru_evictable_data; 556 /* 557 * Number of bytes consumed by ARC buffers that meet the 558 * following criteria: backing buffers of type ARC_BUFC_METADATA, 559 * residing in the arc_mru state, and are eligible for eviction 560 * (e.g. have no outstanding holds on the buffer). 561 */ 562 kstat_named_t arcstat_mru_evictable_metadata; 563 /* 564 * Total number of bytes that *would have been* consumed by ARC 565 * buffers in the arc_mru_ghost state. The key thing to note 566 * here, is the fact that this size doesn't actually indicate 567 * RAM consumption. The ghost lists only consist of headers and 568 * don't actually have ARC buffers linked off of these headers. 569 * Thus, *if* the headers had associated ARC buffers, these 570 * buffers *would have* consumed this number of bytes. 571 */ 572 kstat_named_t arcstat_mru_ghost_size; 573 /* 574 * Number of bytes that *would have been* consumed by ARC 575 * buffers that are eligible for eviction, of type 576 * ARC_BUFC_DATA, and linked off the arc_mru_ghost state. 577 */ 578 kstat_named_t arcstat_mru_ghost_evictable_data; 579 /* 580 * Number of bytes that *would have been* consumed by ARC 581 * buffers that are eligible for eviction, of type 582 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state. 583 */ 584 kstat_named_t arcstat_mru_ghost_evictable_metadata; 585 /* 586 * Total number of bytes consumed by ARC buffers residing in the 587 * arc_mfu state. This includes *all* buffers in the arc_mfu 588 * state; e.g. data, metadata, evictable, and unevictable buffers 589 * are all included in this value. 590 */ 591 kstat_named_t arcstat_mfu_size; 592 /* 593 * Number of bytes consumed by ARC buffers that are eligible for 594 * eviction, of type ARC_BUFC_DATA, and reside in the arc_mfu 595 * state. 596 */ 597 kstat_named_t arcstat_mfu_evictable_data; 598 /* 599 * Number of bytes consumed by ARC buffers that are eligible for 600 * eviction, of type ARC_BUFC_METADATA, and reside in the 601 * arc_mfu state. 602 */ 603 kstat_named_t arcstat_mfu_evictable_metadata; 604 /* 605 * Total number of bytes that *would have been* consumed by ARC 606 * buffers in the arc_mfu_ghost state. See the comment above 607 * arcstat_mru_ghost_size for more details. 608 */ 609 kstat_named_t arcstat_mfu_ghost_size; 610 /* 611 * Number of bytes that *would have been* consumed by ARC 612 * buffers that are eligible for eviction, of type 613 * ARC_BUFC_DATA, and linked off the arc_mfu_ghost state. 614 */ 615 kstat_named_t arcstat_mfu_ghost_evictable_data; 616 /* 617 * Number of bytes that *would have been* consumed by ARC 618 * buffers that are eligible for eviction, of type 619 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state. 620 */ 621 kstat_named_t arcstat_mfu_ghost_evictable_metadata; 622 kstat_named_t arcstat_l2_hits; 623 kstat_named_t arcstat_l2_misses; 624 kstat_named_t arcstat_l2_feeds; 625 kstat_named_t arcstat_l2_rw_clash; 626 kstat_named_t arcstat_l2_read_bytes; 627 kstat_named_t arcstat_l2_write_bytes; 628 kstat_named_t arcstat_l2_writes_sent; 629 kstat_named_t arcstat_l2_writes_done; 630 kstat_named_t arcstat_l2_writes_error; 631 kstat_named_t arcstat_l2_writes_lock_retry; 632 kstat_named_t arcstat_l2_evict_lock_retry; 633 kstat_named_t arcstat_l2_evict_reading; 634 kstat_named_t arcstat_l2_evict_l1cached; 635 kstat_named_t arcstat_l2_free_on_write; 636 kstat_named_t arcstat_l2_abort_lowmem; 637 kstat_named_t arcstat_l2_cksum_bad; 638 kstat_named_t arcstat_l2_io_error; 639 kstat_named_t arcstat_l2_size; 640 kstat_named_t arcstat_l2_asize; 641 kstat_named_t arcstat_l2_hdr_size; 642 kstat_named_t arcstat_memory_throttle_count; 643 kstat_named_t arcstat_meta_used; 644 kstat_named_t arcstat_meta_limit; 645 kstat_named_t arcstat_meta_max; 646 kstat_named_t arcstat_meta_min; 647 kstat_named_t arcstat_sync_wait_for_async; 648 kstat_named_t arcstat_demand_hit_predictive_prefetch; 649 } arc_stats_t; 650 651 static arc_stats_t arc_stats = { 652 { "hits", KSTAT_DATA_UINT64 }, 653 { "misses", KSTAT_DATA_UINT64 }, 654 { "demand_data_hits", KSTAT_DATA_UINT64 }, 655 { "demand_data_misses", KSTAT_DATA_UINT64 }, 656 { "demand_metadata_hits", KSTAT_DATA_UINT64 }, 657 { "demand_metadata_misses", KSTAT_DATA_UINT64 }, 658 { "prefetch_data_hits", KSTAT_DATA_UINT64 }, 659 { "prefetch_data_misses", KSTAT_DATA_UINT64 }, 660 { "prefetch_metadata_hits", KSTAT_DATA_UINT64 }, 661 { "prefetch_metadata_misses", KSTAT_DATA_UINT64 }, 662 { "mru_hits", KSTAT_DATA_UINT64 }, 663 { "mru_ghost_hits", KSTAT_DATA_UINT64 }, 664 { "mfu_hits", KSTAT_DATA_UINT64 }, 665 { "mfu_ghost_hits", KSTAT_DATA_UINT64 }, 666 { "deleted", KSTAT_DATA_UINT64 }, 667 { "mutex_miss", KSTAT_DATA_UINT64 }, 668 { "evict_skip", KSTAT_DATA_UINT64 }, 669 { "evict_not_enough", KSTAT_DATA_UINT64 }, 670 { "evict_l2_cached", KSTAT_DATA_UINT64 }, 671 { "evict_l2_eligible", KSTAT_DATA_UINT64 }, 672 { "evict_l2_ineligible", KSTAT_DATA_UINT64 }, 673 { "evict_l2_skip", KSTAT_DATA_UINT64 }, 674 { "hash_elements", KSTAT_DATA_UINT64 }, 675 { "hash_elements_max", KSTAT_DATA_UINT64 }, 676 { "hash_collisions", KSTAT_DATA_UINT64 }, 677 { "hash_chains", KSTAT_DATA_UINT64 }, 678 { "hash_chain_max", KSTAT_DATA_UINT64 }, 679 { "p", KSTAT_DATA_UINT64 }, 680 { "c", KSTAT_DATA_UINT64 }, 681 { "c_min", KSTAT_DATA_UINT64 }, 682 { "c_max", KSTAT_DATA_UINT64 }, 683 { "size", KSTAT_DATA_UINT64 }, 684 { "compressed_size", KSTAT_DATA_UINT64 }, 685 { "uncompressed_size", KSTAT_DATA_UINT64 }, 686 { "overhead_size", KSTAT_DATA_UINT64 }, 687 { "hdr_size", KSTAT_DATA_UINT64 }, 688 { "data_size", KSTAT_DATA_UINT64 }, 689 { "metadata_size", KSTAT_DATA_UINT64 }, 690 { "other_size", KSTAT_DATA_UINT64 }, 691 { "anon_size", KSTAT_DATA_UINT64 }, 692 { "anon_evictable_data", KSTAT_DATA_UINT64 }, 693 { "anon_evictable_metadata", KSTAT_DATA_UINT64 }, 694 { "mru_size", KSTAT_DATA_UINT64 }, 695 { "mru_evictable_data", KSTAT_DATA_UINT64 }, 696 { "mru_evictable_metadata", KSTAT_DATA_UINT64 }, 697 { "mru_ghost_size", KSTAT_DATA_UINT64 }, 698 { "mru_ghost_evictable_data", KSTAT_DATA_UINT64 }, 699 { "mru_ghost_evictable_metadata", KSTAT_DATA_UINT64 }, 700 { "mfu_size", KSTAT_DATA_UINT64 }, 701 { "mfu_evictable_data", KSTAT_DATA_UINT64 }, 702 { "mfu_evictable_metadata", KSTAT_DATA_UINT64 }, 703 { "mfu_ghost_size", KSTAT_DATA_UINT64 }, 704 { "mfu_ghost_evictable_data", KSTAT_DATA_UINT64 }, 705 { "mfu_ghost_evictable_metadata", KSTAT_DATA_UINT64 }, 706 { "l2_hits", KSTAT_DATA_UINT64 }, 707 { "l2_misses", KSTAT_DATA_UINT64 }, 708 { "l2_feeds", KSTAT_DATA_UINT64 }, 709 { "l2_rw_clash", KSTAT_DATA_UINT64 }, 710 { "l2_read_bytes", KSTAT_DATA_UINT64 }, 711 { "l2_write_bytes", KSTAT_DATA_UINT64 }, 712 { "l2_writes_sent", KSTAT_DATA_UINT64 }, 713 { "l2_writes_done", KSTAT_DATA_UINT64 }, 714 { "l2_writes_error", KSTAT_DATA_UINT64 }, 715 { "l2_writes_lock_retry", KSTAT_DATA_UINT64 }, 716 { "l2_evict_lock_retry", KSTAT_DATA_UINT64 }, 717 { "l2_evict_reading", KSTAT_DATA_UINT64 }, 718 { "l2_evict_l1cached", KSTAT_DATA_UINT64 }, 719 { "l2_free_on_write", KSTAT_DATA_UINT64 }, 720 { "l2_abort_lowmem", KSTAT_DATA_UINT64 }, 721 { "l2_cksum_bad", KSTAT_DATA_UINT64 }, 722 { "l2_io_error", KSTAT_DATA_UINT64 }, 723 { "l2_size", KSTAT_DATA_UINT64 }, 724 { "l2_asize", KSTAT_DATA_UINT64 }, 725 { "l2_hdr_size", KSTAT_DATA_UINT64 }, 726 { "memory_throttle_count", KSTAT_DATA_UINT64 }, 727 { "arc_meta_used", KSTAT_DATA_UINT64 }, 728 { "arc_meta_limit", KSTAT_DATA_UINT64 }, 729 { "arc_meta_max", KSTAT_DATA_UINT64 }, 730 { "arc_meta_min", KSTAT_DATA_UINT64 }, 731 { "sync_wait_for_async", KSTAT_DATA_UINT64 }, 732 { "demand_hit_predictive_prefetch", KSTAT_DATA_UINT64 }, 733 }; 734 735 #define ARCSTAT(stat) (arc_stats.stat.value.ui64) 736 737 #define ARCSTAT_INCR(stat, val) \ 738 atomic_add_64(&arc_stats.stat.value.ui64, (val)) 739 740 #define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1) 741 #define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1) 742 743 #define ARCSTAT_MAX(stat, val) { \ 744 uint64_t m; \ 745 while ((val) > (m = arc_stats.stat.value.ui64) && \ 746 (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \ 747 continue; \ 748 } 749 750 #define ARCSTAT_MAXSTAT(stat) \ 751 ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64) 752 753 /* 754 * We define a macro to allow ARC hits/misses to be easily broken down by 755 * two separate conditions, giving a total of four different subtypes for 756 * each of hits and misses (so eight statistics total). 757 */ 758 #define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \ 759 if (cond1) { \ 760 if (cond2) { \ 761 ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \ 762 } else { \ 763 ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \ 764 } \ 765 } else { \ 766 if (cond2) { \ 767 ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \ 768 } else { \ 769 ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\ 770 } \ 771 } 772 773 kstat_t *arc_ksp; 774 static arc_state_t *arc_anon; 775 static arc_state_t *arc_mru; 776 static arc_state_t *arc_mru_ghost; 777 static arc_state_t *arc_mfu; 778 static arc_state_t *arc_mfu_ghost; 779 static arc_state_t *arc_l2c_only; 780 781 /* 782 * There are several ARC variables that are critical to export as kstats -- 783 * but we don't want to have to grovel around in the kstat whenever we wish to 784 * manipulate them. For these variables, we therefore define them to be in 785 * terms of the statistic variable. This assures that we are not introducing 786 * the possibility of inconsistency by having shadow copies of the variables, 787 * while still allowing the code to be readable. 788 */ 789 #define arc_size ARCSTAT(arcstat_size) /* actual total arc size */ 790 #define arc_p ARCSTAT(arcstat_p) /* target size of MRU */ 791 #define arc_c ARCSTAT(arcstat_c) /* target size of cache */ 792 #define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */ 793 #define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */ 794 #define arc_meta_limit ARCSTAT(arcstat_meta_limit) /* max size for metadata */ 795 #define arc_meta_min ARCSTAT(arcstat_meta_min) /* min size for metadata */ 796 #define arc_meta_used ARCSTAT(arcstat_meta_used) /* size of metadata */ 797 #define arc_meta_max ARCSTAT(arcstat_meta_max) /* max size of metadata */ 798 799 /* compressed size of entire arc */ 800 #define arc_compressed_size ARCSTAT(arcstat_compressed_size) 801 /* uncompressed size of entire arc */ 802 #define arc_uncompressed_size ARCSTAT(arcstat_uncompressed_size) 803 /* number of bytes in the arc from arc_buf_t's */ 804 #define arc_overhead_size ARCSTAT(arcstat_overhead_size) 805 806 static int arc_no_grow; /* Don't try to grow cache size */ 807 static uint64_t arc_tempreserve; 808 static uint64_t arc_loaned_bytes; 809 810 typedef struct arc_callback arc_callback_t; 811 812 struct arc_callback { 813 void *acb_private; 814 arc_done_func_t *acb_done; 815 arc_buf_t *acb_buf; 816 boolean_t acb_compressed; 817 zio_t *acb_zio_dummy; 818 arc_callback_t *acb_next; 819 }; 820 821 typedef struct arc_write_callback arc_write_callback_t; 822 823 struct arc_write_callback { 824 void *awcb_private; 825 arc_done_func_t *awcb_ready; 826 arc_done_func_t *awcb_children_ready; 827 arc_done_func_t *awcb_physdone; 828 arc_done_func_t *awcb_done; 829 arc_buf_t *awcb_buf; 830 }; 831 832 /* 833 * ARC buffers are separated into multiple structs as a memory saving measure: 834 * - Common fields struct, always defined, and embedded within it: 835 * - L2-only fields, always allocated but undefined when not in L2ARC 836 * - L1-only fields, only allocated when in L1ARC 837 * 838 * Buffer in L1 Buffer only in L2 839 * +------------------------+ +------------------------+ 840 * | arc_buf_hdr_t | | arc_buf_hdr_t | 841 * | | | | 842 * | | | | 843 * | | | | 844 * +------------------------+ +------------------------+ 845 * | l2arc_buf_hdr_t | | l2arc_buf_hdr_t | 846 * | (undefined if L1-only) | | | 847 * +------------------------+ +------------------------+ 848 * | l1arc_buf_hdr_t | 849 * | | 850 * | | 851 * | | 852 * | | 853 * +------------------------+ 854 * 855 * Because it's possible for the L2ARC to become extremely large, we can wind 856 * up eating a lot of memory in L2ARC buffer headers, so the size of a header 857 * is minimized by only allocating the fields necessary for an L1-cached buffer 858 * when a header is actually in the L1 cache. The sub-headers (l1arc_buf_hdr and 859 * l2arc_buf_hdr) are embedded rather than allocated separately to save a couple 860 * words in pointers. arc_hdr_realloc() is used to switch a header between 861 * these two allocation states. 862 */ 863 typedef struct l1arc_buf_hdr { 864 kmutex_t b_freeze_lock; 865 zio_cksum_t *b_freeze_cksum; 866 #ifdef ZFS_DEBUG 867 /* 868 * Used for debugging with kmem_flags - by allocating and freeing 869 * b_thawed when the buffer is thawed, we get a record of the stack 870 * trace that thawed it. 871 */ 872 void *b_thawed; 873 #endif 874 875 arc_buf_t *b_buf; 876 uint32_t b_bufcnt; 877 /* for waiting on writes to complete */ 878 kcondvar_t b_cv; 879 uint8_t b_byteswap; 880 881 /* protected by arc state mutex */ 882 arc_state_t *b_state; 883 multilist_node_t b_arc_node; 884 885 /* updated atomically */ 886 clock_t b_arc_access; 887 888 /* self protecting */ 889 refcount_t b_refcnt; 890 891 arc_callback_t *b_acb; 892 void *b_pdata; 893 } l1arc_buf_hdr_t; 894 895 typedef struct l2arc_dev l2arc_dev_t; 896 897 typedef struct l2arc_buf_hdr { 898 /* protected by arc_buf_hdr mutex */ 899 l2arc_dev_t *b_dev; /* L2ARC device */ 900 uint64_t b_daddr; /* disk address, offset byte */ 901 902 list_node_t b_l2node; 903 } l2arc_buf_hdr_t; 904 905 struct arc_buf_hdr { 906 /* protected by hash lock */ 907 dva_t b_dva; 908 uint64_t b_birth; 909 910 arc_buf_contents_t b_type; 911 arc_buf_hdr_t *b_hash_next; 912 arc_flags_t b_flags; 913 914 /* 915 * This field stores the size of the data buffer after 916 * compression, and is set in the arc's zio completion handlers. 917 * It is in units of SPA_MINBLOCKSIZE (e.g. 1 == 512 bytes). 918 * 919 * While the block pointers can store up to 32MB in their psize 920 * field, we can only store up to 32MB minus 512B. This is due 921 * to the bp using a bias of 1, whereas we use a bias of 0 (i.e. 922 * a field of zeros represents 512B in the bp). We can't use a 923 * bias of 1 since we need to reserve a psize of zero, here, to 924 * represent holes and embedded blocks. 925 * 926 * This isn't a problem in practice, since the maximum size of a 927 * buffer is limited to 16MB, so we never need to store 32MB in 928 * this field. Even in the upstream illumos code base, the 929 * maximum size of a buffer is limited to 16MB. 930 */ 931 uint16_t b_psize; 932 933 /* 934 * This field stores the size of the data buffer before 935 * compression, and cannot change once set. It is in units 936 * of SPA_MINBLOCKSIZE (e.g. 2 == 1024 bytes) 937 */ 938 uint16_t b_lsize; /* immutable */ 939 uint64_t b_spa; /* immutable */ 940 941 /* L2ARC fields. Undefined when not in L2ARC. */ 942 l2arc_buf_hdr_t b_l2hdr; 943 /* L1ARC fields. Undefined when in l2arc_only state */ 944 l1arc_buf_hdr_t b_l1hdr; 945 }; 946 947 #define GHOST_STATE(state) \ 948 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \ 949 (state) == arc_l2c_only) 950 951 #define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_FLAG_IN_HASH_TABLE) 952 #define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) 953 #define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_FLAG_IO_ERROR) 954 #define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_FLAG_PREFETCH) 955 #define HDR_COMPRESSION_ENABLED(hdr) \ 956 ((hdr)->b_flags & ARC_FLAG_COMPRESSED_ARC) 957 958 #define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_FLAG_L2CACHE) 959 #define HDR_L2_READING(hdr) \ 960 (((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) && \ 961 ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR)) 962 #define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITING) 963 #define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_FLAG_L2_EVICTED) 964 #define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITE_HEAD) 965 #define HDR_SHARED_DATA(hdr) ((hdr)->b_flags & ARC_FLAG_SHARED_DATA) 966 967 #define HDR_ISTYPE_METADATA(hdr) \ 968 ((hdr)->b_flags & ARC_FLAG_BUFC_METADATA) 969 #define HDR_ISTYPE_DATA(hdr) (!HDR_ISTYPE_METADATA(hdr)) 970 971 #define HDR_HAS_L1HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L1HDR) 972 #define HDR_HAS_L2HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR) 973 974 /* For storing compression mode in b_flags */ 975 #define HDR_COMPRESS_OFFSET (highbit64(ARC_FLAG_COMPRESS_0) - 1) 976 977 #define HDR_GET_COMPRESS(hdr) ((enum zio_compress)BF32_GET((hdr)->b_flags, \ 978 HDR_COMPRESS_OFFSET, SPA_COMPRESSBITS)) 979 #define HDR_SET_COMPRESS(hdr, cmp) BF32_SET((hdr)->b_flags, \ 980 HDR_COMPRESS_OFFSET, SPA_COMPRESSBITS, (cmp)); 981 982 #define ARC_BUF_LAST(buf) ((buf)->b_next == NULL) 983 #define ARC_BUF_SHARED(buf) ((buf)->b_flags & ARC_BUF_FLAG_SHARED) 984 #define ARC_BUF_COMPRESSED(buf) ((buf)->b_flags & ARC_BUF_FLAG_COMPRESSED) 985 986 /* 987 * Other sizes 988 */ 989 990 #define HDR_FULL_SIZE ((int64_t)sizeof (arc_buf_hdr_t)) 991 #define HDR_L2ONLY_SIZE ((int64_t)offsetof(arc_buf_hdr_t, b_l1hdr)) 992 993 /* 994 * Hash table routines 995 */ 996 997 #define HT_LOCK_PAD 64 998 999 struct ht_lock { 1000 kmutex_t ht_lock; 1001 #ifdef _KERNEL 1002 unsigned char pad[(HT_LOCK_PAD - sizeof (kmutex_t))]; 1003 #endif 1004 }; 1005 1006 #define BUF_LOCKS 256 1007 typedef struct buf_hash_table { 1008 uint64_t ht_mask; 1009 arc_buf_hdr_t **ht_table; 1010 struct ht_lock ht_locks[BUF_LOCKS]; 1011 } buf_hash_table_t; 1012 1013 static buf_hash_table_t buf_hash_table; 1014 1015 #define BUF_HASH_INDEX(spa, dva, birth) \ 1016 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask) 1017 #define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)]) 1018 #define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock)) 1019 #define HDR_LOCK(hdr) \ 1020 (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth))) 1021 1022 uint64_t zfs_crc64_table[256]; 1023 1024 /* 1025 * Level 2 ARC 1026 */ 1027 1028 #define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */ 1029 #define L2ARC_HEADROOM 2 /* num of writes */ 1030 /* 1031 * If we discover during ARC scan any buffers to be compressed, we boost 1032 * our headroom for the next scanning cycle by this percentage multiple. 1033 */ 1034 #define L2ARC_HEADROOM_BOOST 200 1035 #define L2ARC_FEED_SECS 1 /* caching interval secs */ 1036 #define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */ 1037 1038 #define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent) 1039 #define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done) 1040 1041 /* L2ARC Performance Tunables */ 1042 uint64_t l2arc_write_max = L2ARC_WRITE_SIZE; /* default max write size */ 1043 uint64_t l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra write during warmup */ 1044 uint64_t l2arc_headroom = L2ARC_HEADROOM; /* number of dev writes */ 1045 uint64_t l2arc_headroom_boost = L2ARC_HEADROOM_BOOST; 1046 uint64_t l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */ 1047 uint64_t l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval milliseconds */ 1048 boolean_t l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */ 1049 boolean_t l2arc_feed_again = B_TRUE; /* turbo warmup */ 1050 boolean_t l2arc_norw = B_TRUE; /* no reads during writes */ 1051 1052 /* 1053 * L2ARC Internals 1054 */ 1055 struct l2arc_dev { 1056 vdev_t *l2ad_vdev; /* vdev */ 1057 spa_t *l2ad_spa; /* spa */ 1058 uint64_t l2ad_hand; /* next write location */ 1059 uint64_t l2ad_start; /* first addr on device */ 1060 uint64_t l2ad_end; /* last addr on device */ 1061 boolean_t l2ad_first; /* first sweep through */ 1062 boolean_t l2ad_writing; /* currently writing */ 1063 kmutex_t l2ad_mtx; /* lock for buffer list */ 1064 list_t l2ad_buflist; /* buffer list */ 1065 list_node_t l2ad_node; /* device list node */ 1066 refcount_t l2ad_alloc; /* allocated bytes */ 1067 }; 1068 1069 static list_t L2ARC_dev_list; /* device list */ 1070 static list_t *l2arc_dev_list; /* device list pointer */ 1071 static kmutex_t l2arc_dev_mtx; /* device list mutex */ 1072 static l2arc_dev_t *l2arc_dev_last; /* last device used */ 1073 static list_t L2ARC_free_on_write; /* free after write buf list */ 1074 static list_t *l2arc_free_on_write; /* free after write list ptr */ 1075 static kmutex_t l2arc_free_on_write_mtx; /* mutex for list */ 1076 static uint64_t l2arc_ndev; /* number of devices */ 1077 1078 typedef struct l2arc_read_callback { 1079 arc_buf_hdr_t *l2rcb_hdr; /* read header */ 1080 blkptr_t l2rcb_bp; /* original blkptr */ 1081 zbookmark_phys_t l2rcb_zb; /* original bookmark */ 1082 int l2rcb_flags; /* original flags */ 1083 } l2arc_read_callback_t; 1084 1085 typedef struct l2arc_write_callback { 1086 l2arc_dev_t *l2wcb_dev; /* device info */ 1087 arc_buf_hdr_t *l2wcb_head; /* head of write buflist */ 1088 } l2arc_write_callback_t; 1089 1090 typedef struct l2arc_data_free { 1091 /* protected by l2arc_free_on_write_mtx */ 1092 void *l2df_data; 1093 size_t l2df_size; 1094 arc_buf_contents_t l2df_type; 1095 list_node_t l2df_list_node; 1096 } l2arc_data_free_t; 1097 1098 static kmutex_t l2arc_feed_thr_lock; 1099 static kcondvar_t l2arc_feed_thr_cv; 1100 static uint8_t l2arc_thread_exit; 1101 1102 static void *arc_get_data_buf(arc_buf_hdr_t *, uint64_t, void *); 1103 static void arc_free_data_buf(arc_buf_hdr_t *, void *, uint64_t, void *); 1104 static void arc_hdr_free_pdata(arc_buf_hdr_t *hdr); 1105 static void arc_hdr_alloc_pdata(arc_buf_hdr_t *); 1106 static void arc_access(arc_buf_hdr_t *, kmutex_t *); 1107 static boolean_t arc_is_overflowing(); 1108 static void arc_buf_watch(arc_buf_t *); 1109 1110 static arc_buf_contents_t arc_buf_type(arc_buf_hdr_t *); 1111 static uint32_t arc_bufc_to_flags(arc_buf_contents_t); 1112 static inline void arc_hdr_set_flags(arc_buf_hdr_t *hdr, arc_flags_t flags); 1113 static inline void arc_hdr_clear_flags(arc_buf_hdr_t *hdr, arc_flags_t flags); 1114 1115 static boolean_t l2arc_write_eligible(uint64_t, arc_buf_hdr_t *); 1116 static void l2arc_read_done(zio_t *); 1117 1118 static uint64_t 1119 buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth) 1120 { 1121 uint8_t *vdva = (uint8_t *)dva; 1122 uint64_t crc = -1ULL; 1123 int i; 1124 1125 ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY); 1126 1127 for (i = 0; i < sizeof (dva_t); i++) 1128 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ vdva[i]) & 0xFF]; 1129 1130 crc ^= (spa>>8) ^ birth; 1131 1132 return (crc); 1133 } 1134 1135 #define HDR_EMPTY(hdr) \ 1136 ((hdr)->b_dva.dva_word[0] == 0 && \ 1137 (hdr)->b_dva.dva_word[1] == 0) 1138 1139 #define HDR_EQUAL(spa, dva, birth, hdr) \ 1140 ((hdr)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \ 1141 ((hdr)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \ 1142 ((hdr)->b_birth == birth) && ((hdr)->b_spa == spa) 1143 1144 static void 1145 buf_discard_identity(arc_buf_hdr_t *hdr) 1146 { 1147 hdr->b_dva.dva_word[0] = 0; 1148 hdr->b_dva.dva_word[1] = 0; 1149 hdr->b_birth = 0; 1150 } 1151 1152 static arc_buf_hdr_t * 1153 buf_hash_find(uint64_t spa, const blkptr_t *bp, kmutex_t **lockp) 1154 { 1155 const dva_t *dva = BP_IDENTITY(bp); 1156 uint64_t birth = BP_PHYSICAL_BIRTH(bp); 1157 uint64_t idx = BUF_HASH_INDEX(spa, dva, birth); 1158 kmutex_t *hash_lock = BUF_HASH_LOCK(idx); 1159 arc_buf_hdr_t *hdr; 1160 1161 mutex_enter(hash_lock); 1162 for (hdr = buf_hash_table.ht_table[idx]; hdr != NULL; 1163 hdr = hdr->b_hash_next) { 1164 if (HDR_EQUAL(spa, dva, birth, hdr)) { 1165 *lockp = hash_lock; 1166 return (hdr); 1167 } 1168 } 1169 mutex_exit(hash_lock); 1170 *lockp = NULL; 1171 return (NULL); 1172 } 1173 1174 /* 1175 * Insert an entry into the hash table. If there is already an element 1176 * equal to elem in the hash table, then the already existing element 1177 * will be returned and the new element will not be inserted. 1178 * Otherwise returns NULL. 1179 * If lockp == NULL, the caller is assumed to already hold the hash lock. 1180 */ 1181 static arc_buf_hdr_t * 1182 buf_hash_insert(arc_buf_hdr_t *hdr, kmutex_t **lockp) 1183 { 1184 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth); 1185 kmutex_t *hash_lock = BUF_HASH_LOCK(idx); 1186 arc_buf_hdr_t *fhdr; 1187 uint32_t i; 1188 1189 ASSERT(!DVA_IS_EMPTY(&hdr->b_dva)); 1190 ASSERT(hdr->b_birth != 0); 1191 ASSERT(!HDR_IN_HASH_TABLE(hdr)); 1192 1193 if (lockp != NULL) { 1194 *lockp = hash_lock; 1195 mutex_enter(hash_lock); 1196 } else { 1197 ASSERT(MUTEX_HELD(hash_lock)); 1198 } 1199 1200 for (fhdr = buf_hash_table.ht_table[idx], i = 0; fhdr != NULL; 1201 fhdr = fhdr->b_hash_next, i++) { 1202 if (HDR_EQUAL(hdr->b_spa, &hdr->b_dva, hdr->b_birth, fhdr)) 1203 return (fhdr); 1204 } 1205 1206 hdr->b_hash_next = buf_hash_table.ht_table[idx]; 1207 buf_hash_table.ht_table[idx] = hdr; 1208 arc_hdr_set_flags(hdr, ARC_FLAG_IN_HASH_TABLE); 1209 1210 /* collect some hash table performance data */ 1211 if (i > 0) { 1212 ARCSTAT_BUMP(arcstat_hash_collisions); 1213 if (i == 1) 1214 ARCSTAT_BUMP(arcstat_hash_chains); 1215 1216 ARCSTAT_MAX(arcstat_hash_chain_max, i); 1217 } 1218 1219 ARCSTAT_BUMP(arcstat_hash_elements); 1220 ARCSTAT_MAXSTAT(arcstat_hash_elements); 1221 1222 return (NULL); 1223 } 1224 1225 static void 1226 buf_hash_remove(arc_buf_hdr_t *hdr) 1227 { 1228 arc_buf_hdr_t *fhdr, **hdrp; 1229 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth); 1230 1231 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx))); 1232 ASSERT(HDR_IN_HASH_TABLE(hdr)); 1233 1234 hdrp = &buf_hash_table.ht_table[idx]; 1235 while ((fhdr = *hdrp) != hdr) { 1236 ASSERT3P(fhdr, !=, NULL); 1237 hdrp = &fhdr->b_hash_next; 1238 } 1239 *hdrp = hdr->b_hash_next; 1240 hdr->b_hash_next = NULL; 1241 arc_hdr_clear_flags(hdr, ARC_FLAG_IN_HASH_TABLE); 1242 1243 /* collect some hash table performance data */ 1244 ARCSTAT_BUMPDOWN(arcstat_hash_elements); 1245 1246 if (buf_hash_table.ht_table[idx] && 1247 buf_hash_table.ht_table[idx]->b_hash_next == NULL) 1248 ARCSTAT_BUMPDOWN(arcstat_hash_chains); 1249 } 1250 1251 /* 1252 * Global data structures and functions for the buf kmem cache. 1253 */ 1254 static kmem_cache_t *hdr_full_cache; 1255 static kmem_cache_t *hdr_l2only_cache; 1256 static kmem_cache_t *buf_cache; 1257 1258 static void 1259 buf_fini(void) 1260 { 1261 int i; 1262 1263 kmem_free(buf_hash_table.ht_table, 1264 (buf_hash_table.ht_mask + 1) * sizeof (void *)); 1265 for (i = 0; i < BUF_LOCKS; i++) 1266 mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock); 1267 kmem_cache_destroy(hdr_full_cache); 1268 kmem_cache_destroy(hdr_l2only_cache); 1269 kmem_cache_destroy(buf_cache); 1270 } 1271 1272 /* 1273 * Constructor callback - called when the cache is empty 1274 * and a new buf is requested. 1275 */ 1276 /* ARGSUSED */ 1277 static int 1278 hdr_full_cons(void *vbuf, void *unused, int kmflag) 1279 { 1280 arc_buf_hdr_t *hdr = vbuf; 1281 1282 bzero(hdr, HDR_FULL_SIZE); 1283 cv_init(&hdr->b_l1hdr.b_cv, NULL, CV_DEFAULT, NULL); 1284 refcount_create(&hdr->b_l1hdr.b_refcnt); 1285 mutex_init(&hdr->b_l1hdr.b_freeze_lock, NULL, MUTEX_DEFAULT, NULL); 1286 multilist_link_init(&hdr->b_l1hdr.b_arc_node); 1287 arc_space_consume(HDR_FULL_SIZE, ARC_SPACE_HDRS); 1288 1289 return (0); 1290 } 1291 1292 /* ARGSUSED */ 1293 static int 1294 hdr_l2only_cons(void *vbuf, void *unused, int kmflag) 1295 { 1296 arc_buf_hdr_t *hdr = vbuf; 1297 1298 bzero(hdr, HDR_L2ONLY_SIZE); 1299 arc_space_consume(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS); 1300 1301 return (0); 1302 } 1303 1304 /* ARGSUSED */ 1305 static int 1306 buf_cons(void *vbuf, void *unused, int kmflag) 1307 { 1308 arc_buf_t *buf = vbuf; 1309 1310 bzero(buf, sizeof (arc_buf_t)); 1311 mutex_init(&buf->b_evict_lock, NULL, MUTEX_DEFAULT, NULL); 1312 arc_space_consume(sizeof (arc_buf_t), ARC_SPACE_HDRS); 1313 1314 return (0); 1315 } 1316 1317 /* 1318 * Destructor callback - called when a cached buf is 1319 * no longer required. 1320 */ 1321 /* ARGSUSED */ 1322 static void 1323 hdr_full_dest(void *vbuf, void *unused) 1324 { 1325 arc_buf_hdr_t *hdr = vbuf; 1326 1327 ASSERT(HDR_EMPTY(hdr)); 1328 cv_destroy(&hdr->b_l1hdr.b_cv); 1329 refcount_destroy(&hdr->b_l1hdr.b_refcnt); 1330 mutex_destroy(&hdr->b_l1hdr.b_freeze_lock); 1331 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 1332 arc_space_return(HDR_FULL_SIZE, ARC_SPACE_HDRS); 1333 } 1334 1335 /* ARGSUSED */ 1336 static void 1337 hdr_l2only_dest(void *vbuf, void *unused) 1338 { 1339 arc_buf_hdr_t *hdr = vbuf; 1340 1341 ASSERT(HDR_EMPTY(hdr)); 1342 arc_space_return(HDR_L2ONLY_SIZE, ARC_SPACE_L2HDRS); 1343 } 1344 1345 /* ARGSUSED */ 1346 static void 1347 buf_dest(void *vbuf, void *unused) 1348 { 1349 arc_buf_t *buf = vbuf; 1350 1351 mutex_destroy(&buf->b_evict_lock); 1352 arc_space_return(sizeof (arc_buf_t), ARC_SPACE_HDRS); 1353 } 1354 1355 /* 1356 * Reclaim callback -- invoked when memory is low. 1357 */ 1358 /* ARGSUSED */ 1359 static void 1360 hdr_recl(void *unused) 1361 { 1362 dprintf("hdr_recl called\n"); 1363 /* 1364 * umem calls the reclaim func when we destroy the buf cache, 1365 * which is after we do arc_fini(). 1366 */ 1367 if (!arc_dead) 1368 cv_signal(&arc_reclaim_thread_cv); 1369 } 1370 1371 static void 1372 buf_init(void) 1373 { 1374 uint64_t *ct; 1375 uint64_t hsize = 1ULL << 12; 1376 int i, j; 1377 1378 /* 1379 * The hash table is big enough to fill all of physical memory 1380 * with an average block size of zfs_arc_average_blocksize (default 8K). 1381 * By default, the table will take up 1382 * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers). 1383 */ 1384 while (hsize * zfs_arc_average_blocksize < physmem * PAGESIZE) 1385 hsize <<= 1; 1386 retry: 1387 buf_hash_table.ht_mask = hsize - 1; 1388 buf_hash_table.ht_table = 1389 kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP); 1390 if (buf_hash_table.ht_table == NULL) { 1391 ASSERT(hsize > (1ULL << 8)); 1392 hsize >>= 1; 1393 goto retry; 1394 } 1395 1396 hdr_full_cache = kmem_cache_create("arc_buf_hdr_t_full", HDR_FULL_SIZE, 1397 0, hdr_full_cons, hdr_full_dest, hdr_recl, NULL, NULL, 0); 1398 hdr_l2only_cache = kmem_cache_create("arc_buf_hdr_t_l2only", 1399 HDR_L2ONLY_SIZE, 0, hdr_l2only_cons, hdr_l2only_dest, hdr_recl, 1400 NULL, NULL, 0); 1401 buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t), 1402 0, buf_cons, buf_dest, NULL, NULL, NULL, 0); 1403 1404 for (i = 0; i < 256; i++) 1405 for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--) 1406 *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY); 1407 1408 for (i = 0; i < BUF_LOCKS; i++) { 1409 mutex_init(&buf_hash_table.ht_locks[i].ht_lock, 1410 NULL, MUTEX_DEFAULT, NULL); 1411 } 1412 } 1413 1414 /* 1415 * This is the size that the buf occupies in memory. If the buf is compressed, 1416 * it will correspond to the compressed size. You should use this method of 1417 * getting the buf size unless you explicitly need the logical size. 1418 */ 1419 int32_t 1420 arc_buf_size(arc_buf_t *buf) 1421 { 1422 return (ARC_BUF_COMPRESSED(buf) ? 1423 HDR_GET_PSIZE(buf->b_hdr) : HDR_GET_LSIZE(buf->b_hdr)); 1424 } 1425 1426 int32_t 1427 arc_buf_lsize(arc_buf_t *buf) 1428 { 1429 return (HDR_GET_LSIZE(buf->b_hdr)); 1430 } 1431 1432 enum zio_compress 1433 arc_get_compression(arc_buf_t *buf) 1434 { 1435 return (ARC_BUF_COMPRESSED(buf) ? 1436 HDR_GET_COMPRESS(buf->b_hdr) : ZIO_COMPRESS_OFF); 1437 } 1438 1439 #define ARC_MINTIME (hz>>4) /* 62 ms */ 1440 1441 static inline boolean_t 1442 arc_buf_is_shared(arc_buf_t *buf) 1443 { 1444 boolean_t shared = (buf->b_data != NULL && 1445 buf->b_data == buf->b_hdr->b_l1hdr.b_pdata); 1446 IMPLY(shared, HDR_SHARED_DATA(buf->b_hdr)); 1447 IMPLY(shared, ARC_BUF_SHARED(buf)); 1448 IMPLY(shared, ARC_BUF_COMPRESSED(buf) || ARC_BUF_LAST(buf)); 1449 1450 /* 1451 * It would be nice to assert arc_can_share() too, but the "hdr isn't 1452 * already being shared" requirement prevents us from doing that. 1453 */ 1454 1455 return (shared); 1456 } 1457 1458 /* 1459 * Free the checksum associated with this header. If there is no checksum, this 1460 * is a no-op. 1461 */ 1462 static inline void 1463 arc_cksum_free(arc_buf_hdr_t *hdr) 1464 { 1465 ASSERT(HDR_HAS_L1HDR(hdr)); 1466 mutex_enter(&hdr->b_l1hdr.b_freeze_lock); 1467 if (hdr->b_l1hdr.b_freeze_cksum != NULL) { 1468 kmem_free(hdr->b_l1hdr.b_freeze_cksum, sizeof (zio_cksum_t)); 1469 hdr->b_l1hdr.b_freeze_cksum = NULL; 1470 } 1471 mutex_exit(&hdr->b_l1hdr.b_freeze_lock); 1472 } 1473 1474 /* 1475 * Return true iff at least one of the bufs on hdr is not compressed. 1476 */ 1477 static boolean_t 1478 arc_hdr_has_uncompressed_buf(arc_buf_hdr_t *hdr) 1479 { 1480 for (arc_buf_t *b = hdr->b_l1hdr.b_buf; b != NULL; b = b->b_next) { 1481 if (!ARC_BUF_COMPRESSED(b)) { 1482 return (B_TRUE); 1483 } 1484 } 1485 return (B_FALSE); 1486 } 1487 1488 /* 1489 * If we've turned on the ZFS_DEBUG_MODIFY flag, verify that the buf's data 1490 * matches the checksum that is stored in the hdr. If there is no checksum, 1491 * or if the buf is compressed, this is a no-op. 1492 */ 1493 static void 1494 arc_cksum_verify(arc_buf_t *buf) 1495 { 1496 arc_buf_hdr_t *hdr = buf->b_hdr; 1497 zio_cksum_t zc; 1498 1499 if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 1500 return; 1501 1502 if (ARC_BUF_COMPRESSED(buf)) { 1503 ASSERT(hdr->b_l1hdr.b_freeze_cksum == NULL || 1504 arc_hdr_has_uncompressed_buf(hdr)); 1505 return; 1506 } 1507 1508 ASSERT(HDR_HAS_L1HDR(hdr)); 1509 1510 mutex_enter(&hdr->b_l1hdr.b_freeze_lock); 1511 if (hdr->b_l1hdr.b_freeze_cksum == NULL || HDR_IO_ERROR(hdr)) { 1512 mutex_exit(&hdr->b_l1hdr.b_freeze_lock); 1513 return; 1514 } 1515 1516 fletcher_2_native(buf->b_data, arc_buf_size(buf), NULL, &zc); 1517 if (!ZIO_CHECKSUM_EQUAL(*hdr->b_l1hdr.b_freeze_cksum, zc)) 1518 panic("buffer modified while frozen!"); 1519 mutex_exit(&hdr->b_l1hdr.b_freeze_lock); 1520 } 1521 1522 static boolean_t 1523 arc_cksum_is_equal(arc_buf_hdr_t *hdr, zio_t *zio) 1524 { 1525 enum zio_compress compress = BP_GET_COMPRESS(zio->io_bp); 1526 boolean_t valid_cksum; 1527 1528 ASSERT(!BP_IS_EMBEDDED(zio->io_bp)); 1529 VERIFY3U(BP_GET_PSIZE(zio->io_bp), ==, HDR_GET_PSIZE(hdr)); 1530 1531 /* 1532 * We rely on the blkptr's checksum to determine if the block 1533 * is valid or not. When compressed arc is enabled, the l2arc 1534 * writes the block to the l2arc just as it appears in the pool. 1535 * This allows us to use the blkptr's checksum to validate the 1536 * data that we just read off of the l2arc without having to store 1537 * a separate checksum in the arc_buf_hdr_t. However, if compressed 1538 * arc is disabled, then the data written to the l2arc is always 1539 * uncompressed and won't match the block as it exists in the main 1540 * pool. When this is the case, we must first compress it if it is 1541 * compressed on the main pool before we can validate the checksum. 1542 */ 1543 if (!HDR_COMPRESSION_ENABLED(hdr) && compress != ZIO_COMPRESS_OFF) { 1544 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF); 1545 uint64_t lsize = HDR_GET_LSIZE(hdr); 1546 uint64_t csize; 1547 1548 void *cbuf = zio_buf_alloc(HDR_GET_PSIZE(hdr)); 1549 csize = zio_compress_data(compress, zio->io_data, cbuf, lsize); 1550 ASSERT3U(csize, <=, HDR_GET_PSIZE(hdr)); 1551 if (csize < HDR_GET_PSIZE(hdr)) { 1552 /* 1553 * Compressed blocks are always a multiple of the 1554 * smallest ashift in the pool. Ideally, we would 1555 * like to round up the csize to the next 1556 * spa_min_ashift but that value may have changed 1557 * since the block was last written. Instead, 1558 * we rely on the fact that the hdr's psize 1559 * was set to the psize of the block when it was 1560 * last written. We set the csize to that value 1561 * and zero out any part that should not contain 1562 * data. 1563 */ 1564 bzero((char *)cbuf + csize, HDR_GET_PSIZE(hdr) - csize); 1565 csize = HDR_GET_PSIZE(hdr); 1566 } 1567 zio_push_transform(zio, cbuf, csize, HDR_GET_PSIZE(hdr), NULL); 1568 } 1569 1570 /* 1571 * Block pointers always store the checksum for the logical data. 1572 * If the block pointer has the gang bit set, then the checksum 1573 * it represents is for the reconstituted data and not for an 1574 * individual gang member. The zio pipeline, however, must be able to 1575 * determine the checksum of each of the gang constituents so it 1576 * treats the checksum comparison differently than what we need 1577 * for l2arc blocks. This prevents us from using the 1578 * zio_checksum_error() interface directly. Instead we must call the 1579 * zio_checksum_error_impl() so that we can ensure the checksum is 1580 * generated using the correct checksum algorithm and accounts for the 1581 * logical I/O size and not just a gang fragment. 1582 */ 1583 valid_cksum = (zio_checksum_error_impl(zio->io_spa, zio->io_bp, 1584 BP_GET_CHECKSUM(zio->io_bp), zio->io_data, zio->io_size, 1585 zio->io_offset, NULL) == 0); 1586 zio_pop_transforms(zio); 1587 return (valid_cksum); 1588 } 1589 1590 /* 1591 * Given a buf full of data, if ZFS_DEBUG_MODIFY is enabled this computes a 1592 * checksum and attaches it to the buf's hdr so that we can ensure that the buf 1593 * isn't modified later on. If buf is compressed or there is already a checksum 1594 * on the hdr, this is a no-op (we only checksum uncompressed bufs). 1595 */ 1596 static void 1597 arc_cksum_compute(arc_buf_t *buf) 1598 { 1599 arc_buf_hdr_t *hdr = buf->b_hdr; 1600 1601 if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 1602 return; 1603 1604 ASSERT(HDR_HAS_L1HDR(hdr)); 1605 1606 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock); 1607 if (hdr->b_l1hdr.b_freeze_cksum != NULL) { 1608 ASSERT(arc_hdr_has_uncompressed_buf(hdr)); 1609 mutex_exit(&hdr->b_l1hdr.b_freeze_lock); 1610 return; 1611 } else if (ARC_BUF_COMPRESSED(buf)) { 1612 mutex_exit(&hdr->b_l1hdr.b_freeze_lock); 1613 return; 1614 } 1615 1616 ASSERT(!ARC_BUF_COMPRESSED(buf)); 1617 hdr->b_l1hdr.b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t), 1618 KM_SLEEP); 1619 fletcher_2_native(buf->b_data, arc_buf_size(buf), NULL, 1620 hdr->b_l1hdr.b_freeze_cksum); 1621 mutex_exit(&hdr->b_l1hdr.b_freeze_lock); 1622 arc_buf_watch(buf); 1623 } 1624 1625 #ifndef _KERNEL 1626 typedef struct procctl { 1627 long cmd; 1628 prwatch_t prwatch; 1629 } procctl_t; 1630 #endif 1631 1632 /* ARGSUSED */ 1633 static void 1634 arc_buf_unwatch(arc_buf_t *buf) 1635 { 1636 #ifndef _KERNEL 1637 if (arc_watch) { 1638 int result; 1639 procctl_t ctl; 1640 ctl.cmd = PCWATCH; 1641 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data; 1642 ctl.prwatch.pr_size = 0; 1643 ctl.prwatch.pr_wflags = 0; 1644 result = write(arc_procfd, &ctl, sizeof (ctl)); 1645 ASSERT3U(result, ==, sizeof (ctl)); 1646 } 1647 #endif 1648 } 1649 1650 /* ARGSUSED */ 1651 static void 1652 arc_buf_watch(arc_buf_t *buf) 1653 { 1654 #ifndef _KERNEL 1655 if (arc_watch) { 1656 int result; 1657 procctl_t ctl; 1658 ctl.cmd = PCWATCH; 1659 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data; 1660 ctl.prwatch.pr_size = arc_buf_size(buf); 1661 ctl.prwatch.pr_wflags = WA_WRITE; 1662 result = write(arc_procfd, &ctl, sizeof (ctl)); 1663 ASSERT3U(result, ==, sizeof (ctl)); 1664 } 1665 #endif 1666 } 1667 1668 static arc_buf_contents_t 1669 arc_buf_type(arc_buf_hdr_t *hdr) 1670 { 1671 arc_buf_contents_t type; 1672 if (HDR_ISTYPE_METADATA(hdr)) { 1673 type = ARC_BUFC_METADATA; 1674 } else { 1675 type = ARC_BUFC_DATA; 1676 } 1677 VERIFY3U(hdr->b_type, ==, type); 1678 return (type); 1679 } 1680 1681 boolean_t 1682 arc_is_metadata(arc_buf_t *buf) 1683 { 1684 return (HDR_ISTYPE_METADATA(buf->b_hdr) != 0); 1685 } 1686 1687 static uint32_t 1688 arc_bufc_to_flags(arc_buf_contents_t type) 1689 { 1690 switch (type) { 1691 case ARC_BUFC_DATA: 1692 /* metadata field is 0 if buffer contains normal data */ 1693 return (0); 1694 case ARC_BUFC_METADATA: 1695 return (ARC_FLAG_BUFC_METADATA); 1696 default: 1697 break; 1698 } 1699 panic("undefined ARC buffer type!"); 1700 return ((uint32_t)-1); 1701 } 1702 1703 void 1704 arc_buf_thaw(arc_buf_t *buf) 1705 { 1706 arc_buf_hdr_t *hdr = buf->b_hdr; 1707 1708 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon); 1709 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 1710 1711 arc_cksum_verify(buf); 1712 1713 /* 1714 * Compressed buffers do not manipulate the b_freeze_cksum or 1715 * allocate b_thawed. 1716 */ 1717 if (ARC_BUF_COMPRESSED(buf)) { 1718 ASSERT(hdr->b_l1hdr.b_freeze_cksum == NULL || 1719 arc_hdr_has_uncompressed_buf(hdr)); 1720 return; 1721 } 1722 1723 ASSERT(HDR_HAS_L1HDR(hdr)); 1724 arc_cksum_free(hdr); 1725 1726 mutex_enter(&hdr->b_l1hdr.b_freeze_lock); 1727 #ifdef ZFS_DEBUG 1728 if (zfs_flags & ZFS_DEBUG_MODIFY) { 1729 if (hdr->b_l1hdr.b_thawed != NULL) 1730 kmem_free(hdr->b_l1hdr.b_thawed, 1); 1731 hdr->b_l1hdr.b_thawed = kmem_alloc(1, KM_SLEEP); 1732 } 1733 #endif 1734 1735 mutex_exit(&hdr->b_l1hdr.b_freeze_lock); 1736 1737 arc_buf_unwatch(buf); 1738 } 1739 1740 void 1741 arc_buf_freeze(arc_buf_t *buf) 1742 { 1743 arc_buf_hdr_t *hdr = buf->b_hdr; 1744 kmutex_t *hash_lock; 1745 1746 if (!(zfs_flags & ZFS_DEBUG_MODIFY)) 1747 return; 1748 1749 if (ARC_BUF_COMPRESSED(buf)) { 1750 ASSERT(hdr->b_l1hdr.b_freeze_cksum == NULL || 1751 arc_hdr_has_uncompressed_buf(hdr)); 1752 return; 1753 } 1754 1755 hash_lock = HDR_LOCK(hdr); 1756 mutex_enter(hash_lock); 1757 1758 ASSERT(HDR_HAS_L1HDR(hdr)); 1759 ASSERT(hdr->b_l1hdr.b_freeze_cksum != NULL || 1760 hdr->b_l1hdr.b_state == arc_anon); 1761 arc_cksum_compute(buf); 1762 mutex_exit(hash_lock); 1763 } 1764 1765 /* 1766 * The arc_buf_hdr_t's b_flags should never be modified directly. Instead, 1767 * the following functions should be used to ensure that the flags are 1768 * updated in a thread-safe way. When manipulating the flags either 1769 * the hash_lock must be held or the hdr must be undiscoverable. This 1770 * ensures that we're not racing with any other threads when updating 1771 * the flags. 1772 */ 1773 static inline void 1774 arc_hdr_set_flags(arc_buf_hdr_t *hdr, arc_flags_t flags) 1775 { 1776 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); 1777 hdr->b_flags |= flags; 1778 } 1779 1780 static inline void 1781 arc_hdr_clear_flags(arc_buf_hdr_t *hdr, arc_flags_t flags) 1782 { 1783 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); 1784 hdr->b_flags &= ~flags; 1785 } 1786 1787 /* 1788 * Setting the compression bits in the arc_buf_hdr_t's b_flags is 1789 * done in a special way since we have to clear and set bits 1790 * at the same time. Consumers that wish to set the compression bits 1791 * must use this function to ensure that the flags are updated in 1792 * thread-safe manner. 1793 */ 1794 static void 1795 arc_hdr_set_compress(arc_buf_hdr_t *hdr, enum zio_compress cmp) 1796 { 1797 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); 1798 1799 /* 1800 * Holes and embedded blocks will always have a psize = 0 so 1801 * we ignore the compression of the blkptr and set the 1802 * arc_buf_hdr_t's compression to ZIO_COMPRESS_OFF. 1803 * Holes and embedded blocks remain anonymous so we don't 1804 * want to uncompress them. Mark them as uncompressed. 1805 */ 1806 if (!zfs_compressed_arc_enabled || HDR_GET_PSIZE(hdr) == 0) { 1807 arc_hdr_clear_flags(hdr, ARC_FLAG_COMPRESSED_ARC); 1808 HDR_SET_COMPRESS(hdr, ZIO_COMPRESS_OFF); 1809 ASSERT(!HDR_COMPRESSION_ENABLED(hdr)); 1810 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF); 1811 } else { 1812 arc_hdr_set_flags(hdr, ARC_FLAG_COMPRESSED_ARC); 1813 HDR_SET_COMPRESS(hdr, cmp); 1814 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, cmp); 1815 ASSERT(HDR_COMPRESSION_ENABLED(hdr)); 1816 } 1817 } 1818 1819 /* 1820 * Looks for another buf on the same hdr which has the data decompressed, copies 1821 * from it, and returns true. If no such buf exists, returns false. 1822 */ 1823 static boolean_t 1824 arc_buf_try_copy_decompressed_data(arc_buf_t *buf) 1825 { 1826 arc_buf_hdr_t *hdr = buf->b_hdr; 1827 boolean_t copied = B_FALSE; 1828 1829 ASSERT(HDR_HAS_L1HDR(hdr)); 1830 ASSERT3P(buf->b_data, !=, NULL); 1831 ASSERT(!ARC_BUF_COMPRESSED(buf)); 1832 1833 for (arc_buf_t *from = hdr->b_l1hdr.b_buf; from != NULL; 1834 from = from->b_next) { 1835 /* can't use our own data buffer */ 1836 if (from == buf) { 1837 continue; 1838 } 1839 1840 if (!ARC_BUF_COMPRESSED(from)) { 1841 bcopy(from->b_data, buf->b_data, arc_buf_size(buf)); 1842 copied = B_TRUE; 1843 break; 1844 } 1845 } 1846 1847 /* 1848 * There were no decompressed bufs, so there should not be a 1849 * checksum on the hdr either. 1850 */ 1851 EQUIV(!copied, hdr->b_l1hdr.b_freeze_cksum == NULL); 1852 1853 return (copied); 1854 } 1855 1856 /* 1857 * Given a buf that has a data buffer attached to it, this function will 1858 * efficiently fill the buf with data of the specified compression setting from 1859 * the hdr and update the hdr's b_freeze_cksum if necessary. If the buf and hdr 1860 * are already sharing a data buf, no copy is performed. 1861 * 1862 * If the buf is marked as compressed but uncompressed data was requested, this 1863 * will allocate a new data buffer for the buf, remove that flag, and fill the 1864 * buf with uncompressed data. You can't request a compressed buf on a hdr with 1865 * uncompressed data, and (since we haven't added support for it yet) if you 1866 * want compressed data your buf must already be marked as compressed and have 1867 * the correct-sized data buffer. 1868 */ 1869 static int 1870 arc_buf_fill(arc_buf_t *buf, boolean_t compressed) 1871 { 1872 arc_buf_hdr_t *hdr = buf->b_hdr; 1873 boolean_t hdr_compressed = (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF); 1874 dmu_object_byteswap_t bswap = hdr->b_l1hdr.b_byteswap; 1875 1876 ASSERT3P(buf->b_data, !=, NULL); 1877 IMPLY(compressed, hdr_compressed); 1878 IMPLY(compressed, ARC_BUF_COMPRESSED(buf)); 1879 1880 if (hdr_compressed == compressed) { 1881 if (!arc_buf_is_shared(buf)) { 1882 bcopy(hdr->b_l1hdr.b_pdata, buf->b_data, 1883 arc_buf_size(buf)); 1884 } 1885 } else { 1886 ASSERT(hdr_compressed); 1887 ASSERT(!compressed); 1888 ASSERT3U(HDR_GET_LSIZE(hdr), !=, HDR_GET_PSIZE(hdr)); 1889 1890 /* 1891 * If the buf is sharing its data with the hdr, unlink it and 1892 * allocate a new data buffer for the buf. 1893 */ 1894 if (arc_buf_is_shared(buf)) { 1895 ASSERT(ARC_BUF_COMPRESSED(buf)); 1896 1897 /* We need to give the buf it's own b_data */ 1898 buf->b_flags &= ~ARC_BUF_FLAG_SHARED; 1899 buf->b_data = 1900 arc_get_data_buf(hdr, HDR_GET_LSIZE(hdr), buf); 1901 arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA); 1902 1903 /* Previously overhead was 0; just add new overhead */ 1904 ARCSTAT_INCR(arcstat_overhead_size, HDR_GET_LSIZE(hdr)); 1905 } else if (ARC_BUF_COMPRESSED(buf)) { 1906 /* We need to reallocate the buf's b_data */ 1907 arc_free_data_buf(hdr, buf->b_data, HDR_GET_PSIZE(hdr), 1908 buf); 1909 buf->b_data = 1910 arc_get_data_buf(hdr, HDR_GET_LSIZE(hdr), buf); 1911 1912 /* We increased the size of b_data; update overhead */ 1913 ARCSTAT_INCR(arcstat_overhead_size, 1914 HDR_GET_LSIZE(hdr) - HDR_GET_PSIZE(hdr)); 1915 } 1916 1917 /* 1918 * Regardless of the buf's previous compression settings, it 1919 * should not be compressed at the end of this function. 1920 */ 1921 buf->b_flags &= ~ARC_BUF_FLAG_COMPRESSED; 1922 1923 /* 1924 * Try copying the data from another buf which already has a 1925 * decompressed version. If that's not possible, it's time to 1926 * bite the bullet and decompress the data from the hdr. 1927 */ 1928 if (arc_buf_try_copy_decompressed_data(buf)) { 1929 /* Skip byteswapping and checksumming (already done) */ 1930 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, !=, NULL); 1931 return (0); 1932 } else { 1933 int error = zio_decompress_data(HDR_GET_COMPRESS(hdr), 1934 hdr->b_l1hdr.b_pdata, buf->b_data, 1935 HDR_GET_PSIZE(hdr), HDR_GET_LSIZE(hdr)); 1936 1937 /* 1938 * Absent hardware errors or software bugs, this should 1939 * be impossible, but log it anyway so we can debug it. 1940 */ 1941 if (error != 0) { 1942 zfs_dbgmsg( 1943 "hdr %p, compress %d, psize %d, lsize %d", 1944 hdr, HDR_GET_COMPRESS(hdr), 1945 HDR_GET_PSIZE(hdr), HDR_GET_LSIZE(hdr)); 1946 return (SET_ERROR(EIO)); 1947 } 1948 } 1949 } 1950 1951 /* Byteswap the buf's data if necessary */ 1952 if (bswap != DMU_BSWAP_NUMFUNCS) { 1953 ASSERT(!HDR_SHARED_DATA(hdr)); 1954 ASSERT3U(bswap, <, DMU_BSWAP_NUMFUNCS); 1955 dmu_ot_byteswap[bswap].ob_func(buf->b_data, HDR_GET_LSIZE(hdr)); 1956 } 1957 1958 /* Compute the hdr's checksum if necessary */ 1959 arc_cksum_compute(buf); 1960 1961 return (0); 1962 } 1963 1964 int 1965 arc_decompress(arc_buf_t *buf) 1966 { 1967 return (arc_buf_fill(buf, B_FALSE)); 1968 } 1969 1970 /* 1971 * Return the size of the block, b_pdata, that is stored in the arc_buf_hdr_t. 1972 */ 1973 static uint64_t 1974 arc_hdr_size(arc_buf_hdr_t *hdr) 1975 { 1976 uint64_t size; 1977 1978 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF && 1979 HDR_GET_PSIZE(hdr) > 0) { 1980 size = HDR_GET_PSIZE(hdr); 1981 } else { 1982 ASSERT3U(HDR_GET_LSIZE(hdr), !=, 0); 1983 size = HDR_GET_LSIZE(hdr); 1984 } 1985 return (size); 1986 } 1987 1988 /* 1989 * Increment the amount of evictable space in the arc_state_t's refcount. 1990 * We account for the space used by the hdr and the arc buf individually 1991 * so that we can add and remove them from the refcount individually. 1992 */ 1993 static void 1994 arc_evictable_space_increment(arc_buf_hdr_t *hdr, arc_state_t *state) 1995 { 1996 arc_buf_contents_t type = arc_buf_type(hdr); 1997 1998 ASSERT(HDR_HAS_L1HDR(hdr)); 1999 2000 if (GHOST_STATE(state)) { 2001 ASSERT0(hdr->b_l1hdr.b_bufcnt); 2002 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 2003 ASSERT3P(hdr->b_l1hdr.b_pdata, ==, NULL); 2004 (void) refcount_add_many(&state->arcs_esize[type], 2005 HDR_GET_LSIZE(hdr), hdr); 2006 return; 2007 } 2008 2009 ASSERT(!GHOST_STATE(state)); 2010 if (hdr->b_l1hdr.b_pdata != NULL) { 2011 (void) refcount_add_many(&state->arcs_esize[type], 2012 arc_hdr_size(hdr), hdr); 2013 } 2014 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL; 2015 buf = buf->b_next) { 2016 if (arc_buf_is_shared(buf)) 2017 continue; 2018 (void) refcount_add_many(&state->arcs_esize[type], 2019 arc_buf_size(buf), buf); 2020 } 2021 } 2022 2023 /* 2024 * Decrement the amount of evictable space in the arc_state_t's refcount. 2025 * We account for the space used by the hdr and the arc buf individually 2026 * so that we can add and remove them from the refcount individually. 2027 */ 2028 static void 2029 arc_evictable_space_decrement(arc_buf_hdr_t *hdr, arc_state_t *state) 2030 { 2031 arc_buf_contents_t type = arc_buf_type(hdr); 2032 2033 ASSERT(HDR_HAS_L1HDR(hdr)); 2034 2035 if (GHOST_STATE(state)) { 2036 ASSERT0(hdr->b_l1hdr.b_bufcnt); 2037 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 2038 ASSERT3P(hdr->b_l1hdr.b_pdata, ==, NULL); 2039 (void) refcount_remove_many(&state->arcs_esize[type], 2040 HDR_GET_LSIZE(hdr), hdr); 2041 return; 2042 } 2043 2044 ASSERT(!GHOST_STATE(state)); 2045 if (hdr->b_l1hdr.b_pdata != NULL) { 2046 (void) refcount_remove_many(&state->arcs_esize[type], 2047 arc_hdr_size(hdr), hdr); 2048 } 2049 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL; 2050 buf = buf->b_next) { 2051 if (arc_buf_is_shared(buf)) 2052 continue; 2053 (void) refcount_remove_many(&state->arcs_esize[type], 2054 arc_buf_size(buf), buf); 2055 } 2056 } 2057 2058 /* 2059 * Add a reference to this hdr indicating that someone is actively 2060 * referencing that memory. When the refcount transitions from 0 to 1, 2061 * we remove it from the respective arc_state_t list to indicate that 2062 * it is not evictable. 2063 */ 2064 static void 2065 add_reference(arc_buf_hdr_t *hdr, void *tag) 2066 { 2067 ASSERT(HDR_HAS_L1HDR(hdr)); 2068 if (!MUTEX_HELD(HDR_LOCK(hdr))) { 2069 ASSERT(hdr->b_l1hdr.b_state == arc_anon); 2070 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 2071 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 2072 } 2073 2074 arc_state_t *state = hdr->b_l1hdr.b_state; 2075 2076 if ((refcount_add(&hdr->b_l1hdr.b_refcnt, tag) == 1) && 2077 (state != arc_anon)) { 2078 /* We don't use the L2-only state list. */ 2079 if (state != arc_l2c_only) { 2080 multilist_remove(&state->arcs_list[arc_buf_type(hdr)], 2081 hdr); 2082 arc_evictable_space_decrement(hdr, state); 2083 } 2084 /* remove the prefetch flag if we get a reference */ 2085 arc_hdr_clear_flags(hdr, ARC_FLAG_PREFETCH); 2086 } 2087 } 2088 2089 /* 2090 * Remove a reference from this hdr. When the reference transitions from 2091 * 1 to 0 and we're not anonymous, then we add this hdr to the arc_state_t's 2092 * list making it eligible for eviction. 2093 */ 2094 static int 2095 remove_reference(arc_buf_hdr_t *hdr, kmutex_t *hash_lock, void *tag) 2096 { 2097 int cnt; 2098 arc_state_t *state = hdr->b_l1hdr.b_state; 2099 2100 ASSERT(HDR_HAS_L1HDR(hdr)); 2101 ASSERT(state == arc_anon || MUTEX_HELD(hash_lock)); 2102 ASSERT(!GHOST_STATE(state)); 2103 2104 /* 2105 * arc_l2c_only counts as a ghost state so we don't need to explicitly 2106 * check to prevent usage of the arc_l2c_only list. 2107 */ 2108 if (((cnt = refcount_remove(&hdr->b_l1hdr.b_refcnt, tag)) == 0) && 2109 (state != arc_anon)) { 2110 multilist_insert(&state->arcs_list[arc_buf_type(hdr)], hdr); 2111 ASSERT3U(hdr->b_l1hdr.b_bufcnt, >, 0); 2112 arc_evictable_space_increment(hdr, state); 2113 } 2114 return (cnt); 2115 } 2116 2117 /* 2118 * Move the supplied buffer to the indicated state. The hash lock 2119 * for the buffer must be held by the caller. 2120 */ 2121 static void 2122 arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *hdr, 2123 kmutex_t *hash_lock) 2124 { 2125 arc_state_t *old_state; 2126 int64_t refcnt; 2127 uint32_t bufcnt; 2128 boolean_t update_old, update_new; 2129 arc_buf_contents_t buftype = arc_buf_type(hdr); 2130 2131 /* 2132 * We almost always have an L1 hdr here, since we call arc_hdr_realloc() 2133 * in arc_read() when bringing a buffer out of the L2ARC. However, the 2134 * L1 hdr doesn't always exist when we change state to arc_anon before 2135 * destroying a header, in which case reallocating to add the L1 hdr is 2136 * pointless. 2137 */ 2138 if (HDR_HAS_L1HDR(hdr)) { 2139 old_state = hdr->b_l1hdr.b_state; 2140 refcnt = refcount_count(&hdr->b_l1hdr.b_refcnt); 2141 bufcnt = hdr->b_l1hdr.b_bufcnt; 2142 update_old = (bufcnt > 0 || hdr->b_l1hdr.b_pdata != NULL); 2143 } else { 2144 old_state = arc_l2c_only; 2145 refcnt = 0; 2146 bufcnt = 0; 2147 update_old = B_FALSE; 2148 } 2149 update_new = update_old; 2150 2151 ASSERT(MUTEX_HELD(hash_lock)); 2152 ASSERT3P(new_state, !=, old_state); 2153 ASSERT(!GHOST_STATE(new_state) || bufcnt == 0); 2154 ASSERT(old_state != arc_anon || bufcnt <= 1); 2155 2156 /* 2157 * If this buffer is evictable, transfer it from the 2158 * old state list to the new state list. 2159 */ 2160 if (refcnt == 0) { 2161 if (old_state != arc_anon && old_state != arc_l2c_only) { 2162 ASSERT(HDR_HAS_L1HDR(hdr)); 2163 multilist_remove(&old_state->arcs_list[buftype], hdr); 2164 2165 if (GHOST_STATE(old_state)) { 2166 ASSERT0(bufcnt); 2167 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 2168 update_old = B_TRUE; 2169 } 2170 arc_evictable_space_decrement(hdr, old_state); 2171 } 2172 if (new_state != arc_anon && new_state != arc_l2c_only) { 2173 2174 /* 2175 * An L1 header always exists here, since if we're 2176 * moving to some L1-cached state (i.e. not l2c_only or 2177 * anonymous), we realloc the header to add an L1hdr 2178 * beforehand. 2179 */ 2180 ASSERT(HDR_HAS_L1HDR(hdr)); 2181 multilist_insert(&new_state->arcs_list[buftype], hdr); 2182 2183 if (GHOST_STATE(new_state)) { 2184 ASSERT0(bufcnt); 2185 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 2186 update_new = B_TRUE; 2187 } 2188 arc_evictable_space_increment(hdr, new_state); 2189 } 2190 } 2191 2192 ASSERT(!HDR_EMPTY(hdr)); 2193 if (new_state == arc_anon && HDR_IN_HASH_TABLE(hdr)) 2194 buf_hash_remove(hdr); 2195 2196 /* adjust state sizes (ignore arc_l2c_only) */ 2197 2198 if (update_new && new_state != arc_l2c_only) { 2199 ASSERT(HDR_HAS_L1HDR(hdr)); 2200 if (GHOST_STATE(new_state)) { 2201 ASSERT0(bufcnt); 2202 2203 /* 2204 * When moving a header to a ghost state, we first 2205 * remove all arc buffers. Thus, we'll have a 2206 * bufcnt of zero, and no arc buffer to use for 2207 * the reference. As a result, we use the arc 2208 * header pointer for the reference. 2209 */ 2210 (void) refcount_add_many(&new_state->arcs_size, 2211 HDR_GET_LSIZE(hdr), hdr); 2212 ASSERT3P(hdr->b_l1hdr.b_pdata, ==, NULL); 2213 } else { 2214 uint32_t buffers = 0; 2215 2216 /* 2217 * Each individual buffer holds a unique reference, 2218 * thus we must remove each of these references one 2219 * at a time. 2220 */ 2221 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL; 2222 buf = buf->b_next) { 2223 ASSERT3U(bufcnt, !=, 0); 2224 buffers++; 2225 2226 /* 2227 * When the arc_buf_t is sharing the data 2228 * block with the hdr, the owner of the 2229 * reference belongs to the hdr. Only 2230 * add to the refcount if the arc_buf_t is 2231 * not shared. 2232 */ 2233 if (arc_buf_is_shared(buf)) 2234 continue; 2235 2236 (void) refcount_add_many(&new_state->arcs_size, 2237 arc_buf_size(buf), buf); 2238 } 2239 ASSERT3U(bufcnt, ==, buffers); 2240 2241 if (hdr->b_l1hdr.b_pdata != NULL) { 2242 (void) refcount_add_many(&new_state->arcs_size, 2243 arc_hdr_size(hdr), hdr); 2244 } else { 2245 ASSERT(GHOST_STATE(old_state)); 2246 } 2247 } 2248 } 2249 2250 if (update_old && old_state != arc_l2c_only) { 2251 ASSERT(HDR_HAS_L1HDR(hdr)); 2252 if (GHOST_STATE(old_state)) { 2253 ASSERT0(bufcnt); 2254 ASSERT3P(hdr->b_l1hdr.b_pdata, ==, NULL); 2255 2256 /* 2257 * When moving a header off of a ghost state, 2258 * the header will not contain any arc buffers. 2259 * We use the arc header pointer for the reference 2260 * which is exactly what we did when we put the 2261 * header on the ghost state. 2262 */ 2263 2264 (void) refcount_remove_many(&old_state->arcs_size, 2265 HDR_GET_LSIZE(hdr), hdr); 2266 } else { 2267 uint32_t buffers = 0; 2268 2269 /* 2270 * Each individual buffer holds a unique reference, 2271 * thus we must remove each of these references one 2272 * at a time. 2273 */ 2274 for (arc_buf_t *buf = hdr->b_l1hdr.b_buf; buf != NULL; 2275 buf = buf->b_next) { 2276 ASSERT3U(bufcnt, !=, 0); 2277 buffers++; 2278 2279 /* 2280 * When the arc_buf_t is sharing the data 2281 * block with the hdr, the owner of the 2282 * reference belongs to the hdr. Only 2283 * add to the refcount if the arc_buf_t is 2284 * not shared. 2285 */ 2286 if (arc_buf_is_shared(buf)) 2287 continue; 2288 2289 (void) refcount_remove_many( 2290 &old_state->arcs_size, arc_buf_size(buf), 2291 buf); 2292 } 2293 ASSERT3U(bufcnt, ==, buffers); 2294 ASSERT3P(hdr->b_l1hdr.b_pdata, !=, NULL); 2295 (void) refcount_remove_many( 2296 &old_state->arcs_size, arc_hdr_size(hdr), hdr); 2297 } 2298 } 2299 2300 if (HDR_HAS_L1HDR(hdr)) 2301 hdr->b_l1hdr.b_state = new_state; 2302 2303 /* 2304 * L2 headers should never be on the L2 state list since they don't 2305 * have L1 headers allocated. 2306 */ 2307 ASSERT(multilist_is_empty(&arc_l2c_only->arcs_list[ARC_BUFC_DATA]) && 2308 multilist_is_empty(&arc_l2c_only->arcs_list[ARC_BUFC_METADATA])); 2309 } 2310 2311 void 2312 arc_space_consume(uint64_t space, arc_space_type_t type) 2313 { 2314 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES); 2315 2316 switch (type) { 2317 case ARC_SPACE_DATA: 2318 ARCSTAT_INCR(arcstat_data_size, space); 2319 break; 2320 case ARC_SPACE_META: 2321 ARCSTAT_INCR(arcstat_metadata_size, space); 2322 break; 2323 case ARC_SPACE_OTHER: 2324 ARCSTAT_INCR(arcstat_other_size, space); 2325 break; 2326 case ARC_SPACE_HDRS: 2327 ARCSTAT_INCR(arcstat_hdr_size, space); 2328 break; 2329 case ARC_SPACE_L2HDRS: 2330 ARCSTAT_INCR(arcstat_l2_hdr_size, space); 2331 break; 2332 } 2333 2334 if (type != ARC_SPACE_DATA) 2335 ARCSTAT_INCR(arcstat_meta_used, space); 2336 2337 atomic_add_64(&arc_size, space); 2338 } 2339 2340 void 2341 arc_space_return(uint64_t space, arc_space_type_t type) 2342 { 2343 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES); 2344 2345 switch (type) { 2346 case ARC_SPACE_DATA: 2347 ARCSTAT_INCR(arcstat_data_size, -space); 2348 break; 2349 case ARC_SPACE_META: 2350 ARCSTAT_INCR(arcstat_metadata_size, -space); 2351 break; 2352 case ARC_SPACE_OTHER: 2353 ARCSTAT_INCR(arcstat_other_size, -space); 2354 break; 2355 case ARC_SPACE_HDRS: 2356 ARCSTAT_INCR(arcstat_hdr_size, -space); 2357 break; 2358 case ARC_SPACE_L2HDRS: 2359 ARCSTAT_INCR(arcstat_l2_hdr_size, -space); 2360 break; 2361 } 2362 2363 if (type != ARC_SPACE_DATA) { 2364 ASSERT(arc_meta_used >= space); 2365 if (arc_meta_max < arc_meta_used) 2366 arc_meta_max = arc_meta_used; 2367 ARCSTAT_INCR(arcstat_meta_used, -space); 2368 } 2369 2370 ASSERT(arc_size >= space); 2371 atomic_add_64(&arc_size, -space); 2372 } 2373 2374 /* 2375 * Given a hdr and a buf, returns whether that buf can share its b_data buffer 2376 * with the hdr's b_pdata. 2377 */ 2378 static boolean_t 2379 arc_can_share(arc_buf_hdr_t *hdr, arc_buf_t *buf) 2380 { 2381 /* 2382 * The criteria for sharing a hdr's data are: 2383 * 1. the hdr's compression matches the buf's compression 2384 * 2. the hdr doesn't need to be byteswapped 2385 * 3. the hdr isn't already being shared 2386 * 4. the buf is either compressed or it is the last buf in the hdr list 2387 * 2388 * Criterion #4 maintains the invariant that shared uncompressed 2389 * bufs must be the final buf in the hdr's b_buf list. Reading this, you 2390 * might ask, "if a compressed buf is allocated first, won't that be the 2391 * last thing in the list?", but in that case it's impossible to create 2392 * a shared uncompressed buf anyway (because the hdr must be compressed 2393 * to have the compressed buf). You might also think that #3 is 2394 * sufficient to make this guarantee, however it's possible 2395 * (specifically in the rare L2ARC write race mentioned in 2396 * arc_buf_alloc_impl()) there will be an existing uncompressed buf that 2397 * is sharable, but wasn't at the time of its allocation. Rather than 2398 * allow a new shared uncompressed buf to be created and then shuffle 2399 * the list around to make it the last element, this simply disallows 2400 * sharing if the new buf isn't the first to be added. 2401 */ 2402 ASSERT3P(buf->b_hdr, ==, hdr); 2403 boolean_t hdr_compressed = HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF; 2404 boolean_t buf_compressed = ARC_BUF_COMPRESSED(buf) != 0; 2405 return (buf_compressed == hdr_compressed && 2406 hdr->b_l1hdr.b_byteswap == DMU_BSWAP_NUMFUNCS && 2407 !HDR_SHARED_DATA(hdr) && 2408 (ARC_BUF_LAST(buf) || ARC_BUF_COMPRESSED(buf))); 2409 } 2410 2411 /* 2412 * Allocate a buf for this hdr. If you care about the data that's in the hdr, 2413 * or if you want a compressed buffer, pass those flags in. Returns 0 if the 2414 * copy was made successfully, or an error code otherwise. 2415 */ 2416 static int 2417 arc_buf_alloc_impl(arc_buf_hdr_t *hdr, void *tag, boolean_t compressed, 2418 boolean_t fill, arc_buf_t **ret) 2419 { 2420 arc_buf_t *buf; 2421 2422 ASSERT(HDR_HAS_L1HDR(hdr)); 2423 ASSERT3U(HDR_GET_LSIZE(hdr), >, 0); 2424 VERIFY(hdr->b_type == ARC_BUFC_DATA || 2425 hdr->b_type == ARC_BUFC_METADATA); 2426 ASSERT3P(ret, !=, NULL); 2427 ASSERT3P(*ret, ==, NULL); 2428 2429 buf = *ret = kmem_cache_alloc(buf_cache, KM_PUSHPAGE); 2430 buf->b_hdr = hdr; 2431 buf->b_data = NULL; 2432 buf->b_next = hdr->b_l1hdr.b_buf; 2433 buf->b_flags = 0; 2434 2435 add_reference(hdr, tag); 2436 2437 /* 2438 * We're about to change the hdr's b_flags. We must either 2439 * hold the hash_lock or be undiscoverable. 2440 */ 2441 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); 2442 2443 /* 2444 * Only honor requests for compressed bufs if the hdr is actually 2445 * compressed. 2446 */ 2447 if (compressed && HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF) 2448 buf->b_flags |= ARC_BUF_FLAG_COMPRESSED; 2449 2450 /* 2451 * If the hdr's data can be shared then we share the data buffer and 2452 * set the appropriate bit in the hdr's b_flags to indicate the hdr is 2453 * sharing it's b_pdata with the arc_buf_t. Otherwise, we allocate a new 2454 * buffer to store the buf's data. 2455 * 2456 * There is one additional restriction here because we're sharing 2457 * hdr -> buf instead of the usual buf -> hdr: the hdr can't be actively 2458 * involved in an L2ARC write, because if this buf is used by an 2459 * arc_write() then the hdr's data buffer will be released when the 2460 * write completes, even though the L2ARC write might still be using it. 2461 */ 2462 boolean_t can_share = arc_can_share(hdr, buf) && !HDR_L2_WRITING(hdr); 2463 2464 /* Set up b_data and sharing */ 2465 if (can_share) { 2466 buf->b_data = hdr->b_l1hdr.b_pdata; 2467 buf->b_flags |= ARC_BUF_FLAG_SHARED; 2468 arc_hdr_set_flags(hdr, ARC_FLAG_SHARED_DATA); 2469 } else { 2470 buf->b_data = 2471 arc_get_data_buf(hdr, arc_buf_size(buf), buf); 2472 ARCSTAT_INCR(arcstat_overhead_size, arc_buf_size(buf)); 2473 } 2474 VERIFY3P(buf->b_data, !=, NULL); 2475 2476 hdr->b_l1hdr.b_buf = buf; 2477 hdr->b_l1hdr.b_bufcnt += 1; 2478 2479 /* 2480 * If the user wants the data from the hdr, we need to either copy or 2481 * decompress the data. 2482 */ 2483 if (fill) { 2484 return (arc_buf_fill(buf, ARC_BUF_COMPRESSED(buf) != 0)); 2485 } 2486 2487 return (0); 2488 } 2489 2490 static char *arc_onloan_tag = "onloan"; 2491 2492 static inline void 2493 arc_loaned_bytes_update(int64_t delta) 2494 { 2495 atomic_add_64(&arc_loaned_bytes, delta); 2496 2497 /* assert that it did not wrap around */ 2498 ASSERT3S(atomic_add_64_nv(&arc_loaned_bytes, 0), >=, 0); 2499 } 2500 2501 /* 2502 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in 2503 * flight data by arc_tempreserve_space() until they are "returned". Loaned 2504 * buffers must be returned to the arc before they can be used by the DMU or 2505 * freed. 2506 */ 2507 arc_buf_t * 2508 arc_loan_buf(spa_t *spa, boolean_t is_metadata, int size) 2509 { 2510 arc_buf_t *buf = arc_alloc_buf(spa, arc_onloan_tag, 2511 is_metadata ? ARC_BUFC_METADATA : ARC_BUFC_DATA, size); 2512 2513 arc_loaned_bytes_update(size); 2514 2515 return (buf); 2516 } 2517 2518 arc_buf_t * 2519 arc_loan_compressed_buf(spa_t *spa, uint64_t psize, uint64_t lsize, 2520 enum zio_compress compression_type) 2521 { 2522 arc_buf_t *buf = arc_alloc_compressed_buf(spa, arc_onloan_tag, 2523 psize, lsize, compression_type); 2524 2525 arc_loaned_bytes_update(psize); 2526 2527 return (buf); 2528 } 2529 2530 2531 /* 2532 * Return a loaned arc buffer to the arc. 2533 */ 2534 void 2535 arc_return_buf(arc_buf_t *buf, void *tag) 2536 { 2537 arc_buf_hdr_t *hdr = buf->b_hdr; 2538 2539 ASSERT3P(buf->b_data, !=, NULL); 2540 ASSERT(HDR_HAS_L1HDR(hdr)); 2541 (void) refcount_add(&hdr->b_l1hdr.b_refcnt, tag); 2542 (void) refcount_remove(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag); 2543 2544 arc_loaned_bytes_update(-arc_buf_size(buf)); 2545 } 2546 2547 /* Detach an arc_buf from a dbuf (tag) */ 2548 void 2549 arc_loan_inuse_buf(arc_buf_t *buf, void *tag) 2550 { 2551 arc_buf_hdr_t *hdr = buf->b_hdr; 2552 2553 ASSERT3P(buf->b_data, !=, NULL); 2554 ASSERT(HDR_HAS_L1HDR(hdr)); 2555 (void) refcount_add(&hdr->b_l1hdr.b_refcnt, arc_onloan_tag); 2556 (void) refcount_remove(&hdr->b_l1hdr.b_refcnt, tag); 2557 2558 arc_loaned_bytes_update(arc_buf_size(buf)); 2559 } 2560 2561 static void 2562 l2arc_free_data_on_write(void *data, size_t size, arc_buf_contents_t type) 2563 { 2564 l2arc_data_free_t *df = kmem_alloc(sizeof (*df), KM_SLEEP); 2565 2566 df->l2df_data = data; 2567 df->l2df_size = size; 2568 df->l2df_type = type; 2569 mutex_enter(&l2arc_free_on_write_mtx); 2570 list_insert_head(l2arc_free_on_write, df); 2571 mutex_exit(&l2arc_free_on_write_mtx); 2572 } 2573 2574 static void 2575 arc_hdr_free_on_write(arc_buf_hdr_t *hdr) 2576 { 2577 arc_state_t *state = hdr->b_l1hdr.b_state; 2578 arc_buf_contents_t type = arc_buf_type(hdr); 2579 uint64_t size = arc_hdr_size(hdr); 2580 2581 /* protected by hash lock, if in the hash table */ 2582 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) { 2583 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 2584 ASSERT(state != arc_anon && state != arc_l2c_only); 2585 2586 (void) refcount_remove_many(&state->arcs_esize[type], 2587 size, hdr); 2588 } 2589 (void) refcount_remove_many(&state->arcs_size, size, hdr); 2590 if (type == ARC_BUFC_METADATA) { 2591 arc_space_return(size, ARC_SPACE_META); 2592 } else { 2593 ASSERT(type == ARC_BUFC_DATA); 2594 arc_space_return(size, ARC_SPACE_DATA); 2595 } 2596 2597 l2arc_free_data_on_write(hdr->b_l1hdr.b_pdata, size, type); 2598 } 2599 2600 /* 2601 * Share the arc_buf_t's data with the hdr. Whenever we are sharing the 2602 * data buffer, we transfer the refcount ownership to the hdr and update 2603 * the appropriate kstats. 2604 */ 2605 static void 2606 arc_share_buf(arc_buf_hdr_t *hdr, arc_buf_t *buf) 2607 { 2608 arc_state_t *state = hdr->b_l1hdr.b_state; 2609 2610 ASSERT(arc_can_share(hdr, buf)); 2611 ASSERT3P(hdr->b_l1hdr.b_pdata, ==, NULL); 2612 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); 2613 2614 /* 2615 * Start sharing the data buffer. We transfer the 2616 * refcount ownership to the hdr since it always owns 2617 * the refcount whenever an arc_buf_t is shared. 2618 */ 2619 refcount_transfer_ownership(&state->arcs_size, buf, hdr); 2620 hdr->b_l1hdr.b_pdata = buf->b_data; 2621 arc_hdr_set_flags(hdr, ARC_FLAG_SHARED_DATA); 2622 buf->b_flags |= ARC_BUF_FLAG_SHARED; 2623 2624 /* 2625 * Since we've transferred ownership to the hdr we need 2626 * to increment its compressed and uncompressed kstats and 2627 * decrement the overhead size. 2628 */ 2629 ARCSTAT_INCR(arcstat_compressed_size, arc_hdr_size(hdr)); 2630 ARCSTAT_INCR(arcstat_uncompressed_size, HDR_GET_LSIZE(hdr)); 2631 ARCSTAT_INCR(arcstat_overhead_size, -arc_buf_size(buf)); 2632 } 2633 2634 static void 2635 arc_unshare_buf(arc_buf_hdr_t *hdr, arc_buf_t *buf) 2636 { 2637 arc_state_t *state = hdr->b_l1hdr.b_state; 2638 2639 ASSERT(arc_buf_is_shared(buf)); 2640 ASSERT3P(hdr->b_l1hdr.b_pdata, !=, NULL); 2641 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); 2642 2643 /* 2644 * We are no longer sharing this buffer so we need 2645 * to transfer its ownership to the rightful owner. 2646 */ 2647 refcount_transfer_ownership(&state->arcs_size, hdr, buf); 2648 arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA); 2649 hdr->b_l1hdr.b_pdata = NULL; 2650 buf->b_flags &= ~ARC_BUF_FLAG_SHARED; 2651 2652 /* 2653 * Since the buffer is no longer shared between 2654 * the arc buf and the hdr, count it as overhead. 2655 */ 2656 ARCSTAT_INCR(arcstat_compressed_size, -arc_hdr_size(hdr)); 2657 ARCSTAT_INCR(arcstat_uncompressed_size, -HDR_GET_LSIZE(hdr)); 2658 ARCSTAT_INCR(arcstat_overhead_size, arc_buf_size(buf)); 2659 } 2660 2661 /* 2662 * Remove an arc_buf_t from the hdr's buf list and return the last 2663 * arc_buf_t on the list. If no buffers remain on the list then return 2664 * NULL. 2665 */ 2666 static arc_buf_t * 2667 arc_buf_remove(arc_buf_hdr_t *hdr, arc_buf_t *buf) 2668 { 2669 ASSERT(HDR_HAS_L1HDR(hdr)); 2670 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); 2671 2672 arc_buf_t **bufp = &hdr->b_l1hdr.b_buf; 2673 arc_buf_t *lastbuf = NULL; 2674 2675 /* 2676 * Remove the buf from the hdr list and locate the last 2677 * remaining buffer on the list. 2678 */ 2679 while (*bufp != NULL) { 2680 if (*bufp == buf) 2681 *bufp = buf->b_next; 2682 2683 /* 2684 * If we've removed a buffer in the middle of 2685 * the list then update the lastbuf and update 2686 * bufp. 2687 */ 2688 if (*bufp != NULL) { 2689 lastbuf = *bufp; 2690 bufp = &(*bufp)->b_next; 2691 } 2692 } 2693 buf->b_next = NULL; 2694 ASSERT3P(lastbuf, !=, buf); 2695 IMPLY(hdr->b_l1hdr.b_bufcnt > 0, lastbuf != NULL); 2696 IMPLY(hdr->b_l1hdr.b_bufcnt > 0, hdr->b_l1hdr.b_buf != NULL); 2697 IMPLY(lastbuf != NULL, ARC_BUF_LAST(lastbuf)); 2698 2699 return (lastbuf); 2700 } 2701 2702 /* 2703 * Free up buf->b_data and pull the arc_buf_t off of the the arc_buf_hdr_t's 2704 * list and free it. 2705 */ 2706 static void 2707 arc_buf_destroy_impl(arc_buf_t *buf) 2708 { 2709 arc_buf_hdr_t *hdr = buf->b_hdr; 2710 2711 /* 2712 * Free up the data associated with the buf but only if we're not 2713 * sharing this with the hdr. If we are sharing it with the hdr, the 2714 * hdr is responsible for doing the free. 2715 */ 2716 if (buf->b_data != NULL) { 2717 /* 2718 * We're about to change the hdr's b_flags. We must either 2719 * hold the hash_lock or be undiscoverable. 2720 */ 2721 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr)); 2722 2723 arc_cksum_verify(buf); 2724 arc_buf_unwatch(buf); 2725 2726 if (arc_buf_is_shared(buf)) { 2727 arc_hdr_clear_flags(hdr, ARC_FLAG_SHARED_DATA); 2728 } else { 2729 uint64_t size = arc_buf_size(buf); 2730 arc_free_data_buf(hdr, buf->b_data, size, buf); 2731 ARCSTAT_INCR(arcstat_overhead_size, -size); 2732 } 2733 buf->b_data = NULL; 2734 2735 ASSERT(hdr->b_l1hdr.b_bufcnt > 0); 2736 hdr->b_l1hdr.b_bufcnt -= 1; 2737 } 2738 2739 arc_buf_t *lastbuf = arc_buf_remove(hdr, buf); 2740 2741 if (ARC_BUF_SHARED(buf) && !ARC_BUF_COMPRESSED(buf)) { 2742 /* 2743 * If the current arc_buf_t is sharing its data buffer with the 2744 * hdr, then reassign the hdr's b_pdata to share it with the new 2745 * buffer at the end of the list. The shared buffer is always 2746 * the last one on the hdr's buffer list. 2747 * 2748 * There is an equivalent case for compressed bufs, but since 2749 * they aren't guaranteed to be the last buf in the list and 2750 * that is an exceedingly rare case, we just allow that space be 2751 * wasted temporarily. 2752 */ 2753 if (lastbuf != NULL) { 2754 /* Only one buf can be shared at once */ 2755 VERIFY(!arc_buf_is_shared(lastbuf)); 2756 /* hdr is uncompressed so can't have compressed buf */ 2757 VERIFY(!ARC_BUF_COMPRESSED(lastbuf)); 2758 2759 ASSERT3P(hdr->b_l1hdr.b_pdata, !=, NULL); 2760 arc_hdr_free_pdata(hdr); 2761 2762 /* 2763 * We must setup a new shared block between the 2764 * last buffer and the hdr. The data would have 2765 * been allocated by the arc buf so we need to transfer 2766 * ownership to the hdr since it's now being shared. 2767 */ 2768 arc_share_buf(hdr, lastbuf); 2769 } 2770 } else if (HDR_SHARED_DATA(hdr)) { 2771 /* 2772 * Uncompressed shared buffers are always at the end 2773 * of the list. Compressed buffers don't have the 2774 * same requirements. This makes it hard to 2775 * simply assert that the lastbuf is shared so 2776 * we rely on the hdr's compression flags to determine 2777 * if we have a compressed, shared buffer. 2778 */ 2779 ASSERT3P(lastbuf, !=, NULL); 2780 ASSERT(arc_buf_is_shared(lastbuf) || 2781 HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF); 2782 } 2783 2784 /* 2785 * Free the checksum if we're removing the last uncompressed buf from 2786 * this hdr. 2787 */ 2788 if (!arc_hdr_has_uncompressed_buf(hdr)) { 2789 arc_cksum_free(hdr); 2790 } 2791 2792 /* clean up the buf */ 2793 buf->b_hdr = NULL; 2794 kmem_cache_free(buf_cache, buf); 2795 } 2796 2797 static void 2798 arc_hdr_alloc_pdata(arc_buf_hdr_t *hdr) 2799 { 2800 ASSERT3U(HDR_GET_LSIZE(hdr), >, 0); 2801 ASSERT(HDR_HAS_L1HDR(hdr)); 2802 ASSERT(!HDR_SHARED_DATA(hdr)); 2803 2804 ASSERT3P(hdr->b_l1hdr.b_pdata, ==, NULL); 2805 hdr->b_l1hdr.b_pdata = arc_get_data_buf(hdr, arc_hdr_size(hdr), hdr); 2806 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS; 2807 ASSERT3P(hdr->b_l1hdr.b_pdata, !=, NULL); 2808 2809 ARCSTAT_INCR(arcstat_compressed_size, arc_hdr_size(hdr)); 2810 ARCSTAT_INCR(arcstat_uncompressed_size, HDR_GET_LSIZE(hdr)); 2811 } 2812 2813 static void 2814 arc_hdr_free_pdata(arc_buf_hdr_t *hdr) 2815 { 2816 ASSERT(HDR_HAS_L1HDR(hdr)); 2817 ASSERT3P(hdr->b_l1hdr.b_pdata, !=, NULL); 2818 2819 /* 2820 * If the hdr is currently being written to the l2arc then 2821 * we defer freeing the data by adding it to the l2arc_free_on_write 2822 * list. The l2arc will free the data once it's finished 2823 * writing it to the l2arc device. 2824 */ 2825 if (HDR_L2_WRITING(hdr)) { 2826 arc_hdr_free_on_write(hdr); 2827 ARCSTAT_BUMP(arcstat_l2_free_on_write); 2828 } else { 2829 arc_free_data_buf(hdr, hdr->b_l1hdr.b_pdata, 2830 arc_hdr_size(hdr), hdr); 2831 } 2832 hdr->b_l1hdr.b_pdata = NULL; 2833 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS; 2834 2835 ARCSTAT_INCR(arcstat_compressed_size, -arc_hdr_size(hdr)); 2836 ARCSTAT_INCR(arcstat_uncompressed_size, -HDR_GET_LSIZE(hdr)); 2837 } 2838 2839 static arc_buf_hdr_t * 2840 arc_hdr_alloc(uint64_t spa, int32_t psize, int32_t lsize, 2841 enum zio_compress compression_type, arc_buf_contents_t type) 2842 { 2843 arc_buf_hdr_t *hdr; 2844 2845 VERIFY(type == ARC_BUFC_DATA || type == ARC_BUFC_METADATA); 2846 2847 hdr = kmem_cache_alloc(hdr_full_cache, KM_PUSHPAGE); 2848 ASSERT(HDR_EMPTY(hdr)); 2849 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL); 2850 ASSERT3P(hdr->b_l1hdr.b_thawed, ==, NULL); 2851 HDR_SET_PSIZE(hdr, psize); 2852 HDR_SET_LSIZE(hdr, lsize); 2853 hdr->b_spa = spa; 2854 hdr->b_type = type; 2855 hdr->b_flags = 0; 2856 arc_hdr_set_flags(hdr, arc_bufc_to_flags(type) | ARC_FLAG_HAS_L1HDR); 2857 arc_hdr_set_compress(hdr, compression_type); 2858 2859 hdr->b_l1hdr.b_state = arc_anon; 2860 hdr->b_l1hdr.b_arc_access = 0; 2861 hdr->b_l1hdr.b_bufcnt = 0; 2862 hdr->b_l1hdr.b_buf = NULL; 2863 2864 /* 2865 * Allocate the hdr's buffer. This will contain either 2866 * the compressed or uncompressed data depending on the block 2867 * it references and compressed arc enablement. 2868 */ 2869 arc_hdr_alloc_pdata(hdr); 2870 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 2871 2872 return (hdr); 2873 } 2874 2875 /* 2876 * Transition between the two allocation states for the arc_buf_hdr struct. 2877 * The arc_buf_hdr struct can be allocated with (hdr_full_cache) or without 2878 * (hdr_l2only_cache) the fields necessary for the L1 cache - the smaller 2879 * version is used when a cache buffer is only in the L2ARC in order to reduce 2880 * memory usage. 2881 */ 2882 static arc_buf_hdr_t * 2883 arc_hdr_realloc(arc_buf_hdr_t *hdr, kmem_cache_t *old, kmem_cache_t *new) 2884 { 2885 ASSERT(HDR_HAS_L2HDR(hdr)); 2886 2887 arc_buf_hdr_t *nhdr; 2888 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev; 2889 2890 ASSERT((old == hdr_full_cache && new == hdr_l2only_cache) || 2891 (old == hdr_l2only_cache && new == hdr_full_cache)); 2892 2893 nhdr = kmem_cache_alloc(new, KM_PUSHPAGE); 2894 2895 ASSERT(MUTEX_HELD(HDR_LOCK(hdr))); 2896 buf_hash_remove(hdr); 2897 2898 bcopy(hdr, nhdr, HDR_L2ONLY_SIZE); 2899 2900 if (new == hdr_full_cache) { 2901 arc_hdr_set_flags(nhdr, ARC_FLAG_HAS_L1HDR); 2902 /* 2903 * arc_access and arc_change_state need to be aware that a 2904 * header has just come out of L2ARC, so we set its state to 2905 * l2c_only even though it's about to change. 2906 */ 2907 nhdr->b_l1hdr.b_state = arc_l2c_only; 2908 2909 /* Verify previous threads set to NULL before freeing */ 2910 ASSERT3P(nhdr->b_l1hdr.b_pdata, ==, NULL); 2911 } else { 2912 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 2913 ASSERT0(hdr->b_l1hdr.b_bufcnt); 2914 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL); 2915 2916 /* 2917 * If we've reached here, We must have been called from 2918 * arc_evict_hdr(), as such we should have already been 2919 * removed from any ghost list we were previously on 2920 * (which protects us from racing with arc_evict_state), 2921 * thus no locking is needed during this check. 2922 */ 2923 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 2924 2925 /* 2926 * A buffer must not be moved into the arc_l2c_only 2927 * state if it's not finished being written out to the 2928 * l2arc device. Otherwise, the b_l1hdr.b_pdata field 2929 * might try to be accessed, even though it was removed. 2930 */ 2931 VERIFY(!HDR_L2_WRITING(hdr)); 2932 VERIFY3P(hdr->b_l1hdr.b_pdata, ==, NULL); 2933 2934 #ifdef ZFS_DEBUG 2935 if (hdr->b_l1hdr.b_thawed != NULL) { 2936 kmem_free(hdr->b_l1hdr.b_thawed, 1); 2937 hdr->b_l1hdr.b_thawed = NULL; 2938 } 2939 #endif 2940 2941 arc_hdr_clear_flags(nhdr, ARC_FLAG_HAS_L1HDR); 2942 } 2943 /* 2944 * The header has been reallocated so we need to re-insert it into any 2945 * lists it was on. 2946 */ 2947 (void) buf_hash_insert(nhdr, NULL); 2948 2949 ASSERT(list_link_active(&hdr->b_l2hdr.b_l2node)); 2950 2951 mutex_enter(&dev->l2ad_mtx); 2952 2953 /* 2954 * We must place the realloc'ed header back into the list at 2955 * the same spot. Otherwise, if it's placed earlier in the list, 2956 * l2arc_write_buffers() could find it during the function's 2957 * write phase, and try to write it out to the l2arc. 2958 */ 2959 list_insert_after(&dev->l2ad_buflist, hdr, nhdr); 2960 list_remove(&dev->l2ad_buflist, hdr); 2961 2962 mutex_exit(&dev->l2ad_mtx); 2963 2964 /* 2965 * Since we're using the pointer address as the tag when 2966 * incrementing and decrementing the l2ad_alloc refcount, we 2967 * must remove the old pointer (that we're about to destroy) and 2968 * add the new pointer to the refcount. Otherwise we'd remove 2969 * the wrong pointer address when calling arc_hdr_destroy() later. 2970 */ 2971 2972 (void) refcount_remove_many(&dev->l2ad_alloc, arc_hdr_size(hdr), hdr); 2973 (void) refcount_add_many(&dev->l2ad_alloc, arc_hdr_size(nhdr), nhdr); 2974 2975 buf_discard_identity(hdr); 2976 kmem_cache_free(old, hdr); 2977 2978 return (nhdr); 2979 } 2980 2981 /* 2982 * Allocate a new arc_buf_hdr_t and arc_buf_t and return the buf to the caller. 2983 * The buf is returned thawed since we expect the consumer to modify it. 2984 */ 2985 arc_buf_t * 2986 arc_alloc_buf(spa_t *spa, void *tag, arc_buf_contents_t type, int32_t size) 2987 { 2988 arc_buf_hdr_t *hdr = arc_hdr_alloc(spa_load_guid(spa), size, size, 2989 ZIO_COMPRESS_OFF, type); 2990 ASSERT(!MUTEX_HELD(HDR_LOCK(hdr))); 2991 2992 arc_buf_t *buf = NULL; 2993 VERIFY0(arc_buf_alloc_impl(hdr, tag, B_FALSE, B_FALSE, &buf)); 2994 arc_buf_thaw(buf); 2995 2996 return (buf); 2997 } 2998 2999 /* 3000 * Allocate a compressed buf in the same manner as arc_alloc_buf. Don't use this 3001 * for bufs containing metadata. 3002 */ 3003 arc_buf_t * 3004 arc_alloc_compressed_buf(spa_t *spa, void *tag, uint64_t psize, uint64_t lsize, 3005 enum zio_compress compression_type) 3006 { 3007 ASSERT3U(lsize, >, 0); 3008 ASSERT3U(lsize, >=, psize); 3009 ASSERT(compression_type > ZIO_COMPRESS_OFF); 3010 ASSERT(compression_type < ZIO_COMPRESS_FUNCTIONS); 3011 3012 arc_buf_hdr_t *hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize, 3013 compression_type, ARC_BUFC_DATA); 3014 ASSERT(!MUTEX_HELD(HDR_LOCK(hdr))); 3015 3016 arc_buf_t *buf = NULL; 3017 VERIFY0(arc_buf_alloc_impl(hdr, tag, B_TRUE, B_FALSE, &buf)); 3018 arc_buf_thaw(buf); 3019 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL); 3020 3021 return (buf); 3022 } 3023 3024 static void 3025 arc_hdr_l2hdr_destroy(arc_buf_hdr_t *hdr) 3026 { 3027 l2arc_buf_hdr_t *l2hdr = &hdr->b_l2hdr; 3028 l2arc_dev_t *dev = l2hdr->b_dev; 3029 uint64_t asize = arc_hdr_size(hdr); 3030 3031 ASSERT(MUTEX_HELD(&dev->l2ad_mtx)); 3032 ASSERT(HDR_HAS_L2HDR(hdr)); 3033 3034 list_remove(&dev->l2ad_buflist, hdr); 3035 3036 ARCSTAT_INCR(arcstat_l2_asize, -asize); 3037 ARCSTAT_INCR(arcstat_l2_size, -HDR_GET_LSIZE(hdr)); 3038 3039 vdev_space_update(dev->l2ad_vdev, -asize, 0, 0); 3040 3041 (void) refcount_remove_many(&dev->l2ad_alloc, asize, hdr); 3042 arc_hdr_clear_flags(hdr, ARC_FLAG_HAS_L2HDR); 3043 } 3044 3045 static void 3046 arc_hdr_destroy(arc_buf_hdr_t *hdr) 3047 { 3048 if (HDR_HAS_L1HDR(hdr)) { 3049 ASSERT(hdr->b_l1hdr.b_buf == NULL || 3050 hdr->b_l1hdr.b_bufcnt > 0); 3051 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 3052 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon); 3053 } 3054 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 3055 ASSERT(!HDR_IN_HASH_TABLE(hdr)); 3056 3057 if (!HDR_EMPTY(hdr)) 3058 buf_discard_identity(hdr); 3059 3060 if (HDR_HAS_L2HDR(hdr)) { 3061 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev; 3062 boolean_t buflist_held = MUTEX_HELD(&dev->l2ad_mtx); 3063 3064 if (!buflist_held) 3065 mutex_enter(&dev->l2ad_mtx); 3066 3067 /* 3068 * Even though we checked this conditional above, we 3069 * need to check this again now that we have the 3070 * l2ad_mtx. This is because we could be racing with 3071 * another thread calling l2arc_evict() which might have 3072 * destroyed this header's L2 portion as we were waiting 3073 * to acquire the l2ad_mtx. If that happens, we don't 3074 * want to re-destroy the header's L2 portion. 3075 */ 3076 if (HDR_HAS_L2HDR(hdr)) 3077 arc_hdr_l2hdr_destroy(hdr); 3078 3079 if (!buflist_held) 3080 mutex_exit(&dev->l2ad_mtx); 3081 } 3082 3083 if (HDR_HAS_L1HDR(hdr)) { 3084 arc_cksum_free(hdr); 3085 3086 while (hdr->b_l1hdr.b_buf != NULL) 3087 arc_buf_destroy_impl(hdr->b_l1hdr.b_buf); 3088 3089 #ifdef ZFS_DEBUG 3090 if (hdr->b_l1hdr.b_thawed != NULL) { 3091 kmem_free(hdr->b_l1hdr.b_thawed, 1); 3092 hdr->b_l1hdr.b_thawed = NULL; 3093 } 3094 #endif 3095 3096 if (hdr->b_l1hdr.b_pdata != NULL) { 3097 arc_hdr_free_pdata(hdr); 3098 } 3099 } 3100 3101 ASSERT3P(hdr->b_hash_next, ==, NULL); 3102 if (HDR_HAS_L1HDR(hdr)) { 3103 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 3104 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL); 3105 kmem_cache_free(hdr_full_cache, hdr); 3106 } else { 3107 kmem_cache_free(hdr_l2only_cache, hdr); 3108 } 3109 } 3110 3111 void 3112 arc_buf_destroy(arc_buf_t *buf, void* tag) 3113 { 3114 arc_buf_hdr_t *hdr = buf->b_hdr; 3115 kmutex_t *hash_lock = HDR_LOCK(hdr); 3116 3117 if (hdr->b_l1hdr.b_state == arc_anon) { 3118 ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1); 3119 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 3120 VERIFY0(remove_reference(hdr, NULL, tag)); 3121 arc_hdr_destroy(hdr); 3122 return; 3123 } 3124 3125 mutex_enter(hash_lock); 3126 ASSERT3P(hdr, ==, buf->b_hdr); 3127 ASSERT(hdr->b_l1hdr.b_bufcnt > 0); 3128 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 3129 ASSERT3P(hdr->b_l1hdr.b_state, !=, arc_anon); 3130 ASSERT3P(buf->b_data, !=, NULL); 3131 3132 (void) remove_reference(hdr, hash_lock, tag); 3133 arc_buf_destroy_impl(buf); 3134 mutex_exit(hash_lock); 3135 } 3136 3137 /* 3138 * Evict the arc_buf_hdr that is provided as a parameter. The resultant 3139 * state of the header is dependent on it's state prior to entering this 3140 * function. The following transitions are possible: 3141 * 3142 * - arc_mru -> arc_mru_ghost 3143 * - arc_mfu -> arc_mfu_ghost 3144 * - arc_mru_ghost -> arc_l2c_only 3145 * - arc_mru_ghost -> deleted 3146 * - arc_mfu_ghost -> arc_l2c_only 3147 * - arc_mfu_ghost -> deleted 3148 */ 3149 static int64_t 3150 arc_evict_hdr(arc_buf_hdr_t *hdr, kmutex_t *hash_lock) 3151 { 3152 arc_state_t *evicted_state, *state; 3153 int64_t bytes_evicted = 0; 3154 3155 ASSERT(MUTEX_HELD(hash_lock)); 3156 ASSERT(HDR_HAS_L1HDR(hdr)); 3157 3158 state = hdr->b_l1hdr.b_state; 3159 if (GHOST_STATE(state)) { 3160 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 3161 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 3162 3163 /* 3164 * l2arc_write_buffers() relies on a header's L1 portion 3165 * (i.e. its b_pdata field) during its write phase. 3166 * Thus, we cannot push a header onto the arc_l2c_only 3167 * state (removing it's L1 piece) until the header is 3168 * done being written to the l2arc. 3169 */ 3170 if (HDR_HAS_L2HDR(hdr) && HDR_L2_WRITING(hdr)) { 3171 ARCSTAT_BUMP(arcstat_evict_l2_skip); 3172 return (bytes_evicted); 3173 } 3174 3175 ARCSTAT_BUMP(arcstat_deleted); 3176 bytes_evicted += HDR_GET_LSIZE(hdr); 3177 3178 DTRACE_PROBE1(arc__delete, arc_buf_hdr_t *, hdr); 3179 3180 ASSERT3P(hdr->b_l1hdr.b_pdata, ==, NULL); 3181 if (HDR_HAS_L2HDR(hdr)) { 3182 /* 3183 * This buffer is cached on the 2nd Level ARC; 3184 * don't destroy the header. 3185 */ 3186 arc_change_state(arc_l2c_only, hdr, hash_lock); 3187 /* 3188 * dropping from L1+L2 cached to L2-only, 3189 * realloc to remove the L1 header. 3190 */ 3191 hdr = arc_hdr_realloc(hdr, hdr_full_cache, 3192 hdr_l2only_cache); 3193 } else { 3194 arc_change_state(arc_anon, hdr, hash_lock); 3195 arc_hdr_destroy(hdr); 3196 } 3197 return (bytes_evicted); 3198 } 3199 3200 ASSERT(state == arc_mru || state == arc_mfu); 3201 evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost; 3202 3203 /* prefetch buffers have a minimum lifespan */ 3204 if (HDR_IO_IN_PROGRESS(hdr) || 3205 ((hdr->b_flags & (ARC_FLAG_PREFETCH | ARC_FLAG_INDIRECT)) && 3206 ddi_get_lbolt() - hdr->b_l1hdr.b_arc_access < 3207 arc_min_prefetch_lifespan)) { 3208 ARCSTAT_BUMP(arcstat_evict_skip); 3209 return (bytes_evicted); 3210 } 3211 3212 ASSERT0(refcount_count(&hdr->b_l1hdr.b_refcnt)); 3213 while (hdr->b_l1hdr.b_buf) { 3214 arc_buf_t *buf = hdr->b_l1hdr.b_buf; 3215 if (!mutex_tryenter(&buf->b_evict_lock)) { 3216 ARCSTAT_BUMP(arcstat_mutex_miss); 3217 break; 3218 } 3219 if (buf->b_data != NULL) 3220 bytes_evicted += HDR_GET_LSIZE(hdr); 3221 mutex_exit(&buf->b_evict_lock); 3222 arc_buf_destroy_impl(buf); 3223 } 3224 3225 if (HDR_HAS_L2HDR(hdr)) { 3226 ARCSTAT_INCR(arcstat_evict_l2_cached, HDR_GET_LSIZE(hdr)); 3227 } else { 3228 if (l2arc_write_eligible(hdr->b_spa, hdr)) { 3229 ARCSTAT_INCR(arcstat_evict_l2_eligible, 3230 HDR_GET_LSIZE(hdr)); 3231 } else { 3232 ARCSTAT_INCR(arcstat_evict_l2_ineligible, 3233 HDR_GET_LSIZE(hdr)); 3234 } 3235 } 3236 3237 if (hdr->b_l1hdr.b_bufcnt == 0) { 3238 arc_cksum_free(hdr); 3239 3240 bytes_evicted += arc_hdr_size(hdr); 3241 3242 /* 3243 * If this hdr is being evicted and has a compressed 3244 * buffer then we discard it here before we change states. 3245 * This ensures that the accounting is updated correctly 3246 * in arc_free_data_buf(). 3247 */ 3248 arc_hdr_free_pdata(hdr); 3249 3250 arc_change_state(evicted_state, hdr, hash_lock); 3251 ASSERT(HDR_IN_HASH_TABLE(hdr)); 3252 arc_hdr_set_flags(hdr, ARC_FLAG_IN_HASH_TABLE); 3253 DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, hdr); 3254 } 3255 3256 return (bytes_evicted); 3257 } 3258 3259 static uint64_t 3260 arc_evict_state_impl(multilist_t *ml, int idx, arc_buf_hdr_t *marker, 3261 uint64_t spa, int64_t bytes) 3262 { 3263 multilist_sublist_t *mls; 3264 uint64_t bytes_evicted = 0; 3265 arc_buf_hdr_t *hdr; 3266 kmutex_t *hash_lock; 3267 int evict_count = 0; 3268 3269 ASSERT3P(marker, !=, NULL); 3270 IMPLY(bytes < 0, bytes == ARC_EVICT_ALL); 3271 3272 mls = multilist_sublist_lock(ml, idx); 3273 3274 for (hdr = multilist_sublist_prev(mls, marker); hdr != NULL; 3275 hdr = multilist_sublist_prev(mls, marker)) { 3276 if ((bytes != ARC_EVICT_ALL && bytes_evicted >= bytes) || 3277 (evict_count >= zfs_arc_evict_batch_limit)) 3278 break; 3279 3280 /* 3281 * To keep our iteration location, move the marker 3282 * forward. Since we're not holding hdr's hash lock, we 3283 * must be very careful and not remove 'hdr' from the 3284 * sublist. Otherwise, other consumers might mistake the 3285 * 'hdr' as not being on a sublist when they call the 3286 * multilist_link_active() function (they all rely on 3287 * the hash lock protecting concurrent insertions and 3288 * removals). multilist_sublist_move_forward() was 3289 * specifically implemented to ensure this is the case 3290 * (only 'marker' will be removed and re-inserted). 3291 */ 3292 multilist_sublist_move_forward(mls, marker); 3293 3294 /* 3295 * The only case where the b_spa field should ever be 3296 * zero, is the marker headers inserted by 3297 * arc_evict_state(). It's possible for multiple threads 3298 * to be calling arc_evict_state() concurrently (e.g. 3299 * dsl_pool_close() and zio_inject_fault()), so we must 3300 * skip any markers we see from these other threads. 3301 */ 3302 if (hdr->b_spa == 0) 3303 continue; 3304 3305 /* we're only interested in evicting buffers of a certain spa */ 3306 if (spa != 0 && hdr->b_spa != spa) { 3307 ARCSTAT_BUMP(arcstat_evict_skip); 3308 continue; 3309 } 3310 3311 hash_lock = HDR_LOCK(hdr); 3312 3313 /* 3314 * We aren't calling this function from any code path 3315 * that would already be holding a hash lock, so we're 3316 * asserting on this assumption to be defensive in case 3317 * this ever changes. Without this check, it would be 3318 * possible to incorrectly increment arcstat_mutex_miss 3319 * below (e.g. if the code changed such that we called 3320 * this function with a hash lock held). 3321 */ 3322 ASSERT(!MUTEX_HELD(hash_lock)); 3323 3324 if (mutex_tryenter(hash_lock)) { 3325 uint64_t evicted = arc_evict_hdr(hdr, hash_lock); 3326 mutex_exit(hash_lock); 3327 3328 bytes_evicted += evicted; 3329 3330 /* 3331 * If evicted is zero, arc_evict_hdr() must have 3332 * decided to skip this header, don't increment 3333 * evict_count in this case. 3334 */ 3335 if (evicted != 0) 3336 evict_count++; 3337 3338 /* 3339 * If arc_size isn't overflowing, signal any 3340 * threads that might happen to be waiting. 3341 * 3342 * For each header evicted, we wake up a single 3343 * thread. If we used cv_broadcast, we could 3344 * wake up "too many" threads causing arc_size 3345 * to significantly overflow arc_c; since 3346 * arc_get_data_buf() doesn't check for overflow 3347 * when it's woken up (it doesn't because it's 3348 * possible for the ARC to be overflowing while 3349 * full of un-evictable buffers, and the 3350 * function should proceed in this case). 3351 * 3352 * If threads are left sleeping, due to not 3353 * using cv_broadcast, they will be woken up 3354 * just before arc_reclaim_thread() sleeps. 3355 */ 3356 mutex_enter(&arc_reclaim_lock); 3357 if (!arc_is_overflowing()) 3358 cv_signal(&arc_reclaim_waiters_cv); 3359 mutex_exit(&arc_reclaim_lock); 3360 } else { 3361 ARCSTAT_BUMP(arcstat_mutex_miss); 3362 } 3363 } 3364 3365 multilist_sublist_unlock(mls); 3366 3367 return (bytes_evicted); 3368 } 3369 3370 /* 3371 * Evict buffers from the given arc state, until we've removed the 3372 * specified number of bytes. Move the removed buffers to the 3373 * appropriate evict state. 3374 * 3375 * This function makes a "best effort". It skips over any buffers 3376 * it can't get a hash_lock on, and so, may not catch all candidates. 3377 * It may also return without evicting as much space as requested. 3378 * 3379 * If bytes is specified using the special value ARC_EVICT_ALL, this 3380 * will evict all available (i.e. unlocked and evictable) buffers from 3381 * the given arc state; which is used by arc_flush(). 3382 */ 3383 static uint64_t 3384 arc_evict_state(arc_state_t *state, uint64_t spa, int64_t bytes, 3385 arc_buf_contents_t type) 3386 { 3387 uint64_t total_evicted = 0; 3388 multilist_t *ml = &state->arcs_list[type]; 3389 int num_sublists; 3390 arc_buf_hdr_t **markers; 3391 3392 IMPLY(bytes < 0, bytes == ARC_EVICT_ALL); 3393 3394 num_sublists = multilist_get_num_sublists(ml); 3395 3396 /* 3397 * If we've tried to evict from each sublist, made some 3398 * progress, but still have not hit the target number of bytes 3399 * to evict, we want to keep trying. The markers allow us to 3400 * pick up where we left off for each individual sublist, rather 3401 * than starting from the tail each time. 3402 */ 3403 markers = kmem_zalloc(sizeof (*markers) * num_sublists, KM_SLEEP); 3404 for (int i = 0; i < num_sublists; i++) { 3405 markers[i] = kmem_cache_alloc(hdr_full_cache, KM_SLEEP); 3406 3407 /* 3408 * A b_spa of 0 is used to indicate that this header is 3409 * a marker. This fact is used in arc_adjust_type() and 3410 * arc_evict_state_impl(). 3411 */ 3412 markers[i]->b_spa = 0; 3413 3414 multilist_sublist_t *mls = multilist_sublist_lock(ml, i); 3415 multilist_sublist_insert_tail(mls, markers[i]); 3416 multilist_sublist_unlock(mls); 3417 } 3418 3419 /* 3420 * While we haven't hit our target number of bytes to evict, or 3421 * we're evicting all available buffers. 3422 */ 3423 while (total_evicted < bytes || bytes == ARC_EVICT_ALL) { 3424 /* 3425 * Start eviction using a randomly selected sublist, 3426 * this is to try and evenly balance eviction across all 3427 * sublists. Always starting at the same sublist 3428 * (e.g. index 0) would cause evictions to favor certain 3429 * sublists over others. 3430 */ 3431 int sublist_idx = multilist_get_random_index(ml); 3432 uint64_t scan_evicted = 0; 3433 3434 for (int i = 0; i < num_sublists; i++) { 3435 uint64_t bytes_remaining; 3436 uint64_t bytes_evicted; 3437 3438 if (bytes == ARC_EVICT_ALL) 3439 bytes_remaining = ARC_EVICT_ALL; 3440 else if (total_evicted < bytes) 3441 bytes_remaining = bytes - total_evicted; 3442 else 3443 break; 3444 3445 bytes_evicted = arc_evict_state_impl(ml, sublist_idx, 3446 markers[sublist_idx], spa, bytes_remaining); 3447 3448 scan_evicted += bytes_evicted; 3449 total_evicted += bytes_evicted; 3450 3451 /* we've reached the end, wrap to the beginning */ 3452 if (++sublist_idx >= num_sublists) 3453 sublist_idx = 0; 3454 } 3455 3456 /* 3457 * If we didn't evict anything during this scan, we have 3458 * no reason to believe we'll evict more during another 3459 * scan, so break the loop. 3460 */ 3461 if (scan_evicted == 0) { 3462 /* This isn't possible, let's make that obvious */ 3463 ASSERT3S(bytes, !=, 0); 3464 3465 /* 3466 * When bytes is ARC_EVICT_ALL, the only way to 3467 * break the loop is when scan_evicted is zero. 3468 * In that case, we actually have evicted enough, 3469 * so we don't want to increment the kstat. 3470 */ 3471 if (bytes != ARC_EVICT_ALL) { 3472 ASSERT3S(total_evicted, <, bytes); 3473 ARCSTAT_BUMP(arcstat_evict_not_enough); 3474 } 3475 3476 break; 3477 } 3478 } 3479 3480 for (int i = 0; i < num_sublists; i++) { 3481 multilist_sublist_t *mls = multilist_sublist_lock(ml, i); 3482 multilist_sublist_remove(mls, markers[i]); 3483 multilist_sublist_unlock(mls); 3484 3485 kmem_cache_free(hdr_full_cache, markers[i]); 3486 } 3487 kmem_free(markers, sizeof (*markers) * num_sublists); 3488 3489 return (total_evicted); 3490 } 3491 3492 /* 3493 * Flush all "evictable" data of the given type from the arc state 3494 * specified. This will not evict any "active" buffers (i.e. referenced). 3495 * 3496 * When 'retry' is set to B_FALSE, the function will make a single pass 3497 * over the state and evict any buffers that it can. Since it doesn't 3498 * continually retry the eviction, it might end up leaving some buffers 3499 * in the ARC due to lock misses. 3500 * 3501 * When 'retry' is set to B_TRUE, the function will continually retry the 3502 * eviction until *all* evictable buffers have been removed from the 3503 * state. As a result, if concurrent insertions into the state are 3504 * allowed (e.g. if the ARC isn't shutting down), this function might 3505 * wind up in an infinite loop, continually trying to evict buffers. 3506 */ 3507 static uint64_t 3508 arc_flush_state(arc_state_t *state, uint64_t spa, arc_buf_contents_t type, 3509 boolean_t retry) 3510 { 3511 uint64_t evicted = 0; 3512 3513 while (refcount_count(&state->arcs_esize[type]) != 0) { 3514 evicted += arc_evict_state(state, spa, ARC_EVICT_ALL, type); 3515 3516 if (!retry) 3517 break; 3518 } 3519 3520 return (evicted); 3521 } 3522 3523 /* 3524 * Evict the specified number of bytes from the state specified, 3525 * restricting eviction to the spa and type given. This function 3526 * prevents us from trying to evict more from a state's list than 3527 * is "evictable", and to skip evicting altogether when passed a 3528 * negative value for "bytes". In contrast, arc_evict_state() will 3529 * evict everything it can, when passed a negative value for "bytes". 3530 */ 3531 static uint64_t 3532 arc_adjust_impl(arc_state_t *state, uint64_t spa, int64_t bytes, 3533 arc_buf_contents_t type) 3534 { 3535 int64_t delta; 3536 3537 if (bytes > 0 && refcount_count(&state->arcs_esize[type]) > 0) { 3538 delta = MIN(refcount_count(&state->arcs_esize[type]), bytes); 3539 return (arc_evict_state(state, spa, delta, type)); 3540 } 3541 3542 return (0); 3543 } 3544 3545 /* 3546 * Evict metadata buffers from the cache, such that arc_meta_used is 3547 * capped by the arc_meta_limit tunable. 3548 */ 3549 static uint64_t 3550 arc_adjust_meta(void) 3551 { 3552 uint64_t total_evicted = 0; 3553 int64_t target; 3554 3555 /* 3556 * If we're over the meta limit, we want to evict enough 3557 * metadata to get back under the meta limit. We don't want to 3558 * evict so much that we drop the MRU below arc_p, though. If 3559 * we're over the meta limit more than we're over arc_p, we 3560 * evict some from the MRU here, and some from the MFU below. 3561 */ 3562 target = MIN((int64_t)(arc_meta_used - arc_meta_limit), 3563 (int64_t)(refcount_count(&arc_anon->arcs_size) + 3564 refcount_count(&arc_mru->arcs_size) - arc_p)); 3565 3566 total_evicted += arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA); 3567 3568 /* 3569 * Similar to the above, we want to evict enough bytes to get us 3570 * below the meta limit, but not so much as to drop us below the 3571 * space allotted to the MFU (which is defined as arc_c - arc_p). 3572 */ 3573 target = MIN((int64_t)(arc_meta_used - arc_meta_limit), 3574 (int64_t)(refcount_count(&arc_mfu->arcs_size) - (arc_c - arc_p))); 3575 3576 total_evicted += arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA); 3577 3578 return (total_evicted); 3579 } 3580 3581 /* 3582 * Return the type of the oldest buffer in the given arc state 3583 * 3584 * This function will select a random sublist of type ARC_BUFC_DATA and 3585 * a random sublist of type ARC_BUFC_METADATA. The tail of each sublist 3586 * is compared, and the type which contains the "older" buffer will be 3587 * returned. 3588 */ 3589 static arc_buf_contents_t 3590 arc_adjust_type(arc_state_t *state) 3591 { 3592 multilist_t *data_ml = &state->arcs_list[ARC_BUFC_DATA]; 3593 multilist_t *meta_ml = &state->arcs_list[ARC_BUFC_METADATA]; 3594 int data_idx = multilist_get_random_index(data_ml); 3595 int meta_idx = multilist_get_random_index(meta_ml); 3596 multilist_sublist_t *data_mls; 3597 multilist_sublist_t *meta_mls; 3598 arc_buf_contents_t type; 3599 arc_buf_hdr_t *data_hdr; 3600 arc_buf_hdr_t *meta_hdr; 3601 3602 /* 3603 * We keep the sublist lock until we're finished, to prevent 3604 * the headers from being destroyed via arc_evict_state(). 3605 */ 3606 data_mls = multilist_sublist_lock(data_ml, data_idx); 3607 meta_mls = multilist_sublist_lock(meta_ml, meta_idx); 3608 3609 /* 3610 * These two loops are to ensure we skip any markers that 3611 * might be at the tail of the lists due to arc_evict_state(). 3612 */ 3613 3614 for (data_hdr = multilist_sublist_tail(data_mls); data_hdr != NULL; 3615 data_hdr = multilist_sublist_prev(data_mls, data_hdr)) { 3616 if (data_hdr->b_spa != 0) 3617 break; 3618 } 3619 3620 for (meta_hdr = multilist_sublist_tail(meta_mls); meta_hdr != NULL; 3621 meta_hdr = multilist_sublist_prev(meta_mls, meta_hdr)) { 3622 if (meta_hdr->b_spa != 0) 3623 break; 3624 } 3625 3626 if (data_hdr == NULL && meta_hdr == NULL) { 3627 type = ARC_BUFC_DATA; 3628 } else if (data_hdr == NULL) { 3629 ASSERT3P(meta_hdr, !=, NULL); 3630 type = ARC_BUFC_METADATA; 3631 } else if (meta_hdr == NULL) { 3632 ASSERT3P(data_hdr, !=, NULL); 3633 type = ARC_BUFC_DATA; 3634 } else { 3635 ASSERT3P(data_hdr, !=, NULL); 3636 ASSERT3P(meta_hdr, !=, NULL); 3637 3638 /* The headers can't be on the sublist without an L1 header */ 3639 ASSERT(HDR_HAS_L1HDR(data_hdr)); 3640 ASSERT(HDR_HAS_L1HDR(meta_hdr)); 3641 3642 if (data_hdr->b_l1hdr.b_arc_access < 3643 meta_hdr->b_l1hdr.b_arc_access) { 3644 type = ARC_BUFC_DATA; 3645 } else { 3646 type = ARC_BUFC_METADATA; 3647 } 3648 } 3649 3650 multilist_sublist_unlock(meta_mls); 3651 multilist_sublist_unlock(data_mls); 3652 3653 return (type); 3654 } 3655 3656 /* 3657 * Evict buffers from the cache, such that arc_size is capped by arc_c. 3658 */ 3659 static uint64_t 3660 arc_adjust(void) 3661 { 3662 uint64_t total_evicted = 0; 3663 uint64_t bytes; 3664 int64_t target; 3665 3666 /* 3667 * If we're over arc_meta_limit, we want to correct that before 3668 * potentially evicting data buffers below. 3669 */ 3670 total_evicted += arc_adjust_meta(); 3671 3672 /* 3673 * Adjust MRU size 3674 * 3675 * If we're over the target cache size, we want to evict enough 3676 * from the list to get back to our target size. We don't want 3677 * to evict too much from the MRU, such that it drops below 3678 * arc_p. So, if we're over our target cache size more than 3679 * the MRU is over arc_p, we'll evict enough to get back to 3680 * arc_p here, and then evict more from the MFU below. 3681 */ 3682 target = MIN((int64_t)(arc_size - arc_c), 3683 (int64_t)(refcount_count(&arc_anon->arcs_size) + 3684 refcount_count(&arc_mru->arcs_size) + arc_meta_used - arc_p)); 3685 3686 /* 3687 * If we're below arc_meta_min, always prefer to evict data. 3688 * Otherwise, try to satisfy the requested number of bytes to 3689 * evict from the type which contains older buffers; in an 3690 * effort to keep newer buffers in the cache regardless of their 3691 * type. If we cannot satisfy the number of bytes from this 3692 * type, spill over into the next type. 3693 */ 3694 if (arc_adjust_type(arc_mru) == ARC_BUFC_METADATA && 3695 arc_meta_used > arc_meta_min) { 3696 bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA); 3697 total_evicted += bytes; 3698 3699 /* 3700 * If we couldn't evict our target number of bytes from 3701 * metadata, we try to get the rest from data. 3702 */ 3703 target -= bytes; 3704 3705 total_evicted += 3706 arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA); 3707 } else { 3708 bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA); 3709 total_evicted += bytes; 3710 3711 /* 3712 * If we couldn't evict our target number of bytes from 3713 * data, we try to get the rest from metadata. 3714 */ 3715 target -= bytes; 3716 3717 total_evicted += 3718 arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA); 3719 } 3720 3721 /* 3722 * Adjust MFU size 3723 * 3724 * Now that we've tried to evict enough from the MRU to get its 3725 * size back to arc_p, if we're still above the target cache 3726 * size, we evict the rest from the MFU. 3727 */ 3728 target = arc_size - arc_c; 3729 3730 if (arc_adjust_type(arc_mfu) == ARC_BUFC_METADATA && 3731 arc_meta_used > arc_meta_min) { 3732 bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA); 3733 total_evicted += bytes; 3734 3735 /* 3736 * If we couldn't evict our target number of bytes from 3737 * metadata, we try to get the rest from data. 3738 */ 3739 target -= bytes; 3740 3741 total_evicted += 3742 arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA); 3743 } else { 3744 bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA); 3745 total_evicted += bytes; 3746 3747 /* 3748 * If we couldn't evict our target number of bytes from 3749 * data, we try to get the rest from data. 3750 */ 3751 target -= bytes; 3752 3753 total_evicted += 3754 arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA); 3755 } 3756 3757 /* 3758 * Adjust ghost lists 3759 * 3760 * In addition to the above, the ARC also defines target values 3761 * for the ghost lists. The sum of the mru list and mru ghost 3762 * list should never exceed the target size of the cache, and 3763 * the sum of the mru list, mfu list, mru ghost list, and mfu 3764 * ghost list should never exceed twice the target size of the 3765 * cache. The following logic enforces these limits on the ghost 3766 * caches, and evicts from them as needed. 3767 */ 3768 target = refcount_count(&arc_mru->arcs_size) + 3769 refcount_count(&arc_mru_ghost->arcs_size) - arc_c; 3770 3771 bytes = arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_DATA); 3772 total_evicted += bytes; 3773 3774 target -= bytes; 3775 3776 total_evicted += 3777 arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_METADATA); 3778 3779 /* 3780 * We assume the sum of the mru list and mfu list is less than 3781 * or equal to arc_c (we enforced this above), which means we 3782 * can use the simpler of the two equations below: 3783 * 3784 * mru + mfu + mru ghost + mfu ghost <= 2 * arc_c 3785 * mru ghost + mfu ghost <= arc_c 3786 */ 3787 target = refcount_count(&arc_mru_ghost->arcs_size) + 3788 refcount_count(&arc_mfu_ghost->arcs_size) - arc_c; 3789 3790 bytes = arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_DATA); 3791 total_evicted += bytes; 3792 3793 target -= bytes; 3794 3795 total_evicted += 3796 arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_METADATA); 3797 3798 return (total_evicted); 3799 } 3800 3801 void 3802 arc_flush(spa_t *spa, boolean_t retry) 3803 { 3804 uint64_t guid = 0; 3805 3806 /* 3807 * If retry is B_TRUE, a spa must not be specified since we have 3808 * no good way to determine if all of a spa's buffers have been 3809 * evicted from an arc state. 3810 */ 3811 ASSERT(!retry || spa == 0); 3812 3813 if (spa != NULL) 3814 guid = spa_load_guid(spa); 3815 3816 (void) arc_flush_state(arc_mru, guid, ARC_BUFC_DATA, retry); 3817 (void) arc_flush_state(arc_mru, guid, ARC_BUFC_METADATA, retry); 3818 3819 (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_DATA, retry); 3820 (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_METADATA, retry); 3821 3822 (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_DATA, retry); 3823 (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_METADATA, retry); 3824 3825 (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_DATA, retry); 3826 (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_METADATA, retry); 3827 } 3828 3829 void 3830 arc_shrink(int64_t to_free) 3831 { 3832 if (arc_c > arc_c_min) { 3833 3834 if (arc_c > arc_c_min + to_free) 3835 atomic_add_64(&arc_c, -to_free); 3836 else 3837 arc_c = arc_c_min; 3838 3839 atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift)); 3840 if (arc_c > arc_size) 3841 arc_c = MAX(arc_size, arc_c_min); 3842 if (arc_p > arc_c) 3843 arc_p = (arc_c >> 1); 3844 ASSERT(arc_c >= arc_c_min); 3845 ASSERT((int64_t)arc_p >= 0); 3846 } 3847 3848 if (arc_size > arc_c) 3849 (void) arc_adjust(); 3850 } 3851 3852 typedef enum free_memory_reason_t { 3853 FMR_UNKNOWN, 3854 FMR_NEEDFREE, 3855 FMR_LOTSFREE, 3856 FMR_SWAPFS_MINFREE, 3857 FMR_PAGES_PP_MAXIMUM, 3858 FMR_HEAP_ARENA, 3859 FMR_ZIO_ARENA, 3860 } free_memory_reason_t; 3861 3862 int64_t last_free_memory; 3863 free_memory_reason_t last_free_reason; 3864 3865 /* 3866 * Additional reserve of pages for pp_reserve. 3867 */ 3868 int64_t arc_pages_pp_reserve = 64; 3869 3870 /* 3871 * Additional reserve of pages for swapfs. 3872 */ 3873 int64_t arc_swapfs_reserve = 64; 3874 3875 /* 3876 * Return the amount of memory that can be consumed before reclaim will be 3877 * needed. Positive if there is sufficient free memory, negative indicates 3878 * the amount of memory that needs to be freed up. 3879 */ 3880 static int64_t 3881 arc_available_memory(void) 3882 { 3883 int64_t lowest = INT64_MAX; 3884 int64_t n; 3885 free_memory_reason_t r = FMR_UNKNOWN; 3886 3887 #ifdef _KERNEL 3888 if (needfree > 0) { 3889 n = PAGESIZE * (-needfree); 3890 if (n < lowest) { 3891 lowest = n; 3892 r = FMR_NEEDFREE; 3893 } 3894 } 3895 3896 /* 3897 * check that we're out of range of the pageout scanner. It starts to 3898 * schedule paging if freemem is less than lotsfree and needfree. 3899 * lotsfree is the high-water mark for pageout, and needfree is the 3900 * number of needed free pages. We add extra pages here to make sure 3901 * the scanner doesn't start up while we're freeing memory. 3902 */ 3903 n = PAGESIZE * (freemem - lotsfree - needfree - desfree); 3904 if (n < lowest) { 3905 lowest = n; 3906 r = FMR_LOTSFREE; 3907 } 3908 3909 /* 3910 * check to make sure that swapfs has enough space so that anon 3911 * reservations can still succeed. anon_resvmem() checks that the 3912 * availrmem is greater than swapfs_minfree, and the number of reserved 3913 * swap pages. We also add a bit of extra here just to prevent 3914 * circumstances from getting really dire. 3915 */ 3916 n = PAGESIZE * (availrmem - swapfs_minfree - swapfs_reserve - 3917 desfree - arc_swapfs_reserve); 3918 if (n < lowest) { 3919 lowest = n; 3920 r = FMR_SWAPFS_MINFREE; 3921 } 3922 3923 3924 /* 3925 * Check that we have enough availrmem that memory locking (e.g., via 3926 * mlock(3C) or memcntl(2)) can still succeed. (pages_pp_maximum 3927 * stores the number of pages that cannot be locked; when availrmem 3928 * drops below pages_pp_maximum, page locking mechanisms such as 3929 * page_pp_lock() will fail.) 3930 */ 3931 n = PAGESIZE * (availrmem - pages_pp_maximum - 3932 arc_pages_pp_reserve); 3933 if (n < lowest) { 3934 lowest = n; 3935 r = FMR_PAGES_PP_MAXIMUM; 3936 } 3937 3938 #if defined(__i386) 3939 /* 3940 * If we're on an i386 platform, it's possible that we'll exhaust the 3941 * kernel heap space before we ever run out of available physical 3942 * memory. Most checks of the size of the heap_area compare against 3943 * tune.t_minarmem, which is the minimum available real memory that we 3944 * can have in the system. However, this is generally fixed at 25 pages 3945 * which is so low that it's useless. In this comparison, we seek to 3946 * calculate the total heap-size, and reclaim if more than 3/4ths of the 3947 * heap is allocated. (Or, in the calculation, if less than 1/4th is 3948 * free) 3949 */ 3950 n = (int64_t)vmem_size(heap_arena, VMEM_FREE) - 3951 (vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC) >> 2); 3952 if (n < lowest) { 3953 lowest = n; 3954 r = FMR_HEAP_ARENA; 3955 } 3956 #endif 3957 3958 /* 3959 * If zio data pages are being allocated out of a separate heap segment, 3960 * then enforce that the size of available vmem for this arena remains 3961 * above about 1/4th (1/(2^arc_zio_arena_free_shift)) free. 3962 * 3963 * Note that reducing the arc_zio_arena_free_shift keeps more virtual 3964 * memory (in the zio_arena) free, which can avoid memory 3965 * fragmentation issues. 3966 */ 3967 if (zio_arena != NULL) { 3968 n = (int64_t)vmem_size(zio_arena, VMEM_FREE) - 3969 (vmem_size(zio_arena, VMEM_ALLOC) >> 3970 arc_zio_arena_free_shift); 3971 if (n < lowest) { 3972 lowest = n; 3973 r = FMR_ZIO_ARENA; 3974 } 3975 } 3976 #else 3977 /* Every 100 calls, free a small amount */ 3978 if (spa_get_random(100) == 0) 3979 lowest = -1024; 3980 #endif 3981 3982 last_free_memory = lowest; 3983 last_free_reason = r; 3984 3985 return (lowest); 3986 } 3987 3988 3989 /* 3990 * Determine if the system is under memory pressure and is asking 3991 * to reclaim memory. A return value of B_TRUE indicates that the system 3992 * is under memory pressure and that the arc should adjust accordingly. 3993 */ 3994 static boolean_t 3995 arc_reclaim_needed(void) 3996 { 3997 return (arc_available_memory() < 0); 3998 } 3999 4000 static void 4001 arc_kmem_reap_now(void) 4002 { 4003 size_t i; 4004 kmem_cache_t *prev_cache = NULL; 4005 kmem_cache_t *prev_data_cache = NULL; 4006 extern kmem_cache_t *zio_buf_cache[]; 4007 extern kmem_cache_t *zio_data_buf_cache[]; 4008 extern kmem_cache_t *range_seg_cache; 4009 4010 #ifdef _KERNEL 4011 if (arc_meta_used >= arc_meta_limit) { 4012 /* 4013 * We are exceeding our meta-data cache limit. 4014 * Purge some DNLC entries to release holds on meta-data. 4015 */ 4016 dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent); 4017 } 4018 #if defined(__i386) 4019 /* 4020 * Reclaim unused memory from all kmem caches. 4021 */ 4022 kmem_reap(); 4023 #endif 4024 #endif 4025 4026 for (i = 0; i < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; i++) { 4027 if (zio_buf_cache[i] != prev_cache) { 4028 prev_cache = zio_buf_cache[i]; 4029 kmem_cache_reap_now(zio_buf_cache[i]); 4030 } 4031 if (zio_data_buf_cache[i] != prev_data_cache) { 4032 prev_data_cache = zio_data_buf_cache[i]; 4033 kmem_cache_reap_now(zio_data_buf_cache[i]); 4034 } 4035 } 4036 kmem_cache_reap_now(buf_cache); 4037 kmem_cache_reap_now(hdr_full_cache); 4038 kmem_cache_reap_now(hdr_l2only_cache); 4039 kmem_cache_reap_now(range_seg_cache); 4040 4041 if (zio_arena != NULL) { 4042 /* 4043 * Ask the vmem arena to reclaim unused memory from its 4044 * quantum caches. 4045 */ 4046 vmem_qcache_reap(zio_arena); 4047 } 4048 } 4049 4050 /* 4051 * Threads can block in arc_get_data_buf() waiting for this thread to evict 4052 * enough data and signal them to proceed. When this happens, the threads in 4053 * arc_get_data_buf() are sleeping while holding the hash lock for their 4054 * particular arc header. Thus, we must be careful to never sleep on a 4055 * hash lock in this thread. This is to prevent the following deadlock: 4056 * 4057 * - Thread A sleeps on CV in arc_get_data_buf() holding hash lock "L", 4058 * waiting for the reclaim thread to signal it. 4059 * 4060 * - arc_reclaim_thread() tries to acquire hash lock "L" using mutex_enter, 4061 * fails, and goes to sleep forever. 4062 * 4063 * This possible deadlock is avoided by always acquiring a hash lock 4064 * using mutex_tryenter() from arc_reclaim_thread(). 4065 */ 4066 static void 4067 arc_reclaim_thread(void) 4068 { 4069 hrtime_t growtime = 0; 4070 callb_cpr_t cpr; 4071 4072 CALLB_CPR_INIT(&cpr, &arc_reclaim_lock, callb_generic_cpr, FTAG); 4073 4074 mutex_enter(&arc_reclaim_lock); 4075 while (!arc_reclaim_thread_exit) { 4076 uint64_t evicted = 0; 4077 4078 /* 4079 * This is necessary in order for the mdb ::arc dcmd to 4080 * show up to date information. Since the ::arc command 4081 * does not call the kstat's update function, without 4082 * this call, the command may show stale stats for the 4083 * anon, mru, mru_ghost, mfu, and mfu_ghost lists. Even 4084 * with this change, the data might be up to 1 second 4085 * out of date; but that should suffice. The arc_state_t 4086 * structures can be queried directly if more accurate 4087 * information is needed. 4088 */ 4089 if (arc_ksp != NULL) 4090 arc_ksp->ks_update(arc_ksp, KSTAT_READ); 4091 4092 mutex_exit(&arc_reclaim_lock); 4093 4094 /* 4095 * We call arc_adjust() before (possibly) calling 4096 * arc_kmem_reap_now(), so that we can wake up 4097 * arc_get_data_buf() sooner. 4098 */ 4099 evicted = arc_adjust(); 4100 4101 int64_t free_memory = arc_available_memory(); 4102 if (free_memory < 0) { 4103 4104 arc_no_grow = B_TRUE; 4105 arc_warm = B_TRUE; 4106 4107 /* 4108 * Wait at least zfs_grow_retry (default 60) seconds 4109 * before considering growing. 4110 */ 4111 growtime = gethrtime() + SEC2NSEC(arc_grow_retry); 4112 4113 arc_kmem_reap_now(); 4114 4115 /* 4116 * If we are still low on memory, shrink the ARC 4117 * so that we have arc_shrink_min free space. 4118 */ 4119 free_memory = arc_available_memory(); 4120 4121 int64_t to_free = 4122 (arc_c >> arc_shrink_shift) - free_memory; 4123 if (to_free > 0) { 4124 #ifdef _KERNEL 4125 to_free = MAX(to_free, ptob(needfree)); 4126 #endif 4127 arc_shrink(to_free); 4128 } 4129 } else if (free_memory < arc_c >> arc_no_grow_shift) { 4130 arc_no_grow = B_TRUE; 4131 } else if (gethrtime() >= growtime) { 4132 arc_no_grow = B_FALSE; 4133 } 4134 4135 mutex_enter(&arc_reclaim_lock); 4136 4137 /* 4138 * If evicted is zero, we couldn't evict anything via 4139 * arc_adjust(). This could be due to hash lock 4140 * collisions, but more likely due to the majority of 4141 * arc buffers being unevictable. Therefore, even if 4142 * arc_size is above arc_c, another pass is unlikely to 4143 * be helpful and could potentially cause us to enter an 4144 * infinite loop. 4145 */ 4146 if (arc_size <= arc_c || evicted == 0) { 4147 /* 4148 * We're either no longer overflowing, or we 4149 * can't evict anything more, so we should wake 4150 * up any threads before we go to sleep. 4151 */ 4152 cv_broadcast(&arc_reclaim_waiters_cv); 4153 4154 /* 4155 * Block until signaled, or after one second (we 4156 * might need to perform arc_kmem_reap_now() 4157 * even if we aren't being signalled) 4158 */ 4159 CALLB_CPR_SAFE_BEGIN(&cpr); 4160 (void) cv_timedwait_hires(&arc_reclaim_thread_cv, 4161 &arc_reclaim_lock, SEC2NSEC(1), MSEC2NSEC(1), 0); 4162 CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_lock); 4163 } 4164 } 4165 4166 arc_reclaim_thread_exit = B_FALSE; 4167 cv_broadcast(&arc_reclaim_thread_cv); 4168 CALLB_CPR_EXIT(&cpr); /* drops arc_reclaim_lock */ 4169 thread_exit(); 4170 } 4171 4172 /* 4173 * Adapt arc info given the number of bytes we are trying to add and 4174 * the state that we are comming from. This function is only called 4175 * when we are adding new content to the cache. 4176 */ 4177 static void 4178 arc_adapt(int bytes, arc_state_t *state) 4179 { 4180 int mult; 4181 uint64_t arc_p_min = (arc_c >> arc_p_min_shift); 4182 int64_t mrug_size = refcount_count(&arc_mru_ghost->arcs_size); 4183 int64_t mfug_size = refcount_count(&arc_mfu_ghost->arcs_size); 4184 4185 if (state == arc_l2c_only) 4186 return; 4187 4188 ASSERT(bytes > 0); 4189 /* 4190 * Adapt the target size of the MRU list: 4191 * - if we just hit in the MRU ghost list, then increase 4192 * the target size of the MRU list. 4193 * - if we just hit in the MFU ghost list, then increase 4194 * the target size of the MFU list by decreasing the 4195 * target size of the MRU list. 4196 */ 4197 if (state == arc_mru_ghost) { 4198 mult = (mrug_size >= mfug_size) ? 1 : (mfug_size / mrug_size); 4199 mult = MIN(mult, 10); /* avoid wild arc_p adjustment */ 4200 4201 arc_p = MIN(arc_c - arc_p_min, arc_p + bytes * mult); 4202 } else if (state == arc_mfu_ghost) { 4203 uint64_t delta; 4204 4205 mult = (mfug_size >= mrug_size) ? 1 : (mrug_size / mfug_size); 4206 mult = MIN(mult, 10); 4207 4208 delta = MIN(bytes * mult, arc_p); 4209 arc_p = MAX(arc_p_min, arc_p - delta); 4210 } 4211 ASSERT((int64_t)arc_p >= 0); 4212 4213 if (arc_reclaim_needed()) { 4214 cv_signal(&arc_reclaim_thread_cv); 4215 return; 4216 } 4217 4218 if (arc_no_grow) 4219 return; 4220 4221 if (arc_c >= arc_c_max) 4222 return; 4223 4224 /* 4225 * If we're within (2 * maxblocksize) bytes of the target 4226 * cache size, increment the target cache size 4227 */ 4228 if (arc_size > arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) { 4229 atomic_add_64(&arc_c, (int64_t)bytes); 4230 if (arc_c > arc_c_max) 4231 arc_c = arc_c_max; 4232 else if (state == arc_anon) 4233 atomic_add_64(&arc_p, (int64_t)bytes); 4234 if (arc_p > arc_c) 4235 arc_p = arc_c; 4236 } 4237 ASSERT((int64_t)arc_p >= 0); 4238 } 4239 4240 /* 4241 * Check if arc_size has grown past our upper threshold, determined by 4242 * zfs_arc_overflow_shift. 4243 */ 4244 static boolean_t 4245 arc_is_overflowing(void) 4246 { 4247 /* Always allow at least one block of overflow */ 4248 uint64_t overflow = MAX(SPA_MAXBLOCKSIZE, 4249 arc_c >> zfs_arc_overflow_shift); 4250 4251 return (arc_size >= arc_c + overflow); 4252 } 4253 4254 /* 4255 * Allocate a block and return it to the caller. If we are hitting the 4256 * hard limit for the cache size, we must sleep, waiting for the eviction 4257 * thread to catch up. If we're past the target size but below the hard 4258 * limit, we'll only signal the reclaim thread and continue on. 4259 */ 4260 static void * 4261 arc_get_data_buf(arc_buf_hdr_t *hdr, uint64_t size, void *tag) 4262 { 4263 void *datap = NULL; 4264 arc_state_t *state = hdr->b_l1hdr.b_state; 4265 arc_buf_contents_t type = arc_buf_type(hdr); 4266 4267 arc_adapt(size, state); 4268 4269 /* 4270 * If arc_size is currently overflowing, and has grown past our 4271 * upper limit, we must be adding data faster than the evict 4272 * thread can evict. Thus, to ensure we don't compound the 4273 * problem by adding more data and forcing arc_size to grow even 4274 * further past it's target size, we halt and wait for the 4275 * eviction thread to catch up. 4276 * 4277 * It's also possible that the reclaim thread is unable to evict 4278 * enough buffers to get arc_size below the overflow limit (e.g. 4279 * due to buffers being un-evictable, or hash lock collisions). 4280 * In this case, we want to proceed regardless if we're 4281 * overflowing; thus we don't use a while loop here. 4282 */ 4283 if (arc_is_overflowing()) { 4284 mutex_enter(&arc_reclaim_lock); 4285 4286 /* 4287 * Now that we've acquired the lock, we may no longer be 4288 * over the overflow limit, lets check. 4289 * 4290 * We're ignoring the case of spurious wake ups. If that 4291 * were to happen, it'd let this thread consume an ARC 4292 * buffer before it should have (i.e. before we're under 4293 * the overflow limit and were signalled by the reclaim 4294 * thread). As long as that is a rare occurrence, it 4295 * shouldn't cause any harm. 4296 */ 4297 if (arc_is_overflowing()) { 4298 cv_signal(&arc_reclaim_thread_cv); 4299 cv_wait(&arc_reclaim_waiters_cv, &arc_reclaim_lock); 4300 } 4301 4302 mutex_exit(&arc_reclaim_lock); 4303 } 4304 4305 VERIFY3U(hdr->b_type, ==, type); 4306 if (type == ARC_BUFC_METADATA) { 4307 datap = zio_buf_alloc(size); 4308 arc_space_consume(size, ARC_SPACE_META); 4309 } else { 4310 ASSERT(type == ARC_BUFC_DATA); 4311 datap = zio_data_buf_alloc(size); 4312 arc_space_consume(size, ARC_SPACE_DATA); 4313 } 4314 4315 /* 4316 * Update the state size. Note that ghost states have a 4317 * "ghost size" and so don't need to be updated. 4318 */ 4319 if (!GHOST_STATE(state)) { 4320 4321 (void) refcount_add_many(&state->arcs_size, size, tag); 4322 4323 /* 4324 * If this is reached via arc_read, the link is 4325 * protected by the hash lock. If reached via 4326 * arc_buf_alloc, the header should not be accessed by 4327 * any other thread. And, if reached via arc_read_done, 4328 * the hash lock will protect it if it's found in the 4329 * hash table; otherwise no other thread should be 4330 * trying to [add|remove]_reference it. 4331 */ 4332 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) { 4333 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 4334 (void) refcount_add_many(&state->arcs_esize[type], 4335 size, tag); 4336 } 4337 4338 /* 4339 * If we are growing the cache, and we are adding anonymous 4340 * data, and we have outgrown arc_p, update arc_p 4341 */ 4342 if (arc_size < arc_c && hdr->b_l1hdr.b_state == arc_anon && 4343 (refcount_count(&arc_anon->arcs_size) + 4344 refcount_count(&arc_mru->arcs_size) > arc_p)) 4345 arc_p = MIN(arc_c, arc_p + size); 4346 } 4347 return (datap); 4348 } 4349 4350 /* 4351 * Free the arc data buffer. 4352 */ 4353 static void 4354 arc_free_data_buf(arc_buf_hdr_t *hdr, void *data, uint64_t size, void *tag) 4355 { 4356 arc_state_t *state = hdr->b_l1hdr.b_state; 4357 arc_buf_contents_t type = arc_buf_type(hdr); 4358 4359 /* protected by hash lock, if in the hash table */ 4360 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) { 4361 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 4362 ASSERT(state != arc_anon && state != arc_l2c_only); 4363 4364 (void) refcount_remove_many(&state->arcs_esize[type], 4365 size, tag); 4366 } 4367 (void) refcount_remove_many(&state->arcs_size, size, tag); 4368 4369 VERIFY3U(hdr->b_type, ==, type); 4370 if (type == ARC_BUFC_METADATA) { 4371 zio_buf_free(data, size); 4372 arc_space_return(size, ARC_SPACE_META); 4373 } else { 4374 ASSERT(type == ARC_BUFC_DATA); 4375 zio_data_buf_free(data, size); 4376 arc_space_return(size, ARC_SPACE_DATA); 4377 } 4378 } 4379 4380 /* 4381 * This routine is called whenever a buffer is accessed. 4382 * NOTE: the hash lock is dropped in this function. 4383 */ 4384 static void 4385 arc_access(arc_buf_hdr_t *hdr, kmutex_t *hash_lock) 4386 { 4387 clock_t now; 4388 4389 ASSERT(MUTEX_HELD(hash_lock)); 4390 ASSERT(HDR_HAS_L1HDR(hdr)); 4391 4392 if (hdr->b_l1hdr.b_state == arc_anon) { 4393 /* 4394 * This buffer is not in the cache, and does not 4395 * appear in our "ghost" list. Add the new buffer 4396 * to the MRU state. 4397 */ 4398 4399 ASSERT0(hdr->b_l1hdr.b_arc_access); 4400 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 4401 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr); 4402 arc_change_state(arc_mru, hdr, hash_lock); 4403 4404 } else if (hdr->b_l1hdr.b_state == arc_mru) { 4405 now = ddi_get_lbolt(); 4406 4407 /* 4408 * If this buffer is here because of a prefetch, then either: 4409 * - clear the flag if this is a "referencing" read 4410 * (any subsequent access will bump this into the MFU state). 4411 * or 4412 * - move the buffer to the head of the list if this is 4413 * another prefetch (to make it less likely to be evicted). 4414 */ 4415 if (HDR_PREFETCH(hdr)) { 4416 if (refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) { 4417 /* link protected by hash lock */ 4418 ASSERT(multilist_link_active( 4419 &hdr->b_l1hdr.b_arc_node)); 4420 } else { 4421 arc_hdr_clear_flags(hdr, ARC_FLAG_PREFETCH); 4422 ARCSTAT_BUMP(arcstat_mru_hits); 4423 } 4424 hdr->b_l1hdr.b_arc_access = now; 4425 return; 4426 } 4427 4428 /* 4429 * This buffer has been "accessed" only once so far, 4430 * but it is still in the cache. Move it to the MFU 4431 * state. 4432 */ 4433 if (now > hdr->b_l1hdr.b_arc_access + ARC_MINTIME) { 4434 /* 4435 * More than 125ms have passed since we 4436 * instantiated this buffer. Move it to the 4437 * most frequently used state. 4438 */ 4439 hdr->b_l1hdr.b_arc_access = now; 4440 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 4441 arc_change_state(arc_mfu, hdr, hash_lock); 4442 } 4443 ARCSTAT_BUMP(arcstat_mru_hits); 4444 } else if (hdr->b_l1hdr.b_state == arc_mru_ghost) { 4445 arc_state_t *new_state; 4446 /* 4447 * This buffer has been "accessed" recently, but 4448 * was evicted from the cache. Move it to the 4449 * MFU state. 4450 */ 4451 4452 if (HDR_PREFETCH(hdr)) { 4453 new_state = arc_mru; 4454 if (refcount_count(&hdr->b_l1hdr.b_refcnt) > 0) 4455 arc_hdr_clear_flags(hdr, ARC_FLAG_PREFETCH); 4456 DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr); 4457 } else { 4458 new_state = arc_mfu; 4459 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 4460 } 4461 4462 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 4463 arc_change_state(new_state, hdr, hash_lock); 4464 4465 ARCSTAT_BUMP(arcstat_mru_ghost_hits); 4466 } else if (hdr->b_l1hdr.b_state == arc_mfu) { 4467 /* 4468 * This buffer has been accessed more than once and is 4469 * still in the cache. Keep it in the MFU state. 4470 * 4471 * NOTE: an add_reference() that occurred when we did 4472 * the arc_read() will have kicked this off the list. 4473 * If it was a prefetch, we will explicitly move it to 4474 * the head of the list now. 4475 */ 4476 if ((HDR_PREFETCH(hdr)) != 0) { 4477 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 4478 /* link protected by hash_lock */ 4479 ASSERT(multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 4480 } 4481 ARCSTAT_BUMP(arcstat_mfu_hits); 4482 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 4483 } else if (hdr->b_l1hdr.b_state == arc_mfu_ghost) { 4484 arc_state_t *new_state = arc_mfu; 4485 /* 4486 * This buffer has been accessed more than once but has 4487 * been evicted from the cache. Move it back to the 4488 * MFU state. 4489 */ 4490 4491 if (HDR_PREFETCH(hdr)) { 4492 /* 4493 * This is a prefetch access... 4494 * move this block back to the MRU state. 4495 */ 4496 ASSERT0(refcount_count(&hdr->b_l1hdr.b_refcnt)); 4497 new_state = arc_mru; 4498 } 4499 4500 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 4501 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 4502 arc_change_state(new_state, hdr, hash_lock); 4503 4504 ARCSTAT_BUMP(arcstat_mfu_ghost_hits); 4505 } else if (hdr->b_l1hdr.b_state == arc_l2c_only) { 4506 /* 4507 * This buffer is on the 2nd Level ARC. 4508 */ 4509 4510 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt(); 4511 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr); 4512 arc_change_state(arc_mfu, hdr, hash_lock); 4513 } else { 4514 ASSERT(!"invalid arc state"); 4515 } 4516 } 4517 4518 /* a generic arc_done_func_t which you can use */ 4519 /* ARGSUSED */ 4520 void 4521 arc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg) 4522 { 4523 if (zio == NULL || zio->io_error == 0) 4524 bcopy(buf->b_data, arg, arc_buf_size(buf)); 4525 arc_buf_destroy(buf, arg); 4526 } 4527 4528 /* a generic arc_done_func_t */ 4529 void 4530 arc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg) 4531 { 4532 arc_buf_t **bufp = arg; 4533 if (zio && zio->io_error) { 4534 arc_buf_destroy(buf, arg); 4535 *bufp = NULL; 4536 } else { 4537 *bufp = buf; 4538 ASSERT(buf->b_data); 4539 } 4540 } 4541 4542 static void 4543 arc_hdr_verify(arc_buf_hdr_t *hdr, blkptr_t *bp) 4544 { 4545 if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp)) { 4546 ASSERT3U(HDR_GET_PSIZE(hdr), ==, 0); 4547 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF); 4548 } else { 4549 if (HDR_COMPRESSION_ENABLED(hdr)) { 4550 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, 4551 BP_GET_COMPRESS(bp)); 4552 } 4553 ASSERT3U(HDR_GET_LSIZE(hdr), ==, BP_GET_LSIZE(bp)); 4554 ASSERT3U(HDR_GET_PSIZE(hdr), ==, BP_GET_PSIZE(bp)); 4555 } 4556 } 4557 4558 static void 4559 arc_read_done(zio_t *zio) 4560 { 4561 arc_buf_hdr_t *hdr = zio->io_private; 4562 kmutex_t *hash_lock = NULL; 4563 arc_callback_t *callback_list; 4564 arc_callback_t *acb; 4565 boolean_t freeable = B_FALSE; 4566 boolean_t no_zio_error = (zio->io_error == 0); 4567 4568 /* 4569 * The hdr was inserted into hash-table and removed from lists 4570 * prior to starting I/O. We should find this header, since 4571 * it's in the hash table, and it should be legit since it's 4572 * not possible to evict it during the I/O. The only possible 4573 * reason for it not to be found is if we were freed during the 4574 * read. 4575 */ 4576 if (HDR_IN_HASH_TABLE(hdr)) { 4577 ASSERT3U(hdr->b_birth, ==, BP_PHYSICAL_BIRTH(zio->io_bp)); 4578 ASSERT3U(hdr->b_dva.dva_word[0], ==, 4579 BP_IDENTITY(zio->io_bp)->dva_word[0]); 4580 ASSERT3U(hdr->b_dva.dva_word[1], ==, 4581 BP_IDENTITY(zio->io_bp)->dva_word[1]); 4582 4583 arc_buf_hdr_t *found = buf_hash_find(hdr->b_spa, zio->io_bp, 4584 &hash_lock); 4585 4586 ASSERT((found == hdr && 4587 DVA_EQUAL(&hdr->b_dva, BP_IDENTITY(zio->io_bp))) || 4588 (found == hdr && HDR_L2_READING(hdr))); 4589 ASSERT3P(hash_lock, !=, NULL); 4590 } 4591 4592 if (no_zio_error) { 4593 /* byteswap if necessary */ 4594 if (BP_SHOULD_BYTESWAP(zio->io_bp)) { 4595 if (BP_GET_LEVEL(zio->io_bp) > 0) { 4596 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_UINT64; 4597 } else { 4598 hdr->b_l1hdr.b_byteswap = 4599 DMU_OT_BYTESWAP(BP_GET_TYPE(zio->io_bp)); 4600 } 4601 } else { 4602 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS; 4603 } 4604 } 4605 4606 arc_hdr_clear_flags(hdr, ARC_FLAG_L2_EVICTED); 4607 if (l2arc_noprefetch && HDR_PREFETCH(hdr)) 4608 arc_hdr_clear_flags(hdr, ARC_FLAG_L2CACHE); 4609 4610 callback_list = hdr->b_l1hdr.b_acb; 4611 ASSERT3P(callback_list, !=, NULL); 4612 4613 if (hash_lock && no_zio_error && hdr->b_l1hdr.b_state == arc_anon) { 4614 /* 4615 * Only call arc_access on anonymous buffers. This is because 4616 * if we've issued an I/O for an evicted buffer, we've already 4617 * called arc_access (to prevent any simultaneous readers from 4618 * getting confused). 4619 */ 4620 arc_access(hdr, hash_lock); 4621 } 4622 4623 /* 4624 * If a read request has a callback (i.e. acb_done is not NULL), then we 4625 * make a buf containing the data according to the parameters which were 4626 * passed in. The implementation of arc_buf_alloc_impl() ensures that we 4627 * aren't needlessly decompressing the data multiple times. 4628 */ 4629 int callback_cnt = 0; 4630 for (acb = callback_list; acb != NULL; acb = acb->acb_next) { 4631 if (!acb->acb_done) 4632 continue; 4633 4634 /* This is a demand read since prefetches don't use callbacks */ 4635 callback_cnt++; 4636 4637 int error = arc_buf_alloc_impl(hdr, acb->acb_private, 4638 acb->acb_compressed, no_zio_error, &acb->acb_buf); 4639 if (no_zio_error) { 4640 zio->io_error = error; 4641 } 4642 } 4643 hdr->b_l1hdr.b_acb = NULL; 4644 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS); 4645 if (callback_cnt == 0) { 4646 ASSERT(HDR_PREFETCH(hdr)); 4647 ASSERT0(hdr->b_l1hdr.b_bufcnt); 4648 ASSERT3P(hdr->b_l1hdr.b_pdata, !=, NULL); 4649 } 4650 4651 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt) || 4652 callback_list != NULL); 4653 4654 if (no_zio_error) { 4655 arc_hdr_verify(hdr, zio->io_bp); 4656 } else { 4657 arc_hdr_set_flags(hdr, ARC_FLAG_IO_ERROR); 4658 if (hdr->b_l1hdr.b_state != arc_anon) 4659 arc_change_state(arc_anon, hdr, hash_lock); 4660 if (HDR_IN_HASH_TABLE(hdr)) 4661 buf_hash_remove(hdr); 4662 freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt); 4663 } 4664 4665 /* 4666 * Broadcast before we drop the hash_lock to avoid the possibility 4667 * that the hdr (and hence the cv) might be freed before we get to 4668 * the cv_broadcast(). 4669 */ 4670 cv_broadcast(&hdr->b_l1hdr.b_cv); 4671 4672 if (hash_lock != NULL) { 4673 mutex_exit(hash_lock); 4674 } else { 4675 /* 4676 * This block was freed while we waited for the read to 4677 * complete. It has been removed from the hash table and 4678 * moved to the anonymous state (so that it won't show up 4679 * in the cache). 4680 */ 4681 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon); 4682 freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt); 4683 } 4684 4685 /* execute each callback and free its structure */ 4686 while ((acb = callback_list) != NULL) { 4687 if (acb->acb_done) 4688 acb->acb_done(zio, acb->acb_buf, acb->acb_private); 4689 4690 if (acb->acb_zio_dummy != NULL) { 4691 acb->acb_zio_dummy->io_error = zio->io_error; 4692 zio_nowait(acb->acb_zio_dummy); 4693 } 4694 4695 callback_list = acb->acb_next; 4696 kmem_free(acb, sizeof (arc_callback_t)); 4697 } 4698 4699 if (freeable) 4700 arc_hdr_destroy(hdr); 4701 } 4702 4703 /* 4704 * "Read" the block at the specified DVA (in bp) via the 4705 * cache. If the block is found in the cache, invoke the provided 4706 * callback immediately and return. Note that the `zio' parameter 4707 * in the callback will be NULL in this case, since no IO was 4708 * required. If the block is not in the cache pass the read request 4709 * on to the spa with a substitute callback function, so that the 4710 * requested block will be added to the cache. 4711 * 4712 * If a read request arrives for a block that has a read in-progress, 4713 * either wait for the in-progress read to complete (and return the 4714 * results); or, if this is a read with a "done" func, add a record 4715 * to the read to invoke the "done" func when the read completes, 4716 * and return; or just return. 4717 * 4718 * arc_read_done() will invoke all the requested "done" functions 4719 * for readers of this block. 4720 */ 4721 int 4722 arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_done_func_t *done, 4723 void *private, zio_priority_t priority, int zio_flags, 4724 arc_flags_t *arc_flags, const zbookmark_phys_t *zb) 4725 { 4726 arc_buf_hdr_t *hdr = NULL; 4727 kmutex_t *hash_lock = NULL; 4728 zio_t *rzio; 4729 uint64_t guid = spa_load_guid(spa); 4730 boolean_t compressed_read = (zio_flags & ZIO_FLAG_RAW) != 0; 4731 4732 ASSERT(!BP_IS_EMBEDDED(bp) || 4733 BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA); 4734 4735 top: 4736 if (!BP_IS_EMBEDDED(bp)) { 4737 /* 4738 * Embedded BP's have no DVA and require no I/O to "read". 4739 * Create an anonymous arc buf to back it. 4740 */ 4741 hdr = buf_hash_find(guid, bp, &hash_lock); 4742 } 4743 4744 if (hdr != NULL && HDR_HAS_L1HDR(hdr) && hdr->b_l1hdr.b_pdata != NULL) { 4745 arc_buf_t *buf = NULL; 4746 *arc_flags |= ARC_FLAG_CACHED; 4747 4748 if (HDR_IO_IN_PROGRESS(hdr)) { 4749 4750 if ((hdr->b_flags & ARC_FLAG_PRIO_ASYNC_READ) && 4751 priority == ZIO_PRIORITY_SYNC_READ) { 4752 /* 4753 * This sync read must wait for an 4754 * in-progress async read (e.g. a predictive 4755 * prefetch). Async reads are queued 4756 * separately at the vdev_queue layer, so 4757 * this is a form of priority inversion. 4758 * Ideally, we would "inherit" the demand 4759 * i/o's priority by moving the i/o from 4760 * the async queue to the synchronous queue, 4761 * but there is currently no mechanism to do 4762 * so. Track this so that we can evaluate 4763 * the magnitude of this potential performance 4764 * problem. 4765 * 4766 * Note that if the prefetch i/o is already 4767 * active (has been issued to the device), 4768 * the prefetch improved performance, because 4769 * we issued it sooner than we would have 4770 * without the prefetch. 4771 */ 4772 DTRACE_PROBE1(arc__sync__wait__for__async, 4773 arc_buf_hdr_t *, hdr); 4774 ARCSTAT_BUMP(arcstat_sync_wait_for_async); 4775 } 4776 if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) { 4777 arc_hdr_clear_flags(hdr, 4778 ARC_FLAG_PREDICTIVE_PREFETCH); 4779 } 4780 4781 if (*arc_flags & ARC_FLAG_WAIT) { 4782 cv_wait(&hdr->b_l1hdr.b_cv, hash_lock); 4783 mutex_exit(hash_lock); 4784 goto top; 4785 } 4786 ASSERT(*arc_flags & ARC_FLAG_NOWAIT); 4787 4788 if (done) { 4789 arc_callback_t *acb = NULL; 4790 4791 acb = kmem_zalloc(sizeof (arc_callback_t), 4792 KM_SLEEP); 4793 acb->acb_done = done; 4794 acb->acb_private = private; 4795 acb->acb_compressed = compressed_read; 4796 if (pio != NULL) 4797 acb->acb_zio_dummy = zio_null(pio, 4798 spa, NULL, NULL, NULL, zio_flags); 4799 4800 ASSERT3P(acb->acb_done, !=, NULL); 4801 acb->acb_next = hdr->b_l1hdr.b_acb; 4802 hdr->b_l1hdr.b_acb = acb; 4803 mutex_exit(hash_lock); 4804 return (0); 4805 } 4806 mutex_exit(hash_lock); 4807 return (0); 4808 } 4809 4810 ASSERT(hdr->b_l1hdr.b_state == arc_mru || 4811 hdr->b_l1hdr.b_state == arc_mfu); 4812 4813 if (done) { 4814 if (hdr->b_flags & ARC_FLAG_PREDICTIVE_PREFETCH) { 4815 /* 4816 * This is a demand read which does not have to 4817 * wait for i/o because we did a predictive 4818 * prefetch i/o for it, which has completed. 4819 */ 4820 DTRACE_PROBE1( 4821 arc__demand__hit__predictive__prefetch, 4822 arc_buf_hdr_t *, hdr); 4823 ARCSTAT_BUMP( 4824 arcstat_demand_hit_predictive_prefetch); 4825 arc_hdr_clear_flags(hdr, 4826 ARC_FLAG_PREDICTIVE_PREFETCH); 4827 } 4828 ASSERT(!BP_IS_EMBEDDED(bp) || !BP_IS_HOLE(bp)); 4829 4830 /* Get a buf with the desired data in it. */ 4831 VERIFY0(arc_buf_alloc_impl(hdr, private, 4832 compressed_read, B_TRUE, &buf)); 4833 } else if (*arc_flags & ARC_FLAG_PREFETCH && 4834 refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) { 4835 arc_hdr_set_flags(hdr, ARC_FLAG_PREFETCH); 4836 } 4837 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr); 4838 arc_access(hdr, hash_lock); 4839 if (*arc_flags & ARC_FLAG_L2CACHE) 4840 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE); 4841 mutex_exit(hash_lock); 4842 ARCSTAT_BUMP(arcstat_hits); 4843 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr), 4844 demand, prefetch, !HDR_ISTYPE_METADATA(hdr), 4845 data, metadata, hits); 4846 4847 if (done) 4848 done(NULL, buf, private); 4849 } else { 4850 uint64_t lsize = BP_GET_LSIZE(bp); 4851 uint64_t psize = BP_GET_PSIZE(bp); 4852 arc_callback_t *acb; 4853 vdev_t *vd = NULL; 4854 uint64_t addr = 0; 4855 boolean_t devw = B_FALSE; 4856 uint64_t size; 4857 4858 if (hdr == NULL) { 4859 /* this block is not in the cache */ 4860 arc_buf_hdr_t *exists = NULL; 4861 arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp); 4862 hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize, 4863 BP_GET_COMPRESS(bp), type); 4864 4865 if (!BP_IS_EMBEDDED(bp)) { 4866 hdr->b_dva = *BP_IDENTITY(bp); 4867 hdr->b_birth = BP_PHYSICAL_BIRTH(bp); 4868 exists = buf_hash_insert(hdr, &hash_lock); 4869 } 4870 if (exists != NULL) { 4871 /* somebody beat us to the hash insert */ 4872 mutex_exit(hash_lock); 4873 buf_discard_identity(hdr); 4874 arc_hdr_destroy(hdr); 4875 goto top; /* restart the IO request */ 4876 } 4877 } else { 4878 /* 4879 * This block is in the ghost cache. If it was L2-only 4880 * (and thus didn't have an L1 hdr), we realloc the 4881 * header to add an L1 hdr. 4882 */ 4883 if (!HDR_HAS_L1HDR(hdr)) { 4884 hdr = arc_hdr_realloc(hdr, hdr_l2only_cache, 4885 hdr_full_cache); 4886 } 4887 ASSERT3P(hdr->b_l1hdr.b_pdata, ==, NULL); 4888 ASSERT(GHOST_STATE(hdr->b_l1hdr.b_state)); 4889 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 4890 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 4891 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL); 4892 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL); 4893 4894 /* 4895 * This is a delicate dance that we play here. 4896 * This hdr is in the ghost list so we access it 4897 * to move it out of the ghost list before we 4898 * initiate the read. If it's a prefetch then 4899 * it won't have a callback so we'll remove the 4900 * reference that arc_buf_alloc_impl() created. We 4901 * do this after we've called arc_access() to 4902 * avoid hitting an assert in remove_reference(). 4903 */ 4904 arc_access(hdr, hash_lock); 4905 arc_hdr_alloc_pdata(hdr); 4906 } 4907 ASSERT3P(hdr->b_l1hdr.b_pdata, !=, NULL); 4908 size = arc_hdr_size(hdr); 4909 4910 /* 4911 * If compression is enabled on the hdr, then will do 4912 * RAW I/O and will store the compressed data in the hdr's 4913 * data block. Otherwise, the hdr's data block will contain 4914 * the uncompressed data. 4915 */ 4916 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF) { 4917 zio_flags |= ZIO_FLAG_RAW; 4918 } 4919 4920 if (*arc_flags & ARC_FLAG_PREFETCH) 4921 arc_hdr_set_flags(hdr, ARC_FLAG_PREFETCH); 4922 if (*arc_flags & ARC_FLAG_L2CACHE) 4923 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE); 4924 if (BP_GET_LEVEL(bp) > 0) 4925 arc_hdr_set_flags(hdr, ARC_FLAG_INDIRECT); 4926 if (*arc_flags & ARC_FLAG_PREDICTIVE_PREFETCH) 4927 arc_hdr_set_flags(hdr, ARC_FLAG_PREDICTIVE_PREFETCH); 4928 ASSERT(!GHOST_STATE(hdr->b_l1hdr.b_state)); 4929 4930 acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP); 4931 acb->acb_done = done; 4932 acb->acb_private = private; 4933 acb->acb_compressed = compressed_read; 4934 4935 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL); 4936 hdr->b_l1hdr.b_acb = acb; 4937 arc_hdr_set_flags(hdr, ARC_FLAG_IO_IN_PROGRESS); 4938 4939 if (HDR_HAS_L2HDR(hdr) && 4940 (vd = hdr->b_l2hdr.b_dev->l2ad_vdev) != NULL) { 4941 devw = hdr->b_l2hdr.b_dev->l2ad_writing; 4942 addr = hdr->b_l2hdr.b_daddr; 4943 /* 4944 * Lock out device removal. 4945 */ 4946 if (vdev_is_dead(vd) || 4947 !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER)) 4948 vd = NULL; 4949 } 4950 4951 if (priority == ZIO_PRIORITY_ASYNC_READ) 4952 arc_hdr_set_flags(hdr, ARC_FLAG_PRIO_ASYNC_READ); 4953 else 4954 arc_hdr_clear_flags(hdr, ARC_FLAG_PRIO_ASYNC_READ); 4955 4956 if (hash_lock != NULL) 4957 mutex_exit(hash_lock); 4958 4959 /* 4960 * At this point, we have a level 1 cache miss. Try again in 4961 * L2ARC if possible. 4962 */ 4963 ASSERT3U(HDR_GET_LSIZE(hdr), ==, lsize); 4964 4965 DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp, 4966 uint64_t, lsize, zbookmark_phys_t *, zb); 4967 ARCSTAT_BUMP(arcstat_misses); 4968 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr), 4969 demand, prefetch, !HDR_ISTYPE_METADATA(hdr), 4970 data, metadata, misses); 4971 4972 if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) { 4973 /* 4974 * Read from the L2ARC if the following are true: 4975 * 1. The L2ARC vdev was previously cached. 4976 * 2. This buffer still has L2ARC metadata. 4977 * 3. This buffer isn't currently writing to the L2ARC. 4978 * 4. The L2ARC entry wasn't evicted, which may 4979 * also have invalidated the vdev. 4980 * 5. This isn't prefetch and l2arc_noprefetch is set. 4981 */ 4982 if (HDR_HAS_L2HDR(hdr) && 4983 !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) && 4984 !(l2arc_noprefetch && HDR_PREFETCH(hdr))) { 4985 l2arc_read_callback_t *cb; 4986 4987 DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr); 4988 ARCSTAT_BUMP(arcstat_l2_hits); 4989 4990 cb = kmem_zalloc(sizeof (l2arc_read_callback_t), 4991 KM_SLEEP); 4992 cb->l2rcb_hdr = hdr; 4993 cb->l2rcb_bp = *bp; 4994 cb->l2rcb_zb = *zb; 4995 cb->l2rcb_flags = zio_flags; 4996 4997 ASSERT(addr >= VDEV_LABEL_START_SIZE && 4998 addr + lsize < vd->vdev_psize - 4999 VDEV_LABEL_END_SIZE); 5000 5001 /* 5002 * l2arc read. The SCL_L2ARC lock will be 5003 * released by l2arc_read_done(). 5004 * Issue a null zio if the underlying buffer 5005 * was squashed to zero size by compression. 5006 */ 5007 ASSERT3U(HDR_GET_COMPRESS(hdr), !=, 5008 ZIO_COMPRESS_EMPTY); 5009 rzio = zio_read_phys(pio, vd, addr, 5010 size, hdr->b_l1hdr.b_pdata, 5011 ZIO_CHECKSUM_OFF, 5012 l2arc_read_done, cb, priority, 5013 zio_flags | ZIO_FLAG_DONT_CACHE | 5014 ZIO_FLAG_CANFAIL | 5015 ZIO_FLAG_DONT_PROPAGATE | 5016 ZIO_FLAG_DONT_RETRY, B_FALSE); 5017 DTRACE_PROBE2(l2arc__read, vdev_t *, vd, 5018 zio_t *, rzio); 5019 ARCSTAT_INCR(arcstat_l2_read_bytes, size); 5020 5021 if (*arc_flags & ARC_FLAG_NOWAIT) { 5022 zio_nowait(rzio); 5023 return (0); 5024 } 5025 5026 ASSERT(*arc_flags & ARC_FLAG_WAIT); 5027 if (zio_wait(rzio) == 0) 5028 return (0); 5029 5030 /* l2arc read error; goto zio_read() */ 5031 } else { 5032 DTRACE_PROBE1(l2arc__miss, 5033 arc_buf_hdr_t *, hdr); 5034 ARCSTAT_BUMP(arcstat_l2_misses); 5035 if (HDR_L2_WRITING(hdr)) 5036 ARCSTAT_BUMP(arcstat_l2_rw_clash); 5037 spa_config_exit(spa, SCL_L2ARC, vd); 5038 } 5039 } else { 5040 if (vd != NULL) 5041 spa_config_exit(spa, SCL_L2ARC, vd); 5042 if (l2arc_ndev != 0) { 5043 DTRACE_PROBE1(l2arc__miss, 5044 arc_buf_hdr_t *, hdr); 5045 ARCSTAT_BUMP(arcstat_l2_misses); 5046 } 5047 } 5048 5049 rzio = zio_read(pio, spa, bp, hdr->b_l1hdr.b_pdata, size, 5050 arc_read_done, hdr, priority, zio_flags, zb); 5051 5052 if (*arc_flags & ARC_FLAG_WAIT) 5053 return (zio_wait(rzio)); 5054 5055 ASSERT(*arc_flags & ARC_FLAG_NOWAIT); 5056 zio_nowait(rzio); 5057 } 5058 return (0); 5059 } 5060 5061 /* 5062 * Notify the arc that a block was freed, and thus will never be used again. 5063 */ 5064 void 5065 arc_freed(spa_t *spa, const blkptr_t *bp) 5066 { 5067 arc_buf_hdr_t *hdr; 5068 kmutex_t *hash_lock; 5069 uint64_t guid = spa_load_guid(spa); 5070 5071 ASSERT(!BP_IS_EMBEDDED(bp)); 5072 5073 hdr = buf_hash_find(guid, bp, &hash_lock); 5074 if (hdr == NULL) 5075 return; 5076 5077 /* 5078 * We might be trying to free a block that is still doing I/O 5079 * (i.e. prefetch) or has a reference (i.e. a dedup-ed, 5080 * dmu_sync-ed block). If this block is being prefetched, then it 5081 * would still have the ARC_FLAG_IO_IN_PROGRESS flag set on the hdr 5082 * until the I/O completes. A block may also have a reference if it is 5083 * part of a dedup-ed, dmu_synced write. The dmu_sync() function would 5084 * have written the new block to its final resting place on disk but 5085 * without the dedup flag set. This would have left the hdr in the MRU 5086 * state and discoverable. When the txg finally syncs it detects that 5087 * the block was overridden in open context and issues an override I/O. 5088 * Since this is a dedup block, the override I/O will determine if the 5089 * block is already in the DDT. If so, then it will replace the io_bp 5090 * with the bp from the DDT and allow the I/O to finish. When the I/O 5091 * reaches the done callback, dbuf_write_override_done, it will 5092 * check to see if the io_bp and io_bp_override are identical. 5093 * If they are not, then it indicates that the bp was replaced with 5094 * the bp in the DDT and the override bp is freed. This allows 5095 * us to arrive here with a reference on a block that is being 5096 * freed. So if we have an I/O in progress, or a reference to 5097 * this hdr, then we don't destroy the hdr. 5098 */ 5099 if (!HDR_HAS_L1HDR(hdr) || (!HDR_IO_IN_PROGRESS(hdr) && 5100 refcount_is_zero(&hdr->b_l1hdr.b_refcnt))) { 5101 arc_change_state(arc_anon, hdr, hash_lock); 5102 arc_hdr_destroy(hdr); 5103 mutex_exit(hash_lock); 5104 } else { 5105 mutex_exit(hash_lock); 5106 } 5107 5108 } 5109 5110 /* 5111 * Release this buffer from the cache, making it an anonymous buffer. This 5112 * must be done after a read and prior to modifying the buffer contents. 5113 * If the buffer has more than one reference, we must make 5114 * a new hdr for the buffer. 5115 */ 5116 void 5117 arc_release(arc_buf_t *buf, void *tag) 5118 { 5119 arc_buf_hdr_t *hdr = buf->b_hdr; 5120 5121 /* 5122 * It would be nice to assert that if it's DMU metadata (level > 5123 * 0 || it's the dnode file), then it must be syncing context. 5124 * But we don't know that information at this level. 5125 */ 5126 5127 mutex_enter(&buf->b_evict_lock); 5128 5129 ASSERT(HDR_HAS_L1HDR(hdr)); 5130 5131 /* 5132 * We don't grab the hash lock prior to this check, because if 5133 * the buffer's header is in the arc_anon state, it won't be 5134 * linked into the hash table. 5135 */ 5136 if (hdr->b_l1hdr.b_state == arc_anon) { 5137 mutex_exit(&buf->b_evict_lock); 5138 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 5139 ASSERT(!HDR_IN_HASH_TABLE(hdr)); 5140 ASSERT(!HDR_HAS_L2HDR(hdr)); 5141 ASSERT(HDR_EMPTY(hdr)); 5142 5143 ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1); 5144 ASSERT3S(refcount_count(&hdr->b_l1hdr.b_refcnt), ==, 1); 5145 ASSERT(!list_link_active(&hdr->b_l1hdr.b_arc_node)); 5146 5147 hdr->b_l1hdr.b_arc_access = 0; 5148 5149 /* 5150 * If the buf is being overridden then it may already 5151 * have a hdr that is not empty. 5152 */ 5153 buf_discard_identity(hdr); 5154 arc_buf_thaw(buf); 5155 5156 return; 5157 } 5158 5159 kmutex_t *hash_lock = HDR_LOCK(hdr); 5160 mutex_enter(hash_lock); 5161 5162 /* 5163 * This assignment is only valid as long as the hash_lock is 5164 * held, we must be careful not to reference state or the 5165 * b_state field after dropping the lock. 5166 */ 5167 arc_state_t *state = hdr->b_l1hdr.b_state; 5168 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 5169 ASSERT3P(state, !=, arc_anon); 5170 5171 /* this buffer is not on any list */ 5172 ASSERT3S(refcount_count(&hdr->b_l1hdr.b_refcnt), >, 0); 5173 5174 if (HDR_HAS_L2HDR(hdr)) { 5175 mutex_enter(&hdr->b_l2hdr.b_dev->l2ad_mtx); 5176 5177 /* 5178 * We have to recheck this conditional again now that 5179 * we're holding the l2ad_mtx to prevent a race with 5180 * another thread which might be concurrently calling 5181 * l2arc_evict(). In that case, l2arc_evict() might have 5182 * destroyed the header's L2 portion as we were waiting 5183 * to acquire the l2ad_mtx. 5184 */ 5185 if (HDR_HAS_L2HDR(hdr)) 5186 arc_hdr_l2hdr_destroy(hdr); 5187 5188 mutex_exit(&hdr->b_l2hdr.b_dev->l2ad_mtx); 5189 } 5190 5191 /* 5192 * Do we have more than one buf? 5193 */ 5194 if (hdr->b_l1hdr.b_bufcnt > 1) { 5195 arc_buf_hdr_t *nhdr; 5196 uint64_t spa = hdr->b_spa; 5197 uint64_t psize = HDR_GET_PSIZE(hdr); 5198 uint64_t lsize = HDR_GET_LSIZE(hdr); 5199 enum zio_compress compress = HDR_GET_COMPRESS(hdr); 5200 arc_buf_contents_t type = arc_buf_type(hdr); 5201 VERIFY3U(hdr->b_type, ==, type); 5202 5203 ASSERT(hdr->b_l1hdr.b_buf != buf || buf->b_next != NULL); 5204 (void) remove_reference(hdr, hash_lock, tag); 5205 5206 if (arc_buf_is_shared(buf) && !ARC_BUF_COMPRESSED(buf)) { 5207 ASSERT3P(hdr->b_l1hdr.b_buf, !=, buf); 5208 ASSERT(ARC_BUF_LAST(buf)); 5209 } 5210 5211 /* 5212 * Pull the data off of this hdr and attach it to 5213 * a new anonymous hdr. Also find the last buffer 5214 * in the hdr's buffer list. 5215 */ 5216 arc_buf_t *lastbuf = arc_buf_remove(hdr, buf); 5217 ASSERT3P(lastbuf, !=, NULL); 5218 5219 /* 5220 * If the current arc_buf_t and the hdr are sharing their data 5221 * buffer, then we must stop sharing that block. 5222 */ 5223 if (arc_buf_is_shared(buf)) { 5224 VERIFY(!arc_buf_is_shared(lastbuf)); 5225 5226 /* 5227 * First, sever the block sharing relationship between 5228 * buf and the arc_buf_hdr_t. 5229 */ 5230 arc_unshare_buf(hdr, buf); 5231 5232 /* 5233 * Now we need to recreate the hdr's b_pdata. Since we 5234 * have lastbuf handy, we try to share with it, but if 5235 * we can't then we allocate a new b_pdata and copy the 5236 * data from buf into it. 5237 */ 5238 if (arc_can_share(hdr, lastbuf)) { 5239 arc_share_buf(hdr, lastbuf); 5240 } else { 5241 arc_hdr_alloc_pdata(hdr); 5242 bcopy(buf->b_data, hdr->b_l1hdr.b_pdata, psize); 5243 } 5244 VERIFY3P(lastbuf->b_data, !=, NULL); 5245 } else if (HDR_SHARED_DATA(hdr)) { 5246 /* 5247 * Uncompressed shared buffers are always at the end 5248 * of the list. Compressed buffers don't have the 5249 * same requirements. This makes it hard to 5250 * simply assert that the lastbuf is shared so 5251 * we rely on the hdr's compression flags to determine 5252 * if we have a compressed, shared buffer. 5253 */ 5254 ASSERT(arc_buf_is_shared(lastbuf) || 5255 HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF); 5256 ASSERT(!ARC_BUF_SHARED(buf)); 5257 } 5258 ASSERT3P(hdr->b_l1hdr.b_pdata, !=, NULL); 5259 ASSERT3P(state, !=, arc_l2c_only); 5260 5261 (void) refcount_remove_many(&state->arcs_size, 5262 arc_buf_size(buf), buf); 5263 5264 if (refcount_is_zero(&hdr->b_l1hdr.b_refcnt)) { 5265 ASSERT3P(state, !=, arc_l2c_only); 5266 (void) refcount_remove_many(&state->arcs_esize[type], 5267 arc_buf_size(buf), buf); 5268 } 5269 5270 hdr->b_l1hdr.b_bufcnt -= 1; 5271 arc_cksum_verify(buf); 5272 arc_buf_unwatch(buf); 5273 5274 mutex_exit(hash_lock); 5275 5276 /* 5277 * Allocate a new hdr. The new hdr will contain a b_pdata 5278 * buffer which will be freed in arc_write(). 5279 */ 5280 nhdr = arc_hdr_alloc(spa, psize, lsize, compress, type); 5281 ASSERT3P(nhdr->b_l1hdr.b_buf, ==, NULL); 5282 ASSERT0(nhdr->b_l1hdr.b_bufcnt); 5283 ASSERT0(refcount_count(&nhdr->b_l1hdr.b_refcnt)); 5284 VERIFY3U(nhdr->b_type, ==, type); 5285 ASSERT(!HDR_SHARED_DATA(nhdr)); 5286 5287 nhdr->b_l1hdr.b_buf = buf; 5288 nhdr->b_l1hdr.b_bufcnt = 1; 5289 (void) refcount_add(&nhdr->b_l1hdr.b_refcnt, tag); 5290 buf->b_hdr = nhdr; 5291 5292 mutex_exit(&buf->b_evict_lock); 5293 (void) refcount_add_many(&arc_anon->arcs_size, 5294 arc_buf_size(buf), buf); 5295 } else { 5296 mutex_exit(&buf->b_evict_lock); 5297 ASSERT(refcount_count(&hdr->b_l1hdr.b_refcnt) == 1); 5298 /* protected by hash lock, or hdr is on arc_anon */ 5299 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node)); 5300 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 5301 arc_change_state(arc_anon, hdr, hash_lock); 5302 hdr->b_l1hdr.b_arc_access = 0; 5303 mutex_exit(hash_lock); 5304 5305 buf_discard_identity(hdr); 5306 arc_buf_thaw(buf); 5307 } 5308 } 5309 5310 int 5311 arc_released(arc_buf_t *buf) 5312 { 5313 int released; 5314 5315 mutex_enter(&buf->b_evict_lock); 5316 released = (buf->b_data != NULL && 5317 buf->b_hdr->b_l1hdr.b_state == arc_anon); 5318 mutex_exit(&buf->b_evict_lock); 5319 return (released); 5320 } 5321 5322 #ifdef ZFS_DEBUG 5323 int 5324 arc_referenced(arc_buf_t *buf) 5325 { 5326 int referenced; 5327 5328 mutex_enter(&buf->b_evict_lock); 5329 referenced = (refcount_count(&buf->b_hdr->b_l1hdr.b_refcnt)); 5330 mutex_exit(&buf->b_evict_lock); 5331 return (referenced); 5332 } 5333 #endif 5334 5335 static void 5336 arc_write_ready(zio_t *zio) 5337 { 5338 arc_write_callback_t *callback = zio->io_private; 5339 arc_buf_t *buf = callback->awcb_buf; 5340 arc_buf_hdr_t *hdr = buf->b_hdr; 5341 uint64_t psize = BP_IS_HOLE(zio->io_bp) ? 0 : BP_GET_PSIZE(zio->io_bp); 5342 5343 ASSERT(HDR_HAS_L1HDR(hdr)); 5344 ASSERT(!refcount_is_zero(&buf->b_hdr->b_l1hdr.b_refcnt)); 5345 ASSERT(hdr->b_l1hdr.b_bufcnt > 0); 5346 5347 /* 5348 * If we're reexecuting this zio because the pool suspended, then 5349 * cleanup any state that was previously set the first time the 5350 * callback was invoked. 5351 */ 5352 if (zio->io_flags & ZIO_FLAG_REEXECUTED) { 5353 arc_cksum_free(hdr); 5354 arc_buf_unwatch(buf); 5355 if (hdr->b_l1hdr.b_pdata != NULL) { 5356 if (arc_buf_is_shared(buf)) { 5357 arc_unshare_buf(hdr, buf); 5358 } else { 5359 arc_hdr_free_pdata(hdr); 5360 } 5361 } 5362 } 5363 ASSERT3P(hdr->b_l1hdr.b_pdata, ==, NULL); 5364 ASSERT(!HDR_SHARED_DATA(hdr)); 5365 ASSERT(!arc_buf_is_shared(buf)); 5366 5367 callback->awcb_ready(zio, buf, callback->awcb_private); 5368 5369 if (HDR_IO_IN_PROGRESS(hdr)) 5370 ASSERT(zio->io_flags & ZIO_FLAG_REEXECUTED); 5371 5372 arc_cksum_compute(buf); 5373 arc_hdr_set_flags(hdr, ARC_FLAG_IO_IN_PROGRESS); 5374 5375 enum zio_compress compress; 5376 if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) { 5377 compress = ZIO_COMPRESS_OFF; 5378 } else { 5379 ASSERT3U(HDR_GET_LSIZE(hdr), ==, BP_GET_LSIZE(zio->io_bp)); 5380 compress = BP_GET_COMPRESS(zio->io_bp); 5381 } 5382 HDR_SET_PSIZE(hdr, psize); 5383 arc_hdr_set_compress(hdr, compress); 5384 5385 /* 5386 * If the hdr is compressed, then copy the compressed 5387 * zio contents into arc_buf_hdr_t. Otherwise, copy the original 5388 * data buf into the hdr. Ideally, we would like to always copy the 5389 * io_data into b_pdata but the user may have disabled compressed 5390 * arc thus the on-disk block may or may not match what we maintain 5391 * in the hdr's b_pdata field. 5392 */ 5393 if (HDR_GET_COMPRESS(hdr) != ZIO_COMPRESS_OFF && 5394 !ARC_BUF_COMPRESSED(buf)) { 5395 ASSERT3U(BP_GET_COMPRESS(zio->io_bp), !=, ZIO_COMPRESS_OFF); 5396 ASSERT3U(psize, >, 0); 5397 arc_hdr_alloc_pdata(hdr); 5398 bcopy(zio->io_data, hdr->b_l1hdr.b_pdata, psize); 5399 } else { 5400 ASSERT3P(buf->b_data, ==, zio->io_orig_data); 5401 ASSERT3U(zio->io_orig_size, ==, arc_buf_size(buf)); 5402 ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1); 5403 5404 /* 5405 * This hdr is not compressed so we're able to share 5406 * the arc_buf_t data buffer with the hdr. 5407 */ 5408 arc_share_buf(hdr, buf); 5409 ASSERT0(bcmp(zio->io_orig_data, hdr->b_l1hdr.b_pdata, 5410 HDR_GET_LSIZE(hdr))); 5411 } 5412 arc_hdr_verify(hdr, zio->io_bp); 5413 } 5414 5415 static void 5416 arc_write_children_ready(zio_t *zio) 5417 { 5418 arc_write_callback_t *callback = zio->io_private; 5419 arc_buf_t *buf = callback->awcb_buf; 5420 5421 callback->awcb_children_ready(zio, buf, callback->awcb_private); 5422 } 5423 5424 /* 5425 * The SPA calls this callback for each physical write that happens on behalf 5426 * of a logical write. See the comment in dbuf_write_physdone() for details. 5427 */ 5428 static void 5429 arc_write_physdone(zio_t *zio) 5430 { 5431 arc_write_callback_t *cb = zio->io_private; 5432 if (cb->awcb_physdone != NULL) 5433 cb->awcb_physdone(zio, cb->awcb_buf, cb->awcb_private); 5434 } 5435 5436 static void 5437 arc_write_done(zio_t *zio) 5438 { 5439 arc_write_callback_t *callback = zio->io_private; 5440 arc_buf_t *buf = callback->awcb_buf; 5441 arc_buf_hdr_t *hdr = buf->b_hdr; 5442 5443 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL); 5444 5445 if (zio->io_error == 0) { 5446 arc_hdr_verify(hdr, zio->io_bp); 5447 5448 if (BP_IS_HOLE(zio->io_bp) || BP_IS_EMBEDDED(zio->io_bp)) { 5449 buf_discard_identity(hdr); 5450 } else { 5451 hdr->b_dva = *BP_IDENTITY(zio->io_bp); 5452 hdr->b_birth = BP_PHYSICAL_BIRTH(zio->io_bp); 5453 } 5454 } else { 5455 ASSERT(HDR_EMPTY(hdr)); 5456 } 5457 5458 /* 5459 * If the block to be written was all-zero or compressed enough to be 5460 * embedded in the BP, no write was performed so there will be no 5461 * dva/birth/checksum. The buffer must therefore remain anonymous 5462 * (and uncached). 5463 */ 5464 if (!HDR_EMPTY(hdr)) { 5465 arc_buf_hdr_t *exists; 5466 kmutex_t *hash_lock; 5467 5468 ASSERT3U(zio->io_error, ==, 0); 5469 5470 arc_cksum_verify(buf); 5471 5472 exists = buf_hash_insert(hdr, &hash_lock); 5473 if (exists != NULL) { 5474 /* 5475 * This can only happen if we overwrite for 5476 * sync-to-convergence, because we remove 5477 * buffers from the hash table when we arc_free(). 5478 */ 5479 if (zio->io_flags & ZIO_FLAG_IO_REWRITE) { 5480 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) 5481 panic("bad overwrite, hdr=%p exists=%p", 5482 (void *)hdr, (void *)exists); 5483 ASSERT(refcount_is_zero( 5484 &exists->b_l1hdr.b_refcnt)); 5485 arc_change_state(arc_anon, exists, hash_lock); 5486 mutex_exit(hash_lock); 5487 arc_hdr_destroy(exists); 5488 exists = buf_hash_insert(hdr, &hash_lock); 5489 ASSERT3P(exists, ==, NULL); 5490 } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) { 5491 /* nopwrite */ 5492 ASSERT(zio->io_prop.zp_nopwrite); 5493 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp)) 5494 panic("bad nopwrite, hdr=%p exists=%p", 5495 (void *)hdr, (void *)exists); 5496 } else { 5497 /* Dedup */ 5498 ASSERT(hdr->b_l1hdr.b_bufcnt == 1); 5499 ASSERT(hdr->b_l1hdr.b_state == arc_anon); 5500 ASSERT(BP_GET_DEDUP(zio->io_bp)); 5501 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0); 5502 } 5503 } 5504 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS); 5505 /* if it's not anon, we are doing a scrub */ 5506 if (exists == NULL && hdr->b_l1hdr.b_state == arc_anon) 5507 arc_access(hdr, hash_lock); 5508 mutex_exit(hash_lock); 5509 } else { 5510 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS); 5511 } 5512 5513 ASSERT(!refcount_is_zero(&hdr->b_l1hdr.b_refcnt)); 5514 callback->awcb_done(zio, buf, callback->awcb_private); 5515 5516 kmem_free(callback, sizeof (arc_write_callback_t)); 5517 } 5518 5519 zio_t * 5520 arc_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, arc_buf_t *buf, 5521 boolean_t l2arc, const zio_prop_t *zp, arc_done_func_t *ready, 5522 arc_done_func_t *children_ready, arc_done_func_t *physdone, 5523 arc_done_func_t *done, void *private, zio_priority_t priority, 5524 int zio_flags, const zbookmark_phys_t *zb) 5525 { 5526 arc_buf_hdr_t *hdr = buf->b_hdr; 5527 arc_write_callback_t *callback; 5528 zio_t *zio; 5529 5530 ASSERT3P(ready, !=, NULL); 5531 ASSERT3P(done, !=, NULL); 5532 ASSERT(!HDR_IO_ERROR(hdr)); 5533 ASSERT(!HDR_IO_IN_PROGRESS(hdr)); 5534 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL); 5535 ASSERT3U(hdr->b_l1hdr.b_bufcnt, >, 0); 5536 if (l2arc) 5537 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE); 5538 if (ARC_BUF_COMPRESSED(buf)) { 5539 ASSERT3U(zp->zp_compress, !=, ZIO_COMPRESS_OFF); 5540 ASSERT3U(HDR_GET_LSIZE(hdr), !=, arc_buf_size(buf)); 5541 zio_flags |= ZIO_FLAG_RAW; 5542 } 5543 callback = kmem_zalloc(sizeof (arc_write_callback_t), KM_SLEEP); 5544 callback->awcb_ready = ready; 5545 callback->awcb_children_ready = children_ready; 5546 callback->awcb_physdone = physdone; 5547 callback->awcb_done = done; 5548 callback->awcb_private = private; 5549 callback->awcb_buf = buf; 5550 5551 /* 5552 * The hdr's b_pdata is now stale, free it now. A new data block 5553 * will be allocated when the zio pipeline calls arc_write_ready(). 5554 */ 5555 if (hdr->b_l1hdr.b_pdata != NULL) { 5556 /* 5557 * If the buf is currently sharing the data block with 5558 * the hdr then we need to break that relationship here. 5559 * The hdr will remain with a NULL data pointer and the 5560 * buf will take sole ownership of the block. 5561 */ 5562 if (arc_buf_is_shared(buf)) { 5563 arc_unshare_buf(hdr, buf); 5564 } else { 5565 arc_hdr_free_pdata(hdr); 5566 } 5567 VERIFY3P(buf->b_data, !=, NULL); 5568 arc_hdr_set_compress(hdr, ZIO_COMPRESS_OFF); 5569 } 5570 ASSERT(!arc_buf_is_shared(buf)); 5571 ASSERT3P(hdr->b_l1hdr.b_pdata, ==, NULL); 5572 5573 zio = zio_write(pio, spa, txg, bp, buf->b_data, 5574 HDR_GET_LSIZE(hdr), arc_buf_size(buf), zp, arc_write_ready, 5575 (children_ready != NULL) ? arc_write_children_ready : NULL, 5576 arc_write_physdone, arc_write_done, callback, 5577 priority, zio_flags, zb); 5578 5579 return (zio); 5580 } 5581 5582 static int 5583 arc_memory_throttle(uint64_t reserve, uint64_t txg) 5584 { 5585 #ifdef _KERNEL 5586 uint64_t available_memory = ptob(freemem); 5587 static uint64_t page_load = 0; 5588 static uint64_t last_txg = 0; 5589 5590 #if defined(__i386) 5591 available_memory = 5592 MIN(available_memory, vmem_size(heap_arena, VMEM_FREE)); 5593 #endif 5594 5595 if (freemem > physmem * arc_lotsfree_percent / 100) 5596 return (0); 5597 5598 if (txg > last_txg) { 5599 last_txg = txg; 5600 page_load = 0; 5601 } 5602 /* 5603 * If we are in pageout, we know that memory is already tight, 5604 * the arc is already going to be evicting, so we just want to 5605 * continue to let page writes occur as quickly as possible. 5606 */ 5607 if (curproc == proc_pageout) { 5608 if (page_load > MAX(ptob(minfree), available_memory) / 4) 5609 return (SET_ERROR(ERESTART)); 5610 /* Note: reserve is inflated, so we deflate */ 5611 page_load += reserve / 8; 5612 return (0); 5613 } else if (page_load > 0 && arc_reclaim_needed()) { 5614 /* memory is low, delay before restarting */ 5615 ARCSTAT_INCR(arcstat_memory_throttle_count, 1); 5616 return (SET_ERROR(EAGAIN)); 5617 } 5618 page_load = 0; 5619 #endif 5620 return (0); 5621 } 5622 5623 void 5624 arc_tempreserve_clear(uint64_t reserve) 5625 { 5626 atomic_add_64(&arc_tempreserve, -reserve); 5627 ASSERT((int64_t)arc_tempreserve >= 0); 5628 } 5629 5630 int 5631 arc_tempreserve_space(uint64_t reserve, uint64_t txg) 5632 { 5633 int error; 5634 uint64_t anon_size; 5635 5636 if (reserve > arc_c/4 && !arc_no_grow) 5637 arc_c = MIN(arc_c_max, reserve * 4); 5638 if (reserve > arc_c) 5639 return (SET_ERROR(ENOMEM)); 5640 5641 /* 5642 * Don't count loaned bufs as in flight dirty data to prevent long 5643 * network delays from blocking transactions that are ready to be 5644 * assigned to a txg. 5645 */ 5646 5647 /* assert that it has not wrapped around */ 5648 ASSERT3S(atomic_add_64_nv(&arc_loaned_bytes, 0), >=, 0); 5649 5650 anon_size = MAX((int64_t)(refcount_count(&arc_anon->arcs_size) - 5651 arc_loaned_bytes), 0); 5652 5653 /* 5654 * Writes will, almost always, require additional memory allocations 5655 * in order to compress/encrypt/etc the data. We therefore need to 5656 * make sure that there is sufficient available memory for this. 5657 */ 5658 error = arc_memory_throttle(reserve, txg); 5659 if (error != 0) 5660 return (error); 5661 5662 /* 5663 * Throttle writes when the amount of dirty data in the cache 5664 * gets too large. We try to keep the cache less than half full 5665 * of dirty blocks so that our sync times don't grow too large. 5666 * Note: if two requests come in concurrently, we might let them 5667 * both succeed, when one of them should fail. Not a huge deal. 5668 */ 5669 5670 if (reserve + arc_tempreserve + anon_size > arc_c / 2 && 5671 anon_size > arc_c / 4) { 5672 uint64_t meta_esize = 5673 refcount_count(&arc_anon->arcs_esize[ARC_BUFC_METADATA]); 5674 uint64_t data_esize = 5675 refcount_count(&arc_anon->arcs_esize[ARC_BUFC_DATA]); 5676 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK " 5677 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n", 5678 arc_tempreserve >> 10, meta_esize >> 10, 5679 data_esize >> 10, reserve >> 10, arc_c >> 10); 5680 return (SET_ERROR(ERESTART)); 5681 } 5682 atomic_add_64(&arc_tempreserve, reserve); 5683 return (0); 5684 } 5685 5686 static void 5687 arc_kstat_update_state(arc_state_t *state, kstat_named_t *size, 5688 kstat_named_t *evict_data, kstat_named_t *evict_metadata) 5689 { 5690 size->value.ui64 = refcount_count(&state->arcs_size); 5691 evict_data->value.ui64 = 5692 refcount_count(&state->arcs_esize[ARC_BUFC_DATA]); 5693 evict_metadata->value.ui64 = 5694 refcount_count(&state->arcs_esize[ARC_BUFC_METADATA]); 5695 } 5696 5697 static int 5698 arc_kstat_update(kstat_t *ksp, int rw) 5699 { 5700 arc_stats_t *as = ksp->ks_data; 5701 5702 if (rw == KSTAT_WRITE) { 5703 return (EACCES); 5704 } else { 5705 arc_kstat_update_state(arc_anon, 5706 &as->arcstat_anon_size, 5707 &as->arcstat_anon_evictable_data, 5708 &as->arcstat_anon_evictable_metadata); 5709 arc_kstat_update_state(arc_mru, 5710 &as->arcstat_mru_size, 5711 &as->arcstat_mru_evictable_data, 5712 &as->arcstat_mru_evictable_metadata); 5713 arc_kstat_update_state(arc_mru_ghost, 5714 &as->arcstat_mru_ghost_size, 5715 &as->arcstat_mru_ghost_evictable_data, 5716 &as->arcstat_mru_ghost_evictable_metadata); 5717 arc_kstat_update_state(arc_mfu, 5718 &as->arcstat_mfu_size, 5719 &as->arcstat_mfu_evictable_data, 5720 &as->arcstat_mfu_evictable_metadata); 5721 arc_kstat_update_state(arc_mfu_ghost, 5722 &as->arcstat_mfu_ghost_size, 5723 &as->arcstat_mfu_ghost_evictable_data, 5724 &as->arcstat_mfu_ghost_evictable_metadata); 5725 } 5726 5727 return (0); 5728 } 5729 5730 /* 5731 * This function *must* return indices evenly distributed between all 5732 * sublists of the multilist. This is needed due to how the ARC eviction 5733 * code is laid out; arc_evict_state() assumes ARC buffers are evenly 5734 * distributed between all sublists and uses this assumption when 5735 * deciding which sublist to evict from and how much to evict from it. 5736 */ 5737 unsigned int 5738 arc_state_multilist_index_func(multilist_t *ml, void *obj) 5739 { 5740 arc_buf_hdr_t *hdr = obj; 5741 5742 /* 5743 * We rely on b_dva to generate evenly distributed index 5744 * numbers using buf_hash below. So, as an added precaution, 5745 * let's make sure we never add empty buffers to the arc lists. 5746 */ 5747 ASSERT(!HDR_EMPTY(hdr)); 5748 5749 /* 5750 * The assumption here, is the hash value for a given 5751 * arc_buf_hdr_t will remain constant throughout it's lifetime 5752 * (i.e. it's b_spa, b_dva, and b_birth fields don't change). 5753 * Thus, we don't need to store the header's sublist index 5754 * on insertion, as this index can be recalculated on removal. 5755 * 5756 * Also, the low order bits of the hash value are thought to be 5757 * distributed evenly. Otherwise, in the case that the multilist 5758 * has a power of two number of sublists, each sublists' usage 5759 * would not be evenly distributed. 5760 */ 5761 return (buf_hash(hdr->b_spa, &hdr->b_dva, hdr->b_birth) % 5762 multilist_get_num_sublists(ml)); 5763 } 5764 5765 static void 5766 arc_state_init(void) 5767 { 5768 arc_anon = &ARC_anon; 5769 arc_mru = &ARC_mru; 5770 arc_mru_ghost = &ARC_mru_ghost; 5771 arc_mfu = &ARC_mfu; 5772 arc_mfu_ghost = &ARC_mfu_ghost; 5773 arc_l2c_only = &ARC_l2c_only; 5774 5775 multilist_create(&arc_mru->arcs_list[ARC_BUFC_METADATA], 5776 sizeof (arc_buf_hdr_t), 5777 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5778 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5779 multilist_create(&arc_mru->arcs_list[ARC_BUFC_DATA], 5780 sizeof (arc_buf_hdr_t), 5781 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5782 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5783 multilist_create(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA], 5784 sizeof (arc_buf_hdr_t), 5785 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5786 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5787 multilist_create(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA], 5788 sizeof (arc_buf_hdr_t), 5789 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5790 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5791 multilist_create(&arc_mfu->arcs_list[ARC_BUFC_METADATA], 5792 sizeof (arc_buf_hdr_t), 5793 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5794 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5795 multilist_create(&arc_mfu->arcs_list[ARC_BUFC_DATA], 5796 sizeof (arc_buf_hdr_t), 5797 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5798 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5799 multilist_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA], 5800 sizeof (arc_buf_hdr_t), 5801 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5802 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5803 multilist_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA], 5804 sizeof (arc_buf_hdr_t), 5805 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5806 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5807 multilist_create(&arc_l2c_only->arcs_list[ARC_BUFC_METADATA], 5808 sizeof (arc_buf_hdr_t), 5809 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5810 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5811 multilist_create(&arc_l2c_only->arcs_list[ARC_BUFC_DATA], 5812 sizeof (arc_buf_hdr_t), 5813 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), 5814 zfs_arc_num_sublists_per_state, arc_state_multilist_index_func); 5815 5816 refcount_create(&arc_anon->arcs_esize[ARC_BUFC_METADATA]); 5817 refcount_create(&arc_anon->arcs_esize[ARC_BUFC_DATA]); 5818 refcount_create(&arc_mru->arcs_esize[ARC_BUFC_METADATA]); 5819 refcount_create(&arc_mru->arcs_esize[ARC_BUFC_DATA]); 5820 refcount_create(&arc_mru_ghost->arcs_esize[ARC_BUFC_METADATA]); 5821 refcount_create(&arc_mru_ghost->arcs_esize[ARC_BUFC_DATA]); 5822 refcount_create(&arc_mfu->arcs_esize[ARC_BUFC_METADATA]); 5823 refcount_create(&arc_mfu->arcs_esize[ARC_BUFC_DATA]); 5824 refcount_create(&arc_mfu_ghost->arcs_esize[ARC_BUFC_METADATA]); 5825 refcount_create(&arc_mfu_ghost->arcs_esize[ARC_BUFC_DATA]); 5826 refcount_create(&arc_l2c_only->arcs_esize[ARC_BUFC_METADATA]); 5827 refcount_create(&arc_l2c_only->arcs_esize[ARC_BUFC_DATA]); 5828 5829 refcount_create(&arc_anon->arcs_size); 5830 refcount_create(&arc_mru->arcs_size); 5831 refcount_create(&arc_mru_ghost->arcs_size); 5832 refcount_create(&arc_mfu->arcs_size); 5833 refcount_create(&arc_mfu_ghost->arcs_size); 5834 refcount_create(&arc_l2c_only->arcs_size); 5835 } 5836 5837 static void 5838 arc_state_fini(void) 5839 { 5840 refcount_destroy(&arc_anon->arcs_esize[ARC_BUFC_METADATA]); 5841 refcount_destroy(&arc_anon->arcs_esize[ARC_BUFC_DATA]); 5842 refcount_destroy(&arc_mru->arcs_esize[ARC_BUFC_METADATA]); 5843 refcount_destroy(&arc_mru->arcs_esize[ARC_BUFC_DATA]); 5844 refcount_destroy(&arc_mru_ghost->arcs_esize[ARC_BUFC_METADATA]); 5845 refcount_destroy(&arc_mru_ghost->arcs_esize[ARC_BUFC_DATA]); 5846 refcount_destroy(&arc_mfu->arcs_esize[ARC_BUFC_METADATA]); 5847 refcount_destroy(&arc_mfu->arcs_esize[ARC_BUFC_DATA]); 5848 refcount_destroy(&arc_mfu_ghost->arcs_esize[ARC_BUFC_METADATA]); 5849 refcount_destroy(&arc_mfu_ghost->arcs_esize[ARC_BUFC_DATA]); 5850 refcount_destroy(&arc_l2c_only->arcs_esize[ARC_BUFC_METADATA]); 5851 refcount_destroy(&arc_l2c_only->arcs_esize[ARC_BUFC_DATA]); 5852 5853 refcount_destroy(&arc_anon->arcs_size); 5854 refcount_destroy(&arc_mru->arcs_size); 5855 refcount_destroy(&arc_mru_ghost->arcs_size); 5856 refcount_destroy(&arc_mfu->arcs_size); 5857 refcount_destroy(&arc_mfu_ghost->arcs_size); 5858 refcount_destroy(&arc_l2c_only->arcs_size); 5859 5860 multilist_destroy(&arc_mru->arcs_list[ARC_BUFC_METADATA]); 5861 multilist_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA]); 5862 multilist_destroy(&arc_mfu->arcs_list[ARC_BUFC_METADATA]); 5863 multilist_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA]); 5864 multilist_destroy(&arc_mru->arcs_list[ARC_BUFC_DATA]); 5865 multilist_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA]); 5866 multilist_destroy(&arc_mfu->arcs_list[ARC_BUFC_DATA]); 5867 multilist_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA]); 5868 } 5869 5870 uint64_t 5871 arc_max_bytes(void) 5872 { 5873 return (arc_c_max); 5874 } 5875 5876 void 5877 arc_init(void) 5878 { 5879 /* 5880 * allmem is "all memory that we could possibly use". 5881 */ 5882 #ifdef _KERNEL 5883 uint64_t allmem = ptob(physmem - swapfs_minfree); 5884 #else 5885 uint64_t allmem = (physmem * PAGESIZE) / 2; 5886 #endif 5887 5888 mutex_init(&arc_reclaim_lock, NULL, MUTEX_DEFAULT, NULL); 5889 cv_init(&arc_reclaim_thread_cv, NULL, CV_DEFAULT, NULL); 5890 cv_init(&arc_reclaim_waiters_cv, NULL, CV_DEFAULT, NULL); 5891 5892 /* Convert seconds to clock ticks */ 5893 arc_min_prefetch_lifespan = 1 * hz; 5894 5895 /* set min cache to 1/32 of all memory, or 64MB, whichever is more */ 5896 arc_c_min = MAX(allmem / 32, 64 << 20); 5897 /* set max to 3/4 of all memory, or all but 1GB, whichever is more */ 5898 if (allmem >= 1 << 30) 5899 arc_c_max = allmem - (1 << 30); 5900 else 5901 arc_c_max = arc_c_min; 5902 arc_c_max = MAX(allmem * 3 / 4, arc_c_max); 5903 5904 /* 5905 * In userland, there's only the memory pressure that we artificially 5906 * create (see arc_available_memory()). Don't let arc_c get too 5907 * small, because it can cause transactions to be larger than 5908 * arc_c, causing arc_tempreserve_space() to fail. 5909 */ 5910 #ifndef _KERNEL 5911 arc_c_min = arc_c_max / 2; 5912 #endif 5913 5914 /* 5915 * Allow the tunables to override our calculations if they are 5916 * reasonable (ie. over 64MB) 5917 */ 5918 if (zfs_arc_max > 64 << 20 && zfs_arc_max < allmem) { 5919 arc_c_max = zfs_arc_max; 5920 arc_c_min = MIN(arc_c_min, arc_c_max); 5921 } 5922 if (zfs_arc_min > 64 << 20 && zfs_arc_min <= arc_c_max) 5923 arc_c_min = zfs_arc_min; 5924 5925 arc_c = arc_c_max; 5926 arc_p = (arc_c >> 1); 5927 arc_size = 0; 5928 5929 /* limit meta-data to 1/4 of the arc capacity */ 5930 arc_meta_limit = arc_c_max / 4; 5931 5932 #ifdef _KERNEL 5933 /* 5934 * Metadata is stored in the kernel's heap. Don't let us 5935 * use more than half the heap for the ARC. 5936 */ 5937 arc_meta_limit = MIN(arc_meta_limit, 5938 vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 2); 5939 #endif 5940 5941 /* Allow the tunable to override if it is reasonable */ 5942 if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max) 5943 arc_meta_limit = zfs_arc_meta_limit; 5944 5945 if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0) 5946 arc_c_min = arc_meta_limit / 2; 5947 5948 if (zfs_arc_meta_min > 0) { 5949 arc_meta_min = zfs_arc_meta_min; 5950 } else { 5951 arc_meta_min = arc_c_min / 2; 5952 } 5953 5954 if (zfs_arc_grow_retry > 0) 5955 arc_grow_retry = zfs_arc_grow_retry; 5956 5957 if (zfs_arc_shrink_shift > 0) 5958 arc_shrink_shift = zfs_arc_shrink_shift; 5959 5960 /* 5961 * Ensure that arc_no_grow_shift is less than arc_shrink_shift. 5962 */ 5963 if (arc_no_grow_shift >= arc_shrink_shift) 5964 arc_no_grow_shift = arc_shrink_shift - 1; 5965 5966 if (zfs_arc_p_min_shift > 0) 5967 arc_p_min_shift = zfs_arc_p_min_shift; 5968 5969 if (zfs_arc_num_sublists_per_state < 1) 5970 zfs_arc_num_sublists_per_state = MAX(boot_ncpus, 1); 5971 5972 /* if kmem_flags are set, lets try to use less memory */ 5973 if (kmem_debugging()) 5974 arc_c = arc_c / 2; 5975 if (arc_c < arc_c_min) 5976 arc_c = arc_c_min; 5977 5978 arc_state_init(); 5979 buf_init(); 5980 5981 arc_reclaim_thread_exit = B_FALSE; 5982 5983 arc_ksp = kstat_create("zfs", 0, "arcstats", "misc", KSTAT_TYPE_NAMED, 5984 sizeof (arc_stats) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL); 5985 5986 if (arc_ksp != NULL) { 5987 arc_ksp->ks_data = &arc_stats; 5988 arc_ksp->ks_update = arc_kstat_update; 5989 kstat_install(arc_ksp); 5990 } 5991 5992 (void) thread_create(NULL, 0, arc_reclaim_thread, NULL, 0, &p0, 5993 TS_RUN, minclsyspri); 5994 5995 arc_dead = B_FALSE; 5996 arc_warm = B_FALSE; 5997 5998 /* 5999 * Calculate maximum amount of dirty data per pool. 6000 * 6001 * If it has been set by /etc/system, take that. 6002 * Otherwise, use a percentage of physical memory defined by 6003 * zfs_dirty_data_max_percent (default 10%) with a cap at 6004 * zfs_dirty_data_max_max (default 4GB). 6005 */ 6006 if (zfs_dirty_data_max == 0) { 6007 zfs_dirty_data_max = physmem * PAGESIZE * 6008 zfs_dirty_data_max_percent / 100; 6009 zfs_dirty_data_max = MIN(zfs_dirty_data_max, 6010 zfs_dirty_data_max_max); 6011 } 6012 } 6013 6014 void 6015 arc_fini(void) 6016 { 6017 mutex_enter(&arc_reclaim_lock); 6018 arc_reclaim_thread_exit = B_TRUE; 6019 /* 6020 * The reclaim thread will set arc_reclaim_thread_exit back to 6021 * B_FALSE when it is finished exiting; we're waiting for that. 6022 */ 6023 while (arc_reclaim_thread_exit) { 6024 cv_signal(&arc_reclaim_thread_cv); 6025 cv_wait(&arc_reclaim_thread_cv, &arc_reclaim_lock); 6026 } 6027 mutex_exit(&arc_reclaim_lock); 6028 6029 /* Use B_TRUE to ensure *all* buffers are evicted */ 6030 arc_flush(NULL, B_TRUE); 6031 6032 arc_dead = B_TRUE; 6033 6034 if (arc_ksp != NULL) { 6035 kstat_delete(arc_ksp); 6036 arc_ksp = NULL; 6037 } 6038 6039 mutex_destroy(&arc_reclaim_lock); 6040 cv_destroy(&arc_reclaim_thread_cv); 6041 cv_destroy(&arc_reclaim_waiters_cv); 6042 6043 arc_state_fini(); 6044 buf_fini(); 6045 6046 ASSERT0(arc_loaned_bytes); 6047 } 6048 6049 /* 6050 * Level 2 ARC 6051 * 6052 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk. 6053 * It uses dedicated storage devices to hold cached data, which are populated 6054 * using large infrequent writes. The main role of this cache is to boost 6055 * the performance of random read workloads. The intended L2ARC devices 6056 * include short-stroked disks, solid state disks, and other media with 6057 * substantially faster read latency than disk. 6058 * 6059 * +-----------------------+ 6060 * | ARC | 6061 * +-----------------------+ 6062 * | ^ ^ 6063 * | | | 6064 * l2arc_feed_thread() arc_read() 6065 * | | | 6066 * | l2arc read | 6067 * V | | 6068 * +---------------+ | 6069 * | L2ARC | | 6070 * +---------------+ | 6071 * | ^ | 6072 * l2arc_write() | | 6073 * | | | 6074 * V | | 6075 * +-------+ +-------+ 6076 * | vdev | | vdev | 6077 * | cache | | cache | 6078 * +-------+ +-------+ 6079 * +=========+ .-----. 6080 * : L2ARC : |-_____-| 6081 * : devices : | Disks | 6082 * +=========+ `-_____-' 6083 * 6084 * Read requests are satisfied from the following sources, in order: 6085 * 6086 * 1) ARC 6087 * 2) vdev cache of L2ARC devices 6088 * 3) L2ARC devices 6089 * 4) vdev cache of disks 6090 * 5) disks 6091 * 6092 * Some L2ARC device types exhibit extremely slow write performance. 6093 * To accommodate for this there are some significant differences between 6094 * the L2ARC and traditional cache design: 6095 * 6096 * 1. There is no eviction path from the ARC to the L2ARC. Evictions from 6097 * the ARC behave as usual, freeing buffers and placing headers on ghost 6098 * lists. The ARC does not send buffers to the L2ARC during eviction as 6099 * this would add inflated write latencies for all ARC memory pressure. 6100 * 6101 * 2. The L2ARC attempts to cache data from the ARC before it is evicted. 6102 * It does this by periodically scanning buffers from the eviction-end of 6103 * the MFU and MRU ARC lists, copying them to the L2ARC devices if they are 6104 * not already there. It scans until a headroom of buffers is satisfied, 6105 * which itself is a buffer for ARC eviction. If a compressible buffer is 6106 * found during scanning and selected for writing to an L2ARC device, we 6107 * temporarily boost scanning headroom during the next scan cycle to make 6108 * sure we adapt to compression effects (which might significantly reduce 6109 * the data volume we write to L2ARC). The thread that does this is 6110 * l2arc_feed_thread(), illustrated below; example sizes are included to 6111 * provide a better sense of ratio than this diagram: 6112 * 6113 * head --> tail 6114 * +---------------------+----------+ 6115 * ARC_mfu |:::::#:::::::::::::::|o#o###o###|-->. # already on L2ARC 6116 * +---------------------+----------+ | o L2ARC eligible 6117 * ARC_mru |:#:::::::::::::::::::|#o#ooo####|-->| : ARC buffer 6118 * +---------------------+----------+ | 6119 * 15.9 Gbytes ^ 32 Mbytes | 6120 * headroom | 6121 * l2arc_feed_thread() 6122 * | 6123 * l2arc write hand <--[oooo]--' 6124 * | 8 Mbyte 6125 * | write max 6126 * V 6127 * +==============================+ 6128 * L2ARC dev |####|#|###|###| |####| ... | 6129 * +==============================+ 6130 * 32 Gbytes 6131 * 6132 * 3. If an ARC buffer is copied to the L2ARC but then hit instead of 6133 * evicted, then the L2ARC has cached a buffer much sooner than it probably 6134 * needed to, potentially wasting L2ARC device bandwidth and storage. It is 6135 * safe to say that this is an uncommon case, since buffers at the end of 6136 * the ARC lists have moved there due to inactivity. 6137 * 6138 * 4. If the ARC evicts faster than the L2ARC can maintain a headroom, 6139 * then the L2ARC simply misses copying some buffers. This serves as a 6140 * pressure valve to prevent heavy read workloads from both stalling the ARC 6141 * with waits and clogging the L2ARC with writes. This also helps prevent 6142 * the potential for the L2ARC to churn if it attempts to cache content too 6143 * quickly, such as during backups of the entire pool. 6144 * 6145 * 5. After system boot and before the ARC has filled main memory, there are 6146 * no evictions from the ARC and so the tails of the ARC_mfu and ARC_mru 6147 * lists can remain mostly static. Instead of searching from tail of these 6148 * lists as pictured, the l2arc_feed_thread() will search from the list heads 6149 * for eligible buffers, greatly increasing its chance of finding them. 6150 * 6151 * The L2ARC device write speed is also boosted during this time so that 6152 * the L2ARC warms up faster. Since there have been no ARC evictions yet, 6153 * there are no L2ARC reads, and no fear of degrading read performance 6154 * through increased writes. 6155 * 6156 * 6. Writes to the L2ARC devices are grouped and sent in-sequence, so that 6157 * the vdev queue can aggregate them into larger and fewer writes. Each 6158 * device is written to in a rotor fashion, sweeping writes through 6159 * available space then repeating. 6160 * 6161 * 7. The L2ARC does not store dirty content. It never needs to flush 6162 * write buffers back to disk based storage. 6163 * 6164 * 8. If an ARC buffer is written (and dirtied) which also exists in the 6165 * L2ARC, the now stale L2ARC buffer is immediately dropped. 6166 * 6167 * The performance of the L2ARC can be tweaked by a number of tunables, which 6168 * may be necessary for different workloads: 6169 * 6170 * l2arc_write_max max write bytes per interval 6171 * l2arc_write_boost extra write bytes during device warmup 6172 * l2arc_noprefetch skip caching prefetched buffers 6173 * l2arc_headroom number of max device writes to precache 6174 * l2arc_headroom_boost when we find compressed buffers during ARC 6175 * scanning, we multiply headroom by this 6176 * percentage factor for the next scan cycle, 6177 * since more compressed buffers are likely to 6178 * be present 6179 * l2arc_feed_secs seconds between L2ARC writing 6180 * 6181 * Tunables may be removed or added as future performance improvements are 6182 * integrated, and also may become zpool properties. 6183 * 6184 * There are three key functions that control how the L2ARC warms up: 6185 * 6186 * l2arc_write_eligible() check if a buffer is eligible to cache 6187 * l2arc_write_size() calculate how much to write 6188 * l2arc_write_interval() calculate sleep delay between writes 6189 * 6190 * These three functions determine what to write, how much, and how quickly 6191 * to send writes. 6192 */ 6193 6194 static boolean_t 6195 l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *hdr) 6196 { 6197 /* 6198 * A buffer is *not* eligible for the L2ARC if it: 6199 * 1. belongs to a different spa. 6200 * 2. is already cached on the L2ARC. 6201 * 3. has an I/O in progress (it may be an incomplete read). 6202 * 4. is flagged not eligible (zfs property). 6203 */ 6204 if (hdr->b_spa != spa_guid || HDR_HAS_L2HDR(hdr) || 6205 HDR_IO_IN_PROGRESS(hdr) || !HDR_L2CACHE(hdr)) 6206 return (B_FALSE); 6207 6208 return (B_TRUE); 6209 } 6210 6211 static uint64_t 6212 l2arc_write_size(void) 6213 { 6214 uint64_t size; 6215 6216 /* 6217 * Make sure our globals have meaningful values in case the user 6218 * altered them. 6219 */ 6220 size = l2arc_write_max; 6221 if (size == 0) { 6222 cmn_err(CE_NOTE, "Bad value for l2arc_write_max, value must " 6223 "be greater than zero, resetting it to the default (%d)", 6224 L2ARC_WRITE_SIZE); 6225 size = l2arc_write_max = L2ARC_WRITE_SIZE; 6226 } 6227 6228 if (arc_warm == B_FALSE) 6229 size += l2arc_write_boost; 6230 6231 return (size); 6232 6233 } 6234 6235 static clock_t 6236 l2arc_write_interval(clock_t began, uint64_t wanted, uint64_t wrote) 6237 { 6238 clock_t interval, next, now; 6239 6240 /* 6241 * If the ARC lists are busy, increase our write rate; if the 6242 * lists are stale, idle back. This is achieved by checking 6243 * how much we previously wrote - if it was more than half of 6244 * what we wanted, schedule the next write much sooner. 6245 */ 6246 if (l2arc_feed_again && wrote > (wanted / 2)) 6247 interval = (hz * l2arc_feed_min_ms) / 1000; 6248 else 6249 interval = hz * l2arc_feed_secs; 6250 6251 now = ddi_get_lbolt(); 6252 next = MAX(now, MIN(now + interval, began + interval)); 6253 6254 return (next); 6255 } 6256 6257 /* 6258 * Cycle through L2ARC devices. This is how L2ARC load balances. 6259 * If a device is returned, this also returns holding the spa config lock. 6260 */ 6261 static l2arc_dev_t * 6262 l2arc_dev_get_next(void) 6263 { 6264 l2arc_dev_t *first, *next = NULL; 6265 6266 /* 6267 * Lock out the removal of spas (spa_namespace_lock), then removal 6268 * of cache devices (l2arc_dev_mtx). Once a device has been selected, 6269 * both locks will be dropped and a spa config lock held instead. 6270 */ 6271 mutex_enter(&spa_namespace_lock); 6272 mutex_enter(&l2arc_dev_mtx); 6273 6274 /* if there are no vdevs, there is nothing to do */ 6275 if (l2arc_ndev == 0) 6276 goto out; 6277 6278 first = NULL; 6279 next = l2arc_dev_last; 6280 do { 6281 /* loop around the list looking for a non-faulted vdev */ 6282 if (next == NULL) { 6283 next = list_head(l2arc_dev_list); 6284 } else { 6285 next = list_next(l2arc_dev_list, next); 6286 if (next == NULL) 6287 next = list_head(l2arc_dev_list); 6288 } 6289 6290 /* if we have come back to the start, bail out */ 6291 if (first == NULL) 6292 first = next; 6293 else if (next == first) 6294 break; 6295 6296 } while (vdev_is_dead(next->l2ad_vdev)); 6297 6298 /* if we were unable to find any usable vdevs, return NULL */ 6299 if (vdev_is_dead(next->l2ad_vdev)) 6300 next = NULL; 6301 6302 l2arc_dev_last = next; 6303 6304 out: 6305 mutex_exit(&l2arc_dev_mtx); 6306 6307 /* 6308 * Grab the config lock to prevent the 'next' device from being 6309 * removed while we are writing to it. 6310 */ 6311 if (next != NULL) 6312 spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER); 6313 mutex_exit(&spa_namespace_lock); 6314 6315 return (next); 6316 } 6317 6318 /* 6319 * Free buffers that were tagged for destruction. 6320 */ 6321 static void 6322 l2arc_do_free_on_write() 6323 { 6324 list_t *buflist; 6325 l2arc_data_free_t *df, *df_prev; 6326 6327 mutex_enter(&l2arc_free_on_write_mtx); 6328 buflist = l2arc_free_on_write; 6329 6330 for (df = list_tail(buflist); df; df = df_prev) { 6331 df_prev = list_prev(buflist, df); 6332 ASSERT3P(df->l2df_data, !=, NULL); 6333 if (df->l2df_type == ARC_BUFC_METADATA) { 6334 zio_buf_free(df->l2df_data, df->l2df_size); 6335 } else { 6336 ASSERT(df->l2df_type == ARC_BUFC_DATA); 6337 zio_data_buf_free(df->l2df_data, df->l2df_size); 6338 } 6339 list_remove(buflist, df); 6340 kmem_free(df, sizeof (l2arc_data_free_t)); 6341 } 6342 6343 mutex_exit(&l2arc_free_on_write_mtx); 6344 } 6345 6346 /* 6347 * A write to a cache device has completed. Update all headers to allow 6348 * reads from these buffers to begin. 6349 */ 6350 static void 6351 l2arc_write_done(zio_t *zio) 6352 { 6353 l2arc_write_callback_t *cb; 6354 l2arc_dev_t *dev; 6355 list_t *buflist; 6356 arc_buf_hdr_t *head, *hdr, *hdr_prev; 6357 kmutex_t *hash_lock; 6358 int64_t bytes_dropped = 0; 6359 6360 cb = zio->io_private; 6361 ASSERT3P(cb, !=, NULL); 6362 dev = cb->l2wcb_dev; 6363 ASSERT3P(dev, !=, NULL); 6364 head = cb->l2wcb_head; 6365 ASSERT3P(head, !=, NULL); 6366 buflist = &dev->l2ad_buflist; 6367 ASSERT3P(buflist, !=, NULL); 6368 DTRACE_PROBE2(l2arc__iodone, zio_t *, zio, 6369 l2arc_write_callback_t *, cb); 6370 6371 if (zio->io_error != 0) 6372 ARCSTAT_BUMP(arcstat_l2_writes_error); 6373 6374 /* 6375 * All writes completed, or an error was hit. 6376 */ 6377 top: 6378 mutex_enter(&dev->l2ad_mtx); 6379 for (hdr = list_prev(buflist, head); hdr; hdr = hdr_prev) { 6380 hdr_prev = list_prev(buflist, hdr); 6381 6382 hash_lock = HDR_LOCK(hdr); 6383 6384 /* 6385 * We cannot use mutex_enter or else we can deadlock 6386 * with l2arc_write_buffers (due to swapping the order 6387 * the hash lock and l2ad_mtx are taken). 6388 */ 6389 if (!mutex_tryenter(hash_lock)) { 6390 /* 6391 * Missed the hash lock. We must retry so we 6392 * don't leave the ARC_FLAG_L2_WRITING bit set. 6393 */ 6394 ARCSTAT_BUMP(arcstat_l2_writes_lock_retry); 6395 6396 /* 6397 * We don't want to rescan the headers we've 6398 * already marked as having been written out, so 6399 * we reinsert the head node so we can pick up 6400 * where we left off. 6401 */ 6402 list_remove(buflist, head); 6403 list_insert_after(buflist, hdr, head); 6404 6405 mutex_exit(&dev->l2ad_mtx); 6406 6407 /* 6408 * We wait for the hash lock to become available 6409 * to try and prevent busy waiting, and increase 6410 * the chance we'll be able to acquire the lock 6411 * the next time around. 6412 */ 6413 mutex_enter(hash_lock); 6414 mutex_exit(hash_lock); 6415 goto top; 6416 } 6417 6418 /* 6419 * We could not have been moved into the arc_l2c_only 6420 * state while in-flight due to our ARC_FLAG_L2_WRITING 6421 * bit being set. Let's just ensure that's being enforced. 6422 */ 6423 ASSERT(HDR_HAS_L1HDR(hdr)); 6424 6425 if (zio->io_error != 0) { 6426 /* 6427 * Error - drop L2ARC entry. 6428 */ 6429 list_remove(buflist, hdr); 6430 arc_hdr_clear_flags(hdr, ARC_FLAG_HAS_L2HDR); 6431 6432 ARCSTAT_INCR(arcstat_l2_asize, -arc_hdr_size(hdr)); 6433 ARCSTAT_INCR(arcstat_l2_size, -HDR_GET_LSIZE(hdr)); 6434 6435 bytes_dropped += arc_hdr_size(hdr); 6436 (void) refcount_remove_many(&dev->l2ad_alloc, 6437 arc_hdr_size(hdr), hdr); 6438 } 6439 6440 /* 6441 * Allow ARC to begin reads and ghost list evictions to 6442 * this L2ARC entry. 6443 */ 6444 arc_hdr_clear_flags(hdr, ARC_FLAG_L2_WRITING); 6445 6446 mutex_exit(hash_lock); 6447 } 6448 6449 atomic_inc_64(&l2arc_writes_done); 6450 list_remove(buflist, head); 6451 ASSERT(!HDR_HAS_L1HDR(head)); 6452 kmem_cache_free(hdr_l2only_cache, head); 6453 mutex_exit(&dev->l2ad_mtx); 6454 6455 vdev_space_update(dev->l2ad_vdev, -bytes_dropped, 0, 0); 6456 6457 l2arc_do_free_on_write(); 6458 6459 kmem_free(cb, sizeof (l2arc_write_callback_t)); 6460 } 6461 6462 /* 6463 * A read to a cache device completed. Validate buffer contents before 6464 * handing over to the regular ARC routines. 6465 */ 6466 static void 6467 l2arc_read_done(zio_t *zio) 6468 { 6469 l2arc_read_callback_t *cb; 6470 arc_buf_hdr_t *hdr; 6471 kmutex_t *hash_lock; 6472 boolean_t valid_cksum; 6473 6474 ASSERT3P(zio->io_vd, !=, NULL); 6475 ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE); 6476 6477 spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd); 6478 6479 cb = zio->io_private; 6480 ASSERT3P(cb, !=, NULL); 6481 hdr = cb->l2rcb_hdr; 6482 ASSERT3P(hdr, !=, NULL); 6483 6484 hash_lock = HDR_LOCK(hdr); 6485 mutex_enter(hash_lock); 6486 ASSERT3P(hash_lock, ==, HDR_LOCK(hdr)); 6487 6488 ASSERT3P(zio->io_data, !=, NULL); 6489 6490 /* 6491 * Check this survived the L2ARC journey. 6492 */ 6493 ASSERT3P(zio->io_data, ==, hdr->b_l1hdr.b_pdata); 6494 zio->io_bp_copy = cb->l2rcb_bp; /* XXX fix in L2ARC 2.0 */ 6495 zio->io_bp = &zio->io_bp_copy; /* XXX fix in L2ARC 2.0 */ 6496 6497 valid_cksum = arc_cksum_is_equal(hdr, zio); 6498 if (valid_cksum && zio->io_error == 0 && !HDR_L2_EVICTED(hdr)) { 6499 mutex_exit(hash_lock); 6500 zio->io_private = hdr; 6501 arc_read_done(zio); 6502 } else { 6503 mutex_exit(hash_lock); 6504 /* 6505 * Buffer didn't survive caching. Increment stats and 6506 * reissue to the original storage device. 6507 */ 6508 if (zio->io_error != 0) { 6509 ARCSTAT_BUMP(arcstat_l2_io_error); 6510 } else { 6511 zio->io_error = SET_ERROR(EIO); 6512 } 6513 if (!valid_cksum) 6514 ARCSTAT_BUMP(arcstat_l2_cksum_bad); 6515 6516 /* 6517 * If there's no waiter, issue an async i/o to the primary 6518 * storage now. If there *is* a waiter, the caller must 6519 * issue the i/o in a context where it's OK to block. 6520 */ 6521 if (zio->io_waiter == NULL) { 6522 zio_t *pio = zio_unique_parent(zio); 6523 6524 ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL); 6525 6526 zio_nowait(zio_read(pio, zio->io_spa, zio->io_bp, 6527 hdr->b_l1hdr.b_pdata, zio->io_size, arc_read_done, 6528 hdr, zio->io_priority, cb->l2rcb_flags, 6529 &cb->l2rcb_zb)); 6530 } 6531 } 6532 6533 kmem_free(cb, sizeof (l2arc_read_callback_t)); 6534 } 6535 6536 /* 6537 * This is the list priority from which the L2ARC will search for pages to 6538 * cache. This is used within loops (0..3) to cycle through lists in the 6539 * desired order. This order can have a significant effect on cache 6540 * performance. 6541 * 6542 * Currently the metadata lists are hit first, MFU then MRU, followed by 6543 * the data lists. This function returns a locked list, and also returns 6544 * the lock pointer. 6545 */ 6546 static multilist_sublist_t * 6547 l2arc_sublist_lock(int list_num) 6548 { 6549 multilist_t *ml = NULL; 6550 unsigned int idx; 6551 6552 ASSERT(list_num >= 0 && list_num <= 3); 6553 6554 switch (list_num) { 6555 case 0: 6556 ml = &arc_mfu->arcs_list[ARC_BUFC_METADATA]; 6557 break; 6558 case 1: 6559 ml = &arc_mru->arcs_list[ARC_BUFC_METADATA]; 6560 break; 6561 case 2: 6562 ml = &arc_mfu->arcs_list[ARC_BUFC_DATA]; 6563 break; 6564 case 3: 6565 ml = &arc_mru->arcs_list[ARC_BUFC_DATA]; 6566 break; 6567 } 6568 6569 /* 6570 * Return a randomly-selected sublist. This is acceptable 6571 * because the caller feeds only a little bit of data for each 6572 * call (8MB). Subsequent calls will result in different 6573 * sublists being selected. 6574 */ 6575 idx = multilist_get_random_index(ml); 6576 return (multilist_sublist_lock(ml, idx)); 6577 } 6578 6579 /* 6580 * Evict buffers from the device write hand to the distance specified in 6581 * bytes. This distance may span populated buffers, it may span nothing. 6582 * This is clearing a region on the L2ARC device ready for writing. 6583 * If the 'all' boolean is set, every buffer is evicted. 6584 */ 6585 static void 6586 l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all) 6587 { 6588 list_t *buflist; 6589 arc_buf_hdr_t *hdr, *hdr_prev; 6590 kmutex_t *hash_lock; 6591 uint64_t taddr; 6592 6593 buflist = &dev->l2ad_buflist; 6594 6595 if (!all && dev->l2ad_first) { 6596 /* 6597 * This is the first sweep through the device. There is 6598 * nothing to evict. 6599 */ 6600 return; 6601 } 6602 6603 if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) { 6604 /* 6605 * When nearing the end of the device, evict to the end 6606 * before the device write hand jumps to the start. 6607 */ 6608 taddr = dev->l2ad_end; 6609 } else { 6610 taddr = dev->l2ad_hand + distance; 6611 } 6612 DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist, 6613 uint64_t, taddr, boolean_t, all); 6614 6615 top: 6616 mutex_enter(&dev->l2ad_mtx); 6617 for (hdr = list_tail(buflist); hdr; hdr = hdr_prev) { 6618 hdr_prev = list_prev(buflist, hdr); 6619 6620 hash_lock = HDR_LOCK(hdr); 6621 6622 /* 6623 * We cannot use mutex_enter or else we can deadlock 6624 * with l2arc_write_buffers (due to swapping the order 6625 * the hash lock and l2ad_mtx are taken). 6626 */ 6627 if (!mutex_tryenter(hash_lock)) { 6628 /* 6629 * Missed the hash lock. Retry. 6630 */ 6631 ARCSTAT_BUMP(arcstat_l2_evict_lock_retry); 6632 mutex_exit(&dev->l2ad_mtx); 6633 mutex_enter(hash_lock); 6634 mutex_exit(hash_lock); 6635 goto top; 6636 } 6637 6638 if (HDR_L2_WRITE_HEAD(hdr)) { 6639 /* 6640 * We hit a write head node. Leave it for 6641 * l2arc_write_done(). 6642 */ 6643 list_remove(buflist, hdr); 6644 mutex_exit(hash_lock); 6645 continue; 6646 } 6647 6648 if (!all && HDR_HAS_L2HDR(hdr) && 6649 (hdr->b_l2hdr.b_daddr > taddr || 6650 hdr->b_l2hdr.b_daddr < dev->l2ad_hand)) { 6651 /* 6652 * We've evicted to the target address, 6653 * or the end of the device. 6654 */ 6655 mutex_exit(hash_lock); 6656 break; 6657 } 6658 6659 ASSERT(HDR_HAS_L2HDR(hdr)); 6660 if (!HDR_HAS_L1HDR(hdr)) { 6661 ASSERT(!HDR_L2_READING(hdr)); 6662 /* 6663 * This doesn't exist in the ARC. Destroy. 6664 * arc_hdr_destroy() will call list_remove() 6665 * and decrement arcstat_l2_size. 6666 */ 6667 arc_change_state(arc_anon, hdr, hash_lock); 6668 arc_hdr_destroy(hdr); 6669 } else { 6670 ASSERT(hdr->b_l1hdr.b_state != arc_l2c_only); 6671 ARCSTAT_BUMP(arcstat_l2_evict_l1cached); 6672 /* 6673 * Invalidate issued or about to be issued 6674 * reads, since we may be about to write 6675 * over this location. 6676 */ 6677 if (HDR_L2_READING(hdr)) { 6678 ARCSTAT_BUMP(arcstat_l2_evict_reading); 6679 arc_hdr_set_flags(hdr, ARC_FLAG_L2_EVICTED); 6680 } 6681 6682 /* Ensure this header has finished being written */ 6683 ASSERT(!HDR_L2_WRITING(hdr)); 6684 6685 arc_hdr_l2hdr_destroy(hdr); 6686 } 6687 mutex_exit(hash_lock); 6688 } 6689 mutex_exit(&dev->l2ad_mtx); 6690 } 6691 6692 /* 6693 * Find and write ARC buffers to the L2ARC device. 6694 * 6695 * An ARC_FLAG_L2_WRITING flag is set so that the L2ARC buffers are not valid 6696 * for reading until they have completed writing. 6697 * The headroom_boost is an in-out parameter used to maintain headroom boost 6698 * state between calls to this function. 6699 * 6700 * Returns the number of bytes actually written (which may be smaller than 6701 * the delta by which the device hand has changed due to alignment). 6702 */ 6703 static uint64_t 6704 l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz) 6705 { 6706 arc_buf_hdr_t *hdr, *hdr_prev, *head; 6707 uint64_t write_asize, write_psize, write_sz, headroom; 6708 boolean_t full; 6709 l2arc_write_callback_t *cb; 6710 zio_t *pio, *wzio; 6711 uint64_t guid = spa_load_guid(spa); 6712 6713 ASSERT3P(dev->l2ad_vdev, !=, NULL); 6714 6715 pio = NULL; 6716 write_sz = write_asize = write_psize = 0; 6717 full = B_FALSE; 6718 head = kmem_cache_alloc(hdr_l2only_cache, KM_PUSHPAGE); 6719 arc_hdr_set_flags(head, ARC_FLAG_L2_WRITE_HEAD | ARC_FLAG_HAS_L2HDR); 6720 6721 /* 6722 * Copy buffers for L2ARC writing. 6723 */ 6724 for (int try = 0; try <= 3; try++) { 6725 multilist_sublist_t *mls = l2arc_sublist_lock(try); 6726 uint64_t passed_sz = 0; 6727 6728 /* 6729 * L2ARC fast warmup. 6730 * 6731 * Until the ARC is warm and starts to evict, read from the 6732 * head of the ARC lists rather than the tail. 6733 */ 6734 if (arc_warm == B_FALSE) 6735 hdr = multilist_sublist_head(mls); 6736 else 6737 hdr = multilist_sublist_tail(mls); 6738 6739 headroom = target_sz * l2arc_headroom; 6740 if (zfs_compressed_arc_enabled) 6741 headroom = (headroom * l2arc_headroom_boost) / 100; 6742 6743 for (; hdr; hdr = hdr_prev) { 6744 kmutex_t *hash_lock; 6745 6746 if (arc_warm == B_FALSE) 6747 hdr_prev = multilist_sublist_next(mls, hdr); 6748 else 6749 hdr_prev = multilist_sublist_prev(mls, hdr); 6750 6751 hash_lock = HDR_LOCK(hdr); 6752 if (!mutex_tryenter(hash_lock)) { 6753 /* 6754 * Skip this buffer rather than waiting. 6755 */ 6756 continue; 6757 } 6758 6759 passed_sz += HDR_GET_LSIZE(hdr); 6760 if (passed_sz > headroom) { 6761 /* 6762 * Searched too far. 6763 */ 6764 mutex_exit(hash_lock); 6765 break; 6766 } 6767 6768 if (!l2arc_write_eligible(guid, hdr)) { 6769 mutex_exit(hash_lock); 6770 continue; 6771 } 6772 6773 if ((write_asize + HDR_GET_LSIZE(hdr)) > target_sz) { 6774 full = B_TRUE; 6775 mutex_exit(hash_lock); 6776 break; 6777 } 6778 6779 if (pio == NULL) { 6780 /* 6781 * Insert a dummy header on the buflist so 6782 * l2arc_write_done() can find where the 6783 * write buffers begin without searching. 6784 */ 6785 mutex_enter(&dev->l2ad_mtx); 6786 list_insert_head(&dev->l2ad_buflist, head); 6787 mutex_exit(&dev->l2ad_mtx); 6788 6789 cb = kmem_alloc( 6790 sizeof (l2arc_write_callback_t), KM_SLEEP); 6791 cb->l2wcb_dev = dev; 6792 cb->l2wcb_head = head; 6793 pio = zio_root(spa, l2arc_write_done, cb, 6794 ZIO_FLAG_CANFAIL); 6795 } 6796 6797 hdr->b_l2hdr.b_dev = dev; 6798 hdr->b_l2hdr.b_daddr = dev->l2ad_hand; 6799 arc_hdr_set_flags(hdr, 6800 ARC_FLAG_L2_WRITING | ARC_FLAG_HAS_L2HDR); 6801 6802 mutex_enter(&dev->l2ad_mtx); 6803 list_insert_head(&dev->l2ad_buflist, hdr); 6804 mutex_exit(&dev->l2ad_mtx); 6805 6806 /* 6807 * We rely on the L1 portion of the header below, so 6808 * it's invalid for this header to have been evicted out 6809 * of the ghost cache, prior to being written out. The 6810 * ARC_FLAG_L2_WRITING bit ensures this won't happen. 6811 */ 6812 ASSERT(HDR_HAS_L1HDR(hdr)); 6813 6814 ASSERT3U(HDR_GET_PSIZE(hdr), >, 0); 6815 ASSERT3P(hdr->b_l1hdr.b_pdata, !=, NULL); 6816 ASSERT3U(arc_hdr_size(hdr), >, 0); 6817 uint64_t size = arc_hdr_size(hdr); 6818 6819 (void) refcount_add_many(&dev->l2ad_alloc, size, hdr); 6820 6821 /* 6822 * Normally the L2ARC can use the hdr's data, but if 6823 * we're sharing data between the hdr and one of its 6824 * bufs, L2ARC needs its own copy of the data so that 6825 * the ZIO below can't race with the buf consumer. To 6826 * ensure that this copy will be available for the 6827 * lifetime of the ZIO and be cleaned up afterwards, we 6828 * add it to the l2arc_free_on_write queue. 6829 */ 6830 void *to_write; 6831 if (!HDR_SHARED_DATA(hdr)) { 6832 to_write = hdr->b_l1hdr.b_pdata; 6833 } else { 6834 arc_buf_contents_t type = arc_buf_type(hdr); 6835 if (type == ARC_BUFC_METADATA) { 6836 to_write = zio_buf_alloc(size); 6837 } else { 6838 ASSERT3U(type, ==, ARC_BUFC_DATA); 6839 to_write = zio_data_buf_alloc(size); 6840 } 6841 6842 bcopy(hdr->b_l1hdr.b_pdata, to_write, size); 6843 l2arc_free_data_on_write(to_write, size, type); 6844 } 6845 wzio = zio_write_phys(pio, dev->l2ad_vdev, 6846 hdr->b_l2hdr.b_daddr, size, to_write, 6847 ZIO_CHECKSUM_OFF, NULL, hdr, 6848 ZIO_PRIORITY_ASYNC_WRITE, 6849 ZIO_FLAG_CANFAIL, B_FALSE); 6850 6851 write_sz += HDR_GET_LSIZE(hdr); 6852 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev, 6853 zio_t *, wzio); 6854 6855 write_asize += size; 6856 /* 6857 * Keep the clock hand suitably device-aligned. 6858 */ 6859 uint64_t asize = vdev_psize_to_asize(dev->l2ad_vdev, 6860 size); 6861 write_psize += asize; 6862 dev->l2ad_hand += asize; 6863 6864 mutex_exit(hash_lock); 6865 6866 (void) zio_nowait(wzio); 6867 } 6868 6869 multilist_sublist_unlock(mls); 6870 6871 if (full == B_TRUE) 6872 break; 6873 } 6874 6875 /* No buffers selected for writing? */ 6876 if (pio == NULL) { 6877 ASSERT0(write_sz); 6878 ASSERT(!HDR_HAS_L1HDR(head)); 6879 kmem_cache_free(hdr_l2only_cache, head); 6880 return (0); 6881 } 6882 6883 ASSERT3U(write_asize, <=, target_sz); 6884 ARCSTAT_BUMP(arcstat_l2_writes_sent); 6885 ARCSTAT_INCR(arcstat_l2_write_bytes, write_asize); 6886 ARCSTAT_INCR(arcstat_l2_size, write_sz); 6887 ARCSTAT_INCR(arcstat_l2_asize, write_asize); 6888 vdev_space_update(dev->l2ad_vdev, write_asize, 0, 0); 6889 6890 /* 6891 * Bump device hand to the device start if it is approaching the end. 6892 * l2arc_evict() will already have evicted ahead for this case. 6893 */ 6894 if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) { 6895 dev->l2ad_hand = dev->l2ad_start; 6896 dev->l2ad_first = B_FALSE; 6897 } 6898 6899 dev->l2ad_writing = B_TRUE; 6900 (void) zio_wait(pio); 6901 dev->l2ad_writing = B_FALSE; 6902 6903 return (write_asize); 6904 } 6905 6906 /* 6907 * This thread feeds the L2ARC at regular intervals. This is the beating 6908 * heart of the L2ARC. 6909 */ 6910 static void 6911 l2arc_feed_thread(void) 6912 { 6913 callb_cpr_t cpr; 6914 l2arc_dev_t *dev; 6915 spa_t *spa; 6916 uint64_t size, wrote; 6917 clock_t begin, next = ddi_get_lbolt(); 6918 6919 CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG); 6920 6921 mutex_enter(&l2arc_feed_thr_lock); 6922 6923 while (l2arc_thread_exit == 0) { 6924 CALLB_CPR_SAFE_BEGIN(&cpr); 6925 (void) cv_timedwait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock, 6926 next); 6927 CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock); 6928 next = ddi_get_lbolt() + hz; 6929 6930 /* 6931 * Quick check for L2ARC devices. 6932 */ 6933 mutex_enter(&l2arc_dev_mtx); 6934 if (l2arc_ndev == 0) { 6935 mutex_exit(&l2arc_dev_mtx); 6936 continue; 6937 } 6938 mutex_exit(&l2arc_dev_mtx); 6939 begin = ddi_get_lbolt(); 6940 6941 /* 6942 * This selects the next l2arc device to write to, and in 6943 * doing so the next spa to feed from: dev->l2ad_spa. This 6944 * will return NULL if there are now no l2arc devices or if 6945 * they are all faulted. 6946 * 6947 * If a device is returned, its spa's config lock is also 6948 * held to prevent device removal. l2arc_dev_get_next() 6949 * will grab and release l2arc_dev_mtx. 6950 */ 6951 if ((dev = l2arc_dev_get_next()) == NULL) 6952 continue; 6953 6954 spa = dev->l2ad_spa; 6955 ASSERT3P(spa, !=, NULL); 6956 6957 /* 6958 * If the pool is read-only then force the feed thread to 6959 * sleep a little longer. 6960 */ 6961 if (!spa_writeable(spa)) { 6962 next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz; 6963 spa_config_exit(spa, SCL_L2ARC, dev); 6964 continue; 6965 } 6966 6967 /* 6968 * Avoid contributing to memory pressure. 6969 */ 6970 if (arc_reclaim_needed()) { 6971 ARCSTAT_BUMP(arcstat_l2_abort_lowmem); 6972 spa_config_exit(spa, SCL_L2ARC, dev); 6973 continue; 6974 } 6975 6976 ARCSTAT_BUMP(arcstat_l2_feeds); 6977 6978 size = l2arc_write_size(); 6979 6980 /* 6981 * Evict L2ARC buffers that will be overwritten. 6982 */ 6983 l2arc_evict(dev, size, B_FALSE); 6984 6985 /* 6986 * Write ARC buffers. 6987 */ 6988 wrote = l2arc_write_buffers(spa, dev, size); 6989 6990 /* 6991 * Calculate interval between writes. 6992 */ 6993 next = l2arc_write_interval(begin, size, wrote); 6994 spa_config_exit(spa, SCL_L2ARC, dev); 6995 } 6996 6997 l2arc_thread_exit = 0; 6998 cv_broadcast(&l2arc_feed_thr_cv); 6999 CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */ 7000 thread_exit(); 7001 } 7002 7003 boolean_t 7004 l2arc_vdev_present(vdev_t *vd) 7005 { 7006 l2arc_dev_t *dev; 7007 7008 mutex_enter(&l2arc_dev_mtx); 7009 for (dev = list_head(l2arc_dev_list); dev != NULL; 7010 dev = list_next(l2arc_dev_list, dev)) { 7011 if (dev->l2ad_vdev == vd) 7012 break; 7013 } 7014 mutex_exit(&l2arc_dev_mtx); 7015 7016 return (dev != NULL); 7017 } 7018 7019 /* 7020 * Add a vdev for use by the L2ARC. By this point the spa has already 7021 * validated the vdev and opened it. 7022 */ 7023 void 7024 l2arc_add_vdev(spa_t *spa, vdev_t *vd) 7025 { 7026 l2arc_dev_t *adddev; 7027 7028 ASSERT(!l2arc_vdev_present(vd)); 7029 7030 /* 7031 * Create a new l2arc device entry. 7032 */ 7033 adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP); 7034 adddev->l2ad_spa = spa; 7035 adddev->l2ad_vdev = vd; 7036 adddev->l2ad_start = VDEV_LABEL_START_SIZE; 7037 adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd); 7038 adddev->l2ad_hand = adddev->l2ad_start; 7039 adddev->l2ad_first = B_TRUE; 7040 adddev->l2ad_writing = B_FALSE; 7041 7042 mutex_init(&adddev->l2ad_mtx, NULL, MUTEX_DEFAULT, NULL); 7043 /* 7044 * This is a list of all ARC buffers that are still valid on the 7045 * device. 7046 */ 7047 list_create(&adddev->l2ad_buflist, sizeof (arc_buf_hdr_t), 7048 offsetof(arc_buf_hdr_t, b_l2hdr.b_l2node)); 7049 7050 vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand); 7051 refcount_create(&adddev->l2ad_alloc); 7052 7053 /* 7054 * Add device to global list 7055 */ 7056 mutex_enter(&l2arc_dev_mtx); 7057 list_insert_head(l2arc_dev_list, adddev); 7058 atomic_inc_64(&l2arc_ndev); 7059 mutex_exit(&l2arc_dev_mtx); 7060 } 7061 7062 /* 7063 * Remove a vdev from the L2ARC. 7064 */ 7065 void 7066 l2arc_remove_vdev(vdev_t *vd) 7067 { 7068 l2arc_dev_t *dev, *nextdev, *remdev = NULL; 7069 7070 /* 7071 * Find the device by vdev 7072 */ 7073 mutex_enter(&l2arc_dev_mtx); 7074 for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) { 7075 nextdev = list_next(l2arc_dev_list, dev); 7076 if (vd == dev->l2ad_vdev) { 7077 remdev = dev; 7078 break; 7079 } 7080 } 7081 ASSERT3P(remdev, !=, NULL); 7082 7083 /* 7084 * Remove device from global list 7085 */ 7086 list_remove(l2arc_dev_list, remdev); 7087 l2arc_dev_last = NULL; /* may have been invalidated */ 7088 atomic_dec_64(&l2arc_ndev); 7089 mutex_exit(&l2arc_dev_mtx); 7090 7091 /* 7092 * Clear all buflists and ARC references. L2ARC device flush. 7093 */ 7094 l2arc_evict(remdev, 0, B_TRUE); 7095 list_destroy(&remdev->l2ad_buflist); 7096 mutex_destroy(&remdev->l2ad_mtx); 7097 refcount_destroy(&remdev->l2ad_alloc); 7098 kmem_free(remdev, sizeof (l2arc_dev_t)); 7099 } 7100 7101 void 7102 l2arc_init(void) 7103 { 7104 l2arc_thread_exit = 0; 7105 l2arc_ndev = 0; 7106 l2arc_writes_sent = 0; 7107 l2arc_writes_done = 0; 7108 7109 mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL); 7110 cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL); 7111 mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL); 7112 mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL); 7113 7114 l2arc_dev_list = &L2ARC_dev_list; 7115 l2arc_free_on_write = &L2ARC_free_on_write; 7116 list_create(l2arc_dev_list, sizeof (l2arc_dev_t), 7117 offsetof(l2arc_dev_t, l2ad_node)); 7118 list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t), 7119 offsetof(l2arc_data_free_t, l2df_list_node)); 7120 } 7121 7122 void 7123 l2arc_fini(void) 7124 { 7125 /* 7126 * This is called from dmu_fini(), which is called from spa_fini(); 7127 * Because of this, we can assume that all l2arc devices have 7128 * already been removed when the pools themselves were removed. 7129 */ 7130 7131 l2arc_do_free_on_write(); 7132 7133 mutex_destroy(&l2arc_feed_thr_lock); 7134 cv_destroy(&l2arc_feed_thr_cv); 7135 mutex_destroy(&l2arc_dev_mtx); 7136 mutex_destroy(&l2arc_free_on_write_mtx); 7137 7138 list_destroy(l2arc_dev_list); 7139 list_destroy(l2arc_free_on_write); 7140 } 7141 7142 void 7143 l2arc_start(void) 7144 { 7145 if (!(spa_mode_global & FWRITE)) 7146 return; 7147 7148 (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0, 7149 TS_RUN, minclsyspri); 7150 } 7151 7152 void 7153 l2arc_stop(void) 7154 { 7155 if (!(spa_mode_global & FWRITE)) 7156 return; 7157 7158 mutex_enter(&l2arc_feed_thr_lock); 7159 cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */ 7160 l2arc_thread_exit = 1; 7161 while (l2arc_thread_exit != 0) 7162 cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock); 7163 mutex_exit(&l2arc_feed_thr_lock); 7164 } 7165